cmake : use list(APPEND ...) instead of set() + dedup linker (#9463)

* cmake : use list(APPEND ...) instead of set() + dedup linker

ggml-ci

* cmake : try fix sycl

* cmake : try to fix sycl 2

* cmake : fix sycl build (#9469)

* try fix sycl build

* use CMAKE_CXX_FLAGS as a string variable

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* one more CMAKE_CXX_FLAGS fix (#9471)

---------

Co-authored-by: Michael Podvitskiy <podvitskiymichael@gmail.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/docker.yml

@@ -37,9 +37,9 @@ jobs:
           - { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          # Note: the full-rocm image is failing due to a "no space left on device" error. It is disabled for now to allow the workflow to complete.
+          # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
+          #- { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          #- { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           #- { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
           - { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }

.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

.gitignore

@@ -61,6 +61,7 @@ llama-batched-swift
 /rpc-server
 out/
 tmp/
+autogen-*.md
 
 # Deprecated
 

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

@@ -39,10 +39,12 @@ BUILD_TARGETS = \
 	llama-tokenize \
 	llama-vdot \
 	llama-cvector-generator \
+	llama-gen-docs \
 	tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
+	tests/test-arg-parser \
 	tests/test-autorelease \
 	tests/test-backend-ops \
 	tests/test-chat-template \
@@ -432,7 +434,7 @@ endif
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 
-ifndef RISCV
+ifndef RISCV_CROSS_COMPILE
 
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
@@ -512,7 +514,12 @@ ifneq ($(filter loongarch64%,$(UNAME_M)),)
 	MK_CXXFLAGS += -mlasx
 endif
 
-else
+ifneq ($(filter riscv64%,$(UNAME_M)),)
+	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
+	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
+endif
+
+else # RISC-V CROSS COMPILATION
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
@@ -923,11 +930,11 @@ OBJ_LLAMA = \
 
 OBJ_COMMON = \
 	common/common.o \
+	common/arg.o \
 	common/console.o \
 	common/ngram-cache.o \
 	common/sampling.o \
 	common/train.o \
-	common/grammar-parser.o \
 	common/build-info.o \
 	common/json-schema-to-grammar.o
 
@@ -1156,6 +1163,11 @@ common/common.o: \
 	include/llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+common/arg.o: \
+	common/arg.cpp \
+	common/arg.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 common/sampling.o: \
 	common/sampling.cpp \
 	common/sampling.h \
@@ -1167,11 +1179,6 @@ common/console.o: \
 	common/console.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common/grammar-parser.o: \
-	common/grammar-parser.cpp \
-	common/grammar-parser.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 common/json-schema-to-grammar.o: \
 	common/json-schema-to-grammar.cpp \
 	common/json-schema-to-grammar.h
@@ -1433,6 +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)
@@ -1448,6 +1456,11 @@ examples/server/%.hpp: examples/server/public/% Makefile
 		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
 	) > $@
 
+llama-gen-docs: examples/gen-docs/gen-docs.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 libllava.a: examples/llava/llava.cpp \
 	examples/llava/llava.h \
 	examples/llava/clip.cpp \
@@ -1505,6 +1518,11 @@ run-benchmark-matmult: llama-benchmark-matmult
 
 .PHONY: run-benchmark-matmult swift
 
+tests/test-arg-parser: tests/test-arg-parser.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 tests/test-llama-grammar: tests/test-llama-grammar.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

README.md

@@ -17,7 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- *add hot topics here*
+- Huggingface GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
 ----
 
@@ -89,6 +89,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
 - [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
+- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -163,6 +164,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
 - [AIKit](https://github.com/sozercan/aikit) (MIT)
 - [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
+- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 

common/CMakeLists.txt

@@ -54,12 +54,12 @@ add_library(${TARGET} STATIC
     base64.hpp
     common.h
     common.cpp
+    arg.h
+    arg.cpp
     sampling.h
     sampling.cpp
     console.h
     console.cpp
-    grammar-parser.h
-    grammar-parser.cpp
     json.hpp
     json-schema-to-grammar.cpp
     train.h

common/arg.h

@@ -0,0 +1,77 @@
+#pragma once
+
+#include "common.h"
+
+#include <set>
+#include <string>
+#include <vector>
+
+//
+// CLI argument parsing
+//
+
+struct llama_arg {
+    std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
+    std::vector<const char *> args;
+    const char * value_hint   = nullptr; // help text or example for arg value
+    const char * value_hint_2 = nullptr; // for second arg value
+    const char * env          = nullptr;
+    std::string help;
+    bool is_sparam = false; // is current arg a sampling param?
+    void (*handler_void)   (gpt_params & params) = nullptr;
+    void (*handler_string) (gpt_params & params, const std::string &) = nullptr;
+    void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr;
+    void (*handler_int)    (gpt_params & params, int) = nullptr;
+
+    llama_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const std::string & help,
+        void (*handler)(gpt_params & params, const std::string &)
+    ) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
+
+    llama_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const std::string & help,
+        void (*handler)(gpt_params & params, int)
+    ) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
+
+    llama_arg(
+        const std::initializer_list<const char *> & args,
+        const std::string & help,
+        void (*handler)(gpt_params & params)
+    ) : args(args), help(help), handler_void(handler) {}
+
+    // support 2 values for arg
+    llama_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const char * value_hint_2,
+        const std::string & help,
+        void (*handler)(gpt_params & params, const std::string &, const std::string &)
+    ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
+
+    llama_arg & set_examples(std::initializer_list<enum llama_example> examples);
+    llama_arg & set_env(const char * env);
+    llama_arg & set_sparam();
+    bool in_example(enum llama_example ex);
+    bool get_value_from_env(std::string & output);
+    bool has_value_from_env();
+    std::string to_string();
+};
+
+struct gpt_params_context {
+    enum llama_example ex = LLAMA_EXAMPLE_COMMON;
+    gpt_params & params;
+    std::vector<llama_arg> options;
+    void(*print_usage)(int, char **) = nullptr;
+    gpt_params_context(gpt_params & params) : params(params) {}
+};
+
+// parse input arguments from CLI
+// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+
+// function to be used by test-arg-parser
+gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/common.h

@@ -4,18 +4,11 @@
 
 #include "llama.h"
 
-#include "sampling.h"
-
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
-#include <cmath>
 #include <string>
 #include <vector>
-#include <random>
-#include <thread>
-#include <unordered_map>
-#include <tuple>
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -54,19 +47,6 @@ struct llama_control_vector_load_info;
 // CPU utils
 //
 
-int32_t cpu_get_num_physical_cores();
-int32_t cpu_get_num_math();
-
-//
-// CLI argument parsing
-//
-
-// dimensionality reduction methods, used by cvector-generator
-enum dimre_method {
-    DIMRE_METHOD_PCA,
-    DIMRE_METHOD_MEAN,
-};
-
 struct cpu_params {
     int      n_threads                   = -1;
     bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
@@ -76,9 +56,94 @@ struct cpu_params {
     uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
 };
 
-struct gpt_params {
-    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
+int32_t cpu_get_num_physical_cores();
+int32_t cpu_get_num_math();
 
+//
+// Common params
+//
+
+enum llama_example {
+    LLAMA_EXAMPLE_COMMON,
+    LLAMA_EXAMPLE_SPECULATIVE,
+    LLAMA_EXAMPLE_MAIN,
+    LLAMA_EXAMPLE_INFILL,
+    LLAMA_EXAMPLE_EMBEDDING,
+    LLAMA_EXAMPLE_PERPLEXITY,
+    LLAMA_EXAMPLE_RETRIEVAL,
+    LLAMA_EXAMPLE_PASSKEY,
+    LLAMA_EXAMPLE_IMATRIX,
+    LLAMA_EXAMPLE_BENCH,
+    LLAMA_EXAMPLE_SERVER,
+    LLAMA_EXAMPLE_CVECTOR_GENERATOR,
+    LLAMA_EXAMPLE_EXPORT_LORA,
+    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_LOOKUP,
+    LLAMA_EXAMPLE_PARALLEL,
+
+    LLAMA_EXAMPLE_COUNT,
+};
+
+enum gpt_sampler_type {
+    GPT_SAMPLER_TYPE_NONE        = 0,
+    GPT_SAMPLER_TYPE_TOP_K       = 1,
+    GPT_SAMPLER_TYPE_TOP_P       = 2,
+    GPT_SAMPLER_TYPE_MIN_P       = 3,
+    GPT_SAMPLER_TYPE_TFS_Z       = 4,
+    GPT_SAMPLER_TYPE_TYPICAL_P   = 5,
+    GPT_SAMPLER_TYPE_TEMPERATURE = 6,
+};
+
+// dimensionality reduction methods, used by cvector-generator
+enum dimre_method {
+    DIMRE_METHOD_PCA,
+    DIMRE_METHOD_MEAN,
+};
+
+// sampler parameters
+struct gpt_sampler_params {
+    uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
+
+    int32_t n_prev            = 64;    // number of previous tokens to remember
+    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t min_keep          = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t top_k             = 40;    // <= 0 to use vocab size
+    float   top_p             = 0.95f; // 1.0 = disabled
+    float   min_p             = 0.05f; // 0.0 = disabled
+    float   tfs_z             = 1.00f; // 1.0 = disabled
+    float   typ_p             = 1.00f; // typical_p, 1.0 = disabled
+    float   temp              = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float   dynatemp_range    = 0.00f; // 0.0 = disabled
+    float   dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat    = 1.00f; // 1.0 = disabled
+    float   penalty_freq      = 0.00f; // 0.0 = disabled
+    float   penalty_present   = 0.00f; // 0.0 = disabled
+    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   mirostat_tau      = 5.00f; // target entropy
+    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,
+        GPT_SAMPLER_TYPE_TFS_Z,
+        GPT_SAMPLER_TYPE_TYPICAL_P,
+        GPT_SAMPLER_TYPE_TOP_P,
+        GPT_SAMPLER_TYPE_MIN_P,
+        GPT_SAMPLER_TYPE_TEMPERATURE
+    };
+
+    std::string grammar; // optional BNF-like grammar to constrain sampling
+
+    std::vector<llama_logit_bias> logit_bias; // logit biases to apply
+
+    // print the parameters into a string
+    std::string print() const;
+};
+
+struct gpt_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =     0; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
@@ -120,26 +185,25 @@ struct gpt_params {
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
-    // // sampling parameters
-    struct llama_sampling_params sparams;
+    struct gpt_sampler_params sparams;
 
-    std::string model                = ""; // model path
-    std::string model_draft          = ""; // draft model for speculative decoding
-    std::string model_alias          = "unknown"; // model alias
-    std::string model_url            = ""; // model url to download
-    std::string hf_token             = ""; // HF token
-    std::string hf_repo              = ""; // HF repo
-    std::string hf_file              = ""; // HF file
-    std::string prompt               = "";
-    std::string prompt_file          = ""; // store the external prompt file name
-    std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state
-    std::string input_prefix         = ""; // string to prefix user inputs with
-    std::string input_suffix         = ""; // string to suffix user inputs with
-    std::string logdir               = ""; // directory in which to save YAML log files
-    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding
-    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
-    std::string logits_file          = ""; // file for saving *all* logits
-    std::string rpc_servers          = ""; // comma separated list of RPC servers
+    std::string model                = ""; // model path                                                    // NOLINT
+    std::string model_draft          = ""; // draft model for speculative decoding                          // NOLINT
+    std::string model_alias          = "unknown"; // model alias                                            // NOLINT
+    std::string model_url            = ""; // model url to download                                         // NOLINT
+    std::string hf_token             = ""; // HF token                                                      // NOLINT
+    std::string hf_repo              = ""; // HF repo                                                       // NOLINT
+    std::string hf_file              = ""; // HF file                                                       // NOLINT
+    std::string prompt               = "";                                                                  // NOLINT
+    std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
+    std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
+    std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
+    std::string input_suffix         = ""; // string to suffix user inputs with                             // NOLINT
+    std::string logdir               = ""; // directory in which to save YAML log files                     // NOLINT
+    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
+    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
+    std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
+    std::string rpc_servers          = ""; // comma separated list of RPC servers                           // NOLINT
 
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
@@ -183,15 +247,14 @@ 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 ignore_eos        = false; // ignore generated EOS tokens
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
-    bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
@@ -201,7 +264,7 @@ struct gpt_params {
     std::string cache_type_v = "f16"; // KV cache data type for the V
 
     // multimodal models (see examples/llava)
-    std::string mmproj = "";        // path to multimodal projector
+    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -217,15 +280,15 @@ struct gpt_params {
     int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
 
     std::string hostname      = "127.0.0.1";
-    std::string public_path   = "";
-    std::string chat_template = "";
-    std::string system_prompt = "";
+    std::string public_path   = "";                                                                         // NOLINT
+    std::string chat_template = "";                                                                         // NOLINT
+    std::string system_prompt = "";                                                                         // NOLINT
     bool enable_chat_template = true;
 
     std::vector<std::string> api_keys;
 
-    std::string ssl_file_key  = "";
-    std::string ssl_file_cert = "";
+    std::string ssl_file_key  = "";                                                                         // NOLINT
+    std::string ssl_file_cert = "";                                                                         // NOLINT
 
     bool endpoint_slots   = true;
     bool endpoint_metrics = false;
@@ -280,14 +343,6 @@ struct gpt_params {
     bool batched_bench_output_jsonl = false;
 };
 
-void gpt_params_parse_from_env(gpt_params & params);
-void gpt_params_handle_model_default(gpt_params & params);
-
-bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
-bool gpt_params_parse      (int argc, char ** argv, gpt_params & params);
-bool gpt_params_find_arg   (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
-void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
-
 std::string gpt_params_get_system_info(const gpt_params & params);
 
 bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);

common/grammar-parser.cpp

@@ -1,539 +0,0 @@
-#include "grammar-parser.h"
-#include <cstdint>
-#include <cwchar>
-#include <string>
-#include <utility>
-#include <stdexcept>
-#include <exception>
-
-namespace grammar_parser {
-    // NOTE: assumes valid utf8 (but checks for overrun)
-    // copied from llama.cpp
-    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
-        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
-        uint8_t  first_byte = static_cast<uint8_t>(*src);
-        uint8_t  highbits   = first_byte >> 4;
-        int      len        = lookup[highbits];
-        uint8_t  mask       = (1 << (8 - len)) - 1;
-        uint32_t value      = first_byte & mask;
-        const char * end    = src + len; // may overrun!
-        const char * pos    = src + 1;
-        for ( ; pos < end && *pos; pos++) {
-            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
-        }
-        return std::make_pair(value, pos);
-    }
-
-    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
-        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
-        auto result = state.symbol_ids.emplace(std::string(src, len), next_id);
-        return result.first->second;
-    }
-
-    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
-        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
-        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
-        return next_id;
-    }
-
-    static void add_rule(
-            parse_state & state,
-            uint32_t      rule_id,
-            const std::vector<llama_grammar_element> & rule) {
-        if (state.rules.size() <= rule_id) {
-            state.rules.resize(rule_id + 1);
-        }
-        state.rules[rule_id] = rule;
-    }
-
-    static bool is_digit_char(char c) {
-        return '0' <= c && c <= '9';
-    }
-
-    static bool is_word_char(char c) {
-        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || is_digit_char(c);
-    }
-
-    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
-        const char * pos   = src;
-        const char * end   = src + size;
-        uint32_t     value = 0;
-        for ( ; pos < end && *pos; pos++) {
-            value <<= 4;
-            char c = *pos;
-            if ('a' <= c && c <= 'f') {
-                value += c - 'a' + 10;
-            } else if ('A' <= c && c <= 'F') {
-                value += c - 'A' + 10;
-            } else if ('0' <= c && c <= '9') {
-                value += c - '0';
-            } else {
-                break;
-            }
-        }
-        if (pos != end) {
-            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
-        }
-        return std::make_pair(value, pos);
-    }
-
-    static const char * parse_space(const char * src, bool newline_ok) {
-        const char * pos = src;
-        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
-                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
-            if (*pos == '#') {
-                while (*pos && *pos != '\r' && *pos != '\n') {
-                    pos++;
-                }
-            } else {
-                pos++;
-            }
-        }
-        return pos;
-    }
-
-    static const char * parse_name(const char * src) {
-        const char * pos = src;
-        while (is_word_char(*pos)) {
-            pos++;
-        }
-        if (pos == src) {
-            throw std::runtime_error(std::string("expecting name at ") + src);
-        }
-        return pos;
-    }
-
-    static const char * parse_int(const char * src) {
-        const char * pos = src;
-        while (is_digit_char(*pos)) {
-            pos++;
-        }
-        if (pos == src) {
-            throw std::runtime_error(std::string("expecting integer at ") + src);
-        }
-        return pos;
-    }
-
-    static std::pair<uint32_t, const char *> parse_char(const char * src) {
-        if (*src == '\\') {
-            switch (src[1]) {
-                case 'x': return parse_hex(src + 2, 2);
-                case 'u': return parse_hex(src + 2, 4);
-                case 'U': return parse_hex(src + 2, 8);
-                case 't': return std::make_pair('\t', src + 2);
-                case 'r': return std::make_pair('\r', src + 2);
-                case 'n': return std::make_pair('\n', src + 2);
-                case '\\':
-                case '"':
-                case '[':
-                case ']':
-                    return std::make_pair(src[1], src + 2);
-                default:
-                    throw std::runtime_error(std::string("unknown escape at ") + src);
-            }
-        } else if (*src) {
-            return decode_utf8(src);
-        }
-        throw std::runtime_error("unexpected end of input");
-    }
-
-    const char * parse_alternates(
-            parse_state       & state,
-            const char        * src,
-            const std::string & rule_name,
-            uint32_t            rule_id,
-            bool                is_nested);
-
-    static const char * parse_sequence(
-            parse_state                        & state,
-            const char                         * src,
-            const std::string                  & rule_name,
-            std::vector<llama_grammar_element> & out_elements,
-            bool                                 is_nested) {
-        size_t last_sym_start = out_elements.size();
-        const char * pos = src;
-
-        auto handle_repetitions = [&](int min_times, int max_times) {
-
-            if (last_sym_start == out_elements.size()) {
-                throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
-            }
-
-            // apply transformation to previous symbol (last_sym_start to end) according to
-            // the following rewrite rules:
-            // S{m,n} --> S S S (m times) S'(n-m)
-            //            S'(x)   ::= S S'(x-1) |
-            //            (... n-m definitions of these S' rules ...)
-            //            S'(1)   ::= S |
-            // S{m,} -->  S S S (m times) S'
-            //            S'     ::= S S' |
-            // S*     --> S{0,}
-            //        --> S'     ::= S S' |
-            // S+     --> S{1,}
-            //        --> S S'
-            //            S'     ::= S S' |
-            // S?     --> S{0,1}
-            //        --> S'
-            //            S'     ::= S |
-
-            std::vector<llama_grammar_element> previous_elements(out_elements.begin() + last_sym_start, out_elements.end());
-            if (min_times == 0) {
-                out_elements.resize(last_sym_start);
-            } else {
-                // Repeat the previous elements (min_times - 1) times
-                for (int i = 1; i < min_times; i++) {
-                    out_elements.insert(out_elements.end(), previous_elements.begin(), previous_elements.end());
-                }
-            }
-
-            uint32_t last_rec_rule_id = 0;
-            auto n_opt = max_times < 0 ? 1 : max_times - min_times;
-
-            std::vector<llama_grammar_element> rec_rule(previous_elements);
-            for (int i = 0; i < n_opt; i++) {
-                rec_rule.resize(previous_elements.size());
-                uint32_t rec_rule_id = generate_symbol_id(state, rule_name);
-                if (i > 0 || max_times < 0) {
-                    rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
-                }
-                rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
-                rec_rule.push_back({LLAMA_GRETYPE_END, 0});
-                add_rule(state, rec_rule_id, rec_rule);
-                last_rec_rule_id = rec_rule_id;
-            }
-            if (n_opt > 0) {
-                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
-            }
-        };
-
-        while (*pos) {
-            if (*pos == '"') { // literal string
-                pos++;
-                last_sym_start = out_elements.size();
-                while (*pos != '"') {
-                    if (!*pos) {
-                        throw std::runtime_error("unexpected end of input");
-                    }
-                    auto char_pair = parse_char(pos);
-                         pos       = char_pair.second;
-                    out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '[') { // char range(s)
-                pos++;
-                enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
-                if (*pos == '^') {
-                    pos++;
-                    start_type = LLAMA_GRETYPE_CHAR_NOT;
-                }
-                last_sym_start = out_elements.size();
-                while (*pos != ']') {
-                    if (!*pos) {
-                        throw std::runtime_error("unexpected end of input");
-                    }
-                    auto char_pair = parse_char(pos);
-                         pos       = char_pair.second;
-                    enum llama_gretype type = last_sym_start < out_elements.size()
-                        ? LLAMA_GRETYPE_CHAR_ALT
-                        : start_type;
-
-                    out_elements.push_back({type, char_pair.first});
-                    if (pos[0] == '-' && pos[1] != ']') {
-                        if (!pos[1]) {
-                            throw std::runtime_error("unexpected end of input");
-                        }
-                        auto endchar_pair = parse_char(pos + 1);
-                             pos          = endchar_pair.second;
-                        out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
-                    }
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (is_word_char(*pos)) { // rule reference
-                const char * name_end    = parse_name(pos);
-                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
-                pos = parse_space(name_end, is_nested);
-                last_sym_start = out_elements.size();
-                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
-            } else if (*pos == '(') { // grouping
-                // parse nested alternates into synthesized rule
-                pos = parse_space(pos + 1, true);
-                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
-                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
-                last_sym_start = out_elements.size();
-                // output reference to synthesized rule
-                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
-                if (*pos != ')') {
-                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '.') { // any char
-                last_sym_start = out_elements.size();
-                out_elements.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '*') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(0, -1);
-            } else if (*pos == '+') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(1, -1);
-            } else if (*pos == '?') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(0, 1);
-            } else if (*pos == '{') {
-                pos = parse_space(pos + 1, is_nested);
-
-                if (!is_digit_char(*pos)) {
-                    throw std::runtime_error(std::string("expecting an int at ") + pos);
-                }
-                const char * int_end = parse_int(pos);
-                int min_times = std::stoul(std::string(pos, int_end - pos));
-                pos = parse_space(int_end, is_nested);
-
-                int max_times = -1;
-
-                if (*pos == '}') {
-                    max_times = min_times;
-                    pos = parse_space(pos + 1, is_nested);
-                } else if (*pos == ',') {
-                    pos = parse_space(pos + 1, is_nested);
-
-                    if (is_digit_char(*pos)) {
-                        const char * int_end = parse_int(pos);
-                        max_times = std::stoul(std::string(pos, int_end - pos));
-                        pos = parse_space(int_end, is_nested);
-                    }
-
-                    if (*pos != '}') {
-                        throw std::runtime_error(std::string("expecting '}' at ") + pos);
-                    }
-                    pos = parse_space(pos + 1, is_nested);
-                } else {
-                    throw std::runtime_error(std::string("expecting ',' at ") + pos);
-                }
-                handle_repetitions(min_times, max_times);
-            } else {
-                break;
-            }
-        }
-        return pos;
-    }
-
-    const char * parse_alternates(
-            parse_state       & state,
-            const char        * src,
-            const std::string & rule_name,
-            uint32_t            rule_id,
-            bool                is_nested) {
-        std::vector<llama_grammar_element> rule;
-        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
-        while (*pos == '|') {
-            rule.push_back({LLAMA_GRETYPE_ALT, 0});
-            pos = parse_space(pos + 1, true);
-            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
-        }
-        rule.push_back({LLAMA_GRETYPE_END, 0});
-        add_rule(state, rule_id, rule);
-        return pos;
-    }
-
-    static const char * parse_rule(parse_state & state, const char * src) {
-        const char * name_end = parse_name(src);
-        const char * pos      = parse_space(name_end, false);
-        size_t       name_len = name_end - src;
-        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
-        const std::string name(src, name_len);
-
-        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
-            throw std::runtime_error(std::string("expecting ::= at ") + pos);
-        }
-        pos = parse_space(pos + 3, true);
-
-        pos = parse_alternates(state, pos, name, rule_id, false);
-
-        if (*pos == '\r') {
-            pos += pos[1] == '\n' ? 2 : 1;
-        } else if (*pos == '\n') {
-            pos++;
-        } else if (*pos) {
-            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
-        }
-        return parse_space(pos, true);
-    }
-
-    parse_state parse(const char * src) {
-        try {
-            parse_state state;
-            const char * pos = parse_space(src, true);
-            while (*pos) {
-                pos = parse_rule(state, pos);
-            }
-            // Validate the state to ensure that all rules are defined
-            for (const auto & rule : state.rules) {
-                if (rule.empty()) {
-                    throw std::runtime_error("Undefined rule");
-                }
-                for (const auto & elem : rule) {
-                    if (elem.type == LLAMA_GRETYPE_RULE_REF) {
-                        // Ensure that the rule at that location exists
-                        if (elem.value >= state.rules.size() || state.rules[elem.value].empty()) {
-                            // Get the name of the rule that is missing
-                            for (const auto & kv : state.symbol_ids) {
-                                if (kv.second == elem.value) {
-                                    throw std::runtime_error("Undefined rule identifier '" + kv.first + "'");
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-            return state;
-        } catch (const std::exception & err) {
-            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
-            return parse_state();
-        }
-    }
-
-    static void print_grammar_char(FILE * file, uint32_t c) {
-        if (0x20 <= c && c <= 0x7f) {
-            fprintf(file, "%c", static_cast<char>(c));
-        } else {
-            // cop out of encoding UTF-8
-            fprintf(file, "<U+%04X>", c);
-        }
-    }
-
-    static bool is_char_element(llama_grammar_element elem) {
-        switch (elem.type) {
-            case LLAMA_GRETYPE_CHAR:           return true;
-            case LLAMA_GRETYPE_CHAR_NOT:       return true;
-            case LLAMA_GRETYPE_CHAR_ALT:       return true;
-            case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
-            case LLAMA_GRETYPE_CHAR_ANY:       return true;
-            default:                           return false;
-        }
-    }
-
-    static void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
-        for (auto elem : rule) {
-            switch (elem.type) {
-                case LLAMA_GRETYPE_END:            fprintf(file, "END");            break;
-                case LLAMA_GRETYPE_ALT:            fprintf(file, "ALT");            break;
-                case LLAMA_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
-                case LLAMA_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
-                case LLAMA_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
-                case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
-                case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
-                case LLAMA_GRETYPE_CHAR_ANY:       fprintf(file, "CHAR_ANY");       break;
-            }
-            switch (elem.type) {
-                case LLAMA_GRETYPE_END:
-                case LLAMA_GRETYPE_ALT:
-                case LLAMA_GRETYPE_RULE_REF:
-                    fprintf(file, "(%u) ", elem.value);
-                    break;
-                case LLAMA_GRETYPE_CHAR:
-                case LLAMA_GRETYPE_CHAR_NOT:
-                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
-                case LLAMA_GRETYPE_CHAR_ALT:
-                case LLAMA_GRETYPE_CHAR_ANY:
-                    fprintf(file, "(\"");
-                    print_grammar_char(file, elem.value);
-                    fprintf(file, "\") ");
-                    break;
-            }
-        }
-        fprintf(file, "\n");
-    }
-
-    static void print_rule(
-            FILE     * file,
-            uint32_t   rule_id,
-            const std::vector<llama_grammar_element> & rule,
-            const std::map<uint32_t, std::string>    & symbol_id_names) {
-        if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) {
-            throw std::runtime_error(
-                "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id));
-        }
-        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
-        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
-            llama_grammar_element elem = rule[i];
-            switch (elem.type) {
-                case LLAMA_GRETYPE_END:
-                    throw std::runtime_error(
-                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
-                        std::to_string(i));
-                case LLAMA_GRETYPE_ALT:
-                    fprintf(file, "| ");
-                    break;
-                case LLAMA_GRETYPE_RULE_REF:
-                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
-                    break;
-                case LLAMA_GRETYPE_CHAR:
-                    fprintf(file, "[");
-                    print_grammar_char(file, elem.value);
-                    break;
-                case LLAMA_GRETYPE_CHAR_NOT:
-                    fprintf(file, "[^");
-                    print_grammar_char(file, elem.value);
-                    break;
-                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
-                    if (i == 0 || !is_char_element(rule[i - 1])) {
-                        throw std::runtime_error(
-                            "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
-                            std::to_string(rule_id) + "," + std::to_string(i));
-                    }
-                    fprintf(file, "-");
-                    print_grammar_char(file, elem.value);
-                    break;
-                case LLAMA_GRETYPE_CHAR_ALT:
-                    if (i == 0 || !is_char_element(rule[i - 1])) {
-                        throw std::runtime_error(
-                            "LLAMA_GRETYPE_CHAR_ALT without preceding char: " +
-                            std::to_string(rule_id) + "," + std::to_string(i));
-                    }
-                    print_grammar_char(file, elem.value);
-                    break;
-                case LLAMA_GRETYPE_CHAR_ANY:
-                    fprintf(file, ".");
-                    break;
-            }
-            if (is_char_element(elem)) {
-                switch (rule[i + 1].type) {
-                    case LLAMA_GRETYPE_CHAR_ALT:
-                    case LLAMA_GRETYPE_CHAR_RNG_UPPER:
-                    case LLAMA_GRETYPE_CHAR_ANY:
-                        break;
-                    default:
-                        fprintf(file, "] ");
-                }
-            }
-        }
-        fprintf(file, "\n");
-    }
-
-    void print_grammar(FILE * file, const parse_state & state) {
-        try {
-            std::map<uint32_t, std::string> symbol_id_names;
-            for (const auto & kv : state.symbol_ids) {
-                symbol_id_names[kv.second] = kv.first;
-            }
-            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
-                // fprintf(file, "%zu: ", i);
-                // print_rule_binary(file, state.rules[i]);
-                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
-                // fprintf(file, "\n");
-            }
-        } catch (const std::exception & err) {
-            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
-        }
-    }
-
-    std::vector<const llama_grammar_element *> parse_state::c_rules() {
-        std::vector<const llama_grammar_element *> ret;
-        ret.reserve(rules.size());
-        for (const auto & rule : rules) {
-            ret.push_back(rule.data());
-        }
-        return ret;
-    }
-}

common/grammar-parser.h

@@ -1,29 +0,0 @@
-// Implements a parser for an extended Backus-Naur form (BNF), producing the
-// binary context-free grammar format specified by llama.h. Supports character
-// ranges, grouping, and repetition operators. As an example, a grammar for
-// arithmetic might look like:
-//
-// root  ::= expr
-// expr  ::= term ([-+*/] term)*
-// term  ::= num | "(" space expr ")" space
-// num   ::= [0-9]+ space
-// space ::= [ \t\n]*
-
-#pragma once
-#include "llama.h"
-#include <vector>
-#include <map>
-#include <cstdint>
-#include <string>
-
-namespace grammar_parser {
-    struct parse_state {
-        std::map<std::string, uint32_t>                 symbol_ids;
-        std::vector<std::vector<llama_grammar_element>> rules;
-
-        std::vector<const llama_grammar_element *> c_rules();
-    };
-
-    parse_state parse(const char * src);
-    void print_grammar(FILE * file, const parse_state & state);
-}

common/sampling.cpp

@@ -1,460 +1,450 @@
-#define LLAMA_API_INTERNAL
 #include "sampling.h"
-#include <random>
 
-struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
-    struct llama_sampling_context * result = new llama_sampling_context();
+#include "common.h"
 
-    result->params  = params;
-    result->grammar = nullptr;
+#include <cmath>
+#include <unordered_map>
 
-    // if there is a grammar, parse it
-    if (!params.grammar.empty()) {
-        result->parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+// the ring buffer works similarly to std::deque, but with a fixed capacity
+// TODO: deduplicate with llama-impl.h
+template<typename T>
+struct ring_buffer {
+    ring_buffer(size_t cap) : capacity(cap), data(cap) {}
 
-        // will be empty (default) if there are parse errors
-        if (result->parsed_grammar.rules.empty()) {
-            fprintf(stderr, "%s: failed to parse grammar\n", __func__);
-            delete result;
-            return nullptr;
+    T & front() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[first];
+    }
+
+    const T & front() const {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[first];
+    }
+
+    T & back() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[pos];
+    }
+
+    const T & back() const {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[pos];
+    }
+
+    void push_back(const T & value) {
+        if (sz == capacity) {
+            // advance the start when buffer is full
+            first = (first + 1) % capacity;
+        } else {
+            sz++;
+        }
+        data[pos] = value;
+        pos = (pos + 1) % capacity;
+    }
+
+    T pop_front() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        T value = data[first];
+        first = (first + 1) % capacity;
+        sz--;
+        return value;
+    }
+
+    const T & rat(size_t i) const {
+        if (i >= sz) {
+            throw std::runtime_error("ring buffer: index out of bounds");
+        }
+        return data[(first + sz - i - 1) % capacity];
+    }
+
+    std::vector<T> to_vector() const {
+        std::vector<T> result;
+        result.reserve(sz);
+        for (size_t i = 0; i < sz; i++) {
+            result.push_back(data[(first + i) % capacity]);
+        }
+        return result;
+    }
+
+    void clear() {
+        // here only reset the status of the buffer
+        sz = 0;
+        first = 0;
+        pos = 0;
+    }
+
+    bool empty() const {
+        return sz == 0;
+    }
+
+    size_t size() const {
+        return sz;
+    }
+
+    size_t capacity = 0;
+    size_t sz = 0;
+    size_t first = 0;
+    size_t pos = 0;
+    std::vector<T> data;
+};
+
+struct gpt_sampler {
+    gpt_sampler_params params;
+
+    struct llama_sampler * grmr;
+    struct llama_sampler * chain;
+
+    ring_buffer<llama_token> prev;
+
+    std::vector<llama_token_data> cur;
+
+    llama_token_data_array cur_p;
+
+    void set_logits(struct llama_context * ctx, int idx) {
+        const auto * logits = llama_get_logits_ith(ctx, idx);
+
+        const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
+        cur.resize(n_vocab);
+
+        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+            cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
         }
 
-        // Ensure that there is a "root" node.
-        if (result->parsed_grammar.symbol_ids.find("root") == result->parsed_grammar.symbol_ids.end()) {
-            fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
-            delete result;
-            return nullptr;
-        }
-
-        std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
-
-        struct llama_grammar * grammar = llama_grammar_init(
-                grammar_rules.data(),
-                grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
-        if (grammar == nullptr) {
-            throw std::runtime_error("Failed to initialize llama_grammar");
-        }
-        result->grammar = grammar;
+        cur_p = { cur.data(), cur.size(), -1, false };
     }
+};
 
