server : add loading html page while model is loading (#9468)

* Adding loading page for '/' server requests

* set content when model is loading

* removed loading html file

* updated cmakelist

* updated makefile

* cleaned up whitespace

* cleanup for PR removed error

* updated server test to handle 503 HTML

* updated server test to handle 503 HTML

* ca†ch 503 before parsing json

* revert test

* account for both api and web browser requests

* precommit corrections

* eol fix

* revert changes to pre-commit

* removed print statement

* made loading message more descriptive

* also support .html files

---------

Co-authored-by: VJHack <flymyplane21@gmail.com>
Co-authored-by: Vinesh Janarthanan <36610342+VJHack@users.noreply.github.com>

.github/workflows/build.yml

@@ -375,7 +375,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -401,7 +401,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -442,7 +442,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -546,7 +546,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -576,7 +576,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -610,7 +610,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -969,14 +969,14 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Install
         id: depends
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP SDK installation"

.github/workflows/server.yml

@@ -173,6 +173,7 @@ jobs:
         if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
           cd examples/server/tests
+          $env:PYTHONIOENCODING = ":replace"
           behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
 
       - name: Slow tests

CMakeLists.txt

@@ -139,10 +139,16 @@ set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location o
 # determining _precisely_ which defines are necessary for the llama-config
 # package.
 #
+set(GGML_TRANSIENT_DEFINES)
 get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
 get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
+if (GGML_DIR_DEFINES)
+    list(APPEND GGML_TRANSIENT_DEFINES ${GGML_DIR_DEFINES})
+endif()
 get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
-set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES})
+if (GGML_TARGET_DEFINES)
+    list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
+endif()
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
 
 set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)

Makefile

@@ -1440,6 +1440,7 @@ llama-server: \
 	examples/server/system-prompts.js.hpp \
 	examples/server/prompt-formats.js.hpp \
 	examples/server/json-schema-to-grammar.mjs.hpp \
+	examples/server/loading.html.hpp \
 	common/json.hpp \
 	common/stb_image.h \
 	$(OBJ_ALL)

common/arg.cpp

@@ -720,6 +720,14 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             params.prompt = value;
         }
     ));
+    add_opt(llama_arg(
+        {"--no-perf"},
+        format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
+        [](gpt_params & params) {
+            params.no_perf = true;
+            params.sparams.no_perf = true;
+        }
+    ).set_env("LLAMA_ARG_NO_PERF"));
     add_opt(llama_arg(
         {"-f", "--file"}, "FNAME",
         "a file containing the prompt (default: none)",

common/common.cpp

@@ -820,7 +820,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
-        llama_perf_reset(lctx, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_reset(lctx);
     }
 
     iparams.model   = model;
@@ -916,6 +916,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
     cparams.flash_attn        = params.flash_attn;
+    cparams.no_perf           = params.no_perf;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
     cparams.type_v = kv_cache_type_from_str(params.cache_type_v);

common/common.h

@@ -124,6 +124,7 @@ struct gpt_sampler_params {
     float   mirostat_eta      = 0.10f; // learning rate
     bool    penalize_nl       = false; // consider newlines as a repeatable token
     bool    ignore_eos        = false;
+    bool    no_perf           = false; // disable performance metrics
 
     std::vector<enum gpt_sampler_type> samplers = {
         GPT_SAMPLER_TYPE_TOP_K,
@@ -246,6 +247,7 @@ struct gpt_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool flash_attn        = false; // flash attention
+    bool no_perf           = false; // disable performance metrics
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool logits_all        = false; // return logits for all tokens in the batch

common/sampling.cpp

@@ -142,7 +142,7 @@ std::string gpt_sampler_params::print() const {
 struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params) {
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
-    lparams.no_perf = false; // TODO: control via params
+    lparams.no_perf = params.no_perf;
 
     auto * result = new gpt_sampler {
         /* .params = */ params,
@@ -257,10 +257,10 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
     // TODO: measure grammar performance
 
     if (gsmpl) {
-        llama_perf_print(gsmpl->chain, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
+        llama_perf_sampler_print(gsmpl->chain);
     }
     if (ctx) {
-        llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_print(ctx);
     }
 }
 

convert_hf_to_gguf.py

@@ -626,6 +626,9 @@ class Model:
         if chkhsh == "4e2b24cc4770243d65a2c9ec19770a72f08cffc161adbb73fcbb6b7dd45a0aae":
             # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct
             res = "exaone"
+        if chkhsh == "fcace8b9cac38ce847670c970cd5892031a753a1ef381abd1d9af00f713da085":
+            # ref: https://huggingface.co/microsoft/phi-2
+            res = "phi-2"
 
         if res is None:
             logger.warning("\n")

convert_hf_to_gguf_update.py

@@ -98,6 +98,7 @@ models = [
     {'name': "bloom",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
     {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
     {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
+    {"name": "phi-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
 ]
 
