llama : refactor src/llama.cpp (#10902)

* llama : scatter llama.cpp into multiple modules (wip)

* llama : control-vector -> adapter

* llama : arch

* llama : mmap

ggml-ci

* ci : remove BUILD_SHARED_LIBS=OFF

ggml-ci

* llama : arch (cont)

ggml-ci

* llama : chat

ggml-ci

* llama : model

ggml-ci

* llama : hparams

ggml-ci

* llama : adapter

ggml-ci

* examples : fix

ggml-ci

* rebase

ggml-ci

* minor

* llama : kv cache

ggml-ci

* llama : impl

ggml-ci

* llama : batch

ggml-ci

* cont

ggml-ci

* llama : context

ggml-ci

* minor

* llama : context (cont)

ggml-ci

* llama : model loader

ggml-ci

* common : update lora

ggml-ci

* llama : quant

ggml-ci

* llama : quant (cont)

ggml-ci

* minor [no ci]

.github/workflows/build.yml

@@ -60,8 +60,7 @@ jobs:
             -DLLAMA_CURL=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DGGML_RPC=ON \
+            -DGGML_RPC=ON
-            -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -123,8 +122,7 @@ jobs:
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_CURL=ON \
             -DGGML_METAL=OFF \
-            -DGGML_RPC=ON \
+            -DGGML_RPC=ON
-            -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -181,7 +179,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -651,23 +649,23 @@ jobs:
       matrix:
         include:
           - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'msvc-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=O'
           - build: 'llvm-arm64-opencl-adreno'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
 
@@ -914,7 +912,7 @@ jobs:
         shell: cmd
         run: |
           call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
-          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
+          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DGGML_RPC=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release

common/arg.cpp

@@ -1512,7 +1512,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--lora"}, "FNAME",
         "path to LoRA adapter (can be repeated to use multiple adapters)",
         [](common_params & params, const std::string & value) {
-            params.lora_adapters.push_back({ std::string(value), 1.0 });
+            params.lora_adapters.push_back({ std::string(value), 1.0, nullptr });
         }
         // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@@ -1520,7 +1520,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--lora-scaled"}, "FNAME", "SCALE",
         "path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters)",
         [](common_params & params, const std::string & fname, const std::string & scale) {
-            params.lora_adapters.push_back({ fname, std::stof(scale) });
+            params.lora_adapters.push_back({ fname, std::stof(scale), nullptr });
         }
         // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));

common/common.cpp

@@ -922,20 +922,21 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
-        common_lora_adapter_container loaded_la;
+        llama_lora_adapter_ptr lora;
-        loaded_la.path = la.path;
+        lora.reset(llama_lora_adapter_init(model, la.path.c_str()));
-        loaded_la.scale = la.scale;
+        if (lora == nullptr) {
-        loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
-        if (loaded_la.adapter == nullptr) {
             LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
             llama_free_model(model);
             return iparams;
         }
-        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
+
+        la.ptr = lora.get();
+        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
+
     if (!params.lora_init_without_apply) {
-        common_lora_adapters_apply(lctx, iparams.lora_adapters);
+        common_lora_adapters_apply(lctx, params.lora_adapters);
     }
 
     if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
@@ -996,17 +997,17 @@ struct common_init_result common_init_from_params(common_params & params) {
         llama_perf_context_reset(lctx);
     }
 
-    iparams.model   = model;
+    iparams.model.reset(model);
-    iparams.context = lctx;
+    iparams.context.reset(lctx);
 
     return iparams;
 }
 
-void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters) {
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_info> & lora) {
     llama_lora_adapter_clear(ctx);
-    for (auto & la : lora_adapters) {
+    for (auto & la : lora) {
         if (la.scale != 0.0f) {
-            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+            llama_lora_adapter_set(ctx, la.ptr, la.scale);
         }
     }
 }

common/common.h

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "llama.h"
+#include "llama-cpp.h"
 
 #include <string>
 #include <vector>
@@ -27,10 +27,8 @@
 struct common_lora_adapter_info {
     std::string path;
     float scale;
-};
 
-struct common_lora_adapter_container : common_lora_adapter_info {
+    struct llama_lora_adapter * ptr;
-    struct llama_lora_adapter * adapter;
 };
 
 using llama_tokens = std::vector<llama_token>;
@@ -478,10 +476,12 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
+// note: defines object's lifetime
 struct common_init_result {
-    struct llama_model   * model   = nullptr;
+    llama_model_ptr   model;
-    struct llama_context * context = nullptr;
+    llama_context_ptr context;
-    std::vector<common_lora_adapter_container> lora_adapters;
+
+    std::vector<llama_lora_adapter_ptr> lora;
 };
 
 struct common_init_result     common_init_from_params(common_params & params);
@@ -503,7 +503,7 @@ struct llama_model * common_load_model_from_hf(
     const struct llama_model_params & params);
 
 // clear LoRA adapters from context, then apply new list of adapters
-void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_info> & lora);
 
 //
 // Batch utils
@@ -640,6 +640,10 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 // Split utils
 //
 
-static const char * const LLM_KV_SPLIT_NO            = "split.no";
+namespace {
-static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+
-static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+const char * const LLM_KV_SPLIT_NO            = "split.no";
+const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+
+}

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -434,12 +434,12 @@ static void print_matrix(struct ggml_tensor * probs) {
     }
 }
 
-struct llama_file {
+struct my_llama_file {
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
     size_t size;
 
-    llama_file(const char * fname, const char * mode) {
+    my_llama_file(const char * fname, const char * mode) {
         fp = std::fopen(fname, mode);
         if (fp == NULL) {
             size = 0;
@@ -500,7 +500,7 @@ struct llama_file {
         return std::string(chars.data(), len);
     }
 
-    ~llama_file() {
+    ~my_llama_file() {
         if (fp) {
             std::fclose(fp);
         }
@@ -508,7 +508,7 @@ struct llama_file {
 };
 
 static bool is_ggml_file(const char * filename) {
-    llama_file file(filename, "rb");
+    my_llama_file file(filename, "rb");
     if (file.size < 4) {
         return false;
     }
@@ -576,7 +576,7 @@ static void load_vocab(const char * filename, const Config * config, struct my_l
     } else {
         // assume llama2.c vocabulary
         LOG_INF("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
-        llama_file file(filename, "rb");
+        my_llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
         }

examples/cvector-generator/cvector-generator.cpp

@@ -415,12 +415,13 @@ int main(int argc, char ** argv) {
     // load the model to get hparams
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
 
     // int n_ctx = llama_n_ctx(ctx);
     int n_layers = llama_n_layer(model);
     int n_embd = llama_n_embd(model);
+
     // get model hint param (a.k.a model arch name)
     char model_hint[128];
     llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
@@ -474,8 +475,6 @@ int main(int argc, char ** argv) {
 
     // done with the model, we can now free it to make gain some memory
     printf("Done evaluate prompts, unload model...\n");
-    llama_free(ctx);
-    llama_free_model(model);
 
     bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;
 

examples/embedding/embedding.cpp

@@ -97,8 +97,9 @@ int main(int argc, char ** argv) {
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
+
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
         return 1;
@@ -316,8 +317,6 @@ int main(int argc, char ** argv) {
 
     // clean up
     llama_batch_free(batch);
-    llama_free(ctx);
-    llama_free_model(model);
     llama_backend_free();
 
     return 0;

examples/eval-callback/eval-callback.cpp

@@ -162,8 +162,9 @@ int main(int argc, char ** argv) {
     // init
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
+
     if (model == nullptr || ctx == nullptr) {
         LOG_ERR("%s : failed to init\n", __func__);
         return 1;
@@ -184,9 +185,6 @@ int main(int argc, char ** argv) {
     LOG("\n");
     llama_perf_context_print(ctx);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     return 0;

examples/gguf-split/gguf-split.cpp

@@ -2,15 +2,14 @@
 #include "common.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdlib>
 #include <fstream>
 #include <string>
 #include <vector>
-
-#include <stdio.h>
-#include <string.h>
 #include <climits>
+
+#include <cstdio>
+#include <cstring>
 #include <stdexcept>
 
 #if defined(_WIN32)

examples/imatrix/imatrix.cpp

@@ -430,9 +430,10 @@ static void process_logits(
 
 static bool compute_imatrix(llama_context * ctx, const common_params & params) {
     const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
     const int n_ctx = llama_n_ctx(ctx);
 
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+
     auto tim1 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenizing the input ..\n", __func__);
 
@@ -618,8 +619,9 @@ int main(int argc, char ** argv) {
     // init
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
+
     if (model == nullptr || ctx == nullptr) {
         LOG_ERR("%s : failed to init\n", __func__);
         return 1;
@@ -655,9 +657,6 @@ int main(int argc, char ** argv) {
     LOG("\n");
     llama_perf_context_print(ctx);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     return 0;

examples/infill/infill.cpp

@@ -131,8 +131,8 @@ int main(int argc, char ** argv) {
     LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
     common_init_result llama_init = common_init_from_params(params);
 
-    model = llama_init.model;
+    model = llama_init.model.get();
-    ctx = llama_init.context;
+    ctx = llama_init.context.get();
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
@@ -581,9 +581,6 @@ int main(int argc, char ** argv) {
     LOG("\n");
     common_perf_print(ctx, smpl);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     common_sampler_free(smpl);
     llama_backend_free();
 

examples/lookahead/lookahead.cpp

@@ -58,8 +58,8 @@ int main(int argc, char ** argv) {
     // load the target model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -474,9 +474,6 @@ int main(int argc, char ** argv) {
 
     llama_batch_free(batch);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     LOG("\n\n");

examples/lookup/lookup-create.cpp

@@ -1,14 +1,9 @@
 #include "arg.h"
 #include "common.h"
 #include "ngram-cache.h"
-#include "ggml.h"
 #include "llama.h"
 
-#include <cstdint>
-#include <fstream>
-#include <iostream>
 #include <string>
-#include <unordered_map>
 #include <vector>
 
 int main(int argc, char ** argv){
@@ -25,16 +20,16 @@ int main(int argc, char ** argv){
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model_ptr & model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_context_ptr & ctx = llama_init.context;
+
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = common_tokenize(ctx, params.prompt, true, true);
+    inp = common_tokenize(ctx.get(), params.prompt, true, true);
     fprintf(stderr, "%s: tokenization done\n", __func__);
 
-
     common_ngram_cache ngram_cache;
     common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
     fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());

examples/lookup/lookup-stats.cpp

@@ -30,12 +30,11 @@ int main(int argc, char ** argv){
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_context_ptr & ctx = llama_init.context;
-    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = common_tokenize(ctx, params.prompt, true, true);
+    inp = common_tokenize(ctx.get(), params.prompt, true, true);
 
     common_ngram_cache ngram_cache_context;
     common_ngram_cache ngram_cache_dynamic;
@@ -66,7 +65,7 @@ int main(int argc, char ** argv){
     }
 
     const int n_input = inp.size();
-    const int n_ctx = llama_n_ctx(ctx);
+    const int n_ctx = llama_n_ctx(ctx.get());
 
     int n_drafted = 0;
     int n_accept  = 0;
@@ -150,9 +149,6 @@ int main(int argc, char ** argv){
     LOG_INF("n_accept     = %d\n", n_accept);
     LOG_INF("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     LOG("\n\n");

examples/lookup/lookup.cpp

@@ -33,8 +33,8 @@ int main(int argc, char ** argv){
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
 
     // tokenize the prompt
     std::vector<llama_token> inp;
@@ -243,9 +243,6 @@ int main(int argc, char ** argv){
 
     llama_batch_free(batch_tgt);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     LOG("\n\n");

examples/main/main.cpp

@@ -145,18 +145,18 @@ int main(int argc, char ** argv) {
     llama_context * ctx = nullptr;
     common_sampler * smpl = nullptr;
 
-    std::vector<common_chat_msg> chat_msgs;
-
     g_model = &model;
     g_ctx = &ctx;
     g_smpl = &smpl;
 
+    std::vector<common_chat_msg> chat_msgs;
+
     // load the model and apply lora adapter, if any
     LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
     common_init_result llama_init = common_init_from_params(params);
 
-    model = llama_init.model;
+    model = llama_init.model.get();
-    ctx = llama_init.context;
+    ctx = llama_init.context.get();
 
     if (model == NULL) {
         LOG_ERR("%s: error: unable to load model\n", __func__);
@@ -889,9 +889,6 @@ int main(int argc, char ** argv) {
 
     common_sampler_free(smpl);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     ggml_threadpool_free_fn(threadpool);

examples/parallel/parallel.cpp

@@ -132,8 +132,8 @@ int main(int argc, char ** argv) {
     // load the target model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
 
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {
@@ -416,9 +416,6 @@ int main(int argc, char ** argv) {
 
     llama_batch_free(batch);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     LOG("\n\n");

examples/perplexity/perplexity.cpp

@@ -1987,8 +1987,9 @@ int main(int argc, char ** argv) {
     // load the model and apply lora adapter, if any
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
+
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
         return 1;
@@ -2023,9 +2024,6 @@ int main(int argc, char ** argv) {
     LOG("\n");
     llama_perf_context_print(ctx);
 
-    llama_free(ctx);
-    llama_free_model(model);
-
     llama_backend_free();
 
     return 0;

examples/quantize-stats/quantize-stats.cpp

@@ -1,7 +1,7 @@
-#include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "llama-impl.h"
+#include "llama-context.h"
+#include "common.h"
 
 #include <algorithm>
 #include <cassert>
@@ -9,11 +9,9 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
-#include <map>
 #include <numeric>
 #include <regex>
 #include <string>
-#include <unordered_map>
 #include <vector>
 #include <thread>
 #include <mutex>
@@ -330,13 +328,13 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const auto &tensors = llama_internal_get_tensor_map(ctx);
+    const auto & tensors = llama_internal_get_tensor_map(ctx);
 
     // check layer tensors
     int included_layers = 0;
     int64_t max_nelements = 0;
     bool is_f16 = false;
-    for (const auto& kv_tensor : tensors) {
+    for (const auto & kv_tensor : tensors) {
         if (!layer_included(params, kv_tensor.first)) {
             continue;
         }
@@ -371,8 +369,8 @@ int main(int argc, char ** argv) {
         if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
             continue;
         }
-        const auto *  qfns     = ggml_get_type_traits(type);
+        const auto * qfns     = ggml_get_type_traits(type);
-        const auto *  qfns_cpu = ggml_get_type_traits_cpu(type);
+        const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
         if (qfns_cpu->from_float && qfns->to_float) {
             if (params.verbose) {
                 printf("testing %s ...\n",  ggml_type_name(type));
@@ -382,7 +380,7 @@ int main(int argc, char ** argv) {
 
             error_stats global_stats {};
 
-            for (const auto& kv_tensor : tensors) {
+            for (const auto & kv_tensor : tensors) {
                 if (!layer_included(params, kv_tensor.first)) {
                     continue;
                 }

examples/retrieval/retrieval.cpp

@@ -151,8 +151,8 @@ int main(int argc, char ** argv) {
     // load the model
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
@@ -298,7 +298,5 @@ int main(int argc, char ** argv) {
 
     // clean up
     llama_batch_free(query_batch);
-    llama_free(ctx);
-    llama_free_model(model);
     llama_backend_free();
 }

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

@@ -30,8 +30,8 @@ int main(int argc, char ** argv) {
     // init
     common_init_result llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model;
+    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context;
+    llama_context * ctx = llama_init.context.get();
 
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
@@ -89,8 +89,6 @@ int main(int argc, char ** argv) {
         if (llama_decode(ctx, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_batch_free(batch);
-            llama_free(ctx);
-            llama_free_model(model);
             return 1;
         }
         n_past += 1;
@@ -98,11 +96,8 @@ int main(int argc, char ** argv) {
 
     printf("\n\n");
 
-    // free old context
-    llama_free(ctx);
-
     // make new context
-    auto * ctx2 = llama_new_context_with_model(model, common_context_params_to_llama(params));
+    llama_context * ctx2 = llama_new_context_with_model(model, common_context_params_to_llama(params));
 
     llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
 
@@ -123,8 +118,6 @@ int main(int argc, char ** argv) {
 
         if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
-            llama_free(ctx2);
-            llama_free_model(model);
             return 1;
         }
 
@@ -148,8 +141,6 @@ int main(int argc, char ** argv) {
         if (llama_decode(ctx2, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_batch_free(batch);
-            llama_free(ctx2);
-            llama_free_model(model);
             return 1;
         }
         n_past += 1;
@@ -157,15 +148,13 @@ int main(int argc, char ** argv) {
 
     printf("\n\n");
 
-    llama_free(ctx2);
-
     if (result0 != result1) {
         fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
         return 1;
     }
 
     // make new context
-    auto * ctx3 = llama_new_context_with_model(model, common_context_params_to_llama(params));
+    llama_context * ctx3 = llama_new_context_with_model(model, common_context_params_to_llama(params));
 
     llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
 
@@ -186,8 +175,6 @@ int main(int argc, char ** argv) {
 
         if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
-            llama_free(ctx3);
-            llama_free_model(model);
             return 1;
         }
 
