sycl : fix wrong variable check by assert (#20903)

* fix wrong variable check by assert

* use GGML api

.github/workflows/copilot-setup-steps.yml

@@ -54,4 +54,3 @@ jobs:
           python3 -m venv .venv
           source .venv/bin/activate
           pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
-          pip install flake8 pyright pre-commit

.github/workflows/gguf-publish.yml

@@ -28,11 +28,11 @@ jobs:
     - name: Set up Python
       uses: actions/setup-python@v6
       with:
-        python-version: '3.9.x'
+        python-version: '3.11'
     - name: Install dependencies
       run: |
         cd gguf-py
-        python -m pip install poetry
+        python -m pip install poetry==2.3.2
         poetry install
 
     - name: Build package

README.md

@@ -17,6 +17,7 @@ LLM inference in C/C++
 
 ## Hot topics
 
+- **Hugging Face cache migration: models downloaded with `-hf` are now stored in the standard Hugging Face cache directory, enabling sharing with other HF tools.**
 - **[guide : using the new WebUI of llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/16938)**
 - [guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
 - [[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)
@@ -241,7 +242,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 <details>
 <summary>Tools</summary>
 
-- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from Hugging Face Hub and convert them to GGML
 - [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
 - [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
@@ -300,13 +301,13 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
 - [Trending](https://huggingface.co/models?library=gguf&sort=trending)
 - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
 
 ```sh
 llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
 ```
 
-By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. For example, you may opt to downloading model checkpoints from ModelScope or other model sharing communities by setting the environment variable, e.g. `MODEL_ENDPOINT=https://www.modelscope.cn/`.
+By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. The `MODEL_ENDPOINT` must point to a Hugging Face compatible API endpoint.
 
 After downloading a model, use the CLI tools to run it locally - see below.
 

common/download.cpp

@@ -460,9 +460,9 @@ static gguf_split_info get_gguf_split_info(const std::string & path) {
     int count = 1;
 
     if (std::regex_match(prefix, m, re_split)) {
-        prefix = m[1].str();
         index = std::stoi(m[2].str());
         count = std::stoi(m[3].str());
+        prefix = m[1].str();
     }
 
     std::string tag;

common/hf-cache.cpp

@@ -590,6 +590,8 @@ void migrate_old_cache_to_hf_cache(const std::string & token, bool offline) {
         return; // -hf is not going to work
     }
 
+    bool warned = false;
+
     for (const auto & entry : fs::directory_iterator(old_cache)) {
         if (!entry.is_regular_file()) {
             continue;
@@ -601,6 +603,19 @@ void migrate_old_cache_to_hf_cache(const std::string & token, bool offline) {
             continue;
         }
 
+        if (!warned) {
+            warned = true;
+            LOG_WRN("================================================================================\n"
+                    "WARNING: Migrating cache to HuggingFace cache directory\n"
+                    "  Old cache: %s\n"
+                    "  New cache: %s\n"
+                    "This one-time migration moves models previously downloaded with -hf\n"
+                    "from the legacy llama.cpp cache to the standard HuggingFace cache.\n"
+                    "Models downloaded with --model-url are not affected.\n"
+                    "================================================================================\n",
+                    old_cache.string().c_str(), get_cache_directory().string().c_str());
+        }
+
         auto repo_id = owner + "/" + repo;
         auto files = get_repo_files(repo_id, token);
 

convert_hf_to_gguf.py

@@ -4572,7 +4572,7 @@ class Qwen2MoeModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("Qwen3ForCausalLM")
+@ModelBase.register("Qwen3ForCausalLM", "Qwen3Model")
 class Qwen3Model(Qwen2Model):
     model_arch = gguf.MODEL_ARCH.QWEN3
 

docs/backend/OPENVINO.md

@@ -1,6 +1,9 @@
 # OpenVINO Backend for llama.cpp
-[OpenVINO](https://docs.openvino.ai/) is an open-source toolkit for optimizing and deploying high-performance AI inference, specifically designed for Intel hardware, including CPUs, GPUs, and NPUs, in the cloud, on-premises, and on the edge.
-This document describes the [OpenVINO backend for llama.cpp](../../src/ggml-openvino), which enables hardware-accelerated inference on **Intel® CPUs, GPUs, and NPUs** while remaining compatible with the existing **GGUF model ecosystem**. The backend translates GGML compute graphs into OpenVINO graphs and leverages graph compilation, kernel fusion, and device-specific optimizations to improve inference performance on supported Intel hardware.
+
+> [!NOTE]
+> Performance and memory optimizations, accuracy validation, broader quantization coverage, broader operator and model support are work in progress.
+
+[OpenVINO](https://docs.openvino.ai/) is an open-source toolkit for optimizing and deploying high-performance AI inference, specifically designed for Intel hardware, including CPUs, GPUs, and NPUs, in the cloud, on-premises, and on the edge. [OpenVINO backend for llama.cpp](../../src/ggml-openvino) enables hardware-accelerated inference on **Intel® CPUs, GPUs, and NPUs** while remaining compatible with the existing **GGUF model ecosystem**. The backend translates GGML compute graphs into OpenVINO graphs and leverages graph compilation, kernel fusion, and device-specific optimizations to improve inference performance on supported Intel hardware.
 