 

convert_lora_to_gguf.py

@@ -363,7 +363,13 @@ if __name__ == '__main__':
                     yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B)))
 
             def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-                dest = super().modify_tensors(data_torch, name, bid)
+                dest = list(super().modify_tensors(data_torch, name, bid))
+                # some archs may have the same tensor for lm_head and output (tie word embeddings)
+                # in this case, adapters targeting lm_head will fail when using llama-export-lora
+                # therefore, we ignore them for now
+                # see: https://github.com/ggerganov/llama.cpp/issues/9065
+                if name == "lm_head.weight" and len(dest) == 0:
+                    raise ValueError("lm_head is present in adapter, but is ignored in base model")
                 for dest_name, dest_data in dest:
                     assert isinstance(dest_data, LoraTorchTensor)
                     lora_a, lora_b = dest_data.get_lora_A_B()

examples/batched-bench/batched-bench.cpp

@@ -187,7 +187,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 

examples/batched.swift/Sources/main.swift

@@ -200,8 +200,8 @@ let t_main_end = ggml_time_us()
 
 print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n\n")
 
-llama_perf_print(UnsafeRawPointer(context), LLAMA_PERF_TYPE_CONTEXT)
-llama_perf_print(UnsafeRawPointer(smpl),    LLAMA_PERF_TYPE_SAMPLER_CHAIN)
+llama_perf_sampler_print(smpl)
+llama_perf_context_print(context)
 
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
     let utf8Count = text.utf8.count

examples/batched/batched.cpp

@@ -229,8 +229,8 @@ int main(int argc, char ** argv) {
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
     LOG_TEE("\n");
-    llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
-    llama_perf_print(ctx,  LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 

examples/embedding/embedding.cpp

@@ -306,7 +306,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     // clean up
     llama_batch_free(batch);

examples/eval-callback/eval-callback.cpp

@@ -182,7 +182,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/imatrix/imatrix.cpp

@@ -637,7 +637,7 @@ int main(int argc, char ** argv) {
     g_collector.save_imatrix();
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/llama-bench/llama-bench.cpp

@@ -1630,7 +1630,7 @@ int main(int argc, char ** argv) {
             fflush(p_err->fout);
         }
 
-        llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_print(ctx);
 
         llama_free(ctx);
 

examples/llava/MobileVLM-README.md

@@ -39,7 +39,7 @@ python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
 3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
 
 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf \
+python ./examples/llava/convert_image_encoder_to_gguf.py \
     -m path/to/clip-vit-large-patch14-336 \
     --llava-projector path/to/MobileVLM-1.7B/llava.projector \
     --output-dir path/to/MobileVLM-1.7B \
@@ -47,7 +47,7 @@ python ./examples/llava/convert_image_encoder_to_gguf \
 ```
 
 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf \
+python ./examples/llava/convert_image_encoder_to_gguf.py \
     -m path/to/clip-vit-large-patch14-336 \
     --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
     --output-dir path/to/MobileVLM-1.7B_V2 \
@@ -57,12 +57,12 @@ python ./examples/llava/convert_image_encoder_to_gguf \
 4. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B
+python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B --skip-unknown
 ```
 
-5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k`
+5. Use `quantize` to convert LLaMA part's DataType from `fp32` to `q4_k`
 ```sh
-./llama-quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
+./llama-quantize path/to/MobileVLM-1.7B/ggml-model-F32.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
 ```
 
 Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directory.

examples/llava/llava-cli.cpp

@@ -308,7 +308,7 @@ int main(int argc, char ** argv) {
         // process the prompt
         process_prompt(ctx_llava, image_embed, &params, params.prompt);
 
-        llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_print(ctx_llava->ctx_llama);
         llava_image_embed_free(image_embed);
         ctx_llava->model = NULL;
         llava_free(ctx_llava);
@@ -325,7 +325,7 @@ int main(int argc, char ** argv) {
             // process the prompt
             process_prompt(ctx_llava, image_embed, &params, params.prompt);
 
