cuda : fix tensor size calculation for non-split buffer (#5145)

updated the n_task calculation to use max number of
threads possible. This has improved the prompt eval
performance by around 5% for DOT kernels and by
around 10% for MMLA kernels on AWS Graviton3.

.github/workflows/build.yml

@@ -72,7 +72,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
@@ -107,7 +107,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -141,7 +141,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose
+          ctest -L main --verbose
 
   # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
   #       how to debug it.
@@ -202,7 +202,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest
@@ -394,7 +394,7 @@ jobs:
         if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512
         run: |
           cd build
-          ctest -C Release --verbose --timeout 900
+          ctest -L main -C Release --verbose --timeout 900
 
       - name: Test (Intel SDE)
         id: cmake_test_sde
@@ -406,7 +406,7 @@ jobs:
           7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
           $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
           cd build
-          & $sde -future -- ctest -C Release --verbose --timeout 900
+          & $sde -future -- ctest -L main -C Release --verbose --timeout 900
 
       - name: Determine tag name
         id: tag

.gitignore

@@ -27,7 +27,7 @@
 lcov-report/
 gcovr-report/
 
-build*/
+build*
 out/
 tmp/
 
@@ -89,20 +89,3 @@ examples/jeopardy/results.txt
 
 poetry.lock
 poetry.toml
-
-# Test binaries
-/tests/test-grammar-parser
-/tests/test-llama-grammar
-/tests/test-double-float
-/tests/test-grad0
-/tests/test-opt
-/tests/test-quantize-fns
-/tests/test-quantize-perf
-/tests/test-sampling
-/tests/test-tokenizer-0-llama
-/tests/test-tokenizer-0-falcon
-/tests/test-tokenizer-1-llama
-/tests/test-tokenizer-1-bpe
-/tests/test-rope
-/tests/test-backend-ops
-/tests/test-autorelease

Makefile

@@ -9,7 +9,7 @@ TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
 	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops tests/test-autorelease
+	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -619,7 +619,7 @@ embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(C
 save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
@@ -748,5 +748,8 @@ tests/test-c.o: tests/test-c.c llama.h
 tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

README.md

@@ -10,11 +10,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
+- ⚠️ Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
 - New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
 - Collecting Apple Silicon performance stats:
   - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
   - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
-- Added Mixtral support: https://github.com/ggerganov/llama.cpp/pull/4406
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
@@ -112,6 +112,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
 - [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
 - [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
+- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 
 
 **Bindings:**

ci/run.sh

@@ -22,9 +22,9 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-rm -v $OUT/*.log
-rm -v $OUT/*.exit
-rm -v $OUT/*.md
+rm -f "$OUT/*.log"
+rm -f "$OUT/*.exit"
+rm -f "$OUT/*.md"
 
 sd=`dirname $0`
 cd $sd/../
@@ -94,7 +94,7 @@ function gg_run_ctest_debug {
     (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
 }
@@ -123,9 +123,9 @@ function gg_run_ctest_release {
     (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
-        (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi
 
     set +e
@@ -141,6 +141,61 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
+function gg_get_model {
+    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
+    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    if [[ -s $gguf_3b ]]; then
+        echo -n "$gguf_3b"
+    elif [[ -s $gguf_7b ]]; then
+        echo -n "$gguf_7b"
+    else
+        echo >&2 "No model found. Can't run gg_run_ctest_with_model."
+        exit 1
+    fi
+}
+
+function gg_run_ctest_with_model_debug {
+    cd ${SRC}
+
+    local model; model=$(gg_get_model)
+    cd build-ci-debug
+    set -e
+    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    set +e
+    cd ..
+}
+
+function gg_run_ctest_with_model_release {
+    cd ${SRC}
+
+    local model; model=$(gg_get_model)
+    cd build-ci-release
+    set -e
+    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    set +e
+    cd ..
+}
+
+function gg_sum_ctest_with_model_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest with model files in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
+function gg_sum_ctest_with_model_release {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest with model files in release mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
 # open_llama_3b_v2
 
 function gg_run_open_llama_3b_v2 {
@@ -183,8 +238,6 @@ function gg_run_open_llama_3b_v2 {
 
     wiki_test_60="${path_wiki}/wiki.test-60.raw"
 
