unify matmul_id shader selection

* tests: allow loading test-backend-ops tests from json

* add error threshold based on op

* add error when file cannot be read

* add graph operator json extraction tool

* add nb parameter for non-contiguous input tensors

* fix view check

* only use view if non-contiguous/permuted, use C++ random instead of rand()

* replace internal API calls with public llama_graph_reserve call

* reduce test description length

* fix nb[0] not getting set for view

* add name to tests

* fix inplace error

* use text file instead of json

* move llama_graph_reserve function to new llama-ext header, move export-graph-ops to tests/

* fix missing declaration

* use pragma once

* fix indent

* fix Windows build

.github/workflows/build.yml

@@ -93,7 +93,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest -L main -E "test-llama-archs" --verbose --timeout 900
 
   macOS-latest-cmake-x64:
     runs-on: macos-15-intel
@@ -469,6 +469,7 @@ jobs:
           cd build
           export GGML_VK_VISIBLE_DEVICES=0
           export GGML_VK_DISABLE_F16=1
+          export GGML_VK_DISABLE_COOPMAT=1
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 4800
 
@@ -1726,6 +1727,22 @@ jobs:
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
+  ggml-ci-x64-linux-intel-vulkan:
+    runs-on: [self-hosted, Linux, X64, Intel]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
   ggml-ci-arm64-cpu-kleidiai:
      runs-on: ubuntu-22.04-arm
 

CODEOWNERS

@@ -11,6 +11,8 @@
 /common/base64.hpp.*                    @ggerganov
 /common/build-info.*                    @ggerganov
 /common/chat.*                          @pwilkin
+/common/chat-auto*.*                    @pwilkin
+/common/chat-diff-analyzer.*            @pwilkin
 /common/chat-peg-parser.*               @aldehir
 /common/common.*                        @ggerganov
 /common/console.*                       @ggerganov
@@ -89,12 +91,13 @@
 /src/llama-vocab.*                      @CISC
 /src/models/                            @CISC
 /tests/                                 @ggerganov
-/tests/test-chat-.*                     @pwilkin
+/tests/test-chat.*                      @pwilkin
 /tools/batched-bench/                   @ggerganov
 /tools/cli/                             @ngxson
 /tools/completion/                      @ggerganov
 /tools/mtmd/                            @ngxson
 /tools/perplexity/                      @ggerganov
+/tools/parser/                          @pwilkin
 /tools/quantize/                        @ggerganov
 /tools/rpc/                             @rgerganov
 /tools/server/*                         @ngxson @ggerganov # no subdir

CONTRIBUTING.md

@@ -39,6 +39,7 @@ Before submitting your PR:
     - For intricate features, consider opening a feature request first to discuss and align expectations
     - When adding support for a new model or feature, focus on **CPU support only** in the initial PR unless you have a good reason not to. Add support for other backends like CUDA in follow-up PRs
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
+- If you are a new contributor, limit your open PRs to 1.
 
 After submitting your PR:
 - Expect requests for modifications to ensure the code meets llama.cpp's standards for quality and long-term maintainability

README.md

@@ -259,6 +259,8 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
 - [Kalavai](https://github.com/kalavai-net/kalavai-client) - Crowdsource end to end LLM deployment at any scale
 - [llmaz](https://github.com/InftyAI/llmaz) - ☸️ Easy, advanced inference platform for large language models on Kubernetes.
+- [LLMKube](https://github.com/defilantech/llmkube) - Kubernetes operator for llama.cpp with multi-GPU and Apple Silicon Metal
+  support"
 </details>
 
 <details>
@@ -287,7 +289,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
-| [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
+| [Hexagon [In Progress]](docs/backend/snapdragon/README.md) | Snapdragon |
 | [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
 
 ## Obtaining and quantizing models

benches/nemotron/nemotron-dgx-spark.md

@@ -0,0 +1,72 @@
+# NVIDIA DGX Spark
+
+## System info
+
+```bash
+uname --all
+Linux spark-17ed 6.11.0-1016-nvidia #16-Ubuntu SMP PREEMPT_DYNAMIC Sun Sep 21 16:52:46 UTC 2025 aarch64 aarch64 aarch64 GNU/Linux
+
+g++ --version
+g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
+
+nvidia-smi
+Fri Mar  6 11:39:45 2026
++-----------------------------------------------------------------------------------------+
+| NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0     |
++-----------------------------------------+------------------------+----------------------+
+| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
+| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
+|                                         |                        |               MIG M. |
+|=========================================+========================+======================|
+|   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
+| N/A   52C    P0             13W /  N/A  | Not Supported          |      0%      Default |
+|                                         |                        |                  N/A |
++-----------------------------------------+------------------------+----------------------+
+```
+
+## ggml-org/nemotron-3-super-120b-GGUF
+
+Model: https://huggingface.co/ggml-org/nemotron-3-super-120b-GGUF
+
+- `llama-batched-bench`
+
+main: n_kv_max = 303104, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = 99, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    1.094 |   468.05 |    1.621 |    19.74 |    2.715 |   200.37 |
+|   512 |     32 |    2 |   1088 |    1.463 |   700.16 |    2.437 |    26.26 |    3.900 |   279.01 |
+|   512 |     32 |    4 |   2176 |    2.647 |   773.76 |    4.043 |    31.66 |    6.689 |   325.29 |
+|   512 |     32 |    8 |   4352 |    5.291 |   774.14 |    6.151 |    41.62 |   11.442 |   380.37 |
+|   512 |     32 |   16 |   8704 |   10.603 |   772.62 |   10.385 |    49.30 |   20.987 |   414.72 |
+|   512 |     32 |   32 |  17408 |   21.231 |   771.69 |   18.235 |    56.16 |   39.466 |   441.09 |
+|  4096 |     32 |    1 |   4128 |    5.340 |   767.05 |    1.616 |    19.81 |    6.956 |   593.47 |
+|  4096 |     32 |    2 |   8256 |   10.673 |   767.55 |    2.454 |    26.08 |   13.127 |   628.94 |
+|  4096 |     32 |    4 |  16512 |   21.348 |   767.46 |    4.072 |    31.44 |   25.420 |   649.57 |
+|  4096 |     32 |    8 |  33024 |   42.714 |   767.15 |    6.277 |    40.78 |   48.991 |   674.08 |
+|  4096 |     32 |   16 |  66048 |   85.385 |   767.54 |   10.596 |    48.32 |   95.981 |   688.14 |
+|  4096 |     32 |   32 | 132096 |  170.819 |   767.32 |   18.619 |    55.00 |  189.437 |   697.31 |
+|  8192 |     32 |    1 |   8224 |   10.690 |   766.32 |    1.619 |    19.76 |   12.310 |   668.10 |
+|  8192 |     32 |    2 |  16448 |   21.382 |   766.24 |    2.467 |    25.94 |   23.850 |   689.65 |
+|  8192 |     32 |    4 |  32896 |   42.782 |   765.92 |    4.098 |    31.23 |   46.881 |   701.69 |
+|  8192 |     32 |    8 |  65792 |   85.582 |   765.77 |    6.368 |    40.20 |   91.951 |   715.52 |
+|  8192 |     32 |   16 | 131584 |  171.066 |   766.21 |   10.774 |    47.52 |  181.840 |   723.62 |
+|  8192 |     32 |   32 | 263168 |  342.140 |   766.19 |   18.969 |    53.98 |  361.109 |   728.78 |
+
+
+- `llama-bench`
+
+| model                   |       size |     params | backend    | n_ubatch | fa |            test |                  t/s |
+| ----------------------- | ---------: | ---------: | ---------- | -------: | -: | --------------: | -------------------: |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |          pp2048 |        768.84 ± 0.90 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |            tg32 |         19.94 ± 0.16 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |  pp2048 @ d4096 |        764.51 ± 0.50 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |    tg32 @ d4096 |         19.95 ± 0.18 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |  pp2048 @ d8192 |        759.53 ± 0.71 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |    tg32 @ d8192 |         19.83 ± 0.18 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 | pp2048 @ d16384 |        747.98 ± 1.58 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |   tg32 @ d16384 |         19.84 ± 0.18 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 | pp2048 @ d32768 |        724.40 ± 2.70 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |   tg32 @ d32768 |         19.45 ± 0.18 |
+
+build: 04a65daab (8268)

common/CMakeLists.txt

@@ -47,10 +47,10 @@ add_library(${TARGET} STATIC
     arg.cpp
     arg.h
     base64.hpp
-    chat-parser.cpp
-    chat-parser.h
-    chat-parser-xml-toolcall.h
-    chat-parser-xml-toolcall.cpp
+    chat-auto-parser-generator.cpp
+    chat-auto-parser-helpers.cpp
+    chat-auto-parser.h
+    chat-diff-analyzer.cpp
     chat-peg-parser.cpp
     chat-peg-parser.h
     chat.cpp
@@ -81,6 +81,8 @@ add_library(${TARGET} STATIC
     preset.cpp
     preset.h
     regex-partial.cpp
+    reasoning-budget.cpp
+    reasoning-budget.h
     regex-partial.h
     sampling.cpp
     sampling.h

common/arg.cpp

@@ -732,23 +732,28 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
         "llama-completion",
         "llama-convert-llama2c-to-ggml",
         "llama-cvector-generator",
+        "llama-debug",
+        "llama-diffusion-cli",
         "llama-embedding",
         "llama-eval-callback",
         "llama-export-lora",
+        "llama-finetune",
+        "llama-fit-params",
+        "llama-gemma3-cli",
         "llama-gen-docs",
         "llama-gguf",
         "llama-gguf-hash",
         "llama-gguf-split",
-        "llama-gritlm",
+        "llama-idle",
         "llama-imatrix",
-        "llama-infill",
-        "llama-mtmd-cli",
-        "llama-llava-clip-quantize-cli",
+        "llama-llava-cli",
         "llama-lookahead",
         "llama-lookup",
         "llama-lookup-create",
         "llama-lookup-merge",
         "llama-lookup-stats",
+        "llama-minicpmv-cli",
+        "llama-mtmd-cli",
         "llama-parallel",
         "llama-passkey",
         "llama-perplexity",
@@ -1279,13 +1284,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
     add_opt(common_arg(
-        {"--ctx-checkpoints", "--swa-checkpoints"}, "N",
+        {"-ctxcp", "--ctx-checkpoints", "--swa-checkpoints"}, "N",
         string_format("max number of context checkpoints to create per slot (default: %d)"
             "[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_ctx_checkpoints),
         [](common_params & params, int value) {
             params.n_ctx_checkpoints = value;
         }
     ).set_env("LLAMA_ARG_CTX_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
+    add_opt(common_arg(
+        {"-cpent", "--checkpoint-every-n-tokens"}, "N",
+        string_format("create a checkpoint every n tokens during prefill (processing), -1 to disable (default: %d)", params.checkpoint_every_nt),
+        [](common_params & params, int value) {
+            params.checkpoint_every_nt = value;
+        }
+    ).set_env("LLAMA_ARG_CHECKPOINT_EVERY_NT").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"-cram", "--cache-ram"}, "N",
         string_format("set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)"
@@ -2420,11 +2432,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 );
             }
             if (split_arg.size() == 1) {
-                std::fill(params.fit_params_target.begin(), params.fit_params_target.end(), std::stoul(split_arg[0]) * 1024*1024);
+                std::fill(params.fit_params_target.begin(), params.fit_params_target.end(), std::stoull(split_arg[0]) * 1024*1024);
                 return;
             }
             for (size_t i = 0; i < split_arg.size(); i++) {
-                params.fit_params_target[i] = std::stoul(split_arg[i]) * 1024*1024;
+                params.fit_params_target[i] = std::stoull(split_arg[i]) * 1024*1024;
             }
         }
     ).set_env("LLAMA_ARG_FIT_TARGET"));
@@ -2659,7 +2671,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.out_file = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_FINETUNE}));
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_FINETUNE,
+                    LLAMA_EXAMPLE_RESULTS, LLAMA_EXAMPLE_EXPORT_GRAPH_OPS}));
     add_opt(common_arg(
         {"-ofreq", "--output-frequency"}, "N",
         string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@@ -2827,6 +2840,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.webui_config_json = read_file(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_CONFIG_FILE"));
+    add_opt(common_arg(
+        {"--webui-mcp-proxy"},
+        {"--no-webui-mcp-proxy"},
+        string_format("experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: %s)", params.webui_mcp_proxy ? "enabled" : "disabled"),
+        [](common_params & params, bool value) {
+            params.webui_mcp_proxy = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_MCP_PROXY"));
     add_opt(common_arg(
         {"--webui"},
         {"--no-webui"},
@@ -2898,6 +2919,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             auto parsed = json::parse(value);
             for (const auto & item : parsed.items()) {
+                if (item.key() == "enable_thinking") {
+                    LOG_WRN("Setting 'enable_thinking' via --chat-template-kwargs is deprecated. "
+                            "Use --reasoning on / --reasoning off instead.\n");
+                }
                 params.default_template_kwargs[item.key()] = item.value().dump();
             }
         }
@@ -3033,14 +3058,39 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.reasoning_format = common_reasoning_format_from_name(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_THINK"));
+    add_opt(common_arg(
+        {"-rea", "--reasoning"}, "[on|off|auto]",
+        "Use reasoning/thinking in the chat ('on', 'off', or 'auto', default: 'auto' (detect from template))",
+        [](common_params & params, const std::string & value) {
+            if (is_truthy(value)) {
+                params.enable_reasoning = 1;
+                params.default_template_kwargs["enable_thinking"] = "true";
+            } else if (is_falsey(value)) {
+                params.enable_reasoning = 0;
+                params.default_template_kwargs["enable_thinking"] = "false";
+            } else if (is_autoy(value)) {
+                params.enable_reasoning = -1;
+            } else {
+                throw std::invalid_argument(
+                    string_format("error: unknown value for --reasoning: '%s'\n", value.c_str()));
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING"));
     add_opt(common_arg(
         {"--reasoning-budget"}, "N",
-        "controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)",
+        "token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)",
         [](common_params & params, int value) {
-            if (value != 0 && value != -1) { throw std::invalid_argument("invalid value"); }
+            if (value < -1) { throw std::invalid_argument("invalid value"); }
             params.reasoning_budget = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_THINK_BUDGET"));
+    add_opt(common_arg(
+        {"--reasoning-budget-message"}, "MESSAGE",
+        "message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)",
+        [](common_params & params, const std::string & value) {
+            params.reasoning_budget_message = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_THINK_BUDGET_MESSAGE"));
     add_opt(common_arg(
         {"--chat-template"}, "JINJA_TEMPLATE",
         string_format(
@@ -3592,6 +3642,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+        {"--check"},
+        string_format("check rather than generate results (default: %s)", params.check ? "true" : "false"),
+        [](common_params & params) {
+            params.check = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_RESULTS}));
     add_opt(common_arg(
         {"--save-logits"},
         string_format("save final logits to files for verification (default: %s)", params.save_logits ? "true" : "false"),

common/chat-auto-parser-generator.cpp

@@ -0,0 +1,450 @@
+#include "chat-auto-parser.h"
+#include "chat-peg-parser.h"
+#include "chat.h"
+#include "common.h"
+#include "json-schema-to-grammar.h"
+#include "nlohmann/json.hpp"
+
+#include <stdexcept>
+#include <string>
+
+using json = nlohmann::ordered_json;
+
+// Helper to iterate over tools/functions
+static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
+    for (const auto & tool : tools) {
+        if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
+            continue;
+        }
+        fn(tool);
+    }
+}
+
+namespace autoparser {
+
+parser_build_context::parser_build_context(common_chat_peg_builder & p, const templates_params & inputs) :
+    p(p),
+    inputs(inputs),
+    reasoning_parser(p.eps()) {}
+
+common_chat_params peg_generator::generate_parser(const common_chat_template &    tmpl,
+                                                  const struct templates_params & inputs) {
+    // Run differential analysis to extract template structure
+    struct autoparser autoparser;
+    autoparser.analyze_template(tmpl);
+    return generate_parser(tmpl, inputs, autoparser);
+}
+
+common_chat_params peg_generator::generate_parser(const common_chat_template &    tmpl,
+                                                  const struct templates_params & inputs,
+                                                  const autoparser &              autoparser) {
+    // Build the parser using the analysis results
+    auto parser = autoparser.build_parser(inputs);
+
+    // Create the result structure
+    common_chat_params data;
+    data.prompt           = common_chat_template_direct_apply(tmpl, inputs);
+    data.format           = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.preserved_tokens = autoparser.preserved_tokens;
+    data.parser           = parser.save();
+
+    // Build grammar if tools are present
+    bool has_tools =
+        autoparser.tools.format.mode != tool_format::NONE && inputs.tools.is_array() && !inputs.tools.empty();
+    std::string trigger_marker = !autoparser.tools.format.section_start.empty() ? autoparser.tools.format.section_start :
+                                                                                  autoparser.tools.format.per_call_start;
+
+    bool has_response_format = !inputs.json_schema.empty() && inputs.json_schema.is_object();
+    bool include_grammar = has_response_format || (has_tools &&
+            ((inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO && !trigger_marker.empty()) ||
+              inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+
+    if (include_grammar) {
+        data.grammar_lazy = !has_response_format && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        // Set grammar triggers based on tool section markers (fall back to per-call markers)
+        if (data.grammar_lazy) {
+            data.grammar_triggers = {
+                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, trigger_marker }
+            };
+        }
+    }
+
+    return data;
+}
+
+common_peg_arena autoparser::build_parser(const templates_params & inputs) const {
+    if (!analysis_complete) {
+        throw std::invalid_argument("Cannot call build_parser on autoparser without performing analysis first, call analyze_template(...)");
+    }
+    return build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        // If the template uses Python dict format (single-quoted strings in JSON structures),
+        // pre-register a json-string rule that accepts both quote styles. This must happen
+        // before any call to p.json() so that all JSON parsing inherits the flexible rule.
+        if (tools.format.uses_python_dicts) {
+            p.rule("json-string", p.quoted_string());
+        }
+
+        parser_build_context ctx(p, inputs);
+        bool                 extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+        bool                 enable_thinking   = inputs.enable_thinking;
+
+        ctx.extracting_reasoning = extract_reasoning && enable_thinking && reasoning.mode != reasoning_mode::NONE;
+        ctx.content              = &content;
+
+        // Build reasoning parser
+        ctx.reasoning_parser = reasoning.build_parser(ctx);
+
+        bool has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+        bool has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
+
+        if (has_response_format) {
+            auto response_format = p.rule("response-format", p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
+            return ctx.reasoning_parser + p.space() + p.choice({
+                p.literal("```json") + p.space() + response_format + p.space() + p.literal("```"),
+                response_format
+            }) + p.end();
+        }
+
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && jinja_caps.supports_tool_calls) {
+            return tools.build_parser(ctx);
+        }
+
+        return content.build_parser(ctx);
+    });
+}
+
+common_peg_parser analyze_reasoning::build_parser(parser_build_context & ctx) const {
+    auto & p = ctx.p;
+
+    if (!ctx.extracting_reasoning) {
+        return p.eps();
+    }
+
+    bool thinking_forced_open   = (mode == reasoning_mode::FORCED_OPEN);
+    bool thinking_forced_closed = (mode == reasoning_mode::FORCED_CLOSED);
+
+    if (thinking_forced_open || thinking_forced_closed) {
+        // Thinking is forced open OR forced closed with enable_thinking=true
+        // In both cases, expect only the closing tag (opening was in template)
+        // However, since we might have incorrectly detected the open/close pattern,
+        // we admit an optional starting marker
+        return p.optional(p.literal(start)) + p.reasoning(p.until(end)) + end;
+    }
+    if (mode == reasoning_mode::TAG_BASED || mode == reasoning_mode::TOOLS_ONLY) {
+        // Standard tag-based reasoning OR tools-only mode (reasoning appears with tools)
+        // Both use the same tag-based pattern if markers are available
+        if (!start.empty() && !end.empty()) {
+            return p.optional(start + p.reasoning(p.until(end)) + end);
+        }
+    } else if (mode == reasoning_mode::DELIMITER) {
+        return p.optional(p.reasoning(p.until(end)) + end);
+    }
+
+    return p.eps();
+}
+
+common_peg_parser analyze_content::build_parser(parser_build_context & ctx) const {
+    auto & p = ctx.p;
+
+    if (is_always_wrapped()) {
+        if (ctx.extracting_reasoning) {
+            return ctx.reasoning_parser + start + p.content(p.until(end)) + end + p.end();
+        }
+        return p.content(p.until(start)) + start + p.content(p.until(end)) + end + p.end();
+    }
+    return ctx.reasoning_parser + p.content(p.rest()) + p.end();
+}
+
+common_peg_parser analyze_content::build_optional_wrapped(parser_build_context & ctx) const {
+    auto & p = ctx.p;
+
+    if (is_always_wrapped()) {
+        return p.optional(start + p.content(p.until(end)) + end);
+    }
+    return p.eps();
+}
+
+common_peg_parser analyze_tools::build_parser(parser_build_context & ctx) const {
+    switch (format.mode) {
+        case tool_format::JSON_NATIVE:
+            return build_tool_parser_json_native(ctx);
+        case tool_format::TAG_WITH_JSON:
+            return build_tool_parser_tag_json(ctx);
+        case tool_format::TAG_WITH_TAGGED:
+            return build_tool_parser_tag_tagged(ctx);
+        default:
+            GGML_ABORT("Unable to create tool parser");
+    }
+}
+
+common_peg_parser analyze_tools::build_tool_parser_json_native(parser_build_context & ctx) const {
+    auto &       p           = ctx.p;
+    const auto & inputs      = ctx.inputs;
+    bool         force_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    // Build effective field names with dot notation if function_field is set
+    std::string name_field = format.name_field;
+    std::string args_field = format.args_field;
+
+    if (!format.function_field.empty() && format.function_field != "function" &&
+        name_field.find('.') == std::string::npos) {
+        name_field = format.function_field + "." + name_field;
+        args_field = format.function_field + "." + args_field;
+    }
+
+    auto tools_parser = p.standard_json_tools(
+        format.section_start, format.section_end, inputs.tools, inputs.parallel_tool_calls,
+        inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED, name_field, args_field, format.tools_array_wrapped,
+        format.fun_name_is_key, format.id_field, format.gen_id_field, format.parameter_order);
+
+    // Handle content wrappers if present
+    if (ctx.content && ctx.content->is_always_wrapped()) {
+        auto wrapped_content = ctx.content->build_optional_wrapped(ctx);
+        return ctx.reasoning_parser + wrapped_content + tools_parser + p.end();
+    }
+
+    std::string tool_start = "{";
+    if (!format.section_start.empty()) {
+        tool_start = format.section_start;
+    } else if (!format.per_call_start.empty()) {
+        tool_start = format.per_call_start;
+    }
+
+    return ctx.reasoning_parser + (force_tools ? p.eps() : p.optional(p.content(p.until(tool_start)))) + tools_parser +
+           p.end();
+}
+
+common_peg_parser analyze_tools::build_tool_parser_tag_json(parser_build_context & ctx) const {
+    auto &       p           = ctx.p;
+    const auto & inputs      = ctx.inputs;
+    bool         force_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    common_peg_parser tool_choice = p.choice();
+
+    foreach_function(inputs.tools, [&](const json & tool) {
+        const auto & func   = tool.at("function");
+        std::string  name   = func.at("name");
+        const auto & schema = func.at("parameters");
+
+        // Build call_id parser based on position (if supported)
+        common_peg_parser call_id_section = p.eps();
+        if (call_id.pos == call_id_position::BETWEEN_FUNC_AND_ARGS && !call_id.prefix.empty() &&
+            !call_id.suffix.empty()) {
+            call_id_section = p.optional(call_id.prefix + p.tool_id(p.until(call_id.suffix))) + call_id.suffix;
+        }
+
+        auto func_parser = p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
+                           call_id_section + p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema));
+        if (!function.close.empty()) {
+            func_parser = func_parser + function.close;
+        }
+        tool_choice |= p.rule("tool-" + name, func_parser);
+    });
+
+    auto require_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    common_peg_parser tool_calls = p.eps();
+
+    if (!format.per_call_start.empty()) {
+        auto wrapped_call = format.per_call_start + tool_choice + format.per_call_end;
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call));
+        } else {
+            tool_calls = p.trigger_rule("tool-call", wrapped_call);
+        }
+        if (!format.section_start.empty()) {
+            tool_calls = p.trigger_rule("tool-calls",
+                                        p.literal(format.section_start) + p.space() + tool_calls + p.space() +
+                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end)));
+        }
+    } else {
+        std::string separator = ", ";  // Default
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call", format.section_start + tool_choice +
+                                                         p.zero_or_more(separator + tool_choice) + format.section_end);
+        } else {
+            tool_calls = p.trigger_rule("tool-call", format.section_start + tool_choice + format.section_end);
+        }
+    }
+
+    if (!require_calls) {
+        tool_calls = p.optional(tool_calls);
+    }
+
+    std::string trigger_marker       = !format.section_start.empty() ? format.section_start : format.per_call_start;
+    auto        content_before_tools = trigger_marker.empty() ? p.eps() : p.until(trigger_marker);
+    return ctx.reasoning_parser + (force_tools ? p.eps() : p.optional(p.content(content_before_tools))) + tool_calls +
+           p.end();
+}
+
+common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_context & ctx) const {
+    auto &       p           = ctx.p;
+    const auto & inputs      = ctx.inputs;
+    bool         force_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    common_peg_parser tool_choice = p.choice();
+
+    foreach_function(inputs.tools, [&](const json & tool) {
+        const auto & func   = tool.at("function");
+        std::string  name   = func.at("name");
+        const auto & params = func.at("parameters");
+
+        if (!params.contains("properties") || !params.at("properties").is_object()) {
+            return;
+        }
+
+        const auto &          properties = params.at("properties");
+        std::set<std::string> required;
+        if (params.contains("required") && params.at("required").is_array()) {
+            params.at("required").get_to(required);
+        }
+
+        // Build parser for each argument, separating required and optional
+        std::vector<common_peg_parser> required_parsers;
+        std::vector<common_peg_parser> optional_parsers;
+        for (const auto & [param_name, param_schema] : properties.items()) {
+            bool        is_required = required.find(param_name) != required.end();
+            std::string type        = "object";
+            auto        type_obj    = param_schema.contains("type") ? param_schema.at("type") : json::object();
+            if (type_obj.is_string()) {
+                type_obj.get_to(type);
+            } else if (type_obj.is_object()) {
+                if (type_obj.contains("type") && type_obj.at("type").is_string()) {
+                    type_obj.at("type").get_to(type);
+                }
+            }
+
+            auto arg = p.tool_arg(
+                p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
+                                arguments.name_suffix) +
+                arguments.value_prefix +
+                (type == "string" ? p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
+                                                                     "tool-" + name + "-arg-" + param_name + "-schema",
+                                                                     param_schema, true)) :
+                                    p.tool_arg_json_value(p.schema(
+                                        p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, format.uses_python_dicts)) +
+                                        p.space()) +
+                p.tool_arg_close(p.literal(arguments.value_suffix)));
+
+            auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
+            if (is_required) {
+                required_parsers.push_back(named_arg);
+            } else {
+                optional_parsers.push_back(named_arg);
+            }
+        }
+
+        // Build required arg sequence in definition order
+        common_peg_parser args_seq = p.eps();
+        for (size_t i = 0; i < required_parsers.size(); i++) {
+            if (i > 0) {
+                args_seq = args_seq + p.space();
+            }
+            args_seq = args_seq + required_parsers[i];
+        }
+
+        // Build optional args with flexible ordering
+        if (!optional_parsers.empty()) {
+            common_peg_parser any_opt = p.choice();
+            for (const auto & opt : optional_parsers) {
+                any_opt |= opt;
+            }
+            args_seq = args_seq + p.repeat(p.space() + any_opt, 0, (int) optional_parsers.size());
+        }
+
+        // Build call_id parser based on position (if supported)
+        common_peg_parser call_id_section = p.eps();
+        bool have_call_id = false;
+        if (call_id.pos == call_id_position::BETWEEN_FUNC_AND_ARGS && !call_id.prefix.empty() &&
+            !call_id.suffix.empty()) {
+            have_call_id = true;
+            call_id_section = p.optional(call_id.prefix + p.tool_id(p.until(call_id.suffix)) + call_id.suffix);
+        }
+
+        bool matched_atomic = false;
+        common_peg_parser func_parser = p.eps();
+        if (!function.name_suffix.empty()) {
+            func_parser = p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
+                call_id_section + p.space() + args_seq;
+            matched_atomic = true;
+        } else if (have_call_id) {
+            func_parser = p.atomic(p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
+                call_id_section) + p.space() + args_seq;
+            matched_atomic = true;
+        } else if (!arguments.name_prefix.empty() && properties.size() > 0) {
+            func_parser = p.atomic(p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
+                call_id_section + p.space() + p.peek(p.literal(arguments.name_prefix))) + args_seq;
+            matched_atomic = true;
+        } else {
+            func_parser = p.tool_open(function.name_prefix + p.tool_name(p.literal(name)) + function.name_suffix) +
+                call_id_section + p.space() + args_seq;
+        }
+
+        if (!function.close.empty()) {
+            func_parser = func_parser + p.space() + p.tool_close(p.literal(function.close));
+        } else if (!format.per_call_end.empty()) {
+            // When there's no func_close but there is a per_call_end marker, use peek() to ensure
+            // we only emit tool_close when we can actually see the closing marker. This prevents
+            // premature closing during partial parsing when we've seen e.g. "</" which could be
+            // either "</tool_call>" (end) or "<arg_key>" prefix that failed to match.
+            func_parser = func_parser + p.tool_close(p.peek(p.literal(format.per_call_end)));
+        } else {
+            func_parser =
+                func_parser + p.tool_close(p.space());  // force this to process tool closing callbacks in mapper
+        }
+        if (!matched_atomic) {
+            func_parser = p.atomic(func_parser);
+        }
+
+        tool_choice |= p.rule("tool-" + name, func_parser);
+    });
+
+    auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+    common_peg_parser tool_calls = p.eps();
+
+    if (!format.per_call_start.empty()) {
+        auto wrapped_call = format.per_call_start + p.space() + tool_choice + p.space() + format.per_call_end;
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call));
+        } else {
+            tool_calls = p.trigger_rule("tool-call", wrapped_call);
+        }
+        if (!format.section_start.empty()) {
+            tool_calls = p.trigger_rule("tool-calls",
+                                        p.literal(format.section_start) + p.space() + tool_calls + p.space() +
+                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end)));
+        }
+    } else {
+        std::string separator = ", ";  // Default
+
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call", format.section_start + p.space() + tool_choice +
+                                                         p.zero_or_more(separator + tool_choice) + p.space() +
+                                                         format.section_end);
+        } else {
+            tool_calls = p.trigger_rule(
+                "tool-call", format.section_start + p.space() + tool_choice + p.space() + format.section_end);
+        }
+    }
+
+    if (!require_tools) {
+        tool_calls = p.optional(tool_calls);
+    }
+
+    std::string trigger_marker       = !format.section_start.empty() ? format.section_start : format.per_call_start;
+    auto        content_before_tools = trigger_marker.empty() ? p.eps() : p.until(trigger_marker);
+    return ctx.reasoning_parser + (force_tools ? p.eps() : p.optional(p.content(content_before_tools))) + tool_calls +
+           p.end();
+}
+
+}  // namespace autoparser

common/chat-auto-parser-helpers.cpp

@@ -0,0 +1,347 @@
+#include "chat-auto-parser-helpers.h"
+
+#include "chat-auto-parser.h"
+#include "chat.h"
+#include "log.h"
+#include "nlohmann/json.hpp"
+
+#include <cctype>
+#include <numeric>
+
+using json = nlohmann::ordered_json;
+
+std::string trim_whitespace(const std::string & str) {
+    size_t start = 0;
+    while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {
+        start++;
+    }
+
+    if (start == str.length()) {
+        return "";
+    }
+
+    size_t end = str.length() - 1;
+    while (end > start && std::isspace(static_cast<unsigned char>(str[end]))) {
+        end--;
+    }
+
+    return str.substr(start, end - start + 1);
+}
+
+std::string trim_leading_whitespace(const std::string & str) {
+    size_t start = 0;
+    while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {
+        start++;
+    }
+
+    return str.substr(start);
+}
+
+std::string trim_trailing_whitespace(const std::string & str) {
+    if (str.empty()) {
+        return "";
+    }
+
+    size_t end = str.length() - 1;
+    while (end > 0 && std::isspace(static_cast<unsigned char>(str[end]))) {
+        end--;
+    }
+
+    // If first char is also whitespace, return empty string
+    if (end == 0 && std::isspace(static_cast<unsigned char>(str[0]))) {
+        return "";
+    }
+
+    return str.substr(0, end + 1);
+}
+
+std::string trim_trailing_newlines(const std::string & str) {
+    size_t end = str.length();
+    while (end > 0 && str[end - 1] == '\n') {
+        end--;
+    }
+
+    return str.substr(0, end);
+}
+
+static size_t common_prefix_len(const std::string & left, const std::string & right) {
+    size_t prefix_len = 0;
+    size_t min_len    = std::min(left.length(), right.length());
+    while (prefix_len < min_len && left[prefix_len] == right[prefix_len]) {
+        prefix_len++;
+    }
+    return prefix_len;
+}
+
+static size_t common_suffix_len(const std::string & left, const std::string & right) {
+    size_t suffix_len = 0;
+    size_t min_len    = std::min(left.length(), right.length());
+    while (suffix_len < min_len && left[left.length() - 1 - suffix_len] == right[right.length() - 1 - suffix_len]) {
+        suffix_len++;
+    }
+    return suffix_len;
+}
+
+diff_split calculate_diff_split(const std::string & left, const std::string & right) {
+    diff_split result;
+
+    auto left_seg = segmentize_markers(left);
+    auto right_seg = segmentize_markers(right);
+
+    if (left_seg.empty()) {
+        result.right = right;
+        return result;
+    }
+    if (right_seg.empty()) {
+        result.left = left;
+        return result;
+    }
+
+    auto left_start = left_seg.begin();
+    auto left_end = --left_seg.end();
+    auto right_start = right_seg.begin();
+    auto right_end = --right_seg.end();
+
+    auto test = [&] () {
+        return left_start != left_end && right_start != right_end;
+    };
+
+    bool left_fully_consumed = false;
+    bool right_fully_consumed = false;
+
+    while (test()) {
+        bool advanced = false;
+        if (*left_start == *right_start) {
+            result.prefix.append(left_start->value);
+            left_start++;
+            right_start++;
+            advanced = true;
+        }
+        if (*left_end == *right_end) {
+            result.suffix = left_end->value + result.suffix;
+            if (left_start != left_end) {
+                left_end--;
+            } else {
+                left_fully_consumed = true;
+            }
+            if (right_start != right_end) {
+                right_end--;
+            } else {
+                right_fully_consumed = true;
+            }
+            advanced = true;
+        }
+        if (!advanced) {
+            break;
+        }
+    }
+
+    if (left_start == left_end && right_start != right_end) {
+        if (*left_start == *right_end) {
+            result.suffix = right_end->value + result.suffix;
+            right_end--;
+            left_fully_consumed = true;
+        } else if (*left_start == *right_start) {
+            result.prefix.append(right_start->value);
+            right_start++;
+            left_fully_consumed = true;
+        }
+    } else if (right_start == right_end && left_start != left_end) {
+        if (*left_end == *right_start) {
+            result.suffix = left_end->value + result.suffix;
+            left_end--;
+            right_fully_consumed = true;
+        } else if (*left_start == *right_start) {
+            result.prefix.append(left_start->value);
+            left_start++;
+            right_fully_consumed = true;
+        }
+    } else if (left_start == left_end && right_start == right_end && *left_start == *right_start && left_start->type == segment_type::MARKER) {
+        result.prefix.append(right_start->value);
+        left_fully_consumed = true;
+        right_fully_consumed = true;
+    }
+
+    auto eat_segment = [](std::string str, const segment & seg) -> std::string { return std::move(str) + seg.value; };
+
+    bool can_have_text_suffix = left_end->type == segment_type::TEXT && right_end->type == segment_type::TEXT;
+    bool can_have_text_prefix = right_start->type == segment_type::TEXT && left_start->type == segment_type::TEXT;
+
+    std::string remainder_left = std::accumulate(left_start, left_fully_consumed ? left_end : ++left_end, std::string(), eat_segment);
+    std::string remainder_right = std::accumulate(right_start, right_fully_consumed ? right_end : ++right_end, std::string(), eat_segment);
+
+    size_t suffix_len = can_have_text_suffix ? common_suffix_len(remainder_left, remainder_right) : 0;
+    // avoid overlaps between prefix and suffix
+    size_t prefix_len = can_have_text_prefix ? common_prefix_len(remainder_left.substr(0, remainder_left.size() - suffix_len),
+        remainder_right.substr(0, remainder_right.size() - suffix_len)) : 0;
+
+    result.prefix.append(remainder_left.substr(0, prefix_len));
+    result.suffix = remainder_left.substr(remainder_left.length() - suffix_len, suffix_len) + result.suffix;
+    result.left = remainder_left.substr(prefix_len, remainder_left.length() - prefix_len - suffix_len);
+    result.right = remainder_right.substr(prefix_len, remainder_right.length() - prefix_len - suffix_len);
+
+    if (result.left == "" && result.right == "") {
+        // degenerate case, no diff
+        result.prefix = left;
+        result.suffix = "";
+        // pick prefix = all as representation
+    }
+    return result;
+}
+
+// Returns the prefix of `full` up until the first occurrence of the common prefix of `left` and `right`
+std::string until_common_prefix(const std::string & full, const std::string & left, const std::string & right) {
+    // Find the common prefix of left and right
+    size_t common_prefix_len = 0;
+    size_t min_len           = std::min(left.length(), right.length());
+    while (common_prefix_len < min_len && left[common_prefix_len] == right[common_prefix_len]) {
+        common_prefix_len++;
+    }
+
+    // If there's no common prefix, return empty string
+    if (common_prefix_len == 0) {
+        return "";
+    }
+
+    // Find the common prefix in the full string
+    std::string common_prefix = left.substr(0, common_prefix_len);
+    size_t      pos           = full.find(common_prefix);
+
+    // If not found, return empty string
+    if (pos == std::string::npos) {
+        return "";
+    }
+
+    // Return everything before the common prefix
+    return full.substr(0, pos);
+}
+
+// Returns the suffix of `full` after the last occurrence of the common suffix of `left` and `right`
+std::string after_common_suffix(const std::string & full, const std::string & left, const std::string & right) {
+    // Find the common suffix of left and right (compare from the end)
+    size_t common_suffix_len = 0;
+    size_t min_len           = std::min(left.length(), right.length());
+    while (common_suffix_len < min_len &&
+           left[left.length() - 1 - common_suffix_len] == right[right.length() - 1 - common_suffix_len]) {
+        common_suffix_len++;
+    }
+
+    // If there's no common suffix, return empty string
+    if (common_suffix_len == 0) {
+        return "";
+    }
+
+    // Extract the common suffix
+    std::string common_suffix = left.substr(left.length() - common_suffix_len);
+
+    // Find the last occurrence of the common suffix in the full string
+    size_t pos = full.rfind(common_suffix);
+
+    // If not found, return empty string
+    if (pos == std::string::npos) {
+        return "";
+    }
+
+    // Return everything after the common suffix
+    return full.substr(pos + common_suffix_len);
+}
+
+// TODO: segmentize will treat a JSON array inside tags as a tag: <calls>[{ "fun": { ... } }]</calls> will be three markers
+// not too worried about that because it hasn't turned out as a problem anywhere, but noting here in case it will
+// Might have to put some restrictions on tag contents as well (like "no { }")
+std::vector<segment> segmentize_markers(const std::string & text) {
+    std::vector<segment> retval;
+    bool in_marker = false;
+    char marker_opener = '\0';
+
+    auto is_marker_opener = [](char c) -> bool { return c == '<' || c == '['; };
+    auto is_marker_closer = [](char op, char c) -> bool { return (op == '<' && c == '>') || (op == '[' && c == ']'); };
+
+    size_t last_border = 0;
+
+    for (size_t cur_pos = 0; cur_pos < text.length(); cur_pos++) {
+        if (!in_marker && is_marker_opener(text[cur_pos])) {
+            if (last_border < cur_pos) {
+                retval.push_back(segment(segment_type::TEXT, text.substr(last_border, cur_pos - last_border)));
+            }
+            last_border = cur_pos;
+            in_marker = true;
+            marker_opener = text[cur_pos];
+        } else if (in_marker && is_marker_closer(marker_opener, text[cur_pos])) {
+            // no need to check because last_border will always be smaller
+                retval.push_back(segment(segment_type::MARKER, text.substr(last_border, cur_pos - last_border + 1)));
+            last_border = cur_pos + 1;
+            in_marker = false;
+            marker_opener = '\0';
+        }
+    }
+    if (last_border < text.length()) {
+            retval.push_back(segment(segment_type::TEXT, text.substr(last_border)));
+    }
+    return retval;
+}
+
+std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segments) {
+    std::vector<segment> result;
+    for (const auto & seg : segments) {
+        if (!trim_whitespace(seg.value).empty()) {
+            result.push_back(seg);
+        }
+    }
+    return result;
+}
+
+namespace autoparser {
+
+std::string apply_template(const common_chat_template & tmpl, const template_params & params) {
+    templates_params tmpl_params;
+    tmpl_params.messages              = params.messages;
+    tmpl_params.tools                 = params.tools;
+    tmpl_params.add_generation_prompt = params.add_generation_prompt;
+    tmpl_params.enable_thinking       = params.enable_thinking;
+
+    if (params.extra_context) {
+        tmpl_params.extra_context = *params.extra_context;
+    }
+    tmpl_params.extra_context["enable_thinking"] = params.enable_thinking;
+
+    try {
+        return common_chat_template_direct_apply(tmpl, tmpl_params);
+    } catch (const std::exception & e) {
+        LOG_DBG("Template application failed: %s\n", e.what());
+        return "";
+    }
+}
+
+std::optional<compare_variants_result> compare_variants(
+    const common_chat_template &                   tmpl,
+    const template_params &                        params_A,
+    const std::function<void(template_params &)> & params_modifier) {
+    // Create variant B by copying A
+    template_params params_B = params_A;
+
+    // Apply modifier to create variant B
+    if (params_modifier) {
+        params_modifier(params_B);
+    }
+
+    // Apply template to both variants
+    std::string output_A = apply_template(tmpl, params_A);
+    std::string output_B = apply_template(tmpl, params_B);
+
+    // Check for template application failures
+    if (output_A.empty() || output_B.empty()) {
+        return std::nullopt;
+    }
+
+    // Calculate diff and return result with both outputs
+    compare_variants_result result;
+    result.diff     = calculate_diff_split(output_A, output_B);
+    result.output_A = output_A;
+    result.output_B = output_B;
+
+    return result;
+}
+
+}  // namespace autoparser
+

common/chat-auto-parser-helpers.h

@@ -0,0 +1,73 @@
+#pragma once
+
+#include "chat-auto-parser.h"
+#include <functional>
+#include <optional>
+#include <string>
+
+std::string trim_whitespace(const std::string & str);
+std::string trim_leading_whitespace(const std::string & str);
+std::string trim_trailing_whitespace(const std::string & str);
+std::string trim_trailing_newlines(const std::string & str);
+
+// calculate a diff split (longest common prefix, longest common suffix excluding prefix,
+// mismatched part on the left, mismatched part on the right) between two strings
+// account for markers - align prefix and suffix endings so that they end on markers
+// * eg.:
+// calculate_diff_split("<html><body><div></div></body></html>", "<html><body><p>Something</p></body><html>") ->
+//  { "prefix": "<html><body>" (not: "<html><body><"), "suffix": "</body></html>", "left": "<div></div>", "right": "<p>Something</p>" }
+// calculate_diff_split("<html><body>Something</body></html>", "<html><body></body><html>") ->
+//  { "prefix": "<html><body>", "suffix": "</body></html>", "left": "Something", "right": "" }
+diff_split calculate_diff_split(const std::string & left, const std::string & right);
+
+// Returns the prefix of `full` up until the first occurrence of the common prefix of `left` and `right`
+// Returns empty string if there's no common prefix
+// * eg.:
+// until_common_prefix("really want a FUNCTION call", "FUNCTION alpha", "FUNCTION beta") -> "really want a "
+// until_common_prefix("<tool_call>", "<something>", "<something_else>") -> ""
+// until_common_prefix("some text", "1234", "abcd") -> ""
+// until_common_prefix("one arg two args three args four", "argument alpha", "argument beta") -> "one ""
+std::string until_common_prefix(const std::string & full, const std::string & left, const std::string & right);
+
+// Returns the suffix of `full` after the last occurrence of the common suffix of `left` and `right`
+// Returns empty string if there's no common suffix
+// Mirror function of `until_common_prefix`
+// * eg.:
+// after_common_suffix("really want a FUNCTION call", "first FUNCTION", "second FUNCTION") -> " call"
+// after_common_suffix("one arg two-args three args four", "alpha-args", "beta-args") -> " three args four"
+std::string after_common_suffix(const std::string & full, const std::string & left, const std::string & right);
+
+// Segmentize text into markers and non-marker fragments
+// * eg.:
+// segmentize_markers("<html><head><title>The site title</title><body><div>Here's some <b>content</b></div></body></html>" ->
+//  [ (MARKER, "<html>"), (MARKER, "<head>"), (MARKER, "<title>"), (TEXT, "The site title"), (MARKER, "</title>"),
+//    (MARKER, "<body>"), (MARKER, "<div>"), (TEXT, "Here's some "), (MARKER, "<b>"), (TEXT, "content"), (MARKER, "</b>"),
+//    (MARKER, "</div>"), (MARKER, "</body>"), (MARKER, "</html>")
+//  ]
+// segmentize_markers("<|tool_call|>[args]{ are here }[/args]<|tool_call_end|>") ->
+//  [ (MARKER, "<|tool_call|>"), (MARKER, "[args]"), (TEXT, "{ are here }"), (MARKER, "[/args]"), (MARKER, "<|tool_call_end|>") ]
+std::vector<segment> segmentize_markers(const std::string & text);
+
+// Prune whitespace-only segments from a vector of segments
+// * eg.:
+// segmentize_markers("<tool_call>\n<function=foo>\n<arg=bar>\n   \n</arg>\n</function>\n</tool_call>") ->
+//  X = [ (MARKER, "<tool_call>"), (TEXT, "\n"), (MARKER, "<function=foo>"), (TEXT, "\n"), (MARKER, "<arg=bar>"), (TEXT, "\n   \n"),
+//        (MARKER, "</arg>"), (TEXT, "\n"), (MARKER, "</function>"), (TEXT, "\n"), (MARKER, "</tool_call>") ]
+// prune_whitespace_segments(X) -> [ (MARKER, "<tool_call>"), (MARKER, "<function=foo>"), (MARKER, "<arg=bar>"), (MARKER, "</arg>"),
+//                                   (MARKER, "</function>"), (MARKER, "</tool_call>") ]
+std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segments);
+
+namespace autoparser {
+
+// Apply a template with the given parameters, returning the rendered string (empty on failure)
+std::string apply_template(const common_chat_template & tmpl, const template_params & params);
+
+// Factorized differential comparison function
+// Takes base params and a single modifier lambda to create variant B
+// Returns compare_variants_result containing diff and both outputs, or std::nullopt on failure
+std::optional<compare_variants_result> compare_variants(
+    const common_chat_template &                   tmpl,
+    const template_params &                        params_A,
+    const std::function<void(template_params &)> & params_modifier);
+
+}  // namespace autoparser

common/chat-auto-parser.h

@@ -0,0 +1,433 @@
+#pragma once
+
+#include "chat.h"
+#include "common.h"
+#include "jinja/caps.h"
+#include "peg-parser.h"
+
+#include <chrono>
+#include <optional>
+#include <string>
+#include <utility>
+#include <vector>
+
+using json = nlohmann::ordered_json;
+
+class common_chat_peg_builder;
+
+// ============================================================================
+// Parameters for template application (low-level, used by diff analysis)
+// ============================================================================
+struct template_params {
+    json                messages;
+    json                tools;
+    bool                add_generation_prompt = false;
+    bool                enable_thinking       = true;
+    std::optional<json> extra_context         = std::nullopt;
+};
+
+struct diff_split {
+    std::string prefix;
+    std::string suffix;
+    std::string left;
+    std::string right;
+
+    bool operator==(struct diff_split & other) const {
+        return prefix == other.prefix && suffix == other.suffix && left == other.left && right == other.right;
+    }
+};
+
+// Result of compare_variants containing diff and original outputs
+struct compare_variants_result {
+    diff_split  diff;
+    std::string output_A;
+    std::string output_B;
+};
+
+namespace autoparser {
+
+// ============================================================================
+// High-level params for parser generation
+// ============================================================================
+
+struct templates_params {
+    json                                  messages;
+    json                                  tools;
+    common_chat_tool_choice               tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
+    json                                  json_schema;
+    bool                                  parallel_tool_calls = true;
+    common_reasoning_format               reasoning_format    = COMMON_REASONING_FORMAT_AUTO;
+    bool                                  stream              = true;
+    std::string                           grammar;
+    bool                                  add_generation_prompt = false;
+    bool                                  enable_thinking       = true;
+    std::chrono::system_clock::time_point now                   = std::chrono::system_clock::now();
+    json                                  extra_context;
+    bool                                  add_bos       = false;
+    bool                                  add_eos       = false;
+    bool                                  is_inference  = true;
+    bool                                  add_inference = false;
+    bool                                  mark_input    = true;  // whether to mark input strings in the jinja context
+};
+
+// ============================================================================
+// Analysis Result Enums
+// ============================================================================
+
+// Reasoning handling mode (derived from R1-R3 comparisons)
+enum class reasoning_mode {
+    NONE,           // No reasoning markers detected
+    TAG_BASED,      // Standard tag-based: <think>...</think>
+    DELIMITER,      // Delimiter-based: [BEGIN FINAL RESPONSE] (reasoning ends at delimiter)
+    FORCED_OPEN,    // Template ends with open reasoning tag (empty start, non-empty end)
+    FORCED_CLOSED,  // Template ends with open reasoning tag on enabled thinking but
+                    // with both opened and closed tag for disabled thinking
+    TOOLS_ONLY      // Only reason on tool calls, not on normal content
+};
+
+inline std::ostream & operator<<(std::ostream & os, const reasoning_mode & mode) {
+    switch (mode) {
+        case reasoning_mode::NONE:
+            return os << "NONE";
+        case reasoning_mode::TAG_BASED:
+            return os << "TAG_BASED";
+        case reasoning_mode::DELIMITER:
+            return os << "DELIMITER";
+        case reasoning_mode::FORCED_OPEN:
+            return os << "FORCED_OPEN";
+        case reasoning_mode::FORCED_CLOSED:
+            return os << "FORCED_CLOSED";
+        case reasoning_mode::TOOLS_ONLY:
+            return os << "TOOLS_ONLY";
+        default:
+            return os << "UNKNOWN";
+    }
+}
+
+// Content wrapping mode (derived from C1 comparison)
+enum class content_mode {
+    PLAIN,                   // No content markers
+    ALWAYS_WRAPPED,          // Content always wrapped with markers
+    WRAPPED_WITH_REASONING,  // Content wrapped only when reasoning present
+};
+
+inline std::ostream & operator<<(std::ostream & os, const content_mode & mode) {
+    switch (mode) {
+        case content_mode::PLAIN:
+            return os << "PLAIN";
+        case content_mode::ALWAYS_WRAPPED:
+            return os << "ALWAYS_WRAPPED";
+        case content_mode::WRAPPED_WITH_REASONING:
+            return os << "WRAPPED_WITH_REASONING";
+        default:
+            return os << "UNKNOWN";
+    }
+}
+
+// Call ID position in tool calls (for non-JSON formats)
+enum class call_id_position {
+    NONE,                   // No call ID support detected
+    PRE_FUNC_NAME,          // Call ID before function name: [CALL_ID]id[FUNC]name{args}
+    BETWEEN_FUNC_AND_ARGS,  // Call ID between function and args: [FUNC]name[CALL_ID]id{args}
+    POST_ARGS,              // Call ID after arguments: [FUNC]name{args}[CALL_ID]id
+};
+
+inline std::ostream & operator<<(std::ostream & os, const call_id_position & pos) {
+    switch (pos) {
+        case call_id_position::NONE:
+            return os << "NONE";
+        case call_id_position::PRE_FUNC_NAME:
+            return os << "PRE_FUNC_NAME";
+        case call_id_position::BETWEEN_FUNC_AND_ARGS:
+            return os << "BETWEEN_FUNC_AND_ARGS";
+        case call_id_position::POST_ARGS:
+            return os << "POST_ARGS";
+        default:
+            return os << "UNKNOWN";
+    }
+}
+
+// Tool call format classification (derived from T1-T5, A1-A3 comparisons)
+enum class tool_format {
+    NONE,             // No tool support detected
+    JSON_NATIVE,      // Pure JSON: {"name": "X", "arguments": {...}}
+    TAG_WITH_JSON,    // Tag-based with JSON args: <function=X>{...}</function>
+    TAG_WITH_TAGGED,  // Tag-based with tagged args: <param=key>value</param>
+};
+
+inline std::ostream & operator<<(std::ostream & os, const tool_format & format) {
+    switch (format) {
+        case tool_format::NONE:
+            return os << "NONE";
+        case tool_format::JSON_NATIVE:
+            return os << "JSON_NATIVE";
+        case tool_format::TAG_WITH_JSON:
+            return os << "TAG_WITH_JSON";
+        case tool_format::TAG_WITH_TAGGED:
+            return os << "TAG_WITH_TAGGED";
+        default:
+            return os << "UNKNOWN";
+    }
+}
+
+// ============================================================================
+// Sub-structs for tool analysis
+// ============================================================================
+
+struct tool_format_analysis {
+    tool_format mode = tool_format::NONE;
+
+    std::string section_start;   // e.g., "<tool_call>", "[TOOL_CALLS]", ""
+    std::string section_end;     // e.g., "</tool_call>", ""
+    std::string per_call_start;  // e.g., "<|tool_call_begin|>", "" (for multi-call templates)
+    std::string per_call_end;    // e.g., "<|tool_call_end|>", ""
+
+    bool fun_name_is_key = false;       // In JSON format function name is JSON key, i.e. { "<funname>": { ... arguments ... } }
+    bool tools_array_wrapped = false;   // Tool calls wrapped in JSON array [...]
+    bool uses_python_dicts = false;     // Tool call args use Python dict format (single-quoted strings)
+
+    std::string              function_field = "function";
+    std::string              name_field     = "name";
+    std::string              args_field     = "arguments";
+    std::string              id_field;
+    std::string              gen_id_field;
+    std::vector<std::string> parameter_order;
+};
+
+struct tool_function_analysis {
+    std::string name_prefix;  // e.g., "<function=", "\"name\": \"", "functions."
+    std::string name_suffix;  // e.g., ">", "\"", ":0"
+    std::string close;        // e.g., "</function>", "" (for tag-based)
+};
+
+struct tool_arguments_analysis {
+    std::string start;          // e.g., "<|tool_call_argument_begin|>", "<args>"
+    std::string end;            // e.g., "<|tool_call_argument_end|>", "</args>"
+    std::string name_prefix;   // e.g., "<param=", "<arg_key>", "\""
+    std::string name_suffix;   // e.g., ">", "</arg_key>", "\":"
+    std::string value_prefix;  // e.g., "", "<arg_value>", ""
+    std::string value_suffix;  // e.g., "</param>", "</arg_value>", ""
+    std::string separator;     // e.g., "", "\n", ","
+};
+
+struct tool_id_analysis {
+    call_id_position pos = call_id_position::NONE;
+
+    std::string prefix;  // e.g., "[CALL_ID]" (marker before call ID value)
+    std::string suffix;  // e.g., "" (marker after call ID value, before next section)
+};
+
+// ============================================================================
+// Parser build context (shared interface for build_parser methods)
+// ============================================================================
+
+struct analyze_content;
+
+struct parser_build_context {
+    common_chat_peg_builder & p;
+    const templates_params &          inputs;
+    common_peg_parser                 reasoning_parser;
+    bool                              extracting_reasoning = false;
+    const analyze_content *           content              = nullptr;
+
+    parser_build_context(common_chat_peg_builder & p, const templates_params & inputs);
+};
+
+// ============================================================================
+// Base class for analyzers with parser building
+// ============================================================================
+
+struct analyze_base {
+    virtual ~analyze_base() = default;
+    virtual common_peg_parser build_parser(parser_build_context & ctx) const = 0;
+
+  protected:
+    const common_chat_template * tmpl = nullptr;
+
+    analyze_base() = default;
+    explicit analyze_base(const common_chat_template & tmpl) : tmpl(&tmpl) {}
+};
+
+// ============================================================================
+// Reasoning analyzer
+// ============================================================================
+
+struct analyze_reasoning : analyze_base {
+    reasoning_mode mode = reasoning_mode::NONE;
+
+    std::string start;  // e.g., "<think>", "[THINK]", "<|START_THINKING|>", ""
+    std::string end;    // e.g., "</think>", "[BEGIN FINAL RESPONSE]", "<|END_THINKING|>"
+
+    analyze_reasoning() = default;
+    analyze_reasoning(const common_chat_template & tmpl, bool supports_tools);
+
+    common_peg_parser build_parser(parser_build_context & ctx) const override;
+
+  private:
+    // Look for reasoning markers in rendered content
+    void compare_reasoning_presence();
+
+    // Compare generation prompt with enable_thinking=true vs false
+    void compare_thinking_enabled();
+
+    // Check if reasoning is always possible or only in tool calls
+    void compare_reasoning_scope();
+};
+
+// ============================================================================
+// Content analyzer
+// ============================================================================
+
+struct analyze_content : analyze_base {
+    content_mode mode = content_mode::PLAIN;
+
+    std::string start;  // e.g., "<response>", ">>>all\n", ""
+    std::string end;    // e.g., "</response>", ""
+
+    bool requires_nonnull_content = false;
+
+    analyze_content() = default;
+    analyze_content(const common_chat_template & tmpl, const analyze_reasoning & reasoning);
+
+    common_peg_parser build_parser(parser_build_context & ctx) const override;
+
+    bool is_always_wrapped() const;
+    common_peg_parser build_optional_wrapped(parser_build_context & ctx) const;
+};
+
+// ============================================================================
+// Tool analyzer
+// ============================================================================
+
+struct analyze_tools : analyze_base {
+    tool_format_analysis    format;
+    tool_function_analysis  function;
+    tool_arguments_analysis arguments;
+    tool_id_analysis        call_id;
+
+    analyze_tools() = default;
+    analyze_tools(const common_chat_template & tmpl,
+                  const jinja::caps &          caps,
+                  const analyze_reasoning &    reasoning);
+
+    common_peg_parser build_parser(parser_build_context & ctx) const override;
+
+  private:
+    // Extract tool calling 'haystack' for further analysis and delegate further analysis based on format
+    void analyze_tool_calls(const analyze_reasoning & reasoning);
+
+    // Analyze format based on position of function and argument name in needle
+    void analyze_tool_call_format(const std::string &       haystack,
+                                  const std::string &       fun_name_needle,
+                                  const std::string &       arg_name_needle,
+                                  const analyze_reasoning & reasoning);
+
+    // Analyze specifics of JSON native format (entire tool call is a JSON object)
+    void analyze_tool_call_format_json_native(const std::string & clean_haystack,
+                                              const std::string & fun_name_needle,
+                                              const std::string & arg_name_needle);
+
+    // Analyze specifics of non-JSON native format (tags for function name or for function name and arguments)
+    void analyze_tool_call_format_non_json(const std::string & clean_haystack,
+                                           const std::string & fun_name_needle);
+
+    // Check for and extract specific per-call markers for non-native-JSON templates with parallel call support
+    void check_per_call_markers();
+
+    // Extract function name markers
+    void extract_function_markers();
+
+    // Delegates to separate functions for: separator analysis, argument name analysis, argument value analysis
+    void analyze_arguments();
+
+    // Extract argument name markers
+    void extract_argument_name_markers();
+
+    // Extract argument value markers
+    void extract_argument_value_markers();
+
+    // Extract argument separator, if specified (eg. <arg=foo>...</arg><sep><arg=bar>...</arg>)
+    void extract_argument_separator();
+
+    // Extract argument wrapper markers, if present (eg. '<args><arg=foo>...</arg><arg=bar>...</arg></args>')
+    void extract_args_markers();
+
+    // Extract call ID markers, if present
+    void extract_call_id_markers();
+
+    // Per-format tool parser builders
+    common_peg_parser build_tool_parser_json_native(parser_build_context & ctx) const;
+    common_peg_parser build_tool_parser_tag_json(parser_build_context & ctx) const;
+    common_peg_parser build_tool_parser_tag_tagged(parser_build_context & ctx) const;
+};
+
+// ============================================================================
+// Main autoparser class
+// ============================================================================
+
+struct autoparser {
+    jinja::caps          jinja_caps;
+    analyze_reasoning    reasoning;
+    analyze_content      content;
+    analyze_tools        tools;
+    bool                 analysis_complete = false;
+
+    // Preserved tokens for tokenizer (union of all non-empty markers)
+    std::vector<std::string> preserved_tokens;
+
+    autoparser() = default;
+
+    // Run full differential analysis on a template
+    void analyze_template(const common_chat_template & tmpl);
+
+    // Build the PEG parser for this template
+    common_peg_arena build_parser(const templates_params & inputs) const;
+
+  private:
+    // Collect tokens from entire analysis to preserve
+    void collect_preserved_tokens();
+};
+
+// ============================================================================
+// Parser generator
+// ============================================================================
+
+class peg_generator {
+  public:
+    static common_chat_params generate_parser(const common_chat_template &    tmpl,
+                                              const struct templates_params & inputs);
+
+    static common_chat_params generate_parser(const common_chat_template &    tmpl,
+                                              const struct templates_params & inputs,
+                                              const autoparser &              autoparser);
+};
+
+}  // namespace autoparser
+
+enum segment_type { TEXT, MARKER };
+
+inline std::ostream & operator<<(std::ostream & os, const segment_type & type) {
+    switch (type) {
+        case segment_type::TEXT:
+            return os << "TEXT";
+        case segment_type::MARKER:
+            return os << "MARKER";
+        default:
+            return os << "UNKNOWN";
+    }
+}
+
+struct segment {
+    segment_type type;
+    std::string  value;
+
+    segment(segment_type type, std::string value) : type(type), value(std::move(value)) {}
+
+    bool operator==(const segment & other) const {
+        return type == other.type && value == other.value;
+    }
+
+    bool operator!=(const segment & other) const {
+        return !(*this == other);
+    }
+};