-            llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
+            llama_perf_context_print(ctx_llava->ctx_llama);
             llava_image_embed_free(image_embed);
             ctx_llava->model = NULL;
             llava_free(ctx_llava);

examples/llava/minicpmv-cli.cpp

@@ -319,7 +319,7 @@ int main(int argc, char ** argv) {
             }
         }
         printf("\n");
-        llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_print(ctx_llava->ctx_llama);
 
         ctx_llava->model = NULL;
         llava_free(ctx_llava);

examples/lookup/lookup.cpp

@@ -240,8 +240,7 @@ int main(int argc, char ** argv){
     LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
     LOG_TEE("\ntarget:\n\n");
-    llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
-    llama_perf_print(ctx,  LLAMA_PERF_TYPE_CONTEXT);
+    gpt_perf_print(ctx, smpl);
 
     gpt_sampler_free(smpl);
 

examples/parallel/parallel.cpp

@@ -415,7 +415,7 @@ int main(int argc, char ** argv) {
     LOG_TEE("\n");
 
     // TODO: print sampling/grammar timings for all clients
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 

examples/passkey/passkey.cpp

@@ -256,7 +256,7 @@ int main(int argc, char ** argv) {
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 

examples/perplexity/perplexity.cpp

@@ -2047,7 +2047,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
     write_logfile(ctx, params, model, results);
 
     llama_free(ctx);

examples/retrieval/retrieval.cpp

@@ -292,7 +292,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     // clean up
     llama_batch_free(query_batch);

examples/server/CMakeLists.txt

@@ -30,6 +30,7 @@ set(PUBLIC_ASSETS
     system-prompts.js
     prompt-formats.js
     json-schema-to-grammar.mjs
+    loading.html
 )
 
 foreach(asset ${PUBLIC_ASSETS})

examples/server/README.md

@@ -407,9 +407,44 @@ Notice that each `probs` is an array of length `n_probs`.
 
     *Options:*
 
-    `content`: Set the text to tokenize.
+    `content`: (Required) The text to tokenize.
 
-    `add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
+    `add_special`: (Optional) Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
+
+    `with_pieces`: (Optional) Boolean indicating whether to return token pieces along with IDs.  Default: `false`
+
+**Response:**
+
+Returns a JSON object with a `tokens` field containing the tokenization result. The `tokens` array contains either just token IDs or objects with `id` and `piece` fields, depending on the `with_pieces` parameter. The piece field is a string if the piece is valid unicode or a list of bytes otherwise.
+
+
+If `with_pieces` is `false`:
+```json
+{
+  "tokens": [123, 456, 789]
+}
+```
+
+If `with_pieces` is `true`:
+```json
+{
+  "tokens": [
+    {"id": 123, "piece": "Hello"},
+    {"id": 456, "piece": " world"},
+    {"id": 789, "piece": "!"}
+  ]
+}
+```
+
+With input 'á' (utf8 hex: C3 A1) on tinyllama/stories260k
+```json
+{
+  "tokens": [
+    {"id": 198, "piece": [195]}, // hex C3
+    {"id": 164, "piece": [161]} // hex A1
+  ]
+}
+```
 
 ### POST `/detokenize`: Convert tokens to text
 

examples/server/public/loading.html

@@ -0,0 +1,12 @@
+<!DOCTYPE html>
+<html>
+    <head>
+        <meta http-equiv="refresh" content="5">
+    </head>
+    <body>
+        <div id="loading">
+            The model is loading. Please wait.<br/>
+            The user interface will appear soon.
+        </div>
+    </body>
+</html>

examples/server/server.cpp

@@ -28,6 +28,7 @@
 #include "system-prompts.js.hpp"
 #include "prompt-formats.js.hpp"
 #include "json-schema-to-grammar.mjs.hpp"
+#include "loading.html.hpp"
 
 #include <atomic>
 #include <chrono>
@@ -2592,10 +2593,16 @@ int main(int argc, char ** argv) {
         return false;
     };
 