@@ -204,8 +191,6 @@ int main(int argc, char ** argv) {
         const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
         if (ncopy != seq_store.size()) {
             fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
-            llama_free(ctx3);
-            llama_free_model(model);
             return 1;
         }
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
@@ -218,8 +203,6 @@ int main(int argc, char ** argv) {
         const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
         if (nset != seq_store.size()) {
             fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
-            llama_free(ctx3);
-            llama_free_model(model);
             return 1;
         }
         fprintf(stderr, "%s : seq 1 restored, %zd bytes\n", __func__, nset);
@@ -239,8 +222,6 @@ int main(int argc, char ** argv) {
         if (llama_decode(ctx3, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_batch_free(batch);
-            llama_free(ctx3);
-            llama_free_model(model);
             return 1;
         }
         n_past += 1;
@@ -253,8 +234,6 @@ int main(int argc, char ** argv) {
     llama_sampler_free(smpl3);
 
     llama_batch_free(batch);
-    llama_free(ctx3);
-    llama_free_model(model);
 
     if (result0 != result2) {
         fprintf(stderr, "\n%s : error : the seq restore generation is different\n", __func__);

examples/server/server.cpp

@@ -98,7 +98,7 @@ struct slot_params {
     int64_t t_max_prompt_ms  = -1; // TODO: implement
     int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
 
-    std::vector<common_lora_adapter_container> lora;
+    std::vector<common_lora_adapter_info> lora;
 
     std::vector<std::string> antiprompt;
     std::vector<std::string> response_fields;
@@ -198,7 +198,7 @@ struct server_task {
     bool metrics_reset_bucket = false;
 
     // used by SERVER_TASK_TYPE_SET_LORA
-    std::vector<common_lora_adapter_container> set_lora;
+    std::vector<common_lora_adapter_info> set_lora;
 
     server_task(server_task_type type) : type(type) {}
 
@@ -206,7 +206,6 @@ struct server_task {
             const llama_model * model,
             const llama_context * ctx,
             const common_params & params_base,
-            const std::vector<common_lora_adapter_container> & lora_base,
             const json & data) {
         slot_params params;
 
@@ -265,12 +264,12 @@ struct server_task {
 
         if (data.contains("lora")) {
             if (data.at("lora").is_array()) {
-                params.lora = parse_lora_request(lora_base, data.at("lora"));
+                params.lora = parse_lora_request(params_base.lora_adapters, data.at("lora"));
             } else {
                 throw std::runtime_error("Error: 'lora' must be an array of objects with 'id' and 'scale' fields");
             }
         } else {
-            params.lora = lora_base;
+            params.lora = params_base.lora_adapters;
         }
 
         // TODO: add more sanity checks for the input parameters
@@ -1132,7 +1131,7 @@ struct server_slot {
 
     common_speculative * spec = nullptr;
 
-    std::vector<common_lora_adapter_container> lora;
+    std::vector<common_lora_adapter_info> lora;
 
     // the index relative to completion multi-task request
     size_t index = 0;
@@ -1627,11 +1626,15 @@ struct server_response {
 struct server_context {
     common_params params_base;
 
+    // note: keep these alive - they determine the lifetime of the model, context, etc.
+    common_init_result llama_init;
+    common_init_result llama_init_dft;
+
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
-    std::vector<common_lora_adapter_container> lora;
 
     llama_model * model_dft = nullptr;
+
     llama_context_params cparams_dft;
 
     llama_batch batch = {};
@@ -1655,21 +1658,6 @@ struct server_context {
     float slot_prompt_similarity = 0.0f;
 
     ~server_context() {
-        if (ctx) {
-            llama_free(ctx);
-            ctx = nullptr;
-        }
-
-        if (model) {
-            llama_free_model(model);
-            model = nullptr;
-        }
-
-        if (model_dft) {
-            llama_free_model(model_dft);
-            model_dft = nullptr;
-        }
-
         // Clear any sampling context
         for (server_slot & slot : slots) {
             common_sampler_free(slot.smpl);
@@ -1692,11 +1680,10 @@ struct server_context {
 
         params_base = params;
 
-        common_init_result llama_init = common_init_from_params(params_base);
+        llama_init = common_init_from_params(params_base);
 
-        model = llama_init.model;
+        model = llama_init.model.get();
-        ctx   = llama_init.context;
+        ctx   = llama_init.context.get();
-        lora  = llama_init.lora_adapters;
 
         if (model == nullptr) {
             SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
@@ -1719,25 +1706,22 @@ struct server_context {
             params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
             params_dft.n_parallel   = 1;
 
-            common_init_result llama_init_dft = common_init_from_params(params_dft);
+            llama_init_dft = common_init_from_params(params_dft);
 
-            model_dft = llama_init_dft.model;
+            model_dft = llama_init_dft.model.get();
 
             if (model_dft == nullptr) {
                 SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str());
                 return false;
             }
 
-            if (!common_speculative_are_compatible(ctx, llama_init_dft.context)) {
+            if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
                 SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str());
 
-                llama_free      (llama_init_dft.context);
-                llama_free_model(llama_init_dft.model);
-
                 return false;
             }
 
-            const int n_ctx_dft = llama_n_ctx(llama_init_dft.context);
+            const int n_ctx_dft = llama_n_ctx(llama_init_dft.context.get());
 
             cparams_dft = common_context_params_to_llama(params_dft);
             cparams_dft.n_batch = n_ctx_dft;
@@ -1745,9 +1729,6 @@ struct server_context {
             // force F16 KV cache for the draft model for extra performance
             cparams_dft.type_k = GGML_TYPE_F16;
             cparams_dft.type_v = GGML_TYPE_F16;
-
-            // the context is not needed - we will create one for each slot
-            llama_free(llama_init_dft.context);
         }
 
         return true;
@@ -1898,7 +1879,7 @@ struct server_context {
         if (!are_lora_equal(task.params.lora, slot.lora)) {
             // if lora is changed, we cannot reuse cached tokens
             slot.cache_tokens.clear();
-            slot.lora = std::move(task.params.lora);
+            slot.lora = task.params.lora;
         }
 
         SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str());
@@ -2592,7 +2573,7 @@ struct server_context {
                 } break;
             case SERVER_TASK_TYPE_SET_LORA:
                 {
-                    lora = std::move(task.set_lora);
+                    params_base.lora_adapters = std::move(task.set_lora);
                     auto res = std::make_unique<server_task_result_apply_lora>();
                     res->id = task.id;
                     queue_results.send(std::move(res));
@@ -3671,7 +3652,6 @@ int main(int argc, char ** argv) {
                                             ctx_server.model,
                                             ctx_server.ctx,
                                             ctx_server.params_base,
-                                            ctx_server.lora,
                                             data);
                 task.id_selected_slot = json_value(data, "id_slot", -1);
 
@@ -4098,8 +4078,9 @@ int main(int argc, char ** argv) {
 
     const auto handle_lora_adapters_list = [&](const httplib::Request &, httplib::Response & res) {
         json result = json::array();
-        for (size_t i = 0; i < ctx_server.lora.size(); ++i) {
+        const auto & loras = ctx_server.params_base.lora_adapters;
-            auto & lora = ctx_server.lora[i];
+        for (size_t i = 0; i < loras.size(); ++i) {
+            auto & lora = loras[i];
             result.push_back({
                 {"id", i},
                 {"path", lora.path},
@@ -4118,7 +4099,7 @@ int main(int argc, char ** argv) {
         }
         server_task task(SERVER_TASK_TYPE_SET_LORA);
         task.id = ctx_server.queue_tasks.get_new_id();
-        task.set_lora = parse_lora_request(ctx_server.lora, body);
+        task.set_lora = parse_lora_request(ctx_server.params_base.lora_adapters, body);
         ctx_server.queue_results.add_waiting_task_id(task.id);
         ctx_server.queue_tasks.post(task);
 

examples/server/utils.hpp

@@ -799,25 +799,25 @@ static std::vector<llama_token_data> get_token_probabilities(llama_context * ctx
 }
 
 static bool are_lora_equal(
-        const std::vector<common_lora_adapter_container> & l1,
+        const std::vector<common_lora_adapter_info> & l1,
-        const std::vector<common_lora_adapter_container> & l2) {
+        const std::vector<common_lora_adapter_info> & l2) {
     if (l1.size() != l2.size()) {
         return false;
     }
     for (size_t i = 0; i < l1.size(); ++i) {
         // we don't check lora.path to reduce the time complexity
-        if (l1[i].scale != l2[i].scale || l1[i].adapter != l2[i].adapter) {
+        if (l1[i].scale != l2[i].scale || l1[i].ptr != l2[i].ptr) {
             return false;
         }
     }
     return true;
 }
 
-// parse lora config from JSON request, returned a copy of base_lora with updated scale
+// parse lora config from JSON request, returned a copy of lora_base with updated scale
-static std::vector<common_lora_adapter_container> parse_lora_request(
+static std::vector<common_lora_adapter_info> parse_lora_request(
-        const std::vector<common_lora_adapter_container> & base_lora,
+        const std::vector<common_lora_adapter_info> & lora_base,
         const json & data) {
-    std::vector<common_lora_adapter_container> lora(base_lora);
+    std::vector<common_lora_adapter_info> lora(lora_base);
     int max_idx = lora.size();
 
     // clear existing value

examples/speculative-simple/speculative-simple.cpp

@@ -34,7 +34,7 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     llama_model * model_tgt = NULL;
-    llama_model * model_dft = NULL;
+    //llama_model * model_dft = NULL;
 
     llama_context * ctx_tgt = NULL;
     llama_context * ctx_dft = NULL;
@@ -42,8 +42,8 @@ int main(int argc, char ** argv) {
     // load the target model
     common_init_result llama_init_tgt = common_init_from_params(params);
 
-    model_tgt = llama_init_tgt.model;
+    model_tgt = llama_init_tgt.model.get();
-    ctx_tgt   = llama_init_tgt.context;
+    ctx_tgt   = llama_init_tgt.context.get();
 
     // load the draft model
     params.devices      = params.speculative.devices;
@@ -59,8 +59,8 @@ int main(int argc, char ** argv) {
     params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
     common_init_result llama_init_dft = common_init_from_params(params);
 
-    model_dft = llama_init_dft.model;
+    //model_dft = llama_init_dft.model.get();
-    ctx_dft   = llama_init_dft.context;
+    ctx_dft   = llama_init_dft.context.get();
 
     if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
         return 1;
@@ -251,12 +251,6 @@ int main(int argc, char ** argv) {
     common_sampler_free(smpl);
     common_speculative_free(spec);
 
-    llama_free(ctx_tgt);
-    llama_free_model(model_tgt);
-
-    llama_free(ctx_dft);
-    llama_free_model(model_dft);
-
     llama_backend_free();
 
     LOG("\n\n");

examples/speculative/speculative.cpp

@@ -72,8 +72,9 @@ int main(int argc, char ** argv) {
 
     // load the target model
     common_init_result llama_init_tgt = common_init_from_params(params);
-    model_tgt = llama_init_tgt.model;
+
-    ctx_tgt = llama_init_tgt.context;
+    model_tgt = llama_init_tgt.model.get();
+    ctx_tgt   = llama_init_tgt.context.get();
 
     // load the draft model
     params.devices = params.speculative.devices;
@@ -85,8 +86,9 @@ int main(int argc, char ** argv) {
 
     params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
     common_init_result llama_init_dft = common_init_from_params(params);
-    model_dft = llama_init_dft.model;
+
-    ctx_dft = llama_init_dft.context;
+    model_dft = llama_init_dft.model.get();
+    ctx_dft   = llama_init_dft.context.get();
 
     const bool vocab_type_tgt = llama_vocab_type(model_tgt);
     LOG_DBG("vocab_type tgt: %d\n", vocab_type_tgt);
@@ -631,12 +633,6 @@ int main(int argc, char ** argv) {
 
     llama_batch_free(batch_dft);
 
-    llama_free(ctx_tgt);
-    llama_free_model(model_tgt);
-
-    llama_free(ctx_dft);
-    llama_free_model(model_dft);
-
     llama_backend_free();
 
     LOG("\n\n");

examples/tts/tts.cpp

@@ -458,8 +458,9 @@ int main(int argc, char ** argv) {
     llama_context * ctx_cts = NULL;
 
     common_init_result llama_init_ttc = common_init_from_params(params);
-    model_ttc = llama_init_ttc.model;
+
-    ctx_ttc = llama_init_ttc.context;
+    model_ttc = llama_init_ttc.model.get();
+    ctx_ttc   = llama_init_ttc.context.get();
 
     // TODO: refactor in a common struct
     params.model     = params.vocoder.model;
@@ -470,8 +471,9 @@ int main(int argc, char ** argv) {
     params.embedding = true;
 
     common_init_result llama_init_cts = common_init_from_params(params);
-    model_cts = llama_init_cts.model;
+
-    ctx_cts = llama_init_cts.context;
+    model_cts = llama_init_cts.model.get();
+    ctx_cts   = llama_init_cts.context.get();
 
     std::vector<common_sampler *> smpl(n_parallel);
     for (int i = 0; i < n_parallel; ++i) {
@@ -920,12 +922,6 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
 
     LOG_INF("%s: audio written to file '%s'\n", __func__, fname.c_str());
 
-    llama_free(ctx_ttc);
-    llama_free_model(model_ttc);
-
-    llama_free(ctx_cts);
-    llama_free_model(model_cts);
-
     llama_backend_free();
 
     return 0;

include/llama-cpp.h

@@ -20,6 +20,11 @@ struct llama_sampler_deleter {
     void operator()(llama_sampler * sampler) { llama_sampler_free(sampler); }
 };
 
+struct llama_lora_adapter_deleter {
+    void operator()(llama_lora_adapter * lora_adapter) { llama_lora_adapter_free(lora_adapter); }
+};
+
 typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;
 typedef std::unique_ptr<llama_context, llama_context_deleter> llama_context_ptr;
 typedef std::unique_ptr<llama_sampler, llama_sampler_deleter> llama_sampler_ptr;
+typedef std::unique_ptr<llama_lora_adapter, llama_lora_adapter_deleter> llama_lora_adapter_ptr;

include/llama.h

@@ -385,6 +385,7 @@ extern "C" {
     } llama_chat_message;
 
     // lora adapter
+    // TODO: rename to llama_adapter_lora
     struct llama_lora_adapter;
 
     // Helpers for getting default parameters
@@ -416,6 +417,7 @@ extern "C" {
                              const char * path_model,
               struct llama_model_params   params);
 
+    // TODO: rename to llama_model_free
     LLAMA_API void llama_free_model(struct llama_model * model);
 
     // TODO: rename to llama_init_from_model
@@ -501,14 +503,19 @@ extern "C" {
             const char * fname_out,
             const llama_model_quantize_params * params);
 
+    //
+    // Adapters
+    //
+
     // Load a LoRA adapter from file
-    // The loaded adapter will be associated to the given model, and will be free when the model is deleted
+    // TODO: rename to llama_adapter_lora_init
     LLAMA_API struct llama_lora_adapter * llama_lora_adapter_init(
             struct llama_model * model,
             const char * path_lora);
 
     // Add a loaded LoRA adapter to given context
     // This will not modify model's weight
+    // TODO: rename to llama_set_adapter_lora
     LLAMA_API int32_t llama_lora_adapter_set(
             struct llama_context * ctx,
             struct llama_lora_adapter * adapter,
@@ -516,16 +523,18 @@ extern "C" {
 
     // Remove a specific LoRA adapter from given context
     // Return -1 if the adapter is not present in the context
+    // TODO: rename to llama_rm_adapter_lora
     LLAMA_API int32_t llama_lora_adapter_remove(
             struct llama_context * ctx,
             struct llama_lora_adapter * adapter);
 
     // Remove all LoRA adapters from given context
-    LLAMA_API void llama_lora_adapter_clear(
+    // TODO: rename to llama_clear_adapter_lora
-            struct llama_context * ctx);
+    LLAMA_API void llama_lora_adapter_clear(struct llama_context * ctx);
 
     // Manually free a LoRA adapter
     // Note: loaded adapters will be free when the associated model is deleted
+    // TODO: rename to llama_adapter_lora_free
     LLAMA_API void llama_lora_adapter_free(struct llama_lora_adapter * adapter);
 
     // Apply a loaded control vector to a llama_context, or if data is NULL, clear
@@ -534,6 +543,7 @@ extern "C" {
     // to an n_embd x n_layers buffer starting from layer 1.
     // il_start and il_end are the layer range the vector should apply to (both inclusive)
     // See llama_control_vector_load in common to load a control vector.
+    // TODO: rename to llama_adapter_cvec_apply
     LLAMA_API int32_t llama_control_vector_apply(
             struct llama_context * lctx,
                      const float * data,
@@ -546,6 +556,8 @@ extern "C" {
     // KV cache
     //
 
+    // TODO: remove llama_kv_cache_view_* API
+
     // Information associated with an individual cell in the KV cache view.
     struct llama_kv_cache_view_cell {
         // The position for this cell. Takes KV cache shifts into account.
@@ -592,8 +604,11 @@ extern "C" {
     LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
 
     // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
+    // TODO: change signature to llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_context * ctx)
     LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
 
+    ///
+
     // Returns the number of tokens in the KV cache (slow, use only for debug)
     // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
     LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
@@ -663,6 +678,9 @@ extern "C" {
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
+    // TODO: the llama_kv_cache_defrag and llama_kv_cache_update API tightly couples llama_context with llama_kv_cache
+    //       how to avoid this?
+
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()

src/CMakeLists.txt

@@ -9,9 +9,21 @@ llama_add_compile_flags()
 add_library(llama
             ../include/llama.h
             llama.cpp
-            llama-vocab.cpp
+            llama-adapter.cpp
+            llama-arch.cpp
+            llama-batch.cpp
+            llama-chat.cpp
+            llama-context.cpp
             llama-grammar.cpp
+            llama-hparams.cpp
+            llama-impl.cpp
+            llama-kv-cache.cpp
+            llama-mmap.cpp
+            llama-model-loader.cpp
+            llama-model.cpp
+            llama-quant.cpp
             llama-sampling.cpp
+            llama-vocab.cpp
             unicode.h
             unicode.cpp
             unicode-data.cpp

src/llama-adapter.cpp

@@ -0,0 +1,334 @@
+#include "llama-adapter.h"
+
+#include "llama-model.h"
+
+#include <algorithm>
+#include <map>
+#include <cassert>
+#include <stdexcept>
+
+// vec
+
+struct ggml_tensor * llama_control_vector::tensor_for(int il) const {
+    if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
+        return nullptr;
+    }
+
+    return tensors[il];
+}
+
+struct ggml_tensor * llama_control_vector::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
+    ggml_tensor * layer_dir = tensor_for(il);
+    if (layer_dir != nullptr) {
+        cur = ggml_add(ctx, cur, layer_dir);
+    }
+
+    return cur;
+}
+
+static bool llama_control_vector_init(struct llama_control_vector & cvec, const llama_model & model) {
+    const auto & hparams = model.hparams;
+
+    GGML_ASSERT(cvec.tensors.empty());
+    GGML_ASSERT(cvec.ctxs.empty());
+    GGML_ASSERT(cvec.bufs.empty());
+
+    // create a context for each buffer type
+    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
+    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
+        auto it = ctx_map.find(buft);
+        if (it == ctx_map.end()) {
+            struct ggml_init_params params = {
+                /*.mem_size   =*/ hparams.n_layer*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+
+            ggml_context * ctx = ggml_init(params);
+            if (!ctx) {
+                return nullptr;
+            }
+
+            ctx_map[buft] = ctx;
+            cvec.ctxs.emplace_back(ctx);
+
+            return ctx;
+        }
+
+        return it->second;
+    };
+
+    // make tensors
+    cvec.tensors.reserve(hparams.n_layer);
+    cvec.tensors.push_back(nullptr); // there's never a tensor for layer 0
+    for (size_t il = 1; il < hparams.n_layer; il++) {
+        ggml_backend_buffer_type_t buft = llama_model_select_buft(model, il);
+        ggml_context * ctx = ctx_for_buft(buft);
+        if (!ctx) {
+            LLAMA_LOG_ERROR("%s: failed to allocate context for control vector\n", __func__);
+            return false;
+        }
+        ggml_tensor * tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hparams.n_embd);
+        cvec.tensors.push_back(tensor);
+    }
+
+    // allocate tensors / buffers and zero
+    cvec.bufs.reserve(ctx_map.size());
+    for (auto it : ctx_map) {
+        ggml_backend_buffer_type_t buft = it.first;
+        ggml_context * ctx = it.second;
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+        if (!buf) {
+            LLAMA_LOG_ERROR("%s: failed to allocate buffer for control vector\n", __func__);
+            return false;
+        }
+        ggml_backend_buffer_clear(buf, 0);
+        cvec.bufs.emplace_back(buf);
+    }
+
+    return true;
+}
+
+int32_t llama_control_vector_apply(
+        struct llama_control_vector & cvec,
+        const llama_model & model,
+        const float * data,
+        size_t len,
+        int32_t n_embd,
+        int32_t il_start,
+        int32_t il_end) {
+    const auto & hparams = model.hparams;
+
+    if (data == nullptr) {
+        // disable the current control vector (but leave allocated for later)
+        cvec.layer_start = -1;
+        cvec.layer_end   = -1;
+        return 0;
+    }
+
+    if (n_embd != (int) hparams.n_embd) {
+        LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
+        return 1;
+    }
+
+    if (cvec.tensors.empty()) {
+        if (!llama_control_vector_init(cvec, model)) {
+            return 1;
+        }
+    }
+
+    cvec.layer_start = il_start;
+    cvec.layer_end   = il_end;
+
+    for (size_t il = 1; il < hparams.n_layer; il++) {
+        assert(cvec.tensors[il] != nullptr);
+
+        const size_t off = n_embd * (il - 1); // buffer doesn't have data for layer 0, since it's never present
+        if (off + n_embd <= len) {
+            ggml_backend_tensor_set(cvec.tensors[il], data + off, 0, n_embd * ggml_element_size(cvec.tensors[il]));
+        }
+    }
+
+    return 0;
+}
+
+// lora
+
+llama_lora_weight * llama_lora_adapter::get_weight(struct ggml_tensor * w) {
+    const std::string name(w->name);
+
+    const auto pos = ab_map.find(name);
+    if (pos != ab_map.end()) {
+        return &pos->second;
+    }
+
+    return nullptr;
+}
+
+void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
+    delete adapter;
+}
+
+static void llama_lora_adapter_init_impl(struct llama_model & model, const char * path_lora, struct llama_lora_adapter & adapter) {
+    LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
+
+    ggml_context * ctx_init;
+    struct gguf_init_params meta_gguf_params = {
+        /* .no_alloc = */ true,
+        /* .ctx      = */ &ctx_init,
+    };
+
+    gguf_context_ptr ctx_gguf { gguf_init_from_file(path_lora, meta_gguf_params) };
+    if (!ctx_gguf) {
+        throw std::runtime_error("failed to load lora adapter file from " + std::string(path_lora));
+    }
+
+    ggml_context_ptr ctx { ctx_init };
+
+    // check metadata
+    {
+        auto get_kv_str = [&](const std::string & key) -> std::string {
+            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
+            return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf.get(), id));
+        };
+        auto get_kv_f32 = [&](const std::string & key) -> float {
+            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
+            return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf.get(), id);
+        };
+        LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
+
+        auto general_type = get_kv_str(llm_kv(LLM_KV_GENERAL_TYPE));
+        if (general_type != "adapter") {
+            throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
+        }
+
+        auto general_arch_str = get_kv_str(llm_kv(LLM_KV_GENERAL_ARCHITECTURE));
+        auto general_arch = llm_arch_from_string(general_arch_str);
+        if (general_arch != model.arch) {
+            throw std::runtime_error("model arch and LoRA arch mismatch");
+        }
+
+        auto adapter_type = get_kv_str(llm_kv(LLM_KV_ADAPTER_TYPE));
+        if (adapter_type != "lora") {
+            throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
+        }
+
+        adapter.alpha = get_kv_f32(llm_kv(LLM_KV_ADAPTER_LORA_ALPHA));
+    }
+
+    int n_tensors = gguf_get_n_tensors(ctx_gguf.get());
+
+    // contexts for each buffer type
+    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
+    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
+        auto it = ctx_map.find(buft);
+        if (it == ctx_map.end()) {
+            // add a new context
+            struct ggml_init_params params = {
+                /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ggml_context * buft_ctx = ggml_init(params);
+            if (!buft_ctx) {
+                return nullptr;
+            }
+            ctx_map[buft] = buft_ctx;
+            adapter.ctxs.emplace_back(buft_ctx);
+            return buft_ctx;
+        };
+        return it->second;
+    };
+
+    // bundle lora_a and lora_b into pairs
+    std::map<std::string, llama_lora_weight> ab_map;
+    auto str_endswith = [](const std::string & str, const std::string & suffix) {
+        return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+    };
+
+    for (ggml_tensor * cur = ggml_get_first_tensor(ctx.get()); cur; cur = ggml_get_next_tensor(ctx.get(), cur)) {
+        std::string name(cur->name);
+        if (str_endswith(name, ".lora_a")) {
+            replace_all(name, ".lora_a", "");
+            if (ab_map.find(name) == ab_map.end()) {
+                ab_map[name] = llama_lora_weight(cur, nullptr);
+            } else {
+                ab_map[name].a = cur;
+            }
+        } else if (str_endswith(name, ".lora_b")) {
+            replace_all(name, ".lora_b", "");
+            if (ab_map.find(name) == ab_map.end()) {
+                ab_map[name] = llama_lora_weight(nullptr, cur);
+            } else {
+                ab_map[name].b = cur;
+            }
+        } else {
+            throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
+        }
+    }
+
+    // add tensors
+    for (auto & it : ab_map) {
+        const std::string & name = it.first;
+        llama_lora_weight & w = it.second;
+
+        if (!w.a || !w.b) {
+            throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component");
+        }
+
+        // device buft and device ctx
+        auto * model_tensor = llama_model_get_tensor(model, name.c_str());
+        if (!model_tensor) {
+            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model");
+        }
+
+        struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
+        // validate tensor shape
+        if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
+            throw std::runtime_error("tensor '" + name + "' has incorrect shape");
+        }
+        if (w.a->ne[1] != w.b->ne[0]) {
+            throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
+        }
+
+        // save tensor to adapter
+        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
+        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
+        ggml_set_name(tensor_a, w.a->name);
+        ggml_set_name(tensor_b, w.b->name);
+        adapter.ab_map[name] = llama_lora_weight(tensor_a, tensor_b);
+    }
+
+    // allocate tensors / buffers and zero
+    {
+        adapter.ctxs.reserve(ctx_map.size());
+        adapter.bufs.reserve(ctx_map.size());
+        for (auto & it : ctx_map) {
+            ggml_backend_buffer_type_t buft = it.first;
+            ggml_context * ctx_dev = it.second;
+            ggml_backend_buffer_ptr buf { ggml_backend_alloc_ctx_tensors_from_buft(ctx_dev, buft) };
+            if (!buf) {
+                throw std::runtime_error("failed to allocate buffer for lora adapter\n");
+            }
+            LLAMA_LOG_INFO("%s: %10s LoRA buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get())/1024.0/1024.0);
+            adapter.bufs.emplace_back(std::move(buf));
+        }
+    }
+
+    // set tensor data
+    {
+        llama_file gguf_file(path_lora, "rb");
+        std::vector<uint8_t> read_buf;
+        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
+            size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
+            size_t size = ggml_nbytes(orig);
+            read_buf.resize(size);
+            gguf_file.seek(offs, SEEK_SET);
+            gguf_file.read_raw(read_buf.data(), size);
+            ggml_backend_tensor_set(dev, read_buf.data(), 0, size);
+        };
+        for (auto & it : adapter.ab_map) {
+            auto orig = ab_map[it.first];
+            auto dev  = it.second;
+            set_tensor(orig.a, dev.a);
+            set_tensor(orig.b, dev.b);
+        }
+    }
+
+    LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
+}
+
+struct llama_lora_adapter * llama_lora_adapter_init(struct llama_model * model, const char * path_lora) {
+    struct llama_lora_adapter * adapter = new llama_lora_adapter();
+
+    try {
+        llama_lora_adapter_init_impl(*model, path_lora, *adapter);
+        return adapter;
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
+
+        delete adapter;
+    }
+
+    return nullptr;
+}

src/llama-adapter.h

@@ -0,0 +1,66 @@
+#pragma once
+
+#include "llama-impl.h"
+#include "llama-hparams.h"
+
+#include "ggml-cpp.h"
+
+#include <unordered_map>
+#include <vector>
+
+//
+// llama_adapter_cvec
+//
+
+// TODO: rename to llama_adapter_cvec
+struct llama_control_vector {
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    std::vector<struct ggml_tensor *> tensors; // per layer
+
+    int32_t layer_start = -1;
+    int32_t layer_end   = -1;
+
+    struct ggml_tensor * tensor_for(int il) const;
+
+    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const;
+};
+
+int32_t llama_control_vector_apply(
+        struct llama_control_vector & cvec,
+        const llama_model & model,
+        const float * data,
+        size_t len,
+        int32_t n_embd,
+        int32_t il_start,
+        int32_t il_end);
+
+//
+// llama_adapter_lora
+//
+
+// TODO: rename to llama_adapter_lora_weight
+struct llama_lora_weight {
+    struct ggml_tensor * a = nullptr;
+    struct ggml_tensor * b = nullptr;
+
+    llama_lora_weight() = default;
+    llama_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {}
+};
+
+// TODO: rename to llama_adapter_lora
+struct llama_lora_adapter {
+    // map tensor name to lora_a_b
+    std::unordered_map<std::string, struct llama_lora_weight> ab_map;
+
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    float alpha;
+
+    llama_lora_adapter() = default;
+    ~llama_lora_adapter() = default;
+
+    llama_lora_weight * get_weight(struct ggml_tensor * w);
+};