-    ./bin/test-autorelease ${model_f16}
-
     ./bin/quantize ${model_f16} ${model_q8_0} q8_0
     ./bin/quantize ${model_f16} ${model_q4_0} q4_0
     ./bin/quantize ${model_f16} ${model_q4_1} q4_1
@@ -507,14 +560,18 @@ function gg_sum_open_llama_7b_v2 {
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
     rm -rf ${SRC}/models-mnt
-
     mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
-    python3 -m pip install -r ${SRC}/requirements.txt
-    python3 -m pip install --editable gguf-py
+    # Create a fresh python3 venv and enter it
+    python3 -m venv "$MNT/venv"
+    source "$MNT/venv/bin/activate"
+
+    pip install -r ${SRC}/requirements.txt --disable-pip-version-check
+    pip install --editable gguf-py --disable-pip-version-check
 fi
 
 ret=0
@@ -529,6 +586,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
         else
             test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
+        test $ret -eq 0 && gg_run ctest_with_model_debug
+        test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi
 

examples/server/CMakeLists.txt

@@ -1,7 +1,7 @@
 set(TARGET server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>

examples/server/oai.hpp

@@ -0,0 +1,208 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+#include "utils.hpp"
+
+#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
+
+using json = nlohmann::json;
+
+inline static json oaicompat_completion_params_parse(
+    const json &body /* openai api json semantics */)
+{
+    json llama_params;
+
+    llama_params["__oaicompat"] = true;
+
+    // Map OpenAI parameters to llama.cpp parameters
+    //
+    // For parameters that are defined by the OpenAI documentation (e.g.
+    // temperature), we explicitly specify OpenAI's intended default; we
+    // need to do that because sometimes OpenAI disagrees with llama.cpp
+    //
+    // https://platform.openai.com/docs/api-reference/chat/create
+    llama_sampling_params default_sparams;
+    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
+    llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
+    llama_params["temperature"]       = json_value(body, "temperature", 0.0);
+    llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);
+    llama_params["top_p"]             = json_value(body, "top_p", 1.0);
+    llama_params["n_predict"]         = json_value(body, "max_tokens", -1);
+    llama_params["logit_bias"]        = json_value(body, "logit_bias",json::object());
+    llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
+    llama_params["presence_penalty"]  = json_value(body, "presence_penalty", 0.0);
+    llama_params["seed"]              = json_value(body, "seed", LLAMA_DEFAULT_SEED);
+    llama_params["stream"]            = json_value(body, "stream", false);
+    llama_params["mirostat"]          = json_value(body, "mirostat", default_sparams.mirostat);
+    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
+    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
+    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", default_sparams.penalize_nl);
+    llama_params["typical_p"]         = json_value(body, "typical_p", default_sparams.typical_p);
+    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
+    llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
+    llama_params["tfs_z"]             = json_value(body, "tfs_z", default_sparams.tfs_z);
+
+    if (body.count("grammar") != 0) {
+        llama_params["grammar"] = json_value(body, "grammar", json::object());
+    }
+
+    // Handle 'stop' field
+    if (body.contains("stop") && body["stop"].is_string()) {
+        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
+    } else {
+        llama_params["stop"] = json_value(body, "stop", json::array());
+    }
+
+    // Ensure there is ChatML-specific end sequence among stop words
+    llama_params["stop"].push_back("<|im_end|>");
+
+    return llama_params;
+}
+
+inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
+{
+    json result = response.result_json;
+
+    bool stopped_word        = result.count("stopped_word") != 0;
+    bool stopped_eos         = json_value(result, "stopped_eos", false);
+    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
+    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
+    std::string content      = json_value(result, "content", std::string(""));
+
+    std::string finish_reason = "length";
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+
+    json choices =
+        streaming ? json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"delta", json::object()}}})
+                  : json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"message", json{{"content", content},
+                                                         {"role", "assistant"}}}}});
+
+    std::time_t t = std::time(0);
+
+    json res =
+        json{{"choices", choices},
+            {"created", t},
+            {"model",
+                json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
+            {"usage",
+                json{{"completion_tokens", num_tokens_predicted},
+                     {"prompt_tokens",     num_prompt_tokens},
+                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
+            {"id", gen_chatcmplid()}};
+
+    if (server_verbose) {
+        res["__verbose"] = result;
+    }
+
+    if (result.contains("completion_probabilities")) {
+        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
+    }
+
+    return res;
+}
+
+// return value is vector as there is one case where we might need to generate two responses
+inline static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
+    json result = response.result_json;
+
+    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
+        return std::vector<json>({response.result_json});
+    }
+
+    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
+    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
+
+    bool stopped_word   = json_value(result, "stopped_word", false);
+    bool stopped_eos    = json_value(result, "stopped_eos", false);
+    bool stopped_limit  = json_value(result, "stopped_limit", false);
+    std::string content = json_value(result, "content", std::string(""));
+
+    std::string finish_reason;
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+    if (stopped_limit) {
+        finish_reason = "length";
+    }
+
+    std::time_t t = std::time(0);
+
+    json choices;
+
+    if (!finish_reason.empty()) {
+        choices = json::array({json{{"finish_reason", finish_reason},
+                                    {"index", 0},
+                                    {"delta", json::object()}}});
+    } else {
+        if (first) {
+            if (content.empty()) {
+                choices = json::array({json{{"finish_reason", nullptr},
+                                            {"index", 0},
+                                            {"delta", json{{"role", "assistant"}}}}});
+            } else {
+                // We have to send this as two updates to conform to openai behavior
+                json initial_ret = json{{"choices", json::array({json{
+                                        {"finish_reason", nullptr},
+                                        {"index", 0},
+                                        {"delta", json{
+                                            {"role", "assistant"}
+                                        }}}})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                json second_ret = json{
+                            {"choices", json::array({json{{"finish_reason", nullptr},
+                                                            {"index", 0},
+                                                            {"delta", json{
+                                                            {"content", content}}}
+                                                            }})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                return std::vector<json>({initial_ret, second_ret});
+            }
+        } else {
+            // Some idiosyncrasy in task processing logic makes several trailing calls
+            // with empty content, we ignore these at the calee site.
+            if (content.empty()) {
+                return std::vector<json>({json::object()});
+            }
+
+            choices = json::array({json{
+                {"finish_reason", nullptr},
+                {"index", 0},
+                {"delta",
+                json{
+                    {"content", content},
+                }},
+            }});
+        }
+    }
+
+    json ret = json{{"choices", choices},
+                    {"created", t},
+                    {"id", gen_chatcmplid()},
+                    {"model", modelname},
+                    {"object", "chat.completion.chunk"}};
+
+    return std::vector<json>({ret});
+}