common/chat-parser-xml-toolcall.cpp

@@ -1,879 +0,0 @@
-#include "chat.h"
-#include "chat-parser.h"
-#include "common.h"
-#include "json-partial.h"
-#include "json-schema-to-grammar.h"
-#include "log.h"
-#include "regex-partial.h"
-
-using json = nlohmann::ordered_json;
-
-class xml_toolcall_syntax_exception : public std::runtime_error {
-  public:
-    xml_toolcall_syntax_exception(const std::string & message) : std::runtime_error(message) {}
-};
-
-template<typename T>
-inline void sort_uniq(std::vector<T> &vec) {
-    std::sort(vec.begin(), vec.end());
-    vec.erase(std::unique(vec.begin(), vec.end()), vec.end());
-}
-
-template<typename T>
-inline bool all_space(const T &str) {
-    return std::all_of(str.begin(), str.end(), [](unsigned char ch) { return std::isspace(ch); });
-}
-
-static size_t utf8_truncate_safe(const std::string_view s) {
-    size_t len = s.size();
-    if (len == 0) return 0;
-    size_t i = len;
-    for (size_t back = 0; back < 4 && i > 0; ++back) {
-        --i;
-        unsigned char c = s[i];
-        if ((c & 0x80) == 0) {
-            return len;
-        } else if ((c & 0xC0) == 0xC0) {
-            size_t expected_len = 0;
-            if ((c & 0xE0) == 0xC0) expected_len = 2;
-            else if ((c & 0xF0) == 0xE0) expected_len = 3;
-            else if ((c & 0xF8) == 0xF0) expected_len = 4;
-            else return i;
-            if (len - i >= expected_len) {
-                return len;
-            } else {
-                return i;
-            }
-        }
-    }
-    return len - std::min(len, size_t(3));
-}
-
-inline void utf8_truncate_safe_resize(std::string &s) {
-    s.resize(utf8_truncate_safe(s));
-}
-
-inline std::string_view utf8_truncate_safe_view(const std::string_view s) {
-    return s.substr(0, utf8_truncate_safe(s));
-}
-
-static std::optional<common_chat_msg_parser::find_regex_result> try_find_2_literal_splited_by_spaces(common_chat_msg_parser & builder, const std::string & literal1, const std::string & literal2) {
-    if (literal1.size() == 0) return builder.try_find_literal(literal2);
-    const auto saved_pos = builder.pos();
-    while (auto res = builder.try_find_literal(literal1)) {
-        builder.consume_spaces();
-        const auto match_len = std::min(literal2.size(), builder.input().size() - builder.pos());
-        if (builder.input().compare(builder.pos(), match_len, literal2, 0, match_len) == 0) {
-            if (res->prelude.size() != res->groups[0].begin - saved_pos) {
-                res->prelude = builder.str({saved_pos, res->groups[0].begin});
-            }
-            builder.move_to(builder.pos() + match_len);
-            res->groups[0].end = builder.pos();
-            GGML_ASSERT(res->groups[0].begin != res->groups[0].end);
-            return res;
-        }
-        builder.move_to(res->groups[0].begin + 1);
-    }
-    builder.move_to(saved_pos);
-    return std::nullopt;
-}
-
-/**
- * make a GBNF that accept any strings except those containing any of the forbidden strings.
- */
-std::string make_gbnf_excluding(std::vector<std::string> forbids) {
-    constexpr auto charclass_escape = [](unsigned char c) -> std::string {
-        if (c == '\\' || c == ']' || c == '^' || c == '-') {
-            std::string s = "\\";
-            s.push_back((char)c);
-            return s;
-        }
-        if (isprint(c)) {
-            return std::string(1, (char)c);
-        }
-        char buf[16];
-        snprintf(buf, 15, "\\x%02X", c);
-        return std::string(buf);
-    };
-    constexpr auto build_expr = [charclass_escape](auto self, const std::vector<std::string>& forbids, int l, int r, int depth) -> std::string {
-        std::vector<std::pair<unsigned char, std::pair<int,int>>> children;
-        int i = l;
-        while (i < r) {
-            const std::string &s = forbids[i];
-            if ((int)s.size() == depth) {
-                ++i;
-                continue;
-            }
-            unsigned char c = (unsigned char)s[depth];
-            int j = i;
-            while (j < r && (int)forbids[j].size() > depth &&
-                   (unsigned char)forbids[j][depth] == c) {
-                ++j;
-            }
-            children.push_back({c, {i, j}});
-            i = j;
-        }
-        std::vector<std::string> alts;
-        if (!children.empty()) {
-            std::string cls;
-            for (auto &ch : children) cls += charclass_escape(ch.first);
-            alts.push_back(std::string("[^") + cls + "]");
-        }
-        for (auto &ch : children) {
-            std::string childExpr = self(self, forbids, ch.second.first, ch.second.second, depth+1);
-            if (!childExpr.empty()) {
-                std::string quoted_ch = "\"";
-                if (ch.first == '\\') quoted_ch += "\\\\";
-                else if (ch.first == '"') quoted_ch += "\\\"";
-                else if (isprint(ch.first)) quoted_ch.push_back(ch.first);
-                else {
-                    char buf[16];
-                    snprintf(buf, 15, "\\x%02X", ch.first);
-                    quoted_ch += buf;
-                }
-                quoted_ch += "\"";
-                std::string branch = quoted_ch + std::string(" ") + childExpr;
-                alts.push_back(branch);
-            }
-        }
-        if (alts.empty()) return "";
-        std::ostringstream oss;
-        oss << "( ";
-        for (size_t k = 0; k < alts.size(); ++k) {
-            if (k) oss << " | ";
-            oss << alts[k];
-        }
-        oss << " )";
-        return oss.str();
-    };
-    if (forbids.empty()) return "( . )*";
-    sort(forbids.begin(), forbids.end());
-    std::string expr = build_expr(build_expr, forbids, 0, forbids.size(), 0);
-    if (expr.empty()) {
-        std::string cls;
-        for (auto &s : forbids) if (!s.empty()) cls += charclass_escape((unsigned char)s[0]);
-        expr = std::string("( [^") + cls + "] )";
-    }
-    if (forbids.size() == 1)
-        return expr + "*";
-    else
-        return std::string("( ") + expr + " )*";
-}
-
-/**
- * Build grammar for xml-style tool call
- * form.scope_start and form.scope_end can be empty.
- * Requires data.format for model-specific hacks.
- */
-void build_grammar_xml_tool_call(common_chat_params & data, const json & tools, const struct xml_tool_call_format & form) {
-    GGML_ASSERT(!form.tool_start.empty());
-    GGML_ASSERT(!form.tool_sep.empty());
-    GGML_ASSERT(!form.key_start.empty());
-    GGML_ASSERT(!form.val_end.empty());
-    GGML_ASSERT(!form.tool_end.empty());
-
-    std::string key_val_sep = form.key_val_sep;
-    if (form.key_val_sep2) {
-        key_val_sep += "\n";
-        key_val_sep += *form.key_val_sep2;
-    }
-    GGML_ASSERT(!key_val_sep.empty());
-
-    if (tools.is_array() && !tools.empty()) {
-        data.grammar = build_grammar([&](const common_grammar_builder &builder) {
-            auto string_arg_val = form.last_val_end ?
-                    builder.add_rule("string-arg-val", make_gbnf_excluding({form.val_end, *form.last_val_end})) :
-                    builder.add_rule("string-arg-val", make_gbnf_excluding({form.val_end}));
-
-            std::vector<std::string> tool_rules;
-            for (const auto & tool : tools) {
-                if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
-                    LOG_WRN("Skipping tool without function: %s", tool.dump(2).c_str());
-                    continue;
-                }
-                const auto & function = tool.at("function");
-                if (!function.contains("name") || !function.at("name").is_string()) {
-                    LOG_WRN("Skipping invalid function (invalid name): %s", function.dump(2).c_str());
-                    continue;
-                }
-                if (!function.contains("parameters") || !function.at("parameters").is_object()) {
-                    LOG_WRN("Skipping invalid function (invalid parameters): %s", function.dump(2).c_str());
-                    continue;
-                }
-                std::string name = function.at("name");
-                auto parameters = function.at("parameters");
-                builder.resolve_refs(parameters);
-
-                struct parameter_rule {
-                    std::string symbol_name;
-                    bool is_required;
-                };
-                std::vector<parameter_rule> arg_rules;
-                if (!parameters.contains("properties") || !parameters.at("properties").is_object()) {
-                    LOG_WRN("Skipping invalid function (invalid properties): %s", function.dump(2).c_str());
-                    continue;
-                } else {
-                    std::vector<std::string> requiredParameters;
-                    if (parameters.contains("required")) {
-                        try { parameters.at("required").get_to(requiredParameters); }
-                        catch (const std::runtime_error&) {
-                            LOG_WRN("Invalid function required parameters, ignoring: %s", function.at("required").dump(2).c_str());
-                        }
-                    }
-                    sort_uniq(requiredParameters);
-                    for (const auto & [key, value] : parameters.at("properties").items()) {
-                        std::string quoted_key = key;
-                        bool required = std::binary_search(requiredParameters.begin(), requiredParameters.end(), key);
-                        if (form.key_start.back() == '"' && key_val_sep[0] == '"') {
-                            quoted_key = gbnf_format_literal(key);
-                            quoted_key = quoted_key.substr(1, quoted_key.size() - 2);
-                        }
-                        arg_rules.push_back(parameter_rule {builder.add_rule("func-" + name + "-kv-" + key,
-                            gbnf_format_literal(form.key_start) + " " +
-                            gbnf_format_literal(quoted_key) + " " +
-                            gbnf_format_literal(key_val_sep) + " " +
-                            ((value.contains("type") && value["type"].is_string() && value["type"] == "string" && (!form.raw_argval || *form.raw_argval)) ?
-                                    (form.raw_argval ?
-                                            string_arg_val :
-                                            "( " + string_arg_val + " | " + builder.add_schema(name + "-arg-" + key, value) + " )"
-                                    ) :
-                                    builder.add_schema(name + "-arg-" + key, value)
-                            )
-                        ), required});
-                    }
-                }
-
-                auto next_arg_with_sep = builder.add_rule(name + "-last-arg-end", form.last_val_end ? gbnf_format_literal(*form.last_val_end) : gbnf_format_literal(form.val_end));
-                decltype(next_arg_with_sep) next_arg = "\"\"";
-                for (auto i = arg_rules.size() - 1; /* i >= 0 && */ i < arg_rules.size(); --i) {
-                    std::string include_this_arg = arg_rules[i].symbol_name + " " + next_arg_with_sep;
-                    next_arg = builder.add_rule(name + "-arg-after-" + std::to_string(i), arg_rules[i].is_required ?
-                            include_this_arg : "( " + include_this_arg + " ) | " + next_arg
-                    );
-                    include_this_arg = gbnf_format_literal(form.val_end) + " " + include_this_arg;
-                    next_arg_with_sep = builder.add_rule(name + "-arg-after-" + std::to_string(i) + "-with-sep", arg_rules[i].is_required ?
-                            include_this_arg : "( " + include_this_arg + " ) | " + next_arg_with_sep
-                    );
-                }
-
-                std::string quoted_name = name;
-                if (form.tool_start.back() == '"' && form.tool_sep[0] == '"') {
-                    quoted_name = gbnf_format_literal(name);
-                    quoted_name = quoted_name.substr(1, quoted_name.size() - 2);
-                }
-                quoted_name = gbnf_format_literal(quoted_name);
-                // Kimi-K2 uses functions.{{ tool_call['function']['name'] }}:{{ loop.index }} as function name
-                if (data.format == COMMON_CHAT_FORMAT_KIMI_K2) {
-                    quoted_name = "\"functions.\" " + quoted_name + " \":\" [0-9]+";
-                }
-                tool_rules.push_back(builder.add_rule(name + "-call",
-                        gbnf_format_literal(form.tool_start) + " " +
-                        quoted_name + " " +
-                        gbnf_format_literal(form.tool_sep) + " " +
-                        next_arg
-                ));
-            }
-
-            auto tool_call_once = builder.add_rule("root-tool-call-once", string_join(tool_rules, " | "));
-            auto tool_call_more = builder.add_rule("root-tool-call-more", gbnf_format_literal(form.tool_end) + " " + tool_call_once);
-            auto call_end = builder.add_rule("root-call-end", form.last_tool_end ? gbnf_format_literal(*form.last_tool_end) : gbnf_format_literal(form.tool_end));
-            auto tool_call_multiple_with_end = builder.add_rule("root-tool-call-multiple-with-end", tool_call_once + " " + tool_call_more + "* " + call_end);
-            builder.add_rule("root",
-                (form.scope_start.empty() ? "" : gbnf_format_literal(form.scope_start) + " ") +
-                tool_call_multiple_with_end  + "?" +
-                (form.scope_end.empty() ? "" : " " + gbnf_format_literal(form.scope_end))
-            );
-        });
-
-        // grammar trigger for tool call
-        data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, form.scope_start + form.tool_start });
-    }
-}
-
-/**
- * Parse XML-Style tool call for given xml_tool_call_format. Return false for invalid syntax and get the position untouched.
- * Throws xml_toolcall_syntax_exception if there is invalid syntax and cannot recover the original status for common_chat_msg_parser.
- * form.scope_start, form.tool_sep and form.scope_end can be empty.
- */
-inline bool parse_xml_tool_calls(common_chat_msg_parser & builder, const struct xml_tool_call_format & form) {
-    GGML_ASSERT(!form.tool_start.empty());
-    GGML_ASSERT(!form.key_start.empty());
-    GGML_ASSERT(!form.key_val_sep.empty());
-    GGML_ASSERT(!form.val_end.empty());
-    GGML_ASSERT(!form.tool_end.empty());
-
-    // Helper to choose return false or throw error
-    constexpr auto return_error = [](common_chat_msg_parser & builder, auto &start_pos, const bool &recovery) {
-        LOG_DBG("Failed to parse XML-Style tool call at position: %s\n", gbnf_format_literal(builder.consume_rest().substr(0, 20)).c_str());
-        if (recovery) {
-            builder.move_to(start_pos);
-            return false;
-        } else throw xml_toolcall_syntax_exception("Tool call parsing failed with unrecoverable errors. Try using a grammar to constrain the model’s output.");
-    };
-    // Drop substring from needle to end from a JSON
-    constexpr auto partial_json = [](std::string &json_str, std::string_view needle = "XML_TOOL_CALL_PARTIAL_FLAG") {
-        auto pos = json_str.rfind(needle);
-        if (pos == std::string::npos) {
-            return false;
-        }
-        for (auto i = pos + needle.size(); i < json_str.size(); ++i) {
-            unsigned char ch = static_cast<unsigned char>(json_str[i]);
-            if (ch != '\'' && ch != '"' && ch != '}' && ch != ':' && !std::isspace(ch)) {
-                return false;
-            }
-        }
-        if (pos != 0 && json_str[pos - 1] == '"') {
-            --pos;
-        }
-        json_str.resize(pos);
-        return true;
-    };
-    // Helper to generate a partial argument JSON
-    constexpr auto gen_partial_json = [partial_json](auto set_partial_arg, auto &arguments, auto &builder, auto &function_name) {
-        auto rest = builder.consume_rest();
-        utf8_truncate_safe_resize(rest);
-        set_partial_arg(rest, "XML_TOOL_CALL_PARTIAL_FLAG");
-        auto tool_str = arguments.dump();
-        if (partial_json(tool_str)) {
-            if (builder.add_tool_call(function_name, "", tool_str)) {
-                return;
-            }
-        }
-        LOG_DBG("Failed to parse partial XML-Style tool call, fallback to non-partial: %s\n", tool_str.c_str());
-    };
-    // Helper to find a close (because there may be form.last_val_end or form.last_tool_end)
-    constexpr auto try_find_close = [](
-            common_chat_msg_parser & builder,
-            const std::string & end,
-            const std::optional<std::string> & alt_end,
-            const std::string & end_next,
-            const std::optional<std::string> & alt_end_next
-    ) {
-        auto saved_pos = builder.pos();
-        auto tc = builder.try_find_literal(end);
-        auto val_end_size = end.size();
-        if (alt_end) {
-            auto pos_1 = builder.pos();
-            builder.move_to(saved_pos);
-            auto tc2 = try_find_2_literal_splited_by_spaces(builder, *alt_end, end_next);
-            if (alt_end_next) {
-                builder.move_to(saved_pos);
-                auto tc3 = try_find_2_literal_splited_by_spaces(builder, *alt_end, *alt_end_next);
-                if (tc3 && (!tc2 || tc2->prelude.size() > tc3->prelude.size())) {
-                    tc2 = tc3;
-                }
-            }
-            if (tc2 && (!tc || tc->prelude.size() > tc2->prelude.size())) {
-                tc = tc2;
-                tc->groups[0].end = std::min(builder.input().size(), tc->groups[0].begin + alt_end->size());
-                builder.move_to(tc->groups[0].end);
-                val_end_size = alt_end->size();
-            } else {
-                builder.move_to(pos_1);
-            }
-        }
-        return std::make_pair(val_end_size, tc);
-    };
-    // Helper to find a val_end or last_val_end, returns matched pattern size
-    const auto try_find_val_end = [try_find_close, &builder, &form]() {
-        return try_find_close(builder, form.val_end, form.last_val_end, form.tool_end, form.last_tool_end);
-    };
-    // Helper to find a tool_end or last_tool_end, returns matched pattern size
-    const auto try_find_tool_end = [try_find_close, &builder, &form]() {
-        return try_find_close(builder, form.tool_end, form.last_tool_end, form.scope_end, std::nullopt);
-    };
-
-    bool recovery = true;
-    const auto start_pos = builder.pos();
-    if (!all_space(form.scope_start)) {
-        if (auto tc = builder.try_find_literal(form.scope_start)) {
-            if (all_space(tc->prelude)) {
-                if (form.scope_start.size() != tc->groups[0].end - tc->groups[0].begin)
-                    throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(form.scope_start));
-            } else {
-                builder.move_to(start_pos);
-                return false;
-            }
-        } else return false;
-    }
-    while (auto tc = builder.try_find_literal(form.tool_start)) {
-        if (!all_space(tc->prelude)) {
-            LOG_DBG("XML-Style tool call: Expected %s, but found %s, trying to match next pattern\n",
-                    gbnf_format_literal(form.tool_start).c_str(),
-                    gbnf_format_literal(tc->prelude).c_str()
-            );
-            builder.move_to(tc->groups[0].begin - tc->prelude.size());
-            break;
-        }
-
-        // Find tool name
-        auto func_name = builder.try_find_literal(all_space(form.tool_sep) ? form.key_start : form.tool_sep);
-        if (!func_name) {
-            auto [sz, tc] = try_find_tool_end();
-            func_name = tc;
-        }
-        if (!func_name) {
-            // Partial tool name not supported
-            throw common_chat_msg_partial_exception("incomplete tool_call");
-        }
-        // If the model generate multiple tool call and the first tool call has no argument
-        if (func_name->prelude.find(form.tool_end) != std::string::npos || (form.last_tool_end ? func_name->prelude.find(*form.last_tool_end) != std::string::npos : false)) {
-            builder.move_to(func_name->groups[0].begin - func_name->prelude.size());
-            auto [sz, tc] = try_find_tool_end();
-            func_name = tc;
-        }
-
-        // Parse tool name
-        builder.move_to(all_space(form.tool_sep) ? func_name->groups[0].begin : func_name->groups[0].end);
-        std::string function_name = string_strip(func_name->prelude);
-        // Kimi-K2 uses functions.{{ tool_call['function']['name'] }}:{{ loop.index }} as function name
-        if (builder.syntax().format == COMMON_CHAT_FORMAT_KIMI_K2) {
-            if (string_starts_with(function_name, "functions.")) {
-                static const std::regex re(":\\d+$");
-                if (std::regex_search(function_name, re)) {
-                    function_name = function_name.substr(10, function_name.rfind(":") - 10);
-                }
-            }
-        }
-
-        // Argument JSON
-        json arguments = json::object();
-
-        // Helper to generate a partial argument JSON
-        const auto gen_partial_args = [&](auto set_partial_arg) {
-            gen_partial_json(set_partial_arg, arguments, builder, function_name);
-        };
-
-        // Parse all arg_key/arg_value pairs
-        while (auto tc = builder.try_find_literal(form.key_start)) {
-            if (!all_space(tc->prelude)) {
-                LOG_DBG("XML-Style tool call: Expected %s, but found %s, trying to match next pattern\n",
-                        gbnf_format_literal(form.key_start).c_str(),
-                        gbnf_format_literal(tc->prelude).c_str()
-                );
-                builder.move_to(tc->groups[0].begin - tc->prelude.size());
-                break;
-            }
-            if (tc->groups[0].end - tc->groups[0].begin != form.key_start.size()) {
-                auto tool_call_arg = arguments.dump();
-                if (tool_call_arg.size() != 0 && tool_call_arg[tool_call_arg.size() - 1] == '}') {
-                    tool_call_arg.resize(tool_call_arg.size() - 1);
-                }
-                builder.add_tool_call(function_name, "", tool_call_arg);
-                throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(form.key_start));
-            }
-
-            // Parse arg_key
-            auto key_res = builder.try_find_literal(form.key_val_sep);
-            if (!key_res) {
-                gen_partial_args([&](auto &rest, auto &needle) {arguments[rest + needle] = "";});
-                throw common_chat_msg_partial_exception("Expected " + gbnf_format_literal(form.key_val_sep) + " after " + gbnf_format_literal(form.key_start));
-            }
-            if (key_res->groups[0].end - key_res->groups[0].begin != form.key_val_sep.size()) {
-                gen_partial_args([&](auto &, auto &needle) {arguments[key_res->prelude + needle] = "";});
-                throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(form.key_val_sep));
-            }
-            auto &key = key_res->prelude;
-            recovery = false;
-
-            // Parse arg_value
-            if (form.key_val_sep2) {
-                if (auto tc = builder.try_find_literal(*form.key_val_sep2)) {
-                    if (!all_space(tc->prelude)) {
-                        LOG_DBG("Failed to parse XML-Style tool call: Unexcepted %s between %s and %s\n",
-                                gbnf_format_literal(tc->prelude).c_str(),
-                                gbnf_format_literal(form.key_val_sep).c_str(),
-                                gbnf_format_literal(*form.key_val_sep2).c_str()
-                        );
-                        return return_error(builder, start_pos, false);
-                    }
-                    if (tc->groups[0].end - tc->groups[0].begin != form.key_val_sep2->size()) {
-                        gen_partial_args([&](auto &, auto &needle) {arguments[key] = needle;});
-                        throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(*form.key_val_sep2));
-                    }
-                } else {
-                    gen_partial_args([&](auto &, auto &needle) {arguments[key] = needle;});
-                    throw common_chat_msg_partial_exception("Expected " + gbnf_format_literal(*form.key_val_sep2) + " after " + gbnf_format_literal(form.key_val_sep));
-                }
-            }
-            auto val_start = builder.pos();
-
-            // Test if arg_val is a partial JSON
-            std::optional<common_json> value_json = std::nullopt;
-            if (!form.raw_argval || !*form.raw_argval) {
-                try { value_json = builder.try_consume_json(); }
-                catch (const std::runtime_error&) { builder.move_to(val_start); }
-                // TODO: Delete this when json_partial adds top-level support for null/true/false
-                if (builder.pos() == val_start) {
-                    const static std::regex number_regex(R"([0-9-][0-9]*(\.\d*)?([eE][+-]?\d*)?)");
-                    builder.consume_spaces();
-                    std::string_view sv = utf8_truncate_safe_view(builder.input());
-                    sv.remove_prefix(builder.pos());
-                    std::string rest = "a";
-                    if (sv.size() < 6) rest = sv;
-                    if (string_starts_with("null", rest) || string_starts_with("true", rest) || string_starts_with("false", rest) || std::regex_match(sv.begin(), sv.end(), number_regex)) {
-                        value_json = {123, {"123", "123"}};
-                        builder.consume_rest();
-                    } else {
-                        builder.move_to(val_start);
-                    }
-                }
-            }
-
-            // If it is a JSON and followed by </arg_value>, parse as json
-            // cannot support streaming because it may be a plain text starting with JSON
-            if (value_json) {
-                auto json_end = builder.pos();
-                builder.consume_spaces();
-                if (builder.pos() == builder.input().size()) {
-                    if (form.raw_argval && !*form.raw_argval && (value_json->json.is_string() || value_json->json.is_object() || value_json->json.is_array())) {
-                        arguments[key] = value_json->json;
-                        auto json_str = arguments.dump();
-                        if (!value_json->healing_marker.json_dump_marker.empty()) {
-                            GGML_ASSERT(std::string::npos != json_str.rfind(value_json->healing_marker.json_dump_marker));
-                            json_str.resize(json_str.rfind(value_json->healing_marker.json_dump_marker));
-                        } else {
-                            GGML_ASSERT(json_str.back() == '}');
-                            json_str.resize(json_str.size() - 1);
-                        }
-                        builder.add_tool_call(function_name, "", json_str);
-                    } else {
-                        gen_partial_args([&](auto &, auto &needle) {arguments[key] = needle;});
-                    }
-                    LOG_DBG("Possible JSON arg_value: %s\n", value_json->json.dump().c_str());
-                    throw common_chat_msg_partial_exception("JSON arg_value detected. Waiting for more tokens for validations.");
-                }
-                builder.move_to(json_end);
-                auto [val_end_size, tc] = try_find_val_end();
-                if (tc && all_space(tc->prelude) && value_json->healing_marker.marker.empty()) {
-                    if (tc->groups[0].end - tc->groups[0].begin != val_end_size) {
-                        gen_partial_args([&](auto &, auto &needle) {arguments[key] = needle;});
-                        LOG_DBG("Possible terminated JSON arg_value: %s\n", value_json->json.dump().c_str());
-                        throw common_chat_msg_partial_exception("Partial literal: " + gbnf_format_literal(form.val_end) + (form.last_val_end ? gbnf_format_literal(*form.last_val_end) : ""));
-                    } else arguments[key] = value_json->json;
-                } else builder.move_to(val_start);
-            }
-
-            // If not, parse as plain text
-            if (val_start == builder.pos()) {
-                if (auto [val_end_size, value_plain] = try_find_val_end(); value_plain) {
-                    auto &value_str = value_plain->prelude;
-                    if (form.trim_raw_argval) value_str = string_strip(value_str);
-                    if (value_plain->groups[0].end - value_plain->groups[0].begin != val_end_size) {
-                        gen_partial_args([&](auto &, auto &needle) {arguments[key] = value_str + needle;});
-                        throw common_chat_msg_partial_exception(
-                                "Expected " + gbnf_format_literal(form.val_end) +
-                                " after " + gbnf_format_literal(form.key_val_sep) +
-                                (form.key_val_sep2 ? " " + gbnf_format_literal(*form.key_val_sep2) : "")
-                        );
-                    }
-                    arguments[key] = value_str;
-                } else {
-                    if (form.trim_raw_argval) {
-                        gen_partial_args([&](auto &rest, auto &needle) {arguments[key] = string_strip(rest) + needle;});
-                    } else {
-                        gen_partial_args([&](auto &rest, auto &needle) {arguments[key] = rest + needle;});
-                    }
-                    throw common_chat_msg_partial_exception(
-                            "Expected " + gbnf_format_literal(form.val_end) +
-                            " after " + gbnf_format_literal(form.key_val_sep) +
-                            (form.key_val_sep2 ? " " + gbnf_format_literal(*form.key_val_sep2) : "")
-                    );
-                }
-            }
-        }
-
-        // Consume closing tag
-        if (auto [tool_end_size, tc] = try_find_tool_end(); tc) {
-            if (!all_space(tc->prelude)) {
-                LOG_DBG("Failed to parse XML-Style tool call: Expected %s, but found %s\n",
-                        gbnf_format_literal(form.tool_end).c_str(),
-                        gbnf_format_literal(tc->prelude).c_str()
-                );
-                return return_error(builder, start_pos, recovery);
-            }
-            if (tc->groups[0].end - tc->groups[0].begin == tool_end_size) {
-                // Add the parsed tool call
-                if (!builder.add_tool_call(function_name, "", arguments.dump())) {
-                    throw common_chat_msg_partial_exception("Failed to add XML-Style tool call");
-                }
-                recovery = false;
-                continue;
-            }
-        }
-
-        auto tool_call_arg = arguments.dump();
-        if (tool_call_arg.size() != 0 && tool_call_arg[tool_call_arg.size() - 1] == '}') {
-            tool_call_arg.resize(tool_call_arg.size() - 1);
-        }
-        builder.add_tool_call(function_name, "", tool_call_arg);
-        throw common_chat_msg_partial_exception("Expected " + gbnf_format_literal(form.tool_end) + " after " + gbnf_format_literal(form.val_end));
-    }
-    if (auto tc = builder.try_find_literal(form.scope_end)) {
-        if (!all_space(tc->prelude)) {
-            LOG_DBG("Failed to parse XML-Style tool call: Expected %s, but found %s\n",
-                    gbnf_format_literal(form.scope_end).c_str(),
-                    gbnf_format_literal(tc->prelude).c_str()
-            );
-            return return_error(builder, start_pos, recovery);
-        }
-    } else {
-        if (all_space(form.scope_end)) return true;
-        builder.consume_spaces();
-        if (builder.pos() == builder.input().size())
-            throw common_chat_msg_partial_exception("incomplete tool calls");
-        LOG_DBG("Failed to parse XML-Style tool call: Expected %s, but found %s\n",
-                gbnf_format_literal(form.scope_end).c_str(),
-                gbnf_format_literal(builder.consume_rest()).c_str()
-        );
-        return return_error(builder, start_pos, recovery);
-    }
-
-    return true;
-}
-
-/**
- * Parse XML-Style tool call for given xml_tool_call_format. Return false for invalid syntax and get the position untouched.
- * May cause std::runtime_error if there is invalid syntax because partial valid tool call is already sent out to client.
- * form.scope_start, form.tool_sep and form.scope_end can be empty.
- */
-bool common_chat_msg_parser::try_consume_xml_tool_calls(const struct xml_tool_call_format & form) {
-    auto pos = pos_;
-    auto tsize = result_.tool_calls.size();
-    try { return parse_xml_tool_calls(*this, form); }
-    catch (const xml_toolcall_syntax_exception&) {}
-    move_to(pos);
-    result_.tool_calls.resize(tsize);
-    return false;
-}
-
-/**
- * Parse content uses reasoning and XML-Style tool call
- * TODO: Note that form.allow_toolcall_in_think is not tested yet. If anyone confirms it works, this comment can be removed.
- */
-inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, const struct xml_tool_call_format & form, const std::string & start_think = "<think>", const std::string & end_think = "</think>") {
-    constexpr auto rstrip = [](std::string &s) {
-        s.resize(std::distance(s.begin(), std::find_if(s.rbegin(), s.rend(), [](unsigned char ch) { return !std::isspace(ch); }).base()));
-    };
-    // Erase substring from l to r, along with additional spaces nearby
-    constexpr auto erase_spaces = [](auto &str, size_t l, size_t r) {
-        while (/* l > -1 && */ --l < str.size() && std::isspace(static_cast<unsigned char>(str[l])));
-        ++l;
-        while (++r < str.size() && std::isspace(static_cast<unsigned char>(str[r])));
-        if (l < r) str[l] = '\n';
-        if (l + 1 < r) str[l + 1] = '\n';
-        if (l != 0) l += 2;
-        str.erase(l, r - l);
-        return l;
-    };
-    constexpr auto trim_suffix = [](std::string &content, std::initializer_list<std::string_view> list) {
-        auto best_match = content.size();
-        for (auto pattern: list) {
-            if (pattern.size() == 0) continue;
-            for (auto match_idx = content.size() - std::min(pattern.size(), content.size()); content.size() > match_idx; match_idx++) {
-                auto match_len = content.size() - match_idx;
-                if (content.compare(match_idx, match_len, pattern.data(), match_len) == 0 && best_match > match_idx) {
-                    best_match = match_idx;
-                }
-            }
-        }
-        if (content.size() > best_match) {
-            content.erase(best_match);
-        }
-    };
-    const auto trim_potential_partial_word = [&start_think, &end_think, &form, trim_suffix](std::string &content) {
-        return trim_suffix(content, {
-            start_think, end_think, form.scope_start, form.tool_start, form.tool_sep, form.key_start,
-            form.key_val_sep, form.key_val_sep2 ? form.key_val_sep2->c_str() : "",
-            form.val_end, form.last_val_end ? form.last_val_end->c_str() : "",
-            form.tool_end, form.last_tool_end ? form.last_tool_end->c_str() : "",
-            form.scope_end
-        });
-    };
-
-
-    // Trim leading spaces without affecting keyword matching
-    static const common_regex spaces_regex("\\s*");
-    {
-        auto tc = builder.consume_regex(spaces_regex);
-        auto spaces = builder.str(tc.groups[0]);
-        auto s1 = spaces.size();
-        trim_potential_partial_word(spaces);
-        auto s2 = spaces.size();
-        builder.move_to(builder.pos() - (s1 - s2));
-    }
-
-    // Parse content
-    bool reasoning_unclosed = builder.syntax().thinking_forced_open;
-    std::string unclosed_reasoning_content("");
-    for (;;) {
-        auto tc = try_find_2_literal_splited_by_spaces(builder, form.scope_start, form.tool_start);
-        std::string content;
-        std::string tool_call_start;
-
-        if (tc) {
-            content = std::move(tc->prelude);
-            tool_call_start = builder.str(tc->groups[0]);
-            LOG_DBG("Matched tool start: %s\n", gbnf_format_literal(tool_call_start).c_str());
-        } else {
-            content = builder.consume_rest();
-            utf8_truncate_safe_resize(content);
-        }
-
-        // Handle unclosed think block
-        if (reasoning_unclosed) {
-            if (auto pos = content.find(end_think); pos == std::string::npos && builder.pos() != builder.input().size()) {
-                unclosed_reasoning_content += content;
-                if (!(form.allow_toolcall_in_think && tc)) {
-                    unclosed_reasoning_content += tool_call_start;
-                    continue;
-                }
-            } else {
-                reasoning_unclosed = false;
-                std::string reasoning_content;
-                if (pos == std::string::npos) {
-                    reasoning_content = std::move(content);
-                } else {
-                    reasoning_content = content.substr(0, pos);
-                    content.erase(0, pos + end_think.size());
-                }
-                if (builder.pos() == builder.input().size() && all_space(content)) {
-                    rstrip(reasoning_content);
-                    trim_potential_partial_word(reasoning_content);
-                    rstrip(reasoning_content);
-                    if (reasoning_content.empty()) {
-                        rstrip(unclosed_reasoning_content);
-                        trim_potential_partial_word(unclosed_reasoning_content);
-                        rstrip(unclosed_reasoning_content);
-                        if (unclosed_reasoning_content.empty()) continue;
-                    }
-                }
-                if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) {
-                    builder.add_content(start_think);
-                    builder.add_content(unclosed_reasoning_content);
-                    builder.add_content(reasoning_content);
-                    if (builder.pos() != builder.input().size() || !all_space(content))
-                        builder.add_content(end_think);
-                } else {
-                    builder.add_reasoning_content(unclosed_reasoning_content);
-                    builder.add_reasoning_content(reasoning_content);
-                }
-                unclosed_reasoning_content.clear();
-            }
-        }
-
-        // Handle multiple think block
-        bool toolcall_in_think = false;
-        for (auto think_start = content.find(start_think); think_start != std::string::npos; think_start = content.find(start_think, think_start)) {
-            if (auto think_end = content.find(end_think, think_start + start_think.size()); think_end != std::string::npos) {
-                if (builder.syntax().reasoning_format != COMMON_REASONING_FORMAT_NONE && !builder.syntax().reasoning_in_content) {
-                    auto reasoning_content = content.substr(think_start + start_think.size(), think_end - think_start - start_think.size());
-                    builder.add_reasoning_content(reasoning_content);
-                    think_start = erase_spaces(content, think_start, think_end + end_think.size() - 1);
-                } else {
-                    think_start = think_end + end_think.size() - 1;
-                }
-            } else {
-                // This <tool_call> start is in thinking block, skip this tool call
-                // This <tool_call> start is in thinking block
-                if (form.allow_toolcall_in_think) {
-                    unclosed_reasoning_content = content.substr(think_start + start_think.size());
-                } else {
-                    unclosed_reasoning_content = content.substr(think_start + start_think.size()) + tool_call_start;
-                }
-                reasoning_unclosed = true;
-                content.resize(think_start);
-                toolcall_in_think = true;
-            }
-        }
-
-        if (builder.syntax().reasoning_format != COMMON_REASONING_FORMAT_NONE && !builder.syntax().reasoning_in_content) {
-            rstrip(content);
-            // Handle unclosed </think> token from content: delete all </think> token
-            if (auto pos = content.rfind(end_think); pos != std::string::npos) {
-                while (pos != std::string::npos) {
-                    pos = erase_spaces(content, pos, pos + end_think.size() - 1);
-                    pos = content.rfind(end_think, pos);
-                }
-            }
-            // Strip if needed
-            if (content.size() > 0 && std::isspace(static_cast<unsigned char>(content[0]))) {
-                content = string_strip(content);
-            }
-        }
-
-        // remove potential partial suffix
-        if (builder.pos() == builder.input().size() && builder.is_partial()) {
-            if (unclosed_reasoning_content.empty()) {
-                rstrip(content);
-                trim_potential_partial_word(content);
-                rstrip(content);
-            } else {
-                rstrip(unclosed_reasoning_content);
-                trim_potential_partial_word(unclosed_reasoning_content);
-                rstrip(unclosed_reasoning_content);
-            }
-        }
-
-        // consume unclosed_reasoning_content if allow_toolcall_in_think is set
-        if (form.allow_toolcall_in_think && !unclosed_reasoning_content.empty()) {
-            if (builder.syntax().reasoning_format != COMMON_REASONING_FORMAT_NONE && !builder.syntax().reasoning_in_content) {
-                builder.add_reasoning_content(unclosed_reasoning_content);
-            } else {
-                if (content.empty()) {
-                    content = start_think + unclosed_reasoning_content;
-                } else {
-                    content += "\n\n" + start_think;
-                    content += unclosed_reasoning_content;
-                }
-            }
-            unclosed_reasoning_content.clear();
-        }
-
-        // Add content
-        if (!content.empty()) {
-            // If there are multiple content blocks
-            if (builder.syntax().reasoning_format != COMMON_REASONING_FORMAT_NONE && !builder.syntax().reasoning_in_content && builder.result().content.size() != 0) {
-                builder.add_content("\n\n");
-            }
-            builder.add_content(content);
-        }
-
-        // This <tool_call> start is in thinking block and toolcall_in_think not set, skip this tool call
-        if (toolcall_in_think && !form.allow_toolcall_in_think) {
-            continue;
-        }
-
-        // There is no tool call and all content is parsed
-        if (!tc) {
-            GGML_ASSERT(builder.pos() == builder.input().size());
-            GGML_ASSERT(unclosed_reasoning_content.empty());
-            if (!form.allow_toolcall_in_think) GGML_ASSERT(!reasoning_unclosed);
-            break;
-        }
-
-        builder.move_to(tc->groups[0].begin);
-        if (builder.try_consume_xml_tool_calls(form)) {
-            auto end_of_tool = builder.pos();
-            builder.consume_spaces();
-            if (builder.pos() != builder.input().size()) {
-                builder.move_to(end_of_tool);
-                if (!builder.result().content.empty()) {
-                    builder.add_content("\n\n");
-                }
-            }
-        } else {
-            static const common_regex next_char_regex(".");
-            auto c = builder.str(builder.consume_regex(next_char_regex).groups[0]);
-            rstrip(c);
-            builder.add_content(c);
-        }
-    }
-}
-
-/**
- * Parse content uses reasoning and XML-Style tool call
- */
-void common_chat_msg_parser::consume_reasoning_with_xml_tool_calls(const struct xml_tool_call_format & form, const std::string & start_think, const std::string & end_think) {
-    parse_msg_with_xml_tool_calls(*this, form, start_think, end_think);
-}

common/chat-parser-xml-toolcall.h

@@ -1,45 +0,0 @@
-#pragma once
-
-#include "chat.h"
-
-#include <nlohmann/json.hpp>
-
-#include <optional>
-#include <string>
-#include <vector>
-
-
-// Sample config:
-// MiniMax-M2 (left): <minimax:tool_call>\n<invoke name="tool-name">\n<parameter name="key">value</parameter>\n...</invoke>\n...</minimax:tool_call>
-// GLM 4.5   (right): <tool_call>function_name\n<arg_key>key</arg_key>\n<arg_value>value</arg_value>\n</tool_call>
-struct xml_tool_call_format {
-    std::string scope_start; // <minimax:tool_call>\n  // \n                      // can be empty
-    std::string tool_start;  // <invoke name=\"        // <tool_call>
-    std::string tool_sep;    // \">\n                  // \n                      // can be empty only for parse_xml_tool_calls
-    std::string key_start;   // <parameter name=\"     // <arg_key>
-    std::string key_val_sep; // \">                    // </arg_key>\n<arg_value>
-    std::string val_end;     // </parameter>\n         // </arg_value>\n
-    std::string tool_end;    // </invoke>\n            // </tool_call>\n
-    std::string scope_end;   // </minimax:tool_call>   //                         // can be empty
-    // Set this if there can be dynamic spaces inside key_val_sep.
-    // e.g. key_val_sep=</arg_key> key_val_sep2=<arg_value> for GLM4.5
-    std::optional<std::string> key_val_sep2 = std::nullopt;
-    // Set true if argval should only be raw string. e.g. Hello "world" hi
-    // Set false if argval should only be json string. e.g. "Hello \"world\" hi"
-    // Defaults to std::nullopt, both will be allowed.
-    std::optional<bool> raw_argval = std::nullopt;
-    std::optional<std::string> last_val_end = std::nullopt;
-    std::optional<std::string> last_tool_end = std::nullopt;
-    bool trim_raw_argval = false;
-    bool allow_toolcall_in_think = false;
-};
-
-// make a GBNF that accept any strings except those containing any of the forbidden strings.
-std::string make_gbnf_excluding(std::vector<std::string> forbids);
-
-/**
- * Build grammar for xml-style tool call
- * form.scope_start and form.scope_end can be empty.
- * Requires data.format for model-specific hacks.
- */
-void build_grammar_xml_tool_call(common_chat_params & data, const nlohmann::ordered_json & tools, const struct xml_tool_call_format & form);

common/chat-parser.h

@@ -1,133 +0,0 @@
-#pragma once
-
-#include "chat.h"
-#include "chat-parser-xml-toolcall.h"
-#include "json-partial.h"
-#include "regex-partial.h"
-
-#include <nlohmann/json_fwd.hpp>
-
-#include <optional>
-#include <string>
-#include <vector>
-
-class common_chat_msg_partial_exception : public std::runtime_error {
-  public:
-    common_chat_msg_partial_exception(const std::string & message) : std::runtime_error(message) {}
-};
-
-class common_chat_msg_parser {
-    std::string input_;
-    bool is_partial_;
-    common_chat_parser_params syntax_; // TODO: rename to params
-    std::string healing_marker_;
-
-    size_t pos_ = 0;
-    common_chat_msg result_;
-
-  public:
-    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
-    const std::string & input() const { return input_; }
-    size_t pos() const { return pos_; }
-    const std::string & healing_marker() const { return healing_marker_; }
-    const bool & is_partial() const { return is_partial_; }
-    const common_chat_msg & result() const { return result_; }
-    const common_chat_parser_params & syntax() const { return syntax_; }
-
-    void move_to(size_t pos) {
-        if (pos > input_.size()) {
-            throw std::runtime_error("Invalid position!");
-        }
-        pos_ = pos;
-    }
-    void move_back(size_t n) {
-        if (pos_ < n) {
-            throw std::runtime_error("Can't move back that far!");
-        }
-        pos_ -= n;
-    }
-
-    // Get the substring of the input at the given range
-    std::string str(const common_string_range & rng) const;
-
-    // Appends to the result.content field
-    void add_content(const std::string & content);
-
-    // Appends to the result.reasoning_content field
-    void add_reasoning_content(const std::string & reasoning_content);
-
-    // Adds a tool call to the result. If the tool call is too incomplete (e.g. name empty), it won't add anything.
-    bool add_tool_call(const std::string & name, const std::string & id, const std::string & arguments);
-
-    // Adds a tool call using the "name", "id" and "arguments" fields of the json object
-    bool add_tool_call(const nlohmann::ordered_json & tool_call);
-
-    // Adds an array of tool calls using their "name", "id" and "arguments" fields.
-    bool add_tool_calls(const nlohmann::ordered_json & arr);
-
-    // Adds a tool call using the short form: { "tool_name": { "arg1": val, "arg2": val } }
-    bool add_tool_call_short_form(const nlohmann::ordered_json & tool_call);
-
-    void finish();
-
-    bool consume_spaces();
-
-    void consume_literal(const std::string & literal);
-
-    bool try_parse_reasoning(const std::string & start_think, const std::string & end_think);
-
-    std::string consume_rest();
-
-    struct find_regex_result {
-        std::string prelude;
-        std::vector<common_string_range> groups;
-    };
-
-    std::optional<find_regex_result> try_find_regex(const common_regex & regex, size_t from = std::string::npos, bool add_prelude_to_content = true);
-
-    bool try_consume_literal(const std::string & literal);
-
-    std::optional<find_regex_result> try_find_literal(const std::string & literal);
-
-    find_regex_result consume_regex(const common_regex & regex);
-
-    std::optional<find_regex_result> try_consume_regex(const common_regex & regex);
-
-    std::optional<common_json> try_consume_json();
-    common_json consume_json();
-
-    struct consume_json_result {
-        nlohmann::ordered_json value;
-        bool is_partial;
-    };
-
-    /*
-        Consume (possibly partial) json and converts specific subtrees to (possibly truncated) JSON strings.
-
-        By default, object keys can't be truncated, nor can string values (their corresponding key is removed,
-        e.g. `{"foo": "bar", "baz": "b` -> `{"foo": "bar"}`
-
-        But one can allow subpaths to be kept truncated, and possibly json-dumped to truncated json strings
-        - with `content_paths={{"foo"}}` -> `{"foo": "b` -> {"foo": "b"}`
-        - with `args_paths={{"foo"}}` -> `{"foo": {"b` -> `{"foo": "{b"}`
-    */
-    consume_json_result consume_json_with_dumped_args(
-        const std::vector<std::vector<std::string>> & args_paths = {},
-        const std::vector<std::vector<std::string>> & content_paths = {}
-    );
-    std::optional<consume_json_result> try_consume_json_with_dumped_args(
-        const std::vector<std::vector<std::string>> & args_paths = {},
-        const std::vector<std::vector<std::string>> & content_paths = {}
-    );
-
-    /**
-     * Parse XML-Style tool call for given xml_tool_call_format. Return false for invalid syntax and get the position untouched.
-     * form.scope_start, form.tool_sep and form.scope_end can be empty.
-     */
-    bool try_consume_xml_tool_calls(const struct xml_tool_call_format & form);
-
-    // Parse content uses reasoning and XML-Style tool call
-    void consume_reasoning_with_xml_tool_calls(const struct xml_tool_call_format & form, const std::string & start_think = "<think>", const std::string & end_think = "</think>");
-
-    void clear_tools();
-};

common/chat-peg-parser.cpp

@@ -1,13 +1,17 @@
 #include "chat-peg-parser.h"
 
+#include "chat-auto-parser.h"
+#include "ggml.h"
+#include "peg-parser.h"
+
 #include <nlohmann/json.hpp>
 
-using json = nlohmann::json;
+using ordered_json = nlohmann::ordered_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
     int count = 0;
     while (!sv.empty() && std::isspace(static_cast<unsigned char>(sv.back()))) {
-        if (max != -1 && count <= max) {
+        if (max != -1 && count >= max) {
             break;
         }
         sv.remove_suffix(1);
@@ -16,109 +20,820 @@ static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
     return sv;
 }
 
-void common_chat_peg_mapper::from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result) {
+static std::string_view trim_leading_space(std::string_view sv, int max = -1) {
+    int count = 0;
+    while (!sv.empty() && std::isspace(static_cast<unsigned char>(sv.front()))) {
+        if (max != -1 && count >= max) {
+            break;
+        }
+        sv.remove_prefix(1);
+        count++;
+    }
+    return sv;
+}
+
+static std::string_view trim(std::string_view sv) {
+    return trim_trailing_space(trim_leading_space(sv, 1));
+}
+
+// Count the number of unclosed '{' braces in a JSON-like string,
+// properly skipping braces inside quoted strings.
+static int json_brace_depth(const std::string & s) {
+    int  depth     = 0;
+    bool in_string = false;
+    bool escaped   = false;
+    for (char c : s) {
+        if (escaped) {
+            escaped = false;
+            continue;
+        }
+        if (c == '\\' && in_string) {
+            escaped = true;
+            continue;
+        }
+        if (c == '"') {
+            in_string = !in_string;
+            continue;
+        }
+        if (!in_string) {
+            if (c == '{') {
+                depth++;
+            } else if (c == '}') {
+                depth--;
+            }
+        }
+    }
+    return depth;
+}
+
+// JSON-escape a string and return the inner content (without surrounding quotes).
+static std::string escape_json_string_inner(const std::string & s) {
+    std::string escaped = ordered_json(s).dump();
+    if (escaped.size() >= 2 && escaped.front() == '"' && escaped.back() == '"') {
+        return escaped.substr(1, escaped.size() - 2);
+    }
+    return escaped;
+}
+
+// Convert Python-style single-quoted strings to JSON double-quoted strings
+// Only converts outer string delimiters, properly handling escape sequences:
+// - {'key': 'value'} -> {"key": "value"}
+// - {'code': 'print(\'hello\')'} -> {"code": "print('hello')"}
+// - {'msg': 'He said "hi"'} -> {"msg": "He said \"hi\""}
+static std::string normalize_quotes_to_json(const std::string & input) {
+    std::string result;
+    result.reserve(input.size() + 16);  // May need extra space for escaping
+
+    bool in_single_quoted = false;
+    bool in_double_quoted = false;
+
+    for (size_t i = 0; i < input.size(); ++i) {
+        char c = input[i];
+
+        // Handle escape sequences
+        if (c == '\\' && i + 1 < input.size()) {
+            char next = input[i + 1];
+
+            if (in_single_quoted) {
+                // Inside a single-quoted string being converted to double quotes
+                if (next == '\'') {
+                    // \' -> ' (escaped single quote becomes unescaped in double-quoted string)
+                    result += '\'';
+                    ++i;
+                    continue;
+                }
+                if (next == '"') {
+                    // \" stays as \" (already escaped, works in double-quoted string)
+                    result += "\\\"";
+                    ++i;
+                    continue;
+                }
+                // Other escapes (\n, \\, etc.): pass through both characters
+                result += c;
+                result += next;
+                ++i;
+                continue;
+            }
+
+            if (in_double_quoted) {
+                // Inside a double-quoted string - pass through escape sequences as-is
+                result += c;
+                result += next;
+                ++i;
+                continue;
+            }
+
+            // Outside any string - just pass through the backslash
+            result += c;
+            continue;
+        }
+
+        // Handle quote characters
+        if (c == '"') {
+            if (in_single_quoted) {
+                // Unescaped double quote inside single-quoted string -> must escape for JSON
+                result += "\\\"";
+            } else {
+                // Double quote as string delimiter or outside strings
+                in_double_quoted = !in_double_quoted;
+                result += c;
+            }
+        } else if (c == '\'') {
+            if (in_double_quoted) {
+                // Single quote inside double-quoted string -> pass through
+                result += c;
+            } else if (in_single_quoted) {
+                // Closing single quote -> convert to double quote
+                in_single_quoted = false;
+                result += '"';
+            } else {
+                // Opening single quote -> convert to double quote
+                in_single_quoted = true;
+                result += '"';
+            }
+        } else {
+            result += c;
+        }
+    }
+
+    return result;
+}
+
+void tag_based_peg_mapper::from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result) {
     arena.visit(result, [this](const common_peg_ast_node & node) {
-        map(node);
+        if (!node.tag.empty()) {
+            tags[node.tag] = std::string(node.text);
+        }
     });
 }
 
-void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
-    bool is_reasoning = node.tag == common_chat_peg_builder::REASONING;
-    bool is_content = node.tag == common_chat_peg_builder::CONTENT;
+tagged_parse_result tagged_peg_parser::parse_and_extract(const std::string & input, common_peg_parse_flags extra_flags) const {
+    common_peg_parse_context ctx(input, flags | extra_flags);
+    auto parse_result = arena.parse(ctx);
 
-    if (is_reasoning) {
-        result.reasoning_content = std::string(trim_trailing_space(node.text));
+    tag_based_peg_mapper mapper;
+    mapper.from_ast(ctx.ast, parse_result);
+
+    return { std::move(parse_result), std::move(mapper.tags) };
+}
+
+tagged_parse_result tagged_peg_parser::parse_anywhere_and_extract(const std::string & input) const {
+    if (input.empty()) {
+        return parse_and_extract(input);
     }
+    for (size_t i = 0; i < input.size(); i++) {
+        common_peg_parse_context ctx(input, flags);
+        auto parse_result = arena.parse(ctx, i);
+        if (parse_result.success() || i == input.size() - 1) {
+            tag_based_peg_mapper mapper;
+            mapper.from_ast(ctx.ast, parse_result);
+            return { std::move(parse_result), std::move(mapper.tags) };
+        }
+    }
+    GGML_ABORT("Should not happen");
+}
 
-    if (is_content) {
-        result.content = std::string(trim_trailing_space(node.text));
+tagged_peg_parser build_tagged_peg_parser(
+    const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn) {
+    common_peg_parser_builder builder;
+    builder.set_root(fn(builder));
+    return { builder.build() };
+}
+
+common_peg_parser common_chat_peg_builder::tag_with_safe_content(const std::string &       tag_name,
+                                                                 const std::string &       marker,
+                                                                 const common_peg_parser & p) {
+    if (marker.empty()) {
+        return zero_or_more(choice({ p, rule(tag_name, content(any())) }));
+    }
+    auto content_chunk = rule(tag_name, content(negate(literal(marker)) + any() + until(marker)));
+    return zero_or_more(choice({ p, content_chunk }));
+}
+
+std::string & common_chat_peg_mapper::args_target() {
+    return (current_tool && !current_tool->name.empty()) ? current_tool->arguments : args_buffer;
+}
+
+void common_chat_peg_mapper::from_ast(const common_peg_ast_arena &    arena,
+                                      const common_peg_parse_result & parse_result_arg) {
+    arena.visit(parse_result_arg, [this](const common_peg_ast_node & node) { map(node); });
+    // Flush any pending tool call that was started but never got a name
+    // This happens during partial parsing when the tool call is incomplete
+    if (pending_tool_call.has_value() && !pending_tool_call->name.empty()) {
+        if (!args_buffer.empty()) {
+            pending_tool_call->arguments = args_buffer;
+        }
+        if (closing_quote_pending && !pending_tool_call->arguments.empty()) {
+            pending_tool_call->arguments += "\"";
+        }
+        result.tool_calls.push_back(pending_tool_call.value());
+        pending_tool_call.reset();
     }
 }
 
-void common_chat_peg_native_mapper::map(const common_peg_ast_node & node) {
-    common_chat_peg_mapper::map(node);
+void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
+    // Handle reasoning/content tags
+    bool is_reasoning = node.tag == common_chat_peg_builder::REASONING;
+    bool is_content   = node.tag == common_chat_peg_builder::CONTENT;
 