-    result->prev.resize(params.n_prev);
-
-    result->n_valid = 0;
-
-    llama_sampling_set_rng_seed(result, params.seed);
-
-    return result;
-}
-
-void llama_sampling_free(struct llama_sampling_context * ctx) {
-    if (ctx->grammar != NULL) {
-        llama_grammar_free(ctx->grammar);
-    }
-
-    delete ctx;
-}
-
-void llama_sampling_reset(llama_sampling_context * ctx) {
-    if (ctx->grammar != NULL) {
-        llama_grammar_free(ctx->grammar);
-        ctx->grammar = NULL;
-    }
-
-    if (!ctx->parsed_grammar.rules.empty()) {
-        std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
-
-        struct llama_grammar * grammar = llama_grammar_init(
-                grammar_rules.data(),
-                grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
-        if (grammar == nullptr) {
-            throw std::runtime_error("Failed to initialize llama_grammar");
-        }
-        ctx->grammar = grammar;
-    }
-
-    std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
-    ctx->cur.clear();
-    ctx->n_valid = 0;
-}
-
-void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
-    if (seed == LLAMA_DEFAULT_SEED) {
-        seed = std::random_device{}();
-    }
-    ctx->rng.seed(seed);
-}
-
-void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
-    if (dst->grammar) {
-        llama_grammar_free(dst->grammar);
-        dst->grammar = nullptr;
-    }
-
-    if (src->grammar) {
-        dst->grammar = llama_grammar_copy(src->grammar);
-    }
-
-    dst->prev = src->prev;
-}
-
-llama_token llama_sampling_last(llama_sampling_context * ctx) {
-    return ctx->prev.back();
-}
-
-std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n) {
-    const int size = ctx_sampling->prev.size();
-
-    n = std::min(n, size);
-
-    std::string result;
-
-    for (int i = size - n; i < size; i++) {
-        result += llama_token_to_piece(ctx_main, ctx_sampling->prev[i]);
-    }
-
-    return result;
-}
-
-std::string llama_sampling_print(const llama_sampling_params & params) {
+std::string gpt_sampler_params::print() const {
     char result[1024];
 
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
             "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
-            params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present,
-            params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp,
-            params.mirostat, params.mirostat_eta, params.mirostat_tau);
+            penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
+            top_k, tfs_z, top_p, min_p, typ_p, temp,
+            mirostat, mirostat_eta, mirostat_tau);
 
     return std::string(result);
 }
 
-std::string llama_sampling_order_print(const llama_sampling_params & params) {
-    std::string result = "CFG -> Penalties ";
-    if (params.mirostat == 0) {
-        for (auto sampler_type : params.samplers_sequence) {
-            const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
-            if (!sampler_type_name.empty()) {
-                result += "-> " + sampler_type_name + " ";
+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 = params.no_perf;
+
+    auto * result = new gpt_sampler {
+        /* .params = */ params,
+        /* .grmr   = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
+        /* .chain  = */ llama_sampler_chain_init(lparams),
+        /* .prev   = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
+        /* .cur    = */ {},
+        /* .cur_p  = */ {},
+    };
+
+    llama_sampler_chain_add(result->chain,
+            llama_sampler_init_logit_bias(
+                llama_n_vocab(model),
+                params.logit_bias.size(),
+                params.logit_bias.data()));
+
+    llama_sampler_chain_add(result->chain,
+            llama_sampler_init_penalties(
+                llama_n_vocab  (model),
+                llama_token_eos(model),
+                llama_token_nl (model),
+                params.penalty_last_n,
+                params.penalty_repeat,
+                params.penalty_freq,
+                params.penalty_present,
+                params.penalize_nl,
+                params.ignore_eos));
+
+    if (params.temp > 0.0f) {
+        if (params.mirostat == 0) {
+            for (const auto & cnstr : params.samplers) {
+                switch (cnstr) {
+                    case GPT_SAMPLER_TYPE_TOP_K:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                        break;
+                    case GPT_SAMPLER_TYPE_TOP_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                        break;
+                    case GPT_SAMPLER_TYPE_MIN_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                        break;
+                    case GPT_SAMPLER_TYPE_TFS_Z:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
+                        break;
+                    case GPT_SAMPLER_TYPE_TYPICAL_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                        break;
+                    case GPT_SAMPLER_TYPE_TEMPERATURE:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                        break;
+                    default:
+                        GGML_ASSERT(false && "unknown sampler type");
+                }
             }
+            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
+            llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
+        } else if (params.mirostat == 1) {
+            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+        } else if (params.mirostat == 2) {
+            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
+        } else {
+            GGML_ASSERT(false && "unknown mirostat version");
         }
     } else {
-        result += "-> mirostat ";
+        llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
+        llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
     }
 
     return result;
 }
 
-std::string llama_sampling_type_to_str(llama_sampler_type sampler_type) {
-    switch (sampler_type) {
-        case llama_sampler_type::TOP_K:       return "top_k";
-        case llama_sampler_type::TFS_Z:       return "tfs_z";
-        case llama_sampler_type::TYPICAL_P:   return "typical_p";
-        case llama_sampler_type::TOP_P:       return "top_p";
-        case llama_sampler_type::MIN_P:       return "min_p";
-        case llama_sampler_type::TEMPERATURE: return "temperature";
+void gpt_sampler_free(struct gpt_sampler * gsmpl) {
+    if (gsmpl) {
+        llama_sampler_free(gsmpl->grmr);
+
+        llama_sampler_free(gsmpl->chain);
+
+        delete gsmpl;
+    }
+}
+
+void gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar) {
+    if (accept_grammar) {
+        llama_sampler_accept(gsmpl->grmr, token);
+    }
+
+    llama_sampler_accept(gsmpl->chain, token);
+
+    gsmpl->prev.push_back(token);
+}
+
+void gpt_sampler_reset(struct gpt_sampler * gsmpl) {
+    llama_sampler_reset(gsmpl->grmr);
+
+    llama_sampler_reset(gsmpl->chain);
+}
+
+struct gpt_sampler * gpt_sampler_clone(gpt_sampler * gsmpl) {
+    return new gpt_sampler {
+        /* .params = */ gsmpl->params,
+        /* .grmr   = */ llama_sampler_clone(gsmpl->grmr),
+        /* .chain  = */ llama_sampler_clone(gsmpl->chain),
+        /* .prev   = */ gsmpl->prev,
+        /* .cur    = */ gsmpl->cur,
+        /* .cur_p  = */ gsmpl->cur_p,
+    };
+}
+
+void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl) {
+    // TODO: measure grammar performance
+
+    if (gsmpl) {
+        llama_perf_sampler_print(gsmpl->chain);
+    }
+    if (ctx) {
+        llama_perf_context_print(ctx);
+    }
+}
+
+llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+    gsmpl->set_logits(ctx, idx);
+
+    auto & grmr  = gsmpl->grmr;
+    auto & chain = gsmpl->chain;
+    auto & cur_p = gsmpl->cur_p; // initialized by set_logits
+
+    if (grammar_first) {
+        llama_sampler_apply(grmr, &cur_p);
+    }
+
+    llama_sampler_apply(chain, &cur_p);
+
+    GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
+
+    const llama_token id = cur_p.data[cur_p.selected].id;
+
+    if (grammar_first) {
+        return id;
+    }
+
+    // check if it the sampled token fits the grammar
+    {
+        llama_token_data       single_token_data       = { id, 1.0f, 0.0f };
+        llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
+
+        llama_sampler_apply(grmr, &single_token_data_array);
+
+        const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
+        if (is_valid) {
+            return id;
+        }
+    }
+
+    // resampling:
+    // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
+    gsmpl->set_logits(ctx, idx);
+
+    llama_sampler_apply(grmr,  &cur_p);
+    llama_sampler_apply(chain, &cur_p);
+
+    GGML_ASSERT(cur_p.selected != -1 && "no selected token during re-sampling - check your sampling configuration");
+
+    return cur_p.data[cur_p.selected].id;
+}
+
+uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl) {
+    return llama_sampler_get_seed(gsmpl->chain);
+}
+
+// helpers
+
+llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl) {
+    return &gsmpl->cur_p;
+}
+
+llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl) {
+    return gsmpl->prev.rat(0);
+}
+
+std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) {
+    std::string result = "\tlogits ";
+
+    for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
+        const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
+        result += std::string("-> ") + llama_sampler_name(smpl) + " ";
+    }
+
+    return result;
+}
+
+std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx_main, int n) {
+    n = std::min(n, (int) gsmpl->prev.size());
+
+    if (n <= 0) {
+        return "";
+    }
+
+    std::string result;
+    result.reserve(8*n); // 8 is the average length of a token [citation needed], TODO: compute this from the vocab
+
+    for (int i = n - 1; i >= 0; i--) {
+        const llama_token id = gsmpl->prev.rat(i);
+
+        GGML_ASSERT(id != LLAMA_TOKEN_NULL && "null token in the sampling history - should not happen");
+
+        result += llama_token_to_piece(ctx_main, id);
+    }
+
+    return result;
+}
+
+char gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr) {
+    switch (cnstr) {
+        case GPT_SAMPLER_TYPE_TOP_K:       return 'k';
+        case GPT_SAMPLER_TYPE_TFS_Z:       return 'f';
+        case GPT_SAMPLER_TYPE_TYPICAL_P:   return 'y';
+        case GPT_SAMPLER_TYPE_TOP_P:       return 'p';
+        case GPT_SAMPLER_TYPE_MIN_P:       return 'm';
+        case GPT_SAMPLER_TYPE_TEMPERATURE: return 't';
+        default : return '?';
+    }
+}
+
+std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr) {
+    switch (cnstr) {
+        case GPT_SAMPLER_TYPE_TOP_K:       return "top_k";
+        case GPT_SAMPLER_TYPE_TFS_Z:       return "tfs_z";
+        case GPT_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
+        case GPT_SAMPLER_TYPE_TOP_P:       return "top_p";
+        case GPT_SAMPLER_TYPE_MIN_P:       return "min_p";
+        case GPT_SAMPLER_TYPE_TEMPERATURE: return "temperature";
         default : return "";
     }
 }
 
-std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
-    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
-        {"top_k",       llama_sampler_type::TOP_K},
-        {"top_p",       llama_sampler_type::TOP_P},
-        {"typical_p",   llama_sampler_type::TYPICAL_P},
-        {"min_p",       llama_sampler_type::MIN_P},
-        {"tfs_z",       llama_sampler_type::TFS_Z},
-        {"temperature", llama_sampler_type::TEMPERATURE}
+std::vector<gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, gpt_sampler_type> sampler_canonical_name_map {
+        { "top_k",       GPT_SAMPLER_TYPE_TOP_K },
+        { "top_p",       GPT_SAMPLER_TYPE_TOP_P },
+        { "typ_p",       GPT_SAMPLER_TYPE_TYPICAL_P },
+        { "min_p",       GPT_SAMPLER_TYPE_MIN_P },
+        { "tfs_z",       GPT_SAMPLER_TYPE_TFS_Z },
+        { "temperature", GPT_SAMPLER_TYPE_TEMPERATURE },
     };
 
     // since samplers names are written multiple ways
     // make it ready for both system names and input names
-    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
-        {"top-k",       llama_sampler_type::TOP_K},
-        {"top-p",       llama_sampler_type::TOP_P},
-        {"nucleus",     llama_sampler_type::TOP_P},
-        {"typical-p",   llama_sampler_type::TYPICAL_P},
-        {"typical",     llama_sampler_type::TYPICAL_P},
-        {"min-p",       llama_sampler_type::MIN_P},
-        {"tfs-z",       llama_sampler_type::TFS_Z},
-        {"tfs",         llama_sampler_type::TFS_Z},
-        {"temp",        llama_sampler_type::TEMPERATURE}
+    std::unordered_map<std::string, gpt_sampler_type> sampler_alt_name_map {
+        { "top-k",       GPT_SAMPLER_TYPE_TOP_K },
+        { "top-p",       GPT_SAMPLER_TYPE_TOP_P },
+        { "nucleus",     GPT_SAMPLER_TYPE_TOP_P },
+        { "typical-p",   GPT_SAMPLER_TYPE_TYPICAL_P },
+        { "typical",     GPT_SAMPLER_TYPE_TYPICAL_P },
+        { "typ-p",       GPT_SAMPLER_TYPE_TYPICAL_P },
+        { "typ",         GPT_SAMPLER_TYPE_TYPICAL_P },
+        { "min-p",       GPT_SAMPLER_TYPE_MIN_P },
+        { "tfs-z",       GPT_SAMPLER_TYPE_TFS_Z },
+        { "tfs",         GPT_SAMPLER_TYPE_TFS_Z },
+        { "temp",        GPT_SAMPLER_TYPE_TEMPERATURE },
     };
 
-    std::vector<llama_sampler_type> sampler_types;
-    sampler_types.reserve(names.size());
-    for (const auto & name : names)
-    {
-        auto sampler_item = sampler_canonical_name_map.find(name);
-        if (sampler_item != sampler_canonical_name_map.end())
-        {
-            sampler_types.push_back(sampler_item->second);
-        }
-        else
-        {
-            if (allow_alt_names)
-            {
-                sampler_item = sampler_alt_name_map.find(name);
-                if (sampler_item != sampler_alt_name_map.end())
-                {
-                    sampler_types.push_back(sampler_item->second);
-                }
-            }
-        }
-    }
-    return sampler_types;
-}
+    std::vector<gpt_sampler_type> samplers;
+    samplers.reserve(names.size());
 
-std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string) {
-    std::unordered_map<char, llama_sampler_type> sampler_name_map {
-        {'k', llama_sampler_type::TOP_K},
-        {'p', llama_sampler_type::TOP_P},
-        {'y', llama_sampler_type::TYPICAL_P},
-        {'m', llama_sampler_type::MIN_P},
-        {'f', llama_sampler_type::TFS_Z},
-        {'t', llama_sampler_type::TEMPERATURE}
-    };
-
-    std::vector<llama_sampler_type> sampler_types;
-    sampler_types.reserve(names_string.size());
-    for (const auto & c : names_string) {
-        const auto sampler_item = sampler_name_map.find(c);
-        if (sampler_item != sampler_name_map.end()) {
-            sampler_types.push_back(sampler_item->second);
-        }
-    }
-    return sampler_types;
-}
-
-// no reasons to expose this function in header
-static void sampler_queue(
-                   struct llama_context * ctx_main,
-            const llama_sampling_params & params,
-                 llama_token_data_array & cur_p,
-                                 size_t   min_keep) {
-    const float         temp              = params.temp;
-    const float         dynatemp_range    = params.dynatemp_range;
-    const float         dynatemp_exponent = params.dynatemp_exponent;
-    const int32_t       top_k             = params.top_k;
-    const float         top_p             = params.top_p;
-    const float         min_p             = params.min_p;
-    const float         tfs_z             = params.tfs_z;
-    const float         typical_p         = params.typical_p;
-    const std::vector<llama_sampler_type> & samplers_sequence = params.samplers_sequence;
-
-    for (auto sampler_type : samplers_sequence) {
-        switch (sampler_type) {
-            case llama_sampler_type::TOP_K    : llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
-            case llama_sampler_type::TFS_Z    : llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
-            case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
-            case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
-            case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case llama_sampler_type::TEMPERATURE:
-                if (dynatemp_range > 0) {
-                    float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
-                    float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
-                    llama_sample_entropy(ctx_main, &cur_p, dynatemp_min, dynatemp_max, dynatemp_exponent);
-                } else {
-                    llama_sample_temp(ctx_main, &cur_p, temp);
-                }
-                break;
-            default : break;
-        }
-    }
-}
-
-static llama_token llama_sampling_sample_impl(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx,
-                  bool is_resampling) {
-    const llama_sampling_params & params = ctx_sampling->params;
-
-    const float   temp            = params.temp;
-    const int     mirostat        = params.mirostat;
-    const float   mirostat_tau    = params.mirostat_tau;
-    const float   mirostat_eta    = params.mirostat_eta;
-
-    std::vector<float> original_logits;
-    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
-    if (ctx_sampling->grammar != NULL && !is_resampling) {
-        GGML_ASSERT(!original_logits.empty());
-    }
-    llama_token id = 0;
-
-    if (temp < 0.0) {
-        // greedy sampling, with probs
-        llama_sample_softmax(ctx_main, &cur_p);
-        id = cur_p.data[0].id;
-    } else if (temp == 0.0) {
-        // greedy sampling, no probs
-        id = llama_sample_token_greedy(ctx_main, &cur_p);
-    } else {
-        if (mirostat == 1) {
-            const int mirostat_m = 100;
-            llama_sample_temp(ctx_main, &cur_p, temp);
-            id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
-        } else if (mirostat == 2) {
-            llama_sample_temp(ctx_main, &cur_p, temp);
-            id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
+    for (const auto & name : names) {
+        auto sampler = sampler_canonical_name_map.find(name);
+        if (sampler != sampler_canonical_name_map.end()) {
+            samplers.push_back(sampler->second);
         } else {
-            // temperature sampling
-            size_t min_keep = std::max(1, params.min_keep);
-
-            sampler_queue(ctx_main, params, cur_p, min_keep);
-
-            id = llama_sample_token_with_rng(ctx_main, &cur_p, ctx_sampling->rng);
-
-            //{
-            //    const int n_top = 10;
-            //    LOG("top %d candidates:\n", n_top);
-
-            //    for (int i = 0; i < n_top; i++) {
-            //        const llama_token id = cur_p.data[i].id;
-            //        (void)id; // To avoid a warning that id is unused when logging is disabled.
-            //        LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
-            //    }
-            //}
-
-            //LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
-        }
-    }
-
-    if (ctx_sampling->grammar != NULL && !is_resampling) {
-        // Get a pointer to the logits
-        float * logits = llama_get_logits_ith(ctx_main, idx);
-
-        // Create an array with a single token data element for the sampled id
-        llama_token_data single_token_data = {id, logits[id], 0.0f};
-        llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
-
-        // Apply grammar constraints to the single token
-        llama_grammar_sample(ctx_sampling->grammar, ctx_main, &single_token_data_array);
-
-        // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
-        bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
-
-        // If the token is not valid according to the grammar, perform resampling
-        if (!is_valid) {
-            LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str());
-
-            // Restore logits from the copy
-            std::copy(original_logits.begin(), original_logits.end(), logits);
-
-            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
-        }
-    }
-
-    ctx_sampling->n_valid = temp == 0.0f ? 0 : cur_p.size;
-
-    return id;
-}
-
-static llama_token_data_array llama_sampling_prepare_impl(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx,
-                  bool apply_grammar,
-                  std::vector<float> * original_logits) {
-    const llama_sampling_params & params = ctx_sampling->params;
-
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
-    const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
-    const float   penalty_repeat  = params.penalty_repeat;
-    const float   penalty_freq    = params.penalty_freq;
-    const float   penalty_present = params.penalty_present;
-
-    const bool    penalize_nl     = params.penalize_nl;
-
-    auto & prev = ctx_sampling->prev;
-    auto & cur  = ctx_sampling->cur;
-
-    // Get a pointer to the logits
-    float * logits = llama_get_logits_ith(ctx_main, idx);
-
-    if (ctx_sampling->grammar != NULL && !apply_grammar) {
-        GGML_ASSERT(original_logits != NULL);
-        // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
-        *original_logits = {logits, logits + n_vocab};
-    }
-
-    // apply params.logit_bias map
-    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-        logits[it->first] += it->second;
-    }
-
-    if (ctx_cfg) {
-        float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
-        llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
-    }
-
-    cur.resize(n_vocab);
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
-    }
-
-    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
-
-    // apply penalties
-    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
-    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
-    if (penalty_tokens_used_size) {
-        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
-
-        llama_sample_repetition_penalties(ctx_main, &cur_p,
-                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
-                penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
-
-        if (!penalize_nl) {
-            for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
-                    cur_p.data[idx].logit = nl_logit;
-                    break;
+            if (allow_alt_names) {
+                sampler = sampler_alt_name_map.find(name);
+                if (sampler != sampler_alt_name_map.end()) {
+                    samplers.push_back(sampler->second);
                 }
             }
         }
     }
 
-    // apply grammar checks before sampling logic
-    if (apply_grammar && ctx_sampling->grammar != NULL) {
-        llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p);
+    return samplers;
+}
+
+std::vector<gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars) {
+    std::unordered_map<char, gpt_sampler_type> sampler_name_map = {
+        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_K),       GPT_SAMPLER_TYPE_TOP_K },
+        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TFS_Z),       GPT_SAMPLER_TYPE_TFS_Z },
+        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TYPICAL_P),   GPT_SAMPLER_TYPE_TYPICAL_P },
+        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_P),       GPT_SAMPLER_TYPE_TOP_P },
+        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_MIN_P),       GPT_SAMPLER_TYPE_MIN_P },
+        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TEMPERATURE), GPT_SAMPLER_TYPE_TEMPERATURE }
+    };
+
+    std::vector<gpt_sampler_type> samplers;
+    samplers.reserve(chars.size());
+
+    for (const auto & c : chars) {
+        const auto sampler = sampler_name_map.find(c);
+        if (sampler != sampler_name_map.end()) {
+            samplers.push_back(sampler->second);
+        }
     }
 
-    return cur_p;
-}
-
-llama_token llama_sampling_sample(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx) {
-    // Call the implementation function with is_resampling set to false by default
-    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
-}
-
-llama_token_data_array llama_sampling_prepare(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx,
-                  bool apply_grammar,
-                  std::vector<float> * original_logits) {
-    return llama_sampling_prepare_impl(ctx_sampling,ctx_main, ctx_cfg, idx, apply_grammar, original_logits);
-}
-
-void llama_sampling_accept(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        llama_token id,
-        bool apply_grammar) {
-    ctx_sampling->prev.erase(ctx_sampling->prev.begin());
-    ctx_sampling->prev.push_back(id);
-
-    if (ctx_sampling->grammar != NULL && apply_grammar) {
-        llama_grammar_accept_token(ctx_sampling->grammar, ctx_main, id);
-    }
+    return samplers;
 }

common/sampling.h

@@ -2,159 +2,82 @@
 
 #include "llama.h"
 
-#include "grammar-parser.h"
-
-#include <random>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-// sampler types
-enum class llama_sampler_type : char {
-    TOP_K       = 'k',
-    TOP_P       = 'p',
-    MIN_P       = 'm',
-    TFS_Z       = 'f',
-    TYPICAL_P   = 'y',
-    TEMPERATURE = 't'
-};
-
-// sampling parameters
-typedef struct llama_sampling_params {
-    int32_t     n_prev                = 64;                 // number of previous tokens to remember
-    int32_t     n_probs               = 0;                  // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t     min_keep              = 0;                  // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t     top_k                 = 40;                 // <= 0 to use vocab size
-    float       top_p                 = 0.95f;              // 1.0 = disabled
-    float       min_p                 = 0.05f;              // 0.0 = disabled
-    float       tfs_z                 = 1.00f;              // 1.0 = disabled
-    float       typical_p             = 1.00f;              // 1.0 = disabled
-    float       temp                  = 0.80f;              // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float       dynatemp_range        = 0.00f;              // 0.0 = disabled
-    float       dynatemp_exponent     = 1.00f;              // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t     penalty_last_n        = 64;                 // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float       penalty_repeat        = 1.00f;              // 1.0 = disabled
-    float       penalty_freq          = 0.00f;              // 0.0 = disabled
-    float       penalty_present       = 0.00f;              // 0.0 = disabled
-    int32_t     mirostat              = 0;                  // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float       mirostat_tau          = 5.00f;              // target entropy
-    float       mirostat_eta          = 0.10f;              // learning rate
-    bool        penalize_nl           = false;              // consider newlines as a repeatable token
-    uint32_t    seed                  = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampling_context
-
-    std::vector<llama_sampler_type> samplers_sequence = {
-        llama_sampler_type::TOP_K,
-        llama_sampler_type::TFS_Z,
-        llama_sampler_type::TYPICAL_P,
-        llama_sampler_type::TOP_P,
-        llama_sampler_type::MIN_P,
-        llama_sampler_type::TEMPERATURE
-    };
-
-    std::string grammar;  // optional BNF-like grammar to constrain sampling
-
-    // Classifier-Free Guidance
-    // https://arxiv.org/abs/2306.17806
-    std::string cfg_negative_prompt; // string to help guidance
-    float       cfg_scale     = 1.f; // how strong is guidance
-
-    std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
-
-    std::vector<llama_token> penalty_prompt_tokens;
-    bool                     use_penalty_prompt_tokens = false;
-} llama_sampling_params;
-
-// general sampler context
-// TODO: move to llama.h
-struct llama_sampling_context {
-    // parameters that will be used for sampling
-    llama_sampling_params params;
-
-    // mirostat sampler state
-    float mirostat_mu;
-
-    llama_grammar * grammar;
-
-    // internal
-    grammar_parser::parse_state parsed_grammar;
-
-    // TODO: replace with ring-buffer
-    std::vector<llama_token>      prev;
-    std::vector<llama_token_data> cur;
-    size_t n_valid; // Number of correct top tokens with correct probabilities.
-
-    std::mt19937 rng;
-};
-
 #include "common.h"
 
-// Create a new sampling context instance.
-struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params);
+#include <string>
+#include <vector>
 
-void llama_sampling_free(struct llama_sampling_context * ctx);
-
-// Reset the sampler context
-// - clear prev tokens
-// - reset grammar
-void llama_sampling_reset(llama_sampling_context * ctx);
-
-// Set the sampler seed
-void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed);
-
-// Copy the sampler context
-void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
-
-// Get the last sampled token
-llama_token llama_sampling_last(llama_sampling_context * ctx);
-
-// Get a string representation of the last sampled tokens
-std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n);
-
-// Print sampling parameters into a string
-std::string llama_sampling_print(const llama_sampling_params & params);
-
-// Print sampling order into a string
-std::string llama_sampling_order_print(const llama_sampling_params & params);
-
-std::string llama_sampling_type_to_str(llama_sampler_type sampler_type);
-
-std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
-std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string);
-
-// this is a common sampling function used across the examples for convenience
-// it can serve as a starting point for implementing your own sampling function
-// Note: When using multiple sequences, it is the caller's responsibility to call
-//       llama_sampling_reset when a sequence ends
+// gpt_sampler extends llama_sampler with additional functionality:
 //
-// required:
-//  - ctx_main:     context to use for sampling
-//  - ctx_sampling: sampling-specific context
+//  - grammar support
+//  - custom sampler logic based on the parameters
+//  - history of the last accepted tokens
+//  - performance metrics
 //
-// optional:
-//  - ctx_cfg:      context to use for classifier-free guidance
-//  - idx:          sample from llama_get_logits_ith(ctx, idx)
+// This goal is to have a common implementation of the sampling logic shared across the examples.
+// For example, depending on the temperature, the sampling chain can be very simple (greedy) or more
+// complex (top-k, top-p, etc).
 //
-// returns:
-//  - token:      sampled token
-//  - candidates: vector of candidate tokens
+// Another example is related to the grammar. In general, the grammar constraints applied on the full
+// vocabulary can be very taxing. To improve performance, the grammar can be applied only to the sampled
+// token in order to verify if it fits the grammar. And only if the token doesn't fit the grammar, the
+// grammar constraints are applied to the full vocabulary and the token is resampled.
+//
+// The gpt_sampler also maintains a container with the last accepted tokens. In the future, this can
+// be moved into the core llama library.
+//
+// For convenience, the gpt_sampler also maintains a container with the current candidate tokens.
+// This can be used to access the probabilities of the rest of the non-sampled tokens.
+//
+// TODO: measure grammar performance
 //
-llama_token llama_sampling_sample(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        struct llama_context * ctx_cfg,
-        int idx = -1);
 
-// Prepares and adjusts the set of token candidates for sampling based on penalties, biases, and sampling parameters.
-llama_token_data_array llama_sampling_prepare(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        struct llama_context * ctx_cfg,
-        int idx = 0,
-        bool apply_grammar = true,
-        std::vector<float> * original_logits = nullptr);
+struct gpt_sampler;
 
-void llama_sampling_accept(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        llama_token id,
-        bool apply_grammar);
+// llama_sampler API overloads
+
+struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params);
+
+void gpt_sampler_free(struct gpt_sampler * gsmpl);
+
+// if accept_grammar is true, the token is accepted both by the sampling chain and the grammar
+void                 gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar);
+void                 gpt_sampler_reset (struct gpt_sampler * gsmpl);
+struct gpt_sampler * gpt_sampler_clone (struct gpt_sampler * gsmpl);
+
+// arguments can be nullptr to skip printing
+void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl);
+
+// extended sampling implementation:
+//
+// - set logits
+// - apply the configured sampler chain
+// - check if the token fits the grammar (if any)
+// - if not: resample by first applying the grammar constraints and then sampling again (slower path)
+//
+// if grammar_first is true, the grammar is applied before the samplers (slower)
+// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
+//
+llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
+
+uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl);
+
+// helpers
+
+// access the internal list of current candidate tokens
+llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl);
+
+// get the last accepted token
+llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl);
+
+// print the sampler chain into a string
+std::string gpt_sampler_print(const struct gpt_sampler * gsmpl);
+
+// get a string representation of the last accepted tokens
+std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx, int n);
+
+char        gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr);
+std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr);
+
+std::vector<enum gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<enum gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars);

convert_hf_to_gguf.py

@@ -302,6 +302,8 @@ class Model:
                             gguf.MODEL_TENSOR.TIME_MIX_FIRST,
                             gguf.MODEL_TENSOR.TIME_MIX_W1,
                             gguf.MODEL_TENSOR.TIME_MIX_W2,
+                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
+                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
                         )
                     )
                     or not new_name.endswith(".weight")
@@ -624,6 +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")
@@ -2769,6 +2774,8 @@ class Rwkv6Model(Model):
         self.gguf_writer.add_tokenizer_model("rwkv")
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+        special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]

convert_hf_to_gguf_update.py

@@ -31,6 +31,7 @@ import re
 import requests
 import sys
 import json
+import shutil
 
 from hashlib import sha256
 from enum import IntEnum, auto
@@ -97,6 +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", },
 ]
 
 
@@ -125,12 +127,27 @@ def download_model(model):
     if tokt == TOKENIZER_TYPE.UGM:
         files.append("spiece.model")
 
-    for file in files:
-        save_path = f"models/tokenizers/{name}/{file}"
-        if os.path.isfile(save_path):
-            logger.info(f"{name}: File {save_path} already exists - skipping")
-            continue
-        download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
+    if os.path.isdir(repo):
+        # If repo is a path on the file system, copy the directory
+        for file in files:
+            src_path = os.path.join(repo, file)
+            dst_path = f"models/tokenizers/{name}/{file}"
+            if os.path.isfile(dst_path):
+                logger.info(f"{name}: File {dst_path} already exists - skipping")
+                continue
+            if os.path.isfile(src_path):
+                shutil.copy2(src_path, dst_path)
+                logger.info(f"{name}: Copied {src_path} to {dst_path}")
+            else:
+                logger.warning(f"{name}: Source file {src_path} does not exist")
+    else:
+        # If repo is a URL, download the files
+        for file in files:
+            save_path = f"models/tokenizers/{name}/{file}"
+            if os.path.isfile(save_path):
+                logger.info(f"{name}: File {save_path} already exists - skipping")
+                continue
+            download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
 
 
 for model in models:

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()

docs/build.md

@@ -380,3 +380,9 @@ For detailed info, such as model/device supports, CANN install, please refer to
 ### Android
 
 To read documentation for how to build on Android, [click here](./android.md)
+
+### Arm CPU optimized mulmat kernels
+
+Llama.cpp includes a set of optimized mulmat kernels for the Arm architecture, leveraging Arm® Neon™, int8mm and SVE instructions. These kernels are enabled at build time through the appropriate compiler cpu-type flags, such as `-DCMAKE_C_FLAGS=-march=armv8.2a+i8mm+sve`. Note that these optimized kernels require the model to be quantized into one of the formats: `Q4_0_4_4` (Arm Neon), `Q4_0_4_8` (int8mm) or `Q4_0_8_8` (SVE). The SVE mulmat kernel specifically requires a vector width of 256 bits. When running on devices with a different vector width, it is recommended to use the `Q4_0_4_8` (int8mm) or `Q4_0_4_4` (Arm Neon) formats for better performance. Refer to [examples/quantize/README.md](../examples/quantize/README.md) for more information on the quantization formats.
+
+To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_LLAMAFILE=OFF` (`cmake`).

examples/batched-bench/batched-bench.cpp

@@ -1,36 +1,13 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-// mutates the input string
-static std::vector<int> parse_list(char * p) {
-    std::vector<int> ret;
-
-    char * q = p;
-
-    while (*p) {
-        if (*p == ',') {
-            *p = '\0';
-            ret.push_back(std::atoi(q));
-            q = p + 1;
-        }
-
-        ++p;
-    }
-
-    ret.push_back(std::atoi(q));
-
-    return ret;
-}
-
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
     LOG_TEE("\n");
@@ -39,8 +16,7 @@ static void print_usage(int argc, char ** argv, const gpt_params & params) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
         return 1;
     }
 
@@ -210,7 +186,8 @@ int main(int argc, char ** argv) {
         }
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 

examples/batched.swift/Sources/main.swift

@@ -27,7 +27,6 @@ guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), mo
     print("Failed to load model")
     exit(1)
 }
-
 defer {
     llama_free_model(model)
 }
@@ -37,7 +36,6 @@ var tokens = tokenize(text: prompt, add_bos: true)
 let n_kv_req = UInt32(tokens.count) + UInt32((n_len - Int(tokens.count)) * n_parallel)
 
 var context_params = llama_context_default_params()
-context_params.seed = 1234
 context_params.n_ctx = n_kv_req
 context_params.n_batch = UInt32(max(n_len, n_parallel))
 context_params.n_threads = 8
@@ -48,11 +46,26 @@ guard context != nil else {
     print("Failed to initialize context")
     exit(1)
 }
-
 defer {
     llama_free(context)
 }
 
+var sparams = llama_sampler_chain_default_params()
+
+let smpl = llama_sampler_chain_init(sparams)
+guard smpl != nil else {
+    print("Failed to initialize sampling")
+    exit(1)
+}
+defer {
+    llama_sampler_free(smpl)
+}
+
+llama_sampler_chain_add(smpl, llama_sampler_init_top_k(40));
+llama_sampler_chain_add(smpl, llama_sampler_init_top_p(0.9, 1));
+llama_sampler_chain_add(smpl, llama_sampler_init_temp (0.4));
+llama_sampler_chain_add(smpl, llama_sampler_init_dist (1234));
+
 let n_ctx = llama_n_ctx(context)
 
 print("\nn_len = \(n_len), n_ctx = \(n_ctx), n_batch = \(context_params.n_batch), n_parallel = \(n_parallel), n_kv_req = \(n_kv_req)\n")
@@ -125,32 +138,7 @@ while n_cur <= n_len {
             continue
         }
 
-        var n_vocab = llama_n_vocab(model)
-        var logits = llama_get_logits_ith(context, i_batch[i])
-
-        var candidates: [llama_token_data] = .init(repeating: llama_token_data(), count: Int(n_vocab))
-
-        for token_id in 0 ..< n_vocab {
-            candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0))
-        }
-
-        var candidates_p: llama_token_data_array = .init(
-            data: &candidates,
-            size: candidates.count,
-            sorted: false
-        )
-
-        let top_k: Int32 = 40
-        let top_p: Float = 0.9
-        let temp: Float = 0.4
-
-        llama_sample_top_k(context, &candidates_p, top_k, 1)
-        llama_sample_top_p(context, &candidates_p, top_p, 1)
-        llama_sample_temp(context, &candidates_p, temp)
-
-        let new_token_id = llama_sample_token(context, &candidates_p)
-
-        // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+        let new_token_id = llama_sampler_sample(smpl, context, i_batch[i])
 