examples/server/utils.hpp

@@ -0,0 +1,507 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+
+#include "../llava/clip.h"
+
+using json = nlohmann::json;
+
+extern bool server_verbose;
+
+#ifndef SERVER_VERBOSE
+#define SERVER_VERBOSE 1
+#endif
+
+#if SERVER_VERBOSE != 1
+#define LOG_VERBOSE(MSG, ...)
+#else
+#define LOG_VERBOSE(MSG, ...)                                            \
+    do                                                                   \
+    {                                                                    \
+        if (server_verbose)                                              \
+        {                                                                \
+            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
+        }                                                                \
+    } while (0)
+#endif
+
+#define LOG_ERROR(  MSG, ...) server_log("ERROR",   __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(   MSG, ...) server_log("INFO",    __func__, __LINE__, MSG, __VA_ARGS__)
+
+//
+// parallel
+//
+
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
+enum task_type {
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
+    TASK_TYPE_NEXT_RESPONSE
+};
+
+struct task_server {
+    int id = -1; // to be filled by llama_server_queue
+    int target_id;
+    task_type type;
+    json data;
+    bool infill_mode = false;
+    bool embedding_mode = false;
+    int multitask_id = -1;
+};
+
+struct task_result {
+    int id;
+    int multitask_id = -1;
+    bool stop;
+    bool error;
+    json result_json;
+};
+
+struct task_multi {
+    int id;
+    std::set<int> subtasks_remaining{};
+    std::vector<task_result> results{};
+};
+
+// TODO: can become bool if we can't find use of more states
+enum slot_state
+{
+    IDLE,
+    PROCESSING,
+};
+
+enum slot_command
+{
+    NONE,
+    LOAD_PROMPT,
+    RELEASE,
+};
+
+struct slot_params
+{
+    bool stream       = true;
+    bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
+
+    uint32_t seed      = -1; // RNG seed
+    int32_t  n_keep    =  0; // number of tokens to keep from initial prompt
+    int32_t  n_predict = -1; // new tokens to predict
+
+    std::vector<std::string> antiprompt;
+
+    json input_prefix;
+    json input_suffix;
+};
+
+struct slot_image
+{
+    int32_t id;
+
+    bool request_encode_image = false;
+    float * image_embedding = nullptr;
+    int32_t image_tokens = 0;
+
+    clip_image_u8 * img_data;
+
+    std::string prefix_prompt; // before of this image
+};
+
+// completion token output with probabilities
+struct completion_token_output
+{
+    struct token_prob
+    {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+    llama_token tok;
+    std::string text_to_send;
+};
+
+static inline void server_log(const char *level, const char *function, int line,
+                       const char *message, const nlohmann::ordered_json &extra)
+{
+    nlohmann::ordered_json log
+    {
+        {"timestamp", time(nullptr)},
+        {"level",     level},
+        {"function",  function},
+        {"line",      line},
+        {"message",   message},
+    };
+
+    if (!extra.empty())
+    {
+        log.merge_patch(extra);
+    }
+
+    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
+    printf("%.*s\n", (int)str.size(), str.data());
+    fflush(stdout);
+}
+
+//
+// server utils
+//
+
+template <typename T>
+static T json_value(const json &body, const std::string &key, const T &default_value)
+{
+    // Fallback null to default value
+    return body.contains(key) && !body.at(key).is_null()
+        ? body.value(key, default_value)
+        : default_value;
+}
+
+inline std::string format_chatml(std::vector<json> messages)
+{
+    std::ostringstream chatml_msgs;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        chatml_msgs << "<|im_start|>"
+                    << json_value(*it, "role",    std::string("user")) << '\n';
+        chatml_msgs << json_value(*it, "content", std::string(""))
+                    << "<|im_end|>\n";
+    }
+
+    chatml_msgs << "<|im_start|>assistant" << '\n';
+
+    return chatml_msgs.str();
+}
+
+//
+// work queue utils
+//
+
+struct llama_server_queue {
+    int id = 0;
+    std::mutex mutex_tasks;
+    // queues
+    std::vector<task_server> queue_tasks;
+    std::vector<task_server> queue_tasks_deferred;
+    std::vector<task_multi> queue_multitasks;
+    std::condition_variable condition_tasks;
+    // callback functions
+    std::function<void(task_server&)> callback_new_task;
+    std::function<void(task_multi&)> callback_finish_multitask;
+    std::function<void(void)> callback_all_task_finished;
+
+    // Add a new task to the end of the queue