 The OpenVINO backend is implemented in `ggml/src/ggml-openvino` and provides a translation layer for core GGML operations. The OpenVINO backend replaces the standard GGML graph execution path with Intel's OpenVINO inference engine. This approach allows the same GGUF model file to run on Intel CPUs, Intel GPUs (integrated and discrete), and Intel NPUs without changes to the model or the rest of the llama.cpp stack. When a `ggml_cgraph` is dispatched to OpenVINO backend, it:
 
@@ -179,31 +182,73 @@ curl -L https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF/resolve/main/L
 
 When using the OpenVINO backend, the first inference token may have slightly higher latency due to on-the-fly conversion to the OpenVINO graph. Subsequent tokens and runs will be faster.
 
+> [!NOTE]
+> Default context size is set to the model training context, which may be very large. For example, 131072 for Llama 3.2 1B, which may result in lower performance, especially on edge/laptop devices. Use `-c` to limit context size in supported llama.cpp tools for better performance. For example, `-c 512`.
+
 ```bash
 # If device is unset or unavailable, defaults to CPU.
 # If the system has multiple GPUs, use GPU.0 or GPU.1 to explicitly target a specific GPU.
 
 # Linux
 export GGML_OPENVINO_DEVICE=GPU
+# Enable stateful execution with GPU device to avoid known stateless execution failures.
+export GGML_OPENVINO_STATEFUL_EXECUTION=1
 # To run llama-simple:
 ./build/ReleaseOV/bin/llama-simple -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -n 50 "The story of AI is "
 # To run in chat mode:
-./build/ReleaseOV/bin/llama-cli -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf
+./build/ReleaseOV/bin/llama-cli -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -c 1024
+# To run llama-bench, -fa 1 is needed
+GGML_OPENVINO_STATEFUL_EXECUTION=1 GGML_OPENVINO_DEVICE=GPU ./build/ReleaseOV/bin/llama-bench -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -fa 1
+
+# NPU: keep context small to avoid failures from very large model context windows.
+export GGML_OPENVINO_DEVICE=NPU
+./build/ReleaseOV/bin/llama-cli -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -c 512
 
 # Windows Command Line
 set GGML_OPENVINO_DEVICE=GPU
+# Enable stateful execution with GPU device to avoid known stateless execution failures.
+set GGML_OPENVINO_STATEFUL_EXECUTION=1
 # Windows PowerShell
 $env:GGML_OPENVINO_DEVICE = "GPU"
+$env:GGML_OPENVINO_STATEFUL_EXECUTION = "1"
 
 # To run llama-simple
 build\ReleaseOV\bin\llama-simple.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -n 50 "The story of AI is "
 # To run in chat mode:
-build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf"
+build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -c 1024
+# To run llama-bench, -fa 1 is needed
+build\ReleaseOV\bin\llama-bench.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -fa 1
 
+# NPU: keep context small to avoid failures from very large model context windows.
+# Windows Command Line
+set GGML_OPENVINO_DEVICE=NPU
+# Windows PowerShell
+$env:GGML_OPENVINO_DEVICE = "NPU"
+build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -c 512
 ```
 > [!NOTE]
 > On systems with multiple GPUs, use `GPU.0` or `GPU.1` to explicitly target specific GPU. See [OpenVINO GPU Device](https://docs.openvino.ai/2026/openvino-workflow/running-inference/inference-devices-and-modes/gpu-device.html) for more details.
 
+### Known Issues and Current Workarounds
+
+- GPU stateless execution is currently affected by a known issue.
+  - Workaround: set `GGML_OPENVINO_STATEFUL_EXECUTION=1` when using GPU device.
+- NPU failures can happen when context size is too large. Recent llama.cpp behavior may resolve context size to the model training context (for example, 131072 for Llama 3.2 1B), which is too large for current NPU usage and can also stress laptop CPU/GPU on larger models. To inspect the selected context size, run `llama-cli` or `llama-server` with `-lv 3`.
+  - Workaround: explicitly set context size, for ex. `-c 1024` for NPU runs. Performance will be better with lower context size.
+- Additional NPU limitations:
+  - Model caching is not yet supported.
+  - `llama-server -np > 1` (multiple parallel sequences) is not supported.
+  - `llama-perplexity` is only supported with `-b 512` or smaller.
+- `--context-shift` with `llama-cli` is currently not supported with OpenVINO backend across CPU, GPU, and NPU devices.
+- Encoder models (embedding, reranking) are not supported with the current OpenVINO backend implementation.
+- `-fa 1` is required when running llama-bench with the OpenVINO backend.
+  - `GGML_OPENVINO_STATEFUL_EXECUTION=1 GGML_OPENVINO_DEVICE=GPU ./llama-bench -fa 1`
+- `llama-server` with OpenVINO backend supports only one chat session/thread, when `GGML_OPENVINO_STATEFUL_EXECUTION=1` is enabled.
+- For Intel GPU, NPU detection in containers, GPU, NPU user-space drivers/libraries must be present inside the image. We will include in a future PR. Until then, you can use this reference Dockerfile: [openvino.Dockerfile](https://github.com/ravi9/llama.cpp/blob/ov-docker-update/.devops/openvino.Dockerfile)
+
+> [!NOTE]
+> The OpenVINO backend is actively under development. Fixes are underway, and this document will continue to be updated as issues are resolved.
+
 