         // is it an end of stream? -> mark the stream as finished
         if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
@@ -210,9 +198,10 @@ if n_parallel > 1 {
 
 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")
+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_print_timings(context)
+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

@@ -1,15 +1,13 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
     LOG_TEE("\n");
@@ -21,8 +19,7 @@ int main(int argc, char ** argv) {
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
         return 1;
     }
 
@@ -65,6 +62,15 @@ int main(int argc, char ** argv) {
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
+    auto sparams = llama_sampler_chain_default_params();
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sparams.top_k));
+    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sparams.top_p, params.sparams.min_keep));
+    llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sparams.temp));
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sparams.seed));
+
     if (ctx == NULL) {
         fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
         return 1;
@@ -164,29 +170,7 @@ int main(int argc, char ** argv) {
                 continue;
             }
 
-            auto   n_vocab = llama_n_vocab(model);
-            auto * logits  = llama_get_logits_ith(ctx, i_batch[i]);
-
-            std::vector<llama_token_data> candidates;
-            candidates.reserve(n_vocab);
-
-            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-            }
-
-            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-            const int   top_k = 40;
-            const float top_p = 0.9f;
-            const float temp  = 0.4f;
-
-            llama_sample_top_k(ctx, &candidates_p, top_k, 1);
-            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-            llama_sample_temp (ctx, &candidates_p, temp);
-
-            const llama_token new_token_id = llama_sample_token(ctx, &candidates_p);
-
-            //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
 
             // is it an end of generation? -> mark the stream as finished
             if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
@@ -244,12 +228,15 @@ int main(int argc, char ** argv) {
     LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
     llama_batch_free(batch);
 
+    llama_sampler_free(smpl);
     llama_free(ctx);
     llama_free_model(model);
 

examples/benchmark/benchmark-matmult.cpp

@@ -183,7 +183,7 @@ int main(int argc, char ** argv)  {
 
     ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
 
-    TENSOR_DUMP(gf->nodes[0]);
+    TENSOR_DUMP(ggml_graph_node(gf, 0));
 
     printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
 
@@ -224,7 +224,7 @@ int main(int argc, char ** argv)  {
 
 
     // Let's use the F32 result from above as a reference for the quantized multiplication
-    float sum_of_F32_reference = tensor_sum_elements(gf->nodes[0]);
+    float sum_of_F32_reference = tensor_sum_elements(ggml_graph_node(gf, 0));
 
     printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
     printf("=====================================================================================\n");
@@ -252,7 +252,7 @@ int main(int argc, char ** argv)  {
 
         // Check that the matrix multiplication result is in the right ballpark
         // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
-        float sum_of_Q4_result = tensor_sum_elements(gf31->nodes[0]);
+        float sum_of_Q4_result = tensor_sum_elements(ggml_graph_node(gf31, 0));
         float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
         float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
 

examples/cvector-generator/cvector-generator.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 #include "ggml.h"
@@ -35,9 +36,7 @@ static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
     return ret;
 }
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     printf("\nexample usage:\n");
     printf("\n    CPU only:   %s -m ./llama-3.Q4_K_M.gguf\n", argv[0]);
     printf("\n    with GPU:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99\n", argv[0]);
@@ -390,8 +389,7 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
         return 1;
     }
 

examples/cvector-generator/pca.hpp

@@ -12,12 +12,9 @@
 
 #include <cstdio>
 #include <ctime>
+#include <random>
 #include <string>
-#include <tuple>
 #include <vector>
-#include <algorithm>
-#include <iostream>
-#include <fstream>
 
 #define DEBUG_POS 5
 
@@ -229,8 +226,8 @@ static ggml_status compute_piter(
         result.eigenvectors.resize(params.n_batch);
         result.distances.resize(params.n_batch);
         // get output nodes
-        for (int i = 0; i < gf->n_nodes; ++i) {
-            auto node = gf->nodes[i];
+        for (int i = 0; i < ggml_graph_n_nodes(gf); ++i) {
+            auto node = ggml_graph_node(gf, i);
             int iter = -1;
             // find b_tensor (without copying data from device)
             if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {

examples/embedding/embedding.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -79,8 +80,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
         return 1;
     }
 
@@ -90,14 +90,6 @@ int main(int argc, char ** argv) {
 
     print_build_info();
 
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
-
     llama_backend_init();
     llama_numa_init(params.numa);
 
@@ -313,8 +305,10 @@ int main(int argc, char ** argv) {
         if (notArray) fprintf(stdout, "\n}\n");
     }
 
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
+
     // clean up
-    llama_print_timings(ctx);
     llama_batch_free(batch);
     llama_free(ctx);
     llama_free_model(model);

examples/eval-callback/eval-callback.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 #include "ggml.h"
@@ -144,15 +145,12 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
     print_build_info();
 
-    std::mt19937 rng(params.seed);
-
     llama_backend_init();
     llama_numa_init(params.numa);
 
@@ -183,7 +181,8 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/export-lora/export-lora.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
@@ -369,7 +370,7 @@ struct lora_merge_ctx {
 
         // write data to output file
         {
-            auto result = gf->nodes[gf->n_nodes - 1];
+            auto * result = ggml_graph_node(gf, -1);
             size_t len = ggml_nbytes(result);
             if (read_buf.size() < len) {
                 read_buf.resize(len);
@@ -391,9 +392,7 @@ struct lora_merge_ctx {
     }
 };
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     printf("\nexample usage:\n");
     printf("\n  %s -m base-model.gguf --lora lora-file.gguf -o merged-model-f16.gguf\n", argv[0]);
     printf("\nNOTE: output model is F16\n");
@@ -403,8 +402,7 @@ static void print_usage(int argc, char ** argv, const gpt_params & params) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
         return 1;
     }
 

examples/gbnf-validator/gbnf-validator.cpp

@@ -1,9 +1,5 @@
-#define LLAMA_API_INTERNAL
-
-#include "grammar-parser.h"
-#include "ggml.h"
-#include "llama.h"
 #include "unicode.h"
+#include "llama-grammar.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -12,29 +8,28 @@
 #include <string>
 #include <vector>
 
-static bool llama_sample_grammar_string(struct llama_grammar * grammar, const std::string & input_str, size_t & error_pos, std::string & error_msg) {
-    auto decoded = decode_utf8(input_str, {});
-    const auto & code_points = decoded.first;
+static bool llama_grammar_validate(struct llama_grammar * grammar, const std::string & input_str, size_t & error_pos, std::string & error_msg) {
+    const auto cpts = unicode_cpts_from_utf8(input_str);
 
     const llama_grammar_rules  & rules      = llama_grammar_get_rules (grammar);
-          llama_grammar_stacks & cur_stacks = llama_grammar_get_stacks(grammar);
+          llama_grammar_stacks & stacks_cur = llama_grammar_get_stacks(grammar);
 
     size_t pos = 0;
-    for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
-        const llama_grammar_stacks prev_stacks = llama_grammar_get_stacks(grammar); // copy
+    for (const auto & cpt : cpts) {
+        const llama_grammar_stacks stacks_prev = llama_grammar_get_stacks(grammar); // copy
 
-        llama_grammar_accept(rules, prev_stacks, *it, cur_stacks);
+        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur);
 
-        if (cur_stacks.empty()) {
+        if (stacks_cur.empty()) {
             error_pos = pos;
-            error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'";
-            cur_stacks = prev_stacks;
+            error_msg = "Unexpected character '" + unicode_cpt_to_utf8(cpt) + "'";
+            stacks_cur = stacks_prev;
             return false;
         }
         ++pos;
     }
 
-    for (const auto & stack : cur_stacks) {
+    for (const auto & stack : stacks_cur) {
         if (stack.empty()) {
             return true;
         }
@@ -85,27 +80,7 @@ int main(int argc, char** argv) {
         grammar_str = buffer.str();
     }
 
-    // Parse the GBNF grammar
-    auto parsed_grammar = grammar_parser::parse(grammar_str.c_str());
-
-    // will be empty (default) if there are parse errors
-    if (parsed_grammar.rules.empty()) {
-        fprintf(stdout, "%s: failed to parse grammar\n", __func__);
-        return 1;
-    }
-
-    // Ensure that there is a "root" node.
-    if (parsed_grammar.symbol_ids.find("root") == parsed_grammar.symbol_ids.end()) {
-        fprintf(stdout, "%s: grammar does not contain a 'root' symbol\n", __func__);
-        return 1;
-    }
-
-    std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
-
-    // Create the LLAMA grammar
-    auto grammar = llama_grammar_init(
-            grammar_rules.data(),
-            grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    llama_grammar * grammar = llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root");
     if (grammar == nullptr) {
         throw std::runtime_error("Failed to initialize llama_grammar");
     }
@@ -122,7 +97,7 @@ int main(int argc, char** argv) {
     // Validate the input string against the grammar
     size_t error_pos;
     std::string error_msg;
-    bool is_valid = llama_sample_grammar_string(grammar, input_str, error_pos, error_msg);
+    bool is_valid = llama_grammar_validate(grammar, input_str, error_pos, error_msg);
 
     if (is_valid) {
         fprintf(stdout, "Input string is valid according to the grammar.\n");
@@ -131,7 +106,7 @@ int main(int argc, char** argv) {
     }
 
     // Clean up
-    llama_grammar_free(grammar);
+    llama_grammar_free_impl(grammar);
 
     return 0;
 }

examples/gen-docs/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-gen-docs)
+add_executable(${TARGET} gen-docs.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gen-docs/gen-docs.cpp

@@ -0,0 +1,52 @@
+#include "arg.h"
+#include "common.h"
+
+#include <fstream>
+#include <string>
+
+// Export usage message (-h) to markdown format
+
+static void export_md(std::string fname, llama_example ex) {
+    std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);
+
+    gpt_params params;
+    auto ctx_arg = gpt_params_parser_init(params, ex);
+
+    file << "| Argument | Explanation |\n";
+    file << "| -------- | ----------- |\n";
+    for (auto & opt : ctx_arg.options) {
+        file << "| `";
+        // args
+        for (const auto & arg : opt.args) {
+        if (arg == opt.args.front()) {
+                file << arg;
+                if (opt.args.size() > 1) file << ", ";
+            } else {
+                file << arg << (arg != opt.args.back() ? ", " : "");
+            }
+        }
+        // value hint
+        if (opt.value_hint) {
+            std::string md_value_hint(opt.value_hint);
+            string_replace_all(md_value_hint, "|", "\\|");
+            file << " " << md_value_hint;
+        }
+        if (opt.value_hint_2) {
+            std::string md_value_hint_2(opt.value_hint_2);
+            string_replace_all(md_value_hint_2, "|", "\\|");
+            file << " " << md_value_hint_2;
+        }
+        // help text
+        std::string md_help(opt.help);
+        string_replace_all(md_help, "\n", "<br/>");
+        string_replace_all(md_help, "|", "\\|");
+        file << "` | " << md_help << " |\n";
+    }
+}
+
+int main(int, char **) {
+    export_md("autogen-main.md", LLAMA_EXAMPLE_MAIN);
+    export_md("autogen-server.md", LLAMA_EXAMPLE_SERVER);
+
+    return 0;
+}

examples/gritlm/gritlm.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -9,7 +10,7 @@
 static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
     std::vector<std::vector<float>> result;
 
-    const llama_model * mdl = llama_get_model(ctx);
+    const llama_model * model = llama_get_model(ctx);
 
     llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
@@ -18,16 +19,16 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         const std::string input_string = instruction + sentences[i];
 
-        std::vector<llama_token> inputs = llama_tokenize(mdl, input_string, true, false);
+        std::vector<llama_token> inputs = llama_tokenize(model, input_string, true, false);
 
         const int32_t n_toks = inputs.size();
 
         // GritLM seems to have EOS = ""
         // https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
-        // inputs.push_back(llama_token_eos(mdl));
+        // inputs.push_back(llama_token_eos(model));
 
         // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = llama_tokenize(mdl, instruction, true, false).size();
+        const int32_t n_inst = llama_tokenize(model, instruction, true, false).size();
 
 #ifdef GRIT_DEBUG
         // debug tokens - should be matching as referenced in the GritLM sample
@@ -51,7 +52,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         llama_decode(ctx, batch);
 
         // get embedding dimensions
-        uint64_t n_embd = llama_n_embd(mdl);
+        uint64_t n_embd = llama_n_embd(model);
 
         // allocate embedding output
         std::vector<float> emb_unorm(n_embd, 0.0f);
@@ -92,11 +93,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
     return result;
 }
 
-static std::string generate(llama_context * ctx, const std::string & prompt, bool stream) {
+static std::string generate(llama_context * ctx, llama_sampler * smpl, const std::string & prompt, bool stream) {
     std::string result;
 
-    const llama_model * mdl = llama_get_model(ctx);
-    llama_token eos_token = llama_token_eos(mdl);
+    const llama_model * model = llama_get_model(ctx);
+    llama_token eos_token = llama_token_eos(model);
 
     llama_kv_cache_clear(ctx);
     llama_set_embeddings(ctx, false);
@@ -104,28 +105,24 @@ static std::string generate(llama_context * ctx, const std::string & prompt, boo
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
-    std::vector<llama_token> inputs = llama_tokenize(mdl, prompt, false, true);
+    std::vector<llama_token> inputs = llama_tokenize(model, prompt, false, true);
     int32_t i_current_token = 0;
 
     while (true) {
         llama_batch_clear(bat);
-        auto n_inputs = (int32_t)inputs.size();
-        for (int32_t i = 0; i < n_inputs; i++) {
-            llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
+        {
+            const int32_t n_inputs = inputs.size();
+
+            for (int32_t i = 0; i < n_inputs; i++) {
+                llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
+            }
         }
         inputs.clear();
 
         llama_decode(ctx, bat);
-        auto logits = llama_get_logits_ith(ctx, bat.n_tokens - 1);
 
-        auto candidates = std::vector<llama_token_data>(llama_n_vocab(mdl));
-        auto n_candidates = (int32_t)candidates.size();
-        for (int32_t token = 0; token < n_candidates; token++) {
-            candidates[token] = llama_token_data{ token, logits[token], 0.0f };
-        }
-        auto candidates_p = llama_token_data_array{ candidates.data(), candidates.size(), false };
+        llama_token token = llama_sampler_sample(smpl, ctx, bat.n_tokens - 1);
 
-        llama_token token = llama_sample_token_greedy(ctx, &candidates_p);
         if (token == eos_token) {
             break;
         }
@@ -157,8 +154,7 @@ static std::string gritlm_instruction(const std::string & instruction) {
 int main(int argc, char * argv[]) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
@@ -167,10 +163,18 @@ int main(int argc, char * argv[]) {
 
     llama_backend_init();
 
-    llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
 
     // create generation context
-    llama_context * ctx = llama_new_context_with_model(mdl, cparams);
+    llama_context * ctx = llama_new_context_with_model(model, cparams);
+
+    auto sparams = llama_sampler_chain_default_params();
+
+    sparams.no_perf = false;
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
 
     // ### Embedding/Representation ###
     // samples taken from: https://github.com/ContextualAI/gritlm#basic
@@ -191,7 +195,7 @@ int main(int argc, char * argv[]) {
         const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
         const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
 
-        const int n_embd = llama_n_embd(mdl);
+        const int n_embd = llama_n_embd(model);
 
         const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
         const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
@@ -208,11 +212,12 @@ int main(int argc, char * argv[]) {
     // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
     {
         const std::string prompt = "<|user|>\nPlease write me a poem about my recent hike of Mt. Fuji at midnight in the style of Shakespeare.\n<|assistant|>\n";
-        std::string response = generate(ctx, prompt, true);
+        std::string response = generate(ctx, smpl, prompt, true);
     }
 
+    llama_sampler_free(smpl);
     llama_free(ctx);
-    llama_free_model(mdl);
+    llama_free_model(model);
     llama_backend_free();
 
     return 0;

examples/imatrix/imatrix.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -17,9 +18,7 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s \\\n"
             "       -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \\\n"
@@ -579,8 +578,7 @@ int main(int argc, char ** argv) {
     params.logits_all = true;
     params.verbosity = 1;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
         return 1;
     }
 
@@ -638,7 +636,8 @@ int main(int argc, char ** argv) {
 
     g_collector.save_imatrix();
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/infill/infill.cpp

@@ -1,8 +1,8 @@
+#include "arg.h"
 #include "common.h"
-
 #include "console.h"
+#include "sampling.h"
 #include "llama.h"
-#include "grammar-parser.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -34,6 +34,7 @@
 
 static llama_context           ** g_ctx;
 static llama_model             ** g_model;
+static gpt_sampler             ** g_smpl;
 static gpt_params               * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
@@ -81,7 +82,7 @@ static void write_logfile(
     yaml_dump_string_multiline(logfile, "output", output.c_str());
     yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
-    llama_dump_timing_info_yaml(logfile, ctx);
+    llama_perf_dump_yaml(logfile, ctx);
     fclose(logfile);
 }
 
@@ -93,7 +94,7 @@ static void sigint_handler(int signo) {
         } else {
             console::cleanup();
             printf("\n");
-            llama_print_timings(*g_ctx);
+            gpt_perf_print(*g_ctx, *g_smpl);
             write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
             _exit(130);
         }
@@ -103,14 +104,14 @@ static void sigint_handler(int signo) {
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    llama_sampling_params & sparams = params.sparams;
     g_params = &params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
         return 1;
     }
 
+    auto & sparams = params.sparams;
+
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("infill", "log"));
     LOG_TEE("Log start\n");
@@ -156,26 +157,19 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
+    print_build_info();
 
     LOG("%s: llama backend init\n", __func__);
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
+    llama_model * model = nullptr;
+    llama_context * ctx = nullptr;
+    gpt_sampler  * smpl = nullptr;
 
     g_model = &model;
     g_ctx = &ctx;
+    g_smpl = &smpl;
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
@@ -305,16 +299,14 @@ int main(int argc, char ** argv) {
             LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
+    smpl = gpt_sampler_init(model, sparams);
+
+    LOG_TEE("sampling seed: %u\n", gpt_sampler_get_seed(smpl));
+    LOG_TEE("sampling: \n%s\n", sparams.print().c_str());
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 
     LOG_TEE("\n#####  Infill mode  #####\n\n");
-    if (params.infill) {
-        printf("\n************\n");
-        printf("no need to specify '--infill', always running infill\n");
-        printf("************\n\n");
-    }
     if (params.interactive) {
         const char *control_message;
         if (params.multiline_input) {
@@ -349,8 +341,6 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> embd;
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
-
     while (n_remain != 0 || params.interactive) {
         // predict
         if (!embd.empty()) {
@@ -421,11 +411,11 @@ int main(int argc, char ** argv) {
         embd.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, nullptr);
+            const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            gpt_sampler_accept(smpl, id, true);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
+            // LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, smpl->prev.to_vector()).c_str());
 
             embd.push_back(id);
 
@@ -444,7 +434,7 @@ int main(int argc, char ** argv) {
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+                gpt_sampler_accept(smpl, embd_inp[n_consumed], false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -476,7 +466,7 @@ int main(int argc, char ** argv) {
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
             // deal with eot token in infill mode
-            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
+            if ((gpt_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if (is_interacting && !params.interactive_first) {
                     // print an eot token
                     printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
@@ -542,7 +532,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of generation tokens in interactive mode
-            else if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
+            else if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -615,7 +605,7 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    llama_sampling_reset(ctx_sampling);
+                    gpt_sampler_reset(smpl);
                 }
                 is_interacting = false;
             }
@@ -638,13 +628,14 @@ int main(int argc, char ** argv) {
         fflush(stdout);
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    gpt_perf_print(ctx, smpl);
     write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
     llama_free(ctx);
     llama_free_model(model);
 
-    llama_sampling_free(ctx_sampling);
+    gpt_sampler_free(smpl);
     llama_backend_free();
 
 #ifndef LOG_DISABLE_LOGS

examples/llama-bench/llama-bench.cpp

@@ -249,6 +249,7 @@ struct cmd_params {
     ggml_sched_priority prio;
     int delay;
     bool verbose;
+    bool progress;
     output_formats output_format;
     output_formats output_format_stderr;
 };
@@ -280,6 +281,7 @@ static const cmd_params cmd_params_defaults = {
     /* prio                 */ GGML_SCHED_PRIO_NORMAL,
     /* delay                */ 0,
     /* verbose              */ false,
+    /* progress             */ false,
     /* output_format        */ MARKDOWN,
     /* output_format_stderr */ NONE,
 };
@@ -319,6 +321,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -o, --output <csv|json|jsonl|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
     printf("  -oe, --output-err <csv|json|jsonl|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
     printf("  -v, --verbose                             (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    printf("  --progress                                (default: %s)\n", cmd_params_defaults.progress ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
 }
@@ -364,6 +367,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     params.numa = cmd_params_defaults.numa;
     params.prio = cmd_params_defaults.prio;
     params.delay = cmd_params_defaults.delay;
+    params.progress = cmd_params_defaults.progress;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -616,6 +620,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             invalid_param = !output_format_from_str(argv[i], params.output_format_stderr);
         } else if (arg == "-v" || arg == "--verbose") {
             params.verbose = true;
+        } else if (arg == "--progress") {
+            params.progress = true;
         } else {
             invalid_param = true;
             break;
@@ -1523,7 +1529,13 @@ int main(int argc, char ** argv) {
     llama_model * lmodel = nullptr;
     const cmd_params_instance * prev_inst = nullptr;
 
+    int params_idx = 0;
+    auto params_count = params_instances.size();
     for (const auto & inst : params_instances) {
+        params_idx ++;
+        if (params.progress) {
+            fprintf(stderr, "llama-bench: benchmark %d/%ld: starting\n", params_idx, params_count);
+        }
         // keep the same model between tests when possible
         if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
             if (lmodel) {
@@ -1556,7 +1568,7 @@ int main(int argc, char ** argv) {
 
         struct ggml_threadpool_params tpp = ggml_threadpool_params_default(t.n_threads);
         if (!parse_cpu_mask(t.cpu_mask, tpp.cpumask)) {
-            LOG_TEE("%s: failed to parse cpu-mask: %s\n", __func__, t.cpu_mask.c_str());
+            fprintf(stderr, "%s: failed to parse cpu-mask: %s\n", __func__, t.cpu_mask.c_str());
             exit(1);
         }
         tpp.strict_cpu = t.cpu_strict;
@@ -1565,7 +1577,7 @@ int main(int argc, char ** argv) {
 
         struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp);
         if (!threadpool) {
-            LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
+            fprintf(stderr, "%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
             exit(1);
         }
 
@@ -1573,10 +1585,16 @@ int main(int argc, char ** argv) {
 
         // warmup run
         if (t.n_prompt > 0) {
+            if (params.progress) {
+                fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup prompt run\n", params_idx, params_count);
+            }
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
             test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
+            if (params.progress) {
+                fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup generation run\n", params_idx, params_count);
+            }
             test_gen(ctx, 1, 0, t.n_threads);
         }
 
@@ -1586,9 +1604,15 @@ int main(int argc, char ** argv) {
             uint64_t t_start = get_time_ns();
 
             if (t.n_prompt > 0) {
+                if (params.progress) {
+                    fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, i + 1, params.reps);
+                }
                 test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
             }
             if (t.n_gen > 0) {
+                if (params.progress) {
+                    fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, i + 1, params.reps);
+                }
                 test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
             }
 
@@ -1606,7 +1630,7 @@ int main(int argc, char ** argv) {
             fflush(p_err->fout);
         }
 
-        llama_print_timings(ctx);
+        llama_perf_context_print(ctx);
 
         llama_free(ctx);
 

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -120,8 +120,8 @@ Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmo
     LOGi("Using %d threads", n_threads);
 
     llama_context_params ctx_params = llama_context_default_params();
-    ctx_params.seed  = 1234;
-    ctx_params.n_ctx = 2048;
+
+    ctx_params.n_ctx           = 2048;
     ctx_params.n_threads       = n_threads;
     ctx_params.n_threads_batch = n_threads;
 
@@ -269,12 +269,6 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
     return env->NewStringUTF(result.str().c_str());
 }
 
-extern "C"
-JNIEXPORT void JNICALL
-Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
-    llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
-}
-
 extern "C"
 JNIEXPORT jlong JNICALL
 Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
@@ -311,6 +305,29 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
     return reinterpret_cast<jlong>(batch);
 }
 
+extern "C"
+JNIEXPORT void JNICALL
+Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
+    llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
+}
+
+extern "C"
+JNIEXPORT jlong JNICALL
+Java_android_llama_cpp_LLamaAndroid_new_1sampler(JNIEnv *, jobject) {
+    auto sparams = llama_sampler_chain_default_params();
+    sparams.no_perf = true;
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+    llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
+
+    return reinterpret_cast<jlong>(smpl);
+}
+
+extern "C"
+JNIEXPORT void JNICALL
+Java_android_llama_cpp_LLamaAndroid_free_1sampler(JNIEnv *, jobject, jlong sampler_pointer) {
+    llama_sampler_free(reinterpret_cast<llama_sampler *>(sampler_pointer));
+}
+
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) {
@@ -381,31 +398,21 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         jobject,
         jlong context_pointer,
         jlong batch_pointer,
+        jlong sampler_pointer,
         jint n_len,
         jobject intvar_ncur
 ) {
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
-    const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
+    const auto batch   = reinterpret_cast<llama_batch   *>(batch_pointer);
+    const auto sampler = reinterpret_cast<llama_sampler *>(sampler_pointer);
     const auto model = llama_get_model(context);
 
     if (!la_int_var) la_int_var = env->GetObjectClass(intvar_ncur);
     if (!la_int_var_value) la_int_var_value = env->GetMethodID(la_int_var, "getValue", "()I");
     if (!la_int_var_inc) la_int_var_inc = env->GetMethodID(la_int_var, "inc", "()V");
 
-    auto n_vocab = llama_n_vocab(model);
-    auto logits = llama_get_logits_ith(context, batch->n_tokens - 1);
-
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-    }
-
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
     // sample the most likely token
-    const auto new_token_id = llama_sample_token_greedy(context, &candidates_p);
+    const auto new_token_id = llama_sampler_sample(sampler, context, -1);
 
     const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
     if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {

examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt

@@ -45,8 +45,10 @@ class LLamaAndroid {
     private external fun free_context(context: Long)
     private external fun backend_init(numa: Boolean)
     private external fun backend_free()
-    private external fun free_batch(batch: Long)
     private external fun new_batch(nTokens: Int, embd: Int, nSeqMax: Int): Long
+    private external fun free_batch(batch: Long)
+    private external fun new_sampler(): Long
+    private external fun free_sampler(sampler: Long)
     private external fun bench_model(
         context: Long,
         model: Long,
@@ -69,6 +71,7 @@ class LLamaAndroid {
     private external fun completion_loop(
         context: Long,
         batch: Long,
+        sampler: Long,
         nLen: Int,
         ncur: IntVar
     ): String?
@@ -101,8 +104,11 @@ class LLamaAndroid {
                     val batch = new_batch(512, 0, 1)
                     if (batch == 0L) throw IllegalStateException("new_batch() failed")
 
+                    val sampler = new_sampler()
+                    if (sampler == 0L) throw IllegalStateException("new_sampler() failed")
+
                     Log.i(tag, "Loaded model $pathToModel")
-                    threadLocalState.set(State.Loaded(model, context, batch))
+                    threadLocalState.set(State.Loaded(model, context, batch, sampler))
                 }
                 else -> throw IllegalStateException("Model already loaded")
             }
@@ -114,7 +120,7 @@ class LLamaAndroid {
             is State.Loaded -> {
                 val ncur = IntVar(completion_init(state.context, state.batch, message, nlen))
                 while (ncur.value <= nlen) {
-                    val str = completion_loop(state.context, state.batch, nlen, ncur)
+                    val str = completion_loop(state.context, state.batch, state.sampler, nlen, ncur)
                     if (str == null) {
                         break
                     }
@@ -138,6 +144,7 @@ class LLamaAndroid {
                     free_context(state.context)
                     free_model(state.model)
                     free_batch(state.batch)
+                    free_sampler(state.sampler);
 
                     threadLocalState.set(State.Idle)
                 }
@@ -161,7 +168,7 @@ class LLamaAndroid {
 
         private sealed interface State {
             data object Idle: State
-            data class Loaded(val model: Long, val context: Long, val batch: Long): State
+            data class Loaded(val model: Long, val context: Long, val batch: Long, val sampler: Long): State
         }
 
         // Enforce only one instance of Llm.

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -24,6 +24,7 @@ func llama_batch_add(_ batch: inout llama_batch, _ id: llama_token, _ pos: llama
 actor LlamaContext {
     private var model: OpaquePointer
     private var context: OpaquePointer
+    private var sampling: UnsafeMutablePointer<llama_sampler>
     private var batch: llama_batch
     private var tokens_list: [llama_token]
     var is_done: Bool = false
@@ -42,9 +43,15 @@ actor LlamaContext {
         self.tokens_list = []
         self.batch = llama_batch_init(512, 0, 1)
         self.temporary_invalid_cchars = []
+        let sparams = llama_sampler_chain_default_params()
+        self.sampling = llama_sampler_chain_init(sparams)
+        llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
+        llama_sampler_chain_add(self.sampling, llama_sampler_init_softmax())
+        llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
     }
 
     deinit {
+        llama_sampler_free(sampling)
         llama_batch_free(batch)
         llama_free(context)
         llama_free_model(model)
@@ -69,7 +76,6 @@ actor LlamaContext {
         print("Using \(n_threads) threads")
 
         var ctx_params = llama_context_default_params()
-        ctx_params.seed  = 1234
         ctx_params.n_ctx = 2048
         ctx_params.n_threads       = Int32(n_threads)
         ctx_params.n_threads_batch = Int32(n_threads)
@@ -144,20 +150,7 @@ actor LlamaContext {
     func completion_loop() -> String {
         var new_token_id: llama_token = 0
 
-        let n_vocab = llama_n_vocab(model)
-        let logits = llama_get_logits_ith(context, batch.n_tokens - 1)
-
-        var candidates = Array<llama_token_data>()
-        candidates.reserveCapacity(Int(n_vocab))
-
-        for token_id in 0..<n_vocab {
-            candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0))
-        }
-        candidates.withUnsafeMutableBufferPointer() { buffer in
-            var candidates_p = llama_token_data_array(data: buffer.baseAddress, size: buffer.count, sorted: false)
-
-            new_token_id = llama_sample_token_greedy(context, &candidates_p)
-        }
+        new_token_id = llama_sampler_sample(sampling, context, batch.n_tokens - 1)
 
         if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
             print("\n")

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/clip.cpp

@@ -2449,7 +2449,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     ggml_backend_graph_compute(ctx->backend, gf);
 
     // the last node is the embedding tensor
-    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embeddings = ggml_graph_node(gf, -1);
 
     // copy the embeddings to the location passed by the user
     ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));

examples/llava/llava-cli.cpp

@@ -1,11 +1,12 @@
-#include "ggml.h"
+#include "arg.h"
+#include "base64.hpp"
 #include "log.h"
 #include "common.h"
+#include "sampling.h"
 #include "clip.h"
 #include "llava.h"
 #include "llama.h"
-
-#include "base64.hpp"
+#include "ggml.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -40,11 +41,11 @@ static bool eval_string(struct llama_context * ctx_llama, const char* str, int n
     return true;
 }
 
-static const char * sample(struct llama_sampling_context * ctx_sampling,
+static const char * sample(struct gpt_sampler * smpl,
                            struct llama_context * ctx_llama,
                            int * n_past) {
-    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
-    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
+    const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
+    gpt_sampler_accept(smpl, id, true);
     static std::string ret;
     if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
@@ -112,9 +113,7 @@ struct llava_context {
     struct llama_model * model = NULL;
 };
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\n example usage:\n");
     LOG_TEE("\n     %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
     LOG_TEE("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
@@ -191,15 +190,15 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     LOG_TEE("\n");
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-    if (!ctx_sampling) {
+    struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
+    if (!smpl) {
         fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
         exit(1);
     }
 
     std::string response = "";
     for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+        const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
         response += tmp;
         if (strcmp(tmp, "</s>") == 0) break;
         if (strstr(tmp, "###")) break; // Yi-VL behavior
@@ -211,7 +210,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         fflush(stdout);
     }
 
-    llama_sampling_free(ctx_sampling);
+    gpt_sampler_free(smpl);
     printf("\n");
 }
 
@@ -280,8 +279,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
         return 1;
     }
 
@@ -293,7 +291,7 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        print_usage(argc, argv, {});
+        print_usage(argc, argv);
         return 1;
     }
     auto model = llava_init(&params);
@@ -310,7 +308,7 @@ int main(int argc, char ** argv) {
         // process the prompt
         process_prompt(ctx_llava, image_embed, &params, params.prompt);
 
-        llama_print_timings(ctx_llava->ctx_llama);
+        llama_perf_context_print(ctx_llava->ctx_llama);
         llava_image_embed_free(image_embed);
         ctx_llava->model = NULL;
         llava_free(ctx_llava);
@@ -327,7 +325,7 @@ int main(int argc, char ** argv) {
             // process the prompt
             process_prompt(ctx_llava, image_embed, &params, params.prompt);
 
-            llama_print_timings(ctx_llava->ctx_llama);
+            llama_perf_context_print(ctx_llava->ctx_llama);
             llava_image_embed_free(image_embed);
             ctx_llava->model = NULL;
             llava_free(ctx_llava);

examples/llava/llava.cpp

@@ -184,7 +184,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
     ggml_build_forward_expand(gf, flatten);
     ggml_graph_compute_with_ctx(model.ctx, gf, 1);
-    struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor* result = ggml_graph_node(gf, -1);
 
     memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
     // append without newline tokens (default behavior in llava_arch when not using unpad ):

examples/llava/minicpmv-cli.cpp

@@ -1,9 +1,11 @@
-#include "ggml.h"
+#include "arg.h"
 #include "log.h"
 #include "common.h"
+#include "sampling.h"
 #include "clip.h"
 #include "llava.h"
 #include "llama.h"
+#include "ggml.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -16,8 +18,8 @@ struct llava_context {
 };
 
 static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+    LOG_TEE("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
@@ -163,11 +165,11 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
     LOG_TEE("%s: image token past: %d\n", __func__, n_past);
 }
 
-static const char * sample(struct llama_sampling_context * ctx_sampling,
+static const char * sample(struct gpt_sampler * smpl,
                            struct llama_context * ctx_llama,
                            int * n_past) {
-    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
-    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
+    const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
+    gpt_sampler_accept(smpl, id, true);
     static std::string ret;
     if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
@@ -214,7 +216,7 @@ static struct llava_context * minicpmv_init(gpt_params * params, const std::stri
     return ctx_llava;
 }
 
-static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
+static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
     std::string user_prompt = prompt;
     int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
     if (!is_first) {
@@ -238,13 +240,13 @@ static struct llama_sampling_context * llama_init(struct llava_context * ctx_lla
 
     LOG_TEE("\n");
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-    return ctx_sampling;
+    struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
+    return smpl;
 }
 