-    auto middleware_server_state = [&res_error, &state](const httplib::Request &, httplib::Response & res) {
+    auto middleware_server_state = [&res_error, &state](const httplib::Request & req, httplib::Response & res) {
         server_state current_state = state.load();
         if (current_state == SERVER_STATE_LOADING_MODEL) {
-            res_error(res, format_error_response("Loading model", ERROR_TYPE_UNAVAILABLE));
+            auto tmp = string_split(req.path, '.');
+            if (req.path == "/" || tmp.back() == "html") {
+                res.set_content(reinterpret_cast<const char*>(loading_html), loading_html_len, "text/html; charset=utf-8");
+                res.status = 503;
+            } else {
+                res_error(res, format_error_response("Loading model", ERROR_TYPE_UNAVAILABLE));
+            }
             return false;
         }
         return true;
@@ -3013,12 +3020,39 @@ int main(int argc, char ** argv) {
     const auto handle_tokenize = [&ctx_server, &res_ok](const httplib::Request & req, httplib::Response & res) {
         const json body = json::parse(req.body);
 
-        std::vector<llama_token> tokens;
+        json tokens_response = json::array();
         if (body.count("content") != 0) {
             const bool add_special = json_value(body, "add_special", false);
-            tokens = ctx_server.tokenize(body.at("content"), add_special);
+            const bool with_pieces = json_value(body, "with_pieces", false);
+            std::vector<llama_token> tokens = ctx_server.tokenize(body.at("content"), add_special);
+
+            if (with_pieces) {
+                for (const auto& token : tokens) {
+                    std::string piece = llama_token_to_piece(ctx_server.ctx, token);
+                    json piece_json;
+
+                    // Check if the piece is valid UTF-8
+                    if (is_valid_utf8(piece)) {
+                        piece_json = piece;
+                    } else {
+                        // If not valid UTF-8, store as array of byte values
+                        piece_json = json::array();
+                        for (unsigned char c : piece) {
+                            piece_json.push_back(static_cast<int>(c));
+                        }
+                    }
+
+                    tokens_response.push_back({
+                        {"id", token},
+                        {"piece", piece_json}
+                    });
+                }
+            } else {
+                tokens_response = tokens;
+            }
         }
-        const json data = format_tokenizer_response(tokens);
+
+        const json data = format_tokenizer_response(tokens_response);
         res_ok(res, data);
     };
 

examples/server/tests/features/server.feature

@@ -105,6 +105,14 @@ Feature: llama.cpp server
     Given first token is removed
     Then  tokens can be detokenized
 
+  Scenario: Tokenize with pieces
+    When  tokenizing with pieces:
+    """
+    What is the capital of Germany?
+    媽
+    """
+    Then  tokens are given with pieces
+
   Scenario: Models available
     Given available models
     Then  1 models are supported

examples/server/tests/features/steps/steps.py

@@ -1,3 +1,6 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
 import asyncio
 import json
 import os
@@ -697,6 +700,32 @@ def step_tokenize_set_add_special(context):
     context.tokenize_add_special = True
 
 
+@step("tokenizing with pieces")
+@async_run_until_complete
+async def step_tokenize_with_pieces(context):
+    context.tokenized_text = context_text(context)
+    async with aiohttp.ClientSession() as session:
+        tokenize_args = {"content": context.tokenized_text, "with_pieces": True}
+        if getattr(context, "tokenize_add_special", None) is not None:
+            tokenize_args["add_special"] = context.tokenize_add_special
+
+        async with session.post(
+            f"{context.base_url}/tokenize", json=tokenize_args
+        ) as response:
+            assert response.status == 200
+            tokenize_json = await response.json()
+            context.tokens_with_pieces = tokenize_json["tokens"]
+
+
+@step("tokens are given with pieces")
+@async_run_until_complete
+async def step_tokenize_with_pieces(context):
+    # Verify that the response contains both token IDs and pieces
+    assert all(
+        "id" in token and "piece" in token for token in context.tokens_with_pieces
+    )
+
+
 @step('tokenizing')
 @async_run_until_complete
 async def step_tokenize(context):

examples/server/utils.hpp

@@ -616,7 +616,40 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
     return res;
 }
 