-    bool is_tool_open = node.tag == common_chat_peg_native_builder::TOOL_OPEN;
-    bool is_tool_name = node.tag == common_chat_peg_native_builder::TOOL_NAME;
-    bool is_tool_id = node.tag == common_chat_peg_native_builder::TOOL_ID;
-    bool is_tool_args = node.tag == common_chat_peg_native_builder::TOOL_ARGS;
+    if (is_reasoning) { // GPT OSS can have more than 1 reasoning block, so concatenate here
+        result.reasoning_content += std::string(node.text);
+    }
+
+    if (is_content) {
+        // Concatenate content from multiple content nodes (e.g., when reasoning markers
+        // are preserved before content markers in reasoning_format=NONE mode)
+        result.content += std::string(node.text);
+    }
+
+    // Handle tool-related tags (supporting both JSON and tagged formats)
+    bool is_tool_open  = node.tag == common_chat_peg_builder::TOOL_OPEN;
+    bool is_tool_close = node.tag == common_chat_peg_builder::TOOL_CLOSE;
+    bool is_tool_name  = node.tag == common_chat_peg_builder::TOOL_NAME;
+    bool is_tool_id    = node.tag == common_chat_peg_builder::TOOL_ID;
+    bool is_tool_args  = node.tag == common_chat_peg_builder::TOOL_ARGS;
+    bool is_arg_open   = node.tag == common_chat_peg_builder::TOOL_ARG_OPEN;
+    bool is_arg_close  = node.tag == common_chat_peg_builder::TOOL_ARG_CLOSE;
+    bool is_arg_name         = node.tag == common_chat_peg_builder::TOOL_ARG_NAME;
+    bool is_arg_value        = node.tag == common_chat_peg_builder::TOOL_ARG_VALUE;
+    bool is_arg_string_value = node.tag == common_chat_peg_builder::TOOL_ARG_STRING_VALUE;
 
     if (is_tool_open) {
-        result.tool_calls.emplace_back();
-        current_tool = &result.tool_calls.back();
+        pending_tool_call     = common_chat_tool_call();
+        current_tool          = &pending_tool_call.value();
+        arg_count             = 0;
+        args_buffer.clear();
+        closing_quote_pending = false;
     }
 
     if (is_tool_id && current_tool) {
-        current_tool->id = std::string(trim_trailing_space(node.text));
+        auto text = trim_trailing_space(node.text);
+        if (text.size() >= 2 && text.front() == '"' && text.back() == '"') {
+            text = text.substr(1, text.size() - 2);
+        }
+        current_tool->id = std::string(text);
     }
 
     if (is_tool_name && current_tool) {
         current_tool->name = std::string(trim_trailing_space(node.text));
+        // Now that we have the name, populate the arguments from the buffer
+        if (!args_buffer.empty()) {
+            current_tool->arguments = args_buffer;
+            args_buffer.clear();
+        } else if (current_tool->arguments.empty()) {
+            current_tool->arguments = "{";
+        }
+        // Add the tool call to results so streaming can see it
+        if (pending_tool_call.has_value()) {
+            result.tool_calls.push_back(pending_tool_call.value());
+            pending_tool_call.reset();
+            current_tool = &result.tool_calls.back();
+        }
     }
 
     if (is_tool_args && current_tool) {
-        current_tool->arguments = std::string(trim_trailing_space(node.text));
-    }
-}
-
-void common_chat_peg_constructed_mapper::map(const common_peg_ast_node & node) {
-    common_chat_peg_mapper::map(node);
-
-    bool is_tool_open = node.tag == common_chat_peg_constructed_builder::TOOL_OPEN;
-    bool is_tool_name = node.tag == common_chat_peg_constructed_builder::TOOL_NAME;
-    bool is_tool_close = node.tag == common_chat_peg_constructed_builder::TOOL_CLOSE;
-    bool is_arg_open = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_OPEN;
-    bool is_arg_close = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_CLOSE;
-    bool is_arg_name = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_NAME;
-    bool is_arg_string = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_STRING_VALUE;
-    bool is_arg_json = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_JSON_VALUE;
-
-    if (is_tool_open) {
-        result.tool_calls.emplace_back();
-        current_tool = &result.tool_calls.back();
-        arg_count = 0;
-    }
-
-    if (is_tool_name) {
-        current_tool->name = std::string(node.text);
-        current_tool->arguments = "{";
+        // For JSON format: arguments come as a complete JSON object
+        // For tagged format: built up from individual arg_name/arg_value nodes
+        auto text = trim_trailing_space(node.text);
+        if (!text.empty() && text.front() == '{') {
+            args_target() = std::string(text);
+        }
     }
 
     if (is_arg_open) {
-        needs_closing_quote = false;
+        closing_quote_pending = false;
     }
 
     if (is_arg_name && current_tool) {
+        std::string arg_entry;
         if (arg_count > 0) {
-            current_tool->arguments += ",";
+            arg_entry = ",";
         }
-        current_tool->arguments += json(trim_trailing_space(node.text)).dump() + ":";
+        arg_entry += ordered_json(trim(node.text)).dump() + ":";
         ++arg_count;
+
+        auto & target = args_target();
+        if (target.empty()) {
+            target = "{";
+        }
+        target += arg_entry;
     }
 
-    if (is_arg_string && current_tool) {
-        // Serialize to JSON, but exclude the end quote
-        std::string dumped = json(trim_trailing_space(node.text)).dump();
-        current_tool->arguments += dumped.substr(0, dumped.size() - 1);
-        needs_closing_quote = true;
+    if ((is_arg_value || is_arg_string_value) && current_tool) {
+        std::string value_content = std::string(trim_trailing_space(trim_leading_space(node.text, 1), 1));
+
+        std::string value_to_add;
+        if (value_content.empty() && is_arg_string_value) {
+            // Empty string value - arg_close will add the closing quote
+            value_to_add          = "\"";
+            closing_quote_pending = true;
+        } else if (!value_content.empty() && is_arg_string_value) {
+            // Schema declares this as string type - always treat as literal string value
+            if (!closing_quote_pending) {
+                value_to_add          = "\"";
+                closing_quote_pending = true;
+            }
+            value_to_add += escape_json_string_inner(value_content);
+        } else if (!value_content.empty()) {
+            // For potential containers, normalize Python-style single quotes to JSON double quotes
+            bool is_potential_container = value_content[0] == '[' || value_content[0] == '{';
+            if (is_potential_container) {
+                value_content = normalize_quotes_to_json(value_content);
+            }
+
+            // Try to parse as JSON value (number, bool, null, object, array)
+            try {
+                ordered_json parsed = ordered_json::parse(value_content);
+                if (parsed.is_string()) {
+                    // Don't add closing quote yet (added by arg_close) for monotonic streaming
+                    std::string escaped = parsed.dump();
+                    if (!escaped.empty() && escaped.back() == '"') {
+                        escaped.pop_back();
+                    }
+                    value_to_add          = escaped;
+                    closing_quote_pending = true;
+                } else {
+                    // Non-string values: use raw content to preserve whitespace for monotonicity
+                    value_to_add = value_content;
+                }
+            } catch (...) {
+                if (node.is_partial && is_potential_container) {
+                    // Partial container: pass through the already-normalized content
+                    value_to_add = value_content;
+                } else {
+                    // Not valid JSON - treat as string value
+                    if (!closing_quote_pending) {
+                        value_to_add          = "\"";
+                        closing_quote_pending = true;
+                    }
+                    value_to_add += escape_json_string_inner(value_content);
+                }
+            }
+        }
+
+        args_target() += value_to_add;
     }
 
     if (is_arg_close && current_tool) {
-        if (needs_closing_quote) {
-            current_tool->arguments += "\"";
-            needs_closing_quote = false;
+        if (closing_quote_pending) {
+            args_target() += "\"";
+            closing_quote_pending = false;
         }
     }
 
-    if (is_arg_json && current_tool) {
-        current_tool->arguments += std::string(trim_trailing_space(node.text));
-    }
-
     if (is_tool_close && current_tool) {
-        if (needs_closing_quote) {
-            current_tool->arguments += "\"";
-            needs_closing_quote = false;
+        // Flush buffer to arguments if tool name was never seen
+        if (current_tool->name.empty() && !args_buffer.empty()) {
+            current_tool->arguments = args_buffer;
+            args_buffer.clear();
+        }
+        // Close any pending string quote
+        if (closing_quote_pending) {
+            current_tool->arguments += "\"";
+            closing_quote_pending = false;
+        }
+        // Close any unclosed braces (accounts for nested objects)
+        for (int d = json_brace_depth(current_tool->arguments); d > 0; d--) {
+            current_tool->arguments += "}";
+        }
+        // Add tool call to results if named; otherwise discard
+        if (pending_tool_call.has_value()) {
+            if (!current_tool->name.empty()) {
+                result.tool_calls.push_back(pending_tool_call.value());
+            }
+            pending_tool_call.reset();
         }
-        current_tool->arguments += "}";
     }
 }
+
+common_peg_parser common_chat_peg_builder::standard_constructed_tools(
+    const std::map<std::string, std::string> & markers,
+    const ordered_json &                       tools,
+    bool                                       parallel_tool_calls,
+    bool                                       force_tool_calls) {
+    if (!tools.is_array() || tools.empty()) {
+        return eps();
+    }
+
+    // Extract markers with defaults
+    auto get_marker = [&markers](const std::string & key, const std::string & default_val = "") -> std::string {
+        auto it = markers.find(key);
+        return it != markers.end() ? it->second : default_val;
+    };
+
+    std::string section_start    = get_marker("tool_call_start_marker", "<tool_call>");
+    std::string section_end      = get_marker("tool_call_end_marker", "</tool_call>");
+    std::string func_opener      = get_marker("function_opener", "<function=");
+    std::string func_name_suffix = get_marker("function_name_suffix", ">");
+    std::string func_closer      = get_marker("function_closer", "</function>");
+    std::string param_key_prefix = get_marker("parameter_key_prefix", "<param=");
+    std::string param_key_suffix = get_marker("parameter_key_suffix", ">");
+    std::string param_closer     = get_marker("parameter_closer", "</param>");
+
+    // Build tool choices for tagged format
+    auto tool_choices = choice();
+
+    for (const auto & tool_def : tools) {
+        if (!tool_def.contains("function")) {
+            continue;
+        }
+        const auto &   function = tool_def.at("function");
+        std::string    name     = function.at("name");
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+
+        // Build argument parsers
+        auto args = eps();
+        if (params.contains("properties") && !params["properties"].empty()) {
+            auto arg_choice = choice();
+            for (const auto & el : params["properties"].items()) {
+                const std::string & prop_name = el.key();
+
+                auto arg_name_parser =
+                    choice({ literal(prop_name), literal("\"" + prop_name + "\""), literal("'" + prop_name + "'") });
+
+                auto arg_rule = tool_arg(tool_arg_open(literal(param_key_prefix)) + tool_arg_name(arg_name_parser) +
+                                         literal(param_key_suffix) + tool_arg_value(until(param_closer)) +
+                                         tool_arg_close(literal(param_closer)));
+                arg_choice |= arg_rule;
+            }
+            args = zero_or_more(arg_choice + space());
+        }
+
+        // Build function parser: <function=name>args</function>
+        auto tool_parser = tool(tool_open(literal(func_opener) + tool_name(literal(name)) + literal(func_name_suffix)) +
+                                space() + tool_args(args) + space() + tool_close(literal(func_closer)));
+
+        tool_choices |= rule("tool-" + name, tool_parser);
+    }
+
+    // Build the section with markers
+    auto section =
+        parallel_tool_calls ?
+            trigger_rule("tool-call", literal(section_start) + space() + one_or_more(tool_choices + space()) +
+                                          literal(section_end)) :
+            trigger_rule("tool-call", literal(section_start) + space() + tool_choices + space() + literal(section_end));
+
+    return force_tool_calls ? section : optional(section);
+}
+
+// Python-style tool calls: name(arg1="value1", arg2=123)
+// Used only by LFM2 for now, so we don't merge it into autoparser
+common_peg_parser common_chat_peg_builder::python_style_tool_calls(
+    const ordered_json & tools,
+    bool                 parallel_tool_calls) {
+    if (!tools.is_array() || tools.empty()) {
+        return eps();
+    }
+
+    auto tool_choices = choice();
+
+    for (const auto & tool_def : tools) {
+        if (!tool_def.contains("function")) {
+            continue;
+        }
+        const auto &   function = tool_def.at("function");
+        std::string    name     = function.at("name");
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+
+        auto args = eps();
+        if (params.contains("properties") && !params["properties"].empty()) {
+            auto arg_choice = choice();
+            for (const auto & el : params["properties"].items()) {
+                const std::string & prop_name = el.key();
+                const auto & prop_def = el.value();
+                bool is_string_type = (prop_def.contains("type") && prop_def["type"] == "string");
+
+                auto arg_name_parser = literal(prop_name);
+
+                common_peg_parser arg_value_parser = eps();
+                auto string_value_parser = choice({
+                    literal("\"") + tool_arg_string_value(string_content('"')) + literal("\""),
+                    literal("'") + tool_arg_string_value(string_content('\'')) + literal("'")
+                });
+
+                if (is_string_type) {
+                    arg_value_parser = string_value_parser;
+                } else {
+                    arg_value_parser = tool_arg_value(python_value());
+                }
+
+                // Full argument: name="value" or name=value
+                auto arg_rule = tool_arg(
+                    tool_arg_open(eps()) +
+                    tool_arg_name(arg_name_parser) +
+                    literal("=") +
+                    arg_value_parser +
+                    tool_arg_close(eps())
+                );
+                arg_choice |= arg_rule;
+            }
+
+            args = arg_choice + zero_or_more("," + space() + arg_choice);
+        }
+
+        auto tool_parser = tool(tool_open(tool_name(literal(name)) + literal("(")) +
+            space() + tool_args(args) + space() + tool_close(literal(")"))
+        );
+
+        tool_choices |= rule("tool-" + name, tool_parser);
+    }
+
+    if (parallel_tool_calls) {
+        return "[" + space() + tool_choices + zero_or_more("," + space() + tool_choices) + space() + "]";
+    }
+    return "[" + space() + tool_choices + space() + "]";
+}
+
+// Helper: Parse dot notation key into prefix and field name
+static std::pair<std::string, std::string> parse_key_spec(const std::string & key) {
+    auto dot_pos = key.find('.');
+    if (dot_pos == std::string::npos) {
+        return {"", key};  // Top-level field
+    }
+    return {key.substr(0, dot_pos), key.substr(dot_pos + 1)};
+}
+
+// Mode 1: function_is_key — parse {"function_name": {...}}
+common_peg_parser common_chat_peg_builder::build_json_tools_function_is_key(
+    const ordered_json & tools,
+    const std::string &  args_key,
+    const std::string &  effective_args_key,
+    const std::string &  call_id_key,
+    const std::string &  gen_call_id_key) {
+
+    auto tool_choices = choice();
+
+    for (const auto & tool_def : tools) {
+        if (!tool_def.contains("function")) {
+            continue;
+        }
+        const auto &   function = tool_def.at("function");
+        std::string    name     = function.at("name");
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+
+        // Build inner object fields
+        std::vector<common_peg_parser> inner_fields;
+
+        if (!call_id_key.empty()) {
+            auto id_parser = atomic(
+                literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
+                literal("\"") + tool_id(string_content('"')) + literal("\"")
+            );
+            inner_fields.push_back(optional(id_parser + space() + optional(literal(",") + space())));
+        }
+
+        if (!gen_call_id_key.empty()) {
+            auto gen_id_parser = atomic(
+                literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
+                choice({
+                    literal("\"") + tool_id(string_content('"')) + literal("\""),
+                    tool_id(json_number())
+                })
+            );
+            inner_fields.push_back(optional(gen_id_parser + space() + optional(literal(",") + space())));
+        }
+
+        // Arguments — either wrapped in args_key or parsed directly
+        common_peg_parser args_parser = eps();
+        if (args_key.empty()) {
+            args_parser = tool_args(schema(json(), "tool-" + name + "-schema", params));
+        } else {
+            args_parser = literal("\"" + effective_args_key + "\"") + space() + literal(":") + space() +
+                          tool_args(schema(json(), "tool-" + name + "-schema", params));
+        }
+        inner_fields.push_back(args_parser);
+
+        // Build inner object parser
+        common_peg_parser inner_object = eps();
+        if (args_key.empty() && inner_fields.size() == 1) {
+            inner_object = inner_fields[0];
+        } else {
+            inner_object = literal("{") + space();
+            for (size_t i = 0; i < inner_fields.size(); i++) {
+                inner_object = inner_object + inner_fields[i];
+                if (i < inner_fields.size() - 1) {
+                    inner_object = inner_object + space();
+                }
+            }
+            inner_object = inner_object + space() + literal("}");
+        }
+
+        auto tool_parser = tool(
+            tool_open(literal("{")) + space() +
+            literal("\"") + tool_name(literal(name)) + literal("\"") +
+            space() + literal(":") + space() +
+            inner_object +
+            space() + tool_close(literal("}"))
+        );
+
+        tool_choices |= rule("tool-" + name, tool_parser);
+    }
+
+    return tool_choices;
+}
+
+// Mode 2: Nested keys (dot notation like "function.name")
+common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
+    const ordered_json & tools,
+    const std::string &  effective_name_key,
+    const std::string &  effective_args_key,
+    const std::string &  call_id_key,
+    const std::string &  gen_call_id_key) {
+
+    auto tool_choices = choice();
+
+    auto name_spec = parse_key_spec(effective_name_key);
+    auto args_spec = parse_key_spec(effective_args_key);
+
+    std::string nested_prefix     = !name_spec.first.empty() ? name_spec.first  : args_spec.first;
+    std::string nested_name_field = !name_spec.first.empty() ? name_spec.second  : effective_name_key;
+    std::string nested_args_field = !args_spec.first.empty() ? args_spec.second  : effective_args_key;
+
+    for (const auto & tool_def : tools) {
+        if (!tool_def.contains("function")) {
+            continue;
+        }
+        const auto &   function = tool_def.at("function");
+        std::string    name     = function.at("name");
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+
+        auto nested_name = literal("\"" + nested_name_field + "\"") + space() + literal(":") + space() +
+                          literal("\"") + tool_name(literal(name)) + literal("\"");
+        auto nested_args = literal("\"" + nested_args_field + "\"") + space() + literal(":") + space() +
+                          tool_args(schema(json(), "tool-" + name + "-schema", params));
+
+        auto nested_object = literal("{") + space() +
+                            nested_name + space() + literal(",") + space() +
+                            nested_args +
+                            space() + literal("}");
+
+        // Format: { id?, "function": {...} }
+        auto tool_parser_body = tool_open(literal("{")) + space();
+
+        if (!call_id_key.empty()) {
+            auto id_spec = parse_key_spec(call_id_key);
+            if (id_spec.first.empty()) {
+                auto id_parser = atomic(
+                    literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
+                    literal("\"") + tool_id(string_content('"')) + literal("\"")
+                );
+                tool_parser_body = tool_parser_body + optional(id_parser + space() + literal(",") + space());
+            }
+        }
+
+        if (!gen_call_id_key.empty()) {
+            auto gen_id_spec = parse_key_spec(gen_call_id_key);
+            if (gen_id_spec.first.empty()) {
+                auto gen_id_parser = atomic(
+                    literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
+                    choice({
+                        literal("\"") + tool_id(string_content('"')) + literal("\""),
+                        tool_id(json_number())
+                    })
+                );
+                tool_parser_body = tool_parser_body + optional(gen_id_parser + space() + literal(",") + space());
+            }
+        }
+
+        auto nested_field = literal("\"" + nested_prefix + "\"") + space() + literal(":") + space() + nested_object;
+        tool_parser_body = tool_parser_body + nested_field + space() + tool_close(literal("}"));
+
+        tool_choices |= rule("tool-" + name, tool(tool_parser_body));
+    }
+
+    return tool_choices;
+}
+
+// Mode 3: Flat keys with optional ID fields and parameter ordering
+common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
+    const ordered_json &             tools,
+    const std::string &              effective_name_key,
+    const std::string &              effective_args_key,
+    const std::string &              call_id_key,
+    const std::string &              gen_call_id_key,
+    const std::vector<std::string> & parameters_order) {
+
+    auto tool_choices    = choice();
+    auto name_key_parser = literal("\"" + effective_name_key + "\"");
+    auto args_key_parser = literal("\"" + effective_args_key + "\"");
+
+    for (const auto & tool_def : tools) {
+        if (!tool_def.contains("function")) {
+            continue;
+        }
+        const auto &   function = tool_def.at("function");
+        std::string    name     = function.at("name");
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
+
+        auto tool_name_ = name_key_parser + space() + literal(":") + space() +
+                         literal("\"") + tool_name(literal(name)) + literal("\"");
+        auto tool_args_ = args_key_parser + space() + literal(":") + space() +
+                         tool_args(schema(json(), "tool-" + name + "-schema", params));
+
+        // Build ID parsers if keys are provided
+        common_peg_parser id_parser = eps();
+        if (!call_id_key.empty()) {
+            id_parser = atomic(
+                literal("\"" + call_id_key + "\"") + space() + literal(":") + space() +
+                choice({
+                    literal("\"") + tool_id(string_content('"')) + literal("\""),
+                    tool_id(json_number())
+                })
+            );
+        }
+
+        common_peg_parser gen_id_parser = eps();
+        if (!gen_call_id_key.empty()) {
+            gen_id_parser = atomic(
+                literal("\"" + gen_call_id_key + "\"") + space() + literal(":") + space() +
+                choice({
+                    literal("\"") + tool_id(string_content('"')) + literal("\""),
+                    tool_id(json_number())
+                })
+            );
+        }
+
+        // Create (parser, key) pairs for all fields, then sort by parameters_order
+        std::vector<std::pair<common_peg_parser, std::string>> parser_pairs;
+        parser_pairs.emplace_back(tool_name_, effective_name_key);
+        parser_pairs.emplace_back(tool_args_, effective_args_key);
+        if (!call_id_key.empty()) {
+            parser_pairs.emplace_back(optional(id_parser), call_id_key);
+        }
+        if (!gen_call_id_key.empty()) {
+            parser_pairs.emplace_back(optional(gen_id_parser), gen_call_id_key);
+        }
+
+        std::sort(parser_pairs.begin(), parser_pairs.end(),
+            [&parameters_order](const auto & a, const auto & b) {
+                auto pos_a = std::find(parameters_order.begin(), parameters_order.end(), a.second);
+                auto pos_b = std::find(parameters_order.begin(), parameters_order.end(), b.second);
+                size_t idx_a = (pos_a == parameters_order.end()) ? parameters_order.size() : std::distance(parameters_order.begin(), pos_a);
+                size_t idx_b = (pos_b == parameters_order.end()) ? parameters_order.size() : std::distance(parameters_order.begin(), pos_b);
+                return idx_a < idx_b;
+            });
+
+        auto ordered_body = tool_open(literal("{")) + space();
+        for (size_t i = 0; i < parser_pairs.size(); i++) {
+            ordered_body = ordered_body + parser_pairs[i].first;
+            if (i < parser_pairs.size() - 1) {
+                ordered_body = ordered_body + space() + literal(",") + space();
+            }
+        }
+        ordered_body = ordered_body + space() + tool_close(literal("}"));
+
+        tool_choices |= rule("tool-" + name, tool(ordered_body));
+    }
+
+    return tool_choices;
+}
+
+common_peg_parser common_chat_peg_builder::standard_json_tools(
+                                                       const std::string &              section_start,
+                                                       const std::string &              section_end,
+                                                       const ordered_json &             tools,
+                                                       bool                             parallel_tool_calls,
+                                                       bool                             force_tool_calls,
+                                                       const std::string &              name_key,
+                                                       const std::string &              args_key,
+                                                       bool                             array_wrapped,
+                                                       bool                             function_is_key,
+                                                       const std::string &              call_id_key,
+                                                       const std::string &              gen_call_id_key,
+                                                       const std::vector<std::string> & parameters_order) {
+    if (!tools.is_array() || tools.empty()) {
+        return eps();
+    }
+
+    std::string effective_name_key = name_key.empty() ? "name" : name_key;
+    std::string effective_args_key = args_key.empty() ? "arguments" : args_key;
+
+    // Dispatch to the appropriate builder based on the JSON layout mode
+    common_peg_parser tool_choices = eps();
+    if (function_is_key) {
+        tool_choices = build_json_tools_function_is_key(tools, args_key, effective_args_key, call_id_key, gen_call_id_key);
+    } else {
+        auto name_spec = parse_key_spec(effective_name_key);
+        auto args_spec = parse_key_spec(effective_args_key);
+        if (!name_spec.first.empty() || !args_spec.first.empty()) {
+            tool_choices = build_json_tools_nested_keys(tools, effective_name_key, effective_args_key, call_id_key, gen_call_id_key);
+        } else {
+            tool_choices = build_json_tools_flat_keys(tools, effective_name_key, effective_args_key, call_id_key, gen_call_id_key, parameters_order);
+        }
+    }
+
+    // Build the section with markers
+    auto tool_calls = tool_choices;
+    if (parallel_tool_calls) {
+        tool_calls = tool_calls + zero_or_more(space() + literal(",") + space() + tool_choices);
+    }
+
+    if (array_wrapped) {
+        tool_calls = literal("[") + space() + tool_calls + space() + literal("]");
+    }
+
+    auto section =
+        trigger_rule("tool-call", literal(section_start) + space() + tool_calls + space() + literal(section_end));
+
+    return force_tool_calls ? section : optional(section);
+}

common/chat-peg-parser.h

@@ -3,22 +3,9 @@
 #include "chat.h"
 #include "peg-parser.h"
 
-class common_chat_peg_builder : public common_peg_parser_builder {
-  public:
-    static constexpr const char * REASONING_BLOCK = "reasoning-block";
-    static constexpr const char * REASONING = "reasoning";
-    static constexpr const char * CONTENT = "content";
-
-    common_peg_parser reasoning_block(const common_peg_parser & p) { return tag(REASONING_BLOCK, p); }
-    common_peg_parser reasoning(const common_peg_parser & p) { return tag(REASONING, p); }
-    common_peg_parser content(const common_peg_parser & p) { return tag(CONTENT, p); }
-};
-
-inline common_peg_arena build_chat_peg_parser(const std::function<common_peg_parser(common_chat_peg_builder & builder)> & fn) {
-    common_chat_peg_builder builder;
-    builder.set_root(fn(builder));
-    return builder.build();
-}
+#include <map>
+#include <optional>
+#include <vector>
 
 class common_chat_peg_mapper {
   public:
@@ -26,80 +13,169 @@ class common_chat_peg_mapper {
 
     common_chat_peg_mapper(common_chat_msg & msg) : result(msg) {}
 
+    virtual ~common_chat_peg_mapper() = default;
+
     virtual void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
     virtual void map(const common_peg_ast_node & node);
+    private:
+      // Tool call handling state
+      std::optional<common_chat_tool_call> pending_tool_call;  // Tool call waiting for name
+      common_chat_tool_call *              current_tool          = nullptr;
+      int                                  arg_count             = 0;
+      bool                                 closing_quote_pending = false;
+      std::string                          args_buffer;  // Buffer to delay arguments until tool name is known
+
+      // Returns a reference to the active argument destination string.
+      // Before tool_name is known, writes go to args_buffer; after, to current_tool->arguments.
+      std::string & args_target();
 };
 
-class common_chat_peg_native_builder : public common_chat_peg_builder {
-  public:
-    static constexpr const char * TOOL = "tool";
-    static constexpr const char * TOOL_OPEN = "tool-open";
-    static constexpr const char * TOOL_CLOSE = "tool-close";
-    static constexpr const char * TOOL_ID = "tool-id";
-    static constexpr const char * TOOL_NAME = "tool-name";
-    static constexpr const char * TOOL_ARGS = "tool-args";
+struct content_structure;
+struct tool_call_structure;
 
+class common_chat_peg_builder : public common_peg_parser_builder {
+  public:
+    // Tag constants (from former common_chat_peg_base_builder)
+    static constexpr const char * REASONING_BLOCK = "reasoning-block";
+    static constexpr const char * REASONING       = "reasoning";
+    static constexpr const char * CONTENT         = "content";
+
+    // Tag constants
+    static constexpr const char * TOOL           = "tool";
+    static constexpr const char * TOOL_OPEN      = "tool-open";
+    static constexpr const char * TOOL_CLOSE     = "tool-close";
+    static constexpr const char * TOOL_ID        = "tool-id";
+    static constexpr const char * TOOL_NAME      = "tool-name";
+    static constexpr const char * TOOL_ARGS      = "tool-args";
+    static constexpr const char * TOOL_ARG       = "tool-arg";
+    static constexpr const char * TOOL_ARG_OPEN  = "tool-arg-open";
+    static constexpr const char * TOOL_ARG_CLOSE = "tool-arg-close";
+    static constexpr const char * TOOL_ARG_NAME         = "tool-arg-name";
+    static constexpr const char * TOOL_ARG_VALUE        = "tool-arg-value";
+    static constexpr const char * TOOL_ARG_STRING_VALUE = "tool-arg-string-value";  // For schema-declared string types
+
+    // Low-level tag methods (from former common_chat_peg_base_builder)
+    common_peg_parser reasoning_block(const common_peg_parser & p) { return tag(REASONING_BLOCK, p); }
+
+    common_peg_parser reasoning(const common_peg_parser & p) { return tag(REASONING, p); }
+
+    common_peg_parser content(const common_peg_parser & p) { return tag(CONTENT, p); }
+
+    common_peg_parser tag_with_safe_content(const std::string &       tag_name,
+                        const std::string &       marker,
+                        const common_peg_parser & p);
+
+    // Low-level tag methods
     common_peg_parser tool(const common_peg_parser & p) { return tag(TOOL, p); }
     common_peg_parser tool_open(const common_peg_parser & p) { return atomic(tag(TOOL_OPEN, p)); }
     common_peg_parser tool_close(const common_peg_parser & p) { return atomic(tag(TOOL_CLOSE, p)); }
     common_peg_parser tool_id(const common_peg_parser & p) { return atomic(tag(TOOL_ID, p)); }
     common_peg_parser tool_name(const common_peg_parser & p) { return atomic(tag(TOOL_NAME, p)); }
     common_peg_parser tool_args(const common_peg_parser & p) { return tag(TOOL_ARGS, p); }
-};
-
-class common_chat_peg_native_mapper : public common_chat_peg_mapper {
-    common_chat_tool_call * current_tool;
-
-  public:
-    common_chat_peg_native_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
-
-    void map(const common_peg_ast_node & node) override;
-};
-
-inline common_peg_arena build_chat_peg_native_parser(const std::function<common_peg_parser(common_chat_peg_native_builder & builder)> & fn) {
-    common_chat_peg_native_builder builder;
-    builder.set_root(fn(builder));
-    return builder.build();
-}
-
-class common_chat_peg_constructed_builder : public common_chat_peg_builder {
-  public:
-    static constexpr const char * TOOL = "tool";
-    static constexpr const char * TOOL_OPEN = "tool-open";
-    static constexpr const char * TOOL_CLOSE = "tool-close";
-    static constexpr const char * TOOL_NAME = "tool-name";
-    static constexpr const char * TOOL_ARG = "tool-arg";
-    static constexpr const char * TOOL_ARG_OPEN = "tool-arg-open";
-    static constexpr const char * TOOL_ARG_CLOSE = "tool-arg-close";
-    static constexpr const char * TOOL_ARG_NAME = "tool-arg-name";
-    static constexpr const char * TOOL_ARG_STRING_VALUE = "tool-arg-string-value";
-    static constexpr const char * TOOL_ARG_JSON_VALUE = "tool-arg-json-value";
-
-    common_peg_parser tool(const common_peg_parser & p) { return tag(TOOL, p); }
-    common_peg_parser tool_open(const common_peg_parser & p) { return atomic(tag(TOOL_OPEN, p)); }
-    common_peg_parser tool_close(const common_peg_parser & p) { return atomic(tag(TOOL_CLOSE, p)); }
-    common_peg_parser tool_name(const common_peg_parser & p) { return atomic(tag(TOOL_NAME, p)); }
     common_peg_parser tool_arg(const common_peg_parser & p) { return tag(TOOL_ARG, p); }
     common_peg_parser tool_arg_open(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_OPEN, p)); }
     common_peg_parser tool_arg_close(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_CLOSE, p)); }
     common_peg_parser tool_arg_name(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_NAME, p)); }
+    common_peg_parser tool_arg_value(const common_peg_parser & p) { return tag(TOOL_ARG_VALUE, p); }
+
+    // Use for schema-declared string types - won't be treated as potential JSON container
     common_peg_parser tool_arg_string_value(const common_peg_parser & p) { return tag(TOOL_ARG_STRING_VALUE, p); }
-    common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return tag(TOOL_ARG_JSON_VALUE, p); }
+    common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_VALUE, p)); }
+
+    // Legacy-compatible helper for building standard JSON tool calls
+    // Used by tests and manual parsers
+    // name_key/args_key: JSON key names for function name and arguments
+    //   Empty or "name"/"arguments" will accept both common variations
+    //   Supports dot notation for nested objects (e.g., "function.name")
+    // array_wrapped: if true, tool calls are wrapped in JSON array [...]
+    // function_is_key: if true, function name is the JSON key (e.g., {"func_name": {...}})
+    // call_id_key: JSON key for string call ID (e.g., "id")
+    // gen_call_id_key: JSON key for generated integer call ID (e.g., "tool_call_id")
+    // parameters_order: order in which JSON fields should be parsed
+    common_peg_parser standard_json_tools(const std::string &              section_start,
+                                          const std::string &              section_end,
+                                          const nlohmann::ordered_json &   tools,
+                                          bool                             parallel_tool_calls,
+                                          bool                             force_tool_calls,
+                                          const std::string &              name_key = "",
+                                          const std::string &              args_key = "",
+                                          bool                             array_wrapped = false,
+                                          bool                             function_is_key = false,
+                                          const std::string &              call_id_key = "",
+                                          const std::string &              gen_call_id_key = "",
+                                          const std::vector<std::string> & parameters_order = {});
+
+    // Legacy-compatible helper for building XML/tagged style tool calls
+    // Used by tests and manual parsers
+    common_peg_parser standard_constructed_tools(const std::map<std::string, std::string> & markers,
+                                                 const nlohmann::ordered_json &             tools,
+                                                 bool                                       parallel_tool_calls,
+                                                 bool                                       force_tool_calls);
+
+    // Helper for Python-style function call format: name(arg1="value1", arg2=123)
+    // Used by LFM2 and similar templates
+    common_peg_parser python_style_tool_calls(const nlohmann::ordered_json & tools,
+                                              bool                           parallel_tool_calls);
+
+  private:
+    // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
+    common_peg_parser build_json_tools_function_is_key(const nlohmann::ordered_json & tools,
+                                                       const std::string &            args_key,
+                                                       const std::string &            effective_args_key,
+                                                       const std::string &            call_id_key,
+                                                       const std::string &            gen_call_id_key);
+
+    common_peg_parser build_json_tools_nested_keys(const nlohmann::ordered_json & tools,
+                                                   const std::string &            effective_name_key,
+                                                   const std::string &            effective_args_key,
+                                                   const std::string &            call_id_key,
+                                                   const std::string &            gen_call_id_key);
+
+    common_peg_parser build_json_tools_flat_keys(const nlohmann::ordered_json &   tools,
+                                                 const std::string &              effective_name_key,
+                                                 const std::string &              effective_args_key,
+                                                 const std::string &              call_id_key,
+                                                 const std::string &              gen_call_id_key,
+                                                 const std::vector<std::string> & parameters_order);
 };
 
-class common_chat_peg_constructed_mapper : public common_chat_peg_mapper {
-    common_chat_tool_call * current_tool;
-    int arg_count = 0;
-    bool needs_closing_quote = false;
-
-  public:
-    common_chat_peg_constructed_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
-
-    void map(const common_peg_ast_node & node) override;
-};
-
-inline common_peg_arena build_chat_peg_constructed_parser(const std::function<common_peg_parser(common_chat_peg_constructed_builder & builder)> & fn) {
-    common_chat_peg_constructed_builder builder;
-    builder.set_root(fn(builder));
-    return builder.build();
+inline common_peg_arena build_chat_peg_parser(
+  const std::function<common_peg_parser(common_chat_peg_builder & builder)> & fn) {
+  common_chat_peg_builder builder;
+  builder.set_root(fn(builder));
+  return builder.build();
 }
+
+class tag_based_peg_mapper {
+  public:
+    std::map<std::string, std::string> tags;
+
+    void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
+};
+
+struct tagged_parse_result {
+    common_peg_parse_result              result;
+    std::map<std::string, std::string> tags;
+};
+
+struct tagged_peg_parser {
+    common_peg_arena arena;
+    common_peg_parse_flags flags = COMMON_PEG_PARSE_FLAG_NONE;
+
+    tagged_peg_parser & withDebug() {
+      flags |= COMMON_PEG_PARSE_FLAG_DEBUG;
+      return *this;
+    }
+
+    tagged_peg_parser & withoutDebug() {
+      flags = flags & ~COMMON_PEG_PARSE_FLAG_DEBUG;
+      return *this;
+    }
+
+    tagged_parse_result parse_and_extract(const std::string & input, common_peg_parse_flags extra_flags = COMMON_PEG_PARSE_FLAG_NONE) const;
+    tagged_parse_result parse_anywhere_and_extract(const std::string & input) const;
+};
+
+tagged_peg_parser build_tagged_peg_parser(
+    const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn);
+

common/chat.h

@@ -3,17 +3,30 @@
 #pragma once
 
 #include "common.h"
+#include "jinja/parser.h"
+#include "nlohmann/json_fwd.hpp"
 #include "peg-parser.h"
-#include <functional>
+#include "jinja/runtime.h"
+#include "jinja/caps.h"
+#include "nlohmann/json.hpp"
+
 #include <chrono>
+#include <functional>
+#include <map>
 #include <string>
 #include <vector>
-#include <map>
+
+using chat_template_caps = jinja::caps;
+using json = nlohmann::ordered_json;
 
 #include <nlohmann/json_fwd.hpp>
 
 struct common_chat_templates;
 
+namespace autoparser {
+struct templates_params;
+}  // namespace autoparser
+
 struct common_chat_tool_call {
     std::string name;
     std::string arguments;
@@ -38,21 +51,85 @@ struct common_chat_msg_content_part {
     }
 };
 
+struct common_chat_template {
+    jinja::program prog;
+    std::string bos_tok;
+    std::string eos_tok;
+    std::string src;
+    chat_template_caps caps;
+
+    common_chat_template(const std::string & src, const std::string & bos_token, const std::string & eos_token) {
+        jinja::lexer lexer;
+        auto lexer_res = lexer.tokenize(src);
+        this->prog = jinja::parse_from_tokens(lexer_res);
+
+        this->src = lexer_res.source;
+        this->bos_tok = bos_token;
+        this->eos_tok = eos_token;
+
+        this->caps = jinja::caps_get(prog);
+        // LOG_INF("%s: caps:\n%s\n", __func__, this->caps.to_string().c_str());
+    }
+
+    const std::string & source() const { return src; }
+    const std::string & bos_token() const { return bos_tok; }
+    const std::string & eos_token() const { return eos_tok; }
+
+    // TODO: this is ugly, refactor it somehow
+    json add_system(const json & messages, const std::string & system_prompt) const {
+        GGML_ASSERT(messages.is_array());
+        auto msgs_copy = messages;
+        if (!caps.supports_system_role) {
+            if (msgs_copy.empty()) {
+                msgs_copy.insert(msgs_copy.begin(), json{
+                    {"role", "user"},
+                    {"content", system_prompt}
+                });
+            } else {
+                auto & first_msg = msgs_copy[0];
+                if (!first_msg.contains("content")) {
+                    first_msg["content"] = "";
+                }
+                first_msg["content"] = system_prompt + "\n\n"
+                    + first_msg["content"].get<std::string>();
+            }
+        } else {
+            if (msgs_copy.empty() || msgs_copy[0].at("role") != "system") {
+                msgs_copy.insert(msgs_copy.begin(), json{
+                    {"role", "system"},
+                    {"content", system_prompt}
+                });
+            } else if (msgs_copy[0].at("role") == "system") {
+                msgs_copy[0]["content"] = system_prompt;
+            }
+        }
+        return msgs_copy;
+    }
+
+    chat_template_caps original_caps() const {
+        return caps;
+    }
+
+};
+
 struct common_chat_msg {
-    std::string role;
-    std::string content;
+    std::string                               role;
+    std::string                               content;
     std::vector<common_chat_msg_content_part> content_parts;
-    std::vector<common_chat_tool_call> tool_calls;
-    std::string reasoning_content;
-    std::string tool_name;
-    std::string tool_call_id;
+    std::vector<common_chat_tool_call>        tool_calls;
+    std::string                               reasoning_content;
+    std::string                               tool_name;
+    std::string                               tool_call_id;
 
     nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
 
     bool empty() const {
-        return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
+        return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() &&
+               tool_name.empty() && tool_call_id.empty();
     }
-    void set_tool_call_ids(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
+
+    void set_tool_call_ids(std::vector<std::string> &           ids_cache,
+                           const std::function<std::string()> & gen_tool_call_id) {
         for (auto i = 0u; i < tool_calls.size(); i++) {
             if (ids_cache.size() <= i) {
                 auto id = tool_calls[i].id;
@@ -64,32 +141,28 @@ struct common_chat_msg {
             tool_calls[i].id = ids_cache[i];
         }
     }
+
     bool operator==(const common_chat_msg & other) const {
-        return role == other.role
-            && content == other.content
-            && content_parts == other.content_parts
-            && tool_calls == other.tool_calls
-            && reasoning_content == other.reasoning_content
-            && tool_name == other.tool_name
-            && tool_call_id == other.tool_call_id;
-    }
-    bool operator!=(const common_chat_msg & other) const {
-        return !(*this == other);
+        return role == other.role && content == other.content && content_parts == other.content_parts &&
+               tool_calls == other.tool_calls && reasoning_content == other.reasoning_content &&
+               tool_name == other.tool_name && tool_call_id == other.tool_call_id;
     }
+
+    bool operator!=(const common_chat_msg & other) const { return !(*this == other); }
 };
 
 struct common_chat_msg_diff {
-    std::string reasoning_content_delta;
-    std::string content_delta;
-    size_t tool_call_index = std::string::npos;
+    std::string           reasoning_content_delta;
+    std::string           content_delta;
+    size_t                tool_call_index = std::string::npos;
     common_chat_tool_call tool_call_delta;
 
-    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & msg_prv, const common_chat_msg & msg_new);
+    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & msg_prv,
+                                                           const common_chat_msg & msg_new);
 
     bool operator==(const common_chat_msg_diff & other) const {
-        return content_delta == other.content_delta
-        && tool_call_index == other.tool_call_index
-        && tool_call_delta == other.tool_call_delta;
+        return content_delta == other.content_delta && tool_call_index == other.tool_call_index &&
+               tool_call_delta == other.tool_call_delta;
     }
 };
 
@@ -107,64 +180,41 @@ enum common_chat_tool_choice {
 
 enum common_chat_format {
     COMMON_CHAT_FORMAT_CONTENT_ONLY,
-    COMMON_CHAT_FORMAT_GENERIC,
-    COMMON_CHAT_FORMAT_MISTRAL_NEMO,
-    COMMON_CHAT_FORMAT_MAGISTRAL,
-    COMMON_CHAT_FORMAT_LLAMA_3_X,
-    COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
-    COMMON_CHAT_FORMAT_DEEPSEEK_R1,
-    COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
-    COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
-    COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
-    COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-    COMMON_CHAT_FORMAT_HERMES_2_PRO,
-    COMMON_CHAT_FORMAT_COMMAND_R7B,
-    COMMON_CHAT_FORMAT_GRANITE,
-    COMMON_CHAT_FORMAT_GPT_OSS,
-    COMMON_CHAT_FORMAT_SEED_OSS,
-    COMMON_CHAT_FORMAT_NEMOTRON_V2,
-    COMMON_CHAT_FORMAT_APERTUS,
-    COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS,
-    COMMON_CHAT_FORMAT_GLM_4_5,
-    COMMON_CHAT_FORMAT_MINIMAX_M2,
-    COMMON_CHAT_FORMAT_KIMI_K2,
-    COMMON_CHAT_FORMAT_APRIEL_1_5,
-    COMMON_CHAT_FORMAT_XIAOMI_MIMO,
-    COMMON_CHAT_FORMAT_SOLAR_OPEN,
-    COMMON_CHAT_FORMAT_EXAONE_MOE,
 
     // These are intended to be parsed by the PEG parser
     COMMON_CHAT_FORMAT_PEG_SIMPLE,
     COMMON_CHAT_FORMAT_PEG_NATIVE,
-    COMMON_CHAT_FORMAT_PEG_CONSTRUCTED,
 
-    COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
+    COMMON_CHAT_FORMAT_COUNT,  // Not a format, just the # formats
 };
 
 struct common_chat_templates_inputs {
-    std::vector<common_chat_msg> messages;
-    std::string grammar;
-    std::string json_schema;
-    bool add_generation_prompt = true;
-    bool use_jinja = true;
+    std::vector<common_chat_msg>          messages;
+    std::string                           grammar;
+    std::string                           json_schema;
+    bool                                  add_generation_prompt = true;
+    bool                                  use_jinja             = true;
     // Parameters below only supported when use_jinja is true
-    std::vector<common_chat_tool> tools;
-    common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
-    bool parallel_tool_calls = false;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool enable_thinking"
-    bool enable_thinking = true;
-    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
-    std::map<std::string, std::string> chat_template_kwargs;
-    bool add_bos = false;
-    bool add_eos = false;
+    std::vector<common_chat_tool>         tools;
+    common_chat_tool_choice               tool_choice         = COMMON_CHAT_TOOL_CHOICE_AUTO;
+    bool                                  parallel_tool_calls = false;
+    common_reasoning_format               reasoning_format    = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool enable_thinking"
+    bool                                  enable_thinking     = true;
+    std::chrono::system_clock::time_point now                 = std::chrono::system_clock::now();
+    std::map<std::string, std::string>    chat_template_kwargs;
+    bool                                  add_bos = false;
+    bool                                  add_eos = false;
 };
 
 struct common_chat_params {
     common_chat_format                  format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     std::string                         prompt;
     std::string                         grammar;
-    bool                                grammar_lazy = false;
+    bool                                grammar_lazy         = false;
     bool                                thinking_forced_open = false;
+    bool                                supports_thinking    = false;
+    std::string                         thinking_start_tag;  // e.g., "<think>"
+    std::string                         thinking_end_tag;    // e.g., "</think>"
     std::vector<common_grammar_trigger> grammar_triggers;
     std::vector<std::string>            preserved_tokens;
     std::vector<std::string>            additional_stops;
@@ -174,13 +224,14 @@ struct common_chat_params {
 // per-message parsing syntax
 // should be derived from common_chat_params
 struct common_chat_parser_params {
-    common_chat_format       format                = COMMON_CHAT_FORMAT_CONTENT_ONLY;
-    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool parse_reasoning"
+    common_chat_format      format               = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    common_reasoning_format reasoning_format     = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool parse_reasoning"
     // Whether reasoning_content should be inlined in the content (e.g. for reasoning_format=deepseek in stream mode)
-    bool                     reasoning_in_content  = false;
-    bool                     thinking_forced_open  = false;
-    bool                     parse_tool_calls      = true;
-    common_peg_arena         parser                = {};
+    bool                    reasoning_in_content = false;
+    bool                    thinking_forced_open = false;
+    bool                    parse_tool_calls     = true;
+    bool                    debug                = false;  // Enable debug output for PEG parser
+    common_peg_arena        parser               = {};
     common_chat_parser_params() = default;
     common_chat_parser_params(const common_chat_params & chat_params) {
         format               = chat_params.format;
@@ -193,45 +244,42 @@ bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
 
 void common_chat_templates_free(struct common_chat_templates * tmpls);
 
-struct common_chat_templates_deleter { void operator()(common_chat_templates * tmpls) { common_chat_templates_free(tmpls); } };
+struct common_chat_templates_deleter {
+    void operator()(common_chat_templates * tmpls) { common_chat_templates_free(tmpls); }
+};
 
 typedef std::unique_ptr<struct common_chat_templates, common_chat_templates_deleter> common_chat_templates_ptr;
 
-common_chat_templates_ptr common_chat_templates_init(
-                                    const struct llama_model * model,
-                                           const std::string & chat_template_override,
-                                           const std::string & bos_token_override = "",
-                                           const std::string & eos_token_override = "");
+common_chat_templates_ptr common_chat_templates_init(const struct llama_model * model,
+                                                     const std::string &        chat_template_override,
+                                                     const std::string &        bos_token_override = "",
+                                                     const std::string &        eos_token_override = "");
 
 bool         common_chat_templates_was_explicit(const struct common_chat_templates * tmpls);
 std::string  common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant = "");
 
-
-struct common_chat_params      common_chat_templates_apply(
-    const struct common_chat_templates * tmpls,
-    const struct common_chat_templates_inputs & inputs);
+struct common_chat_params common_chat_templates_apply(const struct common_chat_templates *        tmpls,
+                                                      const struct common_chat_templates_inputs & inputs);
 
 // Format single message, while taking into account the position of that message in chat history
-std::string common_chat_format_single(
-        const struct common_chat_templates * tmpls,
-        const std::vector<common_chat_msg> & past_msg,
-        const common_chat_msg & new_msg,
-        bool add_ass,
-        bool use_jinja);
+std::string common_chat_format_single(const struct common_chat_templates * tmpls,
+                                      const std::vector<common_chat_msg> & past_msg,
+                                      const common_chat_msg &              new_msg,
+                                      bool                                 add_ass,
+                                      bool                                 use_jinja);
 
 // Returns an example of formatted chat
-std::string common_chat_format_example(
-    const struct common_chat_templates * tmpls,
-    bool use_jinja,
-    const std::map<std::string, std::string> & chat_template_kwargs);
+std::string common_chat_format_example(const struct common_chat_templates *       tmpls,
+                                       bool                                       use_jinja,
+                                       const std::map<std::string, std::string> & chat_template_kwargs);
 
-const char*               common_chat_format_name(common_chat_format format);
-common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
-common_chat_msg           common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
+const char *            common_chat_format_name(common_chat_format format);
+common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_parser_params & params);
+common_chat_msg           common_chat_peg_parse(const common_peg_arena & src_parser, const std::string & input, bool is_partial, const common_chat_parser_params & params);
 
 // used by arg and server
-const char *             common_reasoning_format_name(common_reasoning_format format);
-common_reasoning_format  common_reasoning_format_from_name(const std::string & format);
+const char *            common_reasoning_format_name(common_reasoning_format format);
+common_reasoning_format common_reasoning_format_from_name(const std::string & format);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 
@@ -250,3 +298,10 @@ nlohmann::ordered_json common_chat_msg_diff_to_json_oaicompat(const common_chat_
 
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
+
+std::string common_chat_template_direct_apply(
+    const common_chat_template & tmpl,
+    const autoparser::templates_params & inputs,
+    const std::optional<json> & messages_override = std::nullopt,
+    const std::optional<json> & tools_override = std::nullopt,
+    const std::optional<json> & additional_context = std::nullopt);

common/common.cpp

@@ -676,7 +676,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
 
     size_t offset = 0;
     while (offset < filename.size()) {
-        utf8_parse_result result = parse_utf8_codepoint(filename, offset);
+        utf8_parse_result result = common_parse_utf8_codepoint(filename, offset);
 
         if (result.status != utf8_parse_result::SUCCESS) {
             return false;

common/common.h

@@ -104,6 +104,8 @@ enum llama_example {
     LLAMA_EXAMPLE_DIFFUSION,
     LLAMA_EXAMPLE_FINETUNE,
     LLAMA_EXAMPLE_FIT_PARAMS,
+    LLAMA_EXAMPLE_RESULTS,
+    LLAMA_EXAMPLE_EXPORT_GRAPH_OPS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -234,6 +236,14 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
+    // reasoning budget sampler parameters
+    // these are populated by the server/CLI based on chat template params
+    int32_t                  reasoning_budget_tokens   = -1;   // -1 = disabled, >= 0 = token budget
+    bool                     reasoning_budget_activate_immediately = false;
+    std::vector<llama_token> reasoning_budget_start;           // start tag token sequence
+    std::vector<llama_token> reasoning_budget_end;             // end tag token sequence
+    std::vector<llama_token> reasoning_budget_forced;          // forced sequence (message + end tag)
+
     bool backend_sampling = false;
 
     bool has_logit_bias() const {
@@ -456,6 +466,8 @@ struct common_params {
 
     bool   kl_divergence    = false; // compute KL divergence
 
+    bool check             = false; // check rather than generate results for llama-results
+
     bool usage             = false; // print usage
     bool completion        = false; // print source-able completion script
     bool use_color         = false; // use color to distinguish generations and inputs
@@ -516,14 +528,15 @@ struct common_params {
     std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
-    int32_t port              = 8080;         // server listens on this network port
-    int32_t timeout_read      = 600;          // http read timeout in seconds
-    int32_t timeout_write     = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
-    int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
-    bool    cache_prompt      = true;         // whether to enable prompt caching
-    int32_t n_ctx_checkpoints = 8;            // max number of context checkpoints per slot
-    int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
+    int32_t port                = 8080;          // server listens on this network port
+    int32_t timeout_read        = 600;           // http read timeout in seconds
+    int32_t timeout_write       = timeout_read;  // http write timeout in seconds
+    int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting
+    bool    cache_prompt        = true;  // whether to enable prompt caching
+    int32_t n_ctx_checkpoints   = 32;     // max number of context checkpoints per slot
+    int32_t checkpoint_every_nt = 8192;   // make a checkpoint every n tokens during prefill
+    int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -532,7 +545,9 @@ struct common_params {
     bool use_jinja = true;                                                                                  // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    int enable_reasoning = -1; // -1 = auto, 0 = disable, 1 = enable
     int reasoning_budget = -1;
+    std::string reasoning_budget_message; // message injected before end tag when budget exhausted
     bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
     int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time
 
@@ -545,6 +560,7 @@ struct common_params {
 
     // webui configs
     bool webui = true;
+    bool webui_mcp_proxy = false;
     std::string webui_config_json;
 
     // "advanced" endpoints are disabled by default for better security
@@ -911,7 +927,7 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 // MoE utils
 //
 
-const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate|gate_up)_(ch|)exps";
 
 inline std::string llm_ffn_exps_block_regex(int idx) {
     return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);

common/http.h

@@ -7,6 +7,7 @@ struct common_http_url {
     std::string user;
     std::string password;
     std::string host;
+    int port;
     std::string path;
 };
 
@@ -47,6 +48,20 @@ static common_http_url common_http_parse_url(const std::string & url) {
         parts.host = rest;
         parts.path = "/";
     }
+
+    auto colon_pos = parts.host.find(':');
+
+    if (colon_pos != std::string::npos) {
+        parts.port = std::stoi(parts.host.substr(colon_pos + 1));
+        parts.host = parts.host.substr(0, colon_pos);
+    } else if (parts.scheme == "http") {
+        parts.port = 80;
+    } else if (parts.scheme == "https") {
+        parts.port = 443;
+    } else {
+        throw std::runtime_error("unsupported URL scheme: " + parts.scheme);
+    }
+
     return parts;
 }
 
@@ -68,7 +83,7 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
     }
 #endif
 
-    httplib::Client cli(parts.scheme + "://" + parts.host);
+    httplib::Client cli(parts.scheme + "://" + parts.host + ":" + std::to_string(parts.port));
 
     if (!parts.user.empty()) {
         cli.set_basic_auth(parts.user, parts.password);

common/jinja/caps.cpp

@@ -1,3 +1,4 @@
+#include "log.h"
 #include "value.h"
 #include "runtime.h"
 #include "caps.h"
@@ -36,12 +37,16 @@ static void caps_try_execute(jinja::program & prog,
     auto tools = ctx.get_val("tools");
 
     bool success = false;
+    std::string result;
     try {
         jinja::runtime runtime(ctx);
-        runtime.execute(prog);
+        auto results = runtime.execute(prog);
+        auto parts = jinja::runtime::gather_string_parts(results);
+        result = parts->as_string().str();
         success = true;
     } catch (const std::exception & e) {
         JJ_DEBUG("Exception during execution: %s", e.what());
+        result = "";
         // ignore exceptions during capability analysis
     }
 
@@ -90,6 +95,8 @@ caps caps_get(jinja::program & prog) {
         return v->stats.ops.find(op_name) != v->stats.ops.end();
     };
 
+    JJ_DEBUG("%s\n", ">>> Running capability check: typed content");
+
     // case: typed content support
     caps_try_execute(
         prog,
@@ -120,6 +127,7 @@ caps caps_get(jinja::program & prog) {
         }
     );
 
+    JJ_DEBUG("%s\n", ">>> Running capability check: system prompt");
 
     // case: system prompt support
     caps_try_execute(
@@ -150,7 +158,9 @@ caps caps_get(jinja::program & prog) {
         }
     );
 
-    // case: tools support
+    JJ_DEBUG("%s\n", ">>> Running capability check: single tool support");
+
+    // case: tools support: single call
     caps_try_execute(
         prog,
         [&]() {
@@ -162,10 +172,10 @@ caps caps_get(jinja::program & prog) {
                 },
                 {
                     {"role", "assistant"},
-                    {"content", "Assistant message"},
+                    {"content", ""}, // Some templates expect content to be empty with tool calls
                     {"tool_calls", json::array({
                         {
-                            {"id", "call1"},
+                            {"id", "call00001"},
                             {"type", "function"},
                             {"function", {
                                 {"name", "tool1"},
@@ -173,19 +183,18 @@ caps caps_get(jinja::program & prog) {
                                     {"arg", "value"}
                                 }}
                             }}
-                        },
-                        {
-                            {"id", "call2"},
-                            {"type", "function"},
-                            {"function", {
-                                {"name", "tool2"},
-                                {"arguments", {
-                                    {"arg", "value"}
-                                }}
-                            }}
                         }
                     })}
                 },
+                {
+                    {"role", "tool"},
+                    {"content", "Tool response"},
+                    {"tool_call_id", "call00001"}
+                },
+                {
+                    {"role", "assistant"},
+                    {"content", "The tool response was 'tool response'"}
+                },
                 {
                     {"role", "user"},
                     {"content", "User message"},
@@ -199,7 +208,7 @@ caps caps_get(jinja::program & prog) {
                     {"name", "tool"},
                     {"type", "function"},
                     {"function", {
-                        {"name", "tool"},
+                        {"name", "tool1"},
                         {"description", "Tool description"},
                         {"parameters", {
                             {"type", "object"},
@@ -224,6 +233,7 @@ caps caps_get(jinja::program & prog) {
 
             auto & tool_name = tools->at(0)->at("function")->at("name");
             caps_print_stats(tool_name, "tools[0].function.name");
+            caps_print_stats(tools, "tools");
             if (!tool_name->stats.used) {
                 result.supports_tools = false;
             }
@@ -233,6 +243,93 @@ caps caps_get(jinja::program & prog) {
             if (!tool_calls->stats.used) {
                 result.supports_tool_calls = false;
             }
+        }
+    );
+
+    JJ_DEBUG("%s\n", ">>> Running capability check: parallel tool support");
+
+    // case: tools support: parallel calls
+    caps_try_execute(
+        prog,
+        [&]() {
+            // messages
+            return json::array({
+                {
+                    {"role", "user"},
+                    {"content", "User message"},
+                },
+                {
+                    {"role", "assistant"},
+                    {"content", ""}, // Some templates expect content to be empty with tool calls
+                    {"tool_calls", json::array({
+                        {
+                            {"id", "call00001"},
+                            {"type", "function"},
+                            {"function", {
+                                {"name", "tool1"},
+                                {"arguments", {
+                                    {"arg", "value"}
+                                }}
+                            }}
+                        },
+                        {
+                            {"id", "call00002"},
+                            {"type", "function"},
+                            {"function", {
+                                {"name", "tool1"},
+                                {"arguments", {
+                                    {"arg", "value"}
+                                }}
+                            }}
+                        }
+                    })}
+                },
+                {
+                    {"role", "tool"},
+                    {"content", "Tool response"},
+                    {"tool_call_id", "call00001"}
+                },
+                {
+                    {"role", "assistant"},
+                    {"content", "The tool response was 'tool response'"}
+                },
+                {
+                    {"role", "user"},
+                    {"content", "User message"},
+                },
+            });
+        },
+        [&]() {
+            // tools
+            return json::array({
+                {
+                    {"name", "tool"},
+                    {"type", "function"},
+                    {"function", {
+                        {"name", "tool1"},
+                        {"description", "Tool description"},
+                        {"parameters", {
+                            {"type", "object"},
+                            {"properties", {
+                                {"arg", {
+                                    {"type", "string"},
+                                    {"description", "Arg description"},
+                                }},
+                            }},
+                            {"required", json::array({ "arg" })},
+                        }},
+                    }},
+                },
+            });
+        },
+        [&](bool success, value & messages, value & /*tools*/) {
+            if (!success) {
+                result.supports_parallel_tool_calls = false;
+                return;
+            }
+
+            auto & tool_calls = messages->at(1)->at("tool_calls");;
+            caps_print_stats(tool_calls, "messages[1].tool_calls");
 