src/llama-arch.h

@@ -0,0 +1,391 @@
+#pragma once
+
+#include "ggml.h" // ggml_op
+
+#include <string>
+
+//
+// gguf constants (sync with gguf.py)
+//
+
+enum llm_arch {
+    LLM_ARCH_LLAMA,
+    LLM_ARCH_DECI,
+    LLM_ARCH_FALCON,
+    LLM_ARCH_BAICHUAN,
+    LLM_ARCH_GROK,
+    LLM_ARCH_GPT2,
+    LLM_ARCH_GPTJ,
+    LLM_ARCH_GPTNEOX,
+    LLM_ARCH_MPT,
+    LLM_ARCH_STARCODER,
+    LLM_ARCH_REFACT,
+    LLM_ARCH_BERT,
+    LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_JINA_BERT_V2,
+    LLM_ARCH_BLOOM,
+    LLM_ARCH_STABLELM,
+    LLM_ARCH_QWEN,
+    LLM_ARCH_QWEN2,
+    LLM_ARCH_QWEN2MOE,
+    LLM_ARCH_QWEN2VL,
+    LLM_ARCH_PHI2,
+    LLM_ARCH_PHI3,
+    LLM_ARCH_PLAMO,
+    LLM_ARCH_CODESHELL,
+    LLM_ARCH_ORION,
+    LLM_ARCH_INTERNLM2,
+    LLM_ARCH_MINICPM,
+    LLM_ARCH_MINICPM3,
+    LLM_ARCH_GEMMA,
+    LLM_ARCH_GEMMA2,
+    LLM_ARCH_STARCODER2,
+    LLM_ARCH_MAMBA,
+    LLM_ARCH_XVERSE,
+    LLM_ARCH_COMMAND_R,
+    LLM_ARCH_DBRX,
+    LLM_ARCH_OLMO,
+    LLM_ARCH_OLMO2,
+    LLM_ARCH_OLMOE,
+    LLM_ARCH_OPENELM,
+    LLM_ARCH_ARCTIC,
+    LLM_ARCH_DEEPSEEK,
+    LLM_ARCH_DEEPSEEK2,
+    LLM_ARCH_CHATGLM,
+    LLM_ARCH_BITNET,
+    LLM_ARCH_T5,
+    LLM_ARCH_T5ENCODER,
+    LLM_ARCH_JAIS,
+    LLM_ARCH_NEMOTRON,
+    LLM_ARCH_EXAONE,
+    LLM_ARCH_RWKV6,
+    LLM_ARCH_GRANITE,
+    LLM_ARCH_GRANITE_MOE,
+    LLM_ARCH_CHAMELEON,
+    LLM_ARCH_WAVTOKENIZER_DEC,
+    LLM_ARCH_UNKNOWN,
+};
+
+enum llm_kv {
+    LLM_KV_GENERAL_TYPE,
+    LLM_KV_GENERAL_ARCHITECTURE,
+    LLM_KV_GENERAL_QUANTIZATION_VERSION,
+    LLM_KV_GENERAL_ALIGNMENT,
+    LLM_KV_GENERAL_NAME,
+    LLM_KV_GENERAL_AUTHOR,
+    LLM_KV_GENERAL_VERSION,
+    LLM_KV_GENERAL_URL,
+    LLM_KV_GENERAL_DESCRIPTION,
+    LLM_KV_GENERAL_LICENSE,
+    LLM_KV_GENERAL_SOURCE_URL,
+    LLM_KV_GENERAL_SOURCE_HF_REPO,
+
+    LLM_KV_VOCAB_SIZE,
+    LLM_KV_CONTEXT_LENGTH,
+    LLM_KV_EMBEDDING_LENGTH,
+    LLM_KV_FEATURES_LENGTH,
+    LLM_KV_BLOCK_COUNT,
+    LLM_KV_LEADING_DENSE_BLOCK_COUNT,
+    LLM_KV_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
+    LLM_KV_USE_PARALLEL_RESIDUAL,
+    LLM_KV_TENSOR_DATA_LAYOUT,
+    LLM_KV_EXPERT_COUNT,
+    LLM_KV_EXPERT_USED_COUNT,
+    LLM_KV_EXPERT_SHARED_COUNT,
+    LLM_KV_EXPERT_WEIGHTS_SCALE,
+    LLM_KV_POOLING_TYPE,
+    LLM_KV_LOGIT_SCALE,
+    LLM_KV_DECODER_START_TOKEN_ID,
+    LLM_KV_ATTN_LOGIT_SOFTCAPPING,
+    LLM_KV_FINAL_LOGIT_SOFTCAPPING,
+    LLM_KV_SWIN_NORM,
+    LLM_KV_RESCALE_EVERY_N_LAYERS,
+    LLM_KV_TIME_MIX_EXTRA_DIM,
+    LLM_KV_TIME_DECAY_EXTRA_DIM,
+    LLM_KV_RESIDUAL_SCALE,
+    LLM_KV_EMBEDDING_SCALE,
+
+    LLM_KV_ATTENTION_HEAD_COUNT,
+    LLM_KV_ATTENTION_HEAD_COUNT_KV,
+    LLM_KV_ATTENTION_MAX_ALIBI_BIAS,
+    LLM_KV_ATTENTION_CLAMP_KQV,
+    LLM_KV_ATTENTION_KEY_LENGTH,
+    LLM_KV_ATTENTION_VALUE_LENGTH,
+    LLM_KV_ATTENTION_LAYERNORM_EPS,
+    LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,
+    LLM_KV_ATTENTION_GROUPNORM_EPS,
+    LLM_KV_ATTENTION_GROUPNORM_GROUPS,
+    LLM_KV_ATTENTION_CAUSAL,
+    LLM_KV_ATTENTION_Q_LORA_RANK,
+    LLM_KV_ATTENTION_KV_LORA_RANK,
+    LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
+    LLM_KV_ATTENTION_SLIDING_WINDOW,
+    LLM_KV_ATTENTION_SCALE,
+
+    LLM_KV_ROPE_DIMENSION_COUNT,
+    LLM_KV_ROPE_DIMENSION_SECTIONS,
+    LLM_KV_ROPE_FREQ_BASE,
+    LLM_KV_ROPE_SCALE_LINEAR,
+    LLM_KV_ROPE_SCALING_TYPE,
+    LLM_KV_ROPE_SCALING_FACTOR,
+    LLM_KV_ROPE_SCALING_ATTN_FACTOR,
+    LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
+    LLM_KV_ROPE_SCALING_FINETUNED,
+    LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
+
+    LLM_KV_SPLIT_NO,
+    LLM_KV_SPLIT_COUNT,
+    LLM_KV_SPLIT_TENSORS_COUNT,
+
+    LLM_KV_SSM_INNER_SIZE,
+    LLM_KV_SSM_CONV_KERNEL,
+    LLM_KV_SSM_STATE_SIZE,
+    LLM_KV_SSM_TIME_STEP_RANK,
+    LLM_KV_SSM_DT_B_C_RMS,
+
+    LLM_KV_WKV_HEAD_SIZE,
+
+    LLM_KV_TOKENIZER_MODEL,
+    LLM_KV_TOKENIZER_PRE,
+    LLM_KV_TOKENIZER_LIST,
+    LLM_KV_TOKENIZER_TOKEN_TYPE,
+    LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,
+    LLM_KV_TOKENIZER_SCORES,
+    LLM_KV_TOKENIZER_MERGES,
+    LLM_KV_TOKENIZER_BOS_ID,
+    LLM_KV_TOKENIZER_EOS_ID,
+    LLM_KV_TOKENIZER_EOT_ID,
+    LLM_KV_TOKENIZER_EOM_ID,
+    LLM_KV_TOKENIZER_UNK_ID,
+    LLM_KV_TOKENIZER_SEP_ID,
+    LLM_KV_TOKENIZER_PAD_ID,
+    LLM_KV_TOKENIZER_CLS_ID,
+    LLM_KV_TOKENIZER_MASK_ID,
+    LLM_KV_TOKENIZER_ADD_BOS,
+    LLM_KV_TOKENIZER_ADD_EOS,
+    LLM_KV_TOKENIZER_ADD_PREFIX,
+    LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
+    LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
+    LLM_KV_TOKENIZER_HF_JSON,
+    LLM_KV_TOKENIZER_RWKV,
+    LLM_KV_TOKENIZER_FIM_PRE_ID,
+    LLM_KV_TOKENIZER_FIM_SUF_ID,
+    LLM_KV_TOKENIZER_FIM_MID_ID,
+    LLM_KV_TOKENIZER_FIM_PAD_ID,
+    LLM_KV_TOKENIZER_FIM_REP_ID,
+    LLM_KV_TOKENIZER_FIM_SEP_ID,
+
+    LLM_KV_ADAPTER_TYPE,
+    LLM_KV_ADAPTER_LORA_ALPHA,
+
+    LLM_KV_POSNET_EMBEDDING_LENGTH,
+    LLM_KV_POSNET_BLOCK_COUNT,
+
+    LLM_KV_CONVNEXT_EMBEDDING_LENGTH,
+    LLM_KV_CONVNEXT_BLOCK_COUNT,
+
+    // deprecated:
+    LLM_KV_TOKENIZER_PREFIX_ID,
+    LLM_KV_TOKENIZER_SUFFIX_ID,
+    LLM_KV_TOKENIZER_MIDDLE_ID,
+};
+
+enum llm_tensor {
+    LLM_TENSOR_TOKEN_EMBD,
+    LLM_TENSOR_TOKEN_EMBD_NORM,
+    LLM_TENSOR_TOKEN_TYPES,
+    LLM_TENSOR_POS_EMBD,
+    LLM_TENSOR_OUTPUT,
+    LLM_TENSOR_OUTPUT_NORM,
+    LLM_TENSOR_ROPE_FREQS,
+    LLM_TENSOR_ROPE_FACTORS_LONG,
+    LLM_TENSOR_ROPE_FACTORS_SHORT,
+    LLM_TENSOR_ATTN_Q,
+    LLM_TENSOR_ATTN_K,
+    LLM_TENSOR_ATTN_V,
+    LLM_TENSOR_ATTN_QKV,
+    LLM_TENSOR_ATTN_OUT,
+    LLM_TENSOR_ATTN_NORM,
+    LLM_TENSOR_ATTN_NORM_2,
+    LLM_TENSOR_ATTN_OUT_NORM,
+    LLM_TENSOR_ATTN_POST_NORM,
+    LLM_TENSOR_ATTN_ROT_EMBD,
+    LLM_TENSOR_FFN_GATE_INP,
+    LLM_TENSOR_FFN_GATE_INP_SHEXP,
+    LLM_TENSOR_FFN_NORM,
+    LLM_TENSOR_FFN_POST_NORM,
+    LLM_TENSOR_FFN_GATE,
+    LLM_TENSOR_FFN_DOWN,
+    LLM_TENSOR_FFN_UP,
+    LLM_TENSOR_FFN_ACT,
+    LLM_TENSOR_FFN_DOWN_EXP,  // split experts for backward compatibility
+    LLM_TENSOR_FFN_GATE_EXP,
+    LLM_TENSOR_FFN_UP_EXP,
+    LLM_TENSOR_FFN_NORM_EXPS,
+    LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
+    LLM_TENSOR_FFN_GATE_EXPS,
+    LLM_TENSOR_FFN_UP_EXPS,
+    LLM_TENSOR_FFN_DOWN_SHEXP,
+    LLM_TENSOR_FFN_GATE_SHEXP,
+    LLM_TENSOR_FFN_UP_SHEXP,
+    LLM_TENSOR_ATTN_Q_NORM,
+    LLM_TENSOR_ATTN_K_NORM,
+    LLM_TENSOR_LAYER_OUT_NORM,
+    LLM_TENSOR_SSM_IN,
+    LLM_TENSOR_SSM_CONV1D,
+    LLM_TENSOR_SSM_X,
+    LLM_TENSOR_SSM_DT,
+    LLM_TENSOR_SSM_A,
+    LLM_TENSOR_SSM_D,
+    LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_TIME_MIX_W1,
+    LLM_TENSOR_TIME_MIX_W2,
+    LLM_TENSOR_TIME_MIX_LERP_X,
+    LLM_TENSOR_TIME_MIX_LERP_W,
+    LLM_TENSOR_TIME_MIX_LERP_K,
+    LLM_TENSOR_TIME_MIX_LERP_V,
+    LLM_TENSOR_TIME_MIX_LERP_R,
+    LLM_TENSOR_TIME_MIX_LERP_G,
+    LLM_TENSOR_TIME_MIX_FIRST,
+    LLM_TENSOR_TIME_MIX_DECAY,
+    LLM_TENSOR_TIME_MIX_DECAY_W1,
+    LLM_TENSOR_TIME_MIX_DECAY_W2,
+    LLM_TENSOR_TIME_MIX_KEY,
+    LLM_TENSOR_TIME_MIX_VALUE,
+    LLM_TENSOR_TIME_MIX_RECEPTANCE,
+    LLM_TENSOR_TIME_MIX_GATE,
+    LLM_TENSOR_TIME_MIX_LN,
+    LLM_TENSOR_TIME_MIX_OUTPUT,
+    LLM_TENSOR_CHANNEL_MIX_LERP_K,
+    LLM_TENSOR_CHANNEL_MIX_LERP_R,
+    LLM_TENSOR_CHANNEL_MIX_KEY,
+    LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,
+    LLM_TENSOR_CHANNEL_MIX_VALUE,
+    LLM_TENSOR_ATTN_Q_A,
+    LLM_TENSOR_ATTN_Q_B,
+    LLM_TENSOR_ATTN_KV_A_MQA,
+    LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_Q_A_NORM,
+    LLM_TENSOR_ATTN_KV_A_NORM,
+    LLM_TENSOR_ATTN_SUB_NORM,
+    LLM_TENSOR_FFN_SUB_NORM,
+    LLM_TENSOR_DEC_ATTN_NORM,
+    LLM_TENSOR_DEC_ATTN_Q,
+    LLM_TENSOR_DEC_ATTN_K,
+    LLM_TENSOR_DEC_ATTN_V,
+    LLM_TENSOR_DEC_ATTN_OUT,
+    LLM_TENSOR_DEC_ATTN_REL_B,
+    LLM_TENSOR_DEC_CROSS_ATTN_NORM,
+    LLM_TENSOR_DEC_CROSS_ATTN_Q,
+    LLM_TENSOR_DEC_CROSS_ATTN_K,
+    LLM_TENSOR_DEC_CROSS_ATTN_V,
+    LLM_TENSOR_DEC_CROSS_ATTN_OUT,
+    LLM_TENSOR_DEC_CROSS_ATTN_REL_B,
+    LLM_TENSOR_DEC_FFN_NORM,
+    LLM_TENSOR_DEC_FFN_GATE,
+    LLM_TENSOR_DEC_FFN_DOWN,
+    LLM_TENSOR_DEC_FFN_UP,
+    LLM_TENSOR_DEC_OUTPUT_NORM,
+    LLM_TENSOR_ENC_ATTN_NORM,
+    LLM_TENSOR_ENC_ATTN_Q,
+    LLM_TENSOR_ENC_ATTN_K,
+    LLM_TENSOR_ENC_ATTN_V,
+    LLM_TENSOR_ENC_ATTN_OUT,
+    LLM_TENSOR_ENC_ATTN_REL_B,
+    LLM_TENSOR_ENC_FFN_NORM,
+    LLM_TENSOR_ENC_FFN_GATE,
+    LLM_TENSOR_ENC_FFN_DOWN,
+    LLM_TENSOR_ENC_FFN_UP,
+    LLM_TENSOR_ENC_OUTPUT_NORM,
+    LLM_TENSOR_CLS,
+    LLM_TENSOR_CLS_OUT,
+    LLM_TENSOR_CONV1D,
+    LLM_TENSOR_CONVNEXT_DW,
+    LLM_TENSOR_CONVNEXT_NORM,
+    LLM_TENSOR_CONVNEXT_PW1,
+    LLM_TENSOR_CONVNEXT_PW2,
+    LLM_TENSOR_CONVNEXT_GAMMA,
+    LLM_TENSOR_POS_NET_CONV1,
+    LLM_TENSOR_POS_NET_CONV2,
+    LLM_TENSOR_POS_NET_NORM,
+    LLM_TENSOR_POS_NET_NORM1,
+    LLM_TENSOR_POS_NET_NORM2,
+    LLM_TENSOR_POS_NET_ATTN_NORM,
+    LLM_TENSOR_POS_NET_ATTN_Q,
+    LLM_TENSOR_POS_NET_ATTN_K,
+    LLM_TENSOR_POS_NET_ATTN_V,
+    LLM_TENSOR_POS_NET_ATTN_OUT,
+};
+
+enum llm_tensor_layer {
+    LLM_TENSOR_LAYER_INPUT,
+    LLM_TENSOR_LAYER_REPEATING,
+    LLM_TENSOR_LAYER_OUTPUT,
+};
+
+struct LLM_KV {
+    LLM_KV(llm_arch arch);
+
+    llm_arch arch;
+
+    std::string operator()(llm_kv kv) const;
+};
+
+// helper to handle gguf constants
+// usage:
+//
+//   const auto tn = LLM_TN(LLM_ARCH_LLAMA);
+//
+//   std::string name = tn(LLM_TENSOR_OUTPUT);                     -> "output"
+//   std::string name = tn(LLM_TENSOR_TOKEN_EMBD, "bias");         -> "token_embd.bias"
+//   std::string name = tn(LLM_TENSOR_ATTN_NORM, "weight", 3);     -> "blk.3.attn_norm.weight"
+//
+struct LLM_TN_IMPL {
+    const llm_arch arch;
+    const llm_tensor tensor;
+    const char * const suffix;
+    const int bid;
+    const int xid;
+
+    std::string str() const;
+
+    operator std::string() const {
+        return str();
+    }
+
+    friend bool operator==(const std::string & str, const LLM_TN_IMPL & tn) {
+        return str == tn.str();
+    }
+
+    friend bool operator!=(const std::string & str, const LLM_TN_IMPL & tn) {
+        return str != tn.str();
+    }
+};
+
+struct LLM_TN {
+    LLM_TN(llm_arch arch) : arch(arch) {}
+
+    llm_arch arch;
+
+    LLM_TN_IMPL operator()(llm_tensor tensor, const char * suffix, int bid = -1, int xid = -1) const {
+        return { arch, tensor, suffix, bid, xid };
+    }
+
+    LLM_TN_IMPL operator()(llm_tensor tensor, int bid = -1, int xid = -1) const {
+        return { arch, tensor, nullptr, bid, xid };
+    }
+};
+
+
+struct llm_tensor_info {
+    llm_tensor_layer layer;
+    ggml_op op;
+};
+
+const char * llm_arch_name(llm_arch arch);
+
+llm_arch llm_arch_from_string(const std::string & name);
+
+const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);