+    int post(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        if (task.id == -1) {
+            task.id = id++;
+        }
+        queue_tasks.push_back(std::move(task));
+        condition_tasks.notify_one();
+        return task.id;
+    }
+
+    // Add a new task, but defer until one slot is available
+    void defer(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        queue_tasks_deferred.push_back(std::move(task));
+    }
+
+    // Get the next id for creating anew task
+    int get_new_id() {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        return id++;
+    }
+
+    // Register function to process a new task
+    void on_new_task(std::function<void(task_server&)> callback) {
+        callback_new_task = callback;
+    }
+
+    // Register function to process a multitask
+    void on_finish_multitask(std::function<void(task_multi&)> callback) {
+        callback_finish_multitask = callback;
+    }
+
+    // Register the function to be called when the batch of tasks is finished
+    void on_all_tasks_finished(std::function<void(void)> callback) {
+        callback_all_task_finished = callback;
+    }
+
+    // Call when the state of one slot is changed
+    void notify_slot_changed() {
+        // move deferred tasks back to main loop
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        for (auto & task : queue_tasks_deferred) {
+            queue_tasks.push_back(std::move(task));
+        }
+        queue_tasks_deferred.clear();
+    }
+
+    // Start the main loop. This call is blocking
+    void start_loop() {
+        while (true) {
+            // new task arrived
+            LOG_VERBOSE("have new task", {});
+            {
+                while (true)
+                {
+                    std::unique_lock<std::mutex> lock(mutex_tasks);
+                    if (queue_tasks.empty()) {
+                        lock.unlock();
+                        break;
+                    }
+                    task_server task = queue_tasks.front();
+                    queue_tasks.erase(queue_tasks.begin());
+                    lock.unlock();
+                    LOG_VERBOSE("callback_new_task", {});
+                    callback_new_task(task);
+                }
+                LOG_VERBOSE("callback_all_task_finished", {});
+                // process and update all the multitasks
+                auto queue_iterator = queue_multitasks.begin();
+                while (queue_iterator != queue_multitasks.end())
+                {
+                    if (queue_iterator->subtasks_remaining.empty())
+                    {
+                        // all subtasks done == multitask is done
+                        task_multi current_multitask = *queue_iterator;
+                        callback_finish_multitask(current_multitask);
+                        // remove this multitask
+                        queue_iterator = queue_multitasks.erase(queue_iterator);
+                    }
+                    else
+                    {
+                        ++queue_iterator;
+                    }
+                }
+                // all tasks in the current loop is finished
+                callback_all_task_finished();
+            }
+            LOG_VERBOSE("wait for new task", {});
+            // wait for new task
+            {
+                std::unique_lock<std::mutex> lock(mutex_tasks);
+                if (queue_tasks.empty()) {
+                    condition_tasks.wait(lock, [&]{
+                        return !queue_tasks.empty();
+                    });
+                }
+            }
+        }
+    }
+
+    //
+    // functions to manage multitasks
+    //
+
+    // add a multitask by specifying the id of all subtask (subtask is a task_server)
+    void add_multitask(int multitask_id, std::vector<int>& sub_ids)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        task_multi multi;
+        multi.id = multitask_id;
+        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
+        queue_multitasks.push_back(multi);
+    }
+
+    // updatethe remaining subtasks, while appending results to multitask
+    void update_multitask(int multitask_id, int subtask_id, task_result& result)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        for (auto& multitask : queue_multitasks)
+        {
+            if (multitask.id == multitask_id)
+            {
+                multitask.subtasks_remaining.erase(subtask_id);
+                multitask.results.push_back(result);
+            }
+        }
+    }
+};
+
+struct llama_server_response {
+    typedef std::function<void(int, int, task_result&)> callback_multitask_t;
+    callback_multitask_t callback_update_multitask;