             // check for second tool call usage
             auto & tool_call_1 = tool_calls->at(1)->at("function");
@@ -243,6 +340,8 @@ caps caps_get(jinja::program & prog) {
         }
     );
 
+    JJ_DEBUG("%s\n", ">>> Running capability check: preserve reasoning");
+
     // case: preserve reasoning content in chat history
     caps_try_execute(
         prog,

common/jinja/runtime.cpp

@@ -114,8 +114,10 @@ value binary_expression::execute_impl(context & ctx) {
 
     // Logical operators
     if (op.value == "and") {
+        JJ_DEBUG("Executing logical test: %s AND %s", left->type().c_str(), right->type().c_str());
         return left_val->as_bool() ? right->execute(ctx) : std::move(left_val);
     } else if (op.value == "or") {
+        JJ_DEBUG("Executing logical test: %s OR %s", left->type().c_str(), right->type().c_str());
         return left_val->as_bool() ? std::move(left_val) : right->execute(ctx);
     }
 
@@ -838,7 +840,7 @@ value call_expression::execute_impl(context & ctx) {
     for (auto & arg_stmt : this->args) {
         auto arg_val = arg_stmt->execute(ctx);
         JJ_DEBUG("  Argument type: %s", arg_val->type().c_str());
-        args.push_back(std::move(arg_val));
+        args.push_back(arg_val);
     }
     // execute callee
     value callee_val = callee->execute(ctx);

common/jinja/value.h

@@ -12,8 +12,8 @@
 #include <set>
 #include <sstream>
 #include <string>
-#include <unordered_map>
 #include <vector>
+#include <unordered_map>
 
 namespace jinja {
 

common/json-schema-to-grammar.cpp

@@ -27,11 +27,11 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     if (separator_rule.empty()) {
         if (min_items == 1 && !has_max) {
             return item_rule + "+";
-        } else if (min_items == 0 && !has_max) {
-            return item_rule + "*";
-        } else {
-            return item_rule + "{" + std::to_string(min_items) + "," + (has_max ? std::to_string(max_items) : "") + "}";
         }
+        if (min_items == 0 && !has_max) {
+            return item_rule + "*";
+        }
+        return item_rule + "{" + std::to_string(min_items) + "," + (has_max ? std::to_string(max_items) : "") + "}";
     }
 
     auto result = item_rule + " " + build_repetition("(" + separator_rule + " " + item_rule + ")", min_items == 0 ? 0 : min_items - 1, has_max ? max_items - 1 : max_items);
@@ -41,7 +41,7 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
-static void _build_min_max_int(int64_t min_value, int64_t max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
+static void build_min_max_int(int64_t min_value, int64_t max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
     auto has_min = min_value != std::numeric_limits<int64_t>::min();
     auto has_max = max_value != std::numeric_limits<int64_t>::max();
 
@@ -128,14 +128,14 @@ static void _build_min_max_int(int64_t min_value, int64_t max_value, std::string
     if (has_min && has_max) {
         if (min_value < 0 && max_value < 0) {
             out << "\"-\" (";
-            _build_min_max_int(-max_value, -min_value, out, decimals_left, /* top_level= */ true);
+            build_min_max_int(-max_value, -min_value, out, decimals_left, /* top_level= */ true);
             out << ")";
             return;
         }
 
         if (min_value < 0) {
             out << "\"-\" (";
-            _build_min_max_int(0, -min_value, out, decimals_left, /* top_level= */ true);
+            build_min_max_int(0, -min_value, out, decimals_left, /* top_level= */ true);
             out << ") | ";
             min_value = 0;
         }
@@ -159,7 +159,7 @@ static void _build_min_max_int(int64_t min_value, int64_t max_value, std::string
     if (has_min) {
         if (min_value < 0) {
             out << "\"-\" (";
-            _build_min_max_int(std::numeric_limits<int64_t>::min(), -min_value, out, decimals_left, /* top_level= */ false);
+            build_min_max_int(std::numeric_limits<int64_t>::min(), -min_value, out, decimals_left, /* top_level= */ false);
             out << ") | [0] | [1-9] ";
             more_digits(0, decimals_left - 1);
         } else if (min_value == 0) {
@@ -194,7 +194,7 @@ static void _build_min_max_int(int64_t min_value, int64_t max_value, std::string
             }
             digit_range(c, c);
             out << " (";
-            _build_min_max_int(std::stoll(min_s.substr(1)), std::numeric_limits<int64_t>::max(), out, less_decimals, /* top_level= */ false);
+            build_min_max_int(std::stoll(min_s.substr(1)), std::numeric_limits<int64_t>::max(), out, less_decimals, /* top_level= */ false);
             out << ")";
             if (c < '9') {
                 out << " | ";
@@ -213,10 +213,10 @@ static void _build_min_max_int(int64_t min_value, int64_t max_value, std::string
                 more_digits(0, less_decimals);
                 out << " | ";
             }
-            _build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
+            build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
         } else {
             out << "\"-\" (";
-            _build_min_max_int(-max_value, std::numeric_limits<int64_t>::max(), out, decimals_left, /* top_level= */ false);
+            build_min_max_int(-max_value, std::numeric_limits<int64_t>::max(), out, decimals_left, /* top_level= */ false);
             out << ")";
         }
         return;
@@ -232,7 +232,7 @@ struct BuiltinRule {
     std::vector<std::string> deps;
 };
 
-std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
+static std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
     {"boolean", {"(\"true\" | \"false\") space", {}}},
     {"decimal-part", {"[0-9]{1,16}", {}}},
     {"integral-part", {"[0] | [1-9] [0-9]{0,15}", {}}},
@@ -247,7 +247,7 @@ std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
     {"null", {"\"null\" space", {}}},
 };
 
-std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
+static std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
     {"date", {"[0-9]{4} \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
     {"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9]{3} )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
     {"date-time", {"date \"T\" time", {"date", "time"}}},
@@ -260,22 +260,26 @@ static bool is_reserved_name(const std::string & name) {
     static const std::unordered_set<std::string> RESERVED_NAMES = [] {
         std::unordered_set<std::string> s;
         s.insert("root");
-        for (const auto & p : PRIMITIVE_RULES) s.insert(p.first);
-        for (const auto & p : STRING_FORMAT_RULES) s.insert(p.first);
+        for (const auto & p : PRIMITIVE_RULES) {
+            s.insert(p.first);
+        }
+        for (const auto & p : STRING_FORMAT_RULES) {
+            s.insert(p.first);
+        }
         return s;
     }();
     return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
 }
 
-std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
-std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
-std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
-std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
+static std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
+static std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
+static std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
+static std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
     {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };
 
-std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
-std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
+static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
+static std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
 
 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
     std::smatch match;
@@ -322,19 +326,19 @@ private:
         if (_rules.find(esc_name) == _rules.end() || _rules[esc_name] == rule) {
             _rules[esc_name] = rule;
             return esc_name;
-        } else {
-            int i = 0;
-            while (_rules.find(esc_name + std::to_string(i)) != _rules.end() && _rules[esc_name + std::to_string(i)] != rule) {
-                i++;
-            }
-            std::string key = esc_name + std::to_string(i);
-            _rules[key] = rule;
-            return key;
         }
+        int i = 0;
+        while (_rules.find(esc_name + std::to_string(i)) != _rules.end() && _rules[esc_name + std::to_string(i)] != rule) {
+            i++;
+        }
+        std::string key = esc_name + std::to_string(i);
+        _rules[key] = rule;
+        return key;
     }
 
     std::string _generate_union_rule(const std::string & name, const std::vector<json> & alt_schemas) {
         std::vector<std::string> rules;
+        rules.reserve(alt_schemas.size());
         for (size_t i = 0; i < alt_schemas.size(); i++) {
             rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
         }
@@ -398,6 +402,7 @@ private:
                 flush_literal();
 
                 std::vector<std::string> results;
+                results.reserve(ret.size());
                 for (const auto & item : ret) {
                     results.push_back(to_rule(item));
                 }
@@ -551,7 +556,7 @@ private:
             TrieNode() : is_end_of_string(false) {}
 
             void insert(const std::string & string) {
-                auto node = this;
+                auto *node = this;
                 for (char c : string) {
                     node = &node->children[c];
                 }
@@ -676,7 +681,7 @@ private:
                 if (ks.empty()) {
                     return res;
                 }
-                std::string k = ks[0];
+                const std::string& k = ks[0];
                 std::string kv_rule_name = prop_kv_rule_names[k];
                 std::string comma_ref = "( \",\" space " + kv_rule_name + " )";
                 if (first_is_optional) {
@@ -779,13 +784,13 @@ public:
                         std::string pointer = ref.substr(ref.find('#') + 1);
                         std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
-                            std::string sel = tokens[i];
+                            const std::string& sel = tokens[i];
                             if (target.is_object() && target.contains(sel)) {
                                 target = target[sel];
                             } else if (target.is_array()) {
                                 size_t sel_index;
                                 try {
-                                    sel_index = std::stoul(sel);
+                                    sel_index = std::stoull(sel);
                                 } catch (const std::invalid_argument & e) {
                                     sel_index = target.size();
                                 }
@@ -802,7 +807,7 @@ public:
                         _refs[ref] = target;
                     }
                 } else {
-                    for (auto & kv : n.items()) {
+                    for (const auto & kv : n.items()) {
                         visit_refs(kv.value());
                     }
                 }
@@ -812,7 +817,7 @@ public:
         visit_refs(schema);
     }
 
-    std::string _generate_constant_rule(const json & value) {
+    static std::string _generate_constant_rule(const json & value) {
         return format_literal(value.dump());
     }
 
@@ -823,10 +828,12 @@ public:
 
         if (schema.contains("$ref")) {
             return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
-        } else if (schema.contains("oneOf") || schema.contains("anyOf")) {
+        }
+        if (schema.contains("oneOf") || schema.contains("anyOf")) {
             std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
             return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
-        } else if (schema_type.is_array()) {
+        }
+        if (schema_type.is_array()) {
             std::vector<json> schema_types;
             for (const auto & t : schema_type) {
                 json schema_copy(schema);
@@ -834,15 +841,18 @@ public:
                 schema_types.push_back(schema_copy);
             }
             return _add_rule(rule_name, _generate_union_rule(name, schema_types));
-        } else if (schema.contains("const")) {
+        }
+        if (schema.contains("const")) {
             return _add_rule(rule_name, _generate_constant_rule(schema["const"]) + " space");
-        } else if (schema.contains("enum")) {
+        }
+        if (schema.contains("enum")) {
             std::vector<std::string> enum_values;
             for (const auto & v : schema["enum"]) {
                 enum_values.push_back(_generate_constant_rule(v));
             }
             return _add_rule(rule_name, "(" + string_join(enum_values, " | ") + ") space");
-        } else if ((schema_type.is_null() || schema_type == "object")
+        }
+        if ((schema_type.is_null() || schema_type == "object")
                 && (schema.contains("properties") ||
                     (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
             std::unordered_set<std::string> required;
@@ -863,11 +873,12 @@ public:
                 _build_object_rule(
                     properties, required, name,
                     schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
-        } else if ((schema_type.is_null() || schema_type == "object" || schema_type == "string") && schema.contains("allOf")) {
+        }
+        if ((schema_type.is_null() || schema_type == "object" || schema_type == "string") && schema.contains("allOf")) {
             std::unordered_set<std::string> required;
             std::vector<std::pair<std::string, json>> properties;
             std::map<std::string, size_t> enum_values;
-            std::string hybrid_name = name;
+            const std::string& hybrid_name = name;
             std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
                 if (comp_schema.contains("$ref")) {
                     add_component(_refs[comp_schema["$ref"]], is_required);
@@ -890,9 +901,9 @@ public:
                   // todo warning
                 }
             };
-            for (auto & t : schema["allOf"]) {
+            for (const auto & t : schema["allOf"]) {
                 if (t.contains("anyOf")) {
-                    for (auto & tt : t["anyOf"]) {
+                    for (const auto & tt : t["anyOf"]) {
                         add_component(tt, false);
                     }
                 } else {
@@ -911,7 +922,8 @@ public:
                 }
             }
             return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
-        } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
+        }
+        if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
             json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];
             if (items.is_array()) {
                 std::string rule = "\"[\" space ";
@@ -923,27 +935,31 @@ public:
                 }
                 rule += " \"]\" space";
                 return _add_rule(rule_name, rule);
-            } else {
-                std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
-                int min_items = schema.contains("minItems") ? schema["minItems"].get<int>() : 0;
-                json max_items_json = schema.contains("maxItems") ? schema["maxItems"] : json();
-                int max_items = max_items_json.is_number_integer() ? max_items_json.get<int>() : std::numeric_limits<int>::max();
-
-                return _add_rule(rule_name, "\"[\" space " + build_repetition(item_rule_name, min_items, max_items, "\",\" space") + " \"]\" space");
             }
-        } else if ((schema_type.is_null() || schema_type == "string") && schema.contains("pattern")) {
+            std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
+            int min_items = schema.contains("minItems") ? schema["minItems"].get<int>() : 0;
+            json max_items_json = schema.contains("maxItems") ? schema["maxItems"] : json();
+            int max_items = max_items_json.is_number_integer() ? max_items_json.get<int>() : std::numeric_limits<int>::max();
+
+            return _add_rule(rule_name, "\"[\" space " + build_repetition(item_rule_name, min_items, max_items, "\",\" space") + " \"]\" space");
+        }
+        if ((schema_type.is_null() || schema_type == "string") && schema.contains("pattern")) {
             return _visit_pattern(schema["pattern"], rule_name);
-        } else if ((schema_type.is_null() || schema_type == "string") && std::regex_match(schema_format, std::regex("^uuid[1-5]?$"))) {
+        }
+        if ((schema_type.is_null() || schema_type == "string") && std::regex_match(schema_format, std::regex("^uuid[1-5]?$"))) {
             return _add_primitive(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
-        } else if ((schema_type.is_null() || schema_type == "string") && STRING_FORMAT_RULES.find(schema_format + "-string") != STRING_FORMAT_RULES.end()) {
+        }
+        if ((schema_type.is_null() || schema_type == "string") && STRING_FORMAT_RULES.find(schema_format + "-string") != STRING_FORMAT_RULES.end()) {
             auto prim_name = schema_format + "-string";
             return _add_rule(rule_name, _add_primitive(prim_name, STRING_FORMAT_RULES.at(prim_name)));
-        } else if (schema_type == "string" && (schema.contains("minLength") || schema.contains("maxLength"))) {
+        }
+        if (schema_type == "string" && (schema.contains("minLength") || schema.contains("maxLength"))) {
             std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
             int min_len = schema.contains("minLength") ? schema["minLength"].get<int>() : 0;
             int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
             return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
-        } else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
+        }
+        if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
             int64_t min_value = std::numeric_limits<int64_t>::min();
             int64_t max_value = std::numeric_limits<int64_t>::max();
             if (schema.contains("minimum")) {
@@ -958,19 +974,24 @@ public:
             }
             std::stringstream out;
             out << "(";
-            _build_min_max_int(min_value, max_value, out);
+            build_min_max_int(min_value, max_value, out);
             out << ") space";
             return _add_rule(rule_name, out.str());
-        } else if (schema.empty() || schema_type == "object") {
-            return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
-        } else {
-            if (!schema_type.is_string() || PRIMITIVE_RULES.find(schema_type.get<std::string>()) == PRIMITIVE_RULES.end()) {
-                _errors.push_back("Unrecognized schema: " + schema.dump());
-                return "";
-            }
-            // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
-            return _add_primitive(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
         }
+        if (schema.empty() || schema_type == "object") {
+            return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
+        }
+        if (schema_type.is_null() && schema.is_object()) {
+            // No type constraint and no recognized structural keywords (e.g. {"description": "..."}).
+            // Per JSON Schema semantics this is equivalent to {} and accepts any value.
+            return _add_rule(rule_name, _add_primitive("value", PRIMITIVE_RULES.at("value")));
+        }
+        if (!schema_type.is_string() || PRIMITIVE_RULES.find(schema_type.get<std::string>()) == PRIMITIVE_RULES.end()) {
+            _errors.push_back("Unrecognized schema: " + schema.dump());
+            return "";
+        }
+        // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
+        return _add_primitive(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
     }
 
     void check_errors() {
@@ -985,7 +1006,7 @@ public:
     std::string format_grammar() {
         std::stringstream ss;
         for (const auto & kv : _rules) {
-            ss << kv.first << " ::= " << kv.second << std::endl;
+            ss << kv.first << " ::= " << kv.second << '\n';
         }
         return ss.str();
     }

common/peg-parser.cpp

@@ -1,14 +1,15 @@
-#include "common.h"
 #include "peg-parser.h"
-#include "json-schema-to-grammar.h"
-#include "unicode.h"
 
-#include <nlohmann/json.hpp>
+#include "common.h"
+#include "json-schema-to-grammar.h"
+#include "log.h"
+#include "unicode.h"
 
 #include <algorithm>
 #include <initializer_list>
 #include <map>
 #include <memory>
+#include <nlohmann/json.hpp>
 #include <regex>
 #include <stdexcept>
 #include <unordered_set>
@@ -34,8 +35,7 @@ static bool is_hex_digit(const char c) {
 // This is used in common_peg_until_parser and to build a GBNF exclusion grammar
 struct trie {
     struct node {
-        size_t depth = 0;
-        std::map<unsigned char, size_t> children;
+        std::map<uint32_t, size_t> children;  // Use uint32_t to store Unicode codepoints
         bool is_word;
     };
 
@@ -55,15 +55,22 @@ struct trie {
         size_t current = 0; // Start at root
         size_t pos = start_pos;
 
+        // LOG_DBG("%s: checking at pos %zu, sv='%s'\n", __func__, start_pos, std::string(sv).c_str());
+
         while (pos < sv.size()) {
-            auto it = nodes[current].children.find(sv[pos]);
+            auto result = common_parse_utf8_codepoint(sv, pos);
+            if (result.status != utf8_parse_result::SUCCESS) {
+                break;
+            }
+
+            auto it = nodes[current].children.find(result.codepoint);
             if (it == nodes[current].children.end()) {
                 // Can't continue matching
                 return match_result{match_result::NO_MATCH};
             }
 
             current = it->second;
-            pos++;
+            pos += result.bytes_consumed;
 
             // Check if we've matched a complete word
             if (nodes[current].is_word) {
@@ -82,22 +89,22 @@ struct trie {
     }
 
     struct prefix_and_next {
-        std::string prefix;
-        std::string next_chars;
+        std::vector<uint32_t> prefix;
+        std::vector<uint32_t> next_chars;
     };
 
     std::vector<prefix_and_next> collect_prefix_and_next() {
-        std::string prefix;
+        std::vector<uint32_t>        prefix;
         std::vector<prefix_and_next> result;
         collect_prefix_and_next(0, prefix, result);
         return result;
     }
 
   private:
-    void collect_prefix_and_next(size_t index, std::string & prefix, std::vector<prefix_and_next> & out) {
+    void collect_prefix_and_next(size_t index, std::vector<uint32_t> & prefix, std::vector<prefix_and_next> & out) {
         if (!nodes[index].is_word) {
             if (!nodes[index].children.empty()) {
-                std::string chars;
+                std::vector<uint32_t> chars;
                 chars.reserve(nodes[index].children.size());
                 for (const auto & p : nodes[index].children) {
                     chars.push_back(p.first);
@@ -107,7 +114,7 @@ struct trie {
         }
 
         for (const auto & p : nodes[index].children) {
-            unsigned char ch = p.first;
+            uint32_t ch = p.first;
             auto child = p.second;
             prefix.push_back(ch);
             collect_prefix_and_next(child, prefix, out);
@@ -123,11 +130,19 @@ struct trie {
 
     void insert(const std::string & word) {
         size_t current = 0;
-        for (unsigned char ch : word) {
+        size_t pos     = 0;
+        while (pos < word.length()) {
+            auto result = common_parse_utf8_codepoint(word, pos);
+            if (result.status != utf8_parse_result::SUCCESS) {
+                break;
+            }
+
+            uint32_t ch = result.codepoint;
+            pos += result.bytes_consumed;
+
             auto it = nodes[current].children.find(ch);
             if (it == nodes[current].children.end()) {
                 size_t child = create_node();
-                nodes[child].depth = nodes[current].depth + 1;
                 nodes[current].children[ch] = child;
                 current = child;
             } else {
@@ -286,6 +301,32 @@ struct parser_executor {
     parser_executor(const common_peg_arena & arena, common_peg_parse_context & ctx, size_t start)
         : arena(arena), ctx(ctx), start_pos(start) {}
 
+    std::string debug_indent() const { return std::string(ctx.parse_depth * 2, ' '); }
+
+    std::string debug_input_snippet(size_t pos, size_t len = 60) const {
+        if (pos >= ctx.input.size()) {
+            return "<EOF>";
+        }
+        auto        snippet = ctx.input.substr(pos, len);
+        // Escape newlines for display
+        std::string result;
+        for (char c : snippet) {
+            if (c == '\n') {
+                result += "\\n";
+            } else if (c == '\r') {
+                result += "\\r";
+            } else if (c == '\t') {
+                result += "\\t";
+            } else {
+                result += c;
+            }
+        }
+        if (pos + len < ctx.input.size()) {
+            result += "...";
+        }
+        return result;
+    }
+
     common_peg_parse_result operator()(const common_peg_epsilon_parser & /* p */) const {
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos);
     }
@@ -308,7 +349,7 @@ struct parser_executor {
         auto pos = start_pos;
         for (auto i = 0u; i < p.literal.size(); ++i) {
             if (pos >= ctx.input.size()) {
-                if (!ctx.is_partial) {
+                if (!ctx.is_lenient()) {
                     return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
                 }
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos);
@@ -323,12 +364,32 @@ struct parser_executor {
     }
 
     common_peg_parse_result operator()(const common_peg_sequence_parser & p) {
+        if (ctx.is_debug()) {
+            LOG_DBG("%sSEQ start at %zu '%s' (%zu children)\n", debug_indent().c_str(), start_pos,
+                    debug_input_snippet(start_pos).c_str(), p.children.size());
+        }
+        ctx.parse_depth++;
+
         auto pos = start_pos;
         std::vector<common_peg_ast_id> nodes;
 
-        for (const auto & child_id : p.children) {
+        for (size_t i = 0; i < p.children.size(); i++) {
+            const auto & child_id = p.children[i];
+            if (ctx.is_debug()) {
+                fprintf(stderr, "%sSEQ child %zu: %s\n", debug_indent().c_str(), i, arena.dump(child_id).c_str());
+            }
             auto result = arena.parse(child_id, ctx, pos);
+
+            if (ctx.is_debug()) {
+                fprintf(stderr, "%sSEQ child %zu: %s at %zu->%zu\n", debug_indent().c_str(), i,
+                        common_peg_parse_result_type_name(result.type), result.start, result.end);
+            }
+
             if (result.fail()) {
+                ctx.parse_depth--;
+                if (ctx.is_debug()) {
+                    fprintf(stderr, "%sSEQ -> FAIL\n", debug_indent().c_str());
+                }
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos, result.end);
             }
 
@@ -337,28 +398,65 @@ struct parser_executor {
             }
 
             if (result.need_more_input()) {
+                ctx.parse_depth--;
+                if (ctx.is_debug()) {
+                    fprintf(stderr, "%sSEQ -> NEED_MORE\n", debug_indent().c_str());
+                }
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, result.end, std::move(nodes));
             }
 
             pos = result.end;
         }
 
+        ctx.parse_depth--;
+        if (ctx.is_debug()) {
+            fprintf(stderr, "%sSEQ -> SUCCESS at %zu->%zu\n", debug_indent().c_str(), start_pos, pos);
+        }
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos, std::move(nodes));
     }
 
     common_peg_parse_result operator()(const common_peg_choice_parser & p) {
+        if (ctx.is_debug()) {
+            fprintf(stderr, "%sCHOICE start at %zu '%s' (%zu options)\n", debug_indent().c_str(), start_pos,
+                    debug_input_snippet(start_pos).c_str(), p.children.size());
+        }
+        ctx.parse_depth++;
+
         auto pos = start_pos;
-        for (const auto & child_id : p.children) {
+        for (size_t i = 0; i < p.children.size(); i++) {
+            const auto & child_id = p.children[i];
+            if (ctx.is_debug()) {
+                fprintf(stderr, "%sCHOICE option %zu: %s\n", debug_indent().c_str(), i, arena.dump(child_id).c_str());
+            }
             auto result = arena.parse(child_id, ctx, pos);
+            if (ctx.is_debug()) {
+                fprintf(stderr, "%sCHOICE option %zu: %s\n", debug_indent().c_str(), i,
+                        common_peg_parse_result_type_name(result.type));
+            }
             if (!result.fail()) {
+                ctx.parse_depth--;
+                if (ctx.is_debug()) {
+                    fprintf(stderr, "%sCHOICE -> %s (option %zu)\n", debug_indent().c_str(),
+                            common_peg_parse_result_type_name(result.type), i);
+                }
                 return result;
             }
         }
 
+        ctx.parse_depth--;
+        if (ctx.is_debug()) {
+            fprintf(stderr, "%sCHOICE -> FAIL (no options matched)\n", debug_indent().c_str());
+        }
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
     }
 
     common_peg_parse_result operator()(const common_peg_repetition_parser & p) {
+        if (ctx.is_debug()) {
+            fprintf(stderr, "%sREPEAT start at %zu '%s' (min=%d, max=%d)\n", debug_indent().c_str(), start_pos,
+                    debug_input_snippet(start_pos).c_str(), p.min_count, p.max_count);
+        }
+        ctx.parse_depth++;
+
         auto pos = start_pos;
         int match_count = 0;
         std::vector<common_peg_ast_id> nodes;
@@ -366,14 +464,26 @@ struct parser_executor {
         // Try to match up to max_count times (or unlimited if max_count is -1)
         while (p.max_count == -1 || match_count < p.max_count) {
             if (pos >= ctx.input.size()) {
+                if (ctx.is_debug()) {
+                    fprintf(stderr, "%sREPEAT: at end of input, count=%d\n", debug_indent().c_str(), match_count);
+                }
                 break;
             }
 
             auto result = arena.parse(p.child, ctx, pos);
 
+            if (ctx.is_debug()) {
+                fprintf(stderr, "%sREPEAT iter %d: %s at %zu->%zu, nodes=%zu\n", debug_indent().c_str(), match_count,
+                        common_peg_parse_result_type_name(result.type), result.start, result.end, result.nodes.size());
+                fprintf(stderr, "%sREPEAT CHILD: %s\n", debug_indent().c_str(), arena.dump(p.child).c_str());
+            }
+
             if (result.success()) {
                 // Prevent infinite loop on empty matches
                 if (result.end == pos) {
+                    if (ctx.is_debug()) {
+                        fprintf(stderr, "%s  REPEAT: empty match, stopping\n", debug_indent().c_str());
+                    }
                     break;
                 }
 
@@ -391,21 +501,43 @@ struct parser_executor {
                     nodes.insert(nodes.end(), result.nodes.begin(), result.nodes.end());
                 }
 
+                ctx.parse_depth--;
+                if (ctx.is_debug()) {
+                    fprintf(stderr, "%sREPEAT -> NEED_MORE (count=%d, nodes=%zu)\n", debug_indent().c_str(),
+                            match_count, nodes.size());
+                }
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, result.end, std::move(nodes));
             }
 
             // Child failed - stop trying
+            if (ctx.is_debug()) {
+                fprintf(stderr, "%sREPEAT: child failed, stopping\n", debug_indent().c_str());
+            }
             break;
         }
 
         // Check if we got enough matches
         if (p.min_count > 0 && match_count < p.min_count) {
-            if (pos >= ctx.input.size() && ctx.is_partial) {
+            ctx.parse_depth--;
+            if (pos >= ctx.input.size() && ctx.is_lenient()) {
+                if (ctx.is_debug()) {
+                    fprintf(stderr, "%sREPEAT -> NEED_MORE (not enough matches: %d < %d)\n", debug_indent().c_str(),
+                            match_count, p.min_count);
+                }
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos, std::move(nodes));
             }
+            if (ctx.is_debug()) {
+                fprintf(stderr, "%sREPEAT -> FAIL (not enough matches: %d < %d)\n", debug_indent().c_str(), match_count,
+                        p.min_count);
+            }
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos, pos);
         }
 
+        ctx.parse_depth--;
+        if (ctx.is_debug()) {
+            fprintf(stderr, "%sREPEAT -> SUCCESS (count=%d, nodes=%zu)\n", debug_indent().c_str(), match_count,
+                    nodes.size());
+        }
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos, std::move(nodes));
     }
 
@@ -434,10 +566,10 @@ struct parser_executor {
 
     common_peg_parse_result operator()(const common_peg_any_parser & /* p */) const {
         // Parse a single UTF-8 codepoint (not just a single byte)
-        auto result = parse_utf8_codepoint(ctx.input, start_pos);
+        auto result = common_parse_utf8_codepoint(ctx.input, start_pos);
 
         if (result.status == utf8_parse_result::INCOMPLETE) {
-            if (!ctx.is_partial) {
+            if (!ctx.is_lenient()) {
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
             }
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos);
@@ -468,7 +600,7 @@ struct parser_executor {
 
         // Try to match up to max_count times (or unlimited if max_count is -1)
         while (p.max_count == -1 || match_count < p.max_count) {
-            auto result = parse_utf8_codepoint(ctx.input, pos);
+            auto result = common_parse_utf8_codepoint(ctx.input, pos);
 
             if (result.status == utf8_parse_result::INCOMPLETE) {
                 if (match_count >= p.min_count) {
@@ -476,7 +608,7 @@ struct parser_executor {
                     return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
                 }
                 // Not enough matches yet
-                if (!ctx.is_partial) {
+                if (!ctx.is_lenient()) {
                     return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
                 }
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos);
@@ -517,7 +649,7 @@ struct parser_executor {
 
         // Check if we got enough matches
         if (match_count < p.min_count) {
-            if (pos >= ctx.input.size() && ctx.is_partial) {
+            if (pos >= ctx.input.size() && ctx.is_lenient()) {
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos);
             }
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos, pos);
@@ -526,31 +658,23 @@ struct parser_executor {
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
     }
 
-    static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos) {
+    static common_peg_parse_result handle_escape_sequence(common_peg_parse_context & ctx, size_t start, size_t & pos, const char delimiter) {
         ++pos; // consume '\'
         if (pos >= ctx.input.size()) {
-            if (!ctx.is_partial) {
+            if (!ctx.is_lenient()) {
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
             }
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
         }
 
-        switch (ctx.input[pos]) {
-            case '"':
-            case '\\':
-            case '/':
-            case 'b':
-            case 'f':
-            case 'n':
-            case 'r':
-            case 't':
-                ++pos;
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
-            case 'u':
-                return handle_unicode_escape(ctx, start, pos);
-            default:
-                // Invalid escape sequence
-                return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
+        char c = ctx.input[pos];
+        if (c == delimiter || c == '\\' || c == '/' || c == 'b' || c == 'f' || c == 'n' || c == 'r' || c == 't') {
+            ++pos;
+            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
+        } else if (c == 'u') {
+            return handle_unicode_escape(ctx, start, pos);
+        } else {
+            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
         }
     }
 
@@ -558,7 +682,7 @@ struct parser_executor {
         ++pos; // consume 'u'
         for (int i = 0; i < 4; ++i) {
             if (pos >= ctx.input.size()) {
-                if (!ctx.is_partial) {
+                if (!ctx.is_lenient()) {
                     return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start);
                 }
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start, pos);
@@ -571,28 +695,28 @@ struct parser_executor {
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start, pos);
     }
 
-    common_peg_parse_result operator()(const common_peg_json_string_parser & /* p */) {
+    common_peg_parse_result operator()(const common_peg_string_parser & p) {
         auto pos = start_pos;
 
         // Parse string content (without quotes)
         while (pos < ctx.input.size()) {
             char c = ctx.input[pos];
 
-            if (c == '"') {
-                // Found closing quote - success (don't consume it)
+            if (c == p.delimiter) {
+                // Found closing delimiter - success (don't consume it)
                 return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_SUCCESS, start_pos, pos);
             }
 
             if (c == '\\') {
-                auto result = handle_escape_sequence(ctx, start_pos, pos);
+                auto result = handle_escape_sequence(ctx, start_pos, pos, p.delimiter);
                 if (!result.success()) {
                     return result;
                 }
             } else {
-                auto utf8_result = parse_utf8_codepoint(ctx.input, pos);
+                auto utf8_result = common_parse_utf8_codepoint(ctx.input, pos);
 
                 if (utf8_result.status == utf8_parse_result::INCOMPLETE) {
-                    if (!ctx.is_partial) {
+                    if (!ctx.is_lenient()) {
                         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
                     }
                     return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos);
@@ -607,7 +731,7 @@ struct parser_executor {
         }
 
         // Reached end without finding closing quote
-        if (!ctx.is_partial) {
+        if (!ctx.is_lenient()) {
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos, pos);
         }
         return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, pos);
@@ -621,11 +745,11 @@ struct parser_executor {
         size_t last_valid_pos = start_pos;
 
         while (pos < ctx.input.size()) {
-            auto utf8_result = parse_utf8_codepoint(ctx.input, pos);
+            auto utf8_result = common_parse_utf8_codepoint(ctx.input, pos);
 
             if (utf8_result.status == utf8_parse_result::INCOMPLETE) {
                 // Incomplete UTF-8 sequence
-                if (!ctx.is_partial) {
+                if (!ctx.is_lenient()) {
                     // Input is complete but UTF-8 is incomplete = malformed
                     return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_FAIL, start_pos);
                 }
@@ -655,7 +779,7 @@ struct parser_executor {
             last_valid_pos = pos;
         }
 
-        if (last_valid_pos == ctx.input.size() && ctx.is_partial) {
+        if (last_valid_pos == ctx.input.size() && ctx.is_lenient()) {
             // Reached the end of a partial stream, there might still be more input that we need to consume.
             return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos, last_valid_pos);
         }
@@ -694,6 +818,9 @@ struct parser_executor {
 
     common_peg_parse_result operator()(const common_peg_tag_parser & p) {
         // Parse the child
+        if (ctx.is_debug()) {
+            fprintf(stderr, "%sTAG: %s\n", debug_indent().c_str(), p.tag.c_str());
+        }
         auto result = arena.parse(p.child, ctx, start_pos);
 
         if (!result.fail()) {
@@ -755,6 +882,31 @@ common_peg_parser_id common_peg_arena::resolve_ref(common_peg_parser_id id) {
     return id;
 }
 
+static void bfs_node(common_peg_ast_arena &arena, std::ostringstream & oss, const common_peg_ast_node & node, int indent) {
+    for (int i = 0; i < indent; i++) {
+        oss << "  ";
+    }
+    oss << "NODE " << node.id;
+    if (!node.rule.empty()) {
+        oss << " (rule " << node.rule << ")";
+    }
+    if (!node.tag.empty()) {
+        oss << " (tag " << node.tag << ")";
+    }
+    oss << " ['" << node.text << "']\n";
+    for (const auto child : node.children) {
+        bfs_node(arena, oss, arena.get(child), indent + 1);
+    }
+}
+
+std::string common_peg_ast_arena::dump() {
+    std::ostringstream oss;
+    for (auto & node : nodes_) {
+        bfs_node(*this, oss, node, 0);
+    }
+    return oss.str();
+}
+
 void common_peg_arena::resolve_refs() {
     // Walk through all parsers and replace refs with their corresponding rule IDs
     for (auto & parser : parsers_) {
@@ -785,7 +937,7 @@ void common_peg_arena::resolve_refs() {
                                  std::is_same_v<T, common_peg_ref_parser> ||
                                  std::is_same_v<T, common_peg_until_parser> ||
                                  std::is_same_v<T, common_peg_literal_parser> ||
-                                 std::is_same_v<T, common_peg_json_string_parser> ||
+                                 std::is_same_v<T, common_peg_string_parser> ||
                                  std::is_same_v<T, common_peg_chars_parser> ||
                                  std::is_same_v<T, common_peg_any_parser> ||
                                  std::is_same_v<T, common_peg_space_parser>) {
@@ -803,9 +955,21 @@ void common_peg_arena::resolve_refs() {
 }
 
 std::string common_peg_arena::dump(common_peg_parser_id id) const {
+    std::unordered_set<common_peg_parser_id> visited;
+    return dump_impl(id, visited);
+}
+
+std::string common_peg_arena::dump_impl(common_peg_parser_id                       id,
+                                        std::unordered_set<common_peg_parser_id> & visited) const {
+    // Check for cycles
+    if (visited.count(id)) {
+        return "[cycle]";
+    }
+    visited.insert(id);
+
     const auto & parser = parsers_.at(id);
 
-    return std::visit([this](const auto & p) -> std::string {
+    return std::visit([this, &visited](const auto & p) -> std::string {
         using T = std::decay_t<decltype(p)>;
 
         if constexpr (std::is_same_v<T, common_peg_epsilon_parser>) {
@@ -819,24 +983,27 @@ std::string common_peg_arena::dump(common_peg_parser_id id) const {
         } else if constexpr (std::is_same_v<T, common_peg_sequence_parser>) {
             std::vector<std::string> parts;
             for (const auto & child : p.children) {
-                parts.push_back(dump(child));
+                parts.push_back(dump_impl(child, visited));
             }
             return "Sequence(" + string_join(parts, ", ") + ")";
         } else if constexpr (std::is_same_v<T, common_peg_choice_parser>) {
             std::vector<std::string> parts;
             for (const auto & child : p.children) {
-                parts.push_back(dump(child));
+                parts.push_back(dump_impl(child, visited));
             }
             return "Choice(" + string_join(parts, ", ") + ")";
         } else if constexpr (std::is_same_v<T, common_peg_repetition_parser>) {
             if (p.max_count == -1) {
-                return "Repetition(" + dump(p.child) + ", " + std::to_string(p.min_count) + ", unbounded)";
+                return "Repetition(" + dump_impl(p.child, visited) + ", " + std::to_string(p.min_count) +
+                        ", unbounded)";
             }
-            return "Repetition(" + dump(p.child) + ", " + std::to_string(p.min_count) + ", " + std::to_string(p.max_count) + ")";
+            return "Repetition(" + dump_impl(p.child, visited) + ", " + std::to_string(p.min_count) + ", " + std::to_string(p.max_count) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_and_parser>) {
-            return "And(" + dump(p.child) + ")";
+            return "And(" + dump_impl(p.child, visited) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_not_parser>) {
-            return "Not(" + dump(p.child) + ")";
+            return "Not(" + dump_impl(p.child, visited) + ")";
+        } else if constexpr (std::is_same_v<T, common_peg_atomic_parser>) {
+            return "Atomic(" + dump_impl(p.child, visited) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_any_parser>) {
             return "Any";
         } else if constexpr (std::is_same_v<T, common_peg_space_parser>) {
@@ -846,16 +1013,20 @@ std::string common_peg_arena::dump(common_peg_parser_id id) const {
                 return "CharRepeat(" + p.pattern + ", " + std::to_string(p.min_count) + ", unbounded)";
             }
             return "CharRepeat(" + p.pattern + ", " + std::to_string(p.min_count) + ", " + std::to_string(p.max_count) + ")";
-        } else if constexpr (std::is_same_v<T, common_peg_json_string_parser>) {
-            return "JsonString()";
+        } else if constexpr (std::is_same_v<T, common_peg_string_parser>) {
+            return "String(" + std::string(1, p.delimiter) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_until_parser>) {
             return "Until(" + string_join(p.delimiters, " | ") + ")";
         } else if constexpr (std::is_same_v<T, common_peg_schema_parser>) {
-            return "Schema(" + dump(p.child) + ", " + (p.schema ? p.schema->dump() : "null") + ")";
+            return "Schema(" + dump_impl(p.child, visited) + ", " + (p.schema ? p.schema->dump() : "null") + ")";
         } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
-            return "Rule(" + p.name + ", " + dump(p.child) + ")";
+            return "Rule(" + p.name + ", " + dump_impl(p.child, visited) + ")";
         } else if constexpr (std::is_same_v<T, common_peg_ref_parser>) {
             return "Ref(" + p.name + ")";
+        } else if constexpr (std::is_same_v<T, common_peg_tag_parser>) {
+            return "Tag(" + p.tag + ", " + dump(p.child) + ")";
+        } else if constexpr (std::is_same_v<T, common_peg_atomic_parser>) {
+            return "Atomic(" + dump(p.child) + ")";
         } else {
             return "Unknown";
         }
@@ -1054,7 +1225,32 @@ common_peg_arena common_peg_parser_builder::build() {
     return std::move(arena_);
 }
 
+// String primitives
+
+common_peg_parser common_peg_parser_builder::string_content(char delimiter) {
+    return wrap(arena_.add_parser(common_peg_string_parser{delimiter}));
+}
+
+common_peg_parser common_peg_parser_builder::double_quoted_string() {
+    return rule("double-quoted-string", [this]() {
+        return sequence({literal("\""), string_content('"'), literal("\""), space()});
+    });
+}
+
+common_peg_parser common_peg_parser_builder::single_quoted_string() {
+    return rule("single-quoted-string", [this]() {
+        return sequence({literal("'"), string_content('\''), literal("'"), space()});
+    });
+}
+
+common_peg_parser common_peg_parser_builder::quoted_string() {
+    return rule("quoted-string", [this]() {
+        return choice({double_quoted_string(), single_quoted_string()});
+    });
+}
+
 // JSON parsers
+
 common_peg_parser common_peg_parser_builder::json_number() {
    return rule("json-number", [this]() {
         auto digit1_9 = chars("[1-9]", 1, 1);
@@ -1062,13 +1258,17 @@ common_peg_parser common_peg_parser_builder::json_number() {
         auto int_part = choice({literal("0"), sequence({digit1_9, chars("[0-9]", 0, -1)})});
         auto frac = sequence({literal("."), digits});
         auto exp = sequence({choice({literal("e"), literal("E")}), optional(chars("[+-]", 1, 1)), digits});
-        return sequence({optional(literal("-")), int_part, optional(frac), optional(exp), space()});
+        // Negative lookahead: only commit the number when the next character can't extend it.
+        // At EOF in partial mode, chars returns NEED_MORE → negate propagates NEED_MORE → number not committed.
+        // This prevents premature commits of partial numbers (e.g. "3" when "3.14" is incoming).
+        auto not_number_continuation = negate(chars("[0-9.eE+-]", 1, 1));
+        return sequence({ optional(literal("-")), int_part, optional(frac), optional(exp), not_number_continuation, space() });
     });
 }
 
 common_peg_parser common_peg_parser_builder::json_string() {
     return rule("json-string", [this]() {
-        return sequence({literal("\""), json_string_content(), literal("\""), space()});
+        return sequence({literal("\""), string_content('"'), literal("\""), space()});
     });
 }
 
@@ -1130,8 +1330,81 @@ common_peg_parser common_peg_parser_builder::json() {
     });
 }
 
-common_peg_parser common_peg_parser_builder::json_string_content() {
-    return wrap(arena_.add_parser(common_peg_json_string_parser{}));
+common_peg_parser common_peg_parser_builder::python_string() {
+    return rule("python-string", [this]() {
+        return choice({double_quoted_string(), single_quoted_string()});
+    });
+}
+
+common_peg_parser common_peg_parser_builder::python_number() {
+    return json_number();
+}
+
+common_peg_parser common_peg_parser_builder::python_bool() {
+    return rule("python-bool", [this]() {
+        return sequence({
+            choice({literal("True"), literal("False")}),
+            space()
+        });
+    });
+}
+
+common_peg_parser common_peg_parser_builder::python_null() {
+    return rule("python-none", [this]() {
+        return sequence({literal("None"), space()});
+    });
+}
+
+common_peg_parser common_peg_parser_builder::python_dict() {
+    return rule("python-dict", [this]() {
+        auto ws = space();
+        auto member = sequence({python_string(), ws, literal(":"), ws, python_value()});
+        auto members = sequence({member, zero_or_more(sequence({ws, literal(","), ws, member}))});
+        return sequence({
+            literal("{"),
+            ws,
+            choice({
+                literal("}"),
+                sequence({members, ws, literal("}")})
+            }),
+            ws
+        });
+    });
+}
+
+common_peg_parser common_peg_parser_builder::python_array() {
+    return rule("python-array", [this]() {
+        auto ws = space();
+        auto elements = sequence({python_value(), zero_or_more(sequence({literal(","), ws, python_value()}))});
+        return sequence({
+            literal("["),
+            ws,
+            choice({
+                literal("]"),
+                sequence({elements, ws, literal("]")})
+            }),
+            ws
+        });
+    });
+}
+
+common_peg_parser common_peg_parser_builder::python_value() {
+    return rule("python-value", [this]() {
+        return choice({
+            python_dict(),
+            python_array(),
+            python_string(),
+            python_number(),
+            python_bool(),
+            python_null()
+        });
+    });
+}
+
+common_peg_parser common_peg_parser_builder::marker() {
+    auto sharp_bracket_parser = literal("<") + until(">") + literal(">");
+    auto square_bracket_parser = literal("[") + until("]") + literal("]");
+    return choice({ sharp_bracket_parser, square_bracket_parser });
 }
 
 common_peg_parser common_peg_parser_builder::json_member(const std::string & key, const common_peg_parser & p) {
@@ -1145,17 +1418,54 @@ common_peg_parser common_peg_parser_builder::json_member(const std::string & key
     });
 }
 
-
-static std::string gbnf_escape_char_class(char c) {
-    switch (c) {
-        case '\n': return "\\n";
-        case '\t': return "\\t";
-        case '\r': return "\\r";
-        case '\\': return "\\\\";
-        case ']':  return "\\]";
-        case '[':  return "\\[";
-        default:   return std::string(1, c);
+static std::string gbnf_escape_char_class(uint32_t c) {
+    if (c == '-' || c == ']' || c == '[' || c == '\\') {
+        return "\\" + std::string(1, (char) c);
     }
+    // Escape whitespace control characters
+    if (c == '\n') {
+        return "\\n";
+    }
+    if (c == '\t') {
+        return "\\t";
+    }
+    if (c == '\r') {
+        return "\\r";
+    }
+
+    // Printable ASCII
+    if (c >= 0x20 && c <= 0x7E) {
+        return std::string(1, (char) c);
+    }
+
+    // Hex escape
+    char         buf[16];
+    const char * hex = "0123456789ABCDEF";
+
+    if (c <= 0xFF) {
+        buf[0] = '\\';
+        buf[1] = 'x';
+        buf[2] = hex[(c >> 4) & 0xF];
+        buf[3] = hex[c & 0xF];
+        buf[4] = '\0';
+    } else if (c <= 0xFFFF) {
+        buf[0] = '\\';
+        buf[1] = 'u';
+        buf[2] = hex[(c >> 12) & 0xF];
+        buf[3] = hex[(c >> 8) & 0xF];
+        buf[4] = hex[(c >> 4) & 0xF];
+        buf[5] = hex[c & 0xF];
+        buf[6] = '\0';
+    } else {
+        buf[0] = '\\';
+        buf[1] = 'U';
+        for (int i = 0; i < 8; i++) {
+            buf[2 + i] = hex[(c >> ((7 - i) * 4)) & 0xF];
+        }
+        buf[10] = '\0';
+    }
+
+    return std::string(buf);
 }
 
 static std::string gbnf_excluding_pattern(const std::vector<std::string> & strings) {
@@ -1173,12 +1483,12 @@ static std::string gbnf_excluding_pattern(const std::vector<std::string> & strin
 
         std::string cls;
         cls.reserve(chars.size());
-        for (const auto & ch : chars) {
+        for (uint32_t ch : chars) {
             cls += gbnf_escape_char_class(ch);
         }
 
         if (!pre.empty()) {
-            pattern += gbnf_format_literal(pre) + " [^" + cls + "]";
+            pattern += gbnf_format_literal(common_unicode_cpts_to_utf8(pre)) + " [^" + cls + "]";
         } else {
             pattern += "[^" + cls + "]";
         }
@@ -1208,7 +1518,7 @@ static std::unordered_set<std::string> collect_reachable_rules(
                           std::is_same_v<T, common_peg_chars_parser> ||
                           std::is_same_v<T, common_peg_space_parser> ||
                           std::is_same_v<T, common_peg_any_parser> ||
-                          std::is_same_v<T, common_peg_json_string_parser>) {
+                          std::is_same_v<T, common_peg_string_parser>) {
                 // These parsers do not have any children
             } else if constexpr (std::is_same_v<T, common_peg_sequence_parser>) {
                 for (auto child : p.children) {
@@ -1344,8 +1654,9 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
                     return result + "{" + std::to_string(p.min_count) + "}";
                 }
                 return result + "{" + std::to_string(p.min_count) + "," + std::to_string(p.max_count) + "}";
-            } else if constexpr (std::is_same_v<T, common_peg_json_string_parser>) {
-                return R"(( [^"\\] | "\\" ( ["\\/ bfnrt] | "u" [0-9a-fA-F]{4} ) )*)";
+            } else if constexpr (std::is_same_v<T, common_peg_string_parser>) {
+                const std::string delim(1, p.delimiter);
+                return R"(( [^)" + delim + R"(\\] | "\\" ( [)" + delim + R"(\\/ bfnrt] | "u" [0-9a-fA-F]{4} ) )*)";
             } else if constexpr (std::is_same_v<T, common_peg_until_parser>) {
                 if (p.delimiters.empty()) {
                     return ".*";
@@ -1475,8 +1786,8 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
                 {"min_count", p.min_count},
                 {"max_count", p.max_count}
             };
-        } else if constexpr (std::is_same_v<T, common_peg_json_string_parser>) {
-            return json{{"type", "json_string"}};
+        } else if constexpr (std::is_same_v<T, common_peg_string_parser>) {
+            return json{{"type", "string"}, {"delimiter", std::string(1, p.delimiter)}};
         } else if constexpr (std::is_same_v<T, common_peg_until_parser>) {
             return json{{"type", "until"}, {"delimiters", p.delimiters}};
         } else if constexpr (std::is_same_v<T, common_peg_schema_parser>) {
@@ -1603,8 +1914,15 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         }
         return parser;
     }
-    if (type == "json_string") {
-        return common_peg_json_string_parser{};
+    if (type == "string") {
+        if (!j.contains("delimiter")) {
+            throw std::runtime_error("string parser missing delimiter field.");
+        }
+        std::string delimiter = j["delimiter"];
+        if (delimiter.empty()) {
+            throw std::runtime_error("string parser delimiter is empty.");
+        }
+        return common_peg_string_parser{delimiter[0]};
     }
     if (type == "until") {
         if (!j.contains("delimiters") || !j["delimiters"].is_array()) {

common/peg-parser.h

@@ -4,6 +4,7 @@
 
 #include <memory>
 #include <unordered_map>
+#include <unordered_set>
 #include <string>
 #include <string_view>
 #include <functional>
@@ -111,6 +112,8 @@ class common_peg_ast_arena {
 
     void visit(common_peg_ast_id id, const common_peg_ast_visitor & visitor) const;
     void visit(const common_peg_parse_result & result, const common_peg_ast_visitor & visitor) const;
+
+    std::string dump();
 };
 
 struct common_peg_parse_result {
@@ -136,21 +139,43 @@ struct common_peg_parse_result {
     bool success() const { return type == COMMON_PEG_PARSE_RESULT_SUCCESS; }
 };
 
+enum common_peg_parse_flags {
+    COMMON_PEG_PARSE_FLAG_NONE    = 0,
+    COMMON_PEG_PARSE_FLAG_LENIENT = 1 << 0,
+    COMMON_PEG_PARSE_FLAG_DEBUG   = 1 << 1,
+};
+
+inline common_peg_parse_flags operator|(common_peg_parse_flags a, common_peg_parse_flags b) {
+    return static_cast<common_peg_parse_flags>(int(a) | int(b));
+}
+
+inline common_peg_parse_flags & operator|=(common_peg_parse_flags & a, common_peg_parse_flags b) {
+    return a = a | b;
+}
+
+inline common_peg_parse_flags operator&(common_peg_parse_flags a, common_peg_parse_flags b) {
+    return static_cast<common_peg_parse_flags>(int(a) & int(b));
+}
+
+inline common_peg_parse_flags operator~(common_peg_parse_flags a) {
+    return static_cast<common_peg_parse_flags>(~int(a));
+}
+
 struct common_peg_parse_context {
     std::string input;
-    bool is_partial;
+    common_peg_parse_flags flags;
     common_peg_ast_arena ast;
 
     int parse_depth;
 
-    common_peg_parse_context()
-        : is_partial(false), parse_depth(0) {}
+    common_peg_parse_context(common_peg_parse_flags flags = COMMON_PEG_PARSE_FLAG_NONE)
+        : flags(flags), parse_depth(0) {}
 
-    common_peg_parse_context(const std::string & input)
-        : input(input), is_partial(false), parse_depth(0) {}
+    common_peg_parse_context(const std::string & input, common_peg_parse_flags flags = COMMON_PEG_PARSE_FLAG_NONE)
+        : input(input), flags(flags), parse_depth(0) {}
 
-    common_peg_parse_context(const std::string & input, bool is_partial)
-        : input(input), is_partial(is_partial), parse_depth(0) {}
+    bool is_lenient() const { return flags & COMMON_PEG_PARSE_FLAG_LENIENT; }
+    bool is_debug() const { return flags & COMMON_PEG_PARSE_FLAG_DEBUG; }
 };
 
 class common_peg_arena;
@@ -206,7 +231,9 @@ struct common_peg_chars_parser {
     int max_count;  // -1 for unbounded
 };
 
-struct common_peg_json_string_parser {};
+struct common_peg_string_parser {
+    char delimiter;
+};
 
 struct common_peg_until_parser {
     std::vector<std::string> delimiters;
@@ -254,7 +281,7 @@ using common_peg_parser_variant = std::variant<
     common_peg_any_parser,
     common_peg_space_parser,
     common_peg_chars_parser,
-    common_peg_json_string_parser,
+    common_peg_string_parser,
     common_peg_until_parser,
     common_peg_schema_parser,
     common_peg_rule_parser,
@@ -299,6 +326,8 @@ class common_peg_arena {
     friend class common_peg_parser_builder;
 
   private:
+    std::string dump_impl(common_peg_parser_id id, std::unordered_set<common_peg_parser_id> & visited) const;
+
     common_peg_parser_id add_parser(common_peg_parser_variant parser);
     void add_rule(const std::string & name, common_peg_parser_id id);
 
@@ -404,6 +433,18 @@ class common_peg_parser_builder {
     //   S -> A{n}
     common_peg_parser repeat(const common_peg_parser & p, int n) { return repeat(p, n, n); }
 
+    // Matches a double-quoted string: '"' content '"' space
+    common_peg_parser double_quoted_string();
+
+    // Matches a single-quoted string: "'" content "'" space
+    common_peg_parser single_quoted_string();
+
+    // Matches a string that accepts both double-quoted and single-quoted styles.
+    common_peg_parser quoted_string();
+
+    // Matches string content without the surrounding delimiter.
+    common_peg_parser string_content(char delimiter);
+
     // Creates a complete JSON parser supporting objects, arrays, strings, numbers, booleans, and null.
     //   value -> object | array | string | number | true | false | null
     common_peg_parser json();
@@ -414,14 +455,24 @@ class common_peg_parser_builder {
     common_peg_parser json_bool();
     common_peg_parser json_null();
 
-    // Matches JSON string content without the surrounding quotes.
-    // Useful for extracting content within a JSON string.
-    common_peg_parser json_string_content();
-
     // Matches a JSON object member with a key and associated parser as the
     // value.
     common_peg_parser json_member(const std::string & key, const common_peg_parser & p);
 