src/llama-batch.cpp

@@ -0,0 +1,368 @@
+#include "llama-batch.h"
+
+#include <cstring>
+#include <algorithm>
+
+llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
+    // clear empty sequences
+    // the previous ubatch is assumed to be gone,
+    // so nothing should refer to values in these sequences anymore.
+    for (size_t i = seq.size(); i-- > 0;) {
+        if (seq[i].length == 0) {
+            seq.pop_back();
+        } else {
+            break;
+        }
+    }
+    ubatch_token.resize(!has_embd ? n_ubatch : 0);
+    ubatch_embd.resize(has_embd ? n_embd * n_ubatch : 0);
+    ubatch_pos.resize(n_ubatch);
+    ubatch_n_seq_id.resize(n_ubatch);
+    ubatch_seq_id.resize(n_ubatch);
+    ubatch_output.resize(n_ubatch);
+    llama_ubatch ubatch = {
+        /*equal_seqs   =*/ true,
+        /*n_tokens     =*/ 0,
+        /*n_seq_tokens =*/ 0,
+        /*n_seqs       =*/ 0,
+        /*token        =*/ !has_embd ? ubatch_token.data() : nullptr,
+        /*embd         =*/ has_embd  ? ubatch_embd.data()  : nullptr,
+        /*pos          =*/ ubatch_pos.data(),
+        /*n_seq_id     =*/ ubatch_n_seq_id.data(),
+        /*seq_id       =*/ ubatch_seq_id.data(),
+        /*output       =*/ ubatch_output.data(),
+    };
+    return ubatch;
+}
+
+void llama_sbatch::add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length) {
+    GGML_ASSERT(batch != nullptr);
+    GGML_ASSERT(length <= seq.length);
+    // Can only add sequences of equal lengths to a batch,
+    // otherwise it isn't clear to which sequence a token belongs
+    GGML_ASSERT(seq.n_seq_id == 0 || ubatch.n_seqs == 0 || length == (size_t) ubatch.n_tokens / ubatch.n_seqs);
+    GGML_ASSERT((seq.n_seq_id != 0) == ubatch.equal_seqs);
+    // NOTE: loops are separated for cache-friendliness
+    if (batch->token) {
+        if (ubatch.equal_seqs) {
+            for (size_t i = 0; i < length; ++i) {
+                ubatch.token[ubatch.n_tokens + i] = batch->token[ids[seq.offset + i]];
+            }
+        } else {
+            // simple split
+            ubatch.token = batch->token + seq.offset;
+        }
+    } else {
+        ubatch.token = nullptr;
+    }
+    if (batch->embd) {
+        if (ubatch.equal_seqs) {
+            for (size_t i = 0; i < length; ++i) {
+                memcpy(
+                        ubatch.embd + (n_embd * (ubatch.n_tokens + i)),
+                        batch->embd + (n_embd * ids[seq.offset + i]),
+                        n_embd * sizeof(float)
+                      );
+            }
+        } else {
+            // simple split
+            ubatch.embd = batch->embd + (n_embd * seq.offset);
+        }
+    } else {
+        ubatch.embd = nullptr;
+    }
+    if (ubatch.equal_seqs) {
+        for (size_t i = 0; i < length; ++i) {
+            ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]];
+        }
+    } else {
+        // simple split
+        ubatch.pos = batch->pos + seq.offset;
+    }
+    if (ubatch.equal_seqs) {
+        ubatch.n_seq_id[ubatch.n_seqs] = seq.n_seq_id;
+        if (seq.seq_id) {
+            ubatch.seq_id[ubatch.n_seqs] = seq.seq_id;
+        }
+    } else {
+        // simple split
+        if (batch->n_seq_id) {
+            ubatch.n_seq_id = batch->n_seq_id + seq.offset;
+        } else {
+            for (size_t i = 0; i < length; ++i) {
+                ubatch.n_seq_id[ubatch.n_seqs + i] = 1;
+            }
+        }
+        if (batch->seq_id) {
+            ubatch.seq_id = batch->seq_id + seq.offset;
+        }
+    }
+    if (logits_all) {
+        for (size_t i = 0; i < length; ++i) {
+            ubatch.output[ubatch.n_tokens + i] = 1;
+            out_ids.push_back(ids[seq.offset + i]);
+        }
+    } else if (batch->logits) {
+        if (ubatch.equal_seqs) {
+            for (size_t i = 0; i < length; ++i) {
+                size_t id = ids[seq.offset + i];
+                int8_t is_output = batch->logits[id];
+                ubatch.output[ubatch.n_tokens + i] = is_output;
+                if (is_output) { out_ids.push_back(id); }
+            }
+        } else {
+            // simple split
+            ubatch.output = batch->logits + seq.offset;
+            for (size_t i = 0; i < length; ++i) {
+                if (ubatch.output[i] != 0) { out_ids.push_back(seq.offset + i); }
+            }
+        }
+    } else {
+        // only get last output
+        for (size_t i = 0; i < length; ++i) {
+            size_t id = ids[seq.offset + i];
+            int8_t is_last = id == ids.size() - 1;
+            ubatch.output[ubatch.n_tokens + i] = is_last;
+            if (is_last) { out_ids.push_back(id); }
+        }
+    }
+    if (ubatch.n_tokens == 0 && ubatch.n_seqs == 0) {
+        ubatch.n_seq_tokens = ubatch.equal_seqs ? length : 1;
+    }
+    ubatch.n_tokens += length;
+    ubatch.n_seqs += ubatch.equal_seqs ? 1 : length; // virtual sequences for simple splits
+    seq.offset += length;
+    seq.length -= length;
+    n_tokens -= length;
+    GGML_ASSERT(ubatch.n_tokens == ubatch.n_seq_tokens * ubatch.n_seqs);
+}
+
+llama_ubatch llama_sbatch::split_simple(size_t n_ubatch) {
+    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
+    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
+    ubatch.equal_seqs = false;
+    if (!seq.empty()) {
+        llama_sbatch_seq & s = seq[0];
+        size_t length = s.length < n_ubatch ? s.length : n_ubatch;
+        GGML_ASSERT(seq.size() == 1 && s.n_seq_id == 0); // don't mix with other splits
+        add_seq_to_ubatch(ubatch, s, length);
+    }
+    return ubatch;
+}
+
+llama_ubatch llama_sbatch::split_equal(size_t n_ubatch) {
+    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
+    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
+    if (!seq.empty()) {
+        size_t length = 0;
+        size_t n_tokens_in_ubatch = 0;
+        GGML_ASSERT(seq[0].n_seq_id > 0); // should not be mixed with simple splits
+                                          // smallest first, because it's easier to split this way;
+                                          // starting from the end to pop in constant time.
+        for (size_t i = seq.size(); i-- > 0;) {
+            llama_sbatch_seq & s = seq[i];
+            GGML_ASSERT(s.length > 0);
+            if (length == 0) {
+                length = s.length < n_ubatch ? s.length : n_ubatch;
+            }
+            add_seq_to_ubatch(ubatch, s, length);
+            n_tokens_in_ubatch += length;
+            // shared prompts can't be mixed with any of their sequences,
+            // so it's safer to compute them in their own ubatch
+            if (s.n_seq_id > 1) { break; }
+            // stop when there isn't enough space for another sequence
+            if (length + n_tokens_in_ubatch > n_ubatch) { break; }
+        }
+    }
+    return ubatch;
+}
+
+llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
+    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
+    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
+    if (!seq.empty()) {
+        llama_sbatch_seq & s = seq[seq.size() - 1];
+        size_t length = s.length < n_ubatch ? s.length : n_ubatch;
+        GGML_ASSERT(s.n_seq_id > 0); // should not be mixed with simple splits
+        add_seq_to_ubatch(ubatch, s, length);
+    }
+    return ubatch;
+}
+
+void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+    GGML_ASSERT(batch.n_tokens >= 0);
+    this->batch = &batch;
+    this->n_embd = n_embd;
+    this->logits_all = logits_all;
+
+    n_tokens = batch.n_tokens;
+    ids.resize(n_tokens);
+    out_ids.clear();
+    // TODO: reserve out_ids and seq
+
+    for (size_t i = 0; i < n_tokens; ++i) {
+        ids[i] = i;
+    }
+    if (simple_split) {
+        seq.resize(1);
+        llama_sbatch_seq & s = seq[0];
+        s.n_seq_id = 0;
+        s.seq_id = nullptr;
+        s.offset = 0;
+        s.length = n_tokens;
+        return;
+    }
+    std::sort(ids.begin(), ids.end(),
+            [&batch](size_t a, size_t b) {
+                int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
+                int32_t n_seq_b = batch.n_seq_id ? batch.n_seq_id[b] : 1;
+                // sort by seq_id, then by pos
+                if (n_seq_a == n_seq_b) {
+                    if (batch.seq_id) {
+                        for (int32_t i = 0; i < n_seq_a; ++i) {
+                            llama_seq_id seq_id_a = batch.seq_id[a][i];
+                            llama_seq_id seq_id_b = batch.seq_id[b][i];
+                            // smaller seq_ids go first
+                            if (seq_id_a != seq_id_b) {
+                                return seq_id_a < seq_id_b;
+                            }
+                        }
+                    }
+                    // when all else is equal, sort by pos
+                    if (batch.pos) {
+                        return batch.pos[a] < batch.pos[b];
+                    }
+                    // no pos, sort by id
+                    return a < b;
+                }
+                // shared prompts go first
+                return n_seq_a > n_seq_b;
+            }
+    );
+    // init seq
+    llama_sbatch_seq * last_seq = nullptr;
+
+    for (size_t i = 0; i < n_tokens; ++i) {
+        const size_t bi = ids[i];
+        const int32_t n_seqs = batch.n_seq_id[bi];
+        llama_seq_id * seq_ids = batch.seq_id[bi];
+        if (last_seq != nullptr) {
+            bool same = n_seqs == last_seq->n_seq_id;
+            for (int32_t j = 0; same && j < n_seqs; ++j) {
+                if (seq_ids[j] != last_seq->seq_id[j]) {
+                    same = false;
+                }
+            }
+            if (same) {
+                last_seq->length += 1;
+                continue;
+            }
+        }
+        llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1};
+        seq.push_back(new_seq);
+        last_seq = &seq.back();
+    }
+    // keep shared prompts first at the end, then sort by length descending.
+    std::sort(seq.begin(), seq.end(),
+            [](llama_sbatch_seq & a, llama_sbatch_seq & b) {
+                if (a.n_seq_id == b.n_seq_id) {
+                    return a.length > b.length;
+                }
+                return a.n_seq_id < b.n_seq_id;
+            }
+            );
+}
+
+llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0) {
+    batch = in_batch;
+    GGML_ASSERT(batch.n_tokens > 0);
+    if (!batch.pos) {
+        pos.resize(batch.n_tokens);
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            pos[i] = i + p0;
+        }
+        batch.pos = pos.data();
+    }
+    if (!batch.n_seq_id) {
+        n_seq_id.resize(batch.n_tokens);
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            n_seq_id[i] = seq_id_0.size();
+        }
+        batch.n_seq_id = n_seq_id.data();
+    }
+    if (!batch.seq_id) {
+        seq_id.resize(batch.n_tokens + 1);
+        seq_id[batch.n_tokens] = NULL;
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            seq_id[i] = seq_id_0.data();
+        }
+        batch.seq_id = seq_id.data();
+    }
+    if (!batch.logits) {
+        logits.resize(batch.n_tokens);
+        logits[logits.size() - 1] = true;
+        batch.logits = logits.data();
+    }
+}
+
+//
+// interface implementation
+//
+
+struct llama_batch llama_batch_get_one(
+             llama_token * tokens,
+                 int32_t   n_tokens) {
+    return {
+        /*n_tokens       =*/ n_tokens,
+        /*tokens         =*/ tokens,
+        /*embd           =*/ nullptr,
+        /*pos            =*/ nullptr,
+        /*n_seq_id       =*/ nullptr,
+        /*seq_id         =*/ nullptr,
+        /*logits         =*/ nullptr,
+    };
+}
+
+struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) {
+    llama_batch batch = {
+        /*n_tokens       =*/ 0,
+        /*tokens         =*/ nullptr,
+        /*embd           =*/ nullptr,
+        /*pos            =*/ nullptr,
+        /*n_seq_id       =*/ nullptr,
+        /*seq_id         =*/ nullptr,
+        /*logits         =*/ nullptr,
+    };
+
+    if (embd) {
+        batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
+    } else {
+        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
+    }
+
+    batch.pos      = (llama_pos *)     malloc(sizeof(llama_pos)      * n_tokens_alloc);
+    batch.n_seq_id = (int32_t *)       malloc(sizeof(int32_t)        * n_tokens_alloc);
+    batch.seq_id   = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * (n_tokens_alloc + 1));
+    for (int i = 0; i < n_tokens_alloc; ++i) {
+        batch.seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);
+    }
+    batch.seq_id[n_tokens_alloc] = nullptr;
+
+    batch.logits   = (int8_t *)        malloc(sizeof(int8_t)         * n_tokens_alloc);
+
+    return batch;
+}
+
+void llama_batch_free(struct llama_batch batch) {
+    if (batch.token)    free(batch.token);
+    if (batch.embd)     free(batch.embd);
+    if (batch.pos)      free(batch.pos);
+    if (batch.n_seq_id) free(batch.n_seq_id);
+    if (batch.seq_id) {
+        for (int i = 0; batch.seq_id[i] != nullptr; ++i) {
+            free(batch.seq_id[i]);
+        }
+        free(batch.seq_id);
+    }
+    if (batch.logits)   free(batch.logits);
+}

src/llama-batch.h

@@ -0,0 +1,88 @@
+#pragma once
+
+#include "llama.h"
+
+#include <array>
+#include <vector>
+
+// very similar to llama_batch,
+// but has more metadata about sequences
+struct llama_ubatch {
+    bool equal_seqs;
+    // TODO: whole_seqs for embeddings?
+
+    uint32_t n_tokens; // total tokens (n_seq_tokens * n_seqs)
+    uint32_t n_seq_tokens; // tokens per sequence
+    uint32_t n_seqs;
+
+    llama_token  *  token;    // [n_tokens]
+    float        *  embd;     // [n_embd, n_tokens]
+    llama_pos    *  pos;      // [n_tokens]
+    int32_t      *  n_seq_id; // [n_seqs]
+    llama_seq_id ** seq_id;   // [n_seqs]
+    int8_t       *  output;   // [n_tokens]
+};
+
+struct llama_sbatch_seq {
+    int32_t n_seq_id;
+
+    llama_seq_id * seq_id;
+
+    size_t offset;
+    size_t length;
+};
+
+// sequence-length-aware batch splitting
+struct llama_sbatch {
+    // tokens left in this batch
+    size_t n_tokens;
+
+    size_t n_embd;
+
+    bool logits_all; // TODO: remove once lctx.logits_all is removed too
+
+    // sorted indices into the batch
+    std::vector<size_t> ids;
+    // batch indices of the output
+    std::vector<size_t> out_ids;
+    std::vector<llama_sbatch_seq> seq;
+
+    const llama_batch * batch = nullptr;
+
+    // buffers for the ubatch
+    std::vector<llama_token>    ubatch_token;
+    std::vector<float>          ubatch_embd;
+    std::vector<llama_pos>      ubatch_pos;
+    std::vector<int32_t>        ubatch_n_seq_id;
+    std::vector<llama_seq_id *> ubatch_seq_id;
+    std::vector<int8_t>         ubatch_output;
+
+    llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false);
+
+    void add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length);
+
+    // simple split, unknown number of sequences of unequal lengths
+    llama_ubatch split_simple(size_t n_ubatch);
+
+    // make batches of equal-length sequences
+    llama_ubatch split_equal(size_t n_ubatch);
+
+    // sequence-wise split
+    llama_ubatch split_seq(size_t n_ubatch);
+
+    void from_batch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+};
+
+// temporary allocate memory for the input batch if needed
+struct llama_batch_allocr {
+    struct llama_batch batch;
+
+    std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id *> seq_id;
+    std::vector<int8_t>         logits;
+
+    // optionally fulfill the batch returned by llama_batch_get_one
+    llama_batch_allocr(struct llama_batch in_batch, llama_pos p0);
+};