-static json format_tokenizer_response(const std::vector<llama_token> & tokens) {
+static bool is_valid_utf8(const std::string & str) {
+    const unsigned char* bytes = reinterpret_cast<const unsigned char*>(str.data());
+    const unsigned char* end = bytes + str.length();
+
+    while (bytes < end) {
+        if (*bytes <= 0x7F) {
+            // 1-byte sequence (0xxxxxxx)
+            bytes++;
+        } else if ((*bytes & 0xE0) == 0xC0) {
+            // 2-byte sequence (110xxxxx 10xxxxxx)
+            if (end - bytes < 2 || (bytes[1] & 0xC0) != 0x80)
+                return false;
+            bytes += 2;
+        } else if ((*bytes & 0xF0) == 0xE0) {
+            // 3-byte sequence (1110xxxx 10xxxxxx 10xxxxxx)
+            if (end - bytes < 3 || (bytes[1] & 0xC0) != 0x80 || (bytes[2] & 0xC0) != 0x80)
+                return false;
+            bytes += 3;
+        } else if ((*bytes & 0xF8) == 0xF0) {
+            // 4-byte sequence (11110xxx 10xxxxxx 10xxxxxx 10xxxxxx)
+            if (end - bytes < 4 || (bytes[1] & 0xC0) != 0x80 ||
+                (bytes[2] & 0xC0) != 0x80 || (bytes[3] & 0xC0) != 0x80)
+                return false;
+            bytes += 4;
+        } else {
+            // Invalid UTF-8 lead byte
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static json format_tokenizer_response(const json & tokens) {
     return json {
         {"tokens", tokens}
     };

examples/simple/simple.cpp

@@ -154,8 +154,8 @@ int main(int argc, char ** argv) {
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
     LOG_TEE("\n");
-    llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
-    llama_perf_print(ctx,  LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 

examples/speculative/speculative.cpp

@@ -616,7 +616,7 @@ int main(int argc, char ** argv) {
 
     LOG_TEE("\ndraft:\n\n");
     // TODO: print sampling/grammar timings for all drafts
-    llama_perf_print(ctx_dft, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx_dft);
 
     LOG_TEE("\ntarget:\n\n");
     gpt_perf_print(ctx_tgt, smpl);

ggml/include/ggml-cann.h

@@ -80,6 +80,13 @@ ggml_backend_cann_buffer_type(int32_t device);
  */
 GGML_API GGML_CALL int32_t ggml_backend_cann_get_device_count(void);
 
+/**
+ * @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
+ *
+ * @return A pointer to the host buffer type interface.
+ */
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
+
 /**
  * @brief Retrieves the description of a specific CANN device.
  *

ggml/src/ggml-cann.cpp

@@ -1221,6 +1221,116 @@ ggml_backend_cann_buffer_type(int32_t device) {
     return &ggml_backend_cann_buffer_types[device];
 }
 
+/**
+ * @brief Retrieves the name associated with a CANN host buffer type.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN host buffer type context.
+ *
+ * @param buft Pointer to the host buffer type context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+GGML_CALL static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return "CANN_Host";
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Retrieves the name associated with a CANN host buffer.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN host buffer context.
+ *
+ * @param buft Pointer to the host buffer context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+GGML_CALL static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return "CANN_Host";
+
+    GGML_UNUSED(buffer);
+}
+
+/**
+ * @brief Free resources associated with a CANN host buffer.
+ *
+ * This function frees the resources associated with a CANN host buffer, including
+ * its context.
+ *
+ * @param buffer The CANN host buffer to free.
+ */
+GGML_CALL static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
+    ACL_CHECK(aclrtFreeHost(buffer->context));
+}
+
+/**
+ * @brief Allocates a new CANN host buffer of the specified size.
+ *
+ * This function allocates a new CANN host buffer with the given size.
+ * @param size Size in bytes of the host buffer to allocate.
+ * @return Pointer to the allocated host buffer, or nullptr if allocation fails.
+ */
+static void * ggml_cann_host_malloc(size_t size) {
+    if (getenv("GGML_CANN_NO_PINNED") != nullptr) {
+        return nullptr;
+    }
+
+    void * hostPtr = nullptr;
+    aclError err = aclrtMallocHost((void **) &hostPtr, size);
+    if (err != ACL_SUCCESS) {
+
+        GGML_CANN_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, aclGetRecentErrMsg());
+        return nullptr;
+    }
+    return hostPtr;
+}
+
+/**
+ * @brief Allocates a new CANN host buffer of the specified type and size.
+ *
+ * @param buft Pointer to the host buffer type context.
+ * @param size Size in bytes of the host buffer to allocate.
+ * @return Pointer to the allocated host buffer, or CPU buffer pointer if allocation fails.
+ */
+GGML_CALL static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * hostPtr = ggml_cann_host_malloc(size);
+
+    if (hostPtr == nullptr) {
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(hostPtr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_cann_host_buffer_name;
+    buffer->iface.free_buffer = ggml_backend_cann_host_buffer_free;
+
+    return buffer;
+}
+
+/**
+ * @brief Interface for managing CANN host buffer types in the GGML backend.
+ *
+ * Provides function pointers for allocating, querying properties, and managing
+ * memory for CANN buffer types in the GGML backend.
+ */
+GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_cann_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cann_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_cann_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_cann_buffer_type_host;
+}
+
 /**
  * @brief Computes the forward operation for a given tensor using CANN
  * operations.