+    // Creates a complete Python format parser supporting dicts, arrays, strings, numbers, booleans, and None.
+    // Differs from JSON: uses True/False/None, accepts both single and double-quoted strings.
+    //   value -> dict | array | string | number | True | False | None
+    common_peg_parser python_value();
+    common_peg_parser python_dict();
+    common_peg_parser python_string();
+    common_peg_parser python_array();
+    common_peg_parser python_number();
+    common_peg_parser python_bool();
+    common_peg_parser python_null();
+
+    // A marker, i.e. text delimited by a pair of <> or []
+    common_peg_parser marker();
+
     // Wraps a parser with JSON schema metadata for grammar generation.
     // Used internally to convert JSON schemas to GBNF grammar rules.
     common_peg_parser schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw = false);

common/reasoning-budget.cpp

@@ -0,0 +1,219 @@
+#include "reasoning-budget.h"
+#include "common.h"
+#include "unicode.h"
+
+#include "log.h"
+
+#include <cmath>
+#include <cstdint>
+#include <string>
+#include <vector>
+
+struct token_matcher {
+    std::vector<llama_token> tokens;
+    size_t pos = 0;
+
+    bool advance(llama_token token) {
+        if (tokens.empty()) {
+            return false;
+        }
+
+        if (token == tokens[pos]) {
+            pos++;
+            if (pos >= tokens.size()) {
+                pos = 0;
+                return true;
+            }
+        } else {
+            pos = 0;
+            if (token == tokens[0]) {
+                pos = 1;
+            }
+        }
+        return false;
+    }
+
+    void reset() { pos = 0; }
+};
+
+struct common_reasoning_budget_ctx {
+    const llama_vocab * vocab;
+
+    token_matcher start_matcher;
+    token_matcher end_matcher;
+    std::vector<llama_token> forced_tokens;
+
+    int32_t budget;           // maximum tokens in reasoning block
+    int32_t remaining;        // tokens remaining in budget
+
+    common_reasoning_budget_state state;
+
+    // for forcing
+    size_t force_pos;         // next position in forced_tokens to force
+};
+
+static const char * common_reasoning_budget_name(const struct llama_sampler * /*smpl*/) {
+    return "reasoning-budget";
+}
+
+static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_token token) {
+    auto * ctx = (common_reasoning_budget_ctx *) smpl->ctx;
+
+    switch (ctx->state) {
+        case REASONING_BUDGET_IDLE:
+        {
+            if (ctx->start_matcher.advance(token)) {
+                ctx->state = REASONING_BUDGET_COUNTING;
+                ctx->remaining = ctx->budget;
+                LOG_INF("reasoning-budget: activated, budget=%d tokens\n", ctx->budget);
+
+                if (ctx->remaining <= 0) {
+                    ctx->state = REASONING_BUDGET_FORCING;
+                    ctx->force_pos = 0;
+                    LOG_INF("reasoning-budget: budget=0, forcing immediately\n");
+                }
+            }
+            break;
+        }
+        case REASONING_BUDGET_COUNTING:
+        case REASONING_BUDGET_WAITING_UTF8:
+        {
+            if (ctx->end_matcher.advance(token)) {
+                ctx->state = REASONING_BUDGET_DONE;
+                LOG_INF("reasoning-budget: deactivated (natural end)\n");
+                break;
+            }
+
+            bool utf8_complete = true;
+            if (ctx->vocab != nullptr) {
+                const std::string piece = common_token_to_piece(ctx->vocab, token, false);
+                utf8_complete = common_utf8_is_complete(piece);
+            }
+
+            if (ctx->state == REASONING_BUDGET_WAITING_UTF8) {
+                if (utf8_complete) {
+                    ctx->state = REASONING_BUDGET_FORCING;
+                    ctx->force_pos = 0;
+                    ctx->end_matcher.reset();
+                    LOG_INF("reasoning-budget: UTF-8 complete, now forcing end sequence\n");
+                }
+            } else if (ctx->state == REASONING_BUDGET_COUNTING) {
+                ctx->remaining--;
+                if (ctx->remaining <= 0) {
+                    if (utf8_complete) {
+                        ctx->state = REASONING_BUDGET_FORCING;
+                        ctx->force_pos = 0;
+                        ctx->end_matcher.reset();
+                        LOG_INF("reasoning-budget: budget exhausted, forcing end sequence\n");
+                    } else {
+                        ctx->state = REASONING_BUDGET_WAITING_UTF8;
+                        ctx->end_matcher.reset();
+                        LOG_INF("reasoning-budget: budget exhausted, waiting for UTF-8 completion\n");
+                    }
+                }
+            }
+            break;
+        }
+        case REASONING_BUDGET_FORCING:
+            // force_pos is advanced in apply(), not here.
+            // This ensures the first forced token isn't skipped when the sampler
+            // is initialized directly in FORCING state (e.g. COUNTING + budget=0)
+            break;
+        case REASONING_BUDGET_DONE:
+            break;
+    }
+}
+
+static void common_reasoning_budget_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (common_reasoning_budget_ctx *) smpl->ctx;
+
+    if (ctx->state != REASONING_BUDGET_FORCING) {
+        // passthrough — don't modify logits
+        return;
+    }
+
+    if (ctx->force_pos >= ctx->forced_tokens.size()) {
+        return;
+    }
+
+    const llama_token forced = ctx->forced_tokens[ctx->force_pos];
+
+    // set all logits to -inf except the forced token
+    for (size_t i = 0; i < cur_p->size; i++) {
+        if (cur_p->data[i].id != forced) {
+            cur_p->data[i].logit = -INFINITY;
+        }
+    }
+
+    // advance to next forced token (done here rather than in accept so that
+    // the first forced token isn't skipped when starting in FORCING state)
+    ctx->force_pos++;
+    if (ctx->force_pos >= ctx->forced_tokens.size()) {
+        ctx->state = REASONING_BUDGET_DONE;
+        LOG_INF("reasoning-budget: forced sequence complete, done\n");
+    }
+}
+
+static void common_reasoning_budget_reset(struct llama_sampler * smpl) {
+    auto * ctx = (common_reasoning_budget_ctx *) smpl->ctx;
+    ctx->state = REASONING_BUDGET_IDLE;
+    ctx->remaining = ctx->budget;
+    ctx->start_matcher.reset();
+    ctx->end_matcher.reset();
+    ctx->force_pos = 0;
+}
+
+static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const common_reasoning_budget_ctx *) smpl->ctx;
+    return common_reasoning_budget_init(
+        ctx->vocab,
+        ctx->start_matcher.tokens,
+        ctx->end_matcher.tokens,
+        ctx->forced_tokens,
+        ctx->budget,
+        ctx->state);
+}
+
+static void common_reasoning_budget_free(struct llama_sampler * smpl) {
+    delete (common_reasoning_budget_ctx *) smpl->ctx;
+}
+
+static struct llama_sampler_i common_reasoning_budget_i = {
+    /* .name              = */ common_reasoning_budget_name,
+    /* .accept            = */ common_reasoning_budget_accept,
+    /* .apply             = */ common_reasoning_budget_apply,
+    /* .reset             = */ common_reasoning_budget_reset,
+    /* .clone             = */ common_reasoning_budget_clone,
+    /* .free              = */ common_reasoning_budget_free,
+    /* .backend_init      = */ nullptr,
+    /* .backend_accept    = */ nullptr,
+    /* .backend_apply     = */ nullptr,
+    /* .backend_set_input = */ nullptr,
+};
+
+struct llama_sampler * common_reasoning_budget_init(
+        const struct llama_vocab       * vocab,
+        const std::vector<llama_token> & start_tokens,
+        const std::vector<llama_token> & end_tokens,
+        const std::vector<llama_token> & forced_tokens,
+        int32_t                          budget,
+        common_reasoning_budget_state    initial_state) {
+    // promote COUNTING with budget <= 0 to FORCING
+    if (initial_state == REASONING_BUDGET_COUNTING && budget <= 0) {
+        initial_state = REASONING_BUDGET_FORCING;
+    }
+
+    return llama_sampler_init(
+        /* .iface = */ &common_reasoning_budget_i,
+        /* .ctx   = */ new common_reasoning_budget_ctx {
+            /* .vocab         = */ vocab,
+            /* .start_matcher = */ { start_tokens, 0 },
+            /* .end_matcher   = */ { end_tokens, 0 },
+            /* .forced_tokens = */ forced_tokens,
+            /* .budget        = */ budget,
+            /* .remaining     = */ budget,
+            /* .state         = */ initial_state,
+            /* .force_pos     = */ 0,
+        }
+    );
+}

common/reasoning-budget.h

@@ -0,0 +1,41 @@
+#pragma once
+
+#include "llama.h"
+
+#include <cstdint>
+#include <vector>
+
+enum common_reasoning_budget_state {
+    REASONING_BUDGET_IDLE,         // waiting for start sequence
+    REASONING_BUDGET_COUNTING,     // counting down tokens
+    REASONING_BUDGET_FORCING,      // forcing budget message + end sequence
+    REASONING_BUDGET_WAITING_UTF8, // budget exhausted, waiting for UTF-8 completion
+    REASONING_BUDGET_DONE,         // passthrough forever
+};
+
+// Creates a reasoning budget sampler that limits token generation inside a
+// reasoning block (e.g. between <think> and </think>).
+//
+// State machine: IDLE -> COUNTING -> WAITING_UTF8 -> FORCING -> DONE
+//   IDLE:         passthrough, watching for start_tokens sequence
+//   COUNTING:     counting down remaining tokens, watching for natural end_tokens
+//   WAITING_UTF8: budget exhausted, allowing tokens to complete a UTF-8 sequence
+//   FORCING:      forces forced_tokens token-by-token (all other logits -> -inf)
+//   DONE:         passthrough forever
+//
+// Parameters:
+//   vocab         - vocabulary (used for UTF-8 boundary detection; can be nullptr)
+//   start_tokens  - token sequence that activates counting
+//   end_tokens    - token sequence for natural deactivation
+//   forced_tokens - token sequence forced when budget expires
+//   budget        - max tokens allowed in the reasoning block
+//   initial_state - initial state of the sampler (e.g. IDLE or COUNTING)
+//                   note: COUNTING with budget <= 0 is promoted to FORCING
+//
+struct llama_sampler * common_reasoning_budget_init(
+        const struct llama_vocab       * vocab,
+        const std::vector<llama_token> & start_tokens,
+        const std::vector<llama_token> & end_tokens,
+        const std::vector<llama_token> & forced_tokens,
+        int32_t                          budget,
+        common_reasoning_budget_state    initial_state);

common/sampling.cpp

@@ -2,6 +2,7 @@
 
 #include "common.h"
 #include "log.h"
+#include "reasoning-budget.h"
 
 #include <algorithm>
 #include <cmath>
@@ -250,6 +251,17 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
         }
     }
 
+    // reasoning budget sampler — added first so it can force tokens before other samplers
+    if (params.reasoning_budget_tokens >= 0 && !params.reasoning_budget_forced.empty()) {
+        samplers.push_back(common_reasoning_budget_init(
+            vocab,
+            params.reasoning_budget_start,
+            params.reasoning_budget_end,
+            params.reasoning_budget_forced,
+            params.reasoning_budget_tokens,
+            params.reasoning_budget_activate_immediately ? REASONING_BUDGET_COUNTING : REASONING_BUDGET_IDLE));
+    }
+
     if (params.has_logit_bias()) {
         samplers.push_back(llama_sampler_init_logit_bias(llama_vocab_n_tokens(vocab), params.logit_bias.size(), params.logit_bias.data()));
     }

common/unicode.cpp

@@ -1,14 +1,20 @@
 #include "unicode.h"
 
+#include <algorithm>
+#include <cassert>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
 // implementation adopted from src/unicode.cpp
 
-size_t utf8_sequence_length(unsigned char first_byte) {
+size_t common_utf8_sequence_length(unsigned char first_byte) {
     const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
     uint8_t highbits = static_cast<uint8_t>(first_byte) >> 4;
     return lookup[highbits];
 }
 
-utf8_parse_result parse_utf8_codepoint(std::string_view input, size_t offset) {
+utf8_parse_result common_parse_utf8_codepoint(std::string_view input, size_t offset) {
     if (offset >= input.size()) {
         return utf8_parse_result(utf8_parse_result::INCOMPLETE);
     }
@@ -62,3 +68,57 @@ utf8_parse_result parse_utf8_codepoint(std::string_view input, size_t offset) {
     // Invalid first byte
     return utf8_parse_result(utf8_parse_result::INVALID);
 }
+
+bool common_utf8_is_complete(const std::string & s) {
+    if (s.empty()) {
+        return true;
+    }
+    for (int i = 1; i <= std::min(4, (int)s.size()); i++) {
+        unsigned char c = s[s.size() - i];
+        if ((c & 0xC0) != 0x80) {
+            int expected = (c >= 0xF0) ? 4 : (c >= 0xE0) ? 3 : (c >= 0xC0) ? 2 : 1;
+            return i >= expected;
+        }
+    }
+    return false;
+}
+
+std::string common_unicode_cpts_to_utf8(const std::vector<uint32_t> & cps) {
+    std::string result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        result.append(common_unicode_cpt_to_utf8(cps[i]));
+    }
+    return result;
+}
+
+std::string common_unicode_cpt_to_utf8(uint32_t cpt) {
+    std::string result;
+
+    if (/* 0x00 <= cpt && */ cpt <= 0x7f) {
+        result.push_back(cpt);
+        return result;
+    }
+    if (0x80 <= cpt && cpt <= 0x7ff) {
+        result.push_back(0xc0 | ((cpt >> 6) & 0x1f));
+        result.push_back(0x80 | (cpt & 0x3f));
+        return result;
+    }
+    if (0x800 <= cpt && cpt <= 0xffff) {
+        result.push_back(0xe0 | ((cpt >> 12) & 0x0f));
+        result.push_back(0x80 | ((cpt >> 6) & 0x3f));
+        result.push_back(0x80 | (cpt & 0x3f));
+        return result;
+    }
+    if (0x10000 <= cpt && cpt <= 0x10ffff) {
+        result.push_back(0xf0 | ((cpt >> 18) & 0x07));
+        result.push_back(0x80 | ((cpt >> 12) & 0x3f));
+        result.push_back(0x80 | ((cpt >> 6) & 0x3f));
+        result.push_back(0x80 | (cpt & 0x3f));
+        return result;
+    }
+
+    throw std::invalid_argument("invalid codepoint");
+}
+
+
+

common/unicode.h

@@ -2,6 +2,8 @@
 
 #include <cstdint>
 #include <string_view>
+#include <vector>
+#include <string>
 
 // UTF-8 parsing utilities for streaming-aware unicode support
 
@@ -16,7 +18,13 @@ struct utf8_parse_result {
 
 // Determine the expected length of a UTF-8 sequence from its first byte
 // Returns 0 for invalid first bytes
-size_t utf8_sequence_length(unsigned char first_byte);
+size_t common_utf8_sequence_length(unsigned char first_byte);
+
+// Check if a string ends with a complete UTF-8 sequence.
+bool common_utf8_is_complete(const std::string & s);
 
 // Parse a single UTF-8 codepoint from input
-utf8_parse_result parse_utf8_codepoint(std::string_view input, size_t offset);
+utf8_parse_result common_parse_utf8_codepoint(std::string_view input, size_t offset);
+
+std::string common_unicode_cpts_to_utf8(const std::vector<uint32_t> & cps);
+std::string common_unicode_cpt_to_utf8(uint32_t cpt);

convert_hf_to_gguf.py

@@ -144,6 +144,7 @@ class ModelBase:
         self.metadata_override = metadata_override
         self.model_name = model_name
         self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
+        self._is_nvfp4 = False
 
         # Apply heuristics to figure out typical tensor encoding based on first tensor's dtype
         # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
@@ -271,6 +272,9 @@ class ModelBase:
         return tensors
 
     def dequant_model(self):
+        if self._is_nvfp4:
+            return  # NVFP4 weights are repacked in _generate_nvfp4_tensors
+
         tensors_to_remove: list[str] = []
         new_tensors: dict[str, Callable[[], Tensor]] = {}
 
@@ -516,6 +520,13 @@ class ModelBase:
         raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # skip NVFP4 auxiliary tensors (handled in _generate_nvfp4_tensors)
+        if self._is_nvfp4:
+            if name.endswith((".weight_scale", ".weight_scale_2", ".input_scale", ".k_scale", ".v_scale")):
+                return []
+            if name.endswith(".weight") and name.replace(".weight", ".weight_scale") in self.model_tensors:
+                return []
+
         new_name = self.map_tensor_name(name)
 
         # Handle gate/up expert tensor fusion if enabled
@@ -551,9 +562,135 @@ class ModelBase:
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         return ()
 
+    @staticmethod
+    def _nvfp4_pack(weight: Tensor, scale: Tensor) -> tuple[np.ndarray, list[int]]:
+        """Repack NVFP4 ModelOpt tensors into ggml super-block layout.
+        Preserves original E4M3 scale bits as UE4M3 (strip sign bit).
+        The per-tensor scale2 factor is stored as a separate tensor and applied at inference time via ggml_mul().
+        Returns (raw_data, logical_shape)."""
+
+        out_features = weight.shape[0]
+        n_blocks = scale.shape[1]
+
+        # Unpack ModelOpt nibble-packed weights
+        w = weight.reshape(out_features, n_blocks, 8)
+        vals = torch.stack([w & 0x0F, w >> 4], dim=-1).reshape(out_features, n_blocks, 16)
+
+        # Preserve original E4M3 scale bits as UE4M3 (strip sign bit)
+        d_ue = scale.view(torch.uint8).numpy().reshape(out_features, n_blocks) & 0x7F
+        qs = (vals[:, :, :8] | (vals[:, :, 8:] << 4)).to(torch.uint8).numpy()
+
+        # Pack into super-blocks: [4 UE4M3 scales, 32 qs bytes] = 36 bytes per 64 elements
+        n_super = n_blocks // 4
+        d_grouped = d_ue.reshape(out_features, n_super, 4)
+        qs_grouped = qs.reshape(out_features, n_super, 4, 8).reshape(out_features, n_super, 32)
+        raw = np.concatenate([d_grouped, qs_grouped], axis=-1).reshape(out_features, n_super * 36)
+        return raw, [out_features, n_super * 64]
+
+    @staticmethod
+    def _nvfp4_scale2_is_trivial(scale2: Tensor) -> bool:
+        return scale2.numel() <= 1 and abs(float(scale2.float().sum()) - 1.0) < 1e-6
+
+    def _repack_nvfp4(self, new_name: str, weight: Tensor, scale: Tensor, scale2: Tensor):
+        raw, shape = self._nvfp4_pack(weight, scale)
+        logger.info(f"Repacked {new_name} with shape {shape} and quantization NVFP4")
+        self.gguf_writer.add_tensor(new_name, raw, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
+
+        # Emit per-tensor scale2 as a separate F32 tensor when non-trivial
+        if not self._nvfp4_scale2_is_trivial(scale2):
+            scale2_f32 = scale2.float().numpy().flatten()
+            scale_name = new_name.replace(".weight", ".scale")
+            logger.info(f"  + {scale_name} (per-tensor NVFP4 scale2, shape [{scale2_f32.size}])")
+            self.gguf_writer.add_tensor(scale_name, scale2_f32)
+
+    def _generate_nvfp4_tensors(self):
+        # Per-layer expert merging to avoid holding all experts in memory
+        expert_blocks: dict[tuple[int, str], list[tuple[int, np.ndarray]]] = {}
+        expert_scales: dict[tuple[int, str], list[tuple[int, float]]] = {}
+        expert_shapes: dict[tuple[int, str], list[int]] = {}
+        n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=True) or 0
+
+        for name in list(self.model_tensors.keys()):
+            if not name.endswith(".weight"):
+                continue
+            scale_name = name.replace(".weight", ".weight_scale")
+            scale2_name = name.replace(".weight", ".weight_scale_2")
+            if scale_name not in self.model_tensors:
+                continue
+            # Force eager materialization of lazy tensors
+            weight = LazyTorchTensor.to_eager(self.model_tensors[name]())
+            scale = LazyTorchTensor.to_eager(self.model_tensors[scale_name]())
+            scale2 = LazyTorchTensor.to_eager(self.model_tensors.get(scale2_name, lambda: torch.tensor(1.0))())
+
+            # Check if this is a per-expert tensor
+            m = re.search(r'\.experts\.(\d+)\.(gate_proj|up_proj|down_proj)\.weight$', name)
+            if m:
+                expert_id = int(m.group(1))
+                proj_type = m.group(2)
+                bid_m = re.search(r'\.layers\.(\d+)\.', name)
+                bid = int(bid_m.group(1)) if bid_m else 0
+                key = (bid, proj_type)
+
+                raw, shape = self._nvfp4_pack(weight, scale)
+
+                if key not in expert_blocks:
+                    expert_blocks[key] = []
+                    expert_scales[key] = []
+                    expert_shapes[key] = shape
+                expert_blocks[key].append((expert_id, raw.copy()))
+                # Collect per-expert scale2 (scalar per expert)
+                expert_scales[key].append((expert_id, float(scale2.float().sum())))
+
+                # Flush when all experts for this (layer, proj) are collected
+                if n_experts > 0 and len(expert_blocks[key]) >= n_experts:
+                    self._flush_nvfp4_experts(key, expert_blocks, expert_scales, expert_shapes, bid, proj_type)
+            else:
+                new_name = self.map_tensor_name(name)
+                self._repack_nvfp4(new_name, weight, scale, scale2)
+
+        # Flush any remaining experts (fallback if n_experts was unknown)
+        for (bid, proj_type) in list(expert_blocks.keys()):
+            self._flush_nvfp4_experts((bid, proj_type), expert_blocks, expert_scales, expert_shapes, bid, proj_type)
+
+    def _flush_nvfp4_experts(self, key, expert_blocks, expert_scales, expert_shapes, bid, proj_type):
+        experts = expert_blocks.pop(key)
+        scales = expert_scales.pop(key)
+        shape = expert_shapes.pop(key)
+
+        experts.sort(key=lambda x: x[0])
+        merged = np.stack([e[1] for e in experts], axis=0)
+        merged_name = f"model.layers.{bid}.mlp.experts.{proj_type}.weight"
+        new_name = self.map_tensor_name(merged_name)
+        logger.info(f"Repacked {new_name} with shape [{len(experts)}, {shape[0]}, {shape[1]}] and quantization NVFP4")
+        self.gguf_writer.add_tensor(new_name, merged, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
+
+        # Emit per-expert scale2 tensor if any expert has non-trivial scale2
+        scales.sort(key=lambda x: x[0])
+        scale_vals = np.array([s[1] for s in scales], dtype=np.float32)
+        if not np.allclose(scale_vals, 1.0, atol=1e-6):
+            scale_name = new_name.replace(".weight", ".scale")
+            logger.info(f"  + {scale_name} (per-expert NVFP4 scale2, shape [{len(scales)}])")
+            self.gguf_writer.add_tensor(scale_name, scale_vals)
+
+        del experts, merged
+
     def prepare_tensors(self):
+        # detect NVFP4 quantization (ModelOpt format)
+        quant_algo = (self.hparams.get("quantization_config") or {}).get("quant_algo")
+        quant_config_file = self.dir_model / "hf_quant_config.json"
+
+        if not quant_algo and quant_config_file.is_file():
+            with open(quant_config_file, "r", encoding="utf-8") as f:
+                quant_algo = (json.load(f).get("quantization") or {}).get("quant_algo")
+
+        self._is_nvfp4 = quant_algo == "NVFP4"
+
         self.dequant_model()
 
+        # NVFP4 weights are repacked and written directly to gguf_writer
+        if self._is_nvfp4:
+            self._generate_nvfp4_tensors()
+
         # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
         if self.tensor_map.mapping:
             max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@@ -4303,6 +4440,14 @@ class Qwen2MoeModel(TextModel):
         # process the experts separately
         name = name.replace("language_model.", "") # InternVL
 
+        # NVFP4 expert weights are handled in _generate_nvfp4_tensors
+        if self._is_nvfp4 and "experts" in name:
+            if name.endswith((".weight", ".weight_scale", ".weight_scale_2", ".input_scale")):
+                if name.endswith(".weight") and name.replace(".weight", ".weight_scale") in self.model_tensors:
+                    return
+                if not name.endswith(".weight"):
+                    return
+
         # handle aggregated expert tensors
         # GGUF stores dimensions reversed from PyTorch, so:
         # PyTorch (A,B,C) -> GGUF writes [C,B,A] -> GGML reads ne={C,B,A}
@@ -4390,15 +4535,31 @@ class Qwen3Model(Qwen2Model):
         hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
         self.origin_hf_arch = hparams.get('architectures', [None])[0]
 
-        # a bit hacky, but currently the only way to detect if this is a rerank model
-        # ref: https://huggingface.co/Qwen/Qwen3-Reranker-0.6B
+        if self._is_qwen3_reranker():
+            self._find_rerank_config()
+
+    def _is_qwen3_reranker(self) -> bool:
         readme_path = self.dir_model / "README.md"
         readme_text = ""
         if readme_path.exists():
             with readme_path.open("r", encoding="utf-8") as f:
                 readme_text = f.read()
-        if "# Qwen3-Reranker" in readme_text:
-            self._find_rerank_config()
+
+        name_hints = [
+            str(self.dir_model.name),
+            str(self.hparams.get("_name_or_path", "")),
+            str(self.hparams.get("model_type", "")),
+            str(self.origin_hf_arch or ""),
+        ]
+        name_hints = [hint.lower() for hint in name_hints if hint]
+
+        if "# qwen3-reranker" in readme_text.lower() or "# qwen3-vl-reranker" in readme_text.lower():
+            return True
+
+        if any("qwen3-reranker" in hint or "qwen3-vl-reranker" in hint for hint in name_hints):
+            return True
+
+        return "sequenceclassification" in (self.origin_hf_arch or "").lower()
 
     def set_vocab(self):
         # deal with intern-s1-mini
@@ -4901,7 +5062,7 @@ class Phi2Model(TextModel):
         self.gguf_writer.add_add_bos_token(False)
 
 
-@ModelBase.register("Phi3ForCausalLM")
+@ModelBase.register("Phi3ForCausalLM", "Phi4ForCausalLMV")
 class Phi3MiniModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PHI3
 
@@ -5076,6 +5237,129 @@ class Phi3MiniModel(TextModel):
         yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
         yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith(("model.vision_tower.", "vision_tower.", "model.mm_projector.", "mm_projector.")):
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Phi4ForCausalLMV")
+class Phi4VisionMmprojModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+
+        self.vision_total_layers = int(self.find_vparam(self.n_block_keys))
+        if self.vision_total_layers < 2:
+            raise ValueError(
+                f"Phi-4 vision mmproj conversion requires at least 2 vision layers, got {self.vision_total_layers}"
+            )
+
+        # Phi-4 uses SigLIP2 hidden_states[-2], so export one fewer encoder block and
+        # drop post-layernorm/head weights. This makes the GGUF runtime output match
+        # the feature map consumed by the patched siglip.cpp Phi-4 projector path.
+        self.vision_export_layers = self.vision_total_layers - 1
+        self.vision_last_layer_idx = self.vision_total_layers - 1
+
+        for key in self.n_block_keys:
+            if key in self.hparams_vision:
+                self.hparams_vision[key] = self.vision_export_layers
+                break
+
+        self.block_count = self.vision_export_layers
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.MMPROJ, self.block_count)
+
+        patch_size = self.preprocessor_config.get("patch_size")
+        if patch_size is None:
+            raise KeyError("Phi-4 vision mmproj conversion requires patch_size in preprocessor_config.json")
+
+        self.hparams_vision["patch_size"] = patch_size
+
+        pos_emb_name = next(
+            (
+                name for name in self.model_tensors
+                if name.endswith("vision_model.embeddings.position_embedding.weight")
+            ),
+            None,
+        )
+        if pos_emb_name is None:
+            raise KeyError("Phi-4 vision mmproj conversion could not find position_embedding.weight")
+
+        pos_emb_shape = self.model_tensors[pos_emb_name]().shape
+        base_grid_tokens = int(pos_emb_shape[0])
+        grid_side = math.isqrt(base_grid_tokens)
+        if grid_side * grid_side != base_grid_tokens:
+            raise ValueError(f"Unexpected Phi-4 position embedding shape: {tuple(pos_emb_shape)}")
+
+        self.hparams_vision["image_size"] = grid_side * patch_size
+
+        min_num_patches = self.preprocessor_config.get("min_num_patches", self.global_config.get("min_num_patches"))
+        max_num_patches = self.preprocessor_config.get("max_num_patches", self.global_config.get("max_num_patches"))
+        if min_num_patches is None or max_num_patches is None:
+            raise KeyError("Phi-4 vision mmproj conversion requires min_num_patches and max_num_patches")
+
+        self.min_pixels = int(min_num_patches) * patch_size * patch_size
+        self.max_pixels = int(max_num_patches) * patch_size * patch_size
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PHI4)
+        self.gguf_writer.add_vision_min_pixels(self.min_pixels)
+        self.gguf_writer.add_vision_max_pixels(self.max_pixels)
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith(("model.vision_tower.vision_tower.", "vision_tower.")):
+            if ".vision_model.head." in name:
+                return
+
+            new_name = name.replace("model.vision_tower.vision_tower.", "vision_tower.")
+
+            if ".vision_model.post_layernorm." in new_name:
+                return
+
+            if bid is not None and bid == self.vision_last_layer_idx:
+                return
+
+            if new_name.endswith("vision_model.embeddings.patch_embedding.weight"):
+                assert self.hparams_vision is not None
+                if data_torch.ndim != 2:
+                    raise ValueError(f"Unexpected Phi-4 patch embedding shape: {tuple(data_torch.shape)}")
+
+                patch_area = self.hparams_vision["patch_size"] ** 2
+                in_features = data_torch.shape[1]
+                if in_features % patch_area != 0:
+                    raise ValueError(
+                        f"Phi-4 patch embedding input dim {in_features} is not divisible by patch area {patch_area}"
+                    )
+
+                num_channels = in_features // patch_area
+                patch_size = self.hparams_vision["patch_size"]
+                data_torch = data_torch.view(data_torch.shape[0], patch_size, patch_size, num_channels)
+                data_torch = data_torch.permute(0, 3, 1, 2)
+
+            yield from super().modify_tensors(data_torch, new_name, bid)
+            return
+
+        if name.startswith(("model.mm_projector.", "mm_projector.")):
+            local_name = name
+            local_name = local_name.replace("model.mm_projector.", "")
+            local_name = local_name.replace("mm_projector.", "")
+
+            if not (local_name.startswith("0.") or local_name.startswith("2.")):
+                return
+
+            suffix = ".bias" if local_name.endswith(".bias") else ".weight"
+            mm_idx = int(local_name.split(".", maxsplit=1)[0])
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, mm_idx, suffix=suffix), data_torch)
+            return
+
+        return
+
 
 @ModelBase.register("PhiMoEForCausalLM")
 class PhiMoeModel(Phi3MiniModel):
@@ -9727,20 +10011,35 @@ class NemotronHModel(GraniteHybridModel):
         # M: Mamba2, *: Attention, -: MLP
         # MoE:
         # M: Mamba2, *: Attention, E: Expert
-        hybrid_override_pattern = self.hparams["hybrid_override_pattern"]
-        self._ssm_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == "M"]
-        self._mlp_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == ("E" if self.is_moe else "-")]
+        pattern = self.hparams.get("hybrid_override_pattern") or self.hparams.get("layers_block_type")
+        if pattern is None:
+            self._ssm_layers = []
+            self._mlp_layers = []
+        elif isinstance(pattern, str):
+            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "M"]
+            self._mlp_layers = [i for i, val in enumerate(pattern) if val == ("E" if self.is_moe else "-")]
+        else:
+            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "mamba"]
+            self._mlp_layers = [i for i, val in enumerate(pattern) if val == "moe"]
 
     def get_attn_layers(self):
-        hybrid_override_pattern = self.hparams["hybrid_override_pattern"]
-        assert len(hybrid_override_pattern) == self.block_count, "Mismatch between hybrid override and num_hidden_layers!"
-        return [i for i, val in enumerate(hybrid_override_pattern) if val == "*"]
+        pattern = self.hparams.get("hybrid_override_pattern") or self.hparams.get("layers_block_type")
+        if pattern is None:
+            return []
+        assert len(pattern) == self.block_count, f"Mismatch between pattern ({len(pattern)}) and block_count ({self.block_count})!"
+        if isinstance(pattern, str):
+            return [i for i, val in enumerate(pattern) if val == "*"]
+
+        return [i for i, val in enumerate(pattern) if val == "attention"]
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-        self.gguf_writer.add_key_length(self.head_dim)
-        self.gguf_writer.add_value_length(self.head_dim)
+        head_dim = self.head_dim
+        if head_dim is None:
+            raise ValueError("Could not find the attention head dim in config")
+        self.gguf_writer.add_key_length(head_dim)
+        self.gguf_writer.add_value_length(head_dim)
 
         # Set feed_forward_length
         # NOTE: This will trigger an override warning. This is preferable to
@@ -9768,6 +10067,9 @@ class NemotronHModel(GraniteHybridModel):
             if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
                 self.gguf_writer.add_expert_used_count(n_experts_used)
 
+            if (latent_size := self.hparams.get("moe_latent_size")) is not None:
+                self.gguf_writer.add_moe_latent_size(latent_size)
+
     def set_vocab(self):
         super().set_vocab()
 
@@ -9787,6 +10089,13 @@ class NemotronHModel(GraniteHybridModel):
             name = name[len("language_model."):]
 
         if self.is_moe and bid is not None:
+            # Skip Multi-Token Prediction (MTP) tensors. These are used for
+            # for speculative decoding but we don't include them in this model
+            # conversion. See https://github.com/ggml-org/llama.cpp/pull/18886
+            if name.startswith("mtp."):
+                logger.info(f"gguf: Skipping MTP (Speculative) layer: {name}")
+                return
+
             if name.endswith("mixer.gate.e_score_correction_bias"):
                 new_name = name.replace("e_score_correction_bias", "e_score_correction.bias")
                 yield from ModelBase.modify_tensors(self, data_torch, new_name, bid)

docs/autoparser.md

@@ -0,0 +1,525 @@
+# Auto-Parser Architecture
+
+The auto-parser automatically analyzes chat templates to determine how to parse model outputs, including content, reasoning, and tool calls.
+
+## Overview
+
+The unified auto-parser uses a pure differential, compositional approach (inspired by the `git diff` algorithm) to analyze chat templates:
+
+**Core Philosophy**:
+
+- **Minimize Hardcoded Patterns**: All markers extracted through template comparison (the only heuristic is JSON detection to distinguish `JSON_NATIVE` from tag-based formats)
+- **Compositional Architecture**: Separate analyzer structs for reasoning, content, and tools — each responsible for its own analysis and parser construction
+
+**Analysis + Parser Building in Two Steps**:
+
+1. `autoparser::autoparser tmpl_analysis(tmpl)` — runs all differential comparisons and populates the analysis structs
+2. `autoparser::peg_generator::generate_parser(tmpl, params, tmpl_analysis)` — uses the analysis to build a PEG parser and optional GBNF grammar
+
+## Data Structures
+
+All structs are defined in [common/chat-auto-parser.h](common/chat-auto-parser.h).
+
+### Top-Level: `autoparser` (main analyzer and generator)
+
+[common/chat-auto-parser.h:367-388](common/chat-auto-parser.h#L367-L388) — top-level analysis result aggregating `jinja_caps`, `reasoning`, `content`, and `tools` sub-analyses, plus `preserved_tokens` (union of all non-empty markers).
+
+### `analyze_reasoning`
+
+[common/chat-auto-parser.h:254-274](common/chat-auto-parser.h#L254-L274) — reasoning analysis result: `mode` enum, `start` marker (e.g. `<think>`), and `end` marker (e.g. `</think>`).
+
+### `analyze_content`
+
+[common/chat-auto-parser.h:280-295](common/chat-auto-parser.h#L280-L295) — content analysis result: `mode` enum, `start`/`end` markers, and `requires_nonnull_content` flag.
+
+### `analyze_tools` and its sub-structs
+
+- [common/chat-auto-parser.h:176-194](common/chat-auto-parser.h#L176-L194) — `tool_format_analysis`: `mode` enum, `section_start/end`, `per_call_start/end`, JSON field names (`function_field`, `name_field`, `args_field`, `id_field`, `gen_id_field`), and format flags (`fun_name_is_key`, `tools_array_wrapped`, `uses_python_dicts`)
+- [common/chat-auto-parser.h:196-200](common/chat-auto-parser.h#L196-L200) — `tool_function_analysis`: `name_prefix`, `name_suffix`, `close` markers around function names
+- [common/chat-auto-parser.h:202-210](common/chat-auto-parser.h#L202-L210) — `tool_arguments_analysis`: `start/end` container markers, `name_prefix/suffix`, `value_prefix/suffix`, `separator`
+- [common/chat-auto-parser.h:212-217](common/chat-auto-parser.h#L212-L217) — `tool_id_analysis`: `pos` enum, `prefix`/`suffix` markers around call ID values
+- [common/chat-auto-parser.h:301-361](common/chat-auto-parser.h#L301-L361) — `analyze_tools`: aggregates the four sub-structs above
+
+### Enums
+
+**`reasoning_mode`**: How the template handles reasoning/thinking blocks.
+
+| Value           | Description                                                                       |
+|-----------------|-----------------------------------------------------------------------------------|
+| `NONE`          | No reasoning markers detected                                                     |
+| `TAG_BASED`     | Standard tag-based: `<think>...</think>`                                          |
+| `DELIMITER`     | Delimiter-based: reasoning ends at a delimiter (e.g., `[BEGIN FINAL RESPONSE]`)   |
+| `FORCED_OPEN`   | Template ends with open reasoning tag when `enable_thinking=true`                 |
+| `FORCED_CLOSED` | `enable_thinking=false` emits both tags; `enable_thinking=true` emits only start  |
+| `TOOLS_ONLY`    | Reasoning only appears in tool call responses, not plain content                  |
+
+**`content_mode`**: How the template wraps assistant content.
+
+| Value                    | Description                                                    |
+|--------------------------|----------------------------------------------------------------|
+| `PLAIN`                  | No content markers                                             |
+| `ALWAYS_WRAPPED`         | Content always wrapped: `<response>...</response>`             |
+| `WRAPPED_WITH_REASONING` | Content wrapped only when reasoning is present                 |
+
+**`tool_format`**: Classification of tool call structure.
+
+| Value            | Description                                                      |
+|------------------|------------------------------------------------------------------|
+| `NONE`           | No tool support detected                                         |
+| `JSON_NATIVE`    | Pure JSON: `{"name": "X", "arguments": {...}}`                   |
+| `TAG_WITH_JSON`  | Tag-based with JSON args: `<function=X>{...}</function>`         |
+| `TAG_WITH_TAGGED`| Tag-based with tagged args: `<param=key>value</param>`           |
+
+**`call_id_position`**: Where call IDs appear in tag-based formats.
+
+| Value                    | Description                                  |
+|--------------------------|----------------------------------------------|
+| `NONE`                   | No call ID support detected                  |
+| `PRE_FUNC_NAME`          | Before function name                         |
+| `BETWEEN_FUNC_AND_ARGS`  | Between function name and arguments          |
+| `POST_ARGS`              | After arguments                              |
+
+## Tool Calling Formats
+
+### JSON_NATIVE
+
+**Structure**: The entire tool call (function name, arguments, values) is in JSON format. Optional enclosing tags around the section.
+
+**Detection**: Function name appears inside a JSON structure (quotes preceded by `{` or `:`).
+
+**Examples**:
+
+Standard OpenAI-style:
+
+```json
+<tool_call>
+{"name": "get_weather", "arguments": {"location": "Paris", "unit": "celsius"}}
+</tool_call>
+```
+
+Mistral Nemo with array wrapper:
+
+```json
+[TOOL_CALLS]
+[{"name": "calculate", "arguments": {"expr": "2+2"}}]
+```
+
+Function name as JSON key (Apertus style):
+
+```json
+{"get_weather": {"location": "Paris"}}
+```
+
+---
+
+### TAG_WITH_JSON
+
+**Structure**: Function name is outside JSON, in tag attributes or XML-style tags. Arguments are a JSON object.
+
+**Detection**: Function name not in JSON, but argument names appear in JSON context.
+
+**Examples**:
+
+Functionary v3.1:
+
+```xml
+<function=get_weather>{"location": "Paris", "unit": "celsius"}</function>
+```
+
+MiniMax:
+
+```xml
+<minimax:tool_call>
+<tool_name>calculate</tool_name>
+<arguments>{"expr": "2+2"}</arguments>
+</minimax:tool_call>
+```
+
+---
+
+### TAG_WITH_TAGGED
+
+**Structure**: Both function name and argument names are in XML-style tags. String values are unquoted; non-string values are JSON-formatted.
+
+**Detection**: Neither function name nor argument names appear in a JSON context.
+
+**Examples**:
+
+Qwen/Hermes XML format:
+
+```xml
+<function=get_weather>
+<param=location>Paris</param>
+<param=unit>celsius</param>
+</function>
+```
+
+Mixed types:
+
+```xml
+<function=calculate>
+<param=expr>2+2</param>
+<param=precision>2</param>
+<param=options>{"round": true}</param>
+</function>
+```
+
+String values (`Paris`, `celsius`, `2+2`) are unquoted; `options` (object type) is JSON-formatted.
+
+---
+
+## Analysis Flow
+
+```text
+autoparser::autoparser(tmpl)
+    |
+    |-- Phase 1: analyze_reasoning(tmpl, jinja_caps.supports_tool_calls)
+    |     |-- R1: compare_reasoning_presence()   — with/without reasoning_content field
+    |     |-- R2: compare_thinking_enabled()     — enable_thinking=false vs true
+    |     '-- R3: compare_reasoning_scope()      — reasoning+content vs reasoning+tools
+    |           (only if supports_tool_calls)
+    |
+    |-- Phase 2: analyze_content(tmpl, reasoning)
+    |     '-- C1: compares content-only vs tools output and content-only vs reasoning output
+    |
+    |-- Phase 3: analyze_tools(tmpl, jinja_caps, reasoning)
+    |     (skipped entirely if !jinja_caps.supports_tool_calls)
+    |     |
+    |     |-- T1: analyze_tool_calls()           — no tools vs with tools; classifies format
+    |     |         |-- JSON path → analyze_tool_call_format_json_native()
+    |     |         '-- tag path → analyze_tool_call_format_non_json()
+    |     |
+    |     (if format != NONE and format != JSON_NATIVE:)
+    |     |
+    |     |-- T2: check_per_call_markers()       — 1 call vs 2 calls; moves section→per-call if needed
+    |     |         (only if supports_parallel_tool_calls)
+    |     |
+    |     |-- T3: extract_function_markers()     — func_alpha vs func_beta; extracts name prefix/suffix/close
+    |     |
+    |     |-- T4: analyze_arguments()            — (TAG_WITH_TAGGED only)
+    |     |         |-- A1: extract_argument_name_markers()   — arg_name_A vs arg_name_B
+    |     |         '-- A2: extract_argument_value_markers()  — value "XXXX" vs "YYYY"
+    |     |
+    |     |-- T5: extract_argument_separator()   — 1 arg vs 2 args; finds separator between args
+    |     |
+    |     |-- T6: extract_args_markers()         — 0 args vs 1 arg; finds args container markers
+    |     |
+    |     '-- T7: extract_call_id_markers()      — call_id "call00001" vs "call99999"
+    |
+    '-- collect_preserved_tokens()               — union of all non-empty markers
+    |
+    '-- apply workarounds()                      — post-hoc patches for edge-case templates
+    |
+    v
+autoparser (analysis result)
+    |
+    v
+autoparser::peg_generator::generate_parser(tmpl, inputs, analysis)
+    |-- analysis.build_parser(inputs)            — builds PEG parser arena
+    |     |-- reasoning.build_parser(ctx)        — reasoning parser (mode-dependent)
+    |     |-- content.build_parser(ctx)          — content parser (mode-dependent)
+    |     '-- tools.build_parser(ctx)            — tool parser (dispatches by tool_format)
+    |           |-- build_tool_parser_json_native()
+    |           |-- build_tool_parser_tag_json()
+    |           '-- build_tool_parser_tag_tagged()
+    |
+    |-- Build GBNF grammar (if tools present and trigger_marker non-empty)
+    '-- Set grammar_triggers from section_start or per_call_start
+    |
+    v
+common_chat_params (prompt, parser, grammar, triggers, preserved_tokens)
+```
+
+## Entry Point
+
+The auto-parser is invoked in [common/chat.cpp:1280-1310](common/chat.cpp#L1280-L1310) in `common_chat_templates_apply_jinja`. A few specialized templates are handled first (Ministral/Magistral Large 3, GPT-OSS with `<|channel|>`, Functionary v3.2 with `>>>all`), then the auto-parser handles everything else via `autoparser::autoparser` + `peg_generator::generate_parser`.
+
+## Algorithm Details
+
+### Core Mechanism: Differential Comparison
+
+All analysis phases use the same factorized comparison function declared in [common/chat-auto-parser-helpers.h:68](common/chat-auto-parser-helpers.h#L68):
+
+```cpp
+compare_variants(tmpl, params_A, params_modifier)
+```
+
+This creates variant B by applying a modifier lambda to a copy of `params_A`, renders both through the template, and computes a `diff_split` ([common/chat-auto-parser.h:28-37](common/chat-auto-parser.h#L28-L37)):
+
+- `prefix` — common prefix between A and B
+- `suffix` — common suffix between A and B
+- `left` — unique to variant A
+- `right` — unique to variant B
+
+The diff is computed via `calculate_diff_split()`, which finds the longest-common-prefix and longest-common-suffix, then iteratively moves incomplete `<...>` or `[...]` markers from the prefix/suffix into left/right until stable (tag boundary fixing).
+
+Text is segmentized into markers and non-marker fragments using `segmentize_markers()`, which splits on `<...>` and `[...]` boundaries.
+
+### Phase 1: Reasoning Analysis
+
+**R1 — `compare_reasoning_presence()`**: Compares assistant message with vs without a `reasoning_content` field.
+
+- Searches `diff.right` (output with reasoning) for the reasoning content needle
+- Uses PEG parsers to find surrounding markers:
+  - If both pre/post markers found in `diff.right` → `TAG_BASED` (both tags visible in diff = no forced close)
+  - If both found but post marker only in the full output B → `FORCED_CLOSED`
+  - If only post marker found → `DELIMITER`
+- Sets `reasoning.start` and `reasoning.end`
+
+**R2 — `compare_thinking_enabled()`**: Compares `enable_thinking=false` vs `true` with a generation prompt.
+
+- Detects `FORCED_OPEN`: `enable_thinking=true` adds a non-empty marker at the end of the prompt (where model will start generating) — sets `reasoning.start`, mode = `FORCED_OPEN`
+- Detects `FORCED_CLOSED`: `enable_thinking=false` produces both start+end markers; `enable_thinking=true` produces only start marker
+- Handles the reverse case: if both start and end are still empty, looks for a single-segment diff on each side to extract both markers
+
+**R3 — `compare_reasoning_scope()`**: Compares assistant message with reasoning+text-content vs reasoning+tool-calls.
+
+- Only runs if `jinja_caps.supports_tool_calls`
+- Detects `TOOLS_ONLY`: reasoning content present in B (with tools) but not in A (with text content)
+- Extracts reasoning markers from the tool call output using PEG parsers
+
+### Phase 2: Content Analysis
+
+**C1**: Two comparisons in the `analyze_content` constructor:
+
+- Comparison 1: content-only output vs tool-call output → `diff_tools`
+- Comparison 2: content-only output vs reasoning+empty-content output → `diff_reasoning`
+
+Classification logic:
+
+- `PLAIN`: `diff_tools.left` equals the response string (content is the entire diff, no wrapper)
+- `ALWAYS_WRAPPED`: markers found surrounding the content text in `pure_content` → extracts `start`/`end`
+
+### Phase 3: Tool Call Analysis
+
+**T1 — `analyze_tool_calls()`**: Compares no-tools vs with-tools output.
+
+- Extracts the tool call section as `diff.right`
+- Calls `analyze_tool_call_format()` which first strips reasoning markers from the haystack, then:
+  - Calls `in_json_haystack()` for both function name and argument name needles
+  - `in_json_haystack()` uses a PEG parser to check whether the needle appears in a JSON context (preceded by `{` or `:` with surrounding quotes)
+  - If function name is in JSON → `JSON_NATIVE` → `analyze_tool_call_format_json_native()`
+  - If function name not in JSON, arg name is in JSON → `TAG_WITH_JSON`
+  - If neither in JSON → `TAG_WITH_TAGGED`
+  - `analyze_tool_call_format_json_native()`: parses the JSON object, matches field values to needles to populate `name_field`, `args_field`, `id_field`, `gen_id_field`; detects `tools_array_wrapped`; extracts `section_start`/`section_end`
+  - `analyze_tool_call_format_non_json()`: uses PEG parsers on the haystack to find up to two opening markers (section + per-call) then up to two closing markers
+
+**T2 — `check_per_call_markers()`**: Compares 1 call vs 2 calls.
+
+- Computes a secondary diff of the second call portion vs the common suffix
+- If the second call content starts with `section_start` → the section marker is actually per-call → moves `section_start/end` to `per_call_start/end` and clears the section markers
+
+**T3 — `extract_function_markers()`**: Compares function name `FUN_FIRST` vs `FUN_SECOND` (two different named functions).
+
+- Finds where the function name appears in `diff.left`
+- Extracts `function.name_prefix` from the common prefix up to the function marker, and `function.name_suffix` from after the name up to the next marker
+- Extends `name_suffix` into `diff.suffix` (to the first marker for TAG_WITH_TAGGED; to the first `{` or `[` for TAG_WITH_JSON)
+- Extracts `function.close` from after the last argument value up to the per-call/section end marker
+
+**T4 — `analyze_arguments()`** (TAG_WITH_TAGGED only):
+
+- **A1 `extract_argument_name_markers()`**: Compares `arg_name_A` vs `arg_name_B` (two different argument names).
+  - Finds shared surrounding structure → `arguments.name_prefix`, `arguments.name_suffix`
+- **A2 `extract_argument_value_markers()`**: Compares argument value `"XXXX"` vs `"YYYY"` (same arg, different value).
+  - Finds markers surrounding the value → `arguments.value_prefix`, `arguments.value_suffix`
+
+**T5 — `extract_argument_separator()`**: Compares 1 argument vs 2 arguments (same function).
+
+- Uses `until_common_prefix(diff.right, ARG_FIRST, ARG_SECOND)` to find what separates the two argument blocks
+
+**T6 — `extract_args_markers()`**: Compares 0 arguments vs 1 argument.
+
+- Uses `until_common_prefix()` and `after_common_suffix()` with the empty and single-arg JSON strings as anchors to find container markers (`arguments.start`, `arguments.end`)
+
+**T7 — `extract_call_id_markers()`**: Compares call IDs `"call00001"` vs `"call99999"`.
+
+- Determines whether function name appears in `diff.prefix` or `diff.suffix` to classify position:
+  - Function name in prefix only → `BETWEEN_FUNC_AND_ARGS` or `POST_ARGS` (further distinguished by where `{` appears)
+  - Function name in suffix only → `PRE_FUNC_NAME`
+- Extracts `call_id.prefix` and `call_id.suffix` markers around the call ID value
+- Clears `per_call_end` if it incorrectly incorporated the call ID suffix
+
+### Workarounds
+
+A workaround array in `common/chat-diff-analyzer.cpp` applies post-hoc patches after analysis. Each workaround is a lambda that inspects the template source and overrides analysis results. Current workarounds:
+
+1. **Old Qwen/DeepSeek thinking templates** — source contains `content.split('</think>')`: sets `reasoning.mode = FORCED_OPEN` with `<think>`/`</think>` markers if no reasoning was detected
+2. **Granite 3.3** — source contains specific "Write your thoughts" text: forces `TAG_BASED` reasoning with `<think>`/`</think>` and `WRAPPED_WITH_REASONING` content with `<response>`/`</response>`
+3. **Cohere Command R+** — source contains `<|CHATBOT_TOKEN|>`: sets `ALWAYS_WRAPPED` content mode if no content start is already set
+4. **Functionary 3.1** — source contains `set has_code_interpreter`: forces `PLAIN` content, specific `per_call_start/end`, clears preserved tokens to only keep Functionary-specific markers
+5. **DeepSeek-R1-Distill-Qwen** — source contains `tool▁calls▁begin` markers: overrides tool section/per-call markers with the correct Unicode block characters
+
+### Parser Building
+
+Each analyzer struct (`analyze_reasoning`, `analyze_content`, `analyze_tools`) implements `build_parser(parser_build_context&)`. They share a `parser_build_context` that carries the PEG builder, inference inputs, the pre-built reasoning parser, and a pointer to the content analyzer.
+
+#### Reasoning Parser (`analyze_reasoning::build_parser`)
+
+| Mode                              | Parser                                                              |
+|-----------------------------------|---------------------------------------------------------------------|
+| Not extracting reasoning          | `eps()`                                                             |
+| `FORCED_OPEN` or `FORCED_CLOSED`  | `reasoning(until(end)) + end` — opening tag was in the prompt       |
+| `TAG_BASED` or `TOOLS_ONLY`       | `optional(start + reasoning(until(end)) + end)`                     |
+| `DELIMITER`                       | `optional(reasoning(until(end)) + end)` — no start marker           |
+
+#### Content Parser (`analyze_content::build_parser`)
+
+| Condition                              | Parser                                                                          |
+|----------------------------------------|---------------------------------------------------------------------------------|
+| `json_schema` present                  | `reasoning + space() + content(schema(json(), "response-format", ...)) + end()` |
+| Tools present                          | Dispatches to `analyze_tools::build_parser()`                                   |
+| `ALWAYS_WRAPPED` with reasoning        | `reasoning + start + content(until(end)) + end + end()`                         |
+| `ALWAYS_WRAPPED` without reasoning     | `content(until(start)) + start + content(until(end)) + end + end()`             |
+| Default (PLAIN)                        | `reasoning + content(rest()) + end()`                                           |
+
+#### Tool Parsers (`analyze_tools::build_parser`)
+
+Dispatches by `format.mode`:
+
+**`build_tool_parser_json_native()`**: Calls `p.standard_json_tools()` which internally dispatches to:
+
+- `build_json_tools_function_is_key()` — function name is the JSON key: `{"get_weather": {...}}`
+- `build_json_tools_nested_keys()` — nested: `{"function": {"name": "X", "arguments": {...}}}`
+- `build_json_tools_flat_keys()` — flat: `{"name": "X", "arguments": {...}}`
+
+Handles content wrappers, array wrapping (`tools_array_wrapped`), parallel calls, and `parameter_order`.
+
+**`build_tool_parser_tag_json()`**: For each tool function:
+
+```text
+tool_open(name_prefix + tool_name(literal(name)) + name_suffix) +
+    call_id_section +
+    tool_args(schema(json(), tool_schema))
+  [+ function.close if non-empty]
+```
+
+Wrapped in per-call markers (with optional parallel call repetition) then optionally in section markers.
+
+**`build_tool_parser_tag_tagged()`**: For each tool function, builds one parser per argument:
+
+- String types: `tool_arg_string_value(schema(until(value_suffix), ...))`
+- JSON types: `tool_arg_json_value(schema(json(), ...))`
+- Required args are plain; optional args wrapped in `optional()`
+- Arguments joined with `space()` between consecutive parsers
+
+For closing: uses `function.close` if present; otherwise uses `peek(per_call_end)` to avoid premature close during partial streaming; falls back to `tool_close(space())` to trigger mapper callbacks.
+
+All three tool parsers return:
+
+```text
+reasoning + optional(content(until(trigger_marker))) + tool_calls + end()
+```
+
+### Python Dict Format
+
+When `format.uses_python_dicts` is true (detected when single-quoted strings appear in JSON argument context), `build_parser()` pre-registers a `json-string` rule that accepts both single-quoted and double-quoted strings. This is done before any `p.json()` call so all JSON parsing inherits the flexible rule.
+
+## Mapper
+
+`common_chat_peg_mapper` maps PEG parse results (AST nodes) into `common_chat_msg` structures. Key design:
+
+- **Buffered arguments**: Before `tool_name` is known, argument text goes to `args_buffer`; once the name is set, the buffer is flushed to `current_tool->arguments`
+- **`args_target()`**: Returns a reference to whichever destination is currently active (buffer or tool args), eliminating branching
+- **`closing_quote_pending`**: Tracks whether a closing `"` needs to be appended when a string argument value is finalized (for schema-declared string types in tagged format)
+- **Quote normalization**: Python-style quotes (`'key': 'value'`) are converted to JSON (`"key": "value"`)
+- **Brace auto-closing**: At tool close, unclosed `{` braces are closed automatically
+
+## Files
+
+| File                                      | Purpose                                                              |
+|-------------------------------------------|----------------------------------------------------------------------|
+| `common/chat-auto-parser.h`               | All analysis structs, enums, `autoparser`, `peg_generator`, `templates_params` |
+| `common/chat-auto-parser-generator.cpp`   | Parser generator: `generate_parser()` and `build_parser()` methods   |
+| `common/chat-diff-analyzer.cpp`           | Differential analysis implementation and workarounds                 |
+| `common/chat-auto-parser-helpers.h/cpp`   | `calculate_diff_split()`, `segmentize_markers()`,                    |
+|                                           | `compare_variants()`, string helpers                                 |
+| `common/chat-peg-parser.h/cpp`            | `common_chat_peg_builder`, `common_chat_peg_mapper`, and helpers     |
+| `common/chat.cpp`                         | Entry point: `common_chat_templates_apply_jinja()`                   |
+| `tools/parser/debug-template-parser.cpp`  | Debug tool for template analysis                                     |
+| `tools/parser/template-analysis.cpp`      | Template analysis tool                                               |
+
+## Testing & Debugging
+
+### Debug Tools
+
+**Template Debugger**: `tools/parser/debug-template-parser.cpp`
+
+- Usage: `./bin/llama-debug-template-parser path/to/template.jinja`
+- Shows detected format, markers, generated parser, and GBNF grammar
+
+**Template Analysis**: `tools/parser/template-analysis.cpp`
+
+- Usage: `./bin/llama-template-analysis path/to/template.jinja`
+
+**Debug Logging**: Enable with `LLAMA_LOG_VERBOSITY=2`
+
+- Shows detailed analysis steps, pattern extraction results, and generated parser structure
+
+**PEG Test Builder**: Fluent API for creating test cases — see [tests/test-chat.cpp:947-1043](tests/test-chat.cpp#L947-L1043). Example usage:
+
+```cpp
+auto tst = peg_tester("models/templates/Template.jinja");
+tst.test("input text")
+   .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+   .tools({tool_json})
+   .parallel_tool_calls(true)
+   .enable_thinking(true)
+   .expect(expected_message)
+   .run();
+```
+
+### Tested Templates
+
+The following templates have active tests in `tests/test-chat.cpp`:
+
+| Template | Format | Notes |
+| -------- | ------ | ----- |
+| Ministral-3-14B-Reasoning | Reasoning | `[THINK]...[/THINK]` tags (specialized handler) |
+| NVIDIA-Nemotron-3-Nano-30B | TAG_WITH_TAGGED | Reasoning + tools |
+| CohereForAI Command-R7B | JSON_NATIVE | `<\|START_THINKING\|>`/`<\|START_RESPONSE\|>` markers |
+| Google Gemma 2 2B | Content only | No tool support |
+| Qwen-QwQ-32B | Reasoning | Forced-open thinking |
+| NousResearch Hermes 2 Pro | JSON_NATIVE | `<tool_call>` wrapper |
+| IBM Granite 3.3 | JSON_NATIVE | `<think></think>` + `<response></response>` |
+| ByteDance Seed-OSS | TAG_WITH_TAGGED | Custom `<seed:think>` and `<seed:tool_call>` tags |
+| Qwen3-Coder | TAG_WITH_TAGGED | XML-style tool format |
+| DeepSeek V3.1 | JSON_NATIVE | Forced thinking mode |
+| GLM-4.6 | TAG_WITH_TAGGED | `<tool_call>name\n<arg_key>...<arg_value>...` format |
+| GLM-4.7-Flash | TAG_WITH_TAGGED | Updated GLM format |
+| Kimi-K2-Thinking | JSON_NATIVE | Reasoning + JSON tools |
+| Apertus-8B-Instruct | JSON_NATIVE | Function name as JSON key |
+| MiniMax-M2 | TAG_WITH_JSON | XML invoke with JSON args |
+| NVIDIA-Nemotron-Nano-v2 | JSON_NATIVE | `<TOOLCALL>` wrapper (nested) |
+| CohereForAI Command-R Plus | JSON_NATIVE | Markdown code block format |
+| Mistral-Nemo-Instruct-2407 | JSON_NATIVE | `[TOOL_CALLS]` wrapper with ID field |
+| Functionary v3.1 | TAG_WITH_JSON | `<function=X>` format |
+| Functionary v3.2 | Specialized | `>>>` recipient delimiter (dedicated handler) |
+| Fireworks Firefunction v2 | TAG_WITH_JSON | Fireworks tool format |
+| DeepSeek R1 Distill (Llama/Qwen) | Reasoning | Forced-open thinking |
+| llama-cpp-deepseek-r1 | Reasoning | Forced-open thinking |
+| Kimi-K2 / Kimi-K2-Instruct | JSON_NATIVE | JSON tools with special markers |
+| Llama 3.1/3.2/3.3 | JSON_NATIVE | Standard Llama tool format |
+| OpenAI GPT-OSS | Specialized | Channel-based (dedicated handler) |
+| Apriel 1.5 | JSON_NATIVE | `<tool_calls>` wrapper with JSON array |
+| Apriel 1.6 Thinker | Reasoning | Implicit reasoning start |
+| Mistral Small 3.2 | JSON_NATIVE | `[TOOL_CALLS]func[ARGS]{...}` with call ID |
+| Devstral | JSON_NATIVE | `[TOOL_CALLS]func[ARGS]{...}` without call ID |
+| StepFun 3.5 Flash | TAG_WITH_TAGGED | `<function=X><parameter=Y>` format |
+
+## Adding Support for New Templates
+
+To support a new template format:
+
+1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `llama-debug-template-parser` to verify markers are correctly extracted.
+2. **If differential analysis extracts incorrect markers** — Add a workaround lambda to the `workarounds` vector in `common/chat-diff-analyzer.cpp`. Inspect the template source for a unique identifying substring.
+3. **If it needs fundamentally different handling** — Add a dedicated handler function in `chat.cpp` before the auto-parser block (as done for GPT-OSS, Functionary v3.2, and Ministral).
+
+## Edge Cases and Quirks
+
+1. **Forced Thinking**: When `enable_thinking=true` and the model prompt ends with an open reasoning tag (e.g., `<think>`), the parser enters forced thinking mode and immediately expects reasoning content without waiting for a start marker.
+2. **Per-Call vs Per-Section Markers**: Some templates wrap each tool call individually (`per_call_start/end`); others wrap the entire section (`section_start/end`). T2 (`check_per_call_markers()`) disambiguates by checking if the second call in a two-call output starts with the section marker.
+3. **Python Dict Format**: The Seed template family uses single-quoted JSON (`'key': 'value'`). The `uses_python_dicts` flag causes the PEG builder to register a flexible `json-string` rule accepting both quote styles before any JSON rules are built.
+4. **Tag Boundary Fixing**: `calculate_diff_split()` iteratively adjusts prefix/suffix boundaries to avoid splitting `<tag>` or `[marker]` tokens, ensuring clean extraction.
+5. **Call ID Side Effects**: When a call ID is detected, `per_call_end` may have been incorrectly set to include the call ID suffix. T7 clears `per_call_end` in this case.
+6. **Tool Analysis Gating**: `analyze_tools` is only constructed (and all tool analysis phases run) when `jinja_caps.supports_tool_calls` is true. Within tool analysis, `check_per_call_markers()` (T2) only runs if `jinja_caps.supports_parallel_tool_calls`.
+7. **`analyze_arguments()` Gating**: Within tool analysis, A1 and A2 (argument name/value marker extraction) only run for `TAG_WITH_TAGGED` format. `extract_argument_separator()` and `extract_args_markers()` run for all non-`JSON_NATIVE` formats.

docs/backend/SYCL.md

@@ -9,6 +9,7 @@
 - [Linux](#linux)
 - [Windows](#windows)
 - [Environment Variable](#environment-variable)
+- [Design Rule](#design-rule)
 - [Known Issue](#known-issues)
 - [Q&A](#qa)
 - [TODO](#todo)
@@ -41,6 +42,9 @@ The following releases are verified and recommended:
 
 ## News
 
+- 2026.03
+  - Support Flash-Attention: less memory usage, performance impact depends on LLM.
+
 - 2026.02
   - Remove support for Nvidia & AMD GPU, because the oneAPI plugin for Nvidia & AMD GPU is unavailable: download/installation channels are out of work. User can't build up the software for Nvidia & AMD GPU.
 