-static const char * llama_loop(struct llava_context * ctx_llava,struct llama_sampling_context * ctx_sampling, int &n_past){
+static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampler * smpl, int &n_past){
 
-    const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+    const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
     return tmp;
 }
 
@@ -253,8 +255,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        show_additional_info(argc, argv);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
         return 1;
     }
 
@@ -266,7 +267,6 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     if (params.mmproj.empty() || (params.image.empty())) {
-        gpt_params_print_usage(argc, argv, params);
         show_additional_info(argc, argv);
         return 1;
     }
@@ -278,12 +278,12 @@ int main(int argc, char ** argv) {
         if (!params.prompt.empty()) {
             LOG_TEE("<user>%s\n", params.prompt.c_str());
             LOG_TEE("<assistant>");
-            auto ctx_sampling = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
+            auto smpl = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
             const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
             std::string response = "";
             bool have_tmp = false;
             for (int i = 0; i < max_tgt_len; i++) {
-                auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                auto tmp = llama_loop(ctx_llava, smpl, n_past);
                 response += tmp;
                 if (strcmp(tmp, "</s>") == 0){
                     if(!have_tmp)continue;
@@ -296,18 +296,18 @@ int main(int argc, char ** argv) {
 
                 fflush(stdout);
             }
-            llama_sampling_free(ctx_sampling);
+            gpt_sampler_free(smpl);
         }else {
             while (true) {
                 LOG_TEE("<user>");
                 std::string prompt;
                 std::getline(std::cin, prompt);
                 LOG_TEE("<assistant>");
-                auto ctx_sampling = llama_init(ctx_llava, &params, prompt, n_past, true);
+                auto smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
                 const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
                 std::string response = "";
                 for (int i = 0; i < max_tgt_len; i++) {
-                    auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                    auto tmp = llama_loop(ctx_llava, smpl, n_past);
                     response += tmp;
                     if (strcmp(tmp, "</s>") == 0) break;
                     if (strstr(tmp, "###")) break; // Yi-VL behavior
@@ -315,11 +315,11 @@ int main(int argc, char ** argv) {
                     if (strstr(response.c_str(), "<user>")) break; // minicpm-v
                     fflush(stdout);
                 }
-                llama_sampling_free(ctx_sampling);
+                gpt_sampler_free(smpl);
             }
         }
         printf("\n");
-        llama_print_timings(ctx_llava->ctx_llama);
+        llama_perf_context_print(ctx_llava->ctx_llama);
 
         ctx_llava->model = NULL;
         llava_free(ctx_llava);

examples/lookahead/lookahead.cpp

@@ -1,7 +1,8 @@
+#include "arg.h"
 #include "common.h"
+#include "sampling.h"
 #include "llama.h"
 
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
@@ -37,8 +38,7 @@ struct ngram_container {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
@@ -118,7 +118,7 @@ int main(int argc, char ** argv) {
     llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
 
     // target model sampling context
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+    struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams);
 
     // verification n-grams
     std::vector<ngram_data> ngrams_cur(G);
@@ -159,9 +159,9 @@ int main(int argc, char ** argv) {
 
     // sample first token
     {
-        id = llama_sampling_sample(ctx_sampling, ctx, NULL, 0);
+        id = gpt_sampler_sample(smpl, ctx, 0);
 
-        llama_sampling_accept(ctx_sampling, ctx, id, true);
+        gpt_sampler_accept(smpl, id, true);
 
         {
             const std::string token_str = llama_token_to_piece(ctx, id);
@@ -284,9 +284,9 @@ int main(int argc, char ** argv) {
             }
 
             // sample the next token
-            id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_batch);
+            id = gpt_sampler_sample(smpl, ctx, i_batch);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            gpt_sampler_accept(smpl, id, true);
 
             // print
             {
@@ -361,7 +361,7 @@ int main(int argc, char ** argv) {
                 if (v == 0) {
                     // sample from the last level
                     for (int i = 0; i < W; i++) {
-                        tokens_j[N - 2][i] = llama_sampling_sample(ctx_sampling, ctx, NULL, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
+                        tokens_j[N - 2][i] = gpt_sampler_sample(smpl, ctx, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
                     }
                 } else {
                     for (int i = 0; i < W; i++) {
@@ -468,10 +468,12 @@ int main(int argc, char ** argv) {
     LOG_TEE("n_predict = %d\n", n_predict);
     LOG_TEE("n_accept  = %d\n", n_accept);
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    gpt_perf_print(ctx, smpl);
+
+    gpt_sampler_free(smpl);
 
     llama_kv_cache_view_free(&kvc_view);
-    llama_sampling_free(ctx_sampling);
 
     llama_batch_free(batch);
 

examples/lookup/lookup-create.cpp

@@ -1,7 +1,8 @@
-#include "ggml.h"
-#include "llama.h"
+#include "arg.h"
 #include "common.h"
 #include "ngram-cache.h"
+#include "ggml.h"
+#include "llama.h"
 
 #include <cstdint>
 #include <fstream>
@@ -13,8 +14,7 @@
 int main(int argc, char ** argv){
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
@@ -40,4 +40,6 @@ int main(int argc, char ** argv){
     fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
 
     llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
+
+    return 0;
 }

examples/lookup/lookup-stats.cpp

@@ -1,8 +1,9 @@
-#include "ggml.h"
+#include "arg.h"
 #include "common.h"
-#include "llama.h"
 #include "log.h"
 #include "ngram-cache.h"
+#include "llama.h"
+#include "ggml.h"
 
 #include <cmath>
 #include <cstdint>
@@ -15,8 +16,7 @@
 int main(int argc, char ** argv){
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 

examples/lookup/lookup.cpp

@@ -1,21 +1,20 @@
+#include "arg.h"
 #include "ggml.h"
-#include "llama.h"
 #include "common.h"
 #include "ngram-cache.h"
+#include "sampling.h"
+#include "llama.h"
 
-#include <cmath>
 #include <cstdint>
 #include <cstdio>
 #include <fstream>
 #include <string>
 #include <vector>
-#include <unordered_map>
 
 int main(int argc, char ** argv){
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
@@ -106,7 +105,7 @@ int main(int argc, char ** argv){
 
     bool has_eos = false;
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+    struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams);
 
     std::vector<llama_token> draft;
 
@@ -130,9 +129,9 @@ int main(int argc, char ** argv){
         int i_dft = 0;
         while (true) {
             // sample from the target model
-            llama_token id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_dft);
+            llama_token id = gpt_sampler_sample(smpl, ctx, i_dft);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            gpt_sampler_accept(smpl, id, true);
 
             const std::string token_str = llama_token_to_piece(ctx, id);
 
@@ -240,10 +239,11 @@ int main(int argc, char ** argv){
     LOG_TEE("n_accept     = %d\n", n_accept);
     LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
-    LOG_TEE("\ntarget:\n");
-    llama_print_timings(ctx);
+    LOG_TEE("\ntarget:\n\n");
+    gpt_perf_print(ctx, smpl);
+
+    gpt_sampler_free(smpl);
 
-    llama_sampling_free(ctx_sampling);
     llama_batch_free(batch_tgt);
 
     llama_free(ctx);

examples/main/main.cpp

@@ -1,6 +1,7 @@
+#include "arg.h"
 #include "common.h"
-
 #include "console.h"
+#include "sampling.h"
 #include "llama.h"
 
 #include <cassert>
@@ -33,6 +34,7 @@
 
 static llama_context           ** g_ctx;
 static llama_model             ** g_model;
+static gpt_sampler             ** g_smpl;
 static gpt_params               * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
@@ -40,6 +42,13 @@ static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting  = false;
 static bool need_insert_eot = false;
 
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n  text generation:     %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]);
+    printf("\n  chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]);
+    printf("\n");
+}
+
 static bool file_exists(const std::string & path) {
     std::ifstream f(path.c_str());
     return f.good();
@@ -92,7 +101,7 @@ static void write_logfile(
     yaml_dump_string_multiline(logfile, "output", output.c_str());
     yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
-    llama_dump_timing_info_yaml(logfile, ctx);
+    llama_perf_dump_yaml(logfile, ctx);
     fclose(logfile);
 }
 
@@ -105,7 +114,7 @@ static void sigint_handler(int signo) {
         } else {
             console::cleanup();
             printf("\n");
-            llama_print_timings(*g_ctx);
+            gpt_perf_print(*g_ctx, *g_smpl);
             write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
             _exit(130);
         }
@@ -121,8 +130,7 @@ static void llama_log_callback_logTee(ggml_log_level level, const char * text, v
 
 static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, std::string role, std::string content) {
     llama_chat_msg new_msg{role, content};
-    auto formatted = llama_chat_format_single(
-        model, g_params->chat_template, chat_msgs, new_msg, role == "user");
+    auto formatted = llama_chat_format_single(model, g_params->chat_template, chat_msgs, new_msg, role == "user");
     chat_msgs.push_back({role, content});
     LOG("formatted: %s\n", formatted.c_str());
     return formatted;
@@ -131,13 +139,11 @@ static std::string chat_add_and_format(struct llama_model * model, std::vector<l
 int main(int argc, char ** argv) {
     gpt_params params;
     g_params = &params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) {
         return 1;
     }
 
-    llama_sampling_params & sparams = params.sparams;
+    auto & sparams = params.sparams;
 
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("main", "log"));
@@ -183,27 +189,21 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
+    print_build_info();
 
     LOG("%s: llama backend init\n", __func__);
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-    llama_context * ctx_guidance = NULL;
+    llama_model * model = nullptr;
+    llama_context * ctx = nullptr;
+    gpt_sampler * smpl = nullptr;
+
     std::vector<llama_chat_msg> chat_msgs;
+
     g_model = &model;
     g_ctx = &ctx;
+    g_smpl = &smpl;
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
@@ -211,10 +211,6 @@ int main(int argc, char ** argv) {
 
     model = llama_init.model;
     ctx = llama_init.context;
-    if (sparams.cfg_scale > 1.f) {
-        struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
-        ctx_guidance = llama_new_context_with_model(model, lparams);
-    }
 
     if (model == NULL) {
         LOG_TEE("%s: error: unable to load model\n", __func__);
@@ -251,9 +247,6 @@ int main(int argc, char ** argv) {
     }
 
     llama_attach_threadpool(ctx, threadpool, threadpool_batch);
-    if (ctx_guidance) {
-        llama_attach_threadpool(ctx_guidance, threadpool, threadpool_batch);
-    }
 
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
@@ -337,24 +330,6 @@ int main(int argc, char ** argv) {
     }
 
     // Tokenize negative prompt
-    std::vector<llama_token> guidance_inp;
-    int guidance_offset = 0;
-    int original_prompt_len = 0;
-    if (ctx_guidance) {
-        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
-
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, true, true);
-        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
-
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true, true);
-        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
-
-        original_prompt_len = original_inp.size();
-        guidance_offset = (int)guidance_inp.size() - original_prompt_len;
-        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
-        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
-    }
-
     if ((int) embd_inp.size() > n_ctx - 4) {
         LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
         return 1;
@@ -421,15 +396,6 @@ int main(int argc, char ** argv) {
             LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
-        if (ctx_guidance) {
-            LOG_TEE("\n");
-            LOG_TEE("%s: negative prompt: '%s'\n", __func__, sparams.cfg_negative_prompt.c_str());
-            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
-            for (int i = 0; i < (int) guidance_inp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
-            }
-        }
-
         if (params.n_keep > add_bos) {
             LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
@@ -495,8 +461,17 @@ int main(int argc, char ** argv) {
             }
         }
     }
-    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
-    LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
+
+    smpl = gpt_sampler_init(model, sparams);
+    if (!smpl) {
+        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }
+
+    LOG_TEE("sampling seed: %u\n", gpt_sampler_get_seed(smpl));
+    LOG_TEE("sampling params: \n%s\n", sparams.print().c_str());
+    LOG_TEE("sampler constr: \n%s\n", gpt_sampler_print(smpl).c_str());
+
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
 
     // group-attention state
@@ -543,7 +518,6 @@ int main(int argc, char ** argv) {
     int n_remain           = params.n_predict;
     int n_consumed         = 0;
     int n_session_consumed = 0;
-    int n_past_guidance    = 0;
 
     std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
     std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
@@ -555,7 +529,6 @@ int main(int argc, char ** argv) {
     display = params.display_prompt;
 
     std::vector<llama_token> embd;
-    std::vector<llama_token> embd_guidance;
 
     // tokenized antiprompts
     std::vector<std::vector<llama_token>> antiprompt_ids;
@@ -565,12 +538,6 @@ int main(int argc, char ** argv) {
         antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
     }
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
-    if (!ctx_sampling) {
-        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
-        exit(1);
-    }
-
     if (llama_model_has_encoder(model)) {
         int enc_input_size = embd_inp.size();
         llama_token * enc_input_buf = embd_inp.data();
@@ -612,7 +579,7 @@ int main(int argc, char ** argv) {
                 // if we run out of context:
                 // - take the n_keep first tokens from the original prompt (via n_past)
                 // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) >= n_ctx) {
+                if (n_past + (int) embd.size() >= n_ctx) {
                     if (params.n_predict == -2) {
                         LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                         break;
@@ -629,11 +596,7 @@ int main(int argc, char ** argv) {
 
                     n_past -= n_discard;
 
-                    if (ctx_guidance) {
-                        n_past_guidance -= n_discard;
-                    }
-
-                    LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+                    LOG("after swap: n_past = %d\n", n_past);
 
                     LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
@@ -686,46 +649,6 @@ int main(int argc, char ** argv) {
                 }
             }
 
-            // evaluate tokens in batches
-            // embd is typically prepared beforehand to fit within a batch, but not always
-            if (ctx_guidance) {
-                int input_size = 0;
-                llama_token * input_buf = NULL;
-
-                if (n_past_guidance < (int) guidance_inp.size()) {
-                    // Guidance context should have the same data with these modifications:
-                    //
-                    // * Replace the initial prompt
-                    // * Shift everything by guidance_offset
-                    embd_guidance = guidance_inp;
-                    if (embd.begin() + original_prompt_len < embd.end()) {
-                        embd_guidance.insert(
-                            embd_guidance.end(),
-                            embd.begin() + original_prompt_len,
-                            embd.end()
-                        );
-                    }
-
-                    input_buf  = embd_guidance.data();
-                    input_size = embd_guidance.size();
-
-                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
-                } else {
-                    input_buf  = embd.data();
-                    input_size = embd.size();
-                }
-
-                for (int i = 0; i < input_size; i += params.n_batch) {
-                    int n_eval = std::min(input_size - i, params.n_batch);
-                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
-                        LOG_TEE("%s : failed to eval\n", __func__);
-                        return 1;
-                    }
-
-                    n_past_guidance += n_eval;
-                }
-            }
-
             for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
                 int n_eval = (int) embd.size() - i;
                 if (n_eval > params.n_batch) {
@@ -755,7 +678,6 @@ int main(int argc, char ** argv) {
         }
 
         embd.clear();
-        embd_guidance.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
             // optionally save the session on first sample (for faster prompt loading next time)
@@ -766,11 +688,11 @@ int main(int argc, char ** argv) {
                 LOG("saved session to %s\n", path_session.c_str());
             }
 
-            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
+            const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, /* apply_grammar= */ true);
+            gpt_sampler_accept(smpl, id, /* apply_grammar= */ true);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
+            // LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, smpl->prev.to_vector()).c_str());
 
             embd.push_back(id);
 
@@ -789,7 +711,7 @@ int main(int argc, char ** argv) {
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], /* apply_grammar= */ false);
+                gpt_sampler_accept(smpl, embd_inp[n_consumed], /* apply_grammar= */ false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -832,7 +754,7 @@ int main(int argc, char ** argv) {
             // check for reverse prompt in the last n_prev tokens
             if (!params.antiprompt.empty()) {
                 const int n_prev = 32;
-                const std::string last_output = llama_sampling_prev_str(ctx_sampling, ctx, n_prev);
+                const std::string last_output = gpt_sampler_prev_str(smpl, ctx, n_prev);
 
                 is_antiprompt = false;
                 // Check if each of the reverse prompts appears at the end of the output.
@@ -854,7 +776,7 @@ int main(int argc, char ** argv) {
                 }
 
                 // check for reverse prompt using special tokens
-                llama_token last_token = llama_sampling_last(ctx_sampling);
+                llama_token last_token = gpt_sampler_last(smpl);
                 for (std::vector<llama_token> ids : antiprompt_ids) {
                     if (ids.size() == 1 && last_token == ids[0]) {
                         if (params.interactive) {
@@ -871,7 +793,7 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of generation tokens in interactive mode
-            if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
+            if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
                 LOG("found an EOG token\n");
 
                 if (params.interactive) {
@@ -892,7 +814,7 @@ int main(int argc, char ** argv) {
 
             // if current token is not EOG, we add it to current assistant message
             if (params.conversation) {
-                auto id = llama_sampling_last(ctx_sampling);
+                const auto id = gpt_sampler_last(smpl);
                 assistant_ss << llama_token_to_piece(ctx, id, false);
             }
 
@@ -988,7 +910,7 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    llama_sampling_reset(ctx_sampling);
+                    gpt_sampler_reset(smpl);
                 }
                 is_interacting = false;
             }
@@ -1013,14 +935,15 @@ int main(int argc, char ** argv) {
         llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    gpt_perf_print(ctx, smpl);
     write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
-    if (ctx_guidance) { llama_free(ctx_guidance); }
+    gpt_sampler_free(smpl);
+
     llama_free(ctx);
     llama_free_model(model);
 
-    llama_sampling_free(ctx_sampling);
     llama_backend_free();
 
     ggml_threadpool_free(threadpool);

examples/parallel/parallel.cpp

@@ -1,7 +1,9 @@
 // A basic application simulating a server with multiple clients.
 // The clients submit requests to the server and they are processed in parallel.
 
+#include "arg.h"
 #include "common.h"
+#include "sampling.h"
 #include "llama.h"
 
 #include <cmath>
@@ -50,8 +52,8 @@ static std::vector<std::string> k_prompts = {
 
 struct client {
     ~client() {
-        if (ctx_sampling) {
-            llama_sampling_free(ctx_sampling);
+        if (smpl) {
+            gpt_sampler_free(smpl);
         }
     }
 
@@ -72,7 +74,7 @@ struct client {
     std::string prompt;
     std::string response;
 
-    struct llama_sampling_context * ctx_sampling = nullptr;
+    struct gpt_sampler * smpl = nullptr;
 };
 
 static void print_date_time() {
@@ -100,8 +102,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
     }
 
@@ -161,7 +162,7 @@ int main(int argc, char ** argv) {
     for (size_t i = 0; i < clients.size(); ++i) {
         auto & client = clients[i];
         client.id = i;
-        client.ctx_sampling = llama_sampling_init(params.sparams);
+        client.smpl = gpt_sampler_init(model, params.sparams);
     }
 
     std::vector<llama_token> tokens_system;
@@ -253,7 +254,7 @@ int main(int argc, char ** argv) {
                     client.prompt   = client.input + "\nAssistant:";
                     client.response = "";
 
-                    llama_sampling_reset(client.ctx_sampling);
+                    gpt_sampler_reset(client.smpl);
 
                     // do not prepend BOS because we have a system prompt!
                     std::vector<llama_token> tokens_prompt;
@@ -341,9 +342,9 @@ int main(int argc, char ** argv) {
                 //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
                 //        client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
 
-                const llama_token id = llama_sampling_sample(client.ctx_sampling, ctx, NULL, client.i_batch - i);
+                const llama_token id = gpt_sampler_sample(client.smpl, ctx, client.i_batch - i);
 
-                llama_sampling_accept(client.ctx_sampling, ctx, id, true);
+                gpt_sampler_accept(client.smpl, id, true);
 
                 if (client.n_decoded == 1) {
                     // start measuring generation time after the first token to make sure all concurrent clients
@@ -371,7 +372,7 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1);
+                    llama_kv_cache_seq_rm(ctx,    client.id + 1, -1, -1);
                     llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
@@ -413,7 +414,8 @@ int main(int argc, char ** argv) {
 
     LOG_TEE("\n");
 
-    llama_print_timings(ctx);
+    // TODO: print sampling/grammar timings for all clients
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 

examples/passkey/passkey.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -6,9 +7,7 @@
 #include <string>
 #include <vector>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n", argv[0]);
     LOG_TEE("\n");
@@ -21,13 +20,10 @@ int main(int argc, char ** argv) {
     params.n_keep = 32;
     params.i_pos  = -1;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
         return 1;
     }
 
-    srand(params.seed == LLAMA_DEFAULT_SEED ? time(NULL) : params.seed);
-
     int n_junk = params.n_junk;
     int n_keep = params.n_keep;
     int n_grp  = params.grp_attn_n;
@@ -80,12 +76,17 @@ int main(int argc, char ** argv) {
     GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
-
     if (ctx == NULL) {
         fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
         return 1;
     }
 
+    auto sparams = llama_sampler_chain_default_params();
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
+
     // tokenize the prompt
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(ctx, params.prompt, true);
@@ -217,20 +218,7 @@ int main(int argc, char ** argv) {
     while (n_cur <= n_len) {
         // sample the next token
         {
-            auto   n_vocab = llama_n_vocab(model);
-            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
-
-            std::vector<llama_token_data> candidates;
-            candidates.reserve(n_vocab);
-
-            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-            }
-
-            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-            // sample the most likely token
-            const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
 
             // is it an end of generation?
             if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
@@ -267,10 +255,13 @@ int main(int argc, char ** argv) {
     LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
+    llama_sampler_free(smpl);
+
     llama_batch_free(batch);
 
     llama_free(ctx);

examples/perplexity/perplexity.cpp

@@ -1,18 +1,19 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
+#include <array>
+#include <atomic>
 #include <cmath>
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <fstream>
+#include <mutex>
+#include <random>
 #include <sstream>
 #include <thread>
-#include <mutex>
-#include <atomic>
 #include <vector>
-#include <array>
-#include <fstream>
-#include <sstream>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -76,7 +77,7 @@ static void write_logfile(
     fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
     yaml_dump_vector_float(logfile, "probs", results.probs);
 
-    llama_dump_timing_info_yaml(logfile, ctx);
+    llama_perf_dump_yaml(logfile, ctx);
     fclose(logfile);
 }
 
@@ -1967,8 +1968,7 @@ int main(int argc, char ** argv) {
     params.n_ctx = 512;
     params.logits_all = true;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
         return 1;
     }
 
@@ -2007,14 +2007,6 @@ int main(int argc, char ** argv) {
 
     print_build_info();
 
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
-
     llama_backend_init();
     llama_numa_init(params.numa);
 
@@ -2054,7 +2046,8 @@ int main(int argc, char ** argv) {
         results = perplexity(ctx, params, n_ctx);
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
     write_logfile(ctx, params, model, results);
 
     llama_free(ctx);

examples/quantize-stats/quantize-stats.cpp

@@ -1,7 +1,7 @@
-#define LLAMA_API_INTERNAL
 #include "common.h"
 #include "ggml.h"
 #include "llama.h"
+#include "llama-impl.h"
 
 #include <algorithm>
 #include <cassert>
@@ -319,8 +319,7 @@ int main(int argc, char ** argv) {
         }
 
         auto cparams = llama_context_default_params();
-        cparams.n_ctx      = 256;
-        cparams.seed       = 1;
+        cparams.n_ctx = 256;
 
         ctx = llama_new_context_with_model(model, cparams);
 

examples/quantize/CMakeLists.txt

@@ -1,6 +1,6 @@
 set(TARGET llama-quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/quantize/README.md

@@ -54,6 +54,8 @@ As the models are currently fully loaded into memory, you will need adequate dis
 
 Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
 
+The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interleaved variants of the `Q4_0` format, providing a data layout that is better suited for specific implementations of optimized mulmat kernels. Since these formats differ only in data layout, they have the same quantized size as the `Q4_0` format.
+
 *(outdated)*
 
 | Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |

examples/retrieval/retrieval.cpp

@@ -1,12 +1,11 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <algorithm>
 #include <fstream>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .\n", argv[0]);
     LOG_TEE("\n");
@@ -113,8 +112,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {
         return 1;
     }
 
@@ -293,9 +291,11 @@ int main(int argc, char ** argv) {
         }
     }
 
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
+
     // clean up
     llama_batch_free(query_batch);
-    llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);
     llama_backend_free();

examples/rpc/README.md

@@ -10,20 +10,21 @@ This can be used for distributed LLM inference with `llama.cpp` in the following
 
 ```mermaid
 flowchart TD
-    rpcb---|TCP|srva
-    rpcb---|TCP|srvb
-    rpcb-.-|TCP|srvn
+    rpcb<-->|TCP|srva
+    rpcb<-->|TCP|srvb
+    rpcb<-.->|TCP|srvn
     subgraph hostn[Host N]
-    srvn[rpc-server]-.-backend3["Backend (CUDA,Metal,etc.)"]
+    srvn[rpc-server]<-.->backend3["Backend (CUDA,Metal,etc.)"]
     end
     subgraph hostb[Host B]
-    srvb[rpc-server]---backend2["Backend (CUDA,Metal,etc.)"]
+    srvb[rpc-server]<-->backend2["Backend (CUDA,Metal,etc.)"]
     end
     subgraph hosta[Host A]
-    srva[rpc-server]---backend["Backend (CUDA,Metal,etc.)"]
+    srva[rpc-server]<-->backend["Backend (CUDA,Metal,etc.)"]
     end
     subgraph host[Main Host]
-    ggml[llama.cpp]---rpcb[RPC backend]
+    local["Backend (CUDA,Metal,etc.)"]<-->ggml[llama-cli]
+    ggml[llama-cli]<-->rpcb[RPC backend]
     end
     style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
 ```
@@ -62,17 +63,12 @@ $ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
 This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
 
 
-On the main host build `llama.cpp` only with `-DGGML_RPC=ON`:
-
-```bash
-mkdir build-rpc
-cd build-rpc
-cmake .. -DGGML_RPC=ON
-cmake --build . --config Release
-```
-
-Finally, use the `--rpc` option to specify the host and port of each `rpc-server`:
+On the main host build `llama.cpp` for the local backend and add `-DGGML_RPC=ON` to the build options.
+Finally, when running `llama-cli`, use the `--rpc` option to specify the host and port of each `rpc-server`:
 
 ```bash
 $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
 ```
+
+This way you can offload model layers to both local and remote devices.
+

examples/save-load-state/save-load-state.cpp

@@ -1,17 +1,17 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <vector>
 #include <cstdio>
-#include <chrono>
 
 int main(int argc, char ** argv) {
     gpt_params params;
 
     params.prompt = "The quick brown fox";
+    params.sparams.seed = 1234;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
@@ -38,6 +38,13 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    auto sparams = llama_sampler_chain_default_params();
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_softmax());
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
+
     // tokenize prompt
     auto tokens = llama_tokenize(ctx, params.prompt, true);
 
@@ -64,16 +71,7 @@ int main(int argc, char ** argv) {
     printf("\nfirst run: %s", params.prompt.c_str());
 
     for (auto i = 0; i < params.n_predict; i++) {
-        auto * logits = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(model);
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-        }
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-        auto next_token = llama_sample_token(ctx, &candidates_p);
+        auto next_token     = llama_sampler_sample(smpl, ctx, -1);
         auto next_token_str = llama_token_to_piece(ctx, next_token);
 
         printf("%s", next_token_str.c_str());
@@ -96,6 +94,11 @@ int main(int argc, char ** argv) {
     // make new context
     auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
 
+    llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl2, llama_sampler_init_softmax());
+    llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sparams.seed));
+
     printf("\nsecond run: %s", params.prompt.c_str());
 
     // load state (rng, logits, embedding and kv_cache) from file
@@ -124,15 +127,7 @@ int main(int argc, char ** argv) {
 
     // second run
     for (auto i = 0; i < params.n_predict; i++) {
-        auto * logits = llama_get_logits(ctx2);
-        auto n_vocab = llama_n_vocab(model);
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-        }
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-        auto next_token = llama_sample_token(ctx2, &candidates_p);
+        auto next_token     = llama_sampler_sample(smpl2, ctx2, -1);
         auto next_token_str = llama_token_to_piece(ctx2, next_token);
 
         printf("%s", next_token_str.c_str());
@@ -157,7 +152,12 @@ int main(int argc, char ** argv) {
     }
 
     // make new context
-    auto* ctx3 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
+    auto * ctx3 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
+
+    llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl3, llama_sampler_init_softmax());
+    llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sparams.seed));
 
     printf("\nsingle seq run: %s", params.prompt.c_str());
 
@@ -215,15 +215,7 @@ int main(int argc, char ** argv) {
 
     // third run with seq 1 instead of 0
     for (auto i = 0; i < params.n_predict; i++) {
-        auto * logits = llama_get_logits(ctx3);
-        auto n_vocab = llama_n_vocab(model);
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-        }
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-        auto next_token = llama_sample_token(ctx3, &candidates_p);
+        auto next_token     = llama_sampler_sample(smpl3, ctx3, -1);
         auto next_token_str = llama_token_to_piece(ctx3, next_token);
 
         printf("%s", next_token_str.c_str());
@@ -240,6 +232,10 @@ int main(int argc, char ** argv) {
 
     printf("\n");
 
+    llama_sampler_free(smpl);
+    llama_sampler_free(smpl2);
+    llama_sampler_free(smpl3);
+
     llama_free(ctx3);
     llama_free_model(model);
 

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

@@ -17,262 +17,126 @@ The project is under active development, and we are [looking for feedback and co
 
 ## Usage
 
-```
-usage: ./llama-server [options]
+| Argument | Explanation |
+| -------- | ----------- |
+| `-h, --help, --usage` | print usage and exit |
+| `--version` | show version and build info |
+| `-v, --verbose` | print verbose information |
+| `--verbosity N` | set specific verbosity level (default: 0) |
+| `-t, --threads N` | number of threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
+| `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
+| `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
+| `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask |
+| `--cpu-strict <0\|1>` | use strict CPU placement (default: 0)<br/> |
+| `--prio N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
+| `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50)<br/> |
+| `-Cb, --cpu-mask-batch M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask) |
+| `-Crb, --cpu-range-batch lo-hi` | ranges of CPUs for affinity. Complements --cpu-mask-batch |
+| `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) |
+| `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
+| `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
+| `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
+| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)<br/>(env: LLAMA_ARG_N_PREDICT) |
+| `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
+| `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
+| `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
+| `-fa, --flash-attn` | enable Flash Attention (default: disabled)<br/>(env: LLAMA_ARG_FLASH_ATTN) |
+| `-p, --prompt PROMPT` | prompt to start generation with |
+| `-f, --file FNAME` | a file containing the prompt (default: none) |
+| `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) |
+| `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
+| `--no-escape` | do not process escape sequences |
+| `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) |
+| `-s, --seed SEED` | RNG seed (default: -1, use random seed for < 0) |
+| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
+| `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
+| `--penalize-nl` | penalize newline tokens (default: false) |
+| `--temp N` | temperature (default: 0.8) |
+| `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
+| `--tfs N` | tail free sampling, parameter z (default: 1.0, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
+| `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
+| `--grammar-file FNAME` | file to read grammar from |
+| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model |
+| `--rope-scale N` | RoPE context scaling factor, expands context by a factor of N |
+| `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model) |
+| `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N |
+| `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0) |
+| `-gan, --grp-attn-n N` | group-attention factor (default: 1) |
+| `-gaw, --grp-attn-w N` | group-attention width (default: 512.0) |
+| `-dkvc, --dump-kv-cache` | verbose print of the KV cache |
+| `-nkvo, --no-kv-offload` | disable KV offload |
+| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16) |
+| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16) |
+| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
+| `-np, --parallel N` | number of parallel sequences to decode (default: 1) |
+| `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
+| `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
+| `--mlock` | force system to keep model in RAM rather than swapping or compressing |
+| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock) |
+| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggerganov/llama.cpp/issues/1437 |
+| `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
+| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs |
+| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1 |
+| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0) |
+| `--check-tensors` | check model tensor data for invalid values (default: false) |
+| `--override-kv KEY=TYPE:VALUE` | advanced option to override model metadata by key. may be specified multiple times.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false |
+| `--lora FNAME` | path to LoRA adapter (can be repeated to use multiple adapters) |
+| `--lora-scaled FNAME SCALE` | path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters) |
+| `--control-vector FNAME` | add a control vector<br/>note: this argument can be repeated to add multiple control vectors |
+| `--control-vector-scaled FNAME SCALE` | add a control vector with user defined scaling SCALE<br/>note: this argument can be repeated to add multiple scaled control vectors |
+| `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive |
+| `-a, --alias STRING` | set alias for model name (to be used by REST API) |
+| `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)<br/>(env: LLAMA_ARG_MODEL) |
+| `-mu, --model-url MODEL_URL` | model download url (default: unused)<br/>(env: LLAMA_ARG_MODEL_URL) |
+| `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
+| `-hff, --hf-file FILE` | Hugging Face model file (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
+| `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
+| `--host HOST` | ip address to listen (default: 127.0.0.1)<br/>(env: LLAMA_ARG_HOST) |
+| `--port PORT` | port to listen (default: 8080)<br/>(env: LLAMA_ARG_PORT) |
+| `--path PATH` | path to serve static files from (default: ) |
+| `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
+| `--api-key KEY` | API key to use for authentication (default: none)<br/>(env: LLAMA_API_KEY) |
+| `--api-key-file FNAME` | path to file containing API keys (default: none) |
+| `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key |
+| `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate |
+| `-to, --timeout N` | server read/write timeout in seconds (default: 600) |
+| `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
+| `-spf, --system-prompt-file FNAME` | set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications |
+| `--log-format {text, json}` | log output format: json or text (default: json) |
+| `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
+| `--no-slots` | disables slots monitoring endpoint (default: enabled)<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
+| `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted:<br/>https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
+| `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
+| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) |
+| `--log-test` | Log test |
+| `--log-disable` | Log disable |
+| `--log-enable` | Log enable |
+| `--log-new` | Log new |
+| `--log-append` | Log append |
+| `--log-file FNAME` | Log file |
 