include/llama.h

@@ -343,7 +343,7 @@ extern "C" {
         bool embeddings;  // if true, extract embeddings (together with logits)
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
         bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
-      //bool no_perf;     // whether to measure performance timings, TODO: implement
+        bool no_perf;     // whether to measure performance timings
 
         // Abort callback
         // if it returns true, execution of llama_decode() will be aborted
@@ -1056,6 +1056,9 @@ extern "C" {
     LLAMA_API struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i);
     LLAMA_API int                    llama_sampler_chain_n  (const struct llama_sampler * chain);
 
+    // after removing a sampler, the chain will no longer own it, and it will not be freed when the chain is freed
+    LLAMA_API struct llama_sampler * llama_sampler_chain_remove(   struct llama_sampler * chain, int32_t i);
+
     // available samplers:
 
     LLAMA_API struct llama_sampler * llama_sampler_init_greedy     (void);
@@ -1173,13 +1176,30 @@ extern "C" {
     // NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements.
     //
 
-    enum llama_perf_type {
-        LLAMA_PERF_TYPE_CONTEXT       = 0,
-        LLAMA_PERF_TYPE_SAMPLER_CHAIN = 1,
+    struct llama_perf_context_data {
+        double t_start_ms;
+        double t_load_ms;
+        double t_p_eval_ms;
+        double t_eval_ms;
+
+        int32_t n_p_eval;
+        int32_t n_eval;
     };
 
-    LLAMA_API void llama_perf_print(const void * ctx, enum llama_perf_type type);
-    LLAMA_API void llama_perf_reset(      void * ctx, enum llama_perf_type type);
+    struct llama_perf_sampler_data {
+        double t_sample_ms;
+
+        int32_t n_sample;
+    };
+
+    LLAMA_API struct llama_perf_context_data llama_perf_context      (const struct llama_context * ctx);
+    LLAMA_API void                           llama_perf_context_print(const struct llama_context * ctx);
+    LLAMA_API void                           llama_perf_context_reset(      struct llama_context * ctx);
+
+    // NOTE: the following work only with samplers constructed via llama_sampler_chain_init
+    LLAMA_API struct llama_perf_sampler_data llama_perf_sampler      (const struct llama_sampler * chain);
+    LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
+    LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
     LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);
 

src/llama-sampling.cpp

@@ -349,13 +349,26 @@ void llama_sampler_chain_add(struct llama_sampler * chain, struct llama_sampler
 struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i) {
     const auto * p = (const llama_sampler_chain *) chain->ctx;
 
-    if (i < 0 || i >= (int32_t) p->samplers.size()) {
+    if (i < 0 || (size_t) i >= p->samplers.size()) {
         return nullptr;
     }
 
     return p->samplers[i];
 }
 
+struct llama_sampler * llama_sampler_chain_remove(struct llama_sampler * chain, int32_t i) {
+    auto * p = (llama_sampler_chain *) chain->ctx;
+
+    if (i < 0 || (size_t) i >= p->samplers.size()) {
+        return nullptr;
+    }
+
+    auto * result = p->samplers[i];
+    p->samplers.erase(p->samplers.begin() + i);
+
+    return result;
+}
+
 int llama_sampler_chain_n(const struct llama_sampler * chain) {
     const auto * p = (const llama_sampler_chain *) chain->ctx;
 