@@ -378,17 +382,27 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 ## Windows
 
-### I. Setup Environment
-
-1. Install GPU driver
+### Install GPU driver
 
 Intel GPU drivers instructions guide and download page can be found here: [Get Intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
 
-2. Install Visual Studio
+### Option 1: download the binary package directly
+
+Download the binary package for Windows from: https://github.com/ggml-org/llama.cpp/releases.
+
+Extract the package to local folder, run the llama tools directly. Refer to [Run the inference](#iii-run-the-inference-1).
+
+Note, the package includes the SYCL running time and all depended dll files, no need to install oneAPI package and activte them.
+
+### Option 2: build locally from the source code.
+
+#### I. Setup environment
+
+1. Install Visual Studio
 
 If you already have a recent version of Microsoft Visual Studio, you can skip this step. Otherwise, please refer to the official download page for [Microsoft Visual Studio](https://visualstudio.microsoft.com/).
 
-3. Install Intel® oneAPI Base toolkit
+2. Install Intel® oneAPI Base toolkit
 
 SYCL backend depends on:
   - Intel® oneAPI DPC++/C++ compiler/running-time.
@@ -439,25 +453,25 @@ Output (example):
 [ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Iris(R) Xe Graphics 1.3 [1.3.28044]
 ```
 
-4. Install build tools
+3. Install build tools
 
 a. Download & install cmake for Windows: https://cmake.org/download/ (CMake can also be installed from Visual Studio Installer)
 b. The new Visual Studio will install Ninja as default. (If not, please install it manually: https://ninja-build.org/)
 
 
-### II. Build llama.cpp
+#### II. Build llama.cpp
 
 You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
 
 Choose one of following methods to build from source code.
 
-#### 1. Script
+##### Option 1: Script
 
 ```sh
 .\examples\sycl\win-build-sycl.bat
 ```
 
-#### 2. CMake
+##### Option 2: CMake
 
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
@@ -486,7 +500,7 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-completion
 ```
 
-#### 3. Visual Studio
+##### Option 3: Visual Studio
 
 You have two options to use Visual Studio to build llama.cpp:
 - As CMake Project using CMake presets.
@@ -496,7 +510,7 @@ You have two options to use Visual Studio to build llama.cpp:
 
 All following commands are executed in PowerShell.
 
-##### - Open as a CMake Project
+###### - Open as a CMake Project
 
 You can use Visual Studio to open the `llama.cpp` folder directly as a CMake project. Before compiling, select one of the SYCL CMake presets:
 
@@ -511,7 +525,7 @@ You can use Visual Studio to open the `llama.cpp` folder directly as a CMake pro
     cmake --build build --config Release -j --target llama-completion
     ```
 
-##### - Generating a Visual Studio Solution
+###### - Generating a Visual Studio Solution
 
 You can use Visual Studio solution to build and work on llama.cpp on Windows. You need to convert the CMake Project into a `.sln` file.
 
@@ -599,7 +613,7 @@ found 2 SYCL devices:
 
 ```
 
-#### Choose level-zero devices
+##### Choose level-zero devices
 
 |Chosen Device ID|Setting|
 |-|-|
@@ -607,7 +621,7 @@ found 2 SYCL devices:
 |1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
 |0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"` or `set ONEAPI_DEVICE_SELECTOR="level_zero:*"`|
 
-#### Execute
+##### Execute
 
 Choose one of following methods to run.
 
@@ -665,7 +679,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 ## Environment Variable
 
-#### Build
+### Build
 
 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
@@ -680,23 +694,50 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 1. FP32 or FP16 have different performance impact to LLM. Recommended to test them for better prompt processing performance on your models. You need to rebuild the code after change `GGML_SYCL_F16=OFF/ON`.
 
-#### Runtime
+### Runtime
 
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
+| GGML_SYCL_ENABLE_FLASH_ATTN | 1 (default) or 0| Enable Flash-Attention. It can reduce memory usage. The performance impact depends on the LLM.|
 | GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Support malloc device memory more than 4GB.|
 
+## Design Rule
 
+- Open to all contributors.
+
+- All code change should be useful to user:
+    - Fix bug.
+    - Add new function.
+    - Improve the performance/usage.
+    - Make code be easy to maintain.
+    - ...
+
+- Don't accept the codes of following cases:
+    - Break legacy function.
+    - Reduce the performance of legacy case in default.
+    - Not completed work/the functionality cannot be demonstrated.
+
+- Encourage to use environment variable to control features to be opened/closed.
+    - User can evaluate the feature without rebuild the code.
+    - Recommend the best features to user by setting them be opened as default.
+
+- Design the code based on the published official releases of oneAPI packages: compiler, library, driver, OS kernel.
+
+- Developers need to maintain the code they submit.
 
 ## Known Issues
 
 - `Split-mode:[row]` is not supported.
 
+- Missed the AOT (Ahead-of-Time) in buiding.
+  - Good: build quickly, smaller size of binary file.
+  - Bad: The startup is slow (JIT) in first time, but subsequent performance is unaffected.
+
 ## Q&A
 
 - Error:  `error while loading shared libraries: libsycl.so: cannot open shared object file: No such file or directory`.
@@ -746,7 +787,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
 
 ### **GitHub contribution**:
-Please add the `SYCL :` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
+Please add the `[SYCL]` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
 
 ## TODO
 

docs/backend/VirtGPU/development.md

@@ -55,7 +55,8 @@ LLAMA_MAC_BUILD=$PWD/build/ggml-virtgpu-backend
 cmake -S . -B $LLAMA_MAC_BUILD \
       -DGGML_NATIVE=OFF \
       -DLLAMA_CURL=ON \
-      -DGGML_REMOTINGBACKEND=ONLY \
+      -DGGML_VIRTGPU=ON \
+      -DGGML_VIRTGPU_BACKEND=ONLY \
       -DGGML_METAL=ON
 
 TARGETS="ggml-metal"
@@ -71,6 +72,7 @@ cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $EXTRA_TARGETS
 ```bash
 # Build virglrenderer with APIR support
 mkdir virglrenderer
+cd virglrenderer
 git clone https://gitlab.freedesktop.org/kpouget/virglrenderer -b main-macos src
 cd src
 
@@ -95,7 +97,7 @@ mkdir llama.cpp
 git clone https://github.com/ggml-org/llama.cpp.git src
 cd src
 
-LLAMA_LINUX_BUILD=$PWD//build-virtgpu
+LLAMA_LINUX_BUILD=$PWD/build-virtgpu
 
 cmake -S . -B $LLAMA_LINUX_BUILD \
       -DGGML_VIRTGPU=ON

docs/build.md

@@ -599,7 +599,13 @@ If KleidiAI is enabled, the output will contain a line similar to:
 ```
 load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
 ```
-KleidiAI's microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm and SME. llama.cpp selects the most efficient kernel based on runtime CPU feature detection. However, on platforms that support SME, you must manually enable SME microkernels by setting the environment variable `GGML_KLEIDIAI_SME=1`.
+KleidiAI’s microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm, SVE, and SME. Llama.cpp selects the most efficient kernels at runtime based on detected CPU capabilities.
+On CPUs that support SME, SME microkernels are enabled automatically using runtime detection.
+The environment variable GGML_KLEIDIAI_SME can be used to control SME behavior:
+- Not set: enable SME automatically if supported and detected.
+- 0: disable SME.
+- <n> > 0: enable SME and assume <n> available SME units (override auto detection).
+If SME is not supported by the CPU, SME microkernels are always disabled.
 
 Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`.
 

docs/development/parsing.md

@@ -22,7 +22,7 @@ Below is a contrived example demonstrating how to use the PEG parser to parse
 output from a model that emits arguments as JSON.
 
 ```cpp
-auto parser = build_chat_peg_native_parser([&](common_chat_peg_native_builder & p) {
+auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
     // Build a choice of all available tools
     auto tool_choice = p.choice();
     for (const auto & tool : tools) {
@@ -212,7 +212,7 @@ mapper.from_ast(ctx.ast, result);
 
 ### Native
 
-The `common_chat_peg_native_builder` builds a `native` parser suitable for
+The `common_chat_peg_builder` builds a `native` parser suitable for
 models that emit tool arguments as a direct JSON object.
 
 - **`reasoning(p)`** - Tag node for `reasoning_content`
@@ -225,7 +225,7 @@ models that emit tool arguments as a direct JSON object.
 - **`tool_args(p)`** - Tag the tool arguments
 
 ```cpp
-build_chat_peg_native_parser([&](common_chat_peg_native_parser & p) {
+build_chat_peg_parser([&](common_chat_peg_builder & p) {
     auto get_weather_tool = p.tool(p.sequence({
         p.tool_open(p.literal("{")),
         p.json_member("name", "\"" + p.tool_name(p.literal("get_weather")) + "\""),
@@ -246,7 +246,7 @@ build_chat_peg_native_parser([&](common_chat_peg_native_parser & p) {
 
 ### Constructed
 
-The `common_chat_peg_constructed_builder` builds a `constructed` parser
+The `common_chat_peg_builder` builds a `constructed` parser
 suitable for models that emit tool arguments as separate entities, such as XML
 tags.
 