-general:
-
-  -h,    --help, --usage          print usage and exit
-         --version                show version and build info
-  -v,    --verbose                print verbose information
-         --verbosity N            set specific verbosity level (default: 0)
-         --verbose-prompt         print a verbose prompt before generation (default: false)
-         --no-display-prompt      don't print prompt at generation (default: false)
-  -co,   --color                  colorise output to distinguish prompt and user input from generations (default: false)
-  -s,    --seed SEED              RNG seed (default: -1, use random seed for < 0)
-  -t,    --threads N              number of threads to use during generation (default: 8)
-  -tb,   --threads-batch N        number of threads to use during batch and prompt processing (default: same as --threads)
-  -td,   --threads-draft N        number of threads to use during generation (default: same as --threads)
-  -tbd,  --threads-batch-draft N  number of threads to use during batch and prompt processing (default: same as --threads-draft)
-         --draft N                number of tokens to draft for speculative decoding (default: 5)
-  -ps,   --p-split N              speculative decoding split probability (default: 0.1)
-  -lcs,  --lookup-cache-static FNAME
-                                  path to static lookup cache to use for lookup decoding (not updated by generation)
-  -lcd,  --lookup-cache-dynamic FNAME
-                                  path to dynamic lookup cache to use for lookup decoding (updated by generation)
-  -c,    --ctx-size N             size of the prompt context (default: 0, 0 = loaded from model)
-  -n,    --predict N              number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)
-  -b,    --batch-size N           logical maximum batch size (default: 2048)
-  -ub,   --ubatch-size N          physical maximum batch size (default: 512)
-         --keep N                 number of tokens to keep from the initial prompt (default: 0, -1 = all)
-         --chunks N               max number of chunks to process (default: -1, -1 = all)
-  -fa,   --flash-attn             enable Flash Attention (default: disabled)
-  -p,    --prompt PROMPT          prompt to start generation with
-                                  in conversation mode, this will be used as system prompt
-                                  (default: '')
-  -f,    --file FNAME             a file containing the prompt (default: none)
-         --in-file FNAME          an input file (repeat to specify multiple files)
-  -bf,   --binary-file FNAME      binary file containing the prompt (default: none)
-  -e,    --escape                 process escapes sequences (\n, \r, \t, \', \", \\) (default: true)
-         --no-escape              do not process escape sequences
-  -ptc,  --print-token-count N    print token count every N tokens (default: -1)
-         --prompt-cache FNAME     file to cache prompt state for faster startup (default: none)
-         --prompt-cache-all       if specified, saves user input and generations to cache as well
-                                  not supported with --interactive or other interactive options
-         --prompt-cache-ro        if specified, uses the prompt cache but does not update it
-  -r,    --reverse-prompt PROMPT  halt generation at PROMPT, return control in interactive mode
-                                  can be specified more than once for multiple prompts
-  -sp,   --special                special tokens output enabled (default: false)
-  -cnv,  --conversation           run in conversation mode, does not print special tokens and suffix/prefix
-                                  if suffix/prefix are not specified, default chat template will be used
-                                  (default: false)
-  -i,    --interactive            run in interactive mode (default: false)
-  -if,   --interactive-first      run in interactive mode and wait for input right away (default: false)
-  -mli,  --multiline-input        allows you to write or paste multiple lines without ending each in '\'
-         --in-prefix-bos          prefix BOS to user inputs, preceding the `--in-prefix` string
-         --in-prefix STRING       string to prefix user inputs with (default: empty)
-         --in-suffix STRING       string to suffix after user inputs with (default: empty)
-         --spm-infill             use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled)
-
-sampling:
-
-         --samplers SAMPLERS      samplers that will be used for generation in the order, separated by ';'
-                                  (default: top_k;tfs_z;typical_p;top_p;min_p;temperature)
-         --sampling-seq SEQUENCE  simplified sequence for samplers that will be used (default: kfypmt)
-         --ignore-eos             ignore end of stream token and continue generating (implies --logit-bias EOS-inf)
-         --penalize-nl            penalize newline tokens (default: false)
-         --temp N                 temperature (default: 0.8)
-         --top-k N                top-k sampling (default: 40, 0 = disabled)
-         --top-p N                top-p sampling (default: 0.9, 1.0 = disabled)
-         --min-p N                min-p sampling (default: 0.1, 0.0 = disabled)
-         --tfs N                  tail free sampling, parameter z (default: 1.0, 1.0 = disabled)
-         --typical N              locally typical sampling, parameter p (default: 1.0, 1.0 = disabled)
-         --repeat-last-n N        last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size)
-         --repeat-penalty N       penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled)
-         --presence-penalty N     repeat alpha presence penalty (default: 0.0, 0.0 = disabled)
-         --frequency-penalty N    repeat alpha frequency penalty (default: 0.0, 0.0 = disabled)
-         --dynatemp-range N       dynamic temperature range (default: 0.0, 0.0 = disabled)
-         --dynatemp-exp N         dynamic temperature exponent (default: 1.0)
-         --mirostat N             use Mirostat sampling.
-                                  Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.
-                                  (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)
-         --mirostat-lr N          Mirostat learning rate, parameter eta (default: 0.1)
-         --mirostat-ent N         Mirostat target entropy, parameter tau (default: 5.0)
-         -l TOKEN_ID(+/-)BIAS     modifies the likelihood of token appearing in the completion,
-                                  i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',
-                                  or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'
-         --cfg-negative-prompt PROMPT
-                                  negative prompt to use for guidance (default: '')
-         --cfg-negative-prompt-file FNAME
-                                  negative prompt file to use for guidance
-         --cfg-scale N            strength of guidance (default: 1.0, 1.0 = disable)
-         --chat-template JINJA_TEMPLATE
-                                  set custom jinja chat template (default: template taken from model's metadata)
-                                  if suffix/prefix are specified, template will be disabled
-                                  only commonly used templates are accepted:
-                                  https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
-
-grammar:
-
-         --grammar GRAMMAR        BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '')
-         --grammar-file FNAME     file to read grammar from
-  -j,    --json-schema SCHEMA     JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object
-                                  For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead
-
-embedding:
-
-         --pooling {none,mean,cls,last}
-                                  pooling type for embeddings, use model default if unspecified
-         --attention {causal,non-causal}
-                                  attention type for embeddings, use model default if unspecified
-
-context hacking:
-
-         --rope-scaling {none,linear,yarn}
-                                  RoPE frequency scaling method, defaults to linear unless specified by the model
-         --rope-scale N           RoPE context scaling factor, expands context by a factor of N
-         --rope-freq-base N       RoPE base frequency, used by NTK-aware scaling (default: loaded from model)
-         --rope-freq-scale N      RoPE frequency scaling factor, expands context by a factor of 1/N
-         --yarn-orig-ctx N        YaRN: original context size of model (default: 0 = model training context size)
-         --yarn-ext-factor N      YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)
-         --yarn-attn-factor N     YaRN: scale sqrt(t) or attention magnitude (default: 1.0)
-         --yarn-beta-slow N       YaRN: high correction dim or alpha (default: 1.0)
-         --yarn-beta-fast N       YaRN: low correction dim or beta (default: 32.0)
-  -gan,  --grp-attn-n N           group-attention factor (default: 1)
-  -gaw,  --grp-attn-w N           group-attention width (default: 512.0)
-  -dkvc, --dump-kv-cache          verbose print of the KV cache
-  -nkvo, --no-kv-offload          disable KV offload
-  -ctk,  --cache-type-k TYPE      KV cache data type for K (default: f16)
-  -ctv,  --cache-type-v TYPE      KV cache data type for V (default: f16)
-
-perplexity:
-
-         --all-logits             return logits for all tokens in the batch (default: false)
-         --hellaswag              compute HellaSwag score over random tasks from datafile supplied with -f
-         --hellaswag-tasks N      number of tasks to use when computing the HellaSwag score (default: 400)
-         --winogrande             compute Winogrande score over random tasks from datafile supplied with -f
-         --winogrande-tasks N     number of tasks to use when computing the Winogrande score (default: 0)
-         --multiple-choice        compute multiple choice score over random tasks from datafile supplied with -f
-         --multiple-choice-tasks N
-                                  number of tasks to use when computing the multiple choice score (default: 0)
-         --kl-divergence          computes KL-divergence to logits provided via --kl-divergence-base
-         --ppl-stride N           stride for perplexity calculation (default: 0)
-         --ppl-output-type {0,1}  output type for perplexity calculation (default: 0)
-
-parallel:
-
-  -dt,   --defrag-thold N         KV cache defragmentation threshold (default: -1.0, < 0 - disabled)
-  -np,   --parallel N             number of parallel sequences to decode (default: 1)
-  -ns,   --sequences N            number of sequences to decode (default: 1)
-  -cb,   --cont-batching          enable continuous batching (a.k.a dynamic batching) (default: enabled)
-
-multi-modality:
-
-         --mmproj FILE            path to a multimodal projector file for LLaVA. see examples/llava/README.md
-         --image FILE             path to an image file. use with multimodal models. Specify multiple times for batching
-
-backend:
-
-         --rpc SERVERS            comma separated list of RPC servers
-         --mlock                  force system to keep model in RAM rather than swapping or compressing
-         --no-mmap                do not memory-map model (slower load but may reduce pageouts if not using mlock)
-         --numa TYPE              attempt optimizations that help on some NUMA systems
-                                    - distribute: spread execution evenly over all nodes
-                                    - isolate: only spawn threads on CPUs on the node that execution started on
-                                    - numactl: use the CPU map provided by numactl
-                                  if run without this previously, it is recommended to drop the system page cache before using this
-                                  see https://github.com/ggerganov/llama.cpp/issues/1437
-
-model:
-
-         --check-tensors          check model tensor data for invalid values (default: false)
-         --override-kv KEY=TYPE:VALUE
-                                  advanced option to override model metadata by key. may be specified multiple times.
-                                  types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false
-         --lora FNAME             apply LoRA adapter (implies --no-mmap)
-         --lora-scaled FNAME S    apply LoRA adapter with user defined scaling S (implies --no-mmap)
-         --lora-base FNAME        optional model to use as a base for the layers modified by the LoRA adapter
-         --control-vector FNAME   add a control vector
-                                  note: this argument can be repeated to add multiple control vectors
-         --control-vector-scaled FNAME SCALE
-                                  add a control vector with user defined scaling SCALE
-                                  note: this argument can be repeated to add multiple scaled control vectors
-         --control-vector-layer-range START END
-                                  layer range to apply the control vector(s) to, start and end inclusive
-  -m,    --model FNAME            model path (default: models/$filename with filename from --hf-file
-                                  or --model-url if set, otherwise models/7B/ggml-model-f16.gguf)
-  -md,   --model-draft FNAME      draft model for speculative decoding (default: unused)
-  -mu,   --model-url MODEL_URL    model download url (default: unused)
-  -hfr,  --hf-repo REPO           Hugging Face model repository (default: unused)
-  -hff,  --hf-file FILE           Hugging Face model file (default: unused)
-  -hft,  --hf-token TOKEN         Hugging Face access token (default: value from HF_TOKEN environment variable)
-
-server:
-
-         --host HOST              ip address to listen (default: 127.0.0.1)
-         --port PORT              port to listen (default: 8080)
-         --path PATH              path to serve static files from (default: )
-         --embedding(s)           restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
-         --api-key KEY            API key to use for authentication (default: none)
-         --api-key-file FNAME     path to file containing API keys (default: none)
-         --ssl-key-file FNAME     path to file a PEM-encoded SSL private key
-         --ssl-cert-file FNAME    path to file a PEM-encoded SSL certificate
-         --timeout N              server read/write timeout in seconds (default: 600)
-         --threads-http N         number of threads used to process HTTP requests (default: -1)
-         --system-prompt-file FNAME
-                                  set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications
-         --log-format {text,json}
-                                  log output format: json or text (default: json)
-         --metrics                enable prometheus compatible metrics endpoint (default: disabled)
-         --no-slots               disables slots monitoring endpoint (default: enabled)
-         --slot-save-path PATH    path to save slot kv cache (default: disabled)
-         --chat-template JINJA_TEMPLATE
-                                  set custom jinja chat template (default: template taken from model's metadata)
-                                  only commonly used templates are accepted:
-                                  https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
-  -sps,  --slot-prompt-similarity SIMILARITY
-                                  how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
-         --lora-init-without-apply
-                                  load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)
-
-logging:
-
-         --simple-io              use basic IO for better compatibility in subprocesses and limited consoles
-  -ld,   --logdir LOGDIR          path under which to save YAML logs (no logging if unset)
-         --log-test               Run simple logging test
-         --log-disable            Disable trace logs
-         --log-enable             Enable trace logs
-         --log-file FNAME         Specify a log filename (without extension)
-         --log-new                Create a separate new log file on start. Each log file will have unique name: "<name>.<ID>.log"
-         --log-append             Don't truncate the old log file.
-```
-
-Available environment variables (if specified, these variables will override parameters specified in arguments):
-
-- `LLAMA_CACHE`: cache directory, used by `--hf-repo`
-- `HF_TOKEN`: Hugging Face access token, used when accessing a gated model with `--hf-repo`
-- `LLAMA_ARG_MODEL`: equivalent to `-m`
-- `LLAMA_ARG_MODEL_URL`: equivalent to `-mu`
-- `LLAMA_ARG_MODEL_ALIAS`: equivalent to `-a`
-- `LLAMA_ARG_HF_REPO`: equivalent to `--hf-repo`
-- `LLAMA_ARG_HF_FILE`: equivalent to `--hf-file`
-- `LLAMA_ARG_THREADS`: equivalent to `-t`
-- `LLAMA_ARG_CTX_SIZE`: equivalent to `-c`
-- `LLAMA_ARG_N_PARALLEL`: equivalent to `-np`
-- `LLAMA_ARG_BATCH`: equivalent to `-b`
-- `LLAMA_ARG_UBATCH`: equivalent to `-ub`
-- `LLAMA_ARG_N_GPU_LAYERS`: equivalent to `-ngl`
-- `LLAMA_ARG_THREADS_HTTP`: equivalent to `--threads-http`
-- `LLAMA_ARG_CHAT_TEMPLATE`: equivalent to `--chat-template`
-- `LLAMA_ARG_N_PREDICT`: equivalent to `-n`
-- `LLAMA_ARG_ENDPOINT_METRICS`: if set to `1`, it will enable metrics endpoint (equivalent to `--metrics`)
-- `LLAMA_ARG_ENDPOINT_SLOTS`: if set to `0`, it will **disable** slots endpoint (equivalent to `--no-slots`). This feature is enabled by default.
-- `LLAMA_ARG_EMBEDDINGS`: if set to `1`, it will enable embeddings endpoint (equivalent to `--embeddings`)
-- `LLAMA_ARG_FLASH_ATTN`: if set to `1`, it will enable flash attention (equivalent to `-fa`)
-- `LLAMA_ARG_CONT_BATCHING`: if set to `0`, it will **disable** continuous batching (equivalent to `--no-cont-batching`). This feature is enabled by default.
-- `LLAMA_ARG_DEFRAG_THOLD`: equivalent to `-dt`
-- `LLAMA_ARG_HOST`: equivalent to `--host`
-- `LLAMA_ARG_PORT`: equivalent to `--port`
+Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
 
 Example usage of docker compose with environment variables:
 
@@ -289,7 +153,7 @@ services:
       LLAMA_ARG_MODEL: /models/my_model.gguf
       LLAMA_ARG_CTX_SIZE: 4096
       LLAMA_ARG_N_PARALLEL: 2
-      LLAMA_ARG_ENDPOINT_METRICS: 1  # to disable, either remove or set to 0
+      LLAMA_ARG_ENDPOINT_METRICS: 1
       LLAMA_ARG_PORT: 8080
 ```
 
@@ -470,8 +334,6 @@ node index.js
 
     `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
 
-    `penalty_prompt`: This will replace the `prompt` for the purpose of the penalty evaluation. Can be either `null`, a string or an array of numbers representing tokens. Default: `null`, which is to use the original `prompt`.
-
     `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
 
     `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
@@ -545,9 +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
 
@@ -724,7 +621,6 @@ Example:
             "stopping_word": ""
         },
         "penalize_nl": true,
-        "penalty_prompt_tokens": [],
         "presence_penalty": 0.0,
         "prompt": "Say hello to llama.cpp",
         "repeat_last_n": 64,
@@ -748,8 +644,7 @@ Example:
         "tfs_z": 1.0,
         "top_k": 40,
         "top_p": 0.949999988079071,
-        "typical_p": 1.0,
-        "use_penalty_prompt_tokens": false
+        "typical_p": 1.0
     }
 ]
 ```

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

@@ -1,9 +1,10 @@
 #include "utils.hpp"
 
+#include "arg.h"
 #include "common.h"
+#include "sampling.h"
 #include "json-schema-to-grammar.h"
 #include "llama.h"
-#include "grammar-parser.h"
 
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
@@ -27,6 +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>
@@ -50,15 +52,12 @@ enum stop_type {
     STOP_TYPE_PARTIAL,
 };
 
+// state diagram: https://github.com/ggerganov/llama.cpp/pull/9283
 enum slot_state {
     SLOT_STATE_IDLE,
-    SLOT_STATE_PROCESSING,
-};
-
-enum slot_command {
-    SLOT_COMMAND_NONE,
-    SLOT_COMMAND_LOAD_PROMPT,
-    SLOT_COMMAND_RELEASE,
+    SLOT_STATE_PROCESSING_PROMPT,
+    SLOT_STATE_DONE_PROMPT,
+    SLOT_STATE_GENERATING,
 };
 
 enum server_state {
@@ -135,7 +134,6 @@ struct server_slot {
     struct slot_params params;
 
     slot_state state = SLOT_STATE_IDLE;
-    slot_command command = SLOT_COMMAND_NONE;
 
     // used to determine the slot that has been used the longest
     int64_t t_last_used = -1;
@@ -173,11 +171,13 @@ struct server_slot {
     std::string stopping_word;
 
     // sampling
-    llama_token sampled;
-    struct llama_sampling_params sparams;
-    llama_sampling_context * ctx_sampling = nullptr;
     json json_schema;
 
+    struct gpt_sampler_params sparams;
+    struct gpt_sampler * smpl = nullptr;
+
+    llama_token sampled;
+
     int32_t ga_i = 0;   // group-attention state
     int32_t ga_n = 1;   // group-attention factor
     int32_t ga_w = 512; // group-attention width
@@ -194,6 +194,8 @@ struct server_slot {
     double t_prompt_processing; // ms
     double t_token_generation; // ms
 
+    std::function<void(int)> callback_on_release;
+
     void reset() {
         n_prompt_tokens    = 0;
         generated_text     = "";
@@ -228,25 +230,28 @@ struct server_slot {
         return n_remaining > 0; // no budget
     }
 
-    bool available() const {
-        return state == SLOT_STATE_IDLE && command == SLOT_COMMAND_NONE;
-    }
-
     bool is_processing() const {
-        return (state == SLOT_STATE_IDLE && command == SLOT_COMMAND_LOAD_PROMPT) || state == SLOT_STATE_PROCESSING;
+        return state != SLOT_STATE_IDLE;
     }
 
     void add_token_string(const completion_token_output & token) {
-        if (command == SLOT_COMMAND_RELEASE) {
+        if (!is_processing()) {
             return;
         }
         generated_token_probs.push_back(token);
     }
 
     void release() {
-        if (state == SLOT_STATE_PROCESSING) {
+        if (is_processing()) {
             t_token_generation = (ggml_time_us() - t_start_generation) / 1e3;
-            command = SLOT_COMMAND_RELEASE;
+            state = SLOT_STATE_IDLE;
+            LOG_INFO("slot released", {
+                {"id_slot",   id},
+                {"id_task",   id_task},
+                {"n_past",    n_past},
+                {"truncated", truncated},
+            });
+            callback_on_release(id);
         }
     }
 
@@ -353,6 +358,9 @@ struct server_metrics {
     uint64_t n_tokens_predicted  = 0;
     uint64_t t_tokens_generation = 0;
 
+    uint64_t n_decode_total     = 0;
+    uint64_t n_busy_slots_total = 0;
+
     void init() {
         t_start = ggml_time_us();
     }
@@ -371,6 +379,15 @@ struct server_metrics {
         t_tokens_generation_total  += slot.t_token_generation;
     }
 
+    void on_decoded(const std::vector<server_slot> & slots) {
+        n_decode_total++;
+        for (const auto & slot : slots) {
+            if (slot.is_processing()) {
+                n_busy_slots_total++;
+            }
+        }
+    }
+
     void reset_bucket() {
         n_prompt_tokens_processed = 0;
         t_prompt_processing       = 0;
@@ -432,6 +449,7 @@ struct server_queue {
     void defer(server_task task) {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         queue_tasks_deferred.push_back(std::move(task));
+        condition_tasks.notify_one();
     }
 
     // Get the next id for creating a new task
@@ -452,14 +470,14 @@ struct server_queue {
         callback_update_slots = std::move(callback);
     }
 
-    // Call when the state of one slot is changed
-    void notify_slot_changed() {
-        // move deferred tasks back to main loop
+    // Call when the state of one slot is changed, it will move one task from deferred to main queue
+    void pop_deferred_task() {
         std::unique_lock<std::mutex> lock(mutex_tasks);
-        for (auto & task : queue_tasks_deferred) {
-            queue_tasks.push_back(std::move(task));
+        if (!queue_tasks_deferred.empty()) {
+            queue_tasks.emplace_back(std::move(queue_tasks_deferred.front()));
+            queue_tasks_deferred.pop_front();
         }
-        queue_tasks_deferred.clear();
+        condition_tasks.notify_one();
     }
 
     // end the start_loop routine
@@ -489,7 +507,7 @@ struct server_queue {
                     break;
                 }
                 server_task task = queue_tasks.front();
-                queue_tasks.erase(queue_tasks.begin());
+                queue_tasks.pop_front();
                 lock.unlock();
                 LOG_VERBOSE("callback_new_task", {{"id_task", task.id}});
                 callback_new_task(task);
@@ -598,7 +616,7 @@ struct server_context {
 
     gpt_params params;
 
-    llama_batch batch;
+    llama_batch batch = {};
 
     bool clean_kv_cache = true;
     bool add_bos_token  = true;
@@ -637,8 +655,8 @@ struct server_context {
 
         // Clear any sampling context
         for (server_slot & slot : slots) {
-            if (slot.ctx_sampling != nullptr) {
-                llama_sampling_free(slot.ctx_sampling);
+            if (slot.smpl != nullptr) {
+                gpt_sampler_free(slot.smpl);
             }
         }
 
@@ -717,6 +735,10 @@ struct server_context {
 
             slot.sparams = params.sparams;
 
+            slot.callback_on_release = [this](int) {
+                queue_tasks.pop_deferred_task();
+            };
+
             slot.reset();
 
             slots.push_back(slot);
@@ -798,7 +820,7 @@ struct server_context {
 
             for (server_slot & slot : slots) {
                 // skip the slot if it is not available
-                if (!slot.available()) {
+                if (slot.is_processing()) {
                     continue;
                 }
 
@@ -840,7 +862,7 @@ struct server_context {
             int64_t t_last = ggml_time_us();
             for (server_slot & slot : slots) {
                 // skip the slot if it is not available
-                if (!slot.available()) {
+                if (slot.is_processing()) {
                     continue;
                 }
 
@@ -865,8 +887,8 @@ struct server_context {
     bool launch_slot_with_task(server_slot & slot, const server_task & task) {
         slot_params default_params;
         // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them)
-        llama_sampling_params default_sparams = params.sparams;
-        auto & data = task.data;
+        auto default_sparams = params.sparams;
+        const auto & data = task.data;
 
         if (data.count("__oaicompat") != 0) {
             slot.oaicompat = true;
@@ -883,7 +905,7 @@ struct server_context {
         slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
         slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
         slot.sparams.tfs_z             = json_value(data, "tfs_z",             default_sparams.tfs_z);
-        slot.sparams.typical_p         = json_value(data, "typical_p",         default_sparams.typical_p);
+        slot.sparams.typ_p             = json_value(data, "typical_p",         default_sparams.typ_p);
         slot.sparams.temp              = json_value(data, "temperature",       default_sparams.temp);
         slot.sparams.dynatemp_range    = json_value(data, "dynatemp_range",    default_sparams.dynatemp_range);
         slot.sparams.dynatemp_exponent = json_value(data, "dynatemp_exponent", default_sparams.dynatemp_exponent);
@@ -905,7 +927,8 @@ struct server_context {
         if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) {
             send_error(task, "Either \"json_schema\" or \"grammar\" can be specified, but not both", ERROR_TYPE_INVALID_REQUEST);
             return false;
-        } else if (data.contains("json_schema") && !data.contains("grammar")) {
+        }
+        if (data.contains("json_schema") && !data.contains("grammar")) {
             try {
                 auto schema                = json_value(data, "json_schema", json::object());
                 slot.sparams.grammar       = json_schema_to_grammar(schema);
@@ -955,56 +978,11 @@ struct server_context {
             }
         }
 
-        // penalize user-provided tokens
-        {
-            slot.sparams.penalty_prompt_tokens.clear();
-            slot.sparams.use_penalty_prompt_tokens = false;
-
-            const auto & penalty_prompt = data.find("penalty_prompt");
-
-            if (penalty_prompt != data.end()) {
-                if (penalty_prompt->is_string()) {
-                    const auto penalty_prompt_string = penalty_prompt->get<std::string>();
-                    slot.sparams.penalty_prompt_tokens = llama_tokenize(model, penalty_prompt_string, false);
-
-                    if (slot.params.n_predict > 0) {
-                        slot.sparams.penalty_prompt_tokens.reserve(slot.sparams.penalty_prompt_tokens.size() + slot.params.n_predict);
-                    }
-                    slot.sparams.use_penalty_prompt_tokens = true;
-
-                    LOG_VERBOSE("penalty_prompt_tokens", {
-                        {"id_slot", slot.id},
-                        {"tokens",  slot.sparams.penalty_prompt_tokens},
-                    });
-                }
-                else if (penalty_prompt->is_array()) {
-                    const auto n_tokens = penalty_prompt->size();
-                    slot.sparams.penalty_prompt_tokens.reserve(n_tokens + std::max(0, slot.params.n_predict));
-
-                    const int n_vocab = llama_n_vocab(model);
-                    for (const auto & penalty_token : *penalty_prompt) {
-                        if (penalty_token.is_number_integer()) {
-                            const auto tok = penalty_token.get<llama_token>();
-                            if (tok >= 0 && tok < n_vocab) {
-                                slot.sparams.penalty_prompt_tokens.push_back(tok);
-                            }
-                        }
-                    }
-                    slot.sparams.use_penalty_prompt_tokens = true;
-
-                    LOG_VERBOSE("penalty_prompt_tokens", {
-                        {"id_slot", slot.id},
-                        {"tokens",  slot.sparams.penalty_prompt_tokens},
-                    });
-                }
-            }
-        }
-
         {
             slot.sparams.logit_bias.clear();
 
             if (json_value(data, "ignore_eos", false) && has_eos_token) {
-                slot.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
+                slot.sparams.logit_bias.push_back({llama_token_eos(model), -INFINITY});
             }
 
             const auto & logit_bias = data.find("logit_bias");
@@ -1025,12 +1003,12 @@ struct server_context {
                         if (el[0].is_number_integer()) {
                             llama_token tok = el[0].get<llama_token>();
                             if (tok >= 0 && tok < n_vocab) {
-                                slot.sparams.logit_bias[tok] = bias;
+                                slot.sparams.logit_bias.push_back({tok, bias});
                             }
                         } else if (el[0].is_string()) {
                             auto toks = llama_tokenize(model, el[0].get<std::string>(), false);
                             for (auto tok : toks) {
-                                slot.sparams.logit_bias[tok] = bias;
+                                slot.sparams.logit_bias.push_back({tok, bias});
                             }
                         }
                     }
@@ -1052,33 +1030,34 @@ struct server_context {
         }
 
         {
-            const auto & samplers_sequence = data.find("samplers");
-            if (samplers_sequence != data.end() && samplers_sequence->is_array()) {
+            const auto & samplers = data.find("samplers");
+            if (samplers != data.end() && samplers->is_array()) {
                 std::vector<std::string> sampler_names;
-                for (const auto & sampler_name : *samplers_sequence) {
-                    if (sampler_name.is_string()) {
-                        sampler_names.emplace_back(sampler_name);
+                for (const auto & name : *samplers) {
+                    if (name.is_string()) {
+                        sampler_names.emplace_back(name);
                     }
                 }
-                slot.sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, false);
+                slot.sparams.samplers = gpt_sampler_types_from_names(sampler_names, false);
             } else {
-                slot.sparams.samplers_sequence = default_sparams.samplers_sequence;
+                slot.sparams.samplers = default_sparams.samplers;
             }
         }
 
         {
-            if (slot.ctx_sampling != nullptr) {
-                llama_sampling_free(slot.ctx_sampling);
+            if (slot.smpl != nullptr) {
+                gpt_sampler_free(slot.smpl);
             }
-            slot.ctx_sampling = llama_sampling_init(slot.sparams);
-            if (slot.ctx_sampling == nullptr) {
+
+            slot.smpl = gpt_sampler_init(model, slot.sparams);
+            if (slot.smpl == nullptr) {
                 // for now, the only error that may happen here is invalid grammar
                 send_error(task, "Failed to parse grammar", ERROR_TYPE_INVALID_REQUEST);
                 return false;
             }
         }
 
-        slot.command = SLOT_COMMAND_LOAD_PROMPT;
+        slot.state = SLOT_STATE_PROCESSING_PROMPT;
         slot.prompt_tokens.clear();
 
         LOG_INFO("slot is processing task", {
@@ -1160,11 +1139,6 @@ struct server_context {
         slot.generated_text += token_str;
         slot.has_next_token = true;
 
-        if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1) {
-            // we can change penalty_prompt_tokens because it is always created from scratch each request
-            slot.ctx_sampling->params.penalty_prompt_tokens.push_back(result.tok);
-        }
-
         // check if there is incomplete UTF-8 character at the end
         bool incomplete = false;
         for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i) {
@@ -1282,13 +1256,10 @@ struct server_context {
     }
 
     json get_formated_generation(const server_slot & slot) const {
-        const auto eos_bias   =             slot.sparams.logit_bias.find(llama_token_eos(model));
-        const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() && eos_bias->second < 0.0f && std::isinf(eos_bias->second);
-
-        std::vector<std::string> samplers_sequence;
-        samplers_sequence.reserve(slot.sparams.samplers_sequence.size());
-        for (const auto & sampler_type : slot.sparams.samplers_sequence) {
-            samplers_sequence.emplace_back(llama_sampling_type_to_str(sampler_type));
+        std::vector<std::string> samplers;
+        samplers.reserve(slot.sparams.samplers.size());
+        for (const auto & sampler : slot.sparams.samplers) {
+            samplers.emplace_back(gpt_sampler_type_to_str(sampler));
         }
 
         return json {
@@ -1296,6 +1267,7 @@ struct server_context {
             {"n_predict",                 slot.n_predict},     // Server configured n_predict
             {"model",                     params.model_alias},
             {"seed",                      slot.sparams.seed},
+            {"seed_cur",                  slot.smpl ? gpt_sampler_get_seed(slot.smpl) : 0},
             {"temperature",               slot.sparams.temp},
             {"dynatemp_range",            slot.sparams.dynatemp_range},
             {"dynatemp_exponent",         slot.sparams.dynatemp_exponent},
@@ -1303,13 +1275,11 @@ struct server_context {
             {"top_p",                     slot.sparams.top_p},
             {"min_p",                     slot.sparams.min_p},
             {"tfs_z",                     slot.sparams.tfs_z},
-            {"typical_p",                 slot.sparams.typical_p},
+            {"typical_p",                 slot.sparams.typ_p},
             {"repeat_last_n",             slot.sparams.penalty_last_n},
             {"repeat_penalty",            slot.sparams.penalty_repeat},
             {"presence_penalty",          slot.sparams.penalty_present},
             {"frequency_penalty",         slot.sparams.penalty_freq},
-            {"penalty_prompt_tokens",     slot.sparams.penalty_prompt_tokens},
-            {"use_penalty_prompt_tokens", slot.sparams.use_penalty_prompt_tokens},
             {"mirostat",                  slot.sparams.mirostat},
             {"mirostat_tau",              slot.sparams.mirostat_tau},
             {"mirostat_eta",              slot.sparams.mirostat_eta},
@@ -1318,13 +1288,13 @@ struct server_context {
             {"max_tokens",                slot.params.n_predict}, // User configured n_predict
             {"n_keep",                    slot.params.n_keep},
             {"n_discard",                 slot.params.n_discard},
-            {"ignore_eos",                ignore_eos},
+            {"ignore_eos",                slot.sparams.ignore_eos},
             {"stream",                    slot.params.stream},
-            {"logit_bias",                slot.sparams.logit_bias},
+          //{"logit_bias",                slot.sparams.logit_bias},
             {"n_probs",                   slot.sparams.n_probs},
             {"min_keep",                  slot.sparams.min_keep},
             {"grammar",                   slot.sparams.grammar},
-            {"samplers",                  samplers_sequence}
+            {"samplers",                  samplers},
         };
     }
 
@@ -1622,7 +1592,7 @@ struct server_context {
                         queue_tasks.defer(task);
                         break;
                     }
-                    if (!slot->available()) {
+                    if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
                         queue_tasks.defer(task);
@@ -1728,6 +1698,9 @@ struct server_context {
                         { "n_tokens_predicted",              metrics.n_tokens_predicted},
                         { "t_tokens_generation",             metrics.t_tokens_generation},
 
+                        { "n_decode_total",                  metrics.n_decode_total},
+                        { "n_busy_slots_total",              metrics.n_busy_slots_total},
+
                         { "kv_cache_tokens_count",           llama_get_kv_cache_token_count(ctx)},
                         { "kv_cache_used_cells",             llama_get_kv_cache_used_cells(ctx)},
 
@@ -1747,7 +1720,7 @@ struct server_context {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
                         break;
                     }
-                    if (!slot->available()) {
+                    if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
                         queue_tasks.defer(task);
@@ -1788,7 +1761,7 @@ struct server_context {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
                         break;
                     }
-                    if (!slot->available()) {
+                    if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
                         queue_tasks.defer(task);
@@ -1836,7 +1809,7 @@ struct server_context {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
                         break;
                     }
-                    if (!slot->available()) {
+                    if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
                         queue_tasks.defer(task);
@@ -1876,33 +1849,12 @@ struct server_context {
             system_prompt_update();
         }
 
-        // release slots
-        for (auto & slot : slots) {
-            if (slot.command == SLOT_COMMAND_RELEASE) {
-                slot.state       = SLOT_STATE_IDLE;
-                slot.command     = SLOT_COMMAND_NONE;
-                slot.t_last_used = ggml_time_us();
-
-                LOG_INFO("slot released", {
-                    {"id_slot",         slot.id},
-                    {"id_task",         slot.id_task},
-                    {"n_ctx",           n_ctx},
-                    {"n_past",          slot.n_past},
-                    {"n_system_tokens", system_tokens.size()},
-                    {"n_cache_tokens",  slot.cache_tokens.size()},
-                    {"truncated",       slot.truncated}
-                });
-
-                queue_tasks.notify_slot_changed();
-            }
-        }
-
         // check if all slots are idle
         {
             bool all_idle = true;
 
             for (auto & slot : slots) {
-                if (slot.state != SLOT_STATE_IDLE || slot.command != SLOT_COMMAND_NONE) {
+                if (slot.is_processing()) {
                     all_idle = false;
                     break;
                 }
@@ -1973,7 +1925,7 @@ struct server_context {
 
         // frist, add sampled tokens from any ongoing sequences
         for (auto & slot : slots) {
-            if (slot.state == SLOT_STATE_IDLE) {
+            if (slot.state != SLOT_STATE_GENERATING) {
                 continue;
             }
 
@@ -2015,7 +1967,7 @@ struct server_context {
         if (params.cont_batching || batch.n_tokens == 0) {
             for (auto & slot : slots) {
                 // this slot still has a prompt to be processed
-                if (slot.state == SLOT_STATE_IDLE && slot.command == SLOT_COMMAND_LOAD_PROMPT) {
+                if (slot.state == SLOT_STATE_PROCESSING_PROMPT) {
                     auto & prompt_tokens = slot.prompt_tokens;
 