 ### Docker Build
 
@@ -229,31 +274,42 @@ docker build --build-arg http_proxy=$http_proxy --build-arg https_proxy=$https_p
 Run llama.cpp with OpenVINO backend Docker container.
 Save sample models in `~/models` as [shown above](#3-download-sample-model). It will be mounted to the container in the examples below.
 
+> [!NOTE]
+> Intel GPU, NPU detection in containers will be included in a future PR. Until then, you can use this reference Dockerfile: [openvino.Dockerfile](https://github.com/ravi9/llama.cpp/blob/ov-docker-update/.devops/openvino.Dockerfile).
+
 ```bash
 #  Run Docker container
-docker run --rm -it -v ~/models:/models llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
+docker run --rm -it -v ~/models:/models llama-openvino:light --no-warmup -c 1024 -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
 
 # With Intel GPU access (iGPU or dGPU)
 docker run --rm -it -v ~/models:/models \
 --device=/dev/dri --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) -u $(id -u):$(id -g) \
-llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
+--env=GGML_OPENVINO_DEVICE=GPU --env=GGML_OPENVINO_STATEFUL_EXECUTION=1 \
+llama-openvino:light --no-warmup -c 1024 -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
 
 # With Intel NPU access
-docker run --rm -it --env GGML_OPENVINO_DEVICE=NPU -v ~/models:/models \
+docker run --rm -it -v ~/models:/models \
 --device=/dev/accel --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) -u $(id -u):$(id -g) \
-llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
+--env=GGML_OPENVINO_DEVICE=NPU \
+llama-openvino:light --no-warmup -c 1024 -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
 ```
 
-Run Llama.cpp Server with OpenVINO Backend:
+Run Llama.cpp Server with OpenVINO Backend.
+> [!NOTE]
+> `llama-server` with OpenVINO backend supports only one chat session/thread, when `GGML_OPENVINO_STATEFUL_EXECUTION=1` is enabled.
+
 ```bash
 # Run the Server Docker container
-docker run --rm -it -p 8080:8080 -v ~/models:/models llama-openvino:server --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
-
-# In a NEW terminal, test the server with curl
+docker run --rm -it -p 8080:8080 -v ~/models:/models llama-openvino:server --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf -c 1024
+# Or Using llama-server executable
+./build/ReleaseOV/bin/llama-server -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf --port 8080 -c 1024
 
 # If you are behind a proxy, make sure to set NO_PROXY to avoid proxy for localhost
 export NO_PROXY=localhost,127.0.0.1
 
+# Option 1: Open your browser to http://localhost:8080 to access the web UI for the llama.cpp server.
+# Option 2: In a NEW terminal, test the server with curl
+
 # Test health endpoint
 curl -f http://localhost:8080/health
 
@@ -295,6 +351,7 @@ The OpenVINO backend can be configured using the following environment variables
 export GGML_OPENVINO_CACHE_DIR=/tmp/ov_cache
 export GGML_OPENVINO_PROFILING=1
 export GGML_OPENVINO_DEVICE=GPU
+export GGML_OPENVINO_STATEFUL_EXECUTION=1
 
 ./build/ReleaseOV/bin/llama-simple -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -n 50 "The story of AI is "
 
@@ -302,38 +359,27 @@ export GGML_OPENVINO_DEVICE=GPU
 set GGML_OPENVINO_CACHE_DIR=C:\tmp\ov_cache
 set GGML_OPENVINO_PROFILING=1
 set GGML_OPENVINO_DEVICE=GPU
+set GGML_OPENVINO_STATEFUL_EXECUTION=1
 
 # Windows PowerShell
 $env:GGML_OPENVINO_CACHE_DIR = "C:\tmp\ov_cache"
 $env:GGML_OPENVINO_PROFILING = "1"
 $env:GGML_OPENVINO_DEVICE = "GPU"
+$env:GGML_OPENVINO_STATEFUL_EXECUTION = "1"
 
 build\ReleaseOV\bin\llama-simple.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -n 50 "The story of AI is "
 
 ```
 
-#### llama-bench
-
-```bash
-# -fa 1 is required when running llama-bench with the OpenVINO backend.
-GGML_OPENVINO_DEVICE=GPU ./llama-bench -fa 1
-```
-
-### NPU Notes
-
-- Model caching is not yet supported
-- Does not support llama-server -np > 1 (multiple parallel sequences)
-- Only supports llama-perplexity -b 512 or smaller
-
 ## Llama.cpp Tools
 