+    // for keeping track of all tasks waiting for the result
+    std::set<int> waiting_task_ids;
+    // the main result queue
+    std::vector<task_result> queue_results;
+    std::mutex mutex_results;
+    std::condition_variable condition_results;
+
+    void add_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.insert(task_id);
+    }
+
+    void remove_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.erase(task_id);
+    }
+
+    // This function blocks the thread until there is a response for this task_id
+    task_result recv(int task_id) {
+        while (true)
+        {
+            std::unique_lock<std::mutex> lock(mutex_results);
+            condition_results.wait(lock, [&]{
+                return !queue_results.empty();
+            });
+            LOG_VERBOSE("condition_results unblock", {});
+
+            for (int i = 0; i < (int) queue_results.size(); i++)
+            {
+                if (queue_results[i].id == task_id)
+                {
+                    assert(queue_results[i].multitask_id == -1);
+                    task_result res = queue_results[i];
+                    queue_results.erase(queue_results.begin() + i);
+                    return res;
+                }
+            }
+        }
+
+        // should never reach here
+    }
+
+    // Register the function to update multitask
+    void on_multitask_update(callback_multitask_t callback) {
+        callback_update_multitask = callback;
+    }
+
+    // Send a new result to a waiting task_id
+    void send(task_result result) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        LOG_VERBOSE("send new result", {});
+        for (auto& task_id : waiting_task_ids) {
+            // LOG_TEE("waiting task id %i \n", task_id);
+            // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
+            if (result.multitask_id == task_id)
+            {
+                LOG_VERBOSE("callback_update_multitask", {});
+                callback_update_multitask(task_id, result.id, result);
+                continue;
+            }
+
+            if (result.id == task_id)
+            {
+                LOG_VERBOSE("queue_results.push_back", {});
+                queue_results.push_back(result);
+                condition_results.notify_one();
+                return;
+            }
+        }
+    }
+};
+
+//
+// base64 utils (TODO: move to common in the future)
+//
+
+static const std::string base64_chars =
+             "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+             "abcdefghijklmnopqrstuvwxyz"
+             "0123456789+/";
+
+static inline bool is_base64(uint8_t c)
+{
+    return (isalnum(c) || (c == '+') || (c == '/'));
+}
+
+static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string)
+{
+    int i = 0;
+    int j = 0;
+    int in_ = 0;
+
+    int in_len = encoded_string.size();
+
+    uint8_t char_array_4[4];
+    uint8_t char_array_3[3];
+
+    std::vector<uint8_t> ret;
+
+    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_]))
+    {
+        char_array_4[i++] = encoded_string[in_]; in_++;
+        if (i == 4)
+        {
+            for (i = 0; i <4; i++)
+            {
+                char_array_4[i] = base64_chars.find(char_array_4[i]);
+            }
+
+            char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+            char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+            char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+            for (i = 0; (i < 3); i++)
+            {
+                ret.push_back(char_array_3[i]);
+            }
+            i = 0;
+        }
+    }
+
+    if (i)
+    {
+        for (j = i; j <4; j++)
+        {
+            char_array_4[j] = 0;
+        }
+
+        for (j = 0; j <4; j++)
+        {
+            char_array_4[j] = base64_chars.find(char_array_4[j]);
+        }
+
+        char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+        char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+        char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+        for (j = 0; (j < i - 1); j++)
+        {
+            ret.push_back(char_array_3[j]);
+        }
+    }
+
+    return ret;
+}
+
+//
+// random string / id
+//
+
+static std::string random_string()
+{
+    static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
+
+    std::random_device rd;
+    std::mt19937 generator(rd());
+
+    std::string result(32, ' ');
+
+    for (int i = 0; i < 32; ++i) {
+        result[i] = str[generator() % str.size()];
+    }
+
+    return result;
+}
+
+static std::string gen_chatcmplid()
+{
+    std::stringstream chatcmplid;
+    chatcmplid << "chatcmpl-" << random_string();
+    return chatcmplid.str();
+}