@@ -1656,3 +1669,37 @@ uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl) {
 
     return LLAMA_DEFAULT_SEED;
 }
+
+// perf
+
+struct llama_perf_sampler_data llama_perf_sampler(const struct llama_sampler * chain) {
+    struct llama_perf_sampler_data data = {};
+
+    if (chain == nullptr || chain->iface != &llama_sampler_chain_i) {
+        GGML_ABORT("%s: invalid sampler passed - requires a sampler created with llama_sampler_chain_init()\n", __func__);
+    }
+
+    const auto * ctx = (const struct llama_sampler_chain *) chain->ctx;
+
+    data.t_sample_ms = 1e-3 * ctx->t_sample_us;
+    data.n_sample    = std::max(0, ctx->n_sample);
+
+    return data;
+}
+
+void llama_perf_sampler_print(const struct llama_sampler * chain) {
+    const auto data = llama_perf_sampler(chain);
+
+    LLAMA_LOG_INFO("%s:    sampling time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, data.t_sample_ms, data.n_sample, data.t_sample_ms / data.n_sample, 1e3 / data.t_sample_ms * data.n_sample);
+}
+
+void llama_perf_sampler_reset(struct llama_sampler * chain) {
+    if (chain == nullptr || chain->iface != &llama_sampler_chain_i) {
+        GGML_ABORT("%s: invalid sampler passed - requires a sampler created with llama_sampler_chain_init()\n", __func__);
+    }
+
+    auto * ctx = (struct llama_sampler_chain *) chain->ctx;
+
+    ctx->t_sample_us = ctx->n_sample = 0;
+}

src/llama.cpp

@@ -2156,6 +2156,10 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer
     if (host_buffer) {
         buft = ggml_backend_sycl_host_buffer_type();
     }
+#elif defined(GGML_USE_CANN)
+    if (host_buffer) {
+        buft = ggml_backend_cann_host_buffer_type();
+    }
 #elif defined(GGML_USE_CPU_HBM)
     buft = ggml_backend_cpu_hbm_buffer_type();
 #elif defined(GGML_USE_VULKAN)
@@ -2482,6 +2486,7 @@ struct llama_cparams {
     bool causal_attn;
     bool offload_kqv;
     bool flash_attn;
+    bool no_perf;
 
     enum llama_pooling_type pooling_type;
 
@@ -6657,8 +6662,6 @@ static bool llm_load_tensors(
         bool use_mlock,
         llama_progress_callback progress_callback,
         void * progress_callback_user_data) {
-    model.t_start_us = ggml_time_us();
-
     auto & hparams = model.hparams;
 
     model.split_mode   = split_mode;
@@ -8589,14 +8592,13 @@ static bool llm_load_tensors(
         }
     }
 
-    // loading time will be recalculate after the first eval, so
-    // we take page faults deferred by mmap() into consideration
-    model.t_load_us = ggml_time_us() - model.t_start_us;
     return true;
 }
 
 // Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
 static int llama_model_load(const std::string & fname, llama_model & model, llama_model_params & params) {
+    model.t_start_us = ggml_time_us();
+
     try {
         llama_model_loader ml(fname, params.use_mmap, params.check_tensors, params.kv_overrides);
 
@@ -8658,6 +8660,10 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
         return -1;
     }
 
+    // loading time will be recalculate after the first eval, so
+    // we take page faults deferred by mmap() into consideration
+    model.t_load_us = ggml_time_us() - model.t_start_us;
+
     return 0;
 }
 
@@ -17945,6 +17951,7 @@ struct llama_context_params llama_context_default_params() {
         /*.embeddings                  =*/ false,
         /*.offload_kqv                 =*/ true,
         /*.flash_attn                  =*/ false,
+        /*.no_perf                     =*/ true,
         /*.abort_callback              =*/ nullptr,
         /*.abort_callback_data         =*/ nullptr,
     };
@@ -18155,6 +18162,7 @@ struct llama_context * llama_new_context_with_model(
     cparams.embeddings       = params.embeddings;
     cparams.offload_kqv      = params.offload_kqv;
     cparams.flash_attn       = params.flash_attn;
+    cparams.no_perf          = params.no_perf;
     cparams.pooling_type     = params.pooling_type;
 
     cparams.n_ctx            = params.n_ctx           == 0    ? hparams.n_ctx_train           : params.n_ctx;
@@ -20073,10 +20081,14 @@ void llama_synchronize(struct llama_context * ctx) {
 
     // add the evaluation to the stats
     if (ctx->n_queued_tokens == 1) {
-        ctx->t_eval_us += ggml_time_us() - ctx->t_compute_start_us;
+        if (!ctx->cparams.no_perf) {
+            ctx->t_eval_us += ggml_time_us() - ctx->t_compute_start_us;
+        }
         ctx->n_eval++;
     } else if (ctx->n_queued_tokens > 1) {
-        ctx->t_p_eval_us += ggml_time_us() - ctx->t_compute_start_us;
+        if (!ctx->cparams.no_perf) {
+            ctx->t_p_eval_us += ggml_time_us() - ctx->t_compute_start_us;
+        }
         ctx->n_p_eval += ctx->n_queued_tokens;
     }
 