 The following tools work with the OpenVINO backend on CPU, GPU, NPU:
-- llama-simple
-- llama-run
-- llama-cli
-- llama-server
 - llama-bench
+- llama-cli
+- llama-completion
 - llama-perplexity
+- llama-server
+- llama-simple
 
 ## Work in Progress
 

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

@@ -246,6 +246,10 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary(ggml_metal
                 case GGML_UNARY_OP_EXP:         op_num = OP_UNARY_NUM_EXP;         break;
                 case GGML_UNARY_OP_SOFTPLUS:    op_num = OP_UNARY_NUM_SOFTPLUS;    break;
                 case GGML_UNARY_OP_EXPM1:       op_num = OP_UNARY_NUM_EXPM1;       break;
+                case GGML_UNARY_OP_FLOOR:       op_num = OP_UNARY_NUM_FLOOR;       break;
+                case GGML_UNARY_OP_CEIL:        op_num = OP_UNARY_NUM_CEIL;        break;
+                case GGML_UNARY_OP_ROUND:       op_num = OP_UNARY_NUM_ROUND;       break;
+                case GGML_UNARY_OP_TRUNC:       op_num = OP_UNARY_NUM_TRUNC;       break;
                 default: GGML_ABORT("fatal error");
             } break;
         default: GGML_ABORT("fatal error");

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

@@ -1039,6 +1039,10 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 case GGML_UNARY_OP_EXP:
                 case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_EXPM1:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     return ggml_is_contiguous_rows(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
                 default:
                     return false;

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

@@ -120,6 +120,10 @@
 #define OP_UNARY_NUM_EXP         114
 #define OP_UNARY_NUM_SOFTPLUS    115
 #define OP_UNARY_NUM_EXPM1       116
+#define OP_UNARY_NUM_FLOOR       117
+#define OP_UNARY_NUM_CEIL        118
+#define OP_UNARY_NUM_ROUND       119
+#define OP_UNARY_NUM_TRUNC       120
 
 #define OP_SUM_ROWS_NUM_SUM_ROWS 10
 #define OP_SUM_ROWS_NUM_MEAN     11

ggml/src/ggml-metal/ggml-metal.metal

@@ -1094,6 +1094,22 @@ kernel void kernel_unary_impl(
             // TODO: precise implementation
             dst_ptr[i0] = (T) (exp(x) - 1);
         }
+
+        if (FC_OP == OP_UNARY_NUM_FLOOR) {
+            dst_ptr[i0] = (T) floor(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CEIL) {
+            dst_ptr[i0] = (T) ceil(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ROUND) {
+            dst_ptr[i0] = (T) round(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TRUNC) {
+            dst_ptr[i0] = (T) trunc(x);
+        }
     }
 
 #undef FC_OP

ggml/src/ggml-sycl/add-id.cpp

@@ -56,7 +56,7 @@ void ggml_sycl_add_id(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
   float* dst_d = (float*)dst->data;
 
   const unsigned int max_work_group_size = ggml_sycl_info().max_work_group_sizes[ctx.device];
-  assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
+  GGML_ASSERT(max_work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
 
   int threads = std::min((unsigned int)ne00, max_work_group_size);  // cols
 

requirements/requirements-pydantic.txt

@@ -1,3 +1,3 @@
 docstring_parser~=0.15
 pydantic~=2.11.7
-requests
+requests~=2.32.3

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "refs/tags/v0.38.0"
+HTTPLIB_VERSION = "refs/tags/v0.39.0"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",

src/llama-arch.cpp

@@ -2564,7 +2564,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TOKEN_EMBD,                 {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_POS_EMBD,                   {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_TOKEN_TYPES,                {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
-    {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_INPUT, GGML_OP_MUL}},
+    {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},  // do the norms on the first layer (not the input layer)
     {LLM_TENSOR_OUTPUT,                     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@@ -2725,7 +2725,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_LAUREL_POST_NORM,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     // this tensor is loaded for T5, but never used
     {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
-    {LLM_TENSOR_CONV1D,                     {LLM_TENSOR_LAYER_INPUT,     GGML_OP_IM2COL}},
+    {LLM_TENSOR_CONV1D,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}},
     {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_POS_NET_NORM2,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},

src/llama-context.cpp

@@ -342,14 +342,6 @@ llama_context::llama_context(
 
         if (cparams.pipeline_parallel) {
             LLAMA_LOG_INFO("%s: pipeline parallelism enabled\n", __func__);
-
-            if (!graph_reuse_disable) {
-                // TODO: figure out a way to make graph reuse work with pipeline parallelism
-                // ref: https://github.com/ggml-org/llama.cpp/pull/20463
-                LLAMA_LOG_WARN("%s: graph reuse is currently not compatible with pipeline parallelism - disabling\n", __func__);
-
-                graph_reuse_disable = true;
-            }
         }
 
         sched_reserve();
@@ -1189,6 +1181,13 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
     if (!graph_reuse_disable && res->can_reuse(gparams)) {
         //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);
 