ggml-alloc.c

@@ -335,7 +335,9 @@ bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) {
 }
 
 size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) {
-    return alloc->max_size;
+    // FIXME: changes in the tensor sizes compared to the measure graph may cause allocations to fail
+    // to avoid this, we add a 10% margin to the buffer size
+    return alloc->max_size + alloc->max_size/10;
 }
 
 // graph allocator

ggml-backend.c

@@ -30,7 +30,9 @@ size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
 GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
     // get_alloc_size is optional, defaults to ggml_nbytes
     if (buft->iface.get_alloc_size) {
-        return buft->iface.get_alloc_size(buft, tensor);
+        size_t size = buft->iface.get_alloc_size(buft, tensor);
+        assert(size >= ggml_nbytes(tensor));
+        return size;
     }
     return ggml_nbytes(tensor);
 }

ggml-cuda.cu

@@ -9790,8 +9790,8 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
     // TODO: mmq/mmv support
 #endif
 
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb1  =  dst->nb[1];
+    const size_t nb11 = src1->nb[1];
+    const size_t nb1  =  dst->nb[1];
 
     const struct ggml_tensor * ids = src0;
     const int32_t id = ((int32_t *) dst->op_params)[0];
@@ -10304,15 +10304,11 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t
 
     if (ggml_is_quantized(tensor->type)) {
         // initialize padding to 0 to avoid possible NaN values
-        int64_t row_low = 0;
-        int64_t row_high = ggml_nrows(tensor);
-        int64_t nrows_split = row_high - row_low;
-
-        size_t original_size = ggml_nbytes_split(tensor, nrows_split);
+        size_t original_size = ggml_nbytes(tensor);
         size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
 
         if (padded_size > original_size && tensor->view_src == nullptr) {
-            CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[ctx->device][0]));
+            CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
         }
     }
 }
@@ -10415,12 +10411,7 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend
 }
 
 GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
-    int64_t row_low = 0;
-    int64_t row_high = ggml_nrows(tensor);
-    int64_t nrows_split = row_high - row_low;
-
-    size_t size = ggml_nbytes_split(tensor, nrows_split);
-
+    size_t size = ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
 
     if (ggml_is_quantized(tensor->type)) {

ggml-metal.m

@@ -26,15 +26,6 @@
 
 #define GGML_METAL_MAX_KERNELS 256
 
-struct ggml_metal_buffer {
-    const char * name;
-
-    void   * data;
-    size_t   size;
-
-    id<MTLBuffer> metal;
-};
-
 struct ggml_metal_kernel {
     id<MTLFunction>             function;
     id<MTLComputePipelineState> pipeline;
@@ -172,9 +163,6 @@ struct ggml_metal_context {
 
     dispatch_queue_t d_queue;
 
-    int n_buffers;
-    struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
-
     struct ggml_metal_kernel kernels[GGML_METAL_MAX_KERNELS];
 
     bool support_simdgroup_reduction;
@@ -242,24 +230,20 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     // Show all the Metal device instances in the system
     NSArray * devices = MTLCopyAllDevices();
     for (id<MTLDevice> device in devices) {
-        NSString * s = [device name];
-        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]);
+        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
     }
     [devices release]; // since it was created by a *Copy* C method
 #endif
 
     // Pick and show default Metal device
     id<MTLDevice> device = MTLCreateSystemDefaultDevice();
-    NSString * s = [device name];
-    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]);
+    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
     // Configure context
     struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
     ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
-    ctx->n_buffers = 0;
-
     ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
     // load library
@@ -534,10 +518,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 static void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 
-    for (int i = 0; i < ctx->n_buffers; ++i) {
-        [ctx->buffers[i].metal release];
-    }
-
     for (int i = 0; i < GGML_METAL_MAX_KERNELS; ++i) {
         if (ctx->kernels[i].pipeline) {
             [ctx->kernels[i].pipeline release];
@@ -580,51 +560,30 @@ struct ggml_backend_metal_buffer_context {
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
 //
-static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
+static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs) {
     //GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
 
     const int64_t tsize = ggml_nbytes(t);
 
     ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
 
-    // compatibility with ggml-backend
-    if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) {
-        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
-
-        // find the view that contains the tensor fully
-        for (int i = 0; i < buf_ctx->n_buffers; ++i) {
-            const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
-
-            //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
-            if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
-                *offs = (size_t) ioffs;
-
-                //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
-
-                return buf_ctx->buffers[i].metal;
-            }
-        }
-
-        GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
-
-        return nil;
-    }
+    struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
 
     // find the view that contains the tensor fully
-    for (int i = 0; i < ctx->n_buffers; ++i) {
-        const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
+    for (int i = 0; i < buf_ctx->n_buffers; ++i) {
+        const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
 
-        //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
-        if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
+        //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
+        if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
             *offs = (size_t) ioffs;
 
-            //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+            //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
 
-            return ctx->buffers[i].metal;
+            return buf_ctx->buffers[i].metal;
         }
     }
 
-    GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__);
+    GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
 
     return nil;
 }
@@ -817,9 +776,9 @@ static bool ggml_metal_graph_compute(
             const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
             const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
 
-            id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
-            id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
-            id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
+            id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
+            id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
+            id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(dst,  &offs_dst)  : nil;
 
             //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
             //if (src0) {
@@ -1601,7 +1560,7 @@ static bool ggml_metal_graph_compute(
                                 struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                 size_t offs_src_cur = 0;
-                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
+                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur);
 
                                 [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j];
                             }
@@ -1746,7 +1705,7 @@ static bool ggml_metal_graph_compute(
                                 struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                 size_t offs_src_cur = 0;
-                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
+                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur);
 