src/llama-chat.cpp

@@ -0,0 +1,549 @@
+#include "llama-chat.h"
+
+#include "llama.h"
+
+#include <map>
+#include <sstream>
+
+#if __cplusplus >= 202000L
+    #define LU8(x) (const char*)(u8##x)
+#else
+    #define LU8(x) u8##x
+#endif
+
+// trim whitespace from the beginning and end of a string
+static std::string trim(const std::string & str) {
+    size_t start = 0;
+    size_t end = str.size();
+    while (start < end && isspace(str[start])) {
+        start += 1;
+    }
+    while (end > start && isspace(str[end - 1])) {
+        end -= 1;
+    }
+    return str.substr(start, end - start);
+}
+
+static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
+    { "chatml",            LLM_CHAT_TEMPLATE_CHATML            },
+    { "llama2",            LLM_CHAT_TEMPLATE_LLAMA_2           },
+    { "llama2-sys",        LLM_CHAT_TEMPLATE_LLAMA_2_SYS       },
+    { "llama2-sys-bos",    LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS   },
+    { "llama2-sys-strip",  LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP },
+    { "mistral-v1",        LLM_CHAT_TEMPLATE_MISTRAL_V1        },
+    { "mistral-v3",        LLM_CHAT_TEMPLATE_MISTRAL_V3        },
+    { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },
+    { "mistral-v7",        LLM_CHAT_TEMPLATE_MISTRAL_V7        },
+    { "phi3",              LLM_CHAT_TEMPLATE_PHI_3             },
+    { "falcon3",           LLM_CHAT_TEMPLATE_FALCON_3          },
+    { "zephyr",            LLM_CHAT_TEMPLATE_ZEPHYR            },
+    { "monarch",           LLM_CHAT_TEMPLATE_MONARCH           },
+    { "gemma",             LLM_CHAT_TEMPLATE_GEMMA             },
+    { "orion",             LLM_CHAT_TEMPLATE_ORION             },
+    { "openchat",          LLM_CHAT_TEMPLATE_OPENCHAT          },
+    { "vicuna",            LLM_CHAT_TEMPLATE_VICUNA            },
+    { "vicuna-orca",       LLM_CHAT_TEMPLATE_VICUNA_ORCA       },
+    { "deepseek",          LLM_CHAT_TEMPLATE_DEEPSEEK          },
+    { "deepseek2",         LLM_CHAT_TEMPLATE_DEEPSEEK_2        },
+    { "command-r",         LLM_CHAT_TEMPLATE_COMMAND_R         },
+    { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
+    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
+    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
+    { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
+    { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
+    { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
+    { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
+    { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
+};
+
+llm_chat_template llm_chat_template_from_str(const std::string & name) {
+    return LLM_CHAT_TEMPLATES.at(name);
+}
+
+llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
+    try {
+        return llm_chat_template_from_str(tmpl);
+    } catch (const std::out_of_range &) {
+        // ignore
+    }
+
+    auto tmpl_contains = [&tmpl](const char * haystack) -> bool {
+        return tmpl.find(haystack) != std::string::npos;
+    };
+    if (tmpl_contains("<|im_start|>")) {
+        return LLM_CHAT_TEMPLATE_CHATML;
+    } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {
+        if (tmpl_contains("[SYSTEM_PROMPT]")) {
+            return LLM_CHAT_TEMPLATE_MISTRAL_V7;
+        } else if (
+            // catches official 'v1' template
+            tmpl_contains("' [INST] ' + system_message")
+            // catches official 'v3' and 'v3-tekken' templates
+            || tmpl_contains("[AVAILABLE_TOOLS]")
+        ) {
+            // Official mistral 'v1', 'v3' and 'v3-tekken' templates
+            // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md
+            // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.md
+            if (tmpl_contains(" [INST]")) {
+                return LLM_CHAT_TEMPLATE_MISTRAL_V1;
+            } else if (tmpl_contains("\"[INST]\"")) {
+                return LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN;
+            }
+            return LLM_CHAT_TEMPLATE_MISTRAL_V3;
+        } else {
+            // llama2 template and its variants
+            // [variant] support system message
+            // See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2
+            bool support_system_message = tmpl_contains("<<SYS>>");
+            bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]");
+            bool strip_message = tmpl_contains("content.strip()");
+            if (strip_message) {
+                return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP;
+            } else if (add_bos_inside_history) {
+                return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS;
+            } else if (support_system_message) {
+                return LLM_CHAT_TEMPLATE_LLAMA_2_SYS;
+            } else {
+                return LLM_CHAT_TEMPLATE_LLAMA_2;
+            }
+        }
+    } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
+        return LLM_CHAT_TEMPLATE_PHI_3;
+    } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
+        return LLM_CHAT_TEMPLATE_FALCON_3;
+    } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
+        return LLM_CHAT_TEMPLATE_ZEPHYR;
+    } else if (tmpl_contains("bos_token + message['role']")) {
+        return LLM_CHAT_TEMPLATE_MONARCH;
+    } else if (tmpl_contains("<start_of_turn>")) {
+        return LLM_CHAT_TEMPLATE_GEMMA;
+    } else if (tmpl_contains("'\\n\\nAssistant: ' + eos_token")) {
+        // OrionStarAI/Orion-14B-Chat
+        return LLM_CHAT_TEMPLATE_ORION;
+    } else if (tmpl_contains("GPT4 Correct ")) {
+        // openchat/openchat-3.5-0106
+        return LLM_CHAT_TEMPLATE_OPENCHAT;
+    } else if (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: ")) {
+        // eachadea/vicuna-13b-1.1 (and Orca variant)
+        if (tmpl_contains("SYSTEM: ")) {
+            return LLM_CHAT_TEMPLATE_VICUNA_ORCA;
+        }
+        return LLM_CHAT_TEMPLATE_VICUNA;
+    } else if (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>")) {
+        // deepseek-ai/deepseek-coder-33b-instruct
+        return LLM_CHAT_TEMPLATE_DEEPSEEK;
+    } else if (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>")) {
+        // CohereForAI/c4ai-command-r-plus
+        return LLM_CHAT_TEMPLATE_COMMAND_R;
+    } else if (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>")) {
+        return LLM_CHAT_TEMPLATE_LLAMA_3;
+    } else if (tmpl_contains("[gMASK]sop")) {
+        // chatglm3-6b
+        return LLM_CHAT_TEMPLATE_CHATGML_3;
+    } else if (tmpl_contains("[gMASK]<sop>")) {
+        return LLM_CHAT_TEMPLATE_CHATGML_4;
+    } else if (tmpl_contains(LU8("<用户>"))) {
+        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
+        return LLM_CHAT_TEMPLATE_MINICPM;
+    } else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {
+        return LLM_CHAT_TEMPLATE_DEEPSEEK_2;
+    } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) {
+        // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
+        // EXAONE-3.0-7.8B-Instruct
+        return LLM_CHAT_TEMPLATE_EXAONE_3;
+    } else if (tmpl_contains("rwkv-world")) {
+        return LLM_CHAT_TEMPLATE_RWKV_WORLD;
+    } else if (tmpl_contains("<|start_of_role|>")) {
+        return LLM_CHAT_TEMPLATE_GRANITE;
+    } else if (tmpl_contains("message['role'] + additional_special_tokens[0] + message['content'] + additional_special_tokens[1]")) {
+        return LLM_CHAT_TEMPLATE_GIGACHAT;
+    } else if (tmpl_contains("<|role_start|>")) {
+        return LLM_CHAT_TEMPLATE_MEGREZ;
+    }
+    return LLM_CHAT_TEMPLATE_UNKNOWN;
+}
+
+// Simple version of "llama_apply_chat_template" that only works with strings
+// This function uses heuristic checks to determine commonly used template. It is not a jinja parser.
+int32_t llm_chat_apply_template(
+    llm_chat_template tmpl,
+    const std::vector<const llama_chat_message *> & chat,
+    std::string & dest, bool add_ass) {
+    // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527
+    std::stringstream ss;
+    if (tmpl == LLM_CHAT_TEMPLATE_CHATML) {
+        // chatml template
+        for (auto message : chat) {
+            ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n";
+        }
+        if (add_ass) {
+            ss << "<|im_start|>assistant\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) {
+        // Official mistral 'v7' template
+        // See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7
+        for (auto message : chat) {
+            std::string role(message->role);
+            std::string content(message->content);
+            if (role == "system") {
+                ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]";
+            } else if (role == "user") {
+                ss << "[INST] " << content << "[/INST]";
+            }
+            else {
+                ss << " " << content << "</s>";
+            }
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1
+            || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3
+            || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN) {
+        // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md
+        // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.md
+        std::string leading_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1 ? " " : "";
+        std::string trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN ? "" : " ";
+        bool trim_assistant_message = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3;
+        bool is_inside_turn = false;
+        for (auto message : chat) {
+            if (!is_inside_turn) {
+                ss << leading_space << "[INST]" << trailing_space;
+                is_inside_turn = true;
+            }
+            std::string role(message->role);
+            std::string content(message->content);
+            if (role == "system") {
+                ss << content << "\n\n";
+            } else if (role == "user") {
+                ss << content << leading_space << "[/INST]";
+            } else {
+                ss << trailing_space << (trim_assistant_message ? trim(content) : content) << "</s>";
+                is_inside_turn = false;
+            }
+        }
+    } else if (
+            tmpl == LLM_CHAT_TEMPLATE_LLAMA_2
+            || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS
+            || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS
+            || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP) {
+        // llama2 template and its variants
+        // [variant] support system message
+        // See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2
+        bool support_system_message = tmpl != LLM_CHAT_TEMPLATE_LLAMA_2;
+        // [variant] add BOS inside history
+        bool add_bos_inside_history = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS;
+        // [variant] trim spaces from the input message
+        bool strip_message = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP;
+        // construct the prompt
+        bool is_inside_turn = true; // skip BOS at the beginning
+        ss << "[INST] ";
+        for (auto message : chat) {
+            std::string content = strip_message ? trim(message->content) : message->content;
+            std::string role(message->role);
+            if (!is_inside_turn) {
+                is_inside_turn = true;
+                ss << (add_bos_inside_history ? "<s>[INST] " : "[INST] ");
+            }
+            if (role == "system") {
+                if (support_system_message) {
+                    ss << "<<SYS>>\n" << content << "\n<</SYS>>\n\n";
+                } else {
+                    // if the model does not support system message, we still include it in the first message, but without <<SYS>>
+                    ss << content << "\n";
+                }
+            } else if (role == "user") {
+                ss << content << " [/INST]";
+            } else {
+                ss << content << "</s>";
+                is_inside_turn = false;
+            }
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_PHI_3) {
+        // Phi 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>\n" << message->content << "<|end|>\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_FALCON_3) {
+        // Falcon 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>\n" << message->content << "\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_ZEPHYR) {
+        // zephyr template
+        for (auto message : chat) {
+            ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MONARCH) {
+        // mlabonne/AlphaMonarch-7B template (the <s> is included inside history)
+        for (auto message : chat) {
+            std::string bos = (message == chat.front()) ? "" : "<s>"; // skip BOS for first message
+            ss << bos << message->role << "\n" << message->content << "</s>\n";
+        }
+        if (add_ass) {
+            ss << "<s>assistant\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GEMMA) {
+        // google/gemma-7b-it
+        std::string system_prompt = "";
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                // there is no system message for gemma, but we will merge it with user prompt, so nothing is broken
+                system_prompt = trim(message->content);
+                continue;
+            }
+            // in gemma, "assistant" is "model"
+            role = role == "assistant" ? "model" : message->role;
+            ss << "<start_of_turn>" << role << "\n";
+            if (!system_prompt.empty() && role != "model") {
+                ss << system_prompt << "\n\n";
+                system_prompt = "";
+            }
+            ss << trim(message->content) << "<end_of_turn>\n";
+        }
+        if (add_ass) {
+            ss << "<start_of_turn>model\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_ORION) {
+        // OrionStarAI/Orion-14B-Chat
+        std::string system_prompt = "";
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                // there is no system message support, we will merge it with user prompt
+                system_prompt = message->content;
+                continue;
+            } else if (role == "user") {
+                ss << "Human: ";
+                if (!system_prompt.empty()) {
+                    ss << system_prompt << "\n\n";
+                    system_prompt = "";
+                }
+                ss << message->content << "\n\nAssistant: </s>";
+            } else {
+                ss << message->content << "</s>";
+            }
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_OPENCHAT) {
+        // openchat/openchat-3.5-0106,
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "<|end_of_turn|>";
+            } else {
+                role[0] = toupper(role[0]);
+                ss << "GPT4 Correct " << role << ": " << message->content << "<|end_of_turn|>";
+            }
+        }
+        if (add_ass) {
+            ss << "GPT4 Correct Assistant:";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_VICUNA || tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) {
+        // eachadea/vicuna-13b-1.1 (and Orca variant)
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                // Orca-Vicuna variant uses a system prefix
+                if (tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) {
+                    ss << "SYSTEM: " << message->content << "\n";
+                } else {
+                    ss << message->content << "\n\n";
+                }
+            } else if (role == "user") {
+                ss << "USER: " << message->content << "\n";
+            } else if (role == "assistant") {
+                ss << "ASSISTANT: " << message->content << "</s>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "ASSISTANT:";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK) {
+        // deepseek-ai/deepseek-coder-33b-instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content;
+            } else if (role == "user") {
+                ss << "### Instruction:\n" << message->content << "\n";
+            } else if (role == "assistant") {
+                ss << "### Response:\n" << message->content << "\n<|EOT|>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "### Response:\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_COMMAND_R) {
+        // CohereForAI/c4ai-command-r-plus
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            } else if (role == "user") {
+                ss << "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            } else if (role == "assistant") {
+                ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            }
+        }
+        if (add_ass) {
+            ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA_3) {
+        // Llama 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|start_header_id|>" << role << "<|end_header_id|>\n\n" << trim(message->content) << "<|eot_id|>";
+        }
+        if (add_ass) {
+            ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) {
+        // chatglm3-6b
+        ss << "[gMASK]" << "sop";
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n " << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) {
+        ss << "[gMASK]" << "<sop>";
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n" << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {
+        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "user") {
+                ss << LU8("<用户>");
+                ss << trim(message->content);
+                ss << "<AI>";
+            } else {
+                ss << trim(message->content);
+            }
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK_2) {
+        // DeepSeek-V2
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << message->content << "\n\n";
+            } else if (role == "assistant") {
+                ss << "Assistant: " << message->content << LU8("<｜end▁of▁sentence｜>");
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_3) {
+        // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
+        // EXAONE-3.0-7.8B-Instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "[|system|]" << trim(message->content) << "[|endofturn|]\n";
+            } else if (role == "user") {
+                ss << "[|user|]" << trim(message->content) << "\n";
+            } else if (role == "assistant") {
+                ss << "[|assistant|]" << trim(message->content) << "[|endofturn|]\n";
+            }
+        }
+        if (add_ass) {
+            ss << "[|assistant|]";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
+        // this template requires the model to have "\n\n" as EOT token
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "user") {
+                ss << "User: " << message->content << "\n\nAssistant:";
+            } else {
+                ss << message->content << "\n\n";
+            }
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) {
+        // IBM Granite template
+        for (const auto & message : chat) {
+            std::string role(message->role);
+            ss << "<|start_of_role|>" << role << "<|end_of_role|>";
+            if (role == "assistant_tool_call") {
+                ss << "<|tool_call|>";
+            }
+            ss << message->content << "<|end_of_text|>\n";
+        }
+        if (add_ass) {
+            ss << "<|start_of_role|>assistant<|end_of_role|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {
+        // GigaChat template
+        bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";
+
+        // Handle system message if present
+        if (has_system) {
+            ss << "<s>" << chat[0]->content << "<|message_sep|>";
+        } else {
+            ss << "<s>";
+        }
+
+        // Process remaining messages
+        for (size_t i = has_system ? 1 : 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "user") {
+                ss << "user<|role_sep|>" << chat[i]->content << "<|message_sep|>"
+                << "available functions<|role_sep|>[]<|message_sep|>";
+            } else if (role == "assistant") {
+                ss << "assistant<|role_sep|>" << chat[i]->content << "<|message_sep|>";
+            }
+        }
+
+        // Add generation prompt if needed
+        if (add_ass) {
+            ss << "assistant<|role_sep|>";
+        }
+    }  else if (tmpl == LLM_CHAT_TEMPLATE_MEGREZ) {
+        // Megrez template
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|role_start|>" << role << "<|role_end|>" << message->content << "<|turn_end|>";
+        }
+
+        if (add_ass) {
+            ss << "<|role_start|>assistant<|role_end|>";
+        }
+    } else {
+        // template not supported
+        return -1;
+    }
+    dest = ss.str();
+    return dest.size();
+}
+
+// public interface
+
+int32_t llama_chat_builtin_templates(const char ** output, size_t len) {
+    auto it = LLM_CHAT_TEMPLATES.begin();
+    for (size_t i = 0; i < std::min(len, LLM_CHAT_TEMPLATES.size()); i++) {
+        output[i] = it->first.c_str();
+        std::advance(it, 1);
+    }
+    return (int32_t) LLM_CHAT_TEMPLATES.size();
+}
+

src/llama-chat.h

@@ -0,0 +1,50 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <cstdint>
+
+enum llm_chat_template {
+    LLM_CHAT_TEMPLATE_CHATML,
+    LLM_CHAT_TEMPLATE_LLAMA_2,
+    LLM_CHAT_TEMPLATE_LLAMA_2_SYS,
+    LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS,
+    LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP,
+    LLM_CHAT_TEMPLATE_MISTRAL_V1,
+    LLM_CHAT_TEMPLATE_MISTRAL_V3,
+    LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
+    LLM_CHAT_TEMPLATE_MISTRAL_V7,
+    LLM_CHAT_TEMPLATE_PHI_3,
+    LLM_CHAT_TEMPLATE_FALCON_3,
+    LLM_CHAT_TEMPLATE_ZEPHYR,
+    LLM_CHAT_TEMPLATE_MONARCH,
+    LLM_CHAT_TEMPLATE_GEMMA,
+    LLM_CHAT_TEMPLATE_ORION,
+    LLM_CHAT_TEMPLATE_OPENCHAT,
+    LLM_CHAT_TEMPLATE_VICUNA,
+    LLM_CHAT_TEMPLATE_VICUNA_ORCA,
+    LLM_CHAT_TEMPLATE_DEEPSEEK,
+    LLM_CHAT_TEMPLATE_DEEPSEEK_2,
+    LLM_CHAT_TEMPLATE_COMMAND_R,
+    LLM_CHAT_TEMPLATE_LLAMA_3,
+    LLM_CHAT_TEMPLATE_CHATGML_3,
+    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_MINICPM,
+    LLM_CHAT_TEMPLATE_EXAONE_3,
+    LLM_CHAT_TEMPLATE_RWKV_WORLD,
+    LLM_CHAT_TEMPLATE_GRANITE,
+    LLM_CHAT_TEMPLATE_GIGACHAT,
+    LLM_CHAT_TEMPLATE_MEGREZ,
+    LLM_CHAT_TEMPLATE_UNKNOWN,
+};
+
+struct llama_chat_message;
+
+llm_chat_template llm_chat_template_from_str(const std::string & name);
+
+llm_chat_template llm_chat_detect_template(const std::string & tmpl);
+
+int32_t llm_chat_apply_template(
+    llm_chat_template tmpl,
+    const std::vector<const llama_chat_message *> & chat,
+    std::string & dest, bool add_ass);