                     // we haven't tokenized the prompt yet - do it now:
@@ -2083,8 +2035,6 @@ struct server_context {
                                 {"id_task", slot.id_task}
                             });
 
-                            slot.state = SLOT_STATE_PROCESSING;
-                            slot.command = SLOT_COMMAND_NONE;
                             slot.release();
                             slot.print_timings();
                             send_final_response(slot);
@@ -2094,8 +2044,6 @@ struct server_context {
                         if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING) {
                             // this prompt is too large to process - discard it
                             if (slot.n_prompt_tokens > n_ubatch) {
-                                slot.state = SLOT_STATE_PROCESSING;
-                                slot.command = SLOT_COMMAND_NONE;
                                 slot.release();
                                 send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER);
                                 continue;
@@ -2140,7 +2088,7 @@ struct server_context {
                                 GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx);
                             }
 
-                            llama_sampling_reset(slot.ctx_sampling);
+                            gpt_sampler_reset(slot.smpl);
 
                             if (!slot.params.cache_prompt) {
                                 slot.n_past_se = 0;
@@ -2153,7 +2101,7 @@ struct server_context {
 
                                 // push the prompt into the sampling context (do not apply grammar)
                                 for (int i = 0; i < slot.n_past; ++i) {
-                                    llama_sampling_accept(slot.ctx_sampling, ctx, slot.cache_tokens[i], false);
+                                    gpt_sampler_accept(slot.smpl, slot.cache_tokens[i], false);
                                 }
                             }
                         }
@@ -2206,7 +2154,7 @@ struct server_context {
                         slot.n_past_se = 0;
                         slot.ga_i = 0;
                         // TODO: is the system prompt ever in the sampling context?
-                        llama_sampling_reset(slot.ctx_sampling);
+                        gpt_sampler_reset(slot.smpl);
                     }
 
                     // remove the non-common part from the cache
@@ -2253,10 +2201,9 @@ struct server_context {
                         {"progress", (float) slot.n_prompt_tokens_processed / slot.n_prompt_tokens},
                     });
 
-                    // entire prompt has been processed - start decoding new tokens
+                    // entire prompt has been processed
                     if (slot.n_past == slot.n_prompt_tokens) {
-                        slot.state   = SLOT_STATE_PROCESSING;
-                        slot.command = SLOT_COMMAND_NONE;
+                        slot.state = SLOT_STATE_DONE_PROMPT;
 
                         GGML_ASSERT(batch.n_tokens > 0);
 
@@ -2338,18 +2285,17 @@ struct server_context {
             };
 
             const int ret = llama_decode(ctx, batch_view);
+            metrics.on_decoded(slots);
 
             if (ret != 0) {
                 if (n_batch == 1 || ret < 0) {
                     // if you get here, it means the KV cache is full - try increasing it via the context size
                     LOG_ERROR("failed to decode the batch: KV cache is full - try increasing it via the context size", {
-                        {"i",   i},
-                        {"n_batch",  ret},
-                        {"ret",   ret},
+                        {"i",       i},
+                        {"n_batch", n_batch},
+                        {"ret",     ret},
                     });
                     for (auto & slot : slots) {
-                        slot.state = SLOT_STATE_PROCESSING;
-                        slot.command = SLOT_COMMAND_NONE;
                         slot.release();
                         send_error(slot, "Input prompt is too big compared to KV size. Please try increasing KV size.");
                     }
@@ -2361,31 +2307,38 @@ struct server_context {
                 i -= n_batch;
 
                 LOG_WARNING("failed to find free space in the KV cache, retrying with smaller batch size - try increasing it via the context size or enable defragmentation", {
-                    {"i",   i},
-                    {"n_batch",  n_batch},
-                    {"ret",   ret},
+                    {"i",       i},
+                    {"n_batch", n_batch},
+                    {"ret",     ret},
                 });
 
                 continue; // continue loop of n_batch
             }
 
             for (auto & slot : slots) {
-                if (slot.state != SLOT_STATE_PROCESSING || slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens)) {
+                if (slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens)) {
                     continue; // continue loop of slots
                 }
 
-                // prompt evaluated for embedding
-                if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING) {
-                    send_embedding(slot, batch_view);
-                    slot.release();
-                    slot.i_batch = -1;
+                if (slot.state == SLOT_STATE_DONE_PROMPT) {
+                    if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING) {
+                        // prompt evaluated for embedding
+                        send_embedding(slot, batch_view);
+                        slot.release();
+                        slot.i_batch = -1;
+                        continue; // continue loop of slots
+                    }
+
+                    // prompt evaluated for next-token prediction
+                    slot.state = SLOT_STATE_GENERATING;
+                } else if (slot.state != SLOT_STATE_GENERATING) {
                     continue; // continue loop of slots
                 }
 
                 completion_token_output result;
-                const llama_token id = llama_sampling_sample(slot.ctx_sampling, ctx, NULL, slot.i_batch - i);
+                const llama_token id = gpt_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
 
-                llama_sampling_accept(slot.ctx_sampling, ctx, id, true);
+                gpt_sampler_accept(slot.smpl, id, true);
 
                 slot.n_decoded += 1;
                 if (slot.n_decoded == 1) {
@@ -2394,37 +2347,19 @@ struct server_context {
                     metrics.on_prompt_eval(slot);
                 }
 
-                llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false };
                 result.tok = id;
 
-                const size_t n_probs = std::min(cur_p.size, (size_t) slot.sparams.n_probs);
-                if (n_probs > 0) {
-                    const size_t n_valid = slot.ctx_sampling->n_valid;
+                const auto * cur_p = gpt_sampler_get_candidates(slot.smpl);
 
-                    // Make sure at least n_probs top tokens are at the front of the vector:
-                    if (slot.sparams.temp == 0.0f && n_probs > n_valid) {
-                        llama_sample_top_k(ctx, &cur_p, n_probs, 0);
-                    }
-
-                    if (slot.sparams.temp == 0.0f) {
-                        // With greedy sampling the probabilities have possibly not been calculated.
-                        for (size_t i = 0; i < n_probs; ++i) {
-                            result.probs.push_back({
-                                cur_p.data[i].id,
-                                i == 0 ? 1.0f : 0.0f
-                            });
-                        }
-                    } else {
-                        for (size_t i = 0; i < n_probs; ++i) {
-                            result.probs.push_back({
-                                cur_p.data[i].id,
-                                i >= n_valid ? 0.0f : cur_p.data[i].p // Tokens filtered out due to e.g. top_k have 0 probability.
-                            });
-                        }
-                    }
+                for (size_t i = 0; i < (size_t) slot.sparams.n_probs; ++i) {
+                    result.probs.push_back({
+                        cur_p->data[i].id,
+                        i >= cur_p->size ? 0.0f : cur_p->data[i].p,
+                    });
                 }
 
                 if (!process_token(result, slot)) {
+                    // release slot because of stop condition
                     slot.release();
                     slot.print_timings();
                     send_final_response(slot);
@@ -2492,14 +2427,10 @@ int main(int argc, char ** argv) {
     // own arguments required by this example
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SERVER)) {
         return 1;
     }
 
-    // parse arguments from environment variables
-    gpt_params_parse_from_env(params);
-
     // TODO: not great to use extern vars
     server_log_json = params.log_json;
     server_verbose = params.verbosity > 0;
@@ -2662,10 +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;
@@ -2705,7 +2642,7 @@ int main(int argc, char ** argv) {
         task.type = SERVER_TASK_TYPE_METRICS;
 
         ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
+        ctx_server.queue_tasks.post(task, true); // high-priority task
 
         // get the result
         server_task_result result = ctx_server.queue_results.recv(task.id);
@@ -2737,7 +2674,7 @@ int main(int argc, char ** argv) {
         task.data.push_back({{"reset_bucket", true}});
 
         ctx_server.queue_results.add_waiting_task_id(task.id);
-        ctx_server.queue_tasks.post(task);
+        ctx_server.queue_tasks.post(task, true); // high-priority task
 
         // get the result
         server_task_result result = ctx_server.queue_results.recv(task.id);
@@ -2751,6 +2688,9 @@ int main(int argc, char ** argv) {
         const uint64_t n_tokens_predicted  = data.at("n_tokens_predicted");
         const uint64_t t_tokens_generation = data.at("t_tokens_generation");
 
+        const uint64_t n_decode_total     = data.at("n_decode_total");
+        const uint64_t n_busy_slots_total = data.at("n_busy_slots_total");
+
         const int32_t kv_cache_used_cells = data.at("kv_cache_used_cells");
 
         // metrics definition: https://prometheus.io/docs/practices/naming/#metric-names
@@ -2771,6 +2711,14 @@ int main(int argc, char ** argv) {
                     {"name",  "tokens_predicted_seconds_total"},
                     {"help",  "Predict process time"},
                     {"value",  (uint64_t) data.at("t_tokens_generation_total") / 1.e3}
+            }, {
+                    {"name",  "n_decode_total"},
+                    {"help",  "Total number of llama_decode() calls"},
+                    {"value",  n_decode_total}
+            }, {
+                    {"name",  "n_busy_slots_per_decode"},
+                    {"help",  "Average number of busy slots per llama_decode() call"},
+                    {"value",  (float) n_busy_slots_total / (float) n_decode_total}
             }}},
             {"gauge", {{
                     {"name",  "prompt_tokens_seconds"},
@@ -2837,7 +2785,7 @@ int main(int argc, char ** argv) {
         task.data = {
             { "id_slot", id_slot },
             { "filename", filename },
-            { "filepath", filepath }
+            { "filepath", filepath },
         };
 
         const int id_task = ctx_server.queue_tasks.post(task);
@@ -2867,7 +2815,7 @@ int main(int argc, char ** argv) {
         task.data = {
             { "id_slot", id_slot },
             { "filename", filename },
-            { "filepath", filepath }
+            { "filepath", filepath },
         };
 
         const int id_task = ctx_server.queue_tasks.post(task);
@@ -2945,7 +2893,7 @@ int main(int argc, char ** argv) {
             { "system_prompt",               ctx_server.system_prompt.c_str() },
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
             { "total_slots",                 ctx_server.params.n_parallel },
-            { "chat_template",               curr_tmpl.c_str() }
+            { "chat_template",               curr_tmpl.c_str() },
         };
 
         res_ok(res, data);
@@ -3056,13 +3004,13 @@ int main(int argc, char ** argv) {
         json models = {
             {"object", "list"},
             {"data", {
-                 {
-                     {"id",       params.model_alias},
-                     {"object",   "model"},
-                     {"created",  std::time(0)},
-                     {"owned_by", "llamacpp"},
-                     {"meta",     ctx_server.model_meta()}
-                 },
+                {
+                    {"id",       params.model_alias},
+                    {"object",   "model"},
+                    {"created",  std::time(0)},
+                    {"owned_by", "llamacpp"},
+                    {"meta",     ctx_server.model_meta()}
+                },
              }}
         };
 
@@ -3072,12 +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/embeddings.feature

@@ -9,8 +9,11 @@ Feature: llama.cpp server
     And   a model alias bert-bge-small
     And   42 as server seed
     And   2 slots
-    And   1024 as batch size
-    And   1024 as ubatch size
+    # the bert-bge-small model has context size of 512
+    # since the generated prompts are as big as the batch size, we need to set the batch size to 512
+    # ref: https://huggingface.co/BAAI/bge-small-en-v1.5/blob/5c38ec7c405ec4b44b94cc5a9bb96e735b38267a/config.json#L20
+    And   512 as batch size
+    And   512 as ubatch size
     And   2048 KV cache size
     And   embeddings extraction
     Then  the server is starting

examples/server/tests/features/parallel.feature

@@ -77,6 +77,35 @@ Feature: Parallel
       | disabled  | 128       |
       | enabled   | 64        |
 
+  Scenario Outline: Multi users with number of prompts exceeding number of slots
+    Given a system prompt You are a writer.
+    And   a model tinyllama-2
+    Given a prompt:
+      """
+      Write a very long book.
+      """
+    And a prompt:
+      """
+      Write another a poem.
+      """
+    And a prompt:
+      """
+      What is LLM?
+      """
+    And a prompt:
+      """
+      The sky is blue and I love it.
+      """
+    And <n_predict> max tokens to predict
+    And streaming is <streaming>
+    Given concurrent OAI completions requests
+    Then the server is busy
+    Then the server is idle
+    Then all prompts are predicted with <n_predict> tokens
+    Examples:
+      | streaming | n_predict |
+      | disabled  | 128       |
+      | enabled   | 64        |
 
   Scenario:  Multi users with total number of tokens to predict exceeds the KV Cache size #3969
     Given a prompt:

examples/server/tests/features/passkey.feature

@@ -15,6 +15,7 @@ Feature: Passkey / Self-extend with context shift
     And   <n_junk> as number of junk
     And   <n_predicted> server max tokens to predict
     And   42 as seed
+    And   0.0 temperature
     And   <n_ctx> KV cache size
     And   1 slots
     And   <n_ga> group attention factor to extend context size through self-extend
@@ -22,7 +23,8 @@ Feature: Passkey / Self-extend with context shift
     # Can be override with N_GPU_LAYERS
     And   <ngl> GPU offloaded layers
     Then  the server is starting
-    Then  the server is healthy
+    # Higher timeout because the model may need to be downloaded from the internet
+    Then  the server is healthy with timeout 120 seconds
     Given available models
     Then  model 0 is trained on <n_ctx_train> tokens context
     Given a prefix prompt:

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
@@ -202,17 +205,15 @@ def step_start_server(context):
             time.sleep(0.1)
 
 
-@step("the server is {expecting_status}")
-@async_run_until_complete
-async def step_wait_for_the_server_to_be_started(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
+async def wait_for_server_status_with_timeout(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str, timeout: int):
     match expecting_status:
         case 'healthy':
             await wait_for_slots_status(context, context.base_url, 200,
-                                        timeout=30)
+                                        timeout=timeout)
 
         case 'ready' | 'idle':
             await wait_for_slots_status(context, context.base_url, 200,
-                                        timeout=30,
+                                        timeout=timeout,
                                         params={'fail_on_no_slot': 1},
                                         slots_idle=context.n_slots,
                                         slots_processing=0)
@@ -225,6 +226,18 @@ async def step_wait_for_the_server_to_be_started(context, expecting_status: Lite
             assert False, "unknown status"
 
 
+@step("the server is {expecting_status} with timeout {timeout:d} seconds")
+@async_run_until_complete
+async def step_wait_for_server_status_with_timeout(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str, timeout: int):
+    await wait_for_server_status_with_timeout(context, expecting_status, timeout)
+
+
+@step("the server is {expecting_status}")
+@async_run_until_complete
+async def step_wait_for_server_status(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
+    await wait_for_server_status_with_timeout(context, expecting_status, 30)
+
+
 @step('all slots are {expected_slot_status_string}')
 @async_run_until_complete
 async def step_all_slots_status(context, expected_slot_status_string: Literal['idle', 'busy'] | str):
@@ -687,6 +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

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -6,9 +7,7 @@
 #include <string>
 #include <vector>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32\n", argv[0]);
     LOG_TEE("\n");
@@ -20,8 +19,7 @@ int main(int argc, char ** argv) {
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
         return 1;
     }
 
@@ -55,6 +53,14 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    auto sparams = llama_sampler_chain_default_params();
+
+    sparams.no_perf = false;
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
+
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
@@ -110,20 +116,7 @@ int main(int argc, char ** argv) {
     while (n_cur <= n_predict) {
         // sample the next token
         {
-            auto   n_vocab = llama_n_vocab(model);
-            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
-
-            std::vector<llama_token_data> candidates;
-            candidates.reserve(n_vocab);
-
-            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-            }
-
-            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-            // sample the most likely token
-            const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
 
             // is it an end of generation?
             if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
@@ -160,12 +153,14 @@ int main(int argc, char ** argv) {
     LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
     llama_batch_free(batch);
-
+    llama_sampler_free(smpl);
     llama_free(ctx);
     llama_free_model(model);
 

examples/speculative/speculative.cpp

@@ -1,11 +1,13 @@
+#include "arg.h"
 #include "common.h"
+#include "sampling.h"
 #include "llama.h"
 
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 #include <set>
+#include <random>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -21,14 +23,13 @@ struct seq_draft {
     std::vector<llama_token> tokens;
     std::vector<std::vector<llama_token_data>> dists;
 
-    struct llama_sampling_context * ctx_sampling;
+    struct gpt_sampler * smpl = nullptr;
 };
 
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
         return 1;
     }
 
@@ -43,10 +44,7 @@ int main(int argc, char ** argv) {
     // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
     const float p_split  = params.p_split;
 
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-    std::default_random_engine rng(params.seed);
+    std::default_random_engine rng(params.sparams.seed);
     std::uniform_real_distribution<> u_dist;
 
 #ifndef LOG_DISABLE_LOGS
@@ -179,19 +177,17 @@ int main(int argc, char ** argv) {
     // used to determine end of generation
     bool has_eos = false;
 
-    // target model sampling context
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+    // target model sampling context (reuse the llama_context's sampling instance)
+    struct gpt_sampler * smpl = gpt_sampler_init(model_tgt, params.sparams);
+
+    struct llama_sampler * softmax = llama_sampler_init_softmax();
 
     // draft sequence data
     std::vector<seq_draft> drafts(n_seq_dft);
 
-    params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
-    if (params.sparams.temp == 0) {
-        params.sparams.temp = -1.0f; // force greedy sampling with probs for the draft model
-    }
-
     for (int s = 0; s < n_seq_dft; ++s) {
-        drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
+        // allocate gpt_sampler for each draft sequence
+        drafts[s].smpl = gpt_sampler_init(model_dft, params.sparams);
     }
 
     llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
@@ -233,12 +229,12 @@ int main(int argc, char ** argv) {
                 bool accept = false;
                 if (params.sparams.temp > 0) {
                     // stochastic verification
+                    gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
 
-                    llama_token_data_array dist_tgt = llama_sampling_prepare(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft], true, NULL);
-                    llama_sample_softmax(ctx_tgt, &dist_tgt);
-                    float p_tgt = 0, p_dft = 0;
+                    auto & dist_tgt = *gpt_sampler_get_candidates(smpl);
 
-                    // GGML_ASSERT(dist_tgt.size() == dist_dft.size());
+                    float p_tgt = 0.0f;
+                    float p_dft = 0.0f;
 
                     while (active_seqs.size() > 0) {
                         // randomly select a sequence to verify from active sequences
@@ -257,9 +253,13 @@ int main(int argc, char ** argv) {
                             }
                             continue;
                         }
+
                         LOG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
                         float r = u_dist(rng);
-                        llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), true };
+                        llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), LLAMA_TOKEN_NULL, true };
+
+                        //GGML_ASSERT(dist_tgt.size <= dist_dft.size);
+
                         // acquire the token probabilities assigned by the draft and target models
                         for (size_t i = 0; i < dist_tgt.size; i++) {
                             if (dist_tgt.data[i].id == drafts[s].tokens[i_dft]) {
@@ -278,7 +278,7 @@ int main(int argc, char ** argv) {
                             accept = true;
                             token_id = drafts[s].tokens[i_dft];
                             token_str = llama_token_to_piece(ctx_tgt, token_id);
-                            llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+                            gpt_sampler_accept(smpl, token_id, true);
 
                             LOG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
                             break;
@@ -289,7 +289,6 @@ int main(int argc, char ** argv) {
                             // calculate residual probability
                             GGML_ASSERT(dist_tgt.sorted);
                             GGML_ASSERT(dist_dft.sorted);
-                            float sum_probs = 0.0f;
 
                             // sort dist by id
                             std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
@@ -299,10 +298,18 @@ int main(int argc, char ** argv) {
                                 return a.id < b.id;
                             });
 
+                            float sum_probs = 0.0f;
+
                             for (size_t i = 0; i < dist_tgt.size; i++) {
-                                dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
+                                if (i < dist_dft.size) {
+                                    dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
+                                } else {
+                                    dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p);
+                                }
+
                                 sum_probs += dist_tgt.data[i].p;
                             }
+
                             for (size_t i = 0; i < dist_tgt.size; i++) {
                                 dist_tgt.data[i].p /= sum_probs;
                             }
@@ -332,21 +339,29 @@ int main(int argc, char ** argv) {
                         // all drafted tokens were rejected
                         // sample from the target model
                         LOG("all drafted tokens were rejected, sampling from residual distribution\n");
-                        token_id = llama_sample_token(ctx_tgt, &dist_tgt);
-                        llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+                        std::vector<float> probs(dist_tgt.size);
+                        for (size_t i = 0; i < dist_tgt.size; ++i) {
+                            probs[i] = dist_tgt.data[i].p;
+                        }
+
+                        std::discrete_distribution<> dist(probs.begin(), probs.end());
+
+                        const int idx = dist(rng);
+
+                        token_id = dist_tgt.data[idx].id;
+                        gpt_sampler_accept(smpl, token_id, true);
                         token_str = llama_token_to_piece(ctx_tgt, token_id);
                     }
-
                 } else {
                     // greedy verification
 
                     // sample from the target model
                     LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
-                    token_id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
+                    token_id = gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft]);
 
-                    llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+                    gpt_sampler_accept(smpl, token_id, true);
 
-                    //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
+                    //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, smpl->prev).c_str());
 
                     token_str = llama_token_to_piece(ctx_tgt, token_id);
 
@@ -434,7 +449,10 @@ int main(int argc, char ** argv) {
             break;
         }
 
-        llama_sampling_cp(ctx_sampling, drafts[0].ctx_sampling);
+        if (drafts[0].smpl) {
+            gpt_sampler_free(drafts[0].smpl);
+        }
+        drafts[0].smpl = gpt_sampler_clone(smpl);
 
         int n_seq_cur  = 1;
         int n_past_cur = n_past_dft;
@@ -463,20 +481,20 @@ int main(int argc, char ** argv) {
                     continue;
                 }
 
-                llama_sampling_sample(drafts[s].ctx_sampling, ctx_dft, NULL, drafts[s].i_batch_dft);
+                gpt_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true);
 
-                const auto & cur_p = drafts[s].ctx_sampling->cur;
+                const auto * cur_p = gpt_sampler_get_candidates(drafts[s].smpl);
 
-                for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p.size()); ++k) {
+                for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) {
                     LOG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
-                            k, s, i, cur_p[k].id, cur_p[k].p, llama_token_to_piece(ctx_dft, cur_p[k].id).c_str());
+                            k, s, i, cur_p->data[k].id, cur_p->data[k].p, llama_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
                 }
 
                 std::vector<int> sa(1, s);
 
                 // attempt to split the branch if the probability is high enough
                 for (int f = 1; f < 8; ++f) {
-                    if (n_seq_cur < n_seq_dft && cur_p[f].p > p_split) {
+                    if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_split) {
                         LOG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
                         llama_kv_cache_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
@@ -503,7 +521,10 @@ int main(int argc, char ** argv) {
                         drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
                         drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
 
-                        llama_sampling_cp(drafts[s].ctx_sampling, drafts[n_seq_cur].ctx_sampling);
+                        if (drafts[n_seq_cur].smpl) {
+                            gpt_sampler_free(drafts[n_seq_cur].smpl);
+                        }
+                        drafts[n_seq_cur].smpl = gpt_sampler_clone(drafts[s].smpl);
 
                         sa.push_back(n_seq_cur);
 
@@ -515,15 +536,15 @@ int main(int argc, char ** argv) {
 
                 // add drafted token for each sequence
                 for (int is = 0; is < (int) sa.size(); ++is) {
-                    const llama_token id = cur_p[is].id;
+                    const llama_token id = cur_p->data[is].id;
 
                     const int s = sa[is];
 
-                    llama_sampling_accept(drafts[s].ctx_sampling, ctx_dft, id, true);
+                    gpt_sampler_accept(drafts[s].smpl, id, true);
 
                     drafts[s].tokens.push_back(id);
                     // save cur_p.data into drafts[s].dists
-                    drafts[s].dists.push_back(cur_p);
+                    drafts[s].dists.push_back({cur_p->data, cur_p->data + cur_p->size});
 
                     // add unique drafted tokens to the target batch
                     drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
@@ -593,17 +614,19 @@ int main(int argc, char ** argv) {
     LOG_TEE("n_accept  = %d\n", n_accept);
     LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
-    LOG_TEE("\ndraft:\n");
-    llama_print_timings(ctx_dft);
+    LOG_TEE("\ndraft:\n\n");
+    // TODO: print sampling/grammar timings for all drafts
+    llama_perf_context_print(ctx_dft);
 
-    LOG_TEE("\ntarget:\n");
-    llama_print_timings(ctx_tgt);
+    LOG_TEE("\ntarget:\n\n");
+    gpt_perf_print(ctx_tgt, smpl);
 
-    llama_sampling_free(ctx_sampling);
+    gpt_sampler_free(smpl);
     for (int s = 0; s < n_seq_dft; ++s) {
-        llama_sampling_free(drafts[s].ctx_sampling);
+        gpt_sampler_free(drafts[s].smpl);
     }
 
+    llama_sampler_free(softmax);
     llama_batch_free(batch_dft);
 
     llama_free(ctx_tgt);

examples/sycl/run-llama2.sh

@@ -4,33 +4,23 @@
 #  Copyright (C) 2024 Intel Corporation
 #  SPDX-License-Identifier: MIT
 
-INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 source /opt/intel/oneapi/setvars.sh
 
-if [ $# -gt 0 ]; then
-    GGML_SYCL_DEVICE=$1
-    GGML_SYCL_SINGLE_GPU=1
-else
-    GGML_SYCL_DEVICE=0
-    GGML_SYCL_SINGLE_GPU=0
-fi
-
 #export GGML_SYCL_DEBUG=1
 
-
 #ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
 
-if [ $GGML_SYCL_SINGLE_GPU -eq 1 ]; then
+INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
+MODEL_FILE=llama-2-7b.Q4_0.gguf
+NGL=33
+
+if [ $# -gt 0 ]; then
+    GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -mg $GGML_SYCL_DEVICE -sm none
+
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0
 fi
-
-#use main GPU only
-#ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
-
-#use multiple GPUs with same max compute units
-#ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1725024810,
-        "narHash": "sha256-ODYRm8zHfLTH3soTFWE452ydPYz2iTvr9T8ftDMUQ3E=",
+        "lastModified": 1725234343,
+        "narHash": "sha256-+ebgonl3NbiKD2UD0x4BszCZQ6sTfL4xioaM49o5B3Y=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "af510d4a62d071ea13925ce41c95e3dec816c01d",
+        "rev": "567b938d64d4b4112ee253b9274472dc3a346eb6",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1724819573,
-        "narHash": "sha256-GnR7/ibgIH1vhoy8cYdmXE6iyZqKqFxQSVkFgosBh6w=",
+        "lastModified": 1725634671,
+        "narHash": "sha256-v3rIhsJBOMLR8e/RNWxr828tB+WywYIoajrZKFM+0Gg=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "71e91c409d1e654808b2621f28a327acfdad8dc2",
+        "rev": "574d1eac1c200690e27b8eb4e24887f8df7ac27c",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1722555339,
-        "narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
+        "lastModified": 1725233747,
+        "narHash": "sha256-Ss8QWLXdr2JCBPcYChJhz4xJm+h/xjl4G0c0XlP6a74=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/356624c12086a18f2ea2825fed34523d60ccc4e3.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/356624c12086a18f2ea2825fed34523d60ccc4e3.tar.gz"
       }
     },
     "root": {

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/include/ggml.h

@@ -358,6 +358,7 @@ extern "C" {
 
     struct ggml_object;
     struct ggml_context;
+    struct ggml_cgraph;
 
     // NOTE: always add types at the end of the enum to keep backward compatibility
     enum ggml_type {
@@ -575,23 +576,9 @@ extern "C" {
         GGML_TENSOR_FLAG_PARAM  = 4,
     };
 
-    // ggml object
-    struct ggml_object {
-        size_t offs;
-        size_t size;
-
-        struct ggml_object * next;
-
-        enum ggml_object_type type;
-
-        char padding[4];
-    };
-
-    static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
-
     // n-dimensional tensor
     struct ggml_tensor {
-        enum ggml_type         type;
+        enum ggml_type type;
 
         GGML_DEPRECATED(enum ggml_backend_type backend, "use the buffer type to find the storage location of the tensor");
 
@@ -655,7 +642,7 @@ extern "C" {
 
     struct ggml_threadpool;     // forward declaration, see ggml.c
 
-    typedef struct  ggml_threadpool * ggml_threadpool_t;
+    typedef struct ggml_threadpool * ggml_threadpool_t;
 
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
@@ -671,35 +658,6 @@ extern "C" {
         void *              abort_callback_data;
     };
 
-    enum ggml_cgraph_eval_order {
-        GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
-        GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
-        GGML_CGRAPH_EVAL_ORDER_COUNT
-    };
-
-    typedef uint32_t ggml_bitset_t;
-
-    struct ggml_hash_set {
-        size_t size;
-        ggml_bitset_t * used;
-        struct ggml_tensor ** keys;
-    };
-
-    // computation graph
-    struct ggml_cgraph {
-        int size;
-        int n_nodes;
-        int n_leafs;
-
-        struct ggml_tensor ** nodes;
-        struct ggml_tensor ** grads;
-        struct ggml_tensor ** leafs;
-
-        struct ggml_hash_set visited_hash_set;
-
-        enum ggml_cgraph_eval_order order;
-    };
-
     // scratch buffer
     struct ggml_scratch {
         size_t offs;
@@ -1272,7 +1230,7 @@ extern "C" {
             size_t                nb1,
             size_t                nb2,
             size_t                nb3,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return view(a)
     GGML_API struct ggml_tensor * ggml_set_inplace(
@@ -1282,19 +1240,19 @@ extern "C" {
             size_t                nb1,
             size_t                nb2,
             size_t                nb3,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     GGML_API struct ggml_tensor * ggml_set_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     GGML_API struct ggml_tensor * ggml_set_1d_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set_2d(
@@ -1302,7 +1260,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             size_t                nb1,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return view(a)
     GGML_API struct ggml_tensor * ggml_set_2d_inplace(
@@ -1310,7 +1268,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             size_t                nb1,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // a -> b, return view(b)
     GGML_API struct ggml_tensor * ggml_cpy(
@@ -2017,8 +1975,6 @@ extern "C" {
     typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
     typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
 
-    #define GGML_N_TASKS_MAX -1
-
     GGML_API struct ggml_tensor * ggml_map_custom1(
             struct ggml_context   * ctx,
             struct ggml_tensor    * a,
@@ -2088,30 +2044,35 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * tensor);
 
-
     GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
 
     // graph allocation in a context
-    GGML_API struct ggml_cgraph * ggml_new_graph         (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
-    GGML_API struct ggml_cgraph * ggml_new_graph_custom  (struct ggml_context * ctx, size_t size, bool grads);
-    GGML_API struct ggml_cgraph * ggml_graph_dup         (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
-    GGML_API struct ggml_cgraph   ggml_graph_view        (struct ggml_cgraph * cgraph, int i0, int i1);
-    GGML_API void                 ggml_graph_cpy         (struct ggml_cgraph * src, struct ggml_cgraph * dst);
-    GGML_API void                 ggml_graph_reset       (struct ggml_cgraph * cgraph);  // zero grads
-    GGML_API void                 ggml_graph_clear       (struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
+    GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads);
+    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+    GGML_API void                 ggml_graph_cpy       (struct ggml_cgraph * src, struct ggml_cgraph * dst);
+    GGML_API void                 ggml_graph_reset     (struct ggml_cgraph * cgraph);  // zero grads
+    GGML_API void                 ggml_graph_clear     (struct ggml_cgraph * cgraph);
+
+    GGML_API int                   ggml_graph_size   (struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_tensor *  ggml_graph_node   (struct ggml_cgraph * cgraph, int i); // if i < 0, returns nodes[n_nodes + i]
+    GGML_API struct ggml_tensor ** ggml_graph_nodes  (struct ggml_cgraph * cgraph);
+    GGML_API int                   ggml_graph_n_nodes(struct ggml_cgraph * cgraph);
+
+    GGML_API void   ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
 
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
-    GGML_API struct ggml_threadpool_params   ggml_threadpool_params_default(int n_threads);
-    GGML_API void                            ggml_threadpool_params_init  (struct ggml_threadpool_params *p, int n_threads);
-    GGML_API bool                            ggml_threadpool_params_match (const struct ggml_threadpool_params *p0, const struct ggml_threadpool_params *p1);
-    GGML_API struct ggml_threadpool*         ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_API void                            ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_API int                             ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_API void                            ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_API void                            ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
+    GGML_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
+    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
+    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
@@ -2509,6 +2470,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_gpublas    (void);
     GGML_API int ggml_cpu_has_sse3       (void);
     GGML_API int ggml_cpu_has_ssse3      (void);
+    GGML_API int ggml_cpu_has_riscv_v    (void);
     GGML_API int ggml_cpu_has_sycl       (void);
     GGML_API int ggml_cpu_has_rpc        (void);
     GGML_API int ggml_cpu_has_vsx        (void);

ggml/src/CMakeLists.txt

@@ -26,6 +26,8 @@ if (NOT MSVC)
     endif()
 endif()
 
+unset(GGML_EXTRA_LIBS)
+
 if (APPLE AND GGML_ACCELERATE)
     find_library(ACCELERATE_FRAMEWORK Accelerate)
     if (ACCELERATE_FRAMEWORK)
@@ -35,7 +37,7 @@ if (APPLE AND GGML_ACCELERATE)
         add_compile_definitions(ACCELERATE_NEW_LAPACK)
         add_compile_definitions(ACCELERATE_LAPACK_ILP64)
 
-        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} ${ACCELERATE_FRAMEWORK})
+        list(APPEND GGML_EXTRA_LIBS ${ACCELERATE_FRAMEWORK})
     else()
         message(WARNING "Accelerate framework not found")
     endif()
@@ -87,7 +89,7 @@ if (GGML_METAL)
             COMMENT "Generate assembly for embedded Metal library"
         )
 
-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM})
+    list(APPEND GGML_SOURCES_METAL ${METALLIB_EMBED_ASM})
     else()
         if (GGML_METAL_SHADER_DEBUG)
             # custom command to do the following:
@@ -132,7 +134,7 @@ if (GGML_METAL)
             )
     endif() # GGML_METAL_EMBED_LIBRARY
 
-    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS}
+    list(APPEND GGML_EXTRA_LIBS
         ${FOUNDATION_LIBRARY}
         ${METAL_FRAMEWORK}
         ${METALKIT_FRAMEWORK}
@@ -157,11 +159,11 @@ if (GGML_OPENMP)
 
         add_compile_definitions(GGML_USE_OPENMP)
 
-        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+        list(APPEND GGML_EXTRA_LIBS OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
 
         if (GGML_MUSA)
-            set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} "/usr/lib/llvm-10/include/openmp")
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} "/usr/lib/llvm-10/lib/libomp.so")
+            list(APPEND GGML_EXTRA_INCLUDES "/usr/lib/llvm-10/include/openmp")
+            list(APPEND GGML_EXTRA_LIBS     "/usr/lib/llvm-10/lib/libomp.so")
         endif()
     else()
         message(WARNING "OpenMP not found")
@@ -244,8 +246,8 @@ if (GGML_BLAS)
         set(GGML_HEADERS_BLAS ../include/ggml-blas.h)
         set(GGML_SOURCES_BLAS ggml-blas.cpp)
 