                                 [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
                             }

ggml.c

@@ -16597,7 +16597,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_SOFT_MAX:
             {
-                n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
+                n_tasks = MIN(n_threads, ggml_nrows(node->src[0]));
             } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {

gguf-py/gguf/gguf_reader.py

@@ -107,7 +107,7 @@ class GGUFReader:
         offs, tensors_fields = self._build_tensors_fields(offs, tensor_count)
         new_align = self.fields.get('general.alignment')
         if new_align is not None:
-            if new_align.types != [GGUFValueType.UINT64]:
+            if new_align.types != [GGUFValueType.UINT32]:
                 raise ValueError('Bad type for general.alignment field')
             self.alignment = new_align.parts[-1][0]
         padding = offs % self.alignment

llama.cpp

@@ -7956,10 +7956,57 @@ void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * can
         auto comp = [](const llama_token_data & a, const llama_token_data & b) {
             return a.logit > b.logit;
         };
-        if (k == (int) candidates->size) {
-            std::sort(candidates->data, candidates->data + candidates->size, comp);
-        } else {
+        if (k <= 128) {
             std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp);
+        } else {
+            constexpr int   nbuckets     = 128;
+            constexpr float bucket_low   = -10.0f;
+            constexpr float bucket_high  =  10.0f;
+            constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
+            constexpr float bucker_inter = -bucket_low * bucket_scale;
+
+            std::vector<int> bucket_idx(candidates->size);
+            std::vector<int> histo(nbuckets, 0);
+
+            for (int i = 0; i < (int)candidates->size; ++i) {
+                const float val = candidates->data[i].logit;
+                int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
+                ib = std::max(0, std::min(nbuckets-1, ib));
+                bucket_idx[i] = ib;
+                ++histo[ib];
+            }
+            int nhave = 0;
+            int ib = nbuckets - 1;
+            for ( ; ib >= 0; --ib) {
+                nhave += histo[ib];
+                if (nhave >= k) break;
+            }
+            std::vector<llama_token_data> tmp_tokens(nhave);
+            auto ptr = tmp_tokens.data();
+            std::vector<llama_token_data*> bucket_ptrs;
+            bucket_ptrs.reserve(nbuckets - ib);
+            for (int j = nbuckets - 1; j >= ib; --j) {
+                bucket_ptrs.push_back(ptr);
+                ptr += histo[j];
+            }
+            for (int i = 0; i < (int)candidates->size; ++i) {
+                int j = bucket_idx[i];
+                if (j >= ib) {
+                    *bucket_ptrs[nbuckets-1-j]++ = candidates->data[i];
+                }
+            }
+
+            ptr = tmp_tokens.data();
+            int ndone = 0;
+            for (int j = nbuckets-1; j > ib; --j) {
+                std::sort(ptr, ptr + histo[j], comp);
+                ptr += histo[j];
+                ndone += histo[j];
+            }
+            std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp);
+
+            std::memcpy(candidates->data, tmp_tokens.data(), k*sizeof(llama_token_data));
+
         }
         candidates->sorted = true;
     }

scripts/ci-run.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+set -euo pipefail
+this=$(realpath "$0"); readonly this
+cd "$(dirname "$this")"
+shellcheck "$this"
+
+if (( $# != 1 && $# != 2  )); then
+    cat >&2 <<'EOF'
+usage:
+    ci-run.sh <tmp_dir> [<cache_dir>]
+
+This script wraps ci/run.sh:
+* If <tmp_dir> is a ramdisk, you can reduce writes to your SSD. If <tmp_dir> is not a ramdisk, keep in mind that total writes will increase by the size of <cache_dir>.
+    (openllama_3b_v2: quantized models are about 30GB)
+* Persistent model and data files are synced to and from <cache_dir>,
+    excluding generated .gguf files.
+    (openllama_3b_v2: persistent files are about 6.6GB)
+* <cache_dir> defaults to  ~/.cache/llama.cpp
+EOF
+    exit 1
+fi
+
+cd .. # => llama.cpp repo root
+
+tmp="$1"
+mkdir -p "$tmp"
+tmp=$(realpath "$tmp")
+echo >&2 "Using tmp=$tmp"
+
+cache="${2-$HOME/.cache/llama.cpp}"
+mkdir -p "$cache"
+cache=$(realpath "$cache")
+echo >&2 "Using cache=$cache"
+
+_sync() {
+    local from="$1"; shift
+    local to="$1"; shift
+
+    echo >&2 "Syncing from $from to $to"
+    mkdir -p "$from" "$to"
+    rsync -a "$from" "$to" --delete-during "$@"
+}
+
+_sync "$(realpath .)/" "$tmp/llama.cpp"
+_sync "$cache/ci-mnt/models/" "$tmp/llama.cpp/ci-mnt/models/"
+
+cd "$tmp/llama.cpp"
+bash ci/run.sh ci-out ci-mnt
+
+_sync 'ci-mnt/models/' "$cache/ci-mnt/models/" --exclude='*.gguf' -P

scripts/run-with-preset.py

@@ -46,7 +46,7 @@ Formatting considerations:
 - To define multiple "reverse_prompt" properties simultaneously the expected format is a list of strings.
 - To define a tensor split, pass a list of floats.
 """
-usage = "run_with_preset.py [-h] [yaml_files ...] [--<ARG_NAME> <ARG_VALUE> ...]"
+usage = "run-with-preset.py [-h] [yaml_files ...] [--<ARG_NAME> <ARG_VALUE> ...]"
 epilog = ("  --<ARG_NAME> specify additional CLI ars to be passed to the binary (override all preset files). "
           "Unknown args will be ignored.")
 