src/llama-context.h

@@ -0,0 +1,128 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-batch.h"
+#include "llama-cparams.h"
+#include "llama-model.h"
+#include "llama-kv-cache.h"
+#include "llama-adapter.h"
+
+#include "ggml-cpp.h"
+
+#include <map>
+#include <unordered_map>
+#include <vector>
+#include <set>
+
+struct llama_context {
+    llama_context(const llama_model & model)
+        : model(model)
+        , t_start_us(model.t_start_us)
+        , t_load_us(model.t_load_us) {}
+
+    const struct llama_model & model;
+
+    struct llama_cparams        cparams;
+    struct llama_sbatch         sbatch;  // TODO: revisit if needed
+    struct llama_kv_cache       kv_self;
+    struct llama_control_vector cvec;
+
+    std::unordered_map<struct llama_lora_adapter *, float> lora_adapters;
+
+    std::vector<ggml_backend_ptr> backends;
+    std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
+
+    ggml_backend_t backend_cpu = nullptr;
+
+    ggml_threadpool_t threadpool       = nullptr;
+    ggml_threadpool_t threadpool_batch = nullptr;
+
+    bool has_evaluated_once = false;
+
+    mutable int64_t t_start_us;
+    mutable int64_t t_load_us;
+    mutable int64_t t_p_eval_us = 0;
+    mutable int64_t t_eval_us   = 0;
+
+    mutable int64_t t_compute_start_us = 0;
+    mutable int64_t n_queued_tokens = 0;
+
+    mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
+    mutable int32_t n_eval   = 0; // number of eval calls
+
+    // host buffer for the model output (logits and embeddings)
+    ggml_backend_buffer_ptr buf_output;
+
+    // decode output (2-dimensional array: [n_outputs][n_vocab])
+    size_t  logits_size = 0; // capacity (of floats) for logits
+    float * logits      = nullptr;
+
+    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
+    size_t  output_size = 0; // capacity (of tokens positions) for the output buffers
+    int32_t n_outputs   = 0; // number of actually-used outputs in the current ubatch or last logical batch
+
+    bool logits_all = false;
+
+    // embeddings output (2-dimensional array: [n_outputs][n_embd])
+    // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
+    size_t  embd_size = 0; // capacity (of floats) for embeddings
+    float * embd      = nullptr;
+
+    // sequence embeddings output (map of [n_embd] vectors)
+    // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
+    std::map<llama_seq_id, std::vector<float>> embd_seq;
+
+    // whether we are computing encoder output or decoder output
+    bool is_encoding = false;
+
+    // TODO: find a better way to accommodate mutli-dimension position encoding methods
+    // number of position id each token get, 1 for each token in most cases.
+    // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate.
+    int n_pos_per_token = 1;
+
+    // output of the encoder part of the encoder-decoder models
+    std::vector<float> embd_enc;
+    std::vector<std::set<llama_seq_id>> seq_ids_enc;
+
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
+    ggml_backend_sched_ptr sched;
+
+    ggml_abort_callback abort_callback      = nullptr;
+    void *              abort_callback_data = nullptr;
+
+    // input tensors
+    struct ggml_tensor * inp_tokens;        // I32 [n_batch]
+    struct ggml_tensor * inp_embd;          // F32 [n_embd, n_batch]
+    struct ggml_tensor * inp_pos;           // I32 [n_batch]
+    struct ggml_tensor * inp_out_ids;       // I32 [n_outputs]
+    struct ggml_tensor * inp_KQ_mask;       // F32 [kv_size, n_batch]
+    struct ggml_tensor * inp_KQ_mask_swa;   // F32 [kv_size, n_batch]
+    struct ggml_tensor * inp_K_shift;       // I32 [kv_size]
+    struct ggml_tensor * inp_mean;          // F32 [n_batch, n_batch]
+    struct ggml_tensor * inp_cls;           // I32 [n_batch]
+    struct ggml_tensor * inp_s_copy;        // I32 [kv_size]
+    struct ggml_tensor * inp_s_mask;        // F32 [1, n_kv]
+    struct ggml_tensor * inp_s_seq;         // I32 [n_kv, n_batch]
+    struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
+    struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
+    struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+};
+
+// TODO: make these methods of llama_context
+void llama_set_k_shift(struct llama_context & lctx);
+
+void llama_set_s_copy(struct llama_context & lctx);
+
+void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch);
+
+// Make sure enough space is available for outputs.
+// Returns max number of outputs for which space was reserved.
+size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs);
+
+// make the outputs have the same order they had in the user-provided batch
+void llama_output_reorder(struct llama_context & ctx);
+
+// For internal test use
+// TODO: remove
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(struct llama_context * ctx);

src/llama-cparams.cpp

@@ -0,0 +1 @@
+#include "llama-cparams.h"

src/llama-cparams.h

@@ -0,0 +1,37 @@
+#pragma once
+
+#include "llama.h"
+
+#include <cstdint>
+
+struct llama_cparams {
+    uint32_t n_ctx;           // context size used during inference
+    uint32_t n_batch;
+    uint32_t n_ubatch;
+    uint32_t n_seq_max;
+    int      n_threads;       // number of threads to use for generation
+    int      n_threads_batch; // number of threads to use for batch processing
+
+    float rope_freq_base;
+    float rope_freq_scale;
+
+    uint32_t n_ctx_orig_yarn;
+    // These hyperparameters are not exposed in GGUF, because all
+    // existing YaRN models use the same values for them.
+    float yarn_ext_factor;
+    float yarn_attn_factor;
+    float yarn_beta_fast;
+    float yarn_beta_slow;
+    float defrag_thold;
+
+    bool embeddings;
+    bool causal_attn;
+    bool offload_kqv;
+    bool flash_attn;
+    bool no_perf;
+
+    enum llama_pooling_type pooling_type;
+
+    ggml_backend_sched_eval_callback cb_eval;
+    void * cb_eval_user_data;
+};

src/llama-grammar.cpp

@@ -1,5 +1,6 @@
 #include "llama-grammar.h"
 
+#include "llama-impl.h"
 #include "llama-vocab.h"
 #include "llama-sampling.h"
 

src/llama-grammar.h

@@ -1,8 +1,10 @@
 #pragma once
 
-#include "llama-impl.h"
+#include "llama.h"
 
 #include <map>
+#include <string>
+#include <vector>
 
 struct llama_vocab;
 

src/llama-hparams.cpp

@@ -0,0 +1,71 @@
+#include "llama-hparams.h"
+
+#include "ggml.h"
+
+uint32_t llama_hparams::n_head(uint32_t il) const {
+    if (il < n_layer) {
+        return n_head_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_head_kv(uint32_t il) const {
+    if (il < n_layer) {
+        return n_head_kv_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_ff(uint32_t il) const {
+    if (il < n_layer) {
+        return n_ff_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_gqa(uint32_t il) const {
+    const uint32_t n_head    = this->n_head(il);
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    if (n_head_kv == 0) {
+        return 0;
+    }
+
+    return n_head/n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    return n_embd_head_k * n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    return n_embd_head_v * n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_k_s() const {
+    if (wkv_head_size != 0) {
+        // for RWKV models
+        return 2 * n_embd;
+    }
+
+    // TODO: maybe support other convolution strides than 1
+    // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
+    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
+}
+
+uint32_t llama_hparams::n_embd_v_s() const {
+    if (wkv_head_size != 0) {
+        // corresponds to RWKV's wkv_states size
+        return n_embd * wkv_head_size;
+    }
+
+    // corresponds to Mamba's ssm_states size
+    return ssm_d_state * ssm_d_inner;
+}

src/llama-hparams.h

@@ -0,0 +1,132 @@
+#pragma once
+
+#include "llama.h"
+
+#include <array>
+
+// bump if necessary
+#define LLAMA_MAX_LAYERS  512
+#define LLAMA_MAX_EXPERTS 160  // DeepSeekV2
+
+struct llama_hparams_posnet {
+    uint32_t n_embd;
+    uint32_t n_layer;
+};
+
+struct llama_hparams_convnext {
+    uint32_t n_embd;
+    uint32_t n_layer;
+};
+
+struct llama_hparams {
+    bool vocab_only;
+    bool rope_finetuned;
+    bool use_par_res;
+    bool swin_norm;
+
+    uint32_t n_vocab = 0;
+    uint32_t n_ctx_train; // context size the model was trained on
+    uint32_t n_embd;
+    uint32_t n_embd_features = 0;
+    uint32_t n_layer;
+    uint32_t n_rot;
+    uint32_t n_swa = 0; // sliding window attention (SWA)
+    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
+    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
+    uint32_t n_expert = 0;
+    uint32_t n_expert_used = 0;
+    uint32_t n_vocab_type = 0; // for BERT-style token types
+    uint32_t n_rel_attn_bkts = 0;
+
+    // for WavTokenizer
+    struct llama_hparams_posnet   posnet;
+    struct llama_hparams_convnext convnext;
+
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+
+    uint32_t n_layer_dense_lead = 0;
+    uint32_t n_lora_q           = 0;
+    uint32_t n_lora_kv          = 0;
+    uint32_t n_ff_exp           = 0;
+    uint32_t n_ff_shexp         = 0;
+    uint32_t n_expert_shared    = 0;
+    uint32_t n_norm_groups      = 0;
+
+    float expert_weights_scale = 0.0;
+
+    float f_norm_eps;
+    float f_norm_rms_eps;
+    float f_norm_group_eps;
+
+    float f_attn_logit_softcapping  = 50.0f;
+    float f_final_logit_softcapping = 30.0f;
+
+    // for RWKV
+    uint32_t rescale_every_n_layers = 0;
+    uint32_t time_mix_extra_dim     = 0;
+    uint32_t time_decay_extra_dim   = 0;
+    uint32_t wkv_head_size          = 0;
+
+    float    rope_attn_factor = 1.0f;
+    float    rope_freq_base_train;
+    float    rope_freq_scale_train;
+    uint32_t n_ctx_orig_yarn;
+    float    rope_yarn_log_mul;
+
+    std::array<int, 4> rope_sections;
+
+    // for State Space Models
+    uint32_t ssm_d_conv  = 0;
+    uint32_t ssm_d_inner = 0;
+    uint32_t ssm_d_state = 0;
+    uint32_t ssm_dt_rank = 0;
+
+    bool ssm_dt_b_c_rms = false;
+
+    float f_clamp_kqv      = 0.0f;
+    float f_max_alibi_bias = 0.0f;
+    float f_logit_scale    = 0.0f;
+
+    // Additional scale factors (Granite/Granite MoE)
+    float f_residual_scale  = 0.0f;
+    float f_embedding_scale = 0.0f;
+    float f_attention_scale = 0.0f;
+
+    bool causal_attn   = true;
+    bool use_alibi     = false;
+    bool attn_soft_cap = false;
+
+    // needed by encoder-decoder models (e.g. T5, FLAN-T5)
+    // ref: https://github.com/ggerganov/llama.cpp/pull/8141
+    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
+
+    enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
+    enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
+    enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
+
+    uint32_t n_head(uint32_t il = 0) const;
+
+    uint32_t n_head_kv(uint32_t il = 0) const;
+
+    uint32_t n_ff(uint32_t il = 0) const;
+
+    uint32_t n_gqa(uint32_t il = 0) const;
+
+    // dimension of key embeddings across all k-v heads
+    uint32_t n_embd_k_gqa(uint32_t il = 0) const;
+
+    // dimension of value embeddings across all k-v heads
+    uint32_t n_embd_v_gqa(uint32_t il = 0) const;
+
+    // dimension of the rolling state embeddings
+    // corresponds to Mamba's conv_states size or RWKV's token_shift states size
+    uint32_t n_embd_k_s() const;
+
+    // dimension of the recurrent state embeddings
+    uint32_t n_embd_v_s() const;
+};
+
+static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+

src/llama-impl.cpp

@@ -0,0 +1,166 @@
+#include "llama-impl.h"
+
+#include "llama.h"
+
+#include <cinttypes>
+#include <climits>
+#include <cstdarg>
+#include <cstring>
+#include <vector>
+#include <sstream>
+
+struct llama_logger_state {
+    ggml_log_callback log_callback = llama_log_callback_default;
+    void * log_callback_user_data = nullptr;
+};
+
+static llama_logger_state g_logger_state;
+
+time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+
+time_meas::~time_meas() {
+        if (t_start_us >= 0) {
+            t_acc += ggml_time_us() - t_start_us;
+        }
+    }
+
+void llama_log_set(ggml_log_callback log_callback, void * user_data) {
+    ggml_log_set(log_callback, user_data);
+    g_logger_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
+    g_logger_state.log_callback_user_data = user_data;
+}
+
+static void llama_log_internal_v(ggml_log_level level, const char * format, va_list args) {
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
+    } else {
+        char * buffer2 = new char[len + 1];
+        vsnprintf(buffer2, len + 1, format, args_copy);
+        buffer2[len] = 0;
+        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
+        delete[] buffer2;
+    }
+    va_end(args_copy);
+}
+
+void llama_log_internal(ggml_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    llama_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}
+
+void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+std::string format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
+    for (size_t i = 1; i < ne.size(); i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
+    }
+    return buf;
+}
+
+std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
+    }
+    return buf;
+}
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return format("unknown type %d", type);
+    }
+}
+
+std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        replace_all(val, "\\", "\\\\");
+                        replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}

src/llama-impl.h

@@ -1,10 +1,9 @@
 #pragma once
 
-#include "llama.h"
+#include "ggml.h" // for ggml_log_level
 
 #include <string>
 #include <vector>
-#include <stdexcept>
 
 #ifdef __GNUC__
 #ifdef __MINGW32__
@@ -35,147 +34,28 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 // helpers
 //
 
-struct time_meas {
+template <typename T>
-    time_meas(int64_t & t_acc, bool disable = false) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+struct no_init {
+    T value;
+    no_init() { /* do nothing */ }
+};
 
-    ~time_meas() {
+struct time_meas {
-        if (t_start_us >= 0) {
+    time_meas(int64_t & t_acc, bool disable = false);
-            t_acc += ggml_time_us() - t_start_us;
+    ~time_meas();
-        }
-    }
 
     const int64_t t_start_us;
 
     int64_t & t_acc;
 };
 
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+void replace_all(std::string & s, const std::string & search, const std::string & replace);
-    if (search.empty()) {
-        return;
-    }
-    std::string builder;
-    builder.reserve(s.length());
-    size_t pos = 0;
-    size_t last_pos = 0;
-    while ((pos = s.find(search, last_pos)) != std::string::npos) {
-        builder.append(s, last_pos, pos - last_pos);
-        builder.append(replace);
-        last_pos = pos + search.length();
-    }
-    builder.append(s, last_pos, std::string::npos);
-    s = std::move(builder);
-}
 
-const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+// TODO: rename to llama_format ?
-    struct llama_context * ctx
+LLAMA_ATTRIBUTE_FORMAT(1, 2)
-);
+std::string format(const char * fmt, ...);
 
-// the ring buffer works similarly to std::deque, but with a fixed capacity
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne);
-template<typename T>
+std::string llama_format_tensor_shape(const struct ggml_tensor * t);
-struct ring_buffer {
-    ring_buffer(size_t cap) : capacity(cap), data(cap) {}
 
-    T & front() {
+std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
-        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 (capacity == 0) {
-            throw std::runtime_error("ring buffer: capacity is zero");
-        }
-
-        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;
-    }
-
-    //T & operator[](size_t i) {
-    //    if (i >= sz) {
-    //        throw std::runtime_error("ring buffer: index out of bounds");
-    //    }
-    //    return data[(first + i) % capacity];
-    //}
-
-    //const T & at(size_t i) const {
-    //    if (i >= sz) {
-    //        throw std::runtime_error("ring buffer: index out of bounds");
-    //    }
-    //    return data[(first + i) % capacity];
-    //}
-
-    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;
-};