+        // with pipeline parallelism, the previous graph_compute_async may still be running
+        // on the GPU. we must synchronize before set_inputs to avoid overwriting input tensors
+        // that the previous compute is still reading.
+        if (cparams.pipeline_parallel) {
+            ggml_backend_sched_synchronize(sched.get());
+        }
+
         n_reused++;
     } else {
         res->reset();

src/llama-model.cpp

@@ -3217,8 +3217,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
                     }
 
-                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
-                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd}, 0);
+                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
+                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -3265,7 +3265,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_MODERN_BERT:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
-                    tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
+                    tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
 
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
 
@@ -3348,8 +3348,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     tok_embd  = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0); // word_embeddings
                     type_embd = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0); // token_type_embeddings
 
-                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0); // LayerNorm
-                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd}, 0); //LayerNorm bias
+                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0); // LayerNorm
+                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0); // LayerNorm bias
 
                     cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, 1}, TENSOR_NOT_REQUIRED);
                     cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {1},         TENSOR_NOT_REQUIRED);
@@ -3400,8 +3400,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_BLOOM:
                 {
                     tok_embd   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
-                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
-                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd}, 0);
+                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
+                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
 
                     // output
                     output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
@@ -5780,8 +5780,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     // Block 0, LN0
-                    tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
-                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0);
+                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
+                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
@@ -5895,8 +5895,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     // Block 0, LN0
-                    tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
-                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0);
+                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
+                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
@@ -6067,8 +6067,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {hparams.n_embd, n_vocab}, 0);
 
-                    conv1d   = create_tensor(tn(LLM_TENSOR_CONV1D, "weight"), {7, hparams.n_embd, hparams.posnet.n_embd}, 0);
-                    conv1d_b = create_tensor(tn(LLM_TENSOR_CONV1D, "bias"),   {1, hparams.posnet.n_embd}, 0);
+                    conv1d   = create_tensor(tn(LLM_TENSOR_CONV1D, "weight", 0), {7, hparams.n_embd, hparams.posnet.n_embd}, 0);
+                    conv1d_b = create_tensor(tn(LLM_TENSOR_CONV1D, "bias",   0), {1, hparams.posnet.n_embd}, 0);
 
                     // posnet
                     {
@@ -6133,8 +6133,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     GGML_ASSERT(hparams.posnet.n_embd == hparams.convnext.n_embd);
 
-                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {hparams.posnet.n_embd}, 0);
-                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {hparams.posnet.n_embd}, 0);
+                    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {hparams.posnet.n_embd}, 0);
+                    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {hparams.posnet.n_embd}, 0);
 
                     // convnext
                     {

src/models/bert.cpp

@@ -28,8 +28,8 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
     cb(inpL, "inp_embd", -1);
 
     // embed layer norm
-    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
-    cb(inpL, "inp_norm", -1);
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, 0);
+    cb(inpL, "inp_norm", 0);
 
     auto * inp_attn = build_attn_inp_no_cache();
 

src/models/bloom.cpp

@@ -16,8 +16,8 @@ llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_para
     inpL = build_norm(inpL,
             model.tok_norm,
             model.tok_norm_b,
-            LLM_NORM, -1);
-    cb(inpL, "inp_norm", -1);
+            LLM_NORM, 0);
+    cb(inpL, "inp_norm", 0);
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 

src/models/modern-bert.cpp

@@ -15,8 +15,8 @@ llm_build_modern_bert::llm_build_modern_bert(const llama_model & model, const ll
     cb(inpL, "inp_embd", -1);
 
     // embed layer norm
-    inpL = build_norm(inpL, model.tok_norm, nullptr, LLM_NORM, -1);
-    cb(inpL, "inp_norm", -1);
+    inpL = build_norm(inpL, model.tok_norm, nullptr, LLM_NORM, 0);
+    cb(inpL, "inp_norm", 0);
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 

src/models/rwkv6.cpp

@@ -8,7 +8,7 @@ llm_build_rwkv6::llm_build_rwkv6(const llama_model & model, const llm_graph_para
     ggml_tensor * inpL;
 
     inpL = build_inp_embd(model.tok_embd);
-    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, 0);
 
     auto * rs_inp = build_rs_inp();
 

src/models/rwkv7.cpp

@@ -9,7 +9,7 @@ llm_build_rwkv7::llm_build_rwkv7(const llama_model & model, const llm_graph_para
     ggml_tensor * v_first = nullptr;
 
     inpL = build_inp_embd(model.tok_embd);
-    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, 0);
 
     auto * rs_inp = build_rs_inp();
 

src/models/wavtokenizer-dec.cpp

@@ -93,7 +93,7 @@ llm_build_wavtokenizer_dec::llm_build_wavtokenizer_dec(const llama_model & model
     cur = build_norm(cur,
             model.tok_norm,
             model.tok_norm_b,
-            LLM_NORM, -1);
+            LLM_NORM, 0);
 
     cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
 

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormTextarea.svelte

@@ -26,6 +26,7 @@
 
 	onMount(() => {
 		if (textareaElement) {
+			autoResizeTextarea(textareaElement);
 			textareaElement.focus();
 		}
 	});
@@ -50,8 +51,9 @@
 	<textarea
 		bind:this={textareaElement}
 		bind:value
-		class="text-md max-h-32 min-h-12 w-full resize-none border-0 bg-transparent p-0 leading-6 outline-none placeholder:text-muted-foreground focus-visible:ring-0 focus-visible:ring-offset-0"
+		class="text-md min-h-12 w-full resize-none border-0 bg-transparent p-0 leading-6 outline-none placeholder:text-muted-foreground focus-visible:ring-0 focus-visible:ring-offset-0"
 		class:cursor-not-allowed={disabled}
+		style="max-height: var(--max-message-height);"
 		{disabled}
 		onkeydown={onKeydown}
 		oninput={(event) => {

vendor/cpp-httplib/httplib.cpp

@@ -142,6 +142,12 @@ SSEClient &SSEClient::set_max_reconnect_attempts(int n) {
   return *this;
 }
 
+SSEClient &SSEClient::set_headers(const Headers &headers) {
+  std::lock_guard<std::mutex> lock(headers_mutex_);
+  headers_ = headers;
+  return *this;
+}
+
 bool SSEClient::is_connected() const { return connected_.load(); }
 
 const std::string &SSEClient::last_event_id() const {
@@ -220,7 +226,11 @@ void SSEClient::run_event_loop() {
 
   while (running_.load()) {
     // Build headers, including Last-Event-ID if we have one
-    auto request_headers = headers_;
+    Headers request_headers;
+    {
+      std::lock_guard<std::mutex> lock(headers_mutex_);
+      request_headers = headers_;
+    }
     if (!last_event_id_.empty()) {
       request_headers.emplace("Last-Event-ID", last_event_id_);
     }
@@ -239,19 +249,19 @@ void SSEClient::run_event_loop() {
       continue;
     }
 
-    if (result.status() != 200) {
+    if (result.status() != StatusCode::OK_200) {
       connected_.store(false);
-      // For certain errors, don't reconnect
-      if (result.status() == 204 || // No Content - server wants us to stop
-          result.status() == 404 || // Not Found
-          result.status() == 401 || // Unauthorized
-          result.status() == 403) { // Forbidden
-        if (on_error_) { on_error_(Error::Connection); }
+      if (on_error_) { on_error_(Error::Connection); }
+
+      // For certain errors, don't reconnect.
+      // Note: 401 is intentionally absent so that handlers can refresh
+      // credentials via set_headers() and let the client reconnect.
+      if (result.status() == StatusCode::NoContent_204 ||
+          result.status() == StatusCode::NotFound_404 ||
+          result.status() == StatusCode::Forbidden_403) {
         break;
       }
 
-      if (on_error_) { on_error_(Error::Connection); }
-
       if (!should_reconnect(reconnect_count)) { break; }
       wait_for_reconnect();
       reconnect_count++;
@@ -9168,18 +9178,11 @@ void ClientImpl::setup_redirect_client(ClientType &client) {
   client.set_compress(compress_);
   client.set_decompress(decompress_);
 
-  // Copy authentication settings BEFORE proxy setup
-  if (!basic_auth_username_.empty()) {
-    client.set_basic_auth(basic_auth_username_, basic_auth_password_);
-  }
-  if (!bearer_token_auth_token_.empty()) {
-    client.set_bearer_token_auth(bearer_token_auth_token_);
-  }
-#ifdef CPPHTTPLIB_SSL_ENABLED
-  if (!digest_auth_username_.empty()) {
-    client.set_digest_auth(digest_auth_username_, digest_auth_password_);
-  }
-#endif
+  // NOTE: Authentication credentials (basic auth, bearer token, digest auth)
+  // are intentionally NOT copied to the redirect client. Per RFC 9110 Section
+  // 15.4, credentials must not be forwarded when redirecting to a different
+  // host. This function is only called for cross-host redirects; same-host
+  // redirects are handled directly in ClientImpl::redirect().
 
   // Setup proxy configuration (CRITICAL ORDER - proxy must be set
   // before proxy auth)
@@ -11425,7 +11428,8 @@ void Client::set_follow_location(bool on) {
 
 void Client::set_path_encode(bool on) { cli_->set_path_encode(on); }
 
-[[deprecated("Use set_path_encode instead")]]
+[[deprecated("Use set_path_encode() instead. "
+             "This function will be removed by v1.0.0.")]]
 void Client::set_url_encode(bool on) {
   cli_->set_path_encode(on);
 }
@@ -16330,9 +16334,10 @@ bool WebSocketClient::connect() {
 