-        set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     ${BLAS_LIBRARIES})
-        set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
+        list(APPEND GGML_EXTRA_LIBS     ${BLAS_LIBRARIES})
+        list(APPEND GGML_EXTRA_INCLUDES ${BLAS_INCLUDE_DIRS})
     else()
         message(WARNING "BLAS not found, please refer to "
         "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@@ -368,19 +370,19 @@ if (GGML_CUDA)
         if (GGML_STATIC)
             if (WIN32)
                 # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
-                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+                list(APPEND GGML_EXTRA_LIBS CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
             else ()
                 if (GGML_MUSA)
-                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart_static MUSA::mublas_static)
+                    list(APPEND GGML_EXTRA_LIBS MUSA::musart_static MUSA::mublas_static)
                 else()
-                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                    list(APPEND GGML_EXTRA_LIBS CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
                 endif()
             endif()
         else()
             if (GGML_MUSA)
-                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart MUSA::mublas)
+                list(APPEND GGML_EXTRA_LIBS MUSA::musart MUSA::mublas)
             else()
-                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+                list(APPEND GGML_EXTRA_LIBS CUDA::cudart CUDA::cublas CUDA::cublasLt)
             endif()
         endif()
 
@@ -388,9 +390,9 @@ if (GGML_CUDA)
             # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
         else()
             if (GGML_MUSA)
-                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
+                list(APPEND GGML_EXTRA_LIBS MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
             else()
-                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+                list(APPEND GGML_EXTRA_LIBS CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
             endif()
         endif()
     else()
@@ -495,7 +497,7 @@ if (GGML_HIPBLAS)
 
     if (CXX_IS_HIPCC)
         set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
-        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} hip::device)
+        list(APPEND GGML_EXTRA_LIBS hip::device)
     else()
         set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
     endif()
@@ -504,7 +506,8 @@ if (GGML_HIPBLAS)
         message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
     endif()
 
-    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} PUBLIC hip::host roc::rocblas roc::hipblas)
+    # TODO: this "PUBLIC" here seems wrong
+    list(APPEND GGML_EXTRA_LIBS PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (GGML_SYCL)
@@ -513,7 +516,8 @@ if (GGML_SYCL)
     endif()
 
     check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
-    if ( DEFINED ENV{ONEAPI_ROOT})
+
+    if (DEFINED ENV{ONEAPI_ROOT})
         message(STATUS "Using oneAPI Release SYCL compiler (icpx).")
     elseif(SUPPORTS_SYCL)
         message(WARNING "Using open-source SYCL compiler (clang++). Didn't detect ENV {ONEAPI_ROOT}.
@@ -551,21 +555,27 @@ if (GGML_SYCL)
 
     find_package(DNNL)
     message("-- DNNL found:" ${DNNL_FOUND})
+
     if (GGML_SYCL_TARGET STREQUAL "INTEL")
         add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
     else()
         add_compile_definitions(GGML_SYCL_DNNL=0)
     endif()
+
+    if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
+        list(APPEND GGML_EXTRA_LIBS DNNL::dnnl)
+    endif()
+
     if (WIN32)
         find_package(IntelSYCL REQUIRED)
         find_package(MKL REQUIRED)
-        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+        list(APPEND GGML_EXTRA_LIBS IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
     else()
         if (GGML_SYCL_TARGET STREQUAL "INTEL")
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+            list(APPEND GGML_EXTRA_LIBS OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
         elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
             set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} -fsycl pthread m dl onemkl)
+            list(APPEND GGML_EXTRA_LIBS pthread m dl onemkl)
         endif()
     endif()
     if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
@@ -579,7 +589,7 @@ if (GGML_RPC)
     list(APPEND GGML_CDEF_PUBLIC GGML_USE_RPC)
 
     if (WIN32)
-        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} ws2_32)
+        list(APPEND GGML_EXTRA_LIBS ws2_32)
     endif()
 
     set(GGML_HEADERS_RPC ../include/ggml-rpc.h)
@@ -657,8 +667,8 @@ if (GGML_VULKAN)
         set(GGML_HEADERS_VULKAN ${CMAKE_CURRENT_SOURCE_DIR}/../include/ggml-vulkan.h ${_ggml_vk_header})
         set(GGML_SOURCES_VULKAN ggml-vulkan.cpp ${_ggml_vk_source})
 
-        set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS} Vulkan::Vulkan)
-        set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CMAKE_CURRENT_BINARY_DIR})
+        list(APPEND GGML_EXTRA_LIBS     Vulkan::Vulkan)
+        list(APPEND GGML_EXTRA_INCLUDES ${CMAKE_CURRENT_BINARY_DIR})
     else()
         message(WARNING "Vulkan not found")
     endif()
@@ -817,8 +827,8 @@ if (GGML_KOMPUTE)
 
         list(APPEND GGML_CDEF_PUBLIC GGML_USE_KOMPUTE)
 
-        set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     kompute)
-        set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CMAKE_CURRENT_BINARY_DIR})
+        list(APPEND GGML_EXTRA_LIBS     kompute)
+        list(APPEND GGML_EXTRA_INCLUDES ${CMAKE_CURRENT_BINARY_DIR})
     else()
         message(WARNING "Kompute not found")
     endif()
@@ -883,9 +893,10 @@ if (GGML_CANN)
             message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
             message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
 
-            set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     ${CANN_LIBRARIES} )
-            set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CANN_INCLUDE_DIRS})
-            set(GGML_EXTRA_LIBDIRS  ${GGML_EXTRA_LIBDIRS}  ${CANN_INSTALL_DIR}/lib64)
+            list(APPEND GGML_EXTRA_LIBS     ${CANN_LIBRARIES} )
+            list(APPEND GGML_EXTRA_INCLUDES ${CANN_INCLUDE_DIRS})
+            list(APPEND GGML_EXTRA_LIBDIRS  ${CANN_INSTALL_DIR}/lib64)
+
             list(APPEND GGML_CDEF_PUBLIC GGML_USE_CANN)
         endif()
     else()
@@ -1322,12 +1333,14 @@ if (EMSCRIPTEN)
     set_target_properties(ggml PROPERTIES COMPILE_FLAGS "-msimd128")
 endif()
 
-target_compile_definitions(ggml PUBLIC  ${GGML_CDEF_PUBLIC})
-target_include_directories(ggml PUBLIC ../include)
+target_compile_definitions(ggml PUBLIC    ${GGML_CDEF_PUBLIC})
+target_include_directories(ggml PUBLIC  ../include)
 target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
-target_link_directories(ggml PRIVATE ${GGML_EXTRA_LIBDIRS})
+target_link_directories   (ggml PRIVATE   ${GGML_EXTRA_LIBDIRS})
 target_compile_features   (ggml PRIVATE c_std_11) # don't bump
 
+list(REMOVE_DUPLICATES GGML_EXTRA_LIBS)
+
 target_link_libraries(ggml PRIVATE Threads::Threads ${GGML_EXTRA_LIBS})
 
 find_library(MATH_LIBRARY m)

ggml/src/ggml-backend.c

@@ -827,6 +827,10 @@ GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const
                 op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
+        case GGML_OP_ROPE_BACK:
+            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
+        case GGML_OP_IM2COL_BACK:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
         default:
             return true;
     }

ggml/src/ggml-blas.cpp

@@ -1,3 +1,4 @@
+#include "ggml-impl.h"
 #include "ggml-blas.h"
 #include "ggml-backend-impl.h"
 

ggml/src/ggml-cann.cpp

@@ -30,6 +30,7 @@
 #include <cstring>
 #include <mutex>
 
+#include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cann/aclnn_ops.h"
 #include "ggml-cann/common.h"
@@ -1220,6 +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.
@@ -1942,7 +2053,7 @@ GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) {
         GGML_CANN_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
         return nullptr;
     }
-
+    ggml_cann_set_device(ctx->device);
     ggml_backend_t cann_backend =
         new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
                          /* .interface = */ ggml_backend_cann_interface,

ggml/src/ggml-cann/Doxyfile

@@ -32,7 +32,7 @@ DOXYFILE_ENCODING      = UTF-8
 # title of most generated pages and in a few other places.
 # The default value is: My Project.
 
-PROJECT_NAME           = "llama.cpp"
+PROJECT_NAME           = "ggml"
 
 # The PROJECT_NUMBER tag can be used to enter a project or revision number. This
 # could be handy for archiving the generated documentation or if some version
@@ -44,7 +44,7 @@ PROJECT_NUMBER         =
 # for a project that appears at the top of each page and should give viewer a
 # quick idea about the purpose of the project. Keep the description short.
 
-PROJECT_BRIEF          = "llama inference engine"
+PROJECT_BRIEF          = "Tensor library for machine learning"
 
 # With the PROJECT_LOGO tag one can specify a logo or an icon that is included
 # in the documentation. The maximum height of the logo should not exceed 55

ggml/src/ggml-cuda.cu

@@ -1,5 +1,5 @@
 #include "ggml-cuda.h"
-#include "ggml.h"
+#include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 
 #include "ggml-cuda/common.cuh"
@@ -27,6 +27,7 @@
 #include "ggml-cuda/rope.cuh"
 #include "ggml-cuda/scale.cuh"
 #include "ggml-cuda/softmax.cuh"
+#include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
@@ -2180,6 +2181,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
             ggml_cuda_dup(ctx, dst);
             break;
         case GGML_OP_ADD:
+        case GGML_OP_ADD1: // TODO: more efficient implementation
             ggml_cuda_op_add(ctx, dst);
             break;
         case GGML_OP_SUB:
@@ -2196,6 +2198,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
             break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(dst)) {
+                case GGML_UNARY_OP_NEG:
+                    ggml_cuda_op_neg(ctx, dst);
+                    break;
                 case GGML_UNARY_OP_GELU:
                     ggml_cuda_op_gelu(ctx, dst);
                     break;
@@ -2304,6 +2309,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_POOL_2D:
             ggml_cuda_op_pool2d(ctx, dst);
             break;
+        case GGML_OP_SUM:
+            ggml_cuda_op_sum(ctx, dst);
+            break;
         case GGML_OP_SUM_ROWS:
             ggml_cuda_op_sum_rows(ctx, dst);
             break;
@@ -2544,7 +2552,11 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         for (int i = 0; i < cgraph->n_nodes; i++) {
             ggml_tensor * node = cgraph->nodes[i];
 
-            if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
+            if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+                continue;
+            }
+
+            if (node->src[0] && node->src[0]->buffer && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
                 use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
 #ifndef NDEBUG
                 GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);
@@ -2748,6 +2760,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_NEG:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
@@ -2877,6 +2890,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_TRANSPOSE:
         case GGML_OP_NORM:
         case GGML_OP_ADD:
+        case GGML_OP_ADD1:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -2887,14 +2901,18 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+            return true;
         case GGML_OP_CONT:
+            return op->src[0]->type != GGML_TYPE_BF16;
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_ROPE:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_IM2COL:
+            return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_2D:
+        case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:

ggml/src/ggml-cuda/cross-entropy-loss.cu

@@ -1,6 +1,6 @@
 #include "common.cuh"
 #include "cross-entropy-loss.cuh"
-#include "sumrows.cuh"
+#include "sum.cuh"
 
 #include <cmath>
 #include <cstdint>
@@ -102,5 +102,5 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
     cross_entropy_loss_f32<<<blocks_num, blocks_dim, shmem, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
 
     // Combine results from individual blocks:
-    sum_rows_f32_cuda(dst_tmp.ptr, dst_d, blocks_num.x, 1, stream);
+    sum_f32_cuda(pool, dst_tmp.ptr, dst_d, blocks_num.x, stream);
 }

ggml/src/ggml-cuda/fattn.cu

@@ -152,7 +152,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
     }                                                                       \
 
 static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * Q = dst->src[1];
+    ggml_tensor * Q = dst->src[0];
     ggml_tensor * K = dst->src[1];
     ggml_tensor * V = dst->src[2];
 
@@ -227,7 +227,7 @@ static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, gg
     }                                                                       \
 
 static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * Q = dst->src[1];
+    ggml_tensor * Q = dst->src[0];
     ggml_tensor * K = dst->src[1];
     ggml_tensor * V = dst->src[2];
 

ggml/src/ggml-cuda/mmq.cu

@@ -26,7 +26,11 @@ void ggml_cuda_op_mul_mat_q(
     // nrows_dst == nrows of the matrix that the kernel writes into
     const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
 
-    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst};
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = compute_capability >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
+    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:

ggml/src/ggml-cuda/mmq.cuh

@@ -2742,6 +2742,7 @@ struct mmq_args {
     int64_t ne00; int64_t ne01; int64_t stride01;
     int64_t ne10; int64_t ne11; int64_t stride11;
     int64_t ne0;
+    bool use_stream_k;
 };
 
 template<ggml_type type>
@@ -2777,8 +2778,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
     const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
     const dim3 block_nums_xy_tiling(nty, ntx, 1);
 
-    const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
-    if (!use_stream_k) {
+    if (!args.use_stream_k) {
         if (args.ne01 % mmq_y == 0) {
             constexpr bool need_check = false;
             mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>

ggml/src/ggml-cuda/sum.cu

@@ -0,0 +1,43 @@
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+// On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh.
+// For this reason CUB must be included BEFORE anything else.
+#include <cub/cub.cuh>
+using namespace cub;
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+
+#include "sumrows.cuh"
+#include "sum.cuh"
+
+#include <cstdint>
+
+void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) {
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+    size_t tmp_size = 0;
+    DeviceReduce::Sum(nullptr,       tmp_size, x, dst, ne, stream);
+    ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
+    DeviceReduce::Sum(tmp_alloc.ptr, tmp_size, x, dst, ne, stream);
+#else
+    // Use (inefficient) sum_rows implementation as a fallback.
+    // For AMD there is rocPRIM which could be used as a drop-in replacement via hipcub but this would require C++11 -> C++14.
+    sum_rows_f32_cuda(x, dst, ne, 1, stream);
+    GGML_UNUSED(pool);
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+}
+
+void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const float * src0_d = (const float *) src0->data;
+    float * dst_d = (float *) dst->data;
+
+    const int64_t ne = ggml_nelements(src0);
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t stream = ctx.stream();
+
+    sum_f32_cuda(pool, src0_d, dst_d, ne, stream);
+}

ggml/src/ggml-cuda/sum.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream);
+
+void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/unary.cu

@@ -1,5 +1,15 @@
 #include "unary.cuh"
 
+static __global__ void neg_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = -x[i];
+}
+
 static __global__ void gelu_f32(const float * x, float * dst, const int k) {
     const float GELU_COEF_A    = 0.044715f;
     const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
@@ -119,6 +129,11 @@ static __global__ void cos_f32(const float * x, float * dst, const int k) {
     dst[i] = cosf(x[i]);
 }
 
+static void neg_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE;
+    neg_f32<<<num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
     gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -184,6 +199,20 @@ static void cos_f32_cuda(const float * x, float * dst, const int k, cudaStream_t
     cos_f32<<<num_blocks, CUDA_COS_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    neg_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;

ggml/src/ggml-cuda/unary.cuh

@@ -1,5 +1,6 @@
 #include "common.cuh"
 
+#define CUDA_NEG_BLOCK_SIZE 256
 #define CUDA_GELU_BLOCK_SIZE 256
 #define CUDA_SILU_BLOCK_SIZE 256
 #define CUDA_TANH_BLOCK_SIZE 256
@@ -12,6 +13,8 @@
 #define CUDA_SIN_BLOCK_SIZE 256
 #define CUDA_COS_BLOCK_SIZE 256
 
+void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/vendors/musa.h

@@ -130,42 +130,3 @@
 #define cudaKernelNodeParams musaKernelNodeParams
 #define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
 #define cudaStreamEndCapture musaStreamEndCapture
-
-// XXX: Clang builtins mapping
-#define __vsub4   __vsub4_musa
-#define __vcmpeq4 __vcmpeq4_musa
-#define __vcmpne4 __vcmpne4_musa
-
-#ifndef __has_builtin
-    #define __has_builtin(x) 0
-#endif
-
-typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
-
-static __device__ __forceinline__ int __vsub4_musa(const int a, const int b) {
-    return __vsubss4(a, b);
-}
-
-static __device__ __forceinline__ unsigned int __vcmpeq4_musa(unsigned int a, unsigned int b) {
-    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
-    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
-    unsigned int c;
-    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
-#pragma unroll
-    for (int i = 0; i < 4; ++i) {
-        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
-    }
-    return c;
-}
-
-static __device__ __forceinline__ unsigned int __vcmpne4_musa(unsigned int a, unsigned int b) {
-    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
-    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
-    unsigned int c;
-    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
-#pragma unroll
-    for (int i = 0; i < 4; ++i) {
-        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
-    }
-    return c;
-}

ggml/src/ggml-impl.h

@@ -629,8 +629,16 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 #endif
 
+enum ggml_cgraph_eval_order {
+    GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
+    GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
+    GGML_CGRAPH_EVAL_ORDER_COUNT
+};
+
 // bitset
 
+typedef uint32_t ggml_bitset_t;
+
 static_assert(sizeof(ggml_bitset_t) == 4, "bitset_t constants must be updated");
 #define BITSET_SHR 5 // log2(sizeof(ggml_bitset_t)*8)
 #define BITSET_MASK (sizeof(ggml_bitset_t)*8 - 1)
@@ -656,6 +664,12 @@ static inline void ggml_bitset_clear(ggml_bitset_t * bitset, size_t i) {
 #define GGML_HASHSET_FULL ((size_t)-1)
 #define GGML_HASHSET_ALREADY_EXISTS ((size_t)-2)
 
+struct ggml_hash_set {
+    size_t size;
+    ggml_bitset_t * used;       // whether or not the keys are in use i.e. set
+    struct ggml_tensor ** keys; // actual tensors in the set, keys[i] is only defined if ggml_bitset_get(used, i)
+};
+
 struct ggml_hash_set ggml_hash_set_new(size_t size);
 void                 ggml_hash_set_free(struct ggml_hash_set * hash_set);
 
@@ -745,6 +759,24 @@ static size_t ggml_hash_find_or_insert(struct ggml_hash_set * hash_set, struct g
     GGML_ABORT("fatal error");
 }
 
+// computation graph
+
+struct ggml_cgraph {
+    int size;
+    int n_nodes;
+    int n_leafs;
+
+    struct ggml_tensor ** nodes;
+    struct ggml_tensor ** grads;
+    struct ggml_tensor ** leafs;
+
+    struct ggml_hash_set visited_hash_set;
+
+    enum ggml_cgraph_eval_order order;
+};
+
+struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-kompute.cpp

@@ -1,4 +1,4 @@
-#include "ggml.h"
+#include "ggml-impl.h"
 #include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-kompute.h"

ggml/src/ggml-metal.m

@@ -1,7 +1,7 @@
 #import "ggml-metal.h"
 
+#import "ggml-impl.h"
 #import "ggml-backend-impl.h"
-#import "ggml.h"
 
 #import <Foundation/Foundation.h>
 
@@ -17,8 +17,8 @@
 #define GGML_METAL_LOG_WARN(...)
 #define GGML_METAL_LOG_ERROR(...)
 #else
-#define GGML_METAL_LOG_INFO(...)  ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
-#define GGML_METAL_LOG_WARN(...)  ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
+#define GGML_METAL_LOG_INFO(...)  ggml_metal_log(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
+#define GGML_METAL_LOG_WARN(...)  ggml_metal_log(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
 #define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
 #endif
 
@@ -799,8 +799,9 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
             return ctx->support_simdgroup_reduction;
         case GGML_OP_NORM:
         case GGML_OP_ROPE:
-        case GGML_OP_IM2COL:
             return true;
+        case GGML_OP_IM2COL:
+            return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_1D:
         case GGML_OP_POOL_2D:
             return false;
@@ -881,7 +882,7 @@ static enum ggml_status ggml_metal_graph_compute(
     // create multiple command buffers and enqueue them
     // then, we encode the graph into the command buffers in parallel
 
-    const int n_nodes  = gf->n_nodes;
+    const int n_nodes = gf->n_nodes;
     const int n_cb = ctx->n_cb;
     const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
 
@@ -3038,8 +3039,7 @@ static enum ggml_status ggml_metal_graph_compute(
         if (status != MTLCommandBufferStatusCompleted) {
             GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
             if (status == MTLCommandBufferStatusError) {
-                NSString * error_code = [command_buffer error].localizedDescription;
-                GGML_METAL_LOG_INFO("error: %s\n", [error_code UTF8String]);
+                GGML_METAL_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]);
             }
 
             return GGML_STATUS_FAILED;

ggml/src/ggml-quants.c

@@ -4003,42 +4003,141 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     float sumf = 0;
 
 #if defined(__ARM_FEATURE_SVE)
-    if (ggml_sve_cnt_b == QK8_0) {
-        const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
-        const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
-        svfloat32_t sumv0 = svdup_n_f32(0.0f);
-        svfloat32_t sumv1 = svdup_n_f32(0.0f);
+    const int vector_length = ggml_sve_cnt_b*8;
 
-        for (; ib + 1 < nb; ib += 2) {
-            const block_q4_0 * restrict x0 = &x[ib + 0];
-            const block_q4_0 * restrict x1 = &x[ib + 1];
-            const block_q8_0 * restrict y0 = &y[ib + 0];
-            const block_q8_0 * restrict y1 = &y[ib + 1];
+    // VLA Implementation using switch case
+    switch (vector_length) {
+        case 128:
+            {
+                // predicate for activating higher lanes for 4 float32 elements
+                const svbool_t ph4 = svptrue_pat_b32(SV_VL4);
 
-            // load x
-            const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
-            const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
 
-            // 4-bit -> 8-bit
-            const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04));
-            const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04));
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
 
-            // sub 8
-            const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
-            const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx0r, 0x0F));
+                    const svint8_t qx0h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx0r, 0x04));
+                    const svint8_t qx1l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx1r, 0x0F));
+                    const svint8_t qx1h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx1r, 0x04));
 
-            // load y
-            const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
-            const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+                    // sub 8
+                    const svint8_t qx0ls = svsub_n_s8_x(svptrue_b8(), qx0h, 8);
+                    const svint8_t qx0hs = svsub_n_s8_x(svptrue_b8(), qx0l, 8);
+                    const svint8_t qx1ls = svsub_n_s8_x(svptrue_b8(), qx1h, 8);
+                    const svint8_t qx1hs = svsub_n_s8_x(svptrue_b8(), qx1l, 8);
 
-            // dot product
-            sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-            sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-        }
+                    // load y
+                    const svint8_t qy0h = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy0l = svld1_s8(svptrue_b8(), y0->qs + 16);
+                    const svint8_t qy1h = svld1_s8(svptrue_b8(), y1->qs);
+                    const svint8_t qy1l = svld1_s8(svptrue_b8(), y1->qs + 16);
 
-        sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+                    // dot product
+                    sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
+                                    svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
+                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
+                                    svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
+                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 256:
+            {
+                // predicate for activating higher lanes for 16 int8 elements
+                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                // predicate for activating lower lanes for  16 int8 elements
+                const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
+                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
+
+                    // sub 8
+                    const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
+                    const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 512:
+            {
+                // predicate for activating higher lanes for 32 int8 elements
+                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
+
+                // predicate for activating higher lanes for 16 int8 elements
+                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                // predicate for activating lower lanes for 16 int8 elements from first 32 int8 activated lanes
+                const svbool_t pl16 = svnot_b_z(ph32, ph16);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(ph32, x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(ph32, x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
+                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
+
+                    // sub 8
+                    const svint8_t qx0s = svsub_n_s8_x(ph32, qx0, 8);
+                    const svint8_t qx1s = svsub_n_s8_x(ph32, qx1, 8);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(ph32, y0->qs);
+                    const svint8_t qy1 = svld1_s8(ph32, y1->qs);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
+            } break;
+        default:
+            assert(false && "Unsupported vector length");
+            break;
     }
+
 #elif defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
@@ -5488,29 +5587,124 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     float sumf = 0;
 
 #if defined(__ARM_FEATURE_SVE)
-    if (ggml_sve_cnt_b == QK8_0) {
-        svfloat32_t sumv0 = svdup_n_f32(0.0f);
-        svfloat32_t sumv1 = svdup_n_f32(0.0f);
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
-        for (; ib + 1 < nb; ib += 2) {
-            const block_q8_0 * restrict x0 = &x[ib + 0];
-            const block_q8_0 * restrict x1 = &x[ib + 1];
-            const block_q8_0 * restrict y0 = &y[ib + 0];
-            const block_q8_0 * restrict y1 = &y[ib + 1];
+    const int vector_length = ggml_sve_cnt_b*8;
 
-            // load x
-            const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
-            const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
+    //VLA Implemenation for SVE
+    switch (vector_length) {
+        case 128:
+            {
+                // predicate for activating lanes for 16 Int8 elements
+                const svbool_t ph16 = svptrue_pat_b8 (SV_VL16);
+                const svbool_t pl16 = svptrue_pat_b32(SV_VL4);
 
-            // load y
-            const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
-            const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
 
-            sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-            sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-        }
+                    // load x
+                    const svint8_t qx0_0 = svld1_s8(ph16, x0->qs);
+                    const svint8_t qx0_1 = svld1_s8(ph16, x0->qs+16);
+                    const svint8_t qx1_0 = svld1_s8(ph16, x1->qs);
+                    const svint8_t qx1_1 = svld1_s8(ph16, x1->qs+16);
 
-        sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+                    // load y
+                    const svint8_t qy0_0 = svld1_s8(ph16, y0->qs);
+                    const svint8_t qy0_1 = svld1_s8(ph16, y0->qs+16);
+                    const svint8_t qy1_0 = svld1_s8(ph16, y1->qs);
+                    const svint8_t qy1_1 = svld1_s8(ph16, y1->qs+16);
+
+                    sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
+                                    svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
+                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
+                                    svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
+                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
+            } break;
+        case 256:
+            {
+                //printf("sve256");
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
+                    const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 512:
+            {
+                // predicate for activating high 256 bit
+                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
+                // predicate for activating low 256 bit
+                const svbool_t pl32 = svnot_b_z(svptrue_b8(), ph32);
+
+                // predicate for activating high lanes for 8 float32 elements
+                const svbool_t ph8 = svptrue_pat_b32(SV_VL8);
+                // predicate for activating low lanes for 8 float32 elements
+                const svbool_t pl8 = svnot_b_z(svptrue_b32(), ph8);
+
+                svfloat32_t sumv00 = svdup_n_f32(0.0f);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    //load 32 int8_t in first half of vector and put another 32 int8_t in second vector lower bits
+                    // and add them to make one 64 element vector
+                    // load x
+                    const svint8_t qx_32 = svld1_s8(ph32, x0->qs);
+                          svint8_t qx_64 = svld1_s8(pl32, x0->qs + 2);
+
+                    qx_64 = svadd_s8_x(svptrue_b8(), qx_32, qx_64);
+
+                    // load y
+                    const svint8_t qy_32 = svld1_s8(ph32, y0->qs);
+                          svint8_t qy_64 = svld1_s8(pl32, y0->qs + 2);
+
+                    qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
+
+                    // scale creation
+                    const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
+                    const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
+
+                    // duplicate deq1 in first half of vector and deq2 in second half of vector
+                    const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
+
+                    const svfloat32_t sumvt = svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx_64, qy_64));
+
+                    sumv00 = svmla_f32_m(svptrue_b32(), sumv00, sumvt, temp);
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), sumv00);
+                break;
+            }
+        default:
+            assert(false && "Unsupported vector length");
+            break;
     }
 #elif defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);

ggml/src/ggml-rpc.cpp

@@ -1,5 +1,5 @@
 #include "ggml-rpc.h"
-#include "ggml.h"
+#include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 
 #include <cinttypes>
@@ -883,15 +883,17 @@ ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rp
     }
     result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
     if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
-        return nullptr;
+        result->buffer = nullptr;
     }
 
-    // require that the tensor data does not go beyond the buffer end
-    uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
-    uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
-    uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
-    GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
-    GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
+    if (result->buffer) {
+        // require that the tensor data does not go beyond the buffer end
+        uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
+        uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
+        uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
+        GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
+        GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
+    }
 
     result->op = (ggml_op) tensor->op;
     for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
@@ -1060,7 +1062,7 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, std::vector<u
     const rpc_tensor * tensors = (const rpc_tensor *)(input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t) + sizeof(n_tensors));
     GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
 
-    static size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
+    size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
     struct ggml_init_params params = {
         /*.mem_size   =*/ buf_size,
         /*.mem_buffer =*/ NULL,

ggml/src/ggml-sycl.cpp

@@ -33,7 +33,7 @@
 #include <sycl/half_type.hpp>
 
 #include "ggml-sycl.h"
-#include "ggml.h"
+#include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 
 #include "ggml-sycl/backend.hpp"
@@ -1954,6 +1954,11 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
         SYCL_CHECK(
             CHECK_TRY_ERROR(ptr = (void *)sycl::malloc_device(
                                 look_ahead_size, *qptr)));
+        if (!ptr) {
+            fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, look_ahead_size);
+            return nullptr;
+        }
+
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
 
@@ -4350,6 +4355,10 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
     void * dev_ptr;
     SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device(
                                     size, *stream)));
+    if (!dev_ptr) {
+        fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, size);
+        return nullptr;
+    }
     ggml_backend_sycl_buffer_context * ctx = new  ggml_backend_sycl_buffer_context(buft_ctx->device, dev_ptr, buft_ctx->stream);
     return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size);
 }
@@ -4570,7 +4579,11 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
         */
         SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)sycl::malloc_device(
                                         size, *stream)));
-
+        if (!buf) {
+            char err_buf[1024];
+            snprintf(err_buf, 1023, "%s: can't malloc %lu Bytes memory on device", __func__, size);
+            throw std::runtime_error(err_buf);
+        }
         // set padding to 0 to avoid possible NaN values
         if (size > original_size) {
             /*
@@ -5124,13 +5137,17 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_CLAMP:
+            return true;
         case GGML_OP_CONT:
+            return op->src[0]->type != GGML_TYPE_BF16;
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_ROPE:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_IM2COL:
+            // TODO: add support for the new F32 operations
+            return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:

ggml/src/ggml-vulkan.cpp

@@ -21,7 +21,7 @@
 #include <memory>
 #include <mutex>
 
-#include "ggml.h"
+#include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 
 #include "ggml-vulkan-shaders.hpp"
@@ -787,6 +787,9 @@ static vk_submission ggml_vk_create_submission(vk_device& device, vk_queue& q, s
 
 static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
     if (ctx->seqs.empty()) {
+        if (fence) {
+            ctx->q->queue.submit({}, fence);
+        }
         return;
     }
     VK_LOG_DEBUG("ggml_vk_submit(" << ctx << ", " << fence << ")");
@@ -4616,7 +4619,7 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
-static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
@@ -4626,10 +4629,10 @@ static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
-    });
+    }, dryrun);
 }
 
-static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
@@ -4639,7 +4642,7 @@ static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
-    });
+    }, dryrun);
 }
 
 static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -5658,11 +5661,15 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
     }
 }
 
-static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, bool last_node, bool dryrun){
+static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_tensor* tensor, int tensor_idx, bool use_fence);
+
+// Returns true if node has enqueued work into the queue, false otherwise
+// If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution.
+static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool submit){
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
 
     if (ggml_is_empty(node) || extra == nullptr) {
-        return;
+        return false;
     }
 
     VK_LOG_DEBUG("ggml_vk_build_graph(" << node << ", " << ggml_op_name(node->op) << ")");
@@ -5679,7 +5686,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_PERMUTE:
     case GGML_OP_TRANSPOSE:
     case GGML_OP_NONE:
-        return;
+        return false;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
         case GGML_UNARY_OP_SILU:
@@ -5689,7 +5696,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_UNARY_OP_TANH:
             break;
         default:
-            return;
+            return false;
         }
         break;
     case GGML_OP_REPEAT:
@@ -5726,7 +5733,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     default:
         std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
         GGML_ABORT("fatal error");
-        return;
+        return false;
     }
 
     vk_context compute_ctx;
@@ -5783,11 +5790,11 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
         break;
     case GGML_OP_SIN:
-        ggml_vk_sin(ctx, compute_ctx, src0, node);
+        ggml_vk_sin(ctx, compute_ctx, src0, node, dryrun);
 
         break;
     case GGML_OP_COS:
-        ggml_vk_cos(ctx, compute_ctx, src0, node);
+        ggml_vk_cos(ctx, compute_ctx, src0, node, dryrun);
 
         break;
     case GGML_OP_CLAMP:
@@ -5826,7 +5833,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
             ggml_vk_unary(ctx, compute_ctx, src0, node, dryrun);
             break;
         default:
-            return;
+            return false;
         }
         break;
     case GGML_OP_DIAG_MASK_INF:
@@ -5870,11 +5877,11 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
         break;
     default:
-        return;
+        return false;
     }
 
     if (dryrun) {
-        return;
+        return false;
     }
 
     ctx->tensor_ctxs[node_idx] = compute_ctx;
@@ -5885,14 +5892,34 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     last_node = true;
 #endif
 
-    if (last_node) {
+    if (submit || last_node) {
         ggml_vk_ctx_end(compute_ctx);
-        compute_ctx->exit_tensor_idx = node_idx;
+
+        // TODO probably it'd be better to pass a exit_node flag to ggml_vk_compute_forward
+        if (last_node) {
+            compute_ctx->exit_tensor_idx = node_idx_begin;
+        }
+        else {
+            compute_ctx->exit_tensor_idx = -1;
+        }
+
         ctx->compute_ctx.reset();
+
+        bool ok = ggml_vk_compute_forward(ctx, node_begin, node_idx_begin, false);
+        if (!ok) {
+            if (node->op == GGML_OP_UNARY) {
+                std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl;
+            }
+            else {
+                std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl;
+            }
+        }
+
     }
+    return true;
 }
 
-static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx){
+static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx, bool use_fence = true){
     ggml_tensor_extra_gpu * extra = nullptr;
 
     switch (tensor->op) {
@@ -5960,40 +5987,38 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
 
     VK_LOG_DEBUG("ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")");
 
-#ifdef GGML_VULKAN_CHECK_RESULTS
-    ggml_vk_check_results_0(tensor);
-#endif
-
     vk_context subctx = ctx->tensor_ctxs[tensor_idx].lock();
 
-#ifdef GGML_VULKAN_PERF
-    std::chrono::steady_clock::time_point start;
-#endif // GGML_VULKAN_PERF
+    // always wait for the GPU work to be done for the last submit
+    if (tensor_idx == subctx->exit_tensor_idx) {
+        use_fence = true;
+    }
 
     // Only run if ctx hasn't been submitted yet
     if (!subctx->seqs.empty()) {
+#ifdef GGML_VULKAN_CHECK_RESULTS
+        ggml_vk_check_results_0(tensor);
+        use_fence = true;
+#endif
+
         // Do staging buffer copies
         for (auto& cpy : subctx->in_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
 