@@ -264,7 +264,7 @@ tags.
 - **`tool_arg_json_value(p)`** - Tag JSON value for the argument
 
 ```cpp
-build_chat_peg_constructed_parser([&](common_chat_peg_constructed_builder & p) {
+build_chat_peg_parser([&](common_chat_peg_builder & p) {
     auto location_arg = p.tool_arg(
         p.tool_arg_open("<parameter name=\"" + p.tool_arg_name(p.literal("location")) + "\">"),
         p.tool_arg_string_value(p.until("</parameter>")),

docs/ops.md

@@ -23,7 +23,7 @@ Legend:
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
@@ -31,22 +31,23 @@ Legend:
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
-|                             DIAG | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                             DIAG | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                             FILL | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
-|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -54,7 +55,7 @@ Legend:
 |                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -63,7 +64,7 @@ Legend:
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                              LOG | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              LOG | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
@@ -75,34 +76,34 @@ Legend:
 |                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | 🟡 |
 |                              PAD | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                          POOL_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROPE | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                         SET_ROWS | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
-|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -116,5 +117,5 @@ Legend:
 |                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
-|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |
-|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |

docs/ops/WebGPU.csv

@@ -5023,20 +5023,20 @@
 "WebGPU: WebGPU","ARGMAX","type=f32,ne=[1024,12,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGMAX","type=f32,ne=[2000,10,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGMAX","type=f32,ne=[5438,3,1,1]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[2,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,2,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,2,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=i32,ne=[10,5,4,1],nr=[2,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=i16,ne=[10,5,4,1],nr=[1,1,1,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[2,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,2,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,2,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=i32,ne=[10,5,4,3],nr=[2,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=i16,ne=[10,5,4,3],nr=[1,1,1,2]","support","0","no","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[2,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,2,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,2,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=i32,ne=[10,5,4,1],nr=[2,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=i16,ne=[10,5,4,1],nr=[1,1,1,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[2,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,2,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,2,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=i32,ne=[10,5,4,3],nr=[2,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=i16,ne=[10,5,4,3],nr=[1,1,1,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,1,1,1],v=0","support","0","no","WebGPU"
 "WebGPU: WebGPU","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[2,1,1,1],v=0","support","0","no","WebGPU"
 "WebGPU: WebGPU","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,2,1,1],v=0","support","0","no","WebGPU"

examples/json_schema_to_grammar.py

@@ -633,7 +633,7 @@ class SchemaConverter:
             return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=None))
 
         elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):
-            items = schema.get('items') or schema['prefixItems']
+            items = schema.get('items', schema.get('prefixItems'))
             if isinstance(items, list):
                 return self._add_rule(
                     rule_name,
@@ -689,6 +689,11 @@ class SchemaConverter:
         elif (schema_type == 'object') or (len(schema) == 0):
             return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object']))
 
+        elif schema_type is None and isinstance(schema, dict):
+            # No type constraint and no recognized structural keywords (e.g. {"description": "..."}).
+            # Per JSON Schema semantics this is equivalent to {} and accepts any value.
+            return self._add_rule(rule_name, self._add_primitive('value', PRIMITIVE_RULES['value']))
+
         else:
             assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
             # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero

ggml/include/ggml-rpc.h

@@ -8,7 +8,12 @@ extern "C" {
 
 #define RPC_PROTO_MAJOR_VERSION    3
 #define RPC_PROTO_MINOR_VERSION    6
-#define RPC_PROTO_PATCH_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    1
+
+#ifdef  __cplusplus
+static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION");
+#endif
+
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API

ggml/include/ggml.h

@@ -427,7 +427,8 @@ extern "C" {
         // GGML_TYPE_IQ4_NL_4_8 = 37,
         // GGML_TYPE_IQ4_NL_8_8 = 38,
         GGML_TYPE_MXFP4   = 39, // MXFP4 (1 block)
-        GGML_TYPE_COUNT   = 40,
+        GGML_TYPE_NVFP4   = 40, // NVFP4 (4 blocks, E4M3 scale)
+        GGML_TYPE_COUNT   = 41,
     };
 
     // precision
@@ -463,6 +464,7 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
         GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
         GGML_FTYPE_MOSTLY_MXFP4   = 25, // except 1d tensors
+        GGML_FTYPE_MOSTLY_NVFP4   = 26, // except 1d tensors
     };
 
     // available tensor operations:
@@ -556,6 +558,7 @@ extern "C" {
         GGML_OP_GATED_LINEAR_ATTN,
         GGML_OP_RWKV_WKV7,
         GGML_OP_SOLVE_TRI,
+        GGML_OP_GATED_DELTA_NET,
 
         GGML_OP_UNARY,
 
@@ -2463,6 +2466,17 @@ extern "C" {
         bool                  lower,
         bool                  uni);
 
+    // TODO: add ggml_gated_delta_net_set_bcast() to be able to configure Q, K broadcast type: tiled vs interleaved [TAG_GGML_GDN_BCAST]
+    // ref: https://github.com/ggml-org/llama.cpp/pull/19468#discussion_r2786394306
+    GGML_API struct ggml_tensor * ggml_gated_delta_net(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * q,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * g,
+            struct ggml_tensor  * beta,
+            struct ggml_tensor  * state);
+
     // custom operators
 
     typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);

ggml/src/ggml-blas/ggml-blas.cpp

@@ -339,8 +339,8 @@ static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t
 }
 
 static void ggml_backend_blas_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
+    // no memory to report
+    *free  = 0;
     *total = 0;
 
     GGML_UNUSED(dev);

ggml/src/ggml-common.h

@@ -102,6 +102,9 @@ typedef sycl::half2 ggml_half2;
 #define QI_MXFP4 (QK_MXFP4 / (4 * QR_MXFP4))
 #define QR_MXFP4 2
 
+#define QI_NVFP4 (QK_NVFP4 / (4 * QR_NVFP4))
+#define QR_NVFP4 2
+
 #define QI5_0 (QK5_0 / (4 * QR5_0))
 #define QR5_0 2
 
@@ -194,6 +197,14 @@ typedef struct {
 } block_mxfp4;
 static_assert(sizeof(block_mxfp4) == sizeof(uint8_t) + QK_MXFP4/2, "wrong mxfp4 block size/padding");
 
+#define QK_NVFP4 64
+#define QK_NVFP4_SUB 16  // sub-block size for per-group scales
+typedef struct {
+    uint8_t d[QK_NVFP4/QK_NVFP4_SUB]; // UE4M3 scales (4 bytes, one per 16-element sub-block)
+    uint8_t qs[QK_NVFP4/2];           // packed 4-bit E2M1 values (32 bytes)
+} block_nvfp4;
+static_assert(sizeof(block_nvfp4) == sizeof(uint8_t)*(QK_NVFP4/QK_NVFP4_SUB) + QK_NVFP4/2, "wrong nvfp4 block size/padding");
+
 #define QK5_0 32
 typedef struct {
     ggml_half d;           // delta

ggml/src/ggml-cpu/arch-fallback.h

@@ -15,6 +15,7 @@
 #define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
 #define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -79,6 +80,8 @@
 #define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
+// quants.c
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
@@ -108,6 +111,7 @@
 // ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
@@ -155,6 +159,7 @@
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -201,9 +206,11 @@
 #define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
+#define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x1_generic ggml_quantize_mat_q8_K_4x1
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
@@ -239,6 +246,7 @@
 #elif defined(__s390x__)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -301,6 +309,7 @@
 #define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -650,6 +650,90 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 }
 
+void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_NVFP4 == 0);
+
+    const block_nvfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    // Each NVFP4 super-block (64 elements) spans 2 q8_0 blocks
+    const int nb = n / QK_NVFP4;
+
+    float sumf = 0;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_mxfp4);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    float32x4_t acc = vdupq_n_f32(0.0f);
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const uint8x16_t q4bits_0 = vld1q_u8(x[ib].qs);
+        const uint8x16_t q4bits_1 = vld1q_u8(x[ib].qs + 16);
+
+        const int8x16_t q4_lo_0 = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits_0, m4b));
+        const int8x16_t q4_hi_0 = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits_0, 4));
+        const int8x16_t q4_lo_1 = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits_1, m4b));
+        const int8x16_t q4_hi_1 = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits_1, 4));
+
+        const int8x16_t q8_0a = vld1q_s8(y[2*ib].qs);
+        const int8x16_t q8_0b = vld1q_s8(y[2*ib].qs + 16);
+        const int8x16_t q8_lo_0 = vcombine_s8(vget_low_s8(q8_0a), vget_low_s8(q8_0b));
+        const int8x16_t q8_hi_0 = vcombine_s8(vget_high_s8(q8_0a), vget_high_s8(q8_0b));
+
+        const int8x16_t q8_1a = vld1q_s8(y[2*ib+1].qs);
+        const int8x16_t q8_1b = vld1q_s8(y[2*ib+1].qs + 16);
+        const int8x16_t q8_lo_1 = vcombine_s8(vget_low_s8(q8_1a), vget_low_s8(q8_1b));
+        const int8x16_t q8_hi_1 = vcombine_s8(vget_high_s8(q8_1a), vget_high_s8(q8_1b));
+
+        const int32x4_t p0 = vaddq_s32(
+            ggml_vdotq_s32(vdupq_n_s32(0), q4_lo_0, q8_lo_0),
+            ggml_vdotq_s32(vdupq_n_s32(0), q4_hi_0, q8_hi_0));
+        const int32x4_t p1 = vaddq_s32(
+            ggml_vdotq_s32(vdupq_n_s32(0), q4_lo_1, q8_lo_1),
+            ggml_vdotq_s32(vdupq_n_s32(0), q4_hi_1, q8_hi_1));
+
+        const int32x4_t sums = vpaddq_s32(p0, p1);
+
+        // Decode 4 UE4M3 scales to f32 and multiply with q8 scales
+        const float dy0 = GGML_CPU_FP16_TO_FP32(y[2*ib].d);
+        const float dy1 = GGML_CPU_FP16_TO_FP32(y[2*ib+1].d);
+        const float32x4_t nvsc = {
+            ggml_ue4m3_to_fp32(x[ib].d[0]),
+            ggml_ue4m3_to_fp32(x[ib].d[1]),
+            ggml_ue4m3_to_fp32(x[ib].d[2]),
+            ggml_ue4m3_to_fp32(x[ib].d[3])
+        };
+        const float32x4_t scales = vmulq_f32(nvsc, (float32x4_t){dy0, dy0, dy1, dy1});
+
+        acc = vfmaq_f32(acc, vcvtq_f32_s32(sums), scales);
+    }
+    sumf = vaddvq_f32(acc);
+#else
+    for (int ib = 0; ib < nb; ++ib) {
+        for (int si = 0; si < 4; ++si) {
+            const float d = ggml_ue4m3_to_fp32(x[ib].d[si]);
+            const int q8b = si / 2;
+            const int q8o = (si % 2) * QK_NVFP4_SUB;
+            const float dy = GGML_CPU_FP16_TO_FP32(y[2*ib + q8b].d);
+
+            int sumi_lo = 0, sumi_hi = 0;
+            for (int j = 0; j < QK_NVFP4_SUB/2; ++j) {
+                const uint8_t qv = x[ib].qs[si*(QK_NVFP4_SUB/2) + j];
+                sumi_lo += y[2*ib + q8b].qs[q8o + j +               0] * kvalues_mxfp4[qv & 0xf];
+                sumi_hi += y[2*ib + q8b].qs[q8o + j + QK_NVFP4_SUB/2] * kvalues_mxfp4[qv >>  4];
+            }
+            sumf += dy * d * (sumi_lo + sumi_hi);
+        }
+    }
+#endif
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/ggml-cpu.c

@@ -270,6 +270,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
     },
+    [GGML_TYPE_NVFP4] = {
+        .from_float               = quantize_row_nvfp4,
+        .vec_dot                  = ggml_vec_dot_nvfp4_q8_0,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_Q2_K] = {
         .from_float               = quantize_row_q2_K,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
@@ -2021,6 +2027,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_solve_tri(params, tensor);
             } break;
+        case GGML_OP_GATED_DELTA_NET:
+            {
+                ggml_compute_forward_gated_delta_net(params, tensor);
+            } break;
         case GGML_OP_MAP_CUSTOM1:
             {
                 ggml_compute_forward_map_custom1(params, tensor);
@@ -2200,6 +2210,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_SOLVE_TRI:
+        case GGML_OP_GATED_DELTA_NET:
             {
                 n_tasks = n_threads;
             } break;
@@ -2905,6 +2916,11 @@ struct ggml_cplan ggml_graph_plan(
                     {
                         cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
                     } break;
+                case GGML_OP_GATED_DELTA_NET:
+                    {
+                        const int64_t S_v = node->src[2]->ne[0];
+                        cur = S_v * sizeof(float) * n_tasks;
+                    } break;
                 case GGML_OP_COUNT:
                     {
                         GGML_ABORT("fatal error");

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -520,7 +520,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
             /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
         },
-        /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
         /* .rhs_type           = */ GGML_TYPE_Q4_0,
         /* .op_type            = */ GGML_TYPE_F32,
@@ -631,7 +631,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
             /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
         },
-        /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
         /* .rhs_type           = */ GGML_TYPE_Q4_0,
         /* .op_type            = */ GGML_TYPE_F32,
@@ -801,7 +801,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
             /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
             /* .pack_func_ex          = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
         },
-        /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
         /* .rhs_type           = */ GGML_TYPE_Q8_0,
         /* .op_type            = */ GGML_TYPE_F32,

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -2497,7 +2497,7 @@ class tinyBLAS_Q0_PPC {
                 for (int r = 0; r < 8; r++) {
                     const block_q4_0 * current_blk = rows_base[r] + blk;
                     vector float v_scale = vec_extract_fp32_from_shorth(vec_splats(current_blk->d));
-                    vector signed char v_qs = reinterpret_cast<vector signed char>(vec_xl(0, current_blk->qs));
+                    vector signed char v_qs = vec_xl(0, (const vector signed char *)current_blk->qs);
                     vector signed char c1, c2;
                     unpack_q4_to_q8(v_qs, c1, c2);
                     convert_and_scale_q8(c1, v_scale, hp_res[r][0], hp_res[r][1]);
@@ -2611,14 +2611,14 @@ class tinyBLAS_Q0_PPC {
                 i = (cols >> 2);
                 if (i > 0) {
                     do {
-                        c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset1->qs));
-                        c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset2->qs));
-                        c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset3->qs));
-                        c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset4->qs));
-                        c5[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset5->qs));
-                        c6[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset6->qs));
-                        c7[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset7->qs));
-                        c8[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset8->qs));
+                        c1[1] = vec_xl(0, (const vector signed char *)aoffset1->qs);
+                        c2[1] = vec_xl(0, (const vector signed char *)aoffset2->qs);
+                        c3[1] = vec_xl(0, (const vector signed char *)aoffset3->qs);
+                        c4[1] = vec_xl(0, (const vector signed char *)aoffset4->qs);
+                        c5[1] = vec_xl(0, (const vector signed char *)aoffset5->qs);
+                        c6[1] = vec_xl(0, (const vector signed char *)aoffset6->qs);
+                        c7[1] = vec_xl(0, (const vector signed char *)aoffset7->qs);
+                        c8[1] = vec_xl(0, (const vector signed char *)aoffset8->qs);
 
                         process_q4_elements(c1, & comparray[0]);
                         process_q4_elements(c2, & comparray[1]);
@@ -2657,10 +2657,10 @@ class tinyBLAS_Q0_PPC {
             i = (cols >> 2);
             if (i > 0) {
                 do {
-                    c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset1->qs));
-                    c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset2->qs));
-                    c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset3->qs));
-                    c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset4->qs));
+                    c1[1] = vec_xl(0, (const vector signed char *)aoffset1->qs);
+                    c2[1] = vec_xl(0, (const vector signed char *)aoffset2->qs);
+                    c3[1] = vec_xl(0, (const vector signed char *)aoffset3->qs);
+                    c4[1] = vec_xl(0, (const vector signed char *)aoffset4->qs);
 
                     process_q4_elements(c1, & comparray[0]);
                     process_q4_elements(c2, & comparray[1]);
@@ -2686,9 +2686,9 @@ class tinyBLAS_Q0_PPC {
             if (i > 0) {
                 do {
                     switch(rows) {
-                        case 3: c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset3->qs));
-                        case 2: c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset2->qs));
-                        case 1: c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset1->qs));
+                        case 3: c3[1] = vec_xl(0, (const vector signed char *)aoffset3->qs);
+                        case 2: c2[1] = vec_xl(0, (const vector signed char *)aoffset2->qs);
+                        case 1: c1[1] = vec_xl(0, (const vector signed char *)aoffset1->qs);
                             break;
                     }
                     process_q4_elements(c1, & comparray[0]);

ggml/src/ggml-cpu/ops.cpp

@@ -670,6 +670,7 @@ void ggml_compute_forward_add(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1119,6 +1120,7 @@ void ggml_compute_forward_add1(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1247,6 +1249,7 @@ void ggml_compute_forward_acc(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -2129,12 +2132,12 @@ static void ggml_compute_forward_gelu_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2176,13 +2179,13 @@ static void ggml_compute_forward_gelu_f16(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2325,12 +2328,12 @@ static void ggml_compute_forward_gelu_erf_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2372,13 +2375,13 @@ static void ggml_compute_forward_gelu_erf_f16(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2444,12 +2447,12 @@ static void ggml_compute_forward_gelu_quick_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2491,13 +2494,13 @@ static void ggml_compute_forward_gelu_quick_f16(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2563,12 +2566,12 @@ static void ggml_compute_forward_silu_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2610,13 +2613,13 @@ static void ggml_compute_forward_silu_f16(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2766,7 +2769,7 @@ static void ggml_compute_forward_silu_back_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2802,7 +2805,7 @@ static void ggml_compute_forward_silu_back_f16(
                 (ggml_fp16_t *) ((char *) src1->data + i1*(src1->nb[1])),
                 (ggml_fp16_t *) ((char *) grad->data + i1*(grad->nb[1])));
 
-    #ifndef NDEBUG
+#ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
             const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
@@ -2810,7 +2813,7 @@ static void ggml_compute_forward_silu_back_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-    #endif
+#endif // NDEBUG
     }
 }
 
@@ -2893,7 +2896,7 @@ static void ggml_compute_forward_reglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2953,7 +2956,7 @@ static void ggml_compute_forward_reglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3036,7 +3039,7 @@ static void ggml_compute_forward_geglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3096,7 +3099,7 @@ static void ggml_compute_forward_geglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3179,7 +3182,7 @@ static void ggml_compute_forward_swiglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3239,7 +3242,7 @@ static void ggml_compute_forward_swiglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3330,7 +3333,7 @@ static void ggml_compute_forward_swiglu_oai_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3409,7 +3412,7 @@ static void ggml_compute_forward_geglu_erf_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3469,7 +3472,7 @@ static void ggml_compute_forward_geglu_erf_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3552,7 +3555,7 @@ static void ggml_compute_forward_geglu_quick_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3612,7 +3615,7 @@ static void ggml_compute_forward_geglu_quick_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -4334,6 +4337,7 @@ void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4609,6 +4613,7 @@ void ggml_compute_forward_set(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4831,6 +4836,7 @@ void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5303,7 +5309,7 @@ static void ggml_compute_forward_soft_max_f32(
                     //printf("p[%d] = %f\n", i, p[i]);
                     assert(!isnan(wp[i]));
                 }
-#endif
+#endif // NDEBUG
 
                 float max = -INFINITY;
                 ggml_vec_max_f32(ne00, &max, wp);
@@ -5328,7 +5334,7 @@ static void ggml_compute_forward_soft_max_f32(
                     assert(!isnan(dp[i]));
                     assert(!isinf(dp[i]));
                 }
-#endif
+#endif // NDEBUG
             }
         }
     }
@@ -5402,7 +5408,7 @@ static void ggml_compute_forward_soft_max_ext_back_f32(
             assert(!isnan(dy[i]));
             assert(!isnan(y[i]));
         }
-#endif
+#endif // NDEBUG
         // Jii = yi - yi*yi
         // Jij = -yi*yj
         // J = diag(y)-y.T*y
@@ -5435,7 +5441,7 @@ static void ggml_compute_forward_soft_max_ext_back_f32(
             assert(!isnan(dx[i]));
             assert(!isinf(dx[i]));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -5555,6 +5561,7 @@ void ggml_compute_forward_clamp(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5803,28 +5810,33 @@ static void ggml_compute_forward_rope_flt(
 
     const int32_t * pos = (const int32_t *) src1->data;
 
+    int64_t last_i2 = -1;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) { // batch
         for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len
-
-            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            if (!mrope_used) {
-                const int64_t p = pos[i2];
-                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
-            }
-            else {
-                const int64_t p_t = pos[i2];
-                const int64_t p_h = pos[i2 + ne2];
-                const int64_t p_w = pos[i2 + ne2 * 2];
-                const int64_t p_e = pos[i2 + ne2 * 3];
-                ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
-                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
-            }
-
             for (int64_t i1 = 0; i1 < ne1; i1++) { // attn-heads
-                if (ir++ < ir0) continue;
+                if (ir++ < ir0) continue; // skip rows mapped to other threads
                 if (ir   > ir1) break;
 
+                float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+                if (last_i2 != i2) {
+                    if (!mrope_used) {
+                        const int64_t p = pos[i2];
+                        ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+                    }
+                    else {
+                        const int64_t p_t = pos[i2];
+                        const int64_t p_h = pos[i2 + ne2];
+                        const int64_t p_w = pos[i2 + ne2 * 2];
+                        const int64_t p_e = pos[i2 + ne2 * 3];
+                        ggml_mrope_cache_init(
+                            p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
+                            freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+                    }
+
+                    last_i2 = i2;
+                }
+
                 T * src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
                 T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1);
 
@@ -10375,6 +10387,189 @@ void ggml_compute_forward_solve_tri(const struct ggml_compute_params * params, s
     }
 }
 
+// ggml_compute_forward_gated_delta_net
+static void ggml_compute_forward_gated_delta_net_one_chunk(
+    const ggml_compute_params * params,
+    ggml_tensor * dst,
+    int64_t ir0,
+    int64_t ir1) {
+
+    ggml_tensor * src_q     = dst->src[0];
+    ggml_tensor * src_k     = dst->src[1];
+    ggml_tensor * src_v     = dst->src[2];
+    ggml_tensor * src_g     = dst->src[3];
+    ggml_tensor * src_beta  = dst->src[4];
+    ggml_tensor * src_state = dst->src[5];
+
+    const int64_t S_v      = src_v->ne[0];
+    const int64_t H        = src_v->ne[1];
+    const int64_t n_tokens = src_v->ne[2];
+    const int64_t n_seqs   = src_v->ne[3];
+
+    GGML_ASSERT(ggml_is_contiguous_rows(src_q));
+    GGML_ASSERT(ggml_is_contiguous_rows(src_k));
+    GGML_ASSERT(ggml_is_contiguous_rows(src_v));
+    GGML_ASSERT(ggml_is_contiguous(src_g));
+    GGML_ASSERT(ggml_is_contiguous(src_beta));
+    GGML_ASSERT(ggml_is_contiguous(src_state));
+
+    GGML_ASSERT(src_g->ne[0] == 1 || src_g->ne[0] == S_v);
+    GGML_ASSERT(src_beta->ne[0] == 1);
+
+    GGML_TENSOR_LOCALS(int64_t, neq, src_q, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbq, src_q, nb);
+    GGML_TENSOR_LOCALS(int64_t, nek, src_k, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbk, src_k, nb);
+    GGML_TENSOR_LOCALS(int64_t, nev, src_v, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbv, src_v, nb);
+    GGML_TENSOR_LOCALS(int64_t, neg, src_g, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbg, src_g, nb);
+    GGML_TENSOR_LOCALS(size_t,  nbb, src_beta, nb);
+
+    const bool kda = (neg0 == S_v);
+
+    // scratch layout per thread: [delta(S_v)]
+    const int64_t scratch_per_thread = S_v;
+    const int ith = params->ith;
+
+    float * delta = (float *)params->wdata + ith * scratch_per_thread + CACHE_LINE_SIZE_F32;
+
+    // output layout: [attn_scores | new_states]
+    // attn_scores: S_v * H * n_tokens * n_seqs floats
+    // new_states:  S_v * S_v * H * n_seqs floats
+    const int64_t attn_score_elems = S_v * H * n_tokens * n_seqs;
+    float * attn_out_base  = (float *)dst->data;
+    float * state_out_base = (float *)dst->data + attn_score_elems;
+
+    const float * state_in_base = (const float *)src_state->data;
+
+  //const int64_t rq1 = nev1 / neq1;
+  //const int64_t rk1 = nev1 / nek1;
+    const int64_t rq3 = nev3 / neq3;
+    const int64_t rk3 = nev3 / nek3;
+
+    const float scale = 1.0f / sqrtf((float) S_v);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t iv1 = ir % H; // head_index
+        const int64_t iv3 = ir / H; // sequence
+
+        const int64_t iq1 = iv1 % neq1;
+        const int64_t ik1 = iv1 % nek1;
+
+        const int64_t iq3 = iv3 / rq3;
+        const int64_t ik3 = iv3 / rk3;
+
+        float * s_out = state_out_base + (iv3 * H + iv1) * S_v * S_v;
+
+        // copy input state into output buffer and operate in-place
+        const float * s_in = state_in_base + (iv3 * H + iv1) * S_v * S_v;
+        memcpy(s_out, s_in, S_v * S_v * sizeof(float));
+
+        // attn output pointer for first token of this (head, seq)
+        float * attn_data = attn_out_base + (iv3 * n_tokens * H + iv1) * S_v;
+
+        for (int64_t t = 0; t < n_tokens; t++) {
+            const float * q_d = (const float *)((const char *)src_q->data + iq3 * nbq3 + t * nbq2 + iq1 * nbq1);
+            const float * k_d = (const float *)((const char *)src_k->data + ik3 * nbk3 + t * nbk2 + ik1 * nbk1);
+            const float * v_d = (const float *)((const char *)src_v->data + iv3 * nbv3 + t * nbv2 + iv1 * nbv1);
+
+            const float beta_val = *(const float *)((const char *)src_beta->data + iv3 * nbb3 + t * nbb2 + iv1 * nbb1);
+            const float * g_d    =  (const float *)((const char *)src_g->data    + iv3 * nbg3 + t * nbg2 + iv1 * nbg1);
+
+            if (kda) {
+                for (int64_t i = 0; i < S_v; ++i) {
+                    ggml_vec_scale_f32(S_v, &s_out[i * S_v], expf(g_d[i]));
+                }
+            } else {
+                ggml_vec_scale_f32(S_v * S_v, s_out, expf(g_d[0]));
+            }
+
+            // delta[j] = sum_i S[j][i] * k[i]
+            memset(delta, 0, S_v * sizeof(float));
+            for (int64_t i = 0; i < S_v; ++i) {
+                ggml_vec_mad_f32(S_v, delta, &s_out[i * S_v], k_d[i]);
+            }
+            for (int64_t j = 0; j < S_v; ++j) {
+                delta[j] = (v_d[j] - delta[j]) * beta_val;
+            }
+
+            // outer product: S[j][i] += k[i] * delta[j]
+            for (int64_t i = 0; i < S_v; ++i) {
+                ggml_vec_mad_f32(S_v, &s_out[i * S_v], delta, k_d[i]);
+            }
+
+            // attn_out[j] = sum_i S[j][i] * q[i]
+            memset(attn_data, 0, S_v * sizeof(float));
+            for (int64_t i = 0; i < S_v; ++i) {
+                ggml_vec_mad_f32(S_v, attn_data, &s_out[i * S_v], q_d[i]);
+            }
+            ggml_vec_scale_f32(S_v, attn_data, scale);
+
+            attn_data += S_v * H; // advance to next token
+        }
+    }
+}
+
+
+static void ggml_compute_forward_gated_delta_net_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    ggml_tensor * V = dst->src[2];
+    int64_t nr = V->ne[1] * V->ne[3];
+
+    // disable for NUMA
+    const bool disable_chunking = ggml_is_numa();
+
+    int nth = params->nth;
+    int ith = params->ith;
+
+    // 4x chunks per thread
+    int nth_scaled = nth * 4;
+    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+    if (nth == 1 || nchunk < nth || disable_chunking) {
+      nchunk = nth;
+    }
+
+    if (ith == 0) {
+      ggml_threadpool_chunk_set(params->threadpool, nth);
+    }
+
+    ggml_barrier(params->threadpool);
+
+    const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+    int current_chunk = ith;
+
+    while (current_chunk < nchunk) {
+        const int64_t ir0 = dr * current_chunk;
+        const int64_t ir1 = MIN(ir0 + dr, nr);
+
+        ggml_compute_forward_gated_delta_net_one_chunk(params, dst, ir0, ir1);
+        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+    }
+}
+
+void ggml_compute_forward_gated_delta_net(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_gated_delta_net_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_rwkv_wkv7
 
 static void ggml_compute_forward_rwkv_wkv7_f32(
@@ -10700,7 +10895,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
             assert(!isnan(s0[i]));
             assert(!isnan(s1[i]));
         }
-#endif
+#endif // NDEBUG
 
         float max = -INFINITY;
         ggml_vec_max_f32(nc, &max, s0);
@@ -10719,7 +10914,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
             assert(!isnan(st[i]));
             assert(!isinf(st[i]));
         }
-#endif
+#endif // NDEBUG
     }
     sums[ith] = sum_thread;
     ggml_barrier(params->threadpool);
@@ -10792,7 +10987,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
             assert(!isnan(s0[i]));
             assert(!isnan(s1[i]));
         }
-#endif
+#endif // NDEBUG
 
         // soft_max
         float max = -INFINITY;
@@ -10810,7 +11005,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
             assert(!isnan(ds0[i]));
             assert(!isinf(ds0[i]));
         }
-#endif
+#endif // NDEBUG
     }
 }
 

ggml/src/ggml-cpu/ops.h

@@ -102,6 +102,7 @@ void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, s
 void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_solve_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_gated_delta_net(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/quants.c

@@ -50,6 +50,10 @@ void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, i
     quantize_row_mxfp4_ref(x, y, k);
 }
 
+void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_nvfp4_ref(x, y, k);
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -216,6 +220,42 @@ void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     *s = sumf;
 }
 
+// NVFP4: super-block of 64 elements = 4 sub-blocks of 16 = 2 q8_0 blocks
+void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_NVFP4 == 0);
+
+    const block_nvfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_NVFP4;
+
+    float sumf = 0;
+
+    for (int ib = 0; ib < nb; ++ib) {
+        for (int s_idx = 0; s_idx < 4; ++s_idx) {
+            const float d = ggml_ue4m3_to_fp32(x[ib].d[s_idx]);
+            const int q8_block = s_idx / 2;
+            const int q8_off   = (s_idx % 2) * QK_NVFP4_SUB;
+            const float dy = GGML_CPU_FP16_TO_FP32(y[2*ib + q8_block].d);
+
+            int sumi_lo = 0, sumi_hi = 0;
+            for (int j = 0; j < QK_NVFP4_SUB/2; ++j) {
+                const uint8_t qv = x[ib].qs[s_idx*(QK_NVFP4_SUB/2) + j];
+                sumi_lo += y[2*ib + q8_block].qs[q8_off + j +               0] * kvalues_mxfp4[qv & 0xf];
+                sumi_hi += y[2*ib + q8_block].qs[q8_off + j + QK_NVFP4_SUB/2] * kvalues_mxfp4[qv >>  4];
+            }
+
+            sumf += dy * d * (sumi_lo + sumi_hi);
+        }
+    }
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/quants.h

@@ -20,6 +20,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -42,6 +43,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -73,6 +75,7 @@ void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
 void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);

ggml/src/ggml-cpu/repack.h

@@ -28,13 +28,17 @@ template <int K, int N> struct block {
 // control size
 static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
 static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<4, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<4,16> size/padding");
 static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
 static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+static_assert(sizeof(block<8, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 16, "wrong block<8,16> size/padding");
 
 using block_q4_0x4 = block<4, 4>;
 using block_q4_0x8 = block<4, 8>;
+using block_q4_0x16 = block<4, 16>;
 using block_q8_0x4 = block<8, 4>;
 using block_q8_0x8 = block<8, 8>;
+using block_q8_0x16 = block<8, 16>;
 
 struct block_q4_Kx8 {
     ggml_half d[8];      // super-block scale for quantized scales
@@ -44,7 +48,14 @@ struct block_q4_Kx8 {
 };
 
 static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+struct block_q4_Kx16 {
+    ggml_half d[16];      // super-block scale for quantized scales
+    ggml_half dmin[16];   // super-block scale for quantized mins
+    uint8_t scales[192];  // scales and mins, quantized with 6 bits
+    uint8_t qs[2048];    // 4--bit quants
+};
 
+static_assert(sizeof(block_q4_Kx16) == sizeof(ggml_half) * 32 + K_SCALE_SIZE * 16 + QK_K * 8, "wrong q4_K block size/padding");
 struct block_q2_Kx8 {
     ggml_half d[8];      // super-block scale for quantized scales
     ggml_half dmin[8];   // super-block scale for quantized mins
@@ -53,6 +64,13 @@ struct block_q2_Kx8 {
 };
 
 static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
+struct block_q2_Kx16 {
+    ggml_half d[16];       // Super-block scale for quantized scales
+    ggml_half dmin[16];    // Super-block scale for quantized mins
+    uint8_t   scales[256]; // Sub-block scales (16 cols * 16 sub-blocks)
+    uint8_t   qs[1024];    // Data (16 cols * 64 bytes per block)
+};
+static_assert(sizeof(block_q2_Kx16) == sizeof(ggml_half) * 32 + QK_K + QK_K * 4, "wrong q2_K block size/padding");
 
 struct block_q5_Kx8 {
     ggml_half d[8];              // super-block scale for quantized scales
@@ -97,6 +115,12 @@ struct block_iq4_nlx8 {
 
 static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
 
+struct block_iq4_nlx16 {
+    ggml_half d[16];            // deltas for 16 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 8];  // nibbles / quants for 16 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx16) == 16 * sizeof(ggml_half) + QK4_NL * 8, "wrong iq4_nlx16 block size/padding");
 struct block_mxfp4x4 {
     uint8_t e[4];
     uint8_t qs[QK_MXFP4 * 2];
@@ -109,7 +133,6 @@ struct block_mxfp4x8 {
 };
 static_assert(sizeof(block_mxfp4x8) == 8 + QK_MXFP4 * 4, "wrong mxfp4x8 block size/padding");
 
-
 #if defined(__cplusplus)
 extern "C" {
 #endif
@@ -132,6 +155,8 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -146,10 +171,22 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#if defined __riscv_zvfh
+void ggml_quantize_mat_q8_0_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif
 
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -170,6 +207,8 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
 void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -184,10 +223,22 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
 void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#if defined __riscv_zvfh
+void ggml_quantize_mat_q8_0_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif
 
 #if defined(__cplusplus)
 } // extern "C"

ggml/src/ggml-cuda/gated_delta_net.cu

@@ -0,0 +1,275 @@
+#include "gated_delta_net.cuh"
+
+template <int S_v, bool KDA>
+__global__ void gated_delta_net_cuda(const float * q,
+                                     const float * k,
+                                     const float * v,
+                                     const float * g,
+                                     const float * beta,
+                                     const float * curr_state,
+                                     float *       dst,
+                                     int64_t       H,
+                                     int64_t       n_tokens,
+                                     int64_t       n_seqs,
+                                     int64_t       sq1,
+                                     int64_t       sq2,
+                                     int64_t       sq3,
+                                     int64_t       sv1,
+                                     int64_t       sv2,
+                                     int64_t       sv3,
+                                     int64_t       sb1,
+                                     int64_t       sb2,
+                                     int64_t       sb3,
+                                     const uint3   neqk1_magic,
+                                     const uint3   rq3_magic,
+                                     float         scale) {
+    const uint32_t h_idx    = blockIdx.x;
+    const uint32_t sequence = blockIdx.y;
+    // each warp owns one column, using warp-level primitives to reduce across rows
+    const int      lane     = threadIdx.x;
+    const int      col      = blockIdx.z * blockDim.y + threadIdx.y;
+
+    const uint32_t iq1 = fastmodulo(h_idx, neqk1_magic);
+    const uint32_t iq3 = fastdiv(sequence, rq3_magic);
+
+    const int64_t attn_score_elems = S_v * H * n_tokens * n_seqs;
+    float *       attn_data        = dst;
+    float *       state            = dst + attn_score_elems;
+
+    const int64_t state_offset = (sequence * H + h_idx) * S_v * S_v;
+    state += state_offset;
+    curr_state += state_offset;
+    attn_data += (sequence * n_tokens * H + h_idx) * S_v;
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v;
+    static_assert(S_v % warp_size == 0, "S_v must be a multiple of warp_size");
+    constexpr int rows_per_lane = (S_v + warp_size - 1) / warp_size;
+    float         s_shard[rows_per_lane];
+#pragma unroll
+    for (int r = 0; r < rows_per_lane; r++) {
+        const int i = r * warp_size + lane;
+        s_shard[r]  = curr_state[i * S_v + col];
+    }
+
+    for (int t = 0; t < n_tokens; t++) {
+        const float * q_t = q + iq3 * sq3 + t * sq2 + iq1 * sq1;
+        const float * k_t = k + iq3 * sq3 + t * sq2 + iq1 * sq1;
+        const float * v_t = v + sequence * sv3 + t * sv2 + h_idx * sv1;
+
+        const int64_t gb_offset = sequence * sb3 + t * sb2 + h_idx * sb1;
+        const float * beta_t = beta + gb_offset;
+        const float * g_t    = g    + gb_offset * (KDA ? S_v : 1);
+
+        const float beta_val = *beta_t;
+
+        if constexpr (!KDA) {
+            const float g_val = expf(*g_t);
+
+            // kv[col] = (S^T @ k)[col] = sum_i S[i][col] * k[i]
+            float kv_shard = 0.0f;
+#pragma unroll
+            for (int r = 0; r < rows_per_lane; r++) {
+                const int i = r * warp_size + lane;
+                kv_shard += s_shard[r] * k_t[i];
+            }
+            float kv_col = warp_reduce_sum<warp_size>(kv_shard);
+
+            // delta[col] = (v[col] - g * kv[col]) * beta
+            float delta_col = (v_t[col] - g_val * kv_col) * beta_val;
+
+            // fused: S[i][col] = g * S[i][col] + k[i] * delta[col]
+            // attn[col] = (S^T @ q)[col] = sum_i S[i][col] * q[i]
+            float attn_partial = 0.0f;
+#pragma unroll
+            for (int r = 0; r < rows_per_lane; r++) {
+                const int i = r * warp_size + lane;
+                s_shard[r]  = g_val * s_shard[r] + k_t[i] * delta_col;
+                attn_partial += s_shard[r] * q_t[i];
+            }
+
+            float attn_col = warp_reduce_sum<warp_size>(attn_partial);
+
+            if (lane == 0) {
+                attn_data[col] = attn_col * scale;
+            }
+        } else {
+            // kv[col] = sum_i g[i] * S[i][col] * k[i]
+            float kv_shard = 0.0f;
+#pragma unroll
+            for (int r = 0; r < rows_per_lane; r++) {
+                const int i = r * warp_size + lane;
+                kv_shard += expf(g_t[i]) * s_shard[r] * k_t[i];
+            }
+
+            float kv_col = warp_reduce_sum<warp_size>(kv_shard);
+
+            // delta[col] = (v[col] - kv[col]) * beta
+            float delta_col = (v_t[col] - kv_col) * beta_val;
+
+            // fused: S[i][col] = g[i] * S[i][col] + k[i] * delta[col]
+            // attn[col] = (S^T @ q)[col] = sum_i S[i][col] * q[i]
+            float attn_partial = 0.0f;
+#pragma unroll
+            for (int r = 0; r < rows_per_lane; r++) {
+                const int i = r * warp_size + lane;
+                s_shard[r]  = expf(g_t[i]) * s_shard[r] + k_t[i] * delta_col;
+                attn_partial += s_shard[r] * q_t[i];
+            }
+
+            float attn_col = warp_reduce_sum<warp_size>(attn_partial);
+
+            if (lane == 0) {
+                attn_data[col] = attn_col * scale;
+            }
+        }
+
+        attn_data += S_v * H;
+    }
+
+    // Write state back to global memory
+#pragma unroll
+    for (int r = 0; r < rows_per_lane; r++) {
+        const int i          = r * warp_size + lane;
+        state[i * S_v + col] = s_shard[r];
+    }
+}
+
+static size_t calculate_smem(const int sv, int cc)
+{
+    size_t smem = 0;
+    if ((GGML_CUDA_CC_IS_AMD(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_RDNA4(cc)) || GGML_CUDA_CC_IS_MTHREADS(cc)) {
+        smem = sv * sv * sizeof(float);
+    }
+    return smem;
+}
+
+template <bool KDA>
+static void launch_gated_delta_net(
+        const float * q_d, const float * k_d, const float * v_d,
+        const float * g_d, const float * b_d, const float * s_d,
+        float * dst_d,
+        int64_t S_v,   int64_t H, int64_t n_tokens, int64_t n_seqs,
+        int64_t sq1,   int64_t sq2, int64_t sq3,
+        int64_t sv1,   int64_t sv2, int64_t sv3,
+        int64_t sb1,   int64_t sb2, int64_t sb3,
+        int64_t neqk1, int64_t rq3,
+        float scale, cudaStream_t stream) {
+    //TODO: Add chunked kernel for even faster pre-fill
+    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
+    const int num_warps = 4;
+    dim3      grid_dims(H, n_seqs, (S_v + num_warps - 1) / num_warps);
+    dim3      block_dims(warp_size <= S_v ? warp_size : S_v, num_warps, 1);
+
+    const uint3 neqk1_magic = init_fastdiv_values(neqk1);
+    const uint3 rq3_magic   = init_fastdiv_values(rq3);
+
+    int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    switch (S_v) {
+        case 16:
+            gated_delta_net_cuda<16, KDA><<<grid_dims, block_dims, 0, stream>>>(
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+                n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+            break;
+        case 32:
+            gated_delta_net_cuda<32, KDA><<<grid_dims, block_dims, 0, stream>>>(
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+                n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+            break;
+        case 64: {
+            constexpr int sv = 64;
+            size_t smem = calculate_smem(sv, cc);
+            gated_delta_net_cuda<sv, KDA><<<grid_dims, block_dims, smem, stream>>>(
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+                n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+            break;
+        }
+        case 128: {
+            constexpr int sv = 128;
+            size_t smem = calculate_smem(sv, cc);
+            gated_delta_net_cuda<sv, KDA><<<grid_dims, block_dims, smem, stream>>>(
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+                n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+            break;
+        }
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src_q     = dst->src[0];
+    ggml_tensor * src_k     = dst->src[1];
+    ggml_tensor * src_v     = dst->src[2];
+    ggml_tensor * src_g     = dst->src[3];
+    ggml_tensor * src_beta  = dst->src[4];
+    ggml_tensor * src_state = dst->src[5];
+
+    GGML_TENSOR_LOCALS(int64_t, neq, src_q, ne);
+    GGML_TENSOR_LOCALS(size_t , nbq, src_q, nb);
+    GGML_TENSOR_LOCALS(int64_t, nek, src_k, ne);
+    GGML_TENSOR_LOCALS(size_t , nbk, src_k, nb);
+    GGML_TENSOR_LOCALS(int64_t, nev, src_v, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbv, src_v, nb);
+    GGML_TENSOR_LOCALS(size_t,  nbb, src_beta, nb);
+
+    const int64_t S_v      = nev0;
+    const int64_t H        = nev1;
+    const int64_t n_tokens = nev2;
+    const int64_t n_seqs   = nev3;
+
+    const bool kda = (src_g->ne[0] == S_v);
+
+    GGML_ASSERT(neq1 == nek1);
+    const int64_t neqk1 = neq1;
+
+    const int64_t rq3 = nev3 / neq3;
+
+    const float * q_d = (const float *) src_q->data;
+    const float * k_d = (const float *) src_k->data;
+    const float * v_d = (const float *) src_v->data;
+    const float * g_d = (const float *) src_g->data;
+    const float * b_d = (const float *) src_beta->data;
+
+    const float * s_d   = (const float *) src_state->data;
+    float *       dst_d = (float *) dst->data;
+
+    GGML_ASSERT(ggml_is_contiguous_rows(src_q));
+    GGML_ASSERT(ggml_is_contiguous_rows(src_k));
+    GGML_ASSERT(ggml_is_contiguous_rows(src_v));
+    GGML_ASSERT(ggml_are_same_stride(src_q, src_k));
+    GGML_ASSERT(src_g->ne[0] == 1 || kda);
+    GGML_ASSERT(ggml_is_contiguous(src_g));
+    GGML_ASSERT(ggml_is_contiguous(src_beta));
+    GGML_ASSERT(ggml_is_contiguous(src_state));
+
+    // strides in floats (beta strides used for both g and beta offset computation)
+    const int64_t sq1 = nbq1 / sizeof(float);
+    const int64_t sq2 = nbq2 / sizeof(float);
+    const int64_t sq3 = nbq3 / sizeof(float);
+    const int64_t sv1 = nbv1 / sizeof(float);
+    const int64_t sv2 = nbv2 / sizeof(float);
+    const int64_t sv3 = nbv3 / sizeof(float);
+    const int64_t sb1 = nbb1 / sizeof(float);
+    const int64_t sb2 = nbb2 / sizeof(float);
+    const int64_t sb3 = nbb3 / sizeof(float);
+
+    const float scale = 1.0f / sqrtf((float) S_v);
+
+    cudaStream_t stream = ctx.stream();
+
+    if (kda) {
+        launch_gated_delta_net<true>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d,
+            S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+            sb1, sb2, sb3, neqk1, rq3, scale, stream);
+    } else {
+        launch_gated_delta_net<false>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d,
+            S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+            sb1, sb2, sb3, neqk1, rq3, scale, stream);
+    }
+}

ggml/src/ggml-cuda/gated_delta_net.cuh

@@ -0,0 +1,4 @@
+#include "common.cuh"
+#include "ggml.h"
+
+void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -53,6 +53,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/gated_delta_net.cuh"
 #include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
@@ -204,7 +205,14 @@ static ggml_cuda_device_info ggml_cuda_init() {
     GGML_ASSERT(info.device_count <= GGML_CUDA_MAX_DEVICES);
 
     int64_t total_vram = 0;
-    GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
+    for (int id = 0; id < info.device_count; ++id) {
+        cudaDeviceProp prop;
+        CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
+        total_vram += prop.totalGlobalMem;
+    }
+    GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices (Total VRAM: %zu MiB):\n",
+                  __func__, info.device_count, (size_t)(total_vram / (1024 * 1024)));
+    total_vram = 0;
 
     std::vector<std::pair<int, std::string>> turing_devices_without_mma;
     for (int id = 0; id < info.device_count; ++id) {
@@ -242,6 +250,12 @@ static ggml_cuda_device_info ggml_cuda_init() {
 #else
         info.devices[id].supports_cooperative_launch = false;
 #endif // !(GGML_USE_MUSA)
+
+        // cudaMemGetInfo returns info for the current device
+        size_t free_mem;
+        CUDA_CHECK(cudaSetDevice(id));
+        CUDA_CHECK(cudaMemGetInfo(&free_mem, NULL));
+
 #if defined(GGML_USE_HIP)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
 
@@ -256,22 +270,25 @@ static ggml_cuda_device_info ggml_cuda_init() {
                 info.devices[id].cc += prop.minor * 0x10;
             }
         }
-        GGML_LOG_INFO("  Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d\n",
+        GGML_LOG_INFO("  Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d, VRAM: %zu MiB (%zu MiB free)\n",
                       id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff,
-                      device_vmm ? "yes" : "no", prop.warpSize);
+                      device_vmm ? "yes" : "no", prop.warpSize,
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
 #elif defined(GGML_USE_MUSA)
         // FIXME: Ensure compatibility with varying warp sizes across different MUSA archs.
         info.devices[id].warp_size = 32;
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = GGML_CUDA_CC_OFFSET_MTHREADS + prop.major * 0x100;
         info.devices[id].cc += prop.minor * 0x10;
-        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
-                        id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
+        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB (%zu MiB free)\n",
+                      id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
 #else
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
-        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
-                        id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
+        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB (%zu MiB free)\n",
+                      id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
+                      (size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
         std::string device_name(prop.name);
         if (device_name == "NVIDIA GeForce MX450") {
             turing_devices_without_mma.push_back({ id, device_name });
@@ -2733,6 +2750,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_GATED_LINEAR_ATTN:
             ggml_cuda_op_gated_linear_attn(ctx, dst);
             break;
+        case GGML_OP_GATED_DELTA_NET:
+            ggml_cuda_op_gated_delta_net(ctx, dst);
+            break;
         case GGML_OP_RWKV_WKV7:
             ggml_cuda_op_rwkv_wkv7(ctx, dst);
             break;
@@ -3348,6 +3368,46 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
         return true;
     }
 
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_SSM_CONV && ops.begin()[1] == GGML_OP_UNARY
+     && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_SILU) {
+        const ggml_tensor * ssm_conv = cgraph->nodes[node_idx];
+        const ggml_tensor * silu     = cgraph->nodes[node_idx+1];
+
+        if (ssm_conv->type != GGML_TYPE_F32 || silu->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        return true;
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_UNARY && ops.begin()[1] == GGML_OP_MUL
+     && unary_ops.size() == 1 && (unary_ops.begin()[0] == GGML_UNARY_OP_SILU || unary_ops.begin()[0] == GGML_UNARY_OP_SIGMOID || unary_ops.begin()[0] == GGML_UNARY_OP_SOFTPLUS)) {
+        const ggml_tensor * unary = cgraph->nodes[node_idx];
+        const ggml_tensor * mul   = cgraph->nodes[node_idx+1];
+
+        if (ggml_get_unary_op(unary) != unary_ops.begin()[0]) {
+            return false;
+        }
+
+        if (unary->type != GGML_TYPE_F32 && unary->type != GGML_TYPE_F16) {
+            return false;
+        }
+
+        if (unary->type != mul->type) {
+            return false;
+        }
+
+        const ggml_tensor * other = (mul->src[0] == unary) ? mul->src[1] : mul->src[0];
+        if (other->type != unary->type) {
+            return false;
+        }
+        if (!ggml_is_contiguous_1(other) || !ggml_is_contiguous_1(unary->src[0]) || !ggml_are_same_shape(other, unary)) {
+            return false;
+        }
+
+        return true;
+    }
+
     if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SCALE && ops.begin()[1] == GGML_OP_UNARY && ops.begin()[2] == GGML_OP_SCALE
      && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_TANH) {
         const ggml_tensor *scale  = cgraph->nodes[node_idx];
@@ -3372,6 +3432,69 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
     return false;
 }
 
+// returns whether the write (out) nodes overwrite the read nodes in operation
+static bool ggml_cuda_check_fusion_memory_ranges(ggml_cgraph * cgraph,
+                                                 int           node_idx,
+                                                 int           node_count,
+                                                 int *         out_nodes,
+                                                 int           out_count) {
+    auto nodes_overlap = [&](const ggml_tensor * a, const ggml_tensor * b) {
+        const int64_t a_start = (int64_t) a->data;
+        const int64_t a_end   = a_start + ggml_nbytes(a);
+
+        const int64_t b_start = (int64_t) b->data;
+        const int64_t b_end   = b_start + ggml_nbytes(b);
+
+        if ((b_start <= a_start && a_start < b_end) || (a_start <= b_start && b_start < a_end)) {
+            return true;
+        }
+
+        return false;
+    };
+
+    bool is_ok = true;
+    // for nrows=1, all fusion operations correctly read the src before writing dst or do it elementwise, so we should be ok
+    if (ggml_nrows(cgraph->nodes[node_idx]) == 1) {
+        return true;
+    }
+
+    for (int i = 0; i < out_count; ++i) {
+        const ggml_tensor * dst = cgraph->nodes[out_nodes[i]];
+
+        for (int j = node_idx; j < node_idx + node_count; ++j) {
+            // Loop over all srcs of all nodes in the fusion. If the src overlaps
+            // the destination and the src is not an intermediate node that's being
+            // elided, then disable fusion.
+
+            for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+                const ggml_tensor * src = cgraph->nodes[j]->src[src_idx];
+
+                if (!src || src->op == GGML_OP_NONE) {
+                    continue;
+                }
+
+                if (nodes_overlap(dst, src)) {
+                    bool found = false;
+
+                    for (int k = node_idx; k < j; ++k) {
+                        if (cgraph->nodes[k] == src) {
+                            found = true;
+                            break;
+                        }
+                    }
+
+                    if (!found) {
+                        is_ok = false;
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
+    return is_ok;
+}
+
 static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, const bool use_cuda_graph, const bool cuda_graph_update_required, const void * graph_key) {
     bool graph_evaluated_or_captured = false;
 
@@ -3568,7 +3691,8 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                                 out_nodes[1] = i + ops.size() - 1;
 
                                 if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
-                                        ggml_cuda_should_use_topk_moe(node, logits, weights, ids)) {
+                                    ggml_cuda_should_use_topk_moe(node, logits, weights, ids) &&
+                                    ggml_cuda_check_fusion_memory_ranges(cgraph, i, ops.size(), out_nodes, 2)) {
                                     ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
                                     i += ops.size() - 1;
                                     continue;
@@ -3583,7 +3707,8 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
 
                                 int out_nodes[2] = { i + 1, i + 5 };
                                 if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
-                                        ggml_cuda_should_use_topk_moe(softmax, logits, weights, ids)) {
+                                    ggml_cuda_should_use_topk_moe(softmax, logits, weights, ids) &&
+                                    ggml_cuda_check_fusion_memory_ranges(cgraph, i, ops.size(), out_nodes, 2)) {
                                     ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
                                     i += ops.size() - 1;
                                     continue;
@@ -3836,6 +3961,20 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                         continue;
                     }
 
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SSM_CONV, GGML_OP_UNARY }, { GGML_UNARY_OP_SILU })) {
+                        ggml_cuda_op_ssm_conv(*cuda_ctx, node, cgraph->nodes[i+1]);
+                        i++;
+                        continue;
+                    }
+
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SILU }) ||
+                        ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SIGMOID }) ||
+                        ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SOFTPLUS })) {
+                        ggml_cuda_op_unary_mul(*cuda_ctx, node, cgraph->nodes[i+1]);
+                        i++;
+                        continue;
+                    }
+
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SCALE, GGML_OP_UNARY, GGML_OP_SCALE }, { GGML_UNARY_OP_TANH })) {
                         i += 2;
                         ggml_cuda_op_softcap(*cuda_ctx, cgraph->nodes[i], node);
@@ -4855,6 +4994,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_GATED_LINEAR_ATTN:
         case GGML_OP_RWKV_WKV7:
             return true;
+        case GGML_OP_GATED_DELTA_NET:
+            //TODO: enable once MUSA compiler is solved https://github.com/ggml-org/llama.cpp/pull/19504#issuecomment-4018634327
+#ifdef GGML_USE_MUSA
+            return false;
+#else
+            return true;
+#endif // GGML_USE_MUSA
         case GGML_OP_FLASH_ATTN_EXT:
             return ggml_cuda_flash_attn_ext_supported(dev_ctx->device, op);
         case GGML_OP_CROSS_ENTROPY_LOSS:

ggml/src/ggml-cuda/ssm-conv.cu

@@ -1,6 +1,7 @@
 #include "ssm-conv.cuh"
+#include "unary.cuh"
 
-template <size_t split_d_inner, size_t d_conv>
+template <bool apply_silu, size_t split_d_inner, size_t d_conv>
 static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1,
                                     const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1,
                                     float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2,
@@ -41,11 +42,11 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
         for (size_t j = 0; j < d_conv; j++) {
             sumf += x[(i + j) % d_conv] * w[j];
         }
-        y_block[i * stride_y + tid] = sumf;
+        y_block[i * stride_y + tid] = apply_silu ? ggml_cuda_op_silu_single(sumf) : sumf;
     }
 }
 
-template <size_t split_d_inner, size_t d_conv, int64_t split_n_t>
+template <bool apply_silu, size_t split_d_inner, size_t d_conv, int64_t split_n_t>
 static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1,
                                                const int src0_nb0, const int src0_nb1, const int src0_nb2,
                                                const int src1_nb1, float * __restrict__ dst, const int dst_nb0,
@@ -65,36 +66,49 @@ static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0,
     const int stride_w = src1_nb1 / sizeof(float);
     const int stride_y = dst_nb1 / sizeof(float);
 
-    float x[d_conv] = { 0.0f };
-    float w[d_conv] = { 0.0f };
+    const int64_t local_n_t = min(split_n_t, n_t - bidz * split_n_t);
+    const int     n_cols    = d_conv - 1 + split_n_t;
 
+    extern __shared__ float smem[];
+
+    constexpr int load_cols   = d_conv - 1 + split_n_t;
+    constexpr int total_elems = split_d_inner * load_cols;
+    int row = tid / load_cols;
+    int col = tid % load_cols;
+#pragma unroll
+    for (int idx = 0; idx < total_elems; idx += split_d_inner) {
+        if (row < (int)split_d_inner) {
+            smem[row * n_cols + col] = x_block[row * stride_x + col];
+        }
+
+        col += split_d_inner;
+        row += col / load_cols;
+        col  = col % load_cols;
+        if (idx >= total_elems - tid - split_d_inner) {
+            break;
+        }
+    }
+    __syncthreads();
+
+    // Load weights into registers (done once, small)
+    float w[d_conv] = { 0.0f };
 #pragma unroll
     for (size_t j = 0; j < d_conv; j++) {
         w[j] = w_block[tid * stride_w + j];
     }
 
+    // Compute from shared memory
+    for (int64_t i = 0; i < local_n_t; i++) {
+        float sumf = 0.0f;
 #pragma unroll
-    for (int64_t i = 0; i < split_n_t; i++) {
-        if (bidz * split_n_t + i < n_t) {
-            float sumf = 0.0f;
-
-            if (i == 0) {
-                for (size_t j = 0; j < d_conv; j++) {
-                    x[j] = x_block[tid * stride_x + j];
-                }
-            } else {
-                x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1];
-            }
-
-#pragma unroll
-            for (size_t j = 0; j < d_conv; j++) {
-                sumf += x[(i + j) % d_conv] * w[j];
-            }
-            y_block[i * stride_y + tid] = sumf;
+        for (size_t j = 0; j < d_conv; j++) {
+            sumf += smem[tid * n_cols + i + j] * w[j];
         }
+        y_block[i * stride_y + tid] = apply_silu ? ggml_cuda_op_silu_single(sumf) : sumf;
     }
 }
 
+template <bool apply_silu>
 static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1,
                               const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1,
                               const int dst_nb2, const int64_t nc, const int64_t nr, const int64_t n_t,
@@ -106,12 +120,13 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
         constexpr int kNC = decltype(NC)::value;
         if (n_t <= 32) {
             const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
-            ssm_conv_f32<threads, kNC><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
+            ssm_conv_f32<apply_silu, threads, kNC><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
                                                                        dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
             const int64_t split_n_t = 32;
             dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
-            ssm_conv_long_token_f32<threads, kNC, split_n_t><<<blocks, threads, 0, stream>>>(
+            const size_t  smem_size = threads * (kNC - 1 + split_n_t) * sizeof(float);
+            ssm_conv_long_token_f32<apply_silu, threads, kNC, split_n_t><<<blocks, threads, smem_size, stream>>>(
                 src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         }
     };
@@ -124,27 +139,36 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
     }
 }
 
-void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * silu_dst) {
     const struct ggml_tensor * src0 = dst->src[0];  // conv_x
     const struct ggml_tensor * src1 = dst->src[1];  // conv1d.weight
+    const bool fuse_silu = silu_dst != nullptr;
+
+    // When fusing, write to silu_dst (the node downstream references).
+    const struct ggml_tensor * out = fuse_silu ? silu_dst : dst;
 
     const int64_t nc  = src1->ne[0];                // d_conv
     const int64_t nr  = src0->ne[1];                // d_inner
-    const int64_t n_t = dst->ne[1];                 // tokens per sequence
-    const int64_t n_s = dst->ne[2];                 // number of sequences in the batch
+    const int64_t n_t = out->ne[1];                 // tokens per sequence
+    const int64_t n_s = out->ne[2];                 // number of sequences in the batch
 
-    GGML_ASSERT(dst->ne[0] == nr);
+    GGML_ASSERT(out->ne[0] == nr);
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(src1->nb[0] == sizeof(float));
     GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
 
     const float * src0_d = (const float *) src0->data;
     const float * src1_d = (const float *) src1->data;
-    float *       dst_d  = (float *) dst->data;
+    float *       dst_d  = (float *) out->data;
     cudaStream_t  stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    ssm_conv_f32_cuda(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, dst->nb[0], dst->nb[1],
-                      dst->nb[2], nc, nr, n_t, n_s, stream);
+    GGML_ASSERT(out->type == GGML_TYPE_F32);
+    if (fuse_silu) {
+        ssm_conv_f32_cuda<true>(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
+                          out->nb[2], nc, nr, n_t, n_s, stream);
+    } else {
+        ssm_conv_f32_cuda<false>(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
+                          out->nb[2], nc, nr, n_t, n_s, stream);
+    }
 }

ggml/src/ggml-cuda/ssm-conv.cuh

@@ -1,3 +1,3 @@
 #include "common.cuh"
 
-void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * silu_dst = nullptr);

ggml/src/ggml-cuda/topk-moe.cu

@@ -119,6 +119,18 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         }
     }
 
+    // Sanitize NaN to -FLT_MAX so the iterative argmax produces unique expert IDs.
+    // NaN comparisons always return false, which would cause the same expert to be
+    // selected repeatedly. -FLT_MAX compares normally and is still excluded by the
+    // -INFINITY sentinel used after each selection round.
+    // More relevant for the cuBLAS path. See https://github.com/ggml-org/llama.cpp/issues/19659
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        if (__isnanf(wt[i])) {
+            wt[i] = -FLT_MAX;
+        }
+    }
+
     // selection_wt is only needed when bias is present (selection uses wt + bias)
     // when no bias, we use wt directly for both selection and weight values
     float selection_wt[has_bias ? experts_per_thread : 1];

ggml/src/ggml-cuda/unary.cu

@@ -560,3 +560,58 @@ void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
         leaky_relu_cuda((const float *)src0_d, (float *)dst_d, ggml_nelements(src0), negative_slope, stream);
     }
 }
+
+/* fused unary + mul */
+
+template <float (*op)(float)>
+static void ggml_cuda_op_unary_mul_impl(ggml_backend_cuda_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node) {
+    // unary_node: UNARY op applied to unary_node->src[0]
+    // mul_node:   MUL(a, b) where one of a/b is unary_node
+    // Output goes to mul_node->data
+
+    const ggml_tensor * unary_src = unary_node->src[0];  // input to the unary op
+    const ggml_tensor * other_src = (mul_node->src[0] == unary_node) ? mul_node->src[1] : mul_node->src[0];
+
+    GGML_ASSERT(ggml_is_contiguous_1(unary_src));
+    GGML_ASSERT(unary_src->nb[0] == ggml_element_size(unary_src));
+    GGML_ASSERT(ggml_is_contiguous_1(other_src));
+    GGML_ASSERT(other_src->nb[0] == ggml_element_size(other_src));
+    GGML_ASSERT(ggml_are_same_shape(unary_src, other_src));
+
+    GGML_ASSERT(unary_src->type == GGML_TYPE_F32 || unary_src->type == GGML_TYPE_F16);
+    GGML_ASSERT(unary_src->type == other_src->type);
+    GGML_ASSERT(unary_src->type == mul_node->type);
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t k  = ggml_nelements(mul_node);
+    const int64_t nc = unary_src->ne[0];
+    const int64_t unary_stride = unary_src->nb[1];
+    const int64_t other_stride = other_src->nb[1];
+
+    if (unary_src->type == GGML_TYPE_F16) {
+        unary_gated_cuda<op>((const half *) unary_src->data, (const half *) other_src->data,
+                             (half *) mul_node->data, k, nc,
+                             unary_stride / sizeof(half), other_stride / sizeof(half), stream);
+    } else {
+        unary_gated_cuda<op>((const float *) unary_src->data, (const float *) other_src->data,
+                             (float *) mul_node->data, k, nc,
+                             unary_stride / sizeof(float), other_stride / sizeof(float), stream);
+    }
+}
+
+void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node) {
+    switch (ggml_get_unary_op(unary_node)) {
+        case GGML_UNARY_OP_SILU:
+            ggml_cuda_op_unary_mul_impl<op_silu>(ctx, unary_node, mul_node);
+            break;
+        case GGML_UNARY_OP_SIGMOID:
+            ggml_cuda_op_unary_mul_impl<op_sigmoid>(ctx, unary_node, mul_node);
+            break;
+        case GGML_UNARY_OP_SOFTPLUS:
+            ggml_cuda_op_unary_mul_impl<op_softplus>(ctx, unary_node, mul_node);
+            break;
+        default:
+            GGML_ABORT("Unsupported unary op for fused unary+mul");
+    }
+}

ggml/src/ggml-cuda/unary.cuh

@@ -89,6 +89,8 @@ void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
 void ggml_cuda_op_xielu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node);
+
 __device__ __forceinline__ float ggml_cuda_op_silu_single(float x) {
     return x / (1.0f + expf(-x));
 }

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -1865,15 +1865,26 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
     const struct ggml_tensor * src1 = op->src[1];
     const struct ggml_tensor * dst  = op;
 
-    if (src0->type != GGML_TYPE_F32) {
-        return false;
-    }
-    if (src1->type != GGML_TYPE_F32) {
-        return false;
-    }
-    if (dst->type != GGML_TYPE_F32) {
+    if (src0->type == GGML_TYPE_F32) {
+        if (src1->type != GGML_TYPE_F32) {
+            return false;
+        }
+        if (dst->type != GGML_TYPE_F32) {
+            return false;
+        }
+    }
+    else if (src0->type == GGML_TYPE_F16) {
+        if (src1->type != GGML_TYPE_F16) {
+            return false;
+        }
+        if (dst->type != GGML_TYPE_F16) {
+            return false;
+        }
+    }
+    else {
         return false;
     }
+
     if (!ggml_are_same_shape(src0, dst)) {
         return false;
     }
@@ -2141,6 +2152,44 @@ static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess
     return true;
 }
 
+static bool ggml_hexagon_supported_ssm_conv(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+    const struct ggml_tensor * dst  = op;
+
+    // Only support FP32 for now
+    if (src0->type != GGML_TYPE_F32 || src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    // Check IO tensor shapes and dims
+    if (src0->ne[3] != 1 || src1->ne[2] != 1 || src1->ne[3] != 1 || dst->ne[3] != 1) {
+        return false; // src0 should be effectively 3D
+    }
+
+    const int d_conv = src1->ne[0];
+    const int d_inner = src0->ne[1];
+    const int n_t = dst->ne[1];
+    const int n_s = dst->ne[2];
+
+    if (src0->ne[0] != d_conv - 1 + n_t || src0->ne[1] != d_inner || src0->ne[2] != n_s) {
+        return false;
+    }
+    if (src1->ne[0] != d_conv || src1->ne[1] != d_inner) {
+        return false;
+    }
+    if (dst->ne[0] != d_inner || dst->ne[1] != n_t || dst->ne[2] != n_s) {
+        return false;
+    }
+
+    // TODO: add support for non-contiguous tensors
+    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    return true;
+}
+
 enum dspqbuf_type {
     DSPQBUF_TYPE_DSP_WRITE_CPU_READ = 0,
     DSPQBUF_TYPE_CPU_WRITE_DSP_READ,
@@ -2457,6 +2506,17 @@ static inline size_t init_flash_attn_ext_req(htp_general_req * req, dspqueue_buf
     return n_bufs;
 }
 
+static inline size_t init_ssm_conv_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
+    req->op = HTP_OP_SSM_CONV;
+
+    size_t n_bufs = 0;
+    n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
+    n_bufs += htp_req_buff_init(&req->src1, &bufs[n_bufs], t->src[1], DSPQBUF_TYPE_CONSTANT);
+    n_bufs += htp_req_buff_init(&req->dst,  &bufs[n_bufs], t,         DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
+
+    return n_bufs;
+}
+
 static const char * ggml_backend_hexagon_name(ggml_backend_t backend) {
     auto sess = static_cast<ggml_hexagon_session *>(backend->context);
     return sess->name.c_str();
@@ -2595,6 +2655,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                 ggml_hexagon_dispatch_op<init_argsort_req>(sess, node, flags);
                 break;
 
+            case GGML_OP_SSM_CONV:
+                ggml_hexagon_dispatch_op<init_ssm_conv_req>(sess, node, flags);
+                break;
+
             default:
                 GGML_ABORT("\nggml-hex: graph-compute %s is not supported\n", ggml_op_desc(node));
         }
@@ -3013,6 +3077,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_argsort(sess, op);
             break;
 
+        case GGML_OP_SSM_CONV:
+            supp = ggml_hexagon_supported_ssm_conv(sess, op);
+            break;
+
         default:
             break;
     }

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -31,6 +31,7 @@ add_library(${HTP_LIB} SHARED
     get-rows-ops.c
     cpy-ops.c
     argsort-ops.c
+    ssm-conv.c
 )
 
 target_compile_definitions(${HTP_LIB} PRIVATE

ggml/src/ggml-hexagon/htp/act-ops.c

@@ -693,8 +693,8 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t n_threads  = octx->n_threads;
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     size_t src0_row_size = src0->nb[1];
     size_t src1_row_size = src1->nb[1]; // zero bytes if src1 is not used
@@ -748,13 +748,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
-    uint32_t n_jobs = MIN(n_threads, src0_nrows);
-
     // Prepare context
     struct htp_act_context actx;
     actx.octx = octx;
 
-    actx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+    actx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
 
     actx.src0_row_size = src0_row_size;
     actx.src1_row_size = src1_row_size;
@@ -794,7 +792,7 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
     actx.data_src1 = data_src1;
     actx.data_dst  = (uint8_t *) dst->data;
 
-    worker_pool_run_func(octx->ctx->worker_pool, act_op_func, &actx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, act_op_func, &actx, n_threads);
     return HTP_STATUS_OK;
 }
 

ggml/src/ggml-hexagon/htp/argsort-ops.c

@@ -241,6 +241,9 @@ int op_argsort(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
+    const uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
+    const uint32_t n_threads = MIN(total_rows, octx->n_threads);
+
     // Allocate scratchpad
     // We need 1 row of float + 1 row of int32 per thread.
     uint32_t ne00 = octx->src0.ne[0];
@@ -251,7 +254,7 @@ int op_argsort(struct htp_ops_context * octx) {
     // Make sure we round up to 256 for alignment requirements
     spad_per_thread = hex_round_up(spad_per_thread, 256);
 
-    size_t total_spad_size = spad_per_thread * octx->n_threads;
+    size_t total_spad_size = spad_per_thread * n_threads;
 
     if (octx->ctx->vtcm_size < total_spad_size) {
         FARF(ERROR, "argsort: VTCM size too small. Needed %zu, have %zu", total_spad_size, octx->ctx->vtcm_size);
@@ -267,15 +270,12 @@ int op_argsort(struct htp_ops_context * octx) {
          octx->dst.ne[0], octx->dst.ne[1], octx->dst.ne[2], octx->dst.ne[3],
          octx->src0.data, octx->dst.data);
 
-    uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
-    uint32_t n_jobs = MIN(total_rows, octx->n_threads);
-
     struct htp_argsort_context actx;
     actx.octx = octx;
-    actx.nrows_per_thread = (total_rows + n_jobs - 1) / n_jobs;
+    actx.nrows_per_thread = (total_rows + n_threads - 1) / n_threads;
 
     // Run jobs
-    worker_pool_run_func(octx->ctx->worker_pool, htp_argsort_f32, &actx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, htp_argsort_f32, &actx, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/binary-ops.c

@@ -95,43 +95,87 @@ static inline uint32_t calc_block_size(struct htp_binary_context * bctx, uint32_
 }
 
 // Macro for scalar op switch
-#define COMPUTE_SCALAR_OP(DST, SRC, VAL, N) \
-    switch (octx->op) { \
-        case HTP_OP_ADD: hvx_add_scalar_f32_aa(DST, SRC, VAL, N); break; \
-        case HTP_OP_SUB: hvx_sub_scalar_f32_aa(DST, SRC, VAL, N); break; \
-        case HTP_OP_MUL: hvx_mul_scalar_f32_aa(DST, SRC, VAL, N); break; \
-        case HTP_OP_DIV: hvx_mul_scalar_f32_aa(DST, SRC, 1.0f / (VAL), N); break; \
-        default: break; \
+#define COMPUTE_SCALAR_OP(DST, SRC, VAL, TYPE, N) \
+    if(TYPE == HTP_TYPE_F32) { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
+            case HTP_OP_SUB: hvx_sub_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
+            case HTP_OP_MUL: hvx_mul_scalar_f32_aa(DST, SRC, *(float *)VAL, N); break; \
+            case HTP_OP_DIV: hvx_mul_scalar_f32_aa(DST, SRC, 1.0f / (*(float *)VAL), N); break; \
+            default: break; \
+        } \
+    } \
+    else { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
+            case HTP_OP_SUB: hvx_sub_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
+            case HTP_OP_MUL: hvx_mul_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
+            case HTP_OP_DIV: hvx_div_scalar_f16_aa(DST, SRC, *(_Float16 *)VAL, N); break; \
+            default: break; \
+        } \
     }
 
 // Macro for vector op switch (All Aligned)
-#define COMPUTE_VECTOR_OP_AAA(DST, SRC0, SRC1, N) \
-    switch (octx->op) { \
-        case HTP_OP_ADD: hvx_add_f32_aaa(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_SUB: hvx_sub_f32_aaa(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_MUL: hvx_mul_f32_aaa(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_DIV: hvx_div_f32_aaa(DST, SRC0, SRC1, N); break; \
-        default: break; \
+#define COMPUTE_VECTOR_OP_AAA(DST, SRC0, SRC1, TYPE, N) \
+    if(TYPE == HTP_TYPE_F32) { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f32_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f32_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f32_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f32_aaa(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
+    } \
+    else { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f16_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f16_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f16_aaa(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f16_aaa(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
     }
 
 // Macro for vector op switch (Dst Aligned, Src0 Aligned, Src1 Unaligned)
-#define COMPUTE_VECTOR_OP_AAU(DST, SRC0, SRC1, N) \
-    switch (octx->op) { \
-        case HTP_OP_ADD: hvx_add_f32_aau(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_SUB: hvx_sub_f32_aau(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_MUL: hvx_mul_f32_aau(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_DIV: hvx_div_f32_aau(DST, SRC0, SRC1, N); break; \
-        default: break; \
+#define COMPUTE_VECTOR_OP_AAU(DST, SRC0, SRC1, TYPE, N) \
+    if(TYPE == HTP_TYPE_F32) { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f32_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f32_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f32_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f32_aau(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
+    } \
+    else { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f16_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f16_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f16_aau(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f16_aau(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
     }
 
 // Macro for vector op switch (All Unaligned - generic loop used in element repeat)
-#define COMPUTE_VECTOR_OP_UUU(DST, SRC0, SRC1, N) \
-    switch (octx->op) { \
-        case HTP_OP_ADD: hvx_add_f32_uuu(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_SUB: hvx_sub_f32_uuu(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_MUL: hvx_mul_f32_uuu(DST, SRC0, SRC1, N); break; \
-        case HTP_OP_DIV: hvx_div_f32_uuu(DST, SRC0, SRC1, N); break; \
-        default: break; \
+#define COMPUTE_VECTOR_OP_UUU(DST, SRC0, SRC1, TYPE, N) \
+    if(TYPE == HTP_TYPE_F32) { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f32_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f32_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f32_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f32_uuu(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
+    } \
+    else { \
+        switch (octx->op) { \
+            case HTP_OP_ADD: hvx_add_f16_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_SUB: hvx_sub_f16_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_MUL: hvx_mul_f16_uuu(DST, SRC0, SRC1, N); break; \
+            case HTP_OP_DIV: hvx_div_f16_uuu(DST, SRC0, SRC1, N); break; \
+            default: break; \
+        } \
     }
 
 // 1. Scalar src1 (ne10 == 1)
@@ -140,6 +184,8 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -170,7 +216,7 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -199,13 +245,12 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
         for (uint32_t r = 0; r < current_block_size; r++) {
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_dst  = d_spad + r * bctx->dst_row_size_aligned;
-            float val = *(float *)src1_ptr;
+            COMPUTE_SCALAR_OP(r_dst, r_src0, src1_ptr, src0_type, ne00);
             src1_ptr += s1_stride;
-            COMPUTE_SCALAR_OP(r_dst, r_src0, val, ne00);
         }
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -216,7 +261,7 @@ static void binary_job_scalar(unsigned int nth, unsigned int ith, void * data) {
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
 
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
         }
         ir += current_block_size;
@@ -230,6 +275,8 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -268,8 +315,8 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
-        dma_queue_push(q, dma_make_ptr(s1_spad, src1_base), bctx->src1_row_size_aligned, bctx->src1_dma_stride, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
+        dma_queue_push(q, dma_make_ptr(s1_spad, src1_base), bctx->src1_row_size_aligned, bctx->src1_dma_stride, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -284,7 +331,7 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_src1 = s1_spad + r * bctx->src1_row_size_aligned;
             uint8_t * r_dst  = d_spad  + r * bctx->dst_row_size_aligned;
-            COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, ne00);
+            COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, src0_type, ne00);
         }
 
         uint32_t i03, i02, i01, rem;
@@ -293,7 +340,7 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
         i02 = fastdiv(rem, &bctx->dim1_div);
         i01 = rem - i02 * ne01;
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -310,8 +357,8 @@ static void binary_job_vector_same_shape(unsigned int nth, unsigned int ith, voi
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
              uint8_t * s1_next = (uint8_t *)src1->data + p13 * nb13 + p12 * nb12 + p11 * nb11;
 
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
-             dma_queue_push(q, dma_make_ptr(s1_spad, s1_next), bctx->src1_row_size_aligned, bctx->src1_dma_stride, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
+             dma_queue_push(q, dma_make_ptr(s1_spad, s1_next), bctx->src1_row_size_aligned, bctx->src1_dma_stride, row_size_bytes, next_block_size);
 
              ir_prefetch += next_block_size;
         }
@@ -326,6 +373,8 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -359,7 +408,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -373,7 +422,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
             uint8_t * r_src0 = s0_spad + r * bctx->src0_row_size_aligned;
             uint8_t * r_src1 = (uint8_t *)s1_ptr; // Constant
             uint8_t * r_dst  = d_spad + r * bctx->dst_row_size_aligned;
-            COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, ne00);
+            COMPUTE_VECTOR_OP_AAA(r_dst, r_src0, r_src1, src0_type, ne00);
         }
 
         uint32_t i03, i02, i01, rem;
@@ -382,7 +431,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
         i02 = fastdiv(rem, &bctx->dim1_div);
         i01 = rem - i02 * ne01;
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -392,7 +441,7 @@ static void binary_job_vector_row_broadcast(unsigned int nth, unsigned int ith,
              p02 = fastdiv(prem, &bctx->dim1_div);
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
         }
         ir += current_block_size;
@@ -406,6 +455,8 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t row_size_bytes = (src0_type == HTP_TYPE_F32) ? ne00 * sizeof(float) : ne00 * sizeof(_Float16);
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -435,7 +486,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -462,11 +513,11 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
             uint8_t * r_dst  = d_spad + r * bctx->dst_row_size_aligned;
 
             // Read src1 from DDR (unaligned)
-            COMPUTE_VECTOR_OP_AAU(r_dst, r_src0, r_src1, ne00);
+            COMPUTE_VECTOR_OP_AAU(r_dst, r_src0, r_src1, src0_type, ne00);
         }
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -476,7 +527,7 @@ static void binary_job_vector_complex(unsigned int nth, unsigned int ith, void *
              p02 = fastdiv(prem, &bctx->dim1_div);
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
         }
         ir += current_block_size;
@@ -490,6 +541,9 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
     struct htp_ops_context * octx = bctx->octx;
     htp_binary_preamble;
 
+    const uint32_t src0_type = octx->src0.type;
+    const uint32_t elem_size_bytes = (src0_type == HTP_TYPE_F32) ? sizeof(float) : sizeof(_Float16);
+    const uint32_t row_size_bytes = ne00 * elem_size_bytes;;
     const uint32_t total_rows = ne01 * ne02 * ne03;
     const uint32_t start_row = bctx->nrows_per_thread * ith;
     const uint32_t end_row   = MIN(start_row + bctx->nrows_per_thread, total_rows);
@@ -519,7 +573,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
         uint8_t * d_spad  = dst_spad_base  + spad_idx * dst_spad_half;
 
         dma_queue_push_vtcm_to_ddr(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, 0);
-        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(s0_spad, src0_curr), bctx->src0_row_size_aligned, nb01, row_size_bytes, current_block_size);
         ir_prefetch += current_block_size;
         spad_idx ^= 1;
     }
@@ -549,12 +603,12 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
             for (uint32_t c = 0; c < ne00; c += ne10) {
                 uint32_t len = MIN(ne10, ne00 - c);
                 // Use UUU for speed and simplicity
-                COMPUTE_VECTOR_OP_UUU(r_dst + c * sizeof(float), r_src0 + c * sizeof(float), r_src1_row, len);
+                COMPUTE_VECTOR_OP_UUU(r_dst + c * elem_size_bytes, r_src0 + c * elem_size_bytes, r_src1_row, src0_type, len);
             }
         }
 
         uint8_t * dst_curr = (uint8_t *)dst->data + i03 * nb3 + i02 * nb2 + i01 * nb1;