   Error error;
   sock_ = detail::create_client_socket(
-      host_, std::string(), port_, AF_UNSPEC, false, false, nullptr, 5, 0,
+      host_, std::string(), port_, address_family_, tcp_nodelay_, ipv6_v6only_,
+      socket_options_, connection_timeout_sec_, connection_timeout_usec_,
       read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
-      write_timeout_usec_, std::string(), error);
+      write_timeout_usec_, interface_, error);
 
   if (sock_ == INVALID_SOCKET) { return false; }
 
@@ -16398,6 +16403,27 @@ void WebSocketClient::set_websocket_ping_interval(time_t sec) {
   websocket_ping_interval_sec_ = sec;
 }
 
+void WebSocketClient::set_tcp_nodelay(bool on) { tcp_nodelay_ = on; }
+
+void WebSocketClient::set_address_family(int family) {
+  address_family_ = family;
+}
+
+void WebSocketClient::set_ipv6_v6only(bool on) { ipv6_v6only_ = on; }
+
+void WebSocketClient::set_socket_options(SocketOptions socket_options) {
+  socket_options_ = std::move(socket_options);
+}
+
+void WebSocketClient::set_connection_timeout(time_t sec, time_t usec) {
+  connection_timeout_sec_ = sec;
+  connection_timeout_usec_ = usec;
+}
+
+void WebSocketClient::set_interface(const std::string &intf) {
+  interface_ = intf;
+}
+
 #ifdef CPPHTTPLIB_SSL_ENABLED
 
 void WebSocketClient::set_ca_cert_path(const std::string &path) {

vendor/cpp-httplib/httplib.h

@@ -8,8 +8,8 @@
 #ifndef CPPHTTPLIB_HTTPLIB_H
 #define CPPHTTPLIB_HTTPLIB_H
 
-#define CPPHTTPLIB_VERSION "0.38.0"
-#define CPPHTTPLIB_VERSION_NUM "0x002600"
+#define CPPHTTPLIB_VERSION "0.39.0"
+#define CPPHTTPLIB_VERSION_NUM "0x002700"
 
 #ifdef _WIN32
 #if defined(_WIN32_WINNT) && _WIN32_WINNT < 0x0A00
@@ -1001,8 +1001,8 @@ private:
 
   protected:
     std::streamsize xsputn(const char *s, std::streamsize n) override {
-      sink_.write(s, static_cast<size_t>(n));
-      return n;
+      if (sink_.write(s, static_cast<size_t>(n))) { return n; }
+      return 0;
     }
 
   private:
@@ -1058,9 +1058,12 @@ make_file_provider(const std::string &name, const std::string &filepath,
 
 inline std::pair<size_t, ContentProvider>
 make_file_body(const std::string &filepath) {
-  std::ifstream f(filepath, std::ios::binary | std::ios::ate);
-  if (!f) { return {0, ContentProvider{}}; }
-  auto size = static_cast<size_t>(f.tellg());
+  size_t size = 0;
+  {
+    std::ifstream f(filepath, std::ios::binary | std::ios::ate);
+    if (!f) { return {0, ContentProvider{}}; }
+    size = static_cast<size_t>(f.tellg());
+  }
 
   ContentProvider provider = [filepath](size_t offset, size_t length,
                                         DataSink &sink) -> bool {
@@ -1882,7 +1885,8 @@ private:
 
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
 public:
-  [[deprecated("Use ssl_backend_error() instead")]]
+  [[deprecated("Use ssl_backend_error() instead. "
+               "This function will be removed by v1.0.0.")]]
   uint64_t ssl_openssl_error() const {
     return ssl_backend_error_;
   }
@@ -2362,13 +2366,16 @@ protected:
 
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
 public:
-  [[deprecated("Use load_ca_cert_store() instead")]]
+  [[deprecated("Use load_ca_cert_store() instead. "
+               "This function will be removed by v1.0.0.")]]
   void set_ca_cert_store(X509_STORE *ca_cert_store);
 
-  [[deprecated("Use tls::create_ca_store() instead")]]
+  [[deprecated("Use tls::create_ca_store() instead. "
+               "This function will be removed by v1.0.0.")]]
   X509_STORE *create_ca_cert_store(const char *ca_cert, std::size_t size) const;
 
-  [[deprecated("Use set_server_certificate_verifier(VerifyCallback) instead")]]
+  [[deprecated("Use set_server_certificate_verifier(VerifyCallback) instead. "
+               "This function will be removed by v1.0.0.")]]
   virtual void set_server_certificate_verifier(
       std::function<SSLVerifierResponse(SSL *ssl)> verifier);
 #endif
@@ -2597,14 +2604,17 @@ private:
 
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
 public:
-  [[deprecated("Use tls_context() instead")]]
+  [[deprecated("Use tls_context() instead. "
+               "This function will be removed by v1.0.0.")]]
   SSL_CTX *ssl_context() const;
 
-  [[deprecated("Use set_session_verifier(session_t) instead")]]
+  [[deprecated("Use set_session_verifier(session_t) instead. "
+               "This function will be removed by v1.0.0.")]]
   void set_server_certificate_verifier(
       std::function<SSLVerifierResponse(SSL *ssl)> verifier);
 