-#ifdef GGML_VULKAN_PERF
-        start = std::chrono::steady_clock::now();
-#endif // GGML_VULKAN_PERF
+        ggml_vk_submit(subctx, use_fence ? ctx->fence : vk::Fence{});
 
-        ggml_vk_submit(subctx, ctx->fence);
+        if (use_fence) {
+            VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences");
+
+            ctx->device->device.resetFences({ ctx->fence });
+        }
+#ifdef GGML_VULKAN_CHECK_RESULTS
+        ggml_vk_check_results_1(tensor);
+#endif
     }
 
     if (tensor_idx == subctx->exit_tensor_idx) {
-        VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences");
-
-#ifdef GGML_VULKAN_PERF
-        auto duration = std::chrono::duration_cast<std::chrono::nanoseconds>(std::chrono::steady_clock::now() - start);
-        ctx->device->perf_logger->log_timing(tensor, duration.count());
-#endif // GGML_VULKAN_PERF
-
-        ctx->device->device.resetFences({ ctx->fence });
-
         // Do staging buffer copies
         for (auto& cpy : subctx->out_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
@@ -6482,7 +6507,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_vk_build_graph(ctx, cgraph->nodes[i], i, 0, true);
+        ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false);
     }
     ggml_vk_preallocate_buffers(ctx);
     ggml_pipeline_allocate_descriptor_sets(ctx->device);
@@ -6497,31 +6522,36 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
     // Reserve tensor context space for all nodes
     ctx->tensor_ctxs.resize(cgraph->n_nodes);
 
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_vk_build_graph(ctx, cgraph->nodes[i], i, i == last_node, false);
-    }
+    bool first_node_in_batch = true; // true if next node will be first node in a batch
+    int submit_node_idx = 0; // index to first node in a batch
 
+    // submit work every submit_count node to overlap CPU cmdbuffer generation with GPU execution
+    constexpr int submit_count = 100;
+    int submitted_nodes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_tensor * node = cgraph->nodes[i];
-
-        if (ggml_vk_is_empty(node)) {
-            continue;
+        if (first_node_in_batch) {
+            submit_node_idx = i;
         }
 
-        bool ok = ggml_vk_compute_forward(ctx, node, i);
-        if (!ok) {
-            if (node->op == GGML_OP_UNARY) {
-                std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl;
-            } else {
-                std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl;
+        bool submit = (submitted_nodes >= submit_count) || (i == last_node);
+
+
+        bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit);
+
+        if (enqueued) {
+            ++submitted_nodes;
+
+#ifndef GGML_VULKAN_CHECK_RESULTS
+            if (first_node_in_batch) {
+                first_node_in_batch = false;
             }
-        }
-#ifdef GGML_VULKAN_CHECK_RESULTS
-        else {
-            ggml_vk_check_results_1(node);
-        }
 #endif
-        GGML_ASSERT(ok);
+        }
+
+        if (submit) {
+            first_node_in_batch = true;
+            submitted_nodes = 0;
+        }
     }
 
 #ifdef GGML_VULKAN_PERF
@@ -6602,6 +6632,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
                         return false;
                 }
             } break;
+        case GGML_OP_CONT:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
             {
@@ -6642,7 +6673,6 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
         case GGML_OP_PAD:
-        case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_ARGSORT:

ggml/src/ggml.c

@@ -287,6 +287,7 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
 #define GGML_DEBUG 0
 #define GGML_GELU_FP16
 #define GGML_GELU_QUICK_FP16
+#define GGML_N_TASKS_MAX (-1)
 
 #define GGML_SOFT_MAX_UNROLL 4
 #define GGML_VEC_DOT_UNROLL  2
@@ -1120,21 +1121,21 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
 #define GGML_F32x4_ADD          vaddq_f32
 #define GGML_F32x4_MUL          vmulq_f32
 #define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x)
-#define GGML_F32x4_REDUCE(res, x)              \
-{                                              \
-    int offset = GGML_F32_ARR >> 1;            \
-    for (int i = 0; i < offset; ++i) {         \
-        x[i] = vaddq_f32(x[i], x[offset+i]);   \
-    }                                          \
-    offset >>= 1;                              \
-    for (int i = 0; i < offset; ++i) {         \
-        x[i] = vaddq_f32(x[i], x[offset+i]);   \
-    }                                          \
-    offset >>= 1;                              \
-    for (int i = 0; i < offset; ++i) {         \
-        x[i] = vaddq_f32(x[i], x[offset+i]);   \
-    }                                          \
-    res = GGML_F32x4_REDUCE_ONE(x[0]);         \
+#define GGML_F32x4_REDUCE(res, x)                  \
+{                                                  \
+    int offset = GGML_F32_ARR >> 1;                \
+    for (int i = 0; i < offset; ++i) {             \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+    }                                              \
+    offset >>= 1;                                  \
+    for (int i = 0; i < offset; ++i) {             \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+    }                                              \
+    offset >>= 1;                                  \
+    for (int i = 0; i < offset; ++i) {             \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+    }                                              \
+    (res) = GGML_F32x4_REDUCE_ONE((x)[0]);         \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -1161,30 +1162,30 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
     #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
     #define GGML_F16x8_ADD          vaddq_f16
     #define GGML_F16x8_MUL          vmulq_f16
-    #define GGML_F16x8_REDUCE(res, x)                             \
-    do {                                                          \
-        int offset = GGML_F16_ARR >> 1;                           \
-        for (int i = 0; i < offset; ++i) {                        \
-            x[i] = vaddq_f16(x[i], x[offset+i]);                  \
-        }                                                         \
-        offset >>= 1;                                             \
-        for (int i = 0; i < offset; ++i) {                        \
-            x[i] = vaddq_f16(x[i], x[offset+i]);                  \
-        }                                                         \
-        offset >>= 1;                                             \
-        for (int i = 0; i < offset; ++i) {                        \
-            x[i] = vaddq_f16(x[i], x[offset+i]);                  \
-        }                                                         \
-        const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 (x[0])); \
-        const float32x4_t t1 = vcvt_f32_f16(vget_high_f16(x[0])); \
-        res = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
+    #define GGML_F16x8_REDUCE(res, x)                               \
+    do {                                                            \
+        int offset = GGML_F16_ARR >> 1;                             \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        offset >>= 1;                                               \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        offset >>= 1;                                               \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
+        const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
+        (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
     } while (0)
 
     #define GGML_F16_VEC                GGML_F16x8
     #define GGML_F16_VEC_ZERO           GGML_F16x8_ZERO
     #define GGML_F16_VEC_SET1           GGML_F16x8_SET1
     #define GGML_F16_VEC_LOAD(p, i)     GGML_F16x8_LOAD(p)
-    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), r[i])
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), (r)[i])
     #define GGML_F16_VEC_FMA            GGML_F16x8_FMA
     #define GGML_F16_VEC_ADD            GGML_F16x8_ADD
     #define GGML_F16_VEC_MUL            GGML_F16x8_MUL
@@ -1893,6 +1894,23 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
 #endif
 
+//
+// ggml object
+//
+
+struct ggml_object {
+    size_t offs;
+    size_t size;
+
+    struct ggml_object * next;
+
+    enum ggml_object_type type;
+
+    char padding[4];
+};
+
+static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
+
 //
 // ggml context
 //
@@ -3847,7 +3865,7 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
 
     if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
         GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
-                __func__, cur_end + size_needed, ctx->mem_size);
+                __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
         assert(false);
         return NULL;
     }
@@ -5267,6 +5285,7 @@ struct ggml_tensor * ggml_concat(
     bool is_node = false;
 
     if (a->grad || b->grad) {
+        GGML_ABORT("fatal error"); // TODO: implement
         is_node = true;
     }
 
@@ -5388,6 +5407,7 @@ struct ggml_tensor * ggml_leaky_relu(
     bool is_node = false;
 
     if (!inplace && (a->grad)) {
+        GGML_ABORT("fatal error"); // TODO: not implemented
         is_node = true;
     }
 
@@ -5826,6 +5846,7 @@ static struct ggml_tensor * ggml_set_impl(
     // make a view of the destination
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
+    GGML_ASSERT(offset < (size_t)(1 << 30));
     int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 };
     ggml_set_op_params(result, params, sizeof(params));
 
@@ -6783,14 +6804,12 @@ struct ggml_tensor * ggml_rope_back(
     GGML_ASSERT(ggml_is_vector(b));
     GGML_ASSERT(b->type == GGML_TYPE_I32);
     GGML_ASSERT(a->ne[2] == b->ne[0]);
-    GGML_ASSERT(c == NULL && "freq factors not implemented yet");
-
-    GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet");
 
     bool is_node = false;
 
     if (a->grad) {
-        is_node = false; // TODO: implement backward
+        GGML_ASSERT(false && "backwards pass not implemented");
+        is_node = false;
     }
 
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
@@ -6808,6 +6827,7 @@ struct ggml_tensor * ggml_rope_back(
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
+    result->src[2] = c;
 
     return result;
 }
@@ -7361,6 +7381,11 @@ struct ggml_tensor * ggml_argsort(
         enum ggml_sort_order  order) {
     bool is_node = false;
 
+    if (a->grad) {
+        GGML_ABORT("fatal error"); // TODO: not implemented
+        is_node = true;
+    }
+
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
@@ -8322,8 +8347,7 @@ static void ggml_compute_forward_dup_same_cont(
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
     GGML_ASSERT(src0->type == dst->type);
 
-    const size_t nb00 = src0->nb[0];
-    const size_t nb0 = dst->nb[0];
+    const size_t nb0 = ggml_type_size(src0->type);
 
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
@@ -8337,8 +8361,8 @@ static void ggml_compute_forward_dup_same_cont(
     if (ie0 < ie1) {
         memcpy(
             ((char *)  dst->data + ie0*nb0),
-            ((char *) src0->data + ie0*nb00),
-            (ie1 - ie0) * ggml_type_size(src0->type));
+            ((char *) src0->data + ie0*nb0),
+            (ie1 - ie0) * nb0);
     }
 }
 
@@ -8355,11 +8379,6 @@ static void ggml_compute_forward_dup_f16(
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) {
-        ggml_compute_forward_dup_same_cont(params, dst);
-        return;
-    }
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -8624,11 +8643,6 @@ static void ggml_compute_forward_dup_bf16(
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) {
-        ggml_compute_forward_dup_same_cont(params, dst);
-        return;
-    }
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -8980,11 +8994,6 @@ static void ggml_compute_forward_dup_f32(
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) {
-        ggml_compute_forward_dup_same_cont(params, dst);
-        return;
-    }
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -9294,13 +9303,13 @@ static void ggml_compute_forward_dup_bytes(
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
     GGML_ASSERT(src0->type == dst->type);
 
+    GGML_TENSOR_UNARY_OP_LOCALS;
+
     if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
         ggml_compute_forward_dup_same_cont(params, dst);
         return;
     }
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
-
     const size_t type_size = ggml_type_size(src0->type);
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
@@ -10969,9 +10978,6 @@ static void ggml_compute_forward_sum_f32(
         return;
     }
 
-    assert(ggml_is_scalar(dst));
-
-
     assert(ggml_is_scalar(dst));
     assert(src0->nb[0] == sizeof(float));
 
@@ -13721,7 +13727,7 @@ static void ggml_compute_forward_get_rows_q(
         const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
         const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
 
-        assert(i01 >= 0 && i01 < ne01);
+        GGML_ASSERT(i01 >= 0 && i01 < ne01);
 
         dequantize_row_q(
                 (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
@@ -13762,7 +13768,7 @@ static void ggml_compute_forward_get_rows_f16(
         const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
         const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
 
-        assert(i01 >= 0 && i01 < ne01);
+        GGML_ASSERT(i01 >= 0 && i01 < ne01);
 
         ggml_fp16_to_fp32_row(
                 (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
@@ -13803,7 +13809,7 @@ static void ggml_compute_forward_get_rows_bf16(
         const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
         const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
 
-        assert(i01 >= 0 && i01 < ne01);
+        GGML_ASSERT(i01 >= 0 && i01 < ne01);
 
         ggml_bf16_to_fp32_row(
                 (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
@@ -13844,7 +13850,7 @@ static void ggml_compute_forward_get_rows_f32(
         const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
         const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
 
-        assert(i01 >= 0 && i01 < ne01);
+        GGML_ASSERT(i01 >= 0 && i01 < ne01);
 
         ggml_vec_cpy_f32(nc,
                 (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
@@ -18372,14 +18378,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 if (src0->grad || src1->grad) {
                     GGML_ASSERT(src0->type == tensor->type);
                     GGML_ASSERT(tensor->grad->type == tensor->type);
-                    GGML_ASSERT(tensor->grad->type == src1->grad->type);
+                    GGML_ASSERT(!src1->grad || src1->grad->type == tensor->grad->type);
 
                     tensor_grad_view = ggml_view_4d(ctx,
-                        tensor->grad,
-                        src1->grad->ne[0],
-                        src1->grad->ne[1],
-                        src1->grad->ne[2],
-                        src1->grad->ne[3],
+                        tensor->grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
                         nb1, nb2, nb3, offset);
                 }
 
@@ -18448,9 +18450,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
 
                     memcpy(&offset, tensor->op_params, sizeof(offset));
 
-                    size_t nb1     = tensor->nb[1];
-                    size_t nb2     = tensor->nb[2];
-                    size_t nb3     = tensor->nb[3];
+                    size_t nb1 = tensor->nb[1];
+                    size_t nb2 = tensor->nb[2];
+                    size_t nb3 = tensor->nb[3];
 
                     if (src0->type != src0->grad->type) {
                         // gradient is typically F32, but src0 could be other type
@@ -19146,7 +19148,8 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
     }
 
     for (size_t i = 0; i < src->visited_hash_set.size; ++i) {
-        if (src->visited_hash_set.keys[i]) {
+        // copy all hashset keys (tensors) that are in use
+        if (ggml_bitset_get(src->visited_hash_set.used, i)) {
             ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
         }
     }
@@ -19176,6 +19179,34 @@ void ggml_graph_clear(struct ggml_cgraph * cgraph) {
     ggml_hash_set_reset(&cgraph->visited_hash_set);
 }
 
+int ggml_graph_size(struct ggml_cgraph * cgraph) {
+    return cgraph->size;
+}
+
+struct ggml_tensor * ggml_graph_node(struct ggml_cgraph * cgraph, int i) {
+    if (i < 0) {
+        GGML_ASSERT(cgraph->n_nodes + i >= 0);
+        return cgraph->nodes[cgraph->n_nodes + i];
+    }
+
+    GGML_ASSERT(i < cgraph->n_nodes);
+    return cgraph->nodes[i];
+}
+
+struct ggml_tensor ** ggml_graph_nodes(struct ggml_cgraph * cgraph) {
+    return cgraph->nodes;
+}
+
+int ggml_graph_n_nodes(struct ggml_cgraph * cgraph) {
+    return cgraph->n_nodes;
+}
+
+void ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
+    GGML_ASSERT(cgraph->size > cgraph->n_nodes);
+    cgraph->nodes[cgraph->n_nodes] = tensor;
+    cgraph->n_nodes++;
+}
+
 // Android's libc implementation "bionic" does not support setting affinity
 #if defined(__gnu_linux__)
 static void set_numa_thread_affinity(int thread_n) {
@@ -19556,7 +19587,8 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     return true;
 }
 
-#else // posix?
+#elif defined(__gnu_linux__)
+// TODO: this may not work on BSD, to be verified
 
 static bool ggml_thread_apply_affinity(const bool * mask) {
     cpu_set_t cpuset;
@@ -19611,6 +19643,18 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     return true;
 }
 
+#else // unsupported platforms
+
+static bool ggml_thread_apply_affinity(const bool * mask) {
+    UNUSED(mask);
+    return true;
+}
+
+static bool ggml_thread_apply_priority(int32_t prio) {
+    UNUSED(prio);
+    return true;
+}
+
 #endif
 
 static bool ggml_thread_cpumask_is_valid(const bool * mask) {
@@ -23244,6 +23288,14 @@ int ggml_cpu_has_arm_fma(void) {
 #endif
 }
 
+int ggml_cpu_has_riscv_v(void) {
+#if defined(__riscv_v_intrinsic)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_metal(void) {
 #if defined(GGML_USE_METAL)
     return 1;

ggml/src/llamafile/sgemm.cpp

@@ -1006,6 +1006,10 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
     assert(nth > 0);
     assert(ith < nth);
 
+    // only enable sgemm for prompt processing
+    if (n < 2)
+        return false;
+
     if (Ctype != GGML_TYPE_F32)
         return false;
 

include/llama.h

@@ -33,12 +33,15 @@
 
 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
 
+// TODO: use everywhere in the implementation
+#define LLAMA_TOKEN_NULL -1
+
 #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 #define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 8
+#define LLAMA_SESSION_VERSION 9
 
 #define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
 #define LLAMA_STATE_SEQ_VERSION 2
@@ -53,8 +56,10 @@ extern "C" {
     // TODO: show sample usage
     //
 
+    // struct llama_vocab; // TODO: add in the future
     struct llama_model;
     struct llama_context;
+    struct llama_sampler;
 
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
@@ -201,6 +206,7 @@ extern "C" {
         LLAMA_SPLIT_MODE_ROW     = 2, // split rows across GPUs
     };
 
+    // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
     typedef struct llama_token_data {
         llama_token id; // token id
         float logit;    // log-odds of the token
@@ -208,8 +214,10 @@ extern "C" {
     } llama_token_data;
 
     typedef struct llama_token_data_array {
+        // TODO: consider SoA
         llama_token_data * data;
         size_t size;
+        int64_t selected; // this is the index in the data array (i.e. not the token id)
         bool sorted;
     } llama_token_data_array;
 
@@ -302,7 +310,6 @@ extern "C" {
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
     //       https://github.com/ggerganov/llama.cpp/pull/7544
     struct llama_context_params {
-        uint32_t seed;              // RNG seed, -1 for random
         uint32_t n_ctx;             // text context, 0 = from model
         uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
         uint32_t n_ubatch;          // physical maximum batch size
@@ -330,11 +337,13 @@ extern "C" {
         enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
         enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
 
-        // Keep the booleans together to avoid misalignment during copy-by-value.
+        // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
+        // TODO: move at the end of the struct
         bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
         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
 
         // Abort callback
         // if it returns true, execution of llama_decode() will be aborted
@@ -358,56 +367,14 @@ extern "C" {
         void * kv_overrides;                 // pointer to vector containing overrides
     } llama_model_quantize_params;
 
-    // grammar types
-    struct llama_grammar;
+    typedef struct llama_logit_bias {
+        llama_token token;
+        float bias;
+    } llama_logit_bias;
 
-    // grammar element type
-    enum llama_gretype {
-        // end of rule definition
-        LLAMA_GRETYPE_END            = 0,
-
-        // start of alternate definition for rule
-        LLAMA_GRETYPE_ALT            = 1,
-
-        // non-terminal element: reference to rule
-        LLAMA_GRETYPE_RULE_REF       = 2,
-
-        // terminal element: character (code point)
-        LLAMA_GRETYPE_CHAR           = 3,
-
-        // inverse char(s) ([^a], [^a-b] [^abc])
-        LLAMA_GRETYPE_CHAR_NOT       = 4,
-
-        // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
-        // be an inclusive range ([a-z])
-        LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
-
-        // modifies a preceding LLAMA_GRETYPE_CHAR or
-        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
-        LLAMA_GRETYPE_CHAR_ALT       = 6,
-
-        // any character (.)
-        LLAMA_GRETYPE_CHAR_ANY       = 7,
-    };
-
-    typedef struct llama_grammar_element {
-        enum llama_gretype type;
-        uint32_t           value; // Unicode code point or rule ID
-    } llama_grammar_element;
-
-    // performance timing information
-    struct llama_timings {
-        double t_start_ms;
-        double t_end_ms;
-        double t_load_ms;
-        double t_sample_ms;
-        double t_p_eval_ms;
-        double t_eval_ms;
-
-        int32_t n_sample;
-        int32_t n_p_eval;
-        int32_t n_eval;
-    };
+    typedef struct llama_sampler_chain_params {
+        bool no_perf; // whether to measure performance timings
+    } llama_sampler_chain_params;
 
     // used in chat template
     typedef struct llama_chat_message {
@@ -419,8 +386,10 @@ extern "C" {
     struct llama_lora_adapter;
 
     // Helpers for getting default parameters
-    LLAMA_API struct llama_model_params llama_model_default_params(void);
-    LLAMA_API struct llama_context_params llama_context_default_params(void);
+    // TODO: update API to start accepting pointers to params structs (https://github.com/ggerganov/llama.cpp/discussions/9172)
+    LLAMA_API struct llama_model_params          llama_model_default_params(void);
+    LLAMA_API struct llama_context_params        llama_context_default_params(void);
+    LLAMA_API struct llama_sampler_chain_params  llama_sampler_chain_default_params(void);
     LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void);
 
     // Initialize the llama + ggml backend
@@ -443,10 +412,11 @@ extern "C" {
 
     LLAMA_API struct llama_model * llama_load_model_from_file(
                              const char * path_model,
-            struct llama_model_params     params);
+              struct llama_model_params   params);
 
     LLAMA_API void llama_free_model(struct llama_model * model);
 
+    // TODO: rename to llama_init_from_model
     LLAMA_API struct llama_context * llama_new_context_with_model(
                      struct llama_model * model,
             struct llama_context_params   params);
@@ -462,23 +432,22 @@ extern "C" {
     LLAMA_API bool llama_supports_mlock      (void);
     LLAMA_API bool llama_supports_gpu_offload(void);
 
-    LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
-
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
 
-    LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
-
-    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model * model);
-    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model * model);
-
     LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
     LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
     LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
     LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);
 
+    LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
+
+    LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
+    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model * model);
+    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model * model);
+
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
 
@@ -706,7 +675,7 @@ extern "C" {
     //
 
     // Returns the *actual* size in bytes of the state
-    // (rng, logits, embedding and kv_cache)
+    // (logits, embedding and kv_cache)
     // Only use when saving the state, not when restoring it, otherwise the size may be too small.
     LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
     LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
@@ -1009,121 +978,113 @@ extern "C" {
                                int32_t   length);
 
     //
-    // Grammar
+    // Sampling API
+    //
+    // Sample usage:
+    //
+    //    // prepare the sampling chain at the start
+    //    auto sparams = llama_sampler_chain_default_params();
+    //
+    //    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+    //
+    //    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(50));
+    //    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(0.9, 1));
+    //    llama_sampler_chain_add(smpl, llama_sampler_init_temp (0.8));
+    //
+    //    // typically, the chain should end with a sampler such as "greedy", "dist" or "mirostat"
+    //    // this sampler will be responsible to select the actual token
+    //    llama_sampler_chain_add(smpl, llama_sampler_init_dist(seed));
+    //
+    //    ...
+    //
+    //    // decoding loop:
+    //    while (...) {
+    //        ...
+    //
+    //        llama_decode(ctx, batch);
+    //
+    //        // sample from the logits of the last token in the batch
+    //        const llama_token id = llama_sampler_sample(smpl, ctx, -1);
+    //
+    //        // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.)
+    //        llama_sampler_accept(smpl, id);
+    //        ...
+    //    }
+    //
+    //    llama_sampler_free(smpl);
+    //
+    // TODO: In the future, llama_sampler will be utilized to offload the sampling to the backends (e.g. GPU).
+    // TODO: in the future, the entire sampling API that uses llama_model should start using llama_vocab
     //
 
-    /// Initialize a llama_grammar.
-    ///
-    /// @param rules The rule elements of the grammar to initialize.
-    /// @param n_rules The number of rules.
-    /// @param start_rule_index The index of the root rule (the starting point of the grammar).
-    /// @return The initialized llama_grammar or nullptr if initialization failed.
-    LLAMA_API struct llama_grammar * llama_grammar_init(
-            const llama_grammar_element ** rules,
-                                 size_t    n_rules,
-                                 size_t    start_rule_index);
+    typedef void * llama_sampler_context_t;
 
-    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
+    // user code can implement the interface below in order to create custom llama_sampler
+    struct llama_sampler_i {
+        const char *           (*name)  (const struct llama_sampler * smpl);                                 // can be NULL
+        void                   (*accept)(      struct llama_sampler * smpl, llama_token token);              // can be NULL
+        void                   (*apply) (      struct llama_sampler * smpl, llama_token_data_array * cur_p); // required
+        void                   (*reset) (      struct llama_sampler * smpl);                                 // can be NULL
+        struct llama_sampler * (*clone) (const struct llama_sampler * smpl);                                 // can be NULL if ctx is NULL
+        void                   (*free)  (      struct llama_sampler * smpl);                                 // can be NULL if ctx is NULL
 
-    LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar);
+        // TODO: API for internal libllama usage for appending the sampling to an existing ggml_cgraph
+        //void (*apply_ggml) (struct llama_sampler * smpl, ...);
+    };
 
-    /// @details Apply constraints from grammar
-    LLAMA_API void llama_grammar_sample(
-            const struct llama_grammar * grammar,
-            const struct llama_context * ctx,
-                llama_token_data_array * candidates);
-    LLAMA_API DEPRECATED(void llama_sample_grammar(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-      const struct llama_grammar * grammar),
-        "use llama_grammar_sample instead");
+    struct llama_sampler {
+        struct llama_sampler_i  * iface;
+        llama_sampler_context_t   ctx;
+    };
 
-    /// @details Accepts the sampled token into the grammar
-    LLAMA_API void llama_grammar_accept_token(
-            struct llama_grammar * grammar,
-            struct llama_context * ctx,
-                     llama_token   token);
+    // mirror of llama_sampler_i:
+    LLAMA_API const char *           llama_sampler_name  (const struct llama_sampler * smpl);
+    LLAMA_API void                   llama_sampler_accept(      struct llama_sampler * smpl, llama_token token);
+    LLAMA_API void                   llama_sampler_apply (      struct llama_sampler * smpl, llama_token_data_array * cur_p);
+    LLAMA_API void                   llama_sampler_reset (      struct llama_sampler * smpl);
+    LLAMA_API struct llama_sampler * llama_sampler_clone (const struct llama_sampler * smpl);
+    // important: do not free if the sampler has been added to a llama_sampler_chain (via llama_sampler_chain_add)
+    LLAMA_API void                   llama_sampler_free  (      struct llama_sampler * smpl);
 
-    //
-    // Sampling functions
-    //
+    // llama_sampler_chain
+    // a type of llama_sampler that can chain multiple samplers one after another
 
-    // Sets the current rng seed.
-    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+    LLAMA_API struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params);
 
-    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
-    /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
-    LLAMA_API void llama_sample_repetition_penalties(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-               const llama_token * last_tokens,
-                          size_t   penalty_last_n,
-                           float   penalty_repeat,
-                           float   penalty_freq,
-                           float   penalty_present);
+    // important: takes ownership of the sampler object and will free it when llama_sampler_free is called
+    LLAMA_API void                   llama_sampler_chain_add(      struct llama_sampler * chain, struct llama_sampler * smpl);
+    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);
 
-    /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
-    /// @param logits Logits extracted from the original generation context.
-    /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
-    /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    LLAMA_API void llama_sample_apply_guidance(
-              struct llama_context * ctx,
-                             float * logits,
-                             float * logits_guidance,
-                             float   scale);
+    // 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);
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist       (uint32_t seed);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    LLAMA_API void llama_sample_softmax(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates);
+    LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_k(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-                         int32_t   k,
-                          size_t   min_keep);
+    LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_p(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-                           float   p,
-                          size_t   min_keep);
+    LLAMA_API struct llama_sampler * llama_sampler_init_top_p      (float   p, size_t min_keep);
 
     /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
-    LLAMA_API void llama_sample_min_p(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-                           float   p,
-                          size_t   min_keep);
+    LLAMA_API struct llama_sampler * llama_sampler_init_min_p      (float   p, size_t min_keep);
 
     /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API void llama_sample_tail_free(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-                           float   z,
-                          size_t   min_keep);
+    LLAMA_API struct llama_sampler * llama_sampler_init_tail_free  (float   z, size_t min_keep);
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
-    LLAMA_API void llama_sample_typical(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-                           float   p,
-                          size_t   min_keep);
+    LLAMA_API struct llama_sampler * llama_sampler_init_typical    (float   p, size_t min_keep);
+    LLAMA_API struct llama_sampler * llama_sampler_init_temp       (float   t);
 
-    /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.
-    LLAMA_API void llama_sample_entropy(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates_p,
-                           float   min_temp,
-                           float   max_temp,
-                           float   exponent_val);
-
-    LLAMA_API void llama_sample_temp(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-                           float   temp);
+    /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.
+    LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext   (float   t, float   delta, float exponent);
 
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
@@ -1131,36 +1092,62 @@ extern "C" {
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-                           float   tau,
-                           float   eta,
-                         int32_t   m,
-                           float * mu);
+    LLAMA_API struct llama_sampler * llama_sampler_init_mirostat(
+                             int32_t   n_vocab,
+                            uint32_t   seed,
+                               float   tau,
+                               float   eta,
+                             int32_t   m);
 
     /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat_v2(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-                           float   tau,
-                           float   eta,
-                           float * mu);
+    LLAMA_API struct llama_sampler * llama_sampler_init_mirostat_v2(
+                            uint32_t   seed,
+                               float   tau,
+                               float   eta);
 
-    /// @details Selects the token with the highest probability.
-    ///          Does not compute the token probabilities. Use llama_sample_softmax() instead.
-    LLAMA_API llama_token llama_sample_token_greedy(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates);
+    LLAMA_API struct llama_sampler * llama_sampler_init_grammar(
+            const struct llama_model * model,
+                          const char * grammar_str,
+                          const char * grammar_root);
 
-    /// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx.
-    LLAMA_API llama_token llama_sample_token(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates);
+    LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
+                             int32_t   n_vocab,         // llama_n_vocab()
+                         llama_token   special_eos_id,  // llama_token_eos()
+                         llama_token   linefeed_id,     // llama_token_nl()
+                             int32_t   penalty_last_n,  // last n tokens to penalize (0 = disable penalty, -1 = context size)
+                               float   penalty_repeat,  // 1.0 = disabled
+                               float   penalty_freq,    // 0.0 = disabled
+                               float   penalty_present, // 0.0 = disabled
+                                bool   penalize_nl,     // consider newlines as a repeatable token
+                                bool   ignore_eos);     // ignore the end-of-sequence token
+
+    LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
+                             int32_t   n_vocab,
+                             int32_t   n_logit_bias,
+              const llama_logit_bias * logit_bias);
+
+
+    // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
+    LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);
+
+    /// @details Sample and accept a token from the idx-th output of the last evaluation
+    //
+    // Shorthand for:
+    //    const auto * logits = llama_get_logits_ith(ctx, idx);
+    //    llama_token_data_array cur_p = { ... init from logits ... };
+    //    llama_sampler_apply(smpl, &cur_p);
+    //    auto token = cur_p.data[cur_p.selected].id;
+    //    llama_sampler_accept(smpl, token);
+    //    return token;
+    // Returns the sampled token
+    LLAMA_API llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx);
+
+    // TODO: extend in the future
+    //LLAMA_API void llama_decode_with_sampler(struct llama_context * ctx, struct llama_sampler * smpl, struct llama_batch batch, ...);
 
     //
     // Model split
@@ -1176,12 +1163,6 @@ extern "C" {
     //  Returns the split_prefix length.
     LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
 
-    // Performance information
-    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
-
-    LLAMA_API void llama_print_timings(struct llama_context * ctx);
-    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
-
     // Print system information
     LLAMA_API const char * llama_print_system_info(void);
 
@@ -1189,65 +1170,41 @@ extern "C" {
     // If this is not called, or NULL is supplied, everything is output on stderr.
     LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data);
 
-    LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
+    //
+    // Performance utils
+    //
+    // NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements.
+    //
+
+    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;
+    };
+
+    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);
 
 #ifdef __cplusplus
 }
 #endif
 
-// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
-#ifdef LLAMA_API_INTERNAL
-
-#include <random>
-#include <string>
-#include <vector>
-
-struct ggml_tensor;
-
-const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
-    struct llama_context * ctx
-);
-
-struct llama_partial_utf8 {
-    uint32_t value;    // bit value so far (unshifted)
-    int      n_remain; // num bytes remaining; -1 indicates invalid sequence
-};
-
-struct llama_grammar_candidate {
-    size_t               index;
-    const uint32_t     * code_points;
-    llama_partial_utf8   partial_utf8;
-};
-
-using llama_grammar_rule  = std::vector<      llama_grammar_element>;
-using llama_grammar_stack = std::vector<const llama_grammar_element *>;
-
-using llama_grammar_rules      = std::vector<llama_grammar_rule>;
-using llama_grammar_stacks     = std::vector<llama_grammar_stack>;
-using llama_grammar_candidates = std::vector<llama_grammar_candidate>;
-
-const llama_grammar_rules  & llama_grammar_get_rules (const struct llama_grammar * grammar);
-      llama_grammar_stacks & llama_grammar_get_stacks(      struct llama_grammar * grammar);
-
-void llama_grammar_accept(
-        const llama_grammar_rules  & rules,
-        const llama_grammar_stacks & stacks,
-        const uint32_t chr,
-              llama_grammar_stacks & new_stacks);
-
-std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack(
-        const llama_grammar_rules & rules,
-        const llama_grammar_stack & stack,
-        const llama_grammar_candidates & candidates);
-
-std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
-        const std::string & src,
-        llama_partial_utf8 partial_start);
-
-// Randomly selects a token from the candidates based on their probabilities using given std::mt19937.
-// This is a temporary workaround in order to fix race conditions when sampling with multiple sequences.
-llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng);
-
-#endif // LLAMA_API_INTERNAL
-
 #endif // LLAMA_H

scripts/sync-ggml-am.sh

@@ -5,7 +5,7 @@
 # Usage:
 #
 #   $ cd /path/to/llama.cpp
-#   $ ./scripts/sync-ggml-am.sh -skip hash0,hash1,hash2...
+#   $ ./scripts/sync-ggml-am.sh -skip hash0,hash1,hash2... -C 3
 #
 
 set -e
@@ -25,9 +25,23 @@ lc=$(cat $SRC_LLAMA/scripts/sync-ggml.last)
 echo "Syncing ggml changes since commit $lc"
 
 to_skip=""
-if [ "$1" == "-skip" ]; then
-    to_skip=$2
-fi
+
+# context for git patches in number of lines
+ctx="8"
+
+while [ "$1" != "" ]; do
+    case $1 in
+        -skip )
+            shift
+            to_skip=$1
+            ;;
+        -C )
+            shift
+            ctx=$1
+            ;;
+    esac
+    shift
+done
 
 cd $SRC_GGML
 
@@ -52,7 +66,7 @@ while read c; do
         fi
     fi
 
-    git format-patch -k $c~1..$c --stdout -- \
+    git format-patch -U${ctx} -k $c~1..$c --stdout -- \
         CMakeLists.txt \
         src/CMakeLists.txt \
         cmake/FindSIMD.cmake \
@@ -191,7 +205,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         > ggml-src.patch.tmp
     mv ggml-src.patch.tmp ggml-src.patch
 
-    git am ggml-src.patch
+    git am -C${ctx} ggml-src.patch
 
     rm -v $SRC_LLAMA/ggml-src.patch
 fi