-        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, ne00 * sizeof(float), current_block_size);
+        dma_queue_push(q, dma_make_ptr(dst_curr, d_spad), nb1, bctx->dst_row_size_aligned, row_size_bytes, current_block_size);
 
         if (ir_prefetch < end_row) {
              uint32_t next_block_size = calc_block_size(bctx, ir_prefetch, end_row, ne01, ne02);
@@ -564,7 +618,7 @@ static void binary_job_element_repeat(unsigned int nth, unsigned int ith, void *
              p02 = fastdiv(prem, &bctx->dim1_div);
              p01 = prem - p02 * ne01;
              uint8_t * s0_next = (uint8_t *)src0->data + p03 * nb03 + p02 * nb02 + p01 * nb01;
-             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, ne00 * sizeof(float), next_block_size);
+             dma_queue_push(q, dma_make_ptr(s0_spad, s0_next), bctx->src0_row_size_aligned, nb01, row_size_bytes, next_block_size);
              ir_prefetch += next_block_size;
         }
         ir += current_block_size;
@@ -672,18 +726,20 @@ static void binary_job_add_id(unsigned int nth, unsigned int ith, void * data) {
     dma_queue_flush(q);
 }
 
-static int execute_op_binary_f32(struct htp_ops_context * octx) {
+static int execute_op_binary(struct htp_ops_context * octx) {
     const struct htp_tensor * src0 = &octx->src0;
     const struct htp_tensor * src1 = &octx->src1;
     struct htp_tensor *       dst  = &octx->dst;
 
-    const uint32_t n_threads  = octx->n_threads;
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     // Use packed row sizes for VTCM allocation
-    const size_t src0_row_size = src0->ne[0] * sizeof(float);
-    const size_t src1_row_size = src1->ne[0] * sizeof(float);
-    const size_t dst_row_size  = dst->ne[0] * sizeof(float);
+    const uint32_t src0_type = octx->src0.type;
+    const size_t elem_size = (src0_type == HTP_TYPE_F32) ? sizeof(float) : sizeof(_Float16);
+    const size_t src0_row_size = src0->ne[0] * elem_size;
+    const size_t src1_row_size = src1->ne[0] * elem_size;
+    const size_t dst_row_size  = dst->ne[0] * elem_size;
 
     // Align to VLEN
     const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
@@ -694,7 +750,7 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
     bool is_scalar = !is_add_id && (src1->ne[0] == 1);
 
     // Determine which kernel we will use to alloc memory and dispatch
-    bool use_vector_same = !is_add_id && !is_scalar && src1->ne[0] == src0->ne[0] &&
+    bool use_vector_same = !is_add_id && !is_scalar && ((src0->nb[1] % VLEN) == 0) && (src1->ne[0] == src0->ne[0]) &&
                (src1->ne[1] == src0->ne[1] || src1->ne[1] == 1) &&
                (src1->ne[2] == src0->ne[2] || src1->ne[2] == 1) &&
                (src1->ne[3] == src0->ne[3] || src1->ne[3] == 1);
@@ -726,7 +782,7 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
     }
 
     if (rows_per_buffer < 1) {
-         FARF(ERROR, "binary-f32: VTCM too small\n");
+         FARF(ERROR, "binary: VTCM too small\n");
          return HTP_STATUS_VTCM_TOO_SMALL;
     }
 
@@ -761,16 +817,14 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
-    uint32_t n_jobs = MIN(n_threads, src0_nrows);
-
     dma_queue * q = octx->ctx->dma[0];
     if (is_row_bcast) {
-        dma_queue_push(q, dma_make_ptr(octx->src1_spad.data, (const void *) src1->data), src1_row_size_aligned, 0, src1->ne[0] * sizeof(float), 1);
+        dma_queue_push(q, dma_make_ptr(octx->src1_spad.data, (const void *) src1->data), src1_row_size_aligned, 0, src1->ne[0] * elem_size, 1);
     }
 
     struct htp_binary_context bctx;
     bctx.octx = octx;
-    bctx.nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+    bctx.nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
     bctx.block_max = rows_per_buffer;
     bctx.src0_row_size_aligned = src0_row_size_aligned;
     bctx.src1_row_size_aligned = src1_row_size_aligned;
@@ -814,14 +868,24 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
         dma_queue_pop(q);
     }
 
-    worker_pool_run_func(octx->ctx->worker_pool, worker_func, &bctx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, worker_func, &bctx, n_threads);
 
     return HTP_STATUS_OK;
 }
 
 int op_binary(struct htp_ops_context * octx) {
-    if (octx->src0.type == HTP_TYPE_F32) {
-        return execute_op_binary_f32(octx);
+
+    // Does not support permutations of src1
+    const struct htp_tensor * src1 = &octx->src1;
+    if (src1->nb[1] < src1->nb[0]) {
+        return HTP_STATUS_NO_SUPPORT;
     }
+
+    const uint32_t src0_type = octx->src0.type;
+    if ((src0_type == HTP_TYPE_F32) || (src0_type == HTP_TYPE_F16)) {
+        return execute_op_binary(octx);
+    }
+
     return HTP_STATUS_NO_SUPPORT;
 }
+

ggml/src/ggml-hexagon/htp/cpy-ops.c

@@ -202,6 +202,8 @@ static void cpy_work_func(unsigned int n, unsigned int i, void *data) {
 int op_cpy(struct htp_ops_context * octx) {
     cpy_preamble;
 
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
     struct htp_copy_context ct;
     ct.octx = octx;
 
@@ -227,8 +229,7 @@ int op_cpy(struct htp_ops_context * octx) {
     const bool transposed = (nb00 > nb01) || (nb0 > nb1);
     const bool sameshape  = !transposed && (ne00 == ne0 && ne01 == ne1 && ne02 == ne2 && ne03 == ne3);
 
-    const uint32_t n_jobs = MIN(nr, octx->n_threads);
-    ct.src0_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+    ct.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 
     if (sametype && sameshape) {
         ct.copy = cpy_thread_sametype_sameshape;
@@ -245,7 +246,7 @@ int op_cpy(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    worker_pool_run_func(octx->ctx->worker_pool, cpy_work_func, &ct, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, cpy_work_func, &ct, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/get-rows-ops.c

@@ -82,6 +82,8 @@ static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
 int op_get_rows(struct htp_ops_context * octx) {
     get_rows_preamble;
 
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
     if (octx->src0.type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
@@ -103,9 +105,8 @@ int op_get_rows(struct htp_ops_context * octx) {
     grctx.get_rows_div_ne10      = init_fastdiv_values(octx->src1.ne[0]);
     grctx.get_rows_div_ne10_ne11 = init_fastdiv_values(octx->src1.ne[0] * octx->src1.ne[1]);
 
-    const uint32_t n_jobs = MIN(nr, octx->n_threads);
-    grctx.src1_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+    grctx.src1_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 
-    worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32, &grctx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32, &grctx, n_threads);
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/htp-msg.h

@@ -68,6 +68,7 @@ enum htp_op {
     HTP_OP_SQR,
     HTP_OP_SQRT,
     HTP_OP_SUM_ROWS,
+    HTP_OP_SSM_CONV,
     INVALID
 };
 

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -41,9 +41,6 @@ struct htp_ops_context {
     worker_pool_context_t * wpool;      // worker pool
     uint32_t                n_threads;  // num threads
 
-    uint32_t src0_nrows_per_thread;
-    uint32_t src1_nrows_per_thread;
-
     uint32_t flags;
 };
 
@@ -61,5 +58,6 @@ int op_set_rows(struct htp_ops_context * octx);
 int op_get_rows(struct htp_ops_context * octx);
 int op_cpy(struct htp_ops_context * octx);
 int op_argsort(struct htp_ops_context * octx);
+int op_ssm_conv(struct htp_ops_context * octx);
 
 #endif /* HTP_OPS_H */

ggml/src/ggml-hexagon/htp/hvx-arith.h

@@ -13,14 +13,15 @@
 // Binary operations (add, mul, sub)
 //
 
-#define hvx_arith_loop_body(dst_type, src0_type, src1_type, vec_store, vec_op) \
+#define UNUSED(x) (void)(x)
+
+#define hvx_arith_loop_body(dst_type, src0_type, src1_type, elem_size, vec_store, vec_op) \
     do {                                                                       \
         dst_type * restrict vdst  = (dst_type *) dst;                          \
         src0_type * restrict vsrc0 = (src0_type *) src0;                       \
         src1_type * restrict vsrc1 = (src1_type *) src1;                       \
                                                                                \
-        const uint32_t elem_size = sizeof(float);                              \
-        const uint32_t epv  = 128 / elem_size;                                 \
+        const uint32_t epv  = 128 / (elem_size);                               \
         const uint32_t nvec = n / epv;                                         \
         const uint32_t nloe = n % epv;                                         \
                                                                                \
@@ -32,62 +33,74 @@
         }                                                                      \
         if (nloe) {                                                            \
             HVX_Vector v = vec_op(vsrc0[i], vsrc1[i]);                         \
-            vec_store((void *) &vdst[i], nloe * elem_size, v);                 \
+            vec_store((void *) &vdst[i], nloe * (elem_size), v);               \
         }                                                                      \
     } while(0)
 
 #if __HVX_ARCH__ < 79
-#define HVX_OP_ADD(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(a, b))
-#define HVX_OP_SUB(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(a, b))
-#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+
+#define HVX_OP_ADD_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(a, b))
+#define HVX_OP_SUB_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(a, b))
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+
 #else
-#define HVX_OP_ADD(a, b) Q6_Vsf_vadd_VsfVsf(a, b)
-#define HVX_OP_SUB(a, b) Q6_Vsf_vsub_VsfVsf(a, b)
-#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+
+#define HVX_OP_ADD_F32(a, b) Q6_Vsf_vadd_VsfVsf(a, b)
+#define HVX_OP_SUB_F32(a, b) Q6_Vsf_vsub_VsfVsf(a, b)
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+
 #endif
 
+#define HVX_OP_ADD_F16(a, b) hvx_vec_add_f16_f16(a, b)
+#define HVX_OP_SUB_F16(a, b) hvx_vec_sub_f16_f16(a, b)
+#define HVX_OP_MUL_F16(a, b) hvx_vec_mul_f16_f16(a, b)
+
 // Generic macro to define alignment permutations for an op
-#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO) \
+#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO, ELEM_TYPE) \
 static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src0 % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
 static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src0 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
 static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
 static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
-    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
 static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src0 % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src0 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src1 % 128 == 0); \
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 
-DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD)
-DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB)
-DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f32, HVX_OP_MUL)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD_F32, float)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB_F32, float)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f32, HVX_OP_MUL_F32, float)
+
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f16, HVX_OP_ADD_F16, _Float16)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f16, HVX_OP_SUB_F16, _Float16)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f16, HVX_OP_MUL_F16, _Float16)
 
 // Dispatcher logic
 #define HVX_BINARY_DISPATCHER(OP_NAME) \
@@ -115,6 +128,10 @@ HVX_BINARY_DISPATCHER(hvx_add_f32)
 HVX_BINARY_DISPATCHER(hvx_sub_f32)
 HVX_BINARY_DISPATCHER(hvx_mul_f32)
 
+HVX_BINARY_DISPATCHER(hvx_add_f16)
+HVX_BINARY_DISPATCHER(hvx_sub_f16)
+HVX_BINARY_DISPATCHER(hvx_mul_f16)
+
 // Mul-Mul Optimized
 static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint8_t * restrict src2, const uint32_t num_elems) {
     assert((unsigned long) dst % 128 == 0);
@@ -136,26 +153,25 @@ static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * re
 
     _Pragma("unroll(4)")
     for (; i < nvec; i++) {
-        HVX_Vector v1 = HVX_OP_MUL(vsrc0[i], vsrc1[i]);
+        HVX_Vector v1 = HVX_OP_MUL_F32(vsrc0[i], vsrc1[i]);
         vdst[i] = HVX_OP_MUL(v1, vsrc2[i]);
     }
 
     if (nloe) {
-        HVX_Vector v1 = HVX_OP_MUL(vsrc0[i], vsrc1[i]);
-        HVX_Vector v2 = HVX_OP_MUL(v1, vsrc2[i]);
+        HVX_Vector v1 = HVX_OP_MUL_F32(vsrc0[i], vsrc1[i]);
+        HVX_Vector v2 = HVX_OP_MUL_F32(v1, vsrc2[i]);
         hvx_vec_store_a((void *) &vdst[i], nloe * elem_size, v2);
     }
 }
 
 // Scalar Operations
 
-#define hvx_scalar_loop_body(dst_type, src_type, vec_store, scalar_op_macro)   \
+#define hvx_scalar_loop_body(dst_type, src_type, elem_size, vec_store, scalar_op_macro)   \
     do {                                                                       \
         dst_type * restrict vdst = (dst_type *) dst;                           \
         src_type * restrict vsrc = (src_type *) src;                           \
                                                                                \
-        const uint32_t elem_size = sizeof(float);                              \
-        const uint32_t epv  = 128 / elem_size;                                 \
+        const uint32_t epv  = 128 / (elem_size);                               \
         const uint32_t nvec = n / epv;                                         \
         const uint32_t nloe = n % epv;                                         \
                                                                                \
@@ -169,138 +185,88 @@ static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * re
         if (nloe) {                                                            \
             HVX_Vector v = vsrc[i];                                            \
             v = scalar_op_macro(v);                                            \
-            vec_store((void *) &vdst[i], nloe * elem_size, v);                 \
+            vec_store((void *) &vdst[i], nloe * (elem_size), v);               \
         }                                                                      \
     } while(0)
 
-#define HVX_OP_ADD_SCALAR(v) \
+#define HVX_OP_ADD_SCALAR_F32(v) \
     ({ \
         const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, v); \
-        HVX_Vector out = HVX_OP_ADD(v, val_vec); \
+        HVX_Vector out = HVX_OP_ADD_F32(v, val_vec); \
         Q6_V_vmux_QVV(pred_inf, inf, out); \
     })
 
-#define HVX_OP_MUL_SCALAR(v) HVX_OP_MUL(v, val_vec)
-#define HVX_OP_SUB_SCALAR(v) HVX_OP_SUB(v, val_vec)
+#define HVX_OP_MUL_SCALAR_F32(v) HVX_OP_MUL_F32(v, val_vec)
+#define HVX_OP_SUB_SCALAR_F32(v) HVX_OP_SUB_F32(v, val_vec)
 
-// Add Scalar Variants
+#define HVX_OP_ADD_SCALAR_F16(v) \
+    ({ \
+        const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VhVh(inf, v); \
+        HVX_Vector out = HVX_OP_ADD_F16(v, val_vec); \
+        Q6_V_vmux_QVV(pred_inf, inf, out); \
+    })
 
-static inline void hvx_add_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    const HVX_Vector inf = hvx_vec_splat_f32(INFINITY);
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_ADD_SCALAR);
+#define HVX_OP_MUL_SCALAR_F16(v) HVX_OP_MUL_F16(v, val_vec)
+#define HVX_OP_SUB_SCALAR_F16(v) HVX_OP_SUB_F16(v, val_vec)
+
+// Scalar Variants
+
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(OP_NAME, OP_MACRO, SPLAT_MACRO, ELEM_TYPE) \
+static inline void OP_NAME##_aa(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
+    const HVX_Vector val_vec = SPLAT_MACRO(val); \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src % 128 == 0); \
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_au(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
+    const HVX_Vector val_vec = SPLAT_MACRO(val); \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
+    assert((uintptr_t) dst % 128 == 0); \
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_ua(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
+    const HVX_Vector val_vec = SPLAT_MACRO(val); \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
+    assert((uintptr_t) src % 128 == 0); \
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uu(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, uint32_t n) { \
+    const HVX_Vector val_vec = SPLAT_MACRO(val); \
+    const HVX_Vector inf = SPLAT_MACRO((ELEM_TYPE)INFINITY); UNUSED(inf); \
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
+} \
+
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_add_scalar_f32, HVX_OP_ADD_SCALAR_F32, hvx_vec_splat_f32, float)
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_sub_scalar_f32, HVX_OP_SUB_SCALAR_F32, hvx_vec_splat_f32, float)
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_mul_scalar_f32, HVX_OP_MUL_SCALAR_F32, hvx_vec_splat_f32, float)
+
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_add_scalar_f16, HVX_OP_ADD_SCALAR_F16, hvx_vec_splat_f16, _Float16)
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_sub_scalar_f16, HVX_OP_SUB_SCALAR_F16, hvx_vec_splat_f16, _Float16)
+DEFINE_HVX_BINARY_SCALAR_OP_VARIANTS(hvx_mul_scalar_f16, HVX_OP_MUL_SCALAR_F16, hvx_vec_splat_f16, _Float16)
+
+// Dispatcher logic
+#define HVX_BINARY_SCALAR_DISPATCHER(OP_NAME, ELEM_TYPE) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src, const ELEM_TYPE val, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) { \
+        OP_NAME##_aa(dst, src, val, num_elems); \
+    } else if (hex_is_aligned((void *) dst, 128)) { \
+        OP_NAME##_au(dst, src, val, num_elems); \
+    } else if (hex_is_aligned((void *) src, 128)) { \
+        OP_NAME##_ua(dst, src, val, num_elems); \
+    } else { \
+        OP_NAME##_uu(dst, src, val, num_elems); \
+    } \
 }
 
-static inline void hvx_add_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    const HVX_Vector inf = hvx_vec_splat_f32(INFINITY);
-    assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_ADD_SCALAR);
-}
+HVX_BINARY_SCALAR_DISPATCHER(hvx_add_scalar_f32, float)
+HVX_BINARY_SCALAR_DISPATCHER(hvx_sub_scalar_f32, float)
+HVX_BINARY_SCALAR_DISPATCHER(hvx_mul_scalar_f32, float)
 
-static inline void hvx_add_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    const HVX_Vector inf = hvx_vec_splat_f32(INFINITY);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_ADD_SCALAR);
-}
-
-static inline void hvx_add_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    static const float kInf = INFINITY;
-    const HVX_Vector inf = hvx_vec_splat_f32(kInf);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_ADD_SCALAR);
-}
-
-// Sub Scalar Variants
-
-static inline void hvx_sub_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_SUB_SCALAR);
-}
-
-static inline void hvx_sub_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_SUB_SCALAR);
-}
-
-static inline void hvx_sub_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_SUB_SCALAR);
-}
-
-static inline void hvx_sub_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_SUB_SCALAR);
-}
-
-// Mul Scalar Variants
-
-static inline void hvx_mul_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_MUL_SCALAR);
-}
-
-static inline void hvx_mul_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_MUL_SCALAR);
-}
-
-static inline void hvx_mul_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_MUL_SCALAR);
-}
-
-static inline void hvx_mul_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
-    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_MUL_SCALAR);
-}
-
-static inline void hvx_add_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
-        hvx_add_scalar_f32_aa(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) dst, 128)) {
-        hvx_add_scalar_f32_au(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) src, 128)) {
-        hvx_add_scalar_f32_ua(dst, src, val, num_elems);
-    } else {
-        hvx_add_scalar_f32_uu(dst, src, val, num_elems);
-    }
-}
-
-static inline void hvx_mul_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
-        hvx_mul_scalar_f32_aa(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) dst, 128)) {
-        hvx_mul_scalar_f32_au(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) src, 128)) {
-        hvx_mul_scalar_f32_ua(dst, src, val, num_elems);
-    } else {
-        hvx_mul_scalar_f32_uu(dst, src, val, num_elems);
-    }
-}
-
-static inline void hvx_sub_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
-        hvx_sub_scalar_f32_aa(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) dst, 128)) {
-        hvx_sub_scalar_f32_au(dst, src, val, num_elems);
-    } else if (hex_is_aligned((void *) src, 128)) {
-        hvx_sub_scalar_f32_ua(dst, src, val, num_elems);
-    } else {
-        hvx_sub_scalar_f32_uu(dst, src, val, num_elems);
-    }
-}
+HVX_BINARY_SCALAR_DISPATCHER(hvx_add_scalar_f16, _Float16)
+HVX_BINARY_SCALAR_DISPATCHER(hvx_sub_scalar_f16, _Float16)
+HVX_BINARY_SCALAR_DISPATCHER(hvx_mul_scalar_f16, _Float16)
 
 // MIN Scalar variants
 
@@ -310,24 +276,24 @@ static inline void hvx_min_scalar_f32_aa(uint8_t * restrict dst, const uint8_t *
     const HVX_Vector val_vec = hvx_vec_splat_f32(val);
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_MIN_SCALAR);
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(float), hvx_vec_store_a, HVX_OP_MIN_SCALAR);
 }
 
 static inline void hvx_min_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
     const HVX_Vector val_vec = hvx_vec_splat_f32(val);
     assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_MIN_SCALAR);
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(float), hvx_vec_store_a, HVX_OP_MIN_SCALAR);
 }
 
 static inline void hvx_min_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
     const HVX_Vector val_vec = hvx_vec_splat_f32(val);
     assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_MIN_SCALAR);
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(float), hvx_vec_store_u, HVX_OP_MIN_SCALAR);
 }
 
 static inline void hvx_min_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
     const HVX_Vector val_vec = hvx_vec_splat_f32(val);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_MIN_SCALAR);
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(float), hvx_vec_store_u, HVX_OP_MIN_SCALAR);
 }
 
 static inline void hvx_min_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
@@ -357,27 +323,27 @@ static inline void hvx_clamp_scalar_f32_aa(uint8_t * restrict dst, const uint8_t
     const HVX_Vector max_vec = hvx_vec_splat_f32(max);
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_CLAMP_SCALAR);
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(float), hvx_vec_store_a, HVX_OP_CLAMP_SCALAR);
 }
 
 static inline void hvx_clamp_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, uint32_t n) {
     const HVX_Vector min_vec = hvx_vec_splat_f32(min);
     const HVX_Vector max_vec = hvx_vec_splat_f32(max);
     assert((unsigned long) dst % 128 == 0);
-    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_CLAMP_SCALAR);
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(float), hvx_vec_store_a, HVX_OP_CLAMP_SCALAR);
 }
 
 static inline void hvx_clamp_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, uint32_t n) {
     const HVX_Vector min_vec = hvx_vec_splat_f32(min);
     const HVX_Vector max_vec = hvx_vec_splat_f32(max);
     assert((unsigned long) src % 128 == 0);
-    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u, HVX_OP_CLAMP_SCALAR);
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(float), hvx_vec_store_u, HVX_OP_CLAMP_SCALAR);
 }
 
 static inline void hvx_clamp_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, uint32_t n) {
     const HVX_Vector min_vec = hvx_vec_splat_f32(min);
     const HVX_Vector max_vec = hvx_vec_splat_f32(max);
-    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_CLAMP_SCALAR);
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(float), hvx_vec_store_u, HVX_OP_CLAMP_SCALAR);
 }
 
 static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float min, const float max, const int num_elems) {
@@ -396,7 +362,7 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
 // Square
 //
 
-#define hvx_sqr_loop_body(dst_type, src_type, vec_store)           \
+#define hvx_sqr_f32_loop_body(dst_type, src_type, vec_store)           \
     do {                                                                   \
         dst_type * restrict vdst  = (dst_type *) dst;                      \
         src_type * restrict vsrc = (src_type *) src;                       \
@@ -410,10 +376,10 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
                                                                            \
         _Pragma("unroll(4)")                                               \
         for (; i < nvec; i++) {                                            \
-            vdst[i] = HVX_OP_MUL(vsrc[i], vsrc[i]);                        \
+            vdst[i] = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                        \
         }                                                                  \
         if (nloe) {                                                        \
-            HVX_Vector v = HVX_OP_MUL(vsrc[i], vsrc[i]);                   \
+            HVX_Vector v = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                   \
             vec_store((void *) &vdst[i], nloe * elem_size, v);             \
         }                                                                  \
     } while(0)
@@ -421,21 +387,21 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
 static inline void hvx_sqr_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src % 128 == 0);
-    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+    hvx_sqr_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
 }
 
 static inline void hvx_sqr_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
     assert((unsigned long) dst % 128 == 0);
-    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+    hvx_sqr_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
 }
 
 static inline void hvx_sqr_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
     assert((unsigned long) src % 128 == 0);
-    hvx_sqr_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+    hvx_sqr_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
 }
 
 static inline void hvx_sqr_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    hvx_sqr_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+    hvx_sqr_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
 }
 
 static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
@@ -454,17 +420,24 @@ static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict
     }
 }
 
-#undef HVX_OP_ADD
-#undef HVX_OP_SUB
-#undef HVX_OP_MUL
+#undef HVX_OP_ADD_F32
+#undef HVX_OP_SUB_F32
+#undef HVX_OP_MUL_F32
+#undef HVX_OP_ADD_F16
+#undef HVX_OP_SUB_F16
+#undef HVX_OP_MUL_F16
 #undef hvx_arith_loop_body
-#undef HVX_OP_ADD_SCALAR
-#undef HVX_OP_SUB_SCALAR
-#undef HVX_OP_MUL_SCALAR
+#undef HVX_OP_ADD_SCALAR_F32
+#undef HVX_OP_SUB_SCALAR_F32
+#undef HVX_OP_MUL_SCALAR_F32
+#undef HVX_OP_ADD_SCALAR_F16
+#undef HVX_OP_SUB_SCALAR_F16
+#undef HVX_OP_MUL_SCALAR_F16
 #undef hvx_scalar_loop_body
 #undef HVX_OP_MIN_SCALAR
 #undef HVX_OP_CLAMP_SCALAR
 #undef DEFINE_HVX_BINARY_OP_VARIANTS
 #undef HVX_BINARY_DISPATCHER
+#undef UNUSED
 
 #endif // HVX_ARITH_H

ggml/src/ggml-hexagon/htp/hvx-base.h

@@ -189,4 +189,52 @@ static inline HVX_VectorPair hvx_vec_mpyacc_f32_f16(HVX_VectorPair acc, HVX_Vect
 
 #endif
 
+#if __HVX_ARCH__ < 79
+
+static inline HVX_Vector hvx_vec_add_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    const HVX_Vector negone = Q6_Vh_vsplat_R(0xBC00); // -1.0 in IEEE FP16
+    const HVX_Vector one    = Q6_Vh_vsplat_R(0x3C00); //  1.0 in IEEE FP16
+    HVX_VectorPair a_p = Q6_Wqf32_vmpy_VhfVhf(a, one);
+    HVX_VectorPair b_p = Q6_Wqf32_vmpy_VhfVhf(b, negone);
+    HVX_Vector a0 = Q6_Vqf32_vsub_Vqf32Vqf32(Q6_V_lo_W(a_p), Q6_V_lo_W(b_p));
+    HVX_Vector a1 = Q6_Vqf32_vsub_Vqf32Vqf32(Q6_V_hi_W(a_p), Q6_V_hi_W(b_p));
+    return Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(a1, a0));
+}
+
+static inline HVX_Vector hvx_vec_sub_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    const HVX_Vector negone = Q6_Vh_vsplat_R(0xBC00); // -1.0 in IEEE FP16
+    const HVX_Vector one    = Q6_Vh_vsplat_R(0x3C00); //  1.0 in IEEE FP16
+    HVX_VectorPair a_p = Q6_Wqf32_vmpy_VhfVhf(a, one);
+    HVX_VectorPair b_p = Q6_Wqf32_vmpy_VhfVhf(b, negone);
+    HVX_Vector a0 = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(a_p), Q6_V_lo_W(b_p));
+    HVX_Vector a1 = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_hi_W(a_p), Q6_V_hi_W(b_p));
+    return Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(a1, a0));
+}
+
+static inline HVX_Vector hvx_vec_mul_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    return Q6_Vhf_equals_Wqf32(Q6_Wqf32_vmpy_VhfVhf(a, b));
+}
+
+#else
+
+static inline HVX_Vector hvx_vec_add_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    return Q6_Vhf_vadd_VhfVhf(a, b);
+}
+
+static inline HVX_Vector hvx_vec_sub_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    return Q6_Vhf_vsub_VhfVhf(a, b);
+}
+
+static inline HVX_Vector hvx_vec_mul_f16_f16(HVX_Vector a, HVX_Vector b)
+{
+    return Q6_Vhf_vmpy_VhfVhf(a, b);
+}
+
+#endif // __HVX_ARCH__ < 79
+
 #endif /* HVX_BASE_H */

ggml/src/ggml-hexagon/htp/hvx-div.h

@@ -15,11 +15,144 @@
 #include "hvx-arith.h"
 
 #if __HVX_ARCH__ < 79
-#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
 #else
-#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
 #endif
 
+// Compute div by scaler in f32. Requires first by expanding fp32 to fp16 and converting the result back to fp32.
+static inline HVX_Vector hvx_div_mul_f16_const_using_f32(HVX_Vector vec1_hf, HVX_Vector vec2_sf_const, HVX_Vector vec_hf_one_1_0) {
+#if __HVX_ARCH__ < 79
+    HVX_VectorPair src_to_f32 = Q6_Wqf32_vmpy_VhfVhf(vec1_hf, vec_hf_one_1_0);
+    HVX_Vector src_to_f32_0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(src_to_f32));
+    HVX_Vector src_to_f32_1 = Q6_Vsf_equals_Vqf32(Q6_V_hi_W(src_to_f32));
+#else
+    HVX_VectorPair src_to_f32 = Q6_Wsf_vmpy_VhfVhf(vec1_hf, vec_hf_one_1_0);
+    HVX_Vector src_to_f32_0 = Q6_V_lo_W(src_to_f32);
+    HVX_Vector src_to_f32_1 = Q6_V_hi_W(src_to_f32);
+#endif
+
+    HVX_Vector div_f32_0 = HVX_OP_MUL_F32(src_to_f32_0, vec2_sf_const);
+    HVX_Vector div_f32_1 = HVX_OP_MUL_F32(src_to_f32_1, vec2_sf_const);
+
+#if __HVX_ARCH__ < 79
+    HVX_Vector res = hvx_vec_f32_to_f16(div_f32_0, div_f32_1);
+#else
+    HVX_Vector res = Q6_Vhf_vcvt_VsfVsf(div_f32_0, div_f32_1);
+#endif
+    return res;
+}
+
+#define hvx_div_scaler_f16_loop_body(dst_type, src_type, vec_store)                     \
+    do {                                                                                \
+        dst_type * restrict vdst = (dst_type *) dst;                                    \
+        src_type * restrict vsrc = (src_type *) src;                                    \
+        HVX_Vector hf_one = Q6_Vh_vsplat_R(0x3C00);                                     \
+                                                                                        \
+        const uint32_t nvec = n / VLEN_FP16;                                            \
+        const uint32_t nloe = n % VLEN_FP16;                                            \
+                                                                                        \
+        uint32_t i = 0;                                                                 \
+                                                                                        \
+        _Pragma("unroll(4)")                                                            \
+        for (; i < nvec; i++) {                                                         \
+            HVX_Vector res = hvx_div_mul_f16_const_using_f32(vsrc[i], val_vec_f32, hf_one); \
+            vdst[i] = res;                                                              \
+        }                                                                               \
+        if (nloe) {                                                                     \
+            HVX_Vector res = hvx_div_mul_f16_const_using_f32(vsrc[i], val_vec_f32, hf_one); \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, res);                      \
+        }                                                                               \
+    } while(0)
+
+static inline void hvx_div_scalar_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
+    const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src % 128 == 0);
+    hvx_div_scaler_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+static inline void hvx_div_scalar_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
+    const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    assert((uintptr_t) dst % 128 == 0);
+    hvx_div_scaler_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+static inline void hvx_div_scalar_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
+    const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    assert((uintptr_t) src % 128 == 0);
+    hvx_div_scaler_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+static inline void hvx_div_scalar_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
+    const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    hvx_div_scaler_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+// Compute div by using hvx_vec_inverse_f32_guard. Requires first by exapnding fp32 to fp16 and convert the result back to fp32.
+static inline HVX_Vector hvx_vec_div_f16_using_f32(HVX_Vector vec1, HVX_Vector vec2, HVX_Vector f32_nan_inf_mask, HVX_Vector vec_hf_one_1_0) {
+#if __HVX_ARCH__ < 79
+    // Convert first input to fp32
+    HVX_VectorPair vec1_to_f32   = Q6_Wqf32_vmpy_VhfVhf(vec1, vec_hf_one_1_0);  // *1.0
+    HVX_Vector     vec1_to_f32_0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(vec1_to_f32));
+    HVX_Vector     vec1_to_f32_1 = Q6_Vsf_equals_Vqf32(Q6_V_hi_W(vec1_to_f32));
+
+    // Convert second input to fp32
+    HVX_VectorPair vec2_to_f32   = Q6_Wqf32_vmpy_VhfVhf(vec2, vec_hf_one_1_0);  // *1.0
+    HVX_Vector     vec2_to_f32_0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(vec2_to_f32));
+    HVX_Vector     vec2_to_f32_1 = Q6_Vsf_equals_Vqf32(Q6_V_hi_W(vec2_to_f32));
+#else
+    // Convert first input to fp32
+    HVX_VectorPair vec1_to_f32   = Q6_Wsf_vmpy_VhfVhf(vec1, vec_hf_one_1_0);  // *1.0
+    HVX_Vector     vec1_to_f32_0 = Q6_V_lo_W(vec1_to_f32);
+    HVX_Vector     vec1_to_f32_1 = Q6_V_hi_W(vec1_to_f32);
+
+    // Convert second input to fp32
+    HVX_VectorPair vec2_to_f32   = Q6_Wsf_vmpy_VhfVhf(vec2, vec_hf_one_1_0);  // *1.0
+    HVX_Vector     vec2_to_f32_0 = Q6_V_lo_W(vec2_to_f32);
+    HVX_Vector     vec2_to_f32_1 = Q6_V_hi_W(vec2_to_f32);
+#endif
+
+    // Inverse second input in fp32
+    HVX_Vector     vec2_inv_f32_0 = hvx_vec_inverse_f32_guard(vec2_to_f32_0, f32_nan_inf_mask);
+    HVX_Vector     vec2_inv_f32_1 = hvx_vec_inverse_f32_guard(vec2_to_f32_1, f32_nan_inf_mask);
+
+    // Multiply first input by inverse of second, in fp32
+    HVX_Vector     div_f32_0 = HVX_OP_MUL_F32(vec1_to_f32_0, vec2_inv_f32_0);
+    HVX_Vector     div_f32_1 = HVX_OP_MUL_F32(vec1_to_f32_1, vec2_inv_f32_1);
+
+    // Convert back to fp16
+#if __HVX_ARCH__ < 79
+    HVX_Vector     recip = hvx_vec_f32_to_f16(div_f32_0, div_f32_1);
+#else
+    HVX_Vector     recip = Q6_Vhf_vcvt_VsfVsf(div_f32_0, div_f32_1);
+#endif
+
+    return recip;
+}
+
+#define hvx_div_f16_loop_body(dst_type, src0_type, src1_type, vec_store)                  \
+    do {                                                                                  \
+        dst_type * restrict vdst = (dst_type *) dst;                                      \
+        src0_type * restrict vsrc0 = (src0_type *) src0;                                  \
+        src1_type * restrict vsrc1 = (src1_type *) src1;                                  \
+                                                                                          \
+        const HVX_Vector nan_inf_mask = Q6_V_vsplat_R(0x7f800000);                        \
+        const HVX_Vector hf_one = Q6_Vh_vsplat_R(0x3C00);                                 \
+                                                                                          \
+        const uint32_t nvec = n / VLEN_FP16;                                              \
+        const uint32_t nloe = n % VLEN_FP16;                                              \
+                                                                                          \
+        uint32_t i = 0;                                                                   \
+                                                                                          \
+        _Pragma("unroll(4)")                                                              \
+        for (; i < nvec; i++) {                                                           \
+            HVX_Vector res = hvx_vec_div_f16_using_f32(vsrc0[i], vsrc1[i], nan_inf_mask, hf_one); \
+            vdst[i] = res;                                                                \
+        }                                                                                 \
+        if (nloe) {                                                                       \
+            HVX_Vector res = hvx_vec_div_f16_using_f32(vsrc0[i], vsrc1[i], nan_inf_mask, hf_one); \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, res);                        \
+        }                                                                                 \
+    } while(0)
+
 #define hvx_div_f32_loop_body(dst_type, src0_type, src1_type, vec_store)             \
     do {                                                                             \
         dst_type * restrict vdst = (dst_type *) dst;                                 \
@@ -36,81 +169,83 @@
         _Pragma("unroll(4)")                                                         \
         for (; i < nvec; i++) {                                                      \
             HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
-            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            HVX_Vector res = HVX_OP_MUL_F32(vsrc0[i], inv_src1);                     \
             vdst[i] = res;                                                           \
         }                                                                            \
         if (nloe) {                                                                  \
             HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
-            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            HVX_Vector res = HVX_OP_MUL_F32(vsrc0[i], inv_src1);                     \
             vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, res);                   \
         }                                                                            \
     } while(0)
 
-// 3-letter suffix variants
-static inline void hvx_div_f32_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) dst % 128 == 0);
-    assert((uintptr_t) src0 % 128 == 0);
-    assert((uintptr_t) src1 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a);
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_DIV_OP_VARIANTS(OP_NAME, OP_LOOP_BODY) \
+static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u); \
+} \
+static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u); \
+} \
+static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src1 % 128 == 0); \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u); \
+} \
+static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u); \
+} \
+
+// Dispatcher logic
+#define HVX_DIV_DISPATCHER(OP_NAME) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128)) { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_aau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_auu(dst, src0, src1, num_elems); \
+        } \
+    } else { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems); \
+        } \
+    } \
 }
 
-static inline void hvx_div_f32_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) dst % 128 == 0);
-    assert((uintptr_t) src0 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a);
-}
+DEFINE_HVX_DIV_OP_VARIANTS(hvx_div_f32, hvx_div_f32_loop_body)
+DEFINE_HVX_DIV_OP_VARIANTS(hvx_div_f16, hvx_div_f16_loop_body)
 
-static inline void hvx_div_f32_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) dst % 128 == 0);
-    assert((uintptr_t) src1 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a);
-}
+HVX_DIV_DISPATCHER(hvx_div_f32)
+HVX_DIV_DISPATCHER(hvx_div_f16)
 
-static inline void hvx_div_f32_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) dst % 128 == 0);
-    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a);
-}
-
-static inline void hvx_div_f32_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) src0 % 128 == 0);
-    assert((uintptr_t) src1 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u);
-}
-
-static inline void hvx_div_f32_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) src0 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u);
-}
-
-static inline void hvx_div_f32_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((uintptr_t) src1 % 128 == 0);
-    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u);
-}
-
-static inline void hvx_div_f32_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u);
-}
-
-static inline void hvx_div_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128)) {
-        if (hex_is_aligned((void *) src0, 128)) {
-            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aaa(dst, src0, src1, num_elems);
-            else                                    hvx_div_f32_aau(dst, src0, src1, num_elems);
-        } else {
-            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aua(dst, src0, src1, num_elems);
-            else                                    hvx_div_f32_auu(dst, src0, src1, num_elems);
-        }
-    } else {
-        if (hex_is_aligned((void *) src0, 128)) {
-            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uaa(dst, src0, src1, num_elems);
-            else                                    hvx_div_f32_uau(dst, src0, src1, num_elems);
-        } else {
-            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uua(dst, src0, src1, num_elems);
-            else                                    hvx_div_f32_uuu(dst, src0, src1, num_elems);
-        }
-    }
-}
-
-#undef HVX_OP_MUL
+#undef HVX_OP_MUL_F32
 
 #endif // HVX_DIV_H

ggml/src/ggml-hexagon/htp/hvx-inverse.h

@@ -137,40 +137,74 @@ static inline HVX_Vector hvx_vec_inverse_f32_guard(HVX_Vector v_sf, HVX_Vector n
         }                                                                    \
     } while(0)
 
-static inline void hvx_inverse_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src % 128 == 0);
-    hvx_inverse_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+static inline HVX_Vector hvx_vec_inverse_f16_guard(HVX_Vector v_sf, HVX_Vector nan_inf_mask) {
+    HVX_Vector out = hvx_vec_inverse_f16(v_sf);
+
+    HVX_Vector     masked_out = Q6_V_vand_VV(out, nan_inf_mask);
+    const HVX_VectorPred pred = Q6_Q_vcmp_eq_VhVh(nan_inf_mask, masked_out);
+
+    return Q6_V_vmux_QVV(pred, Q6_V_vzero(), out);
 }
 
-static inline void hvx_inverse_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    hvx_inverse_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+#define hvx_inverse_f16_loop_body(dst_type, src_type, vec_store)             \
+    do {                                                                     \
+        dst_type * restrict vdst = (dst_type *) dst;                         \
+        src_type * restrict vsrc = (src_type *) src;                         \
+                                                                             \
+        const HVX_Vector nan_inf_mask = Q6_Vh_vsplat_R(0x7c00);              \
+                                                                             \
+        const uint32_t nvec = n / VLEN_FP16;                                 \
+        const uint32_t nloe = n % VLEN_FP16;                                 \
+                                                                             \
+        uint32_t i = 0;                                                      \
+                                                                             \
+        _Pragma("unroll(4)")                                                 \
+        for (; i < nvec; i++) {                                              \
+             vdst[i] = hvx_vec_inverse_f16_guard(vsrc[i], nan_inf_mask);     \
+        }                                                                    \
+        if (nloe) {                                                          \
+            HVX_Vector v = hvx_vec_inverse_f16_guard(vsrc[i], nan_inf_mask); \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, v);             \
+        }                                                                    \
+    } while(0)
+
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_INV_OP_VARIANTS(OP_NAME, OP_LOOP_BODY) \
+static inline void OP_NAME##_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_Vector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    OP_LOOP_BODY(HVX_Vector, HVX_UVector, hvx_vec_store_a); \
+} \
+static inline void OP_NAME##_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
+    assert((uintptr_t) src % 128 == 0); \
+    OP_LOOP_BODY(HVX_UVector, HVX_Vector, hvx_vec_store_u); \
+} \
+static inline void OP_NAME##_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) { \
+    OP_LOOP_BODY(HVX_UVector, HVX_UVector, hvx_vec_store_u); \
+} \
+
+// Dispatcher logic
+#define HVX_INV_DISPATCHER(OP_NAME) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) { \
+        OP_NAME##_aa(dst, src, num_elems); \
+    } else if (hex_is_aligned((void *) dst, 128)) { \
+        OP_NAME##_au(dst, src, num_elems); \
+    } else if (hex_is_aligned((void *) src, 128)) { \
+        OP_NAME##_ua(dst, src, num_elems); \
+    } else { \
+        OP_NAME##_uu(dst, src, num_elems); \
+    } \
 }
 
-static inline void hvx_inverse_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    assert((unsigned long) src % 128 == 0);
-    hvx_inverse_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
-}
+DEFINE_HVX_INV_OP_VARIANTS(hvx_inverse_f32, hvx_inverse_f32_loop_body)
+DEFINE_HVX_INV_OP_VARIANTS(hvx_inverse_f16, hvx_inverse_f16_loop_body)
 
-static inline void hvx_inverse_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
-    hvx_inverse_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
-}
-
-static inline void hvx_inverse_f32(uint8_t * restrict dst, uint8_t * restrict src, const int num_elems) {
-    if ((unsigned long) dst % 128 == 0) {
-        if ((unsigned long) src % 128 == 0) {
-            hvx_inverse_f32_aa(dst, src, num_elems);
-        } else {
-            hvx_inverse_f32_au(dst, src, num_elems);
-        }
-    } else {
-        if ((unsigned long) src % 128 == 0) {
-            hvx_inverse_f32_ua(dst, src, num_elems);
-        } else {
-            hvx_inverse_f32_uu(dst, src, num_elems);
-        }
-    }
-}
+HVX_INV_DISPATCHER(hvx_inverse_f32)
+HVX_INV_DISPATCHER(hvx_inverse_f16)
 
 #endif // HVX_INVERSE_H

ggml/src/ggml-hexagon/htp/hvx-utils.h

@@ -15,4 +15,12 @@
 #include "hvx-div.h"
 #include "hvx-base.h"
 
+#ifndef GATHER_TYPE
+#    if defined(__hexagon__)
+#        define GATHER_TYPE(_a) (intptr_t) _a
+#    else
+#        define GATHER_TYPE(_a) (HVX_Vector *) _a
+#    endif
+#endif
+
 #endif /* HVX_UTILS_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -757,6 +757,47 @@ static void proc_sum_rows_req(struct htp_context * ctx, struct htp_general_req *
     send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
+static void proc_ssm_conv_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+
+    // We've written to the output buffer, we'd also need to flush it
+    rsp_bufs[0].fd     = bufs[2].fd;
+    rsp_bufs[0].ptr    = bufs[2].ptr;
+    rsp_bufs[0].offset = bufs[2].offset;
+    rsp_bufs[0].size   = bufs[2].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup OP context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.src1                   = req->src1;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.src1.data = (uint32_t) bufs[1].ptr;
+    octx.dst.data  = (uint32_t) bufs[2].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_ssm_conv(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
+}
+
 static void proc_activations_req(struct htp_context *     ctx,
                                  struct htp_general_req * req,
                                  struct dspqueue_buffer * bufs,
@@ -1142,6 +1183,14 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
                 proc_argsort_req(ctx, &req, bufs);
                 break;
 
+            case HTP_OP_SSM_CONV:
+                if (n_bufs != 3) {
+                    FARF(ERROR, "Bad ssm-conv-req buffer list");
+                    continue;
+                }
+                proc_ssm_conv_req(ctx, &req, bufs);
+                break;
+
             default:
                 FARF(ERROR, "Unknown Op %u", req.op);
                 break;

ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -400,7 +400,9 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t n_threads = octx->n_threads;
+    const uint32_t ne0 = dst->ne[0];
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads = MIN(octx->n_threads, src0_nrows);
 
     const size_t src0_row_size = src0->nb[1];
     const size_t dst_row_size  = dst->nb[1];
@@ -465,17 +467,14 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
     rctx.dst_row_size_aligned  = dst_row_size_aligned;
     rctx.theta_cache_offset    = theta_cache_size_aligned;
 
-    uint32_t ne0 = dst->ne[0];
-    uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
     rctx.src0_nrows = src0_nrows;
+    rctx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
 
     FARF(HIGH, "rope-f32 n-rows %u n-dims %d ne0 %u ext-factor %.6f theta-scale %.6f attn-factor %.6f\n", rctx.src0_nrows, rctx.n_dims, ne0,
          rctx.ext_factor, rctx.theta_scale, rctx.attn_factor);
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t n_jobs = MIN(n_threads, src0_nrows);
-        rctx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
-        worker_pool_run_func(octx->ctx->worker_pool, rope_job_f32, &rctx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, rope_job_f32, &rctx, n_threads);
     }
 
     return err;

ggml/src/ggml-hexagon/htp/set-rows-ops.c

@@ -128,6 +128,8 @@ static void set_rows_thread_f16_f32(unsigned int nth, unsigned int ith, void *da
 int op_set_rows(struct htp_ops_context * octx) {
     set_rows_preamble;
 
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
     if (octx->src0.type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
@@ -149,15 +151,14 @@ int op_set_rows(struct htp_ops_context * octx) {
     srctx.div_ne12 = init_fastdiv_values(ne12);
     srctx.div_ne11 = init_fastdiv_values(ne11);
 
-    const uint32_t n_jobs = MIN(nr, octx->n_threads);
-    srctx.src0_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+    srctx.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 
     switch(octx->dst.type) {
     case HTP_TYPE_F32:
-        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f32_f32, &srctx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f32_f32, &srctx, n_threads);
         break;
     case HTP_TYPE_F16:
-        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f16_f32, &srctx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f16_f32, &srctx, n_threads);
         break;
     default:
         return HTP_STATUS_NO_SUPPORT;

ggml/src/ggml-hexagon/htp/softmax-ops.c

@@ -353,7 +353,8 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const uint32_t n_threads = octx->n_threads;
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     const size_t src0_row_size = src0->nb[1];
     const size_t src1_row_size = src0_row_size;
@@ -393,12 +394,9 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {
     octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
     octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
 
-    uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
-
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t n_jobs             = MIN(n_threads, src0_nrows);
-        smctx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
-        worker_pool_run_func(octx->ctx->worker_pool, softmax_job_f32, &smctx, n_jobs);
+        smctx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
+        worker_pool_run_func(octx->ctx->worker_pool, softmax_job_f32, &smctx, n_threads);
     }
 
     return err;

ggml/src/ggml-hexagon/htp/ssm-conv.c

@@ -0,0 +1,339 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_mem.h>
+#include <HAP_perf.h>
+#include <HAP_ps.h>
+#include <hexagon_protos.h>
+#include <hexagon_types.h>
+#include <math.h>
+#include <qurt_thread.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "hex-dma.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+
+#define htp_ssm_conv_tensors_preamble                        \
+    struct htp_tensor * restrict src0    = &octx->src0;      \
+    struct htp_tensor * restrict src1    = &octx->src1;      \
+    struct htp_tensor * restrict dst     = &octx->dst;       \
+    struct htp_spad * restrict src0_spad = &octx->src0_spad; \
+    struct htp_spad * restrict src1_spad = &octx->src1_spad; \
+    struct htp_spad * restrict dst_spad  = &octx->dst_spad;  \
+                                                             \
+    const uint32_t ne00 = src0->ne[0];                       \
+    const uint32_t ne01 = src0->ne[1];                       \
+    const uint32_t ne02 = src0->ne[2];                       \
+    const uint32_t ne03 = src0->ne[3];                       \
+                                                             \
+    const uint32_t ne10 = src1->ne[0];                       \
+    const uint32_t ne11 = src1->ne[1];                       \
+    const uint32_t ne12 = src1->ne[2];                       \
+    const uint32_t ne13 = src1->ne[3];                       \
+                                                             \
+    const uint32_t ne0 = dst->ne[0];                         \
+    const uint32_t ne1 = dst->ne[1];                         \
+    const uint32_t ne2 = dst->ne[2];                         \
+    const uint32_t ne3 = dst->ne[3];                         \
+                                                             \
+    const uint32_t nb00 = src0->nb[0];                       \
+    const uint32_t nb01 = src0->nb[1];                       \
+    const uint32_t nb02 = src0->nb[2];                       \
+    const uint32_t nb03 = src0->nb[3];                       \
+                                                             \
+    const uint32_t nb10 = src1->nb[0];                       \
+    const uint32_t nb11 = src1->nb[1];                       \
+    const uint32_t nb12 = src1->nb[2];                       \
+    const uint32_t nb13 = src1->nb[3];                       \
+                                                             \
+    const uint32_t nb0 = dst->nb[0];                         \
+    const uint32_t nb1 = dst->nb[1];                         \
+    const uint32_t nb2 = dst->nb[2];                         \
+    const uint32_t nb3 = dst->nb[3];
+
+struct htp_ssm_conv_context {
+    struct htp_ops_context * octx;
+    uint32_t nrows_per_thread;
+    uint64_t t_start;
+};
+
+#define htp_ssm_conv_preamble                            \
+    struct htp_ssm_conv_context * scctx = (struct htp_ssm_conv_context *) data; \
+    struct htp_ops_context * octx = scctx->octx;         \
+    htp_ssm_conv_tensors_preamble;                       \
+    dma_queue * dma_queue         = octx->ctx->dma[ith];
+
+// Scalar FP32 SSM_CONV implementation
+static void ssm_conv_thread_f32_f32(unsigned int nth, unsigned int ith, void *data) {
+    htp_ssm_conv_preamble;
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const uint32_t d_conv  = src1->ne[0];
+    const uint32_t d_inner = src0->ne[1];
+    const uint32_t n_t     = dst->ne[1];
+    const uint32_t n_s     = dst->ne[2];
+
+    const uint32_t src0_stride_inner = src0->nb[1] / sizeof(float); // stride for inner dimension
+    const uint32_t src0_stride_seq   = src0->nb[2] / sizeof(float); // stride for sequence dimension
+    const uint32_t src1_stride_inner = src1->nb[1] / sizeof(float); // stride for inner dimension
+    const uint32_t dst_stride_token  = dst->nb[1]  / sizeof(float); // stride for token dimension
+    const uint32_t dst_stride_seq    = dst->nb[2]  / sizeof(float); // stride for sequence dimension
+
+    const float * src0_data = (const float *) src0->data;
+    const float * src1_data = (const float *) src1->data;
+    float *       dst_data  = (float *) dst->data;
+
+    // Calculate row range for this thread
+    const uint32_t d_inner_per_thread = scctx->nrows_per_thread;
+    const uint32_t d_inner_start = d_inner_per_thread * ith;
+    const uint32_t d_inner_end   = MIN(d_inner_start + d_inner_per_thread, d_inner);
+
+    // No work for this thread
+    if (d_inner_start >= d_inner_end) {
+        return;
+    }
+
+    for (uint32_t i3 = 0; i3 < n_s; ++i3) {
+        for (uint32_t i2 = 0; i2 < n_t; ++i2) {
+            for (uint32_t i1 = d_inner_start; i1 < d_inner_end; ++i1) {
+                float sumf = 0.0f;
+
+                for (uint32_t i0 = 0; i0 < d_conv; ++i0) {
+                    const uint32_t src0_idx = (i2 + i0) + i1 * src0_stride_inner + i3 * src0_stride_seq;
+                    const uint32_t src1_idx = i0 + i1 * src1_stride_inner;
+
+                    sumf += src0_data[src0_idx] * src1_data[src1_idx];
+                }
+
+                const uint32_t dst_idx = i1 + i2 * dst_stride_token + i3 * dst_stride_seq;
+                dst_data[dst_idx] = sumf;
+            }
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "ssm-conv-f32 %d/%d: %ux%ux%ux%u (%u:%u) * %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n",
+         ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], d_inner_start, d_inner_end,
+         src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], dst->ne[0], dst->ne[1],
+         dst->ne[2], dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+// HVX FP32 SSM_CONV implementation - vectorizes across d_inner dimension
+static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void *data) {
+    htp_ssm_conv_preamble;
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const int nc  = src1->ne[0]; // d_conv
+    const int ncs = src0->ne[0]; // d_conv - 1 + n_t
+
+    const uint32_t d_conv  = src1->ne[0];
+    const uint32_t d_inner = src0->ne[1];
+    const uint32_t n_t     = dst->ne[1];
+    const uint32_t n_s     = dst->ne[2];
+
+    const float * src0_data = (const float *) src0->data;
+    const float * src1_data = (const float *) src1->data;
+    float *       dst_data  = (float *) dst->data;
+
+    // Calculate row range for this thread
+    const int dr = scctx->nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    const uint32_t ir1 = MIN(ir0 + dr, d_inner);
+    const int      ir  = ir1 - ir0;
+
+    if (ir0 >= ir1) {
+        return;  // No work for this thread
+    }
+
+    // src0 and src1 gather offsets
+    uint32_t __attribute__((aligned(VLEN))) src0_offsets[VLEN_FP32] = { 0 };
+    uint32_t __attribute__((aligned(VLEN))) src1_offsets[VLEN_FP32] = { 0 };
+
+    for (uint32_t i = 0; i < VLEN_FP32; ++i) {
+        src0_offsets[i] = i * (ncs)    * sizeof(float);
+        src1_offsets[i] = i * (d_conv) * sizeof(float);
+    }
+
+    const uint32_t src0_gather_len = VLEN * ncs;
+    const uint32_t src1_gather_len = VLEN * d_conv;
+
+    // gather scratchpads
+    HVX_Vector * src0_vec = (HVX_Vector *) (octx->ctx->vtcm_base + ith * VLEN*2 + 0);
+    HVX_Vector * src1_vec = (HVX_Vector *) (octx->ctx->vtcm_base + ith * VLEN*2 + VLEN);
+
+    float * data_src0 = (float *) ((char *) src0->data + ir0 * src0->nb[1]);
+    float * data_src1 = (float *) ((char *) src1->data + ir0 * src1->nb[1]);
+
+    uint8_t * spad_src0 = octx->src0_spad.data + ith * octx->src0_spad.size_per_thread;
+    uint8_t * spad_src1 = octx->src1_spad.data + ith * octx->src1_spad.size_per_thread;
+
+    // copy src1 workload to VTCM
+    dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src1, data_src1), nb11, nb11, ir);
+
+    // FARF(HIGH, "ssm-conv-src1-fetch %d: ir0 %u size %u\n", ith, ir0, nb11 * ir);
+
+    for (uint32_t i3 = 0; i3 < n_s; ++i3) {
+        float * src0_data_ptr = (float *) ((char *) data_src0 + i3 * (src0->nb[2]));
+
+        // copy src0 workload to VTCM
+        dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0, src0_data_ptr), nb01, nb01, ir);
+
+        // FARF(HIGH, "ssm-conv-src0-fetch %d: ir0 %u i3 %u size %u\n", ith, ir0, i3, nb01 * ir);
+
+        dma_queue_flush(dma_queue);
+
+        for (uint32_t i2 = 0; i2 < n_t; ++i2) {
+            float * dst_ptr = (float *) ((char *) dst->data + ir0 * (dst->nb[0]) + i2 * (dst->nb[1]) + i3 * (dst->nb[2]));
+
+            const uint32_t nvec = ir / VLEN_FP32;
+            const uint32_t nloe = ir % VLEN_FP32;
+            uint32_t i1 = 0;
+
+            for (uint32_t vi1 = 0; vi1 < nvec; vi1++) {
+                HVX_Vector acc_vec = Q6_V_vsplat_R(0);
+
+                for (uint32_t i0 = 0; i0 < d_conv; ++i0) {
+                    Q6_vgather_ARMVw(src0_vec, GATHER_TYPE(spad_src0 + (i0 + i1 * ncs) * sizeof(float) + i2 * (src0->nb[0])),
+                                     src0_gather_len, (*(const HVX_Vector *) src0_offsets));
+                    Q6_vgather_ARMVw(src1_vec, GATHER_TYPE(spad_src1 + (i0 + i1 * nc) * sizeof(float)),
+                                     src1_gather_len, (*(const HVX_Vector *) src1_offsets));
+
+                    HVX_Vector prod = Q6_Vqf32_vmpy_VsfVsf(*(const HVX_Vector *) src0_vec, *(const HVX_Vector *) src1_vec);
+                    acc_vec = Q6_Vqf32_vadd_Vqf32Vqf32(acc_vec, prod);
+                }
+
+                *(HVX_UVector *) (dst_ptr + i1) = Q6_Vsf_equals_Vqf32(acc_vec);
+                i1 += VLEN_FP32;
+            }
+
+            if (nloe) {
+                HVX_Vector acc_vec = Q6_V_vsplat_R(0);
+
+                for (uint32_t i0 = 0; i0 < d_conv; ++i0) {
+                    Q6_vgather_ARMVw(src0_vec, GATHER_TYPE(spad_src0 + (i0 + i1 * ncs) * sizeof(float) + i2 * (src0->nb[0])),
+                                     src0_gather_len, (*(const HVX_Vector *) src0_offsets));
+                    Q6_vgather_ARMVw(src1_vec, GATHER_TYPE(spad_src1 + (i0 + i1 * nc) * sizeof(float)),
+                                     src1_gather_len, (*(const HVX_Vector *) src1_offsets));
+
+                    HVX_Vector prod = Q6_Vqf32_vmpy_VsfVsf(*(const HVX_Vector *) src0_vec, *(const HVX_Vector *) src1_vec);
+                    acc_vec = Q6_Vqf32_vadd_Vqf32Vqf32(acc_vec, prod);
+                }
+
+                hvx_vec_store_u(dst_ptr + i1, (ir - i1) * 4, Q6_Vsf_equals_Vqf32(acc_vec));
+            }
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "ssm-conv-f32-hvx %d/%d: %ux%ux%ux%u (%u:%u) * %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n",
+         ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ir0, ir1,
+         src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], dst->ne[0], dst->ne[1],
+         dst->ne[2], dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+int op_ssm_conv_f32(struct htp_ops_context * octx) {
+    htp_ssm_conv_tensors_preamble;
+
+    if (src0->type != HTP_TYPE_F32 || src1->type != HTP_TYPE_F32 || dst->type != HTP_TYPE_F32) {
+        FARF(ERROR, "ssm_conv: only (F32 x F32 -> F32) OPs supported");
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    struct htp_ssm_conv_context scctx = { 0 };
+    scctx.octx = octx;
+
+    const uint32_t d_conv  = src1->ne[0];
+    const uint32_t d_inner = src0->ne[1];
+    const uint32_t n_t     = dst->ne[1];  // tokens per sequence
+    const uint32_t n_s     = dst->ne[2];  // number of sequences in the batch
+
+    const uint32_t n_threads = MIN(octx->n_threads, d_inner);
+
+    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
+        uint32_t use_hvx = 0;
+        if (d_inner >= VLEN_FP32 && d_inner % VLEN_FP32 == 0) {
+            int is_aligned = hex_is_aligned((void *) src0->data, VLEN) &&
+                             hex_is_aligned((void *) src1->data, VLEN) &&
+                             hex_is_aligned((void *) dst->data, VLEN);
+
+            if (is_aligned) {
+                use_hvx = 1;
+            }
+        }
+
+        if (use_hvx) {
+            scctx.nrows_per_thread  = (d_inner + n_threads - 1) / n_threads; // d_inner chunks per thread
+            scctx.nrows_per_thread += (scctx.nrows_per_thread & 1); // round up to even
+
+            octx->src0_spad.size_per_thread = hex_round_up(scctx.nrows_per_thread * nb01, 256);
+            octx->src1_spad.size_per_thread = hex_round_up(scctx.nrows_per_thread * nb11, 256);
+            octx->dst_spad.size_per_thread  = hex_round_up(scctx.nrows_per_thread * sizeof(float), 256);
+
+            octx->src0_spad.size = octx->src0_spad.size_per_thread * n_threads;
+            octx->src1_spad.size = octx->src1_spad.size_per_thread * n_threads;
+            octx->dst_spad.size  = octx->dst_spad.size_per_thread  * n_threads;
+
+            // Compute gather scratchpad size for src0 and src1
+            const size_t gather_spad_size = n_threads * VLEN * 2;
+
+            octx->src0_spad.data = octx->ctx->vtcm_base + gather_spad_size;
+            octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+            octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
+
+            FARF(HIGH, "ssm_conv-f32: gather-spad:%zu spad-per-thread:(%u:%u:%u) spad-sizes:(%u:%u:%u) spad-data:(%p:%p:%p)\n",
+                gather_spad_size, octx->src0_spad.size_per_thread, octx->src1_spad.size_per_thread,
+                octx->dst_spad.size_per_thread, octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size,
+                octx->src0_spad.data, octx->src1_spad.data, octx->dst_spad.data);
+
+            const size_t total_spad_size =
+                gather_spad_size + octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;
+
+            if (total_spad_size > octx->ctx->vtcm_size) {
+                FARF(HIGH, "ssm_conv-f32: HVX scratchpad size %zu exceeds VTCM size %zu", total_spad_size,
+                     octx->ctx->vtcm_size);
+                use_hvx = 0;
+            }
+        }
+
+        FARF(HIGH, "ssm-conv-f32: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : use_hvx %d\n", src0->ne[0],
+             src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], dst->ne[0],
+             dst->ne[1], dst->ne[2], dst->ne[3], use_hvx);
+
+        if (use_hvx) {
+            worker_pool_run_func(octx->ctx->worker_pool, ssm_conv_thread_f32_f32_hvx, &scctx, n_threads);
+        } else {
+            worker_pool_run_func(octx->ctx->worker_pool, ssm_conv_thread_f32_f32, &scctx, n_threads);
+        }
+    }
+
+    return HTP_STATUS_OK;
+}
+
+int op_ssm_conv(struct htp_ops_context * octx) {
+    int                 err = HTP_STATUS_OK;
+    struct htp_tensor * dst = &octx->dst;
+
+    switch (dst->type) {
+        case HTP_TYPE_F32:
+            err = op_ssm_conv_f32(octx);
+            break;
+        default:
+            err = HTP_STATUS_NO_SUPPORT;
+            break;
+    }
+
+    return err;
+}

ggml/src/ggml-hexagon/htp/sum-rows-ops.c

@@ -102,11 +102,9 @@ int op_sum_rows(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
-    const int      n_threads  = octx->n_threads;
     const uint32_t src0_nrows = ne01 * ne02 * ne03;
-
-    uint32_t n_jobs = MIN(n_threads, src0_nrows);
-    uint32_t rows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+    const uint32_t n_threads = MIN(octx->n_threads, src0_nrows);
+    const uint32_t rows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
 
     bool opt_path = false;
     if ((0 == hex_is_aligned((void *) src0->data, VLEN)) && !(nb01 & (VLEN - 1))) {
@@ -124,7 +122,7 @@ int op_sum_rows(struct htp_ops_context * octx) {
         .opt_path        = opt_path,
     };
 
-    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_thread_f32, &smctx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_thread_f32, &smctx, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -301,8 +301,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
             return HTP_STATUS_NO_SUPPORT;
     }
 
-    const int      n_threads  = octx->n_threads;
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads  = MIN(octx->n_threads, src0_nrows);
 
     const size_t src0_row_size = src0->nb[1];
     const size_t dst_row_size  = dst->nb[1];
@@ -338,11 +338,9 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
          octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t n_jobs = MIN(n_threads, src0_nrows);
-
         struct htp_unary_context uctx = {
             .octx                  = octx,
-            .src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs,
+            .src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads,
             .src0_nrows            = src0_nrows,
 
             .data_src0             = (const uint8_t *)src0->data,
@@ -361,7 +359,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
             .nc                    = src0->ne[0],
         };
 
-        worker_pool_run_func(octx->ctx->worker_pool, unary_job_f32_per_thread, &uctx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, unary_job_f32_per_thread, &uctx, n_threads);
     }
 
     return err;

ggml/src/ggml-hip/CMakeLists.txt

@@ -11,6 +11,10 @@ endif()
 list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
 list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
 
+if (NOT DEFINED CMAKE_HIP_FLAGS_DEBUG)
+    set(CMAKE_HIP_FLAGS_DEBUG "-g -O2")
+endif()
+
 # CMake on Windows doesn't support the HIP language yet
 if (WIN32)
     set(CXX_IS_HIPCC TRUE)

ggml/src/ggml-impl.h

@@ -491,6 +491,61 @@ static inline float ggml_e8m0_to_fp32_half(uint8_t x) {
 #define GGML_E8M0_TO_FP32(x) ggml_e8m0_to_fp32(x)
 #define GGML_E8M0_TO_FP32_HALF(x) ggml_e8m0_to_fp32_half(x)
 
+// UE4M3: unsigned, 4 exp bits (bias=7), 3 mantissa bits
+// Returns value * 0.5 to match kvalues_mxfp4 convention (kvalues = 2 * E2M1_float)
+static inline float ggml_ue4m3_to_fp32(uint8_t x) {
+    if (x == 0 || x == 0x7F) {
+        return 0.0f;
+    }
+    int   exp = (x >> 3) & 0xF;
+    int   man = x & 0x7;
+    float raw;
+    if (exp == 0) {
+        raw = ldexpf((float) man, -9);
+    } else {
+        raw = ldexpf(1.0f + (float) man / 8.0f, exp - 7);
+    }
+    return raw * 0.5f;
+}
+
+static inline uint8_t ggml_fp32_to_ue4m3(float x) {
+    if (!(x > 0.0f)) {
+        return 0;
+    }
+    if (x > 448.0f) {
+        x = 448.0f;
+    }
+    uint32_t bits;
+    memcpy(&bits, &x, 4);
+    int fp32_exp  = ((bits >> 23) & 0xFF) - 127;
+    int fp32_man  = (bits >> 20) & 0x7;
+    int ue4m3_exp = fp32_exp + 7;
+    if (ue4m3_exp <= 0) {
+        // subnormal: value = man * 2^-9, man = round(x * 2^9)
+        int man = (int) (x * 512.0f + 0.5f);
+        if (man > 7) {
+            man = 7;
+        }
+        if (man < 1) {
+            return 0;
+        }
+        return (uint8_t) man;
+    }
+    if (ue4m3_exp >= 15) {
+        return 0x7E;
+    }
+    int round_bit = (bits >> 19) & 1;
+    int ue4m3_man = fp32_man + round_bit;
+    if (ue4m3_man > 7) {
+        ue4m3_man = 0;
+        ue4m3_exp++;
+        if (ue4m3_exp >= 15) {
+            return 0x7E;
+        }
+    }
+    return (uint8_t) ((ue4m3_exp << 3) | ue4m3_man);
+}
+
 /**
  * Converts brain16 to float32.
  *

ggml/src/ggml-metal/ggml-metal-context.m

@@ -47,7 +47,7 @@ struct ggml_metal {
     uint64_t fuse_cnt[GGML_OP_COUNT];
 
     // capture state
-    bool capture_next_compute;
+    int capture_compute;
     bool capture_started;
 
     id<MTLCaptureScope> capture_scope;
@@ -75,6 +75,10 @@ struct ggml_metal {
     // abort ggml_metal_graph_compute if callback returns true
     ggml_abort_callback abort_callback;
     void *              abort_callback_data;
+
+    // error state - set when a command buffer fails during synchronize
+    // once set, graph_compute will return GGML_STATUS_FAILED until the backend is recreated
+    bool has_error;
 };
 
 ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
@@ -154,10 +158,19 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
     GGML_LOG_INFO("%s: use concurrency    = %s\n", __func__, res->use_concurrency    ? "true" : "false");
     GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false");
 
-    res->capture_next_compute = false;
+    res->capture_compute = 0;
     res->capture_started = false;
     res->capture_scope = nil;
 
+    {
+        const char * val = getenv("GGML_METAL_CAPTURE_COMPUTE");
+        if (val) {
+            res->capture_compute = atoi(val);
+        }
+    }
+
+    res->has_error = false;
+
     res->gf = nil;
     res->encode_async = nil;
     for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
@@ -246,7 +259,8 @@ void ggml_metal_synchronize(ggml_metal_t ctx) {
                 if (status == MTLCommandBufferStatusError) {
                     GGML_LOG_ERROR("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);
                 }
-                GGML_ABORT("fatal error");
+                ctx->has_error = true;
+                return;
             }
         }
     }
@@ -262,7 +276,15 @@ void ggml_metal_synchronize(ggml_metal_t ctx) {
                 if (status == MTLCommandBufferStatusError) {
                     GGML_LOG_ERROR("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);
                 }
-                GGML_ABORT("fatal error");
+
+                // release this and all remaining command buffers before returning
+                for (size_t j = i; j < ctx->cmd_bufs_ext.count; ++j) {
+                    [ctx->cmd_bufs_ext[j] release];
+                }
+                [ctx->cmd_bufs_ext removeAllObjects];
+
+                ctx->has_error = true;
+                return;
             }
 
             [cmd_buf release];
@@ -414,6 +436,11 @@ bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, con
 }
 
 enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph * gf) {
+    if (ctx->has_error) {
+        GGML_LOG_ERROR("%s: backend is in error state from a previous command buffer failure - recreate the backend to recover\n", __func__);
+        return GGML_STATUS_FAILED;
+    }
+
     // number of nodes encoded by the main thread (empirically determined)
     const int n_main = MAX(64, 0.1*gf->n_nodes);
 
@@ -438,9 +465,13 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
 
         ctx->n_nodes_per_cb = (ctx->n_nodes_1 + ctx->n_cb - 1) / ctx->n_cb;
 
-        const bool use_capture = ctx->capture_next_compute;
+        if (ctx->capture_compute >= 0) {
+            ctx->capture_compute--;
+        }
+
+        const bool use_capture = ctx->capture_compute == 0;
         if (use_capture) {
-            ctx->capture_next_compute = false;
+            ctx->capture_compute = -1;
 
             // make sure all previous computations have finished before starting the capture
             if (ctx->cmd_buf_last) {
@@ -449,6 +480,10 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
             }
 
             if (!ctx->capture_started) {
+                NSString * path = [NSString stringWithFormat:@"/tmp/perf-metal-%d.gputrace", getpid()];
+
+                GGML_LOG_WARN("%s: capturing graph in %s\n", __func__, [path UTF8String]);
+
                 // create capture scope
                 id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
                 ctx->capture_scope = [[MTLCaptureManager sharedCaptureManager] newCaptureScopeWithDevice:device];
@@ -456,7 +491,7 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
                 MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new];
                 descriptor.captureObject = ctx->capture_scope;
                 descriptor.destination = MTLCaptureDestinationGPUTraceDocument;
-                descriptor.outputURL = [NSURL fileURLWithPath:[NSString stringWithFormat:@"/tmp/perf-metal.gputrace"]];
+                descriptor.outputURL = [NSURL fileURLWithPath:path];
 
                 NSError * error = nil;
                 if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) {
@@ -519,7 +554,7 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
 
         // enter here only when capturing in order to wait for all computation to finish
         // otherwise, we leave the graph to compute asynchronously
-        if (!use_capture && ctx->capture_started) {
+        if (use_capture && ctx->capture_started) {
             // wait for completion and check status of each command buffer
             // needed to detect if the device ran out-of-memory for example (#1881)
             {
@@ -571,6 +606,8 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
 
             [ctx->capture_scope endScope];
             [[MTLCaptureManager sharedCaptureManager] stopCapture];
+
+            ctx->capture_started = false;
         }
     }
 
@@ -663,7 +700,7 @@ void ggml_metal_set_n_cb(ggml_metal_t ctx, int n_cb) {
             idx_end,
             ctx->use_fusion,
             ctx->use_concurrency,
-            ctx->capture_next_compute,
+            ctx->capture_compute,
             ctx->debug_graph,
             ctx->debug_fusion);
 
@@ -698,5 +735,5 @@ bool ggml_metal_supports_family(ggml_metal_t ctx, int family) {
 }
 
 void ggml_metal_capture_next_compute(ggml_metal_t ctx) {
-    ctx->capture_next_compute = true;
+    ctx->capture_compute = 1;
 }

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -577,6 +577,41 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv(ggml_metal_
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_gated_delta_net(ggml_metal_library_t lib, const ggml_tensor * op) {
+    char base[256];
+    char name[256];
+
+    // v is src[2], dimensions: S_v = ne[0], H = ne[1]
+    const int ne20 = op->src[2]->ne[0]; // S_v
+    const int ne21 = op->src[2]->ne[1]; // H
+    const int ne30 = op->src[3]->ne[0]; // G
+
+    const int nsg = op->src[2]->ne[0]/32;
+
+    GGML_ASSERT(op->src[5]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->ne[0] == ne20 * ne21);
+    GGML_ASSERT(ne20 % 32 == 0);
+
+    snprintf(base, 256, "kernel_gated_delta_net_%s_%d", ggml_type_name(op->src[0]->type), nsg);
+    snprintf(name, 256, "%s_ne20=%d_ne30=%d", base, ne20, ne30);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, ne20, FC_GATED_DELTA_NET + 0);
+        ggml_metal_cv_set_int16(cv, ne30, FC_GATED_DELTA_NET + 1);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    res.nsg = nsg;
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri(ggml_metal_library_t lib, const ggml_tensor * op) {
     char base[256];
     char name[256];
@@ -1435,10 +1470,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(ggml_metal_l
 
     const bool is_c4 = (op->src[0]->ne[0] % 4 == 0) && (op->src[1]->ne[0] % 4 == 0);
 
+    const bool is_cb = op->src[0]->ne[0] != op->src[1]->ne[0];
     const bool is_rb = ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]) && (ggml_nrows(op->src[1]) == 1) && ggml_nelements(op) < 65536;
 
     snprintf(base, 256, "kernel_bin_fuse_%s_%s_%s%s", t0_str, t1_str, t_str, is_c4 ? "_4" : "");
-    snprintf(name, 256, "%s_op=%d_nf=%d_rb=%d", base, op_num, n_fuse, is_rb);
+    snprintf(name, 256, "%s_op=%d_nf=%d_rb=%d_cb=%d", base, op_num, n_fuse, is_rb, is_cb);
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
     if (!res.pipeline) {
@@ -1447,6 +1483,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(ggml_metal_l
         ggml_metal_cv_set_int16(cv, op_num, FC_BIN + 0);
         ggml_metal_cv_set_int16(cv, n_fuse, FC_BIN + 1);
         ggml_metal_cv_set_bool (cv, is_rb,  FC_BIN + 2);
+        ggml_metal_cv_set_bool (cv, is_cb,  FC_BIN + 3);
 
         res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
 
@@ -1717,12 +1754,29 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale(ggml_met
     char base[256];
     char name[256];
 
-    snprintf(base, 256, "kernel_upscale_%s", ggml_type_name(op->src[0]->type));
-    snprintf(name, 256, "%s", base);
+    const int32_t mode_flags = ggml_get_op_params_i32(op, 0);
+    const ggml_scale_mode mode = (ggml_scale_mode) (mode_flags & 0xFF);
+
+    const bool antialias = (mode_flags & GGML_SCALE_FLAG_ANTIALIAS);
+
+    if (mode == GGML_SCALE_MODE_BILINEAR) {
+        snprintf(base, 256, "kernel_upscale_bilinear_%s", ggml_type_name(op->src[0]->type));
+    } else if (mode == GGML_SCALE_MODE_BICUBIC) {
+        snprintf(base, 256, "kernel_upscale_bicubic_%s", ggml_type_name(op->src[0]->type));
+    } else {
+        snprintf(base, 256, "kernel_upscale_nearest_%s", ggml_type_name(op->src[0]->type));
+    }
+    snprintf(name, 256, "%s_aa=%d", base, antialias);
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
     if (!res.pipeline) {
-        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_bool(cv, antialias, FC_UPSCALE + 0);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
     }
 
     return res;

ggml/src/ggml-metal/ggml-metal-device.h

@@ -125,6 +125,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_gated_delta_net   (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext        (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm            (ggml_metal_library_t lib, const struct ggml_tensor * op);

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1108,7 +1108,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                    op->type == GGML_TYPE_F32 &&
                    (op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
         case GGML_OP_UPSCALE:
-            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST && !(op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS);
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_POOL_1D:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_POOL_2D:
@@ -1155,10 +1155,12 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
             return true;
+        case GGML_OP_GATED_DELTA_NET:
+            return op->src[2]->ne[0] % 32 == 0;
         case GGML_OP_SOLVE_TRI:
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
-            return has_simdgroup_reduction;
+            return has_simdgroup_reduction && op->src[0]->type != GGML_TYPE_NVFP4;
         case GGML_OP_SET:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
@@ -1216,7 +1218,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 };
             }
         case GGML_OP_GET_ROWS:
-            return true;
+            return op->src[0]->type != GGML_TYPE_NVFP4;
         case GGML_OP_SET_ROWS:
             {
                 if (op->src[0]->type != GGML_TYPE_F32) {

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -35,7 +35,7 @@
 #define N_R0_Q4_K 2
 #define N_SG_Q4_K 2
 
-#define N_R0_Q5_K 2
+#define N_R0_Q5_K 1
 #define N_SG_Q5_K 2
 
 #define N_R0_Q6_K 2
@@ -83,6 +83,8 @@
 #define FC_UNARY                       1200
 #define FC_BIN                         1300
 #define FC_SUM_ROWS                    1400
+#define FC_UPSCALE                     1500
+#define FC_GATED_DELTA_NET             1600
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPSG 8
@@ -792,6 +794,44 @@ typedef struct {
     uint64_t nb0;
 } ggml_metal_kargs_ssm_scan;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne20;
+    int32_t  ne21;
+    int32_t  ne22;
+    int32_t  ne23;
+    uint64_t nb20;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
+    int32_t  ns02;
+    int32_t  ns12;
+    int32_t  ns22;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_gated_delta_net;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -890,6 +930,7 @@ typedef struct {
     float    sf1;
     float    sf2;
     float    sf3;
+    float    poffs;
 } ggml_metal_kargs_upscale;
 
 typedef struct {

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -333,6 +333,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_rwkv(ctx, idx);
             } break;
+        case GGML_OP_GATED_DELTA_NET:
+            {
+                n_fuse = ggml_metal_op_gated_delta_net(ctx, idx);
+            } break;
         case GGML_OP_SOLVE_TRI:
             {
                 n_fuse = ggml_metal_op_solve_tri(ctx, idx);
@@ -1562,6 +1566,81 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_gated_delta_net(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    auto pipeline = ggml_metal_library_get_pipeline_gated_delta_net(lib, op);
+
+    int ida = 0;
+
+    ggml_metal_kargs_gated_delta_net args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne10 =*/ ne10,
+        /*.ne11 =*/ ne11,
+        /*.ne12 =*/ ne12,
+        /*.ne13 =*/ ne13,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.ne20 =*/ ne20,
+        /*.ne21 =*/ ne21,
+        /*.ne22 =*/ ne22,
+        /*.ne23 =*/ ne23,
+        /*.nb20 =*/ nb20,
+        /*.nb21 =*/ nb21,
+        /*.nb22 =*/ nb22,
+        /*.nb23 =*/ nb23,
+        /*.ns02 =*/ (int32_t) (nb02/sizeof(float)),
+        /*.ns12 =*/ (int32_t) (nb12/sizeof(float)),
+        /*.ns22 =*/ (int32_t) (nb22/sizeof(float)),
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.ne3  =*/ ne3,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+    };
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args),                  ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), ida++); // q
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), ida++); // k
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), ida++); // v
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[3]), ida++); // gate
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[4]), ida++); // beta
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[5]), ida++); // state
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         ida++); // dst
+
+    const int nsg = pipeline.nsg;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, op->src[2]->ne[0]/nsg, op->src[2]->ne[1], op->src[2]->ne[3], 32, nsg, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -1963,6 +2042,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
           (
            op->src[0]->type == GGML_TYPE_F32  || // TODO: helper function
            op->src[0]->type == GGML_TYPE_F16  ||
+           op->src[0]->type == GGML_TYPE_BF16 ||
            op->src[0]->type == GGML_TYPE_Q4_0 ||
            op->src[0]->type == GGML_TYPE_Q4_1 ||
            op->src[0]->type == GGML_TYPE_Q5_0 ||
@@ -1977,6 +2057,8 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
            op->src[0]->type == GGML_TYPE_Q4_K ||
            op->src[0]->type == GGML_TYPE_Q5_K ||
            op->src[0]->type == GGML_TYPE_Q6_K ||
+           op->src[0]->type == GGML_TYPE_Q2_K ||
+           op->src[0]->type == GGML_TYPE_Q3_K ||
            false) && (ne11 >= 4 && ne11 <= 8)
          )
         )
@@ -3098,9 +3180,7 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);
 
     if (pipeline.cnt) {
-        const int n = pipeline.c4 ? ggml_nelements(op)/4 : ggml_nelements(op);
-
-        ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+        ggml_metal_encoder_dispatch_threadgroups(enc, args.ne0, ggml_nrows(op), 1, 1, 1, 1);
     } else {
         const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
@@ -3729,32 +3809,43 @@ int ggml_metal_op_upscale(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    const float sf0 = (float)ne0/op->src[0]->ne[0];
-    const float sf1 = (float)ne1/op->src[0]->ne[1];
-    const float sf2 = (float)ne2/op->src[0]->ne[2];
-    const float sf3 = (float)ne3/op->src[0]->ne[3];
+    float sf0 = (float)ne0/op->src[0]->ne[0];
+    float sf1 = (float)ne1/op->src[0]->ne[1];
+    float sf2 = (float)ne2/op->src[0]->ne[2];
+    float sf3 = (float)ne3/op->src[0]->ne[3];
+
+    const int32_t mode_flags = ggml_get_op_params_i32(op, 0);
+
+    float poffs = 0.5f;
+
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        poffs = 0.0f;
+        sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+        sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+    }
 
     ggml_metal_kargs_upscale args = {
-        /*.ne00 =*/ ne00,
-        /*.ne01 =*/ ne01,
-        /*.ne02 =*/ ne02,
-        /*.ne03 =*/ ne03,
-        /*.nb00 =*/ nb00,
-        /*.nb01 =*/ nb01,
-        /*.nb02 =*/ nb02,
-        /*.nb03 =*/ nb03,
-        /*.ne0 =*/ ne0,
-        /*.ne1 =*/ ne1,
-        /*.ne2 =*/ ne2,
-        /*.ne3 =*/ ne3,
-        /*.nb0 =*/ nb0,
-        /*.nb1 =*/ nb1,
-        /*.nb2 =*/ nb2,
-        /*.nb3 =*/ nb3,
-        /*.sf0 =*/ sf0,
-        /*.sf1 =*/ sf1,
-        /*.sf2 =*/ sf2,
-        /*.sf3 =*/ sf3
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.nb0   =*/ nb0,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
+        /*.sf0   =*/ sf0,
+        /*.sf1   =*/ sf1,
+        /*.sf2   =*/ sf2,
+        /*.sf3   =*/ sf3,
+        /*.poffs =*/ poffs,
     };
 
     auto pipeline = ggml_metal_library_get_pipeline_upscale(lib, op);