tests/.gitignore

@@ -0,0 +1,3 @@
+*
+!*.*
+test-c.o

tests/CMakeLists.txt

@@ -1,6 +1,6 @@
 function(llama_build_executable source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
-    add_executable(${TEST_TARGET} ${source})
+    add_executable(${TEST_TARGET} ${source} get-model.cpp)
     install(TARGETS ${TEST_TARGET} RUNTIME)
     target_link_libraries(${TEST_TARGET} PRIVATE common)
 endfunction()
@@ -8,14 +8,20 @@ endfunction()
 function(llama_test_executable name source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
     add_test(NAME ${name} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
+    set_property(TEST ${name} PROPERTY LABELS "main")
 endfunction()
 
 function(llama_build_and_test_executable source)
+    llama_build_and_test_executable_with_label(${source} "main")
+endfunction()
+
+function(llama_build_and_test_executable_with_label source label)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
-    add_executable(${TEST_TARGET} ${source})
+    add_executable(${TEST_TARGET} ${source} get-model.cpp)
     install(TARGETS ${TEST_TARGET} RUNTIME)
     target_link_libraries(${TEST_TARGET} PRIVATE common)
     add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
+    set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${label})
 endfunction()
 
 # llama_build_and_test_executable(test-double-float.cpp) # SLOW
@@ -49,10 +55,12 @@ llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp)
 # llama_build_and_test_executable(test-opt.cpp) # SLOW
 llama_build_and_test_executable(test-backend-ops.cpp)
-llama_build_and_test_executable(test-autorelease.cpp)
 
 llama_build_and_test_executable(test-rope.cpp)
 
+llama_build_and_test_executable_with_label(test-model-load-cancel.cpp "model")
+llama_build_and_test_executable_with_label(test-autorelease.cpp "model")
+
 # dummy executable - not installed
 get_filename_component(TEST_TARGET test-c.c NAME_WE)
 add_executable(${TEST_TARGET} test-c.c)

tests/get-model.cpp

@@ -0,0 +1,21 @@
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+#include "get-model.h"
+
+char * get_model_or_exit(int argc, char *argv[]) {
+    char * model_path;
+    if (argc > 1) {
+        model_path = argv[1];
+
+    } else {
+        model_path = getenv("LLAMACPP_TEST_MODELFILE");
+        if (!model_path || strlen(model_path) == 0) {
+            fprintf(stderr, "\033[33mWARNING: No model file provided. Skipping this test. Set LLAMACPP_TEST_MODELFILE=<gguf_model_path> to silence this warning and run this test.\n\033[0m");
+            exit(EXIT_SUCCESS);
+        }
+    }
+
+    return model_path;
+}

tests/get-model.h

@@ -0,0 +1,2 @@
+#pragma once
+char * get_model_or_exit(int, char*[]);

tests/test-autorelease.cpp

@@ -5,19 +5,15 @@
 #include <thread>
 
 #include "llama.h"
+#include "get-model.h"
 
 // This creates a new context inside a pthread and then tries to exit cleanly.
 int main(int argc, char ** argv) {
-    if (argc < 2) {
-        printf("Usage: %s model.gguf\n", argv[0]);
-        return 0; // intentionally return success
-    }
+    auto * model_path = get_model_or_exit(argc, argv);
 
-    const std::string fname = argv[1];
-
-    std::thread([&fname]() {
+    std::thread([&model_path]() {
         llama_backend_init(false);
-        auto * model = llama_load_model_from_file(fname.c_str(), llama_model_default_params());
+        auto * model = llama_load_model_from_file(model_path, llama_model_default_params());
         auto * ctx = llama_new_context_with_model(model, llama_context_default_params());
         llama_free(ctx);
         llama_free_model(model);

tests/test-model-load-cancel.cpp

@@ -0,0 +1,27 @@
+#include "llama.h"
+#include "get-model.h"
+
+#include <cstdlib>
+
+int main(int argc, char *argv[] ) {
+    auto * model_path = get_model_or_exit(argc, argv);
+    auto * file = fopen(model_path, "r");
+    if (file == nullptr) {
+        fprintf(stderr, "no model at '%s' found\n", model_path);
+        return EXIT_FAILURE;
+    }
+
+    fprintf(stderr, "using '%s'\n", model_path);
+    fclose(file);
+
+    llama_backend_init(false);
+    auto params = llama_model_params{};
+    params.use_mmap = false;
+    params.progress_callback = [](float progress, void * ctx){
+        (void) ctx;
+        return progress > 0.50;
+    };
+    auto * model = llama_load_model_from_file(model_path, params);
+    llama_backend_free();
+    return model == nullptr ? EXIT_SUCCESS : EXIT_FAILURE;
+}