-  [[deprecated("Use Result::ssl_backend_error() instead")]]
+  [[deprecated("Use Result::ssl_backend_error() instead. "
+               "This function will be removed by v1.0.0.")]]
   long get_verify_result() const;
 #endif
 };
@@ -2656,18 +2666,22 @@ private:
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
 public:
   [[deprecated("Use SSLServer(PemMemory) or "
-               "SSLServer(ContextSetupCallback) instead")]]
+               "SSLServer(ContextSetupCallback) instead. "
+               "This constructor will be removed by v1.0.0.")]]
   SSLServer(X509 *cert, EVP_PKEY *private_key,
             X509_STORE *client_ca_cert_store = nullptr);
 
-  [[deprecated("Use SSLServer(ContextSetupCallback) instead")]]
+  [[deprecated("Use SSLServer(ContextSetupCallback) instead. "
+               "This constructor will be removed by v1.0.0.")]]
   SSLServer(
       const std::function<bool(SSL_CTX &ssl_ctx)> &setup_ssl_ctx_callback);
 
-  [[deprecated("Use tls_context() instead")]]
+  [[deprecated("Use tls_context() instead. "
+               "This function will be removed by v1.0.0.")]]
   SSL_CTX *ssl_context() const;
 
-  [[deprecated("Use update_certs_pem() instead")]]
+  [[deprecated("Use update_certs_pem() instead. "
+               "This function will be removed by v1.0.0.")]]
   void update_certs(X509 *cert, EVP_PKEY *private_key,
                     X509_STORE *client_ca_cert_store = nullptr);
 #endif
@@ -2752,18 +2766,22 @@ private:
 
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
 public:
-  [[deprecated("Use SSLClient(host, port, PemMemory) instead")]]
+  [[deprecated("Use SSLClient(host, port, PemMemory) instead. "
+               "This constructor will be removed by v1.0.0.")]]
   explicit SSLClient(const std::string &host, int port, X509 *client_cert,
                      EVP_PKEY *client_key,
                      const std::string &private_key_password = std::string());
 
-  [[deprecated("Use Result::ssl_backend_error() instead")]]
+  [[deprecated("Use Result::ssl_backend_error() instead. "
+               "This function will be removed by v1.0.0.")]]
   long get_verify_result() const;
 
-  [[deprecated("Use tls_context() instead")]]
+  [[deprecated("Use tls_context() instead. "
+               "This function will be removed by v1.0.0.")]]
   SSL_CTX *ssl_context() const;
 
-  [[deprecated("Use set_session_verifier(session_t) instead")]]
+  [[deprecated("Use set_session_verifier(session_t) instead. "
+               "This function will be removed by v1.0.0.")]]
   void set_server_certificate_verifier(
       std::function<SSLVerifierResponse(SSL *ssl)> verifier) override;
 
@@ -3641,6 +3659,9 @@ public:
   SSEClient &set_reconnect_interval(int ms);
   SSEClient &set_max_reconnect_attempts(int n);
 
+  // Update headers (thread-safe)
+  SSEClient &set_headers(const Headers &headers);
+
   // State accessors
   bool is_connected() const;
   const std::string &last_event_id() const;
@@ -3665,6 +3686,7 @@ private:
   Client &client_;
   std::string path_;
   Headers headers_;
+  mutable std::mutex headers_mutex_;
 
   // Callbacks
   MessageHandler on_message_;
@@ -3785,6 +3807,12 @@ public:
   void set_read_timeout(time_t sec, time_t usec = 0);
   void set_write_timeout(time_t sec, time_t usec = 0);
   void set_websocket_ping_interval(time_t sec);
+  void set_tcp_nodelay(bool on);
+  void set_address_family(int family);
+  void set_ipv6_v6only(bool on);
+  void set_socket_options(SocketOptions socket_options);
+  void set_connection_timeout(time_t sec, time_t usec = 0);
+  void set_interface(const std::string &intf);
 
 #ifdef CPPHTTPLIB_SSL_ENABLED
   void set_ca_cert_path(const std::string &path);
@@ -3810,6 +3838,13 @@ private:
   time_t write_timeout_usec_ = CPPHTTPLIB_CLIENT_WRITE_TIMEOUT_USECOND;
   time_t websocket_ping_interval_sec_ =
       CPPHTTPLIB_WEBSOCKET_PING_INTERVAL_SECOND;
+  int address_family_ = AF_UNSPEC;
+  bool tcp_nodelay_ = CPPHTTPLIB_TCP_NODELAY;
+  bool ipv6_v6only_ = CPPHTTPLIB_IPV6_V6ONLY;
+  SocketOptions socket_options_ = nullptr;
+  time_t connection_timeout_sec_ = CPPHTTPLIB_CONNECTION_TIMEOUT_SECOND;
+  time_t connection_timeout_usec_ = CPPHTTPLIB_CONNECTION_TIMEOUT_USECOND;
+  std::string interface_;
 
 #ifdef CPPHTTPLIB_SSL_ENABLED
   bool is_ssl_ = false;