@@ -20684,65 +20696,40 @@ const char * llama_print_system_info(void) {
     return s.c_str();
 }
 
-void llama_perf_print(const void * ctx, enum llama_perf_type type) {
-    switch (type) {
-        case LLAMA_PERF_TYPE_CONTEXT:
-            {
-                const auto * p = (const struct llama_context *) ctx;
+struct llama_perf_context_data llama_perf_context(const struct llama_context * ctx) {
+    struct llama_perf_context_data data = {};
 
-                const double t_start_ms   = 1e-3 * p->t_start_us;
-                const double t_end_ms     = 1.00 * ggml_time_ms();
-                const double t_load_ms    = 1e-3 * p->t_load_us;
-                const double t_p_eval_ms  = 1e-3 * p->t_p_eval_us;
-                const double t_eval_ms    = 1e-3 * p->t_eval_us;
-
-                const int32_t n_p_eval  = std::max(0, p->n_p_eval);
-                const int32_t n_eval    = std::max(1, p->n_eval);
-
-                LLAMA_LOG_INFO("%s:        load time = %10.2f ms\n", __func__, t_load_ms);
-                LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
-                        __func__, t_p_eval_ms, n_p_eval, t_p_eval_ms / n_p_eval, 1e3 / t_p_eval_ms * n_p_eval);
-                LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-                        __func__, t_eval_ms, n_eval, t_eval_ms / n_eval, 1e3 / t_eval_ms * n_eval);
-                LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - t_start_ms), (n_p_eval + n_eval));
-            } break;
-        case LLAMA_PERF_TYPE_SAMPLER_CHAIN:
-            {
-                const auto * smpl = (const struct llama_sampler *) ctx;
-                const auto * p = (const struct llama_sampler_chain *) smpl->ctx;
-
-                const double t_sampler_ms = 1e-3 * p->t_sample_us;
-
-                const int32_t n_sampler = std::max(0, p->n_sample);
-
-                LLAMA_LOG_INFO("%s:    sampling time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-                        __func__, t_sampler_ms, n_sampler, t_sampler_ms / n_sampler, 1e3 / t_sampler_ms * n_sampler);
-            } break;
-        default:
-            GGML_ABORT("invalid perf type");
+    if (ctx == nullptr) {
+        return data;
     }
+
+    data.t_start_ms  = 1e-3 * ctx->t_start_us;
+    data.t_load_ms   = 1e-3 * ctx->t_load_us;
+    data.t_p_eval_ms = 1e-3 * ctx->t_p_eval_us;
+    data.t_eval_ms   = 1e-3 * ctx->t_eval_us;
+    data.n_p_eval    = std::max(1, ctx->n_p_eval);
+    data.n_eval      = std::max(1, ctx->n_eval);
+
+    return data;
 }
 
-void llama_perf_reset(void * ctx, enum llama_perf_type type) {
-    switch (type) {
-        case LLAMA_PERF_TYPE_CONTEXT:
-            {
-                auto * p = (struct llama_context *) ctx;
+void llama_perf_context_print(const struct llama_context * ctx) {
+    const auto data = llama_perf_context(ctx);
 
-                p->t_start_us  = ggml_time_us();
-                p->t_eval_us   = p->n_eval = 0;
-                p->t_p_eval_us = p->n_p_eval = 0;
-            } break;
-        case LLAMA_PERF_TYPE_SAMPLER_CHAIN:
-            {
-                auto * smpl = (struct llama_sampler *) ctx;
-                auto * p = (struct llama_sampler_chain *) smpl->ctx;
+    const double t_end_ms = 1e-3 * ggml_time_us();
 
-                p->t_sample_us = p->n_sample = 0;
-            } break;
-        default:
-            GGML_ABORT("invalid perf type");
-    }
+    LLAMA_LOG_INFO("%s:        load time = %10.2f ms\n", __func__, data.t_load_ms);
+    LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval);
+    LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
+    LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
+}
+
+void llama_perf_context_reset(struct llama_context * ctx) {
+    ctx->t_start_us  = ggml_time_us();
+    ctx->t_eval_us   = ctx->n_eval = 0;
+    ctx->t_p_eval_us = ctx->n_p_eval = 0;
 }
 
 void llama_perf_dump_yaml(FILE * stream, const llama_context * ctx) {