src/llama-kv-cache.cpp

@@ -0,0 +1,718 @@
+#include "llama-kv-cache.h"
+
+#include "llama-impl.h"
+#include "llama-batch.h"
+#include "llama-cparams.h"
+#include "llama-model.h"
+
+#include <algorithm>
+#include <limits>
+#include <map>
+
+static const llama_kv_cache_slot_info llama_kv_cache_slot_info_failed{false};
+
+uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams) {
+    // the FA kernels require padding to avoid extra runtime boundary checks
+    return cparams.flash_attn ? 256u : 32u;
+}
+
+bool llama_kv_cache_init(
+             struct llama_kv_cache & cache,
+                 const llama_model & model,
+               const llama_cparams & cparams,
+                         ggml_type   type_k,
+                         ggml_type   type_v,
+                          uint32_t   kv_size,
+                              bool   offload) {
+    const struct llama_hparams & hparams = model.hparams;
+
+    const int32_t n_layer = hparams.n_layer;
+
+    cache.has_shift = false;
+
+    cache.recurrent = llama_model_is_recurrent(&model);
+    cache.v_trans   = !cache.recurrent && !cparams.flash_attn;
+    cache.can_shift = !cache.recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
+
+    LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n",
+            __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, cache.can_shift);
+
+    cache.head = 0;
+    cache.size = kv_size;
+    cache.used = 0;
+
+    cache.type_k = type_k;
+    cache.type_v = type_v;
+
+    cache.cells.clear();
+    cache.cells.resize(kv_size);
+
+    // create a context for each buffer type
+    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
+    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
+        auto it = ctx_map.find(buft);
+        if (it == ctx_map.end()) {
+            struct ggml_init_params params = {
+                /*.mem_size   =*/ size_t(2u*n_layer*ggml_tensor_overhead()),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ggml_context * ctx = ggml_init(params);
+            if (!ctx) {
+                return nullptr;
+            }
+            ctx_map[buft] = ctx;
+            cache.ctxs.emplace_back(ctx);
+            return ctx;
+        }
+        return it->second;
+    };
+
+    cache.k_l.reserve(n_layer);
+    cache.v_l.reserve(n_layer);
+
+    for (int i = 0; i < n_layer; i++) {
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
+        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
+
+        LLAMA_LOG_DEBUG("%s: layer %d: n_embd_k_gqa = %d, n_embd_v_gqa = %d\n", __func__, i, n_embd_k_gqa, n_embd_v_gqa);
+
+        ggml_backend_buffer_type_t buft;
+        if (offload) {
+            auto * dev = model.dev_layer.at(i).dev;
+            buft = ggml_backend_dev_buffer_type(dev);
+        } else {
+            buft = ggml_backend_cpu_buffer_type();
+        }
+        ggml_context * ctx = ctx_for_buft(buft);
+
+        if (!ctx) {
+            LLAMA_LOG_ERROR("%s: failed to create ggml context for kv cache\n", __func__);
+            return false;
+        }
+
+        ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
+        ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
+        ggml_format_name(k, "cache_k_l%d", i);
+        ggml_format_name(v, "cache_v_l%d", i);
+        cache.k_l.push_back(k);
+        cache.v_l.push_back(v);
+    }
+
+    // allocate tensors and initialize the buffers to avoid NaNs in the padding
+    for (auto it : ctx_map) {
+        auto * buft = it.first;
+        auto * ctx  = it.second;
+
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+        if (!buf) {
+            LLAMA_LOG_ERROR("%s: failed to allocate buffer for kv cache\n", __func__);
+            return false;
+        }
+        ggml_backend_buffer_clear(buf, 0);
+        LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
+        cache.bufs.emplace_back(buf);
+    }
+
+    return true;
+}
+
+struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
+           struct llama_kv_cache & cache,
+       const struct llama_ubatch & batch) {
+    const uint32_t n_tokens = batch.n_tokens;
+    const uint32_t n_seqs   = batch.n_seqs;
+    const uint32_t n_seq_tokens = batch.n_seq_tokens;
+
+    if (cache.recurrent) {
+        // For recurrent state architectures (like Mamba or RWKV),
+        // each cache cell can store the state for a whole sequence.
+        // A slot should be always be contiguous.
+
+        // can only process batches with an equal number of new tokens in each sequence
+        GGML_ASSERT(batch.equal_seqs);
+
+        int32_t min = cache.size - 1;
+        int32_t max = 0;
+
+        // everything should fit if all seq_ids are smaller than the max
+        for (uint32_t s = 0; s < n_seqs; ++s) {
+            const uint32_t n_seq_id = batch.n_seq_id[s];
+            for (uint32_t j = 0; j < n_seq_id; ++j) {
+                const llama_seq_id seq_id = batch.seq_id[s][j];
+
+                if (seq_id < 0 || (uint32_t) seq_id >= cache.size) {
+                    // too big seq_id
+                    // TODO: would it be possible to resize the cache instead?
+                    LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, cache.size);
+                    return llama_kv_cache_slot_info_failed;
+                }
+                if (j > 0) {
+                    llama_kv_cell & seq = cache.cells[seq_id];
+                    if (seq.tail >= 0) {
+                        llama_kv_cell & cell = cache.cells[seq.tail];
+                        // clear cells from seq_ids that become shared
+                        // (should not normally happen, but let's handle it anyway)
+                        cell.seq_id.erase(seq_id);
+                        seq.tail = -1;
+                        if (cell.seq_id.empty()) {
+                            cell.pos = -1;
+                            cell.src = -1;
+                            cache.used -= 1;
+                        }
+                    }
+                }
+            }
+        }
+
+#ifndef NDEBUG
+        {
+            std::vector<int32_t> tails_verif;
+            tails_verif.assign(cache.size, -1);
+            for (uint32_t i = 0; i < cache.size; ++i) {
+                llama_kv_cell & cell = cache.cells[i];
+                for (llama_seq_id seq_id : cell.seq_id) {
+                    if (tails_verif[seq_id] != -1) {
+                        LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]);
+                    }
+                    tails_verif[seq_id] = i;
+                }
+            }
+            for (uint32_t i = 0; i < cache.size; ++i) {
+                if (tails_verif[i] != cache.cells[i].tail) {
+                    LLAMA_LOG_ERROR("%s: wrong tail for seq_id %d, (%d instead of %d)\n", __func__, i, cache.cells[i].tail, tails_verif[i]);
+                }
+            }
+        }
+#endif
+
+        // find next empty cell
+        uint32_t next_empty_cell = cache.head;
+
+        for (uint32_t i = 0; i < cache.size; ++i) {
+            if (next_empty_cell >= cache.size) { next_empty_cell -= cache.size; }
+            llama_kv_cell & cell = cache.cells[next_empty_cell];
+            if (cell.is_empty()) { break; }
+            next_empty_cell += 1;
+        }
+
+        // find usable cell range
+        for (uint32_t s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = batch.seq_id[s][0];
+            llama_kv_cell & seq_meta = cache.cells[seq_id];
+            bool has_cell = false;
+            if (seq_meta.tail >= 0) {
+                llama_kv_cell & cell = cache.cells[seq_meta.tail];
+                GGML_ASSERT(cell.has_seq_id(seq_id));
+                // does this seq_id "own" the cell?
+                if (cell.seq_id.size() == 1) { has_cell = true; }
+            }
+            if (!has_cell) {
+                llama_kv_cell & empty_cell = cache.cells[next_empty_cell];
+                GGML_ASSERT(empty_cell.is_empty());
+                // copy old tail into the empty cell
+                if (seq_meta.tail >= 0) {
+                    llama_kv_cell & orig_cell = cache.cells[seq_meta.tail];
+                    empty_cell.pos = orig_cell.pos;
+                    empty_cell.src = orig_cell.src;
+                    orig_cell.seq_id.erase(seq_id);
+                    empty_cell.seq_id.insert(seq_id); // will be overwritten
+                }
+                seq_meta.tail = next_empty_cell;
+                // find next empty cell
+                if (s + 1 < n_seqs) {
+                    next_empty_cell += 1;
+                    for (uint32_t i = 0; i < cache.size; ++i) {
+                        if (next_empty_cell >= cache.size) { next_empty_cell -= cache.size; }
+                        llama_kv_cell & cell = cache.cells[next_empty_cell];
+                        if (cell.is_empty()) { break; }
+                        next_empty_cell += 1;
+                    }
+                }
+            }
+            if (min > seq_meta.tail) { min = seq_meta.tail; }
+            if (max < seq_meta.tail) { max = seq_meta.tail; }
+        }
+
+        // gather and re-order
+        for (uint32_t s = 0; s < n_seqs; ++s) {
+            int32_t dst_id = s + min;
+            int32_t src_id = cache.cells[batch.seq_id[s][0]].tail;
+            if (dst_id != src_id) {
+                llama_kv_cell & dst_cell = cache.cells[dst_id];
+                llama_kv_cell & src_cell = cache.cells[src_id];
+
+                std::swap(dst_cell.pos, src_cell.pos);
+                std::swap(dst_cell.src, src_cell.src);
+                std::swap(dst_cell.seq_id, src_cell.seq_id);
+
+                // swap tails (assuming they NEVER overlap)
+                for (const llama_seq_id seq_id : src_cell.seq_id) {
+                    cache.cells[seq_id].tail = src_id;
+                }
+                for (const llama_seq_id seq_id : dst_cell.seq_id) {
+                    cache.cells[seq_id].tail = dst_id;
+                }
+            }
+        }
+
+        // update the pos of the used seqs
+        for (uint32_t s = 0; s < n_seqs; ++s) {
+            const llama_pos last_pos = batch.pos[n_seq_tokens * s + n_seq_tokens - 1];
+            int32_t cell_id = s + min;
+            llama_kv_cell & cell = cache.cells[cell_id];
+
+            if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) {
+                // What should happen when the pos backtracks or skips a value?
+                // Clearing the state mid-batch would require special-casing which isn't done.
+                LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n",
+                    __func__, last_pos, cell.pos, batch.seq_id[s][0], n_seq_tokens);
+            }
+            cell.pos = last_pos;
+            cell.seq_id.clear();
+            for (int32_t j = 0; j < batch.n_seq_id[s]; ++j) {
+                const llama_seq_id seq_id = batch.seq_id[s][j];
+                cell.seq_id.insert(seq_id);
+                cache.cells[seq_id].tail = cell_id;
+            }
+        }
+
+        // allow getting the range of used cells, from head to head + n
+        cache.head = min;
+        cache.n    = max - min + 1;
+        cache.used = std::count_if(cache.cells.begin(), cache.cells.end(),
+            [](const llama_kv_cell& cell){ return !cell.is_empty(); });
+
+        // sanity check
+        return llama_kv_cache_slot_info(cache.n >= n_seqs);
+    }
+    // otherwise, one cell per token.
+
+    if (n_tokens > cache.size) {
+        LLAMA_LOG_ERROR("%s: n_tokens=%d > cache.size=%d\n", __func__, n_tokens, cache.size);
+        return llama_kv_cache_slot_info_failed;
+    }
+
+    uint32_t n_tested = 0;
+
+    while (true) {
+        if (cache.head + n_tokens > cache.size) {
+            n_tested += cache.size - cache.head;
+            cache.head = 0;
+            continue;
+        }
+
+        bool found = true;
+        for (uint32_t i = 0; i < n_tokens; i++) {
+            if (cache.cells[cache.head + i].pos >= 0) {
+                found = false;
+                cache.head += i + 1;
+                n_tested   += i + 1;
+                break;
+            }
+        }
+
+        if (found) {
+            break;
+        }
+
+        if (n_tested >= cache.size) {
+            //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
+            return llama_kv_cache_slot_info_failed;
+        }
+    }
+
+    for (uint32_t s = 0; s < n_seqs; s++) {
+        for (uint32_t i = 0; i < n_seq_tokens; ++i) {
+            uint32_t k = s*n_seq_tokens + i;
+            cache.cells[cache.head + k].pos = batch.pos[k];
+
+            for (int32_t j = 0; j < batch.n_seq_id[s]; j++) {
+                cache.cells[cache.head + k].seq_id.insert(batch.seq_id[s][j]);
+            }
+        }
+    }
+
+    cache.used += n_tokens;
+
+    return llama_kv_cache_slot_info(cache.head, cache.head + n_tokens);
+}
+
+uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) {
+    for (uint32_t i = cache.size; i > 0; --i) {
+        const llama_kv_cell & cell = cache.cells[i - 1];
+
+        if (cell.pos >= 0 && !cell.is_empty()) {
+            return i;
+        }
+    }
+
+    return 0;
+}
+
+void llama_kv_cache_clear(struct llama_kv_cache & cache) {
+    for (int32_t i = 0; i < (int32_t) cache.size; ++i) {
+        cache.cells[i].pos = -1;
+        cache.cells[i].seq_id.clear();
+        cache.cells[i].src = -1;
+        cache.cells[i].tail = -1;
+    }
+    cache.head = 0;
+    cache.used = 0;
+
+    for (auto & buf : cache.bufs) {
+        ggml_backend_buffer_clear(buf.get(), 0);
+    }
+}
+
+bool llama_kv_cache_seq_rm(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1) {
+    uint32_t new_head = cache.size;
+
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+
+    // models like Mamba or RWKV can't have a state partially erased
+    if (cache.recurrent) {
+        if (seq_id >= (int64_t) cache.size) {
+            // could be fatal
+            return false;
+        }
+        if (0 <= seq_id) {
+            int32_t & tail_id = cache.cells[seq_id].tail;
+            if (tail_id >= 0) {
+                const llama_kv_cell & cell = cache.cells[tail_id];
+                // partial intersection is invalid
+                if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) {
+                    return false;
+                }
+                // invalidate tails which will be cleared
+                if (p0 <= cell.pos && cell.pos < p1) {
+                    tail_id = -1;
+                }
+            }
+        } else {
+            // seq_id is negative, then the range should include everything or nothing
+            if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits<llama_pos>::max())) {
+                return false;
+            }
+        }
+    }
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            if (seq_id < 0) {
+                cache.cells[i].seq_id.clear();
+            } else if (cache.cells[i].has_seq_id(seq_id)) {
+                cache.cells[i].seq_id.erase(seq_id);
+            } else {
+                continue;
+            }
+            if (cache.cells[i].is_empty()) {
+                // keep count of the number of used cells
+                if (cache.cells[i].pos >= 0) cache.used--;
+
+                cache.cells[i].pos = -1;
+                cache.cells[i].src = -1;
+                if (new_head == cache.size) new_head = i;
+            }
+        }
+    }
+
+    // If we freed up a slot, set head to it so searching can start there.
+    if (new_head != cache.size && new_head < cache.head) cache.head = new_head;
+
+    return true;
+}
+
+void llama_kv_cache_seq_cp(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id_src,
+                 llama_seq_id   seq_id_dst,
+                    llama_pos   p0,
+                    llama_pos   p1) {
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+
+    if (cache.recurrent) {
+        if ((uint32_t) seq_id_dst < cache.size && (uint32_t) seq_id_src < cache.size) {
+            llama_kv_cell & tail_src = cache.cells[seq_id_src];
+            llama_kv_cell & tail_dst = cache.cells[seq_id_dst];
+            if (tail_dst.tail >= 0) {
+                // clear destination seq_id if it wasn't empty
+                llama_kv_cell & cell_dst = cache.cells[tail_dst.tail];
+
+                cell_dst.seq_id.erase(seq_id_dst);
+                tail_dst.tail = -1;
+                if (cell_dst.seq_id.empty()) {
+                    cell_dst.pos = -1;
+                    cell_dst.delta = -1;
+                    cell_dst.src = -1;
+                    cache.used -= 1;
+                }
+            }
+            if (tail_src.tail >= 0) {
+                llama_kv_cell & cell_src = cache.cells[tail_src.tail];
+
+                cell_src.seq_id.insert(seq_id_dst);
+                tail_dst.tail = tail_src.tail;
+            }
+        }
+
+        return;
+    }
+    // otherwise, this is the KV cache of a Transformer-like model
+
+    cache.head = 0;
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id_src) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            cache.cells[i].seq_id.insert(seq_id_dst);
+        }
+    }
+}
+
+void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id seq_id) {
+    uint32_t new_head = cache.size;
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.recurrent && (llama_seq_id) i != seq_id) {
+            cache.cells[i].tail = -1;
+        }
+        if (!cache.cells[i].has_seq_id(seq_id)) {
+            if (cache.cells[i].pos >= 0) cache.used--;
+            cache.cells[i].pos = -1;
+            cache.cells[i].src = -1;
+            cache.cells[i].seq_id.clear();
+            if (new_head == cache.size) new_head = i;
+        } else {
+            cache.cells[i].seq_id.clear();
+            cache.cells[i].seq_id.insert(seq_id);
+        }
+    }
+
+    // If we freed up a slot, set head to it so searching can start there.
+    if (new_head != cache.size && new_head < cache.head) cache.head = new_head;
+}
+
+void llama_kv_cache_seq_add(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1,
+                    llama_pos   delta) {
+    uint32_t new_head = cache.size;
+
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+    // If there is no range then return early to avoid looping over the cache.
+    if (p0 == p1) return;
+
+    if (cache.recurrent) {
+        // for Mamba-like or RWKV models, only the pos needs to be shifted
+        if (0 <= seq_id && seq_id < (int64_t) cache.size) {
+            const int32_t tail_id = cache.cells[seq_id].tail;
+            if (tail_id >= 0) {
+                llama_kv_cell & cell = cache.cells[tail_id];
+                if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) {
+                    cell.pos += delta;
+                }
+            }
+        }
+        return;
+    }
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            cache.has_shift = true;
+            cache.cells[i].pos   += delta;
+            cache.cells[i].delta += delta;
+
+            if (cache.cells[i].pos < 0) {
+                if (!cache.cells[i].is_empty()) {
+                    cache.used--;
+                }
+                cache.cells[i].pos = -1;
+                cache.cells[i].seq_id.clear();
+                if (new_head == cache.size) {
+                    new_head = i;
+                }
+            }
+        }
+    }
+
+    // If we freed up a slot, set head to it so searching can start there.
+    // Otherwise we just start the next search from the beginning.
+    cache.head = new_head != cache.size ? new_head : 0;
+}
+
+void llama_kv_cache_seq_div(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1,
+                          int   d) {
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+    // If there is no range then return early to avoid looping over the cache.
+    if (p0 == p1) return;
+
+    if (cache.recurrent) {
+        // for Mamba-like or RWKV models, only the pos needs to be changed
+        if (0 <= seq_id && seq_id < (int64_t) cache.size) {
+            const int32_t tail_id = cache.cells[seq_id].tail;
+            if (tail_id >= 0) {
+                llama_kv_cell & cell = cache.cells[tail_id];
+                if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) {
+                    cell.pos /= d;
+                }
+            }
+        }
+        return;
+    }
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            cache.has_shift = true;
+
+            {
+                llama_pos p_old = cache.cells[i].pos;
+                cache.cells[i].pos   /= d;
+                cache.cells[i].delta += cache.cells[i].pos - p_old;
+            }
+        }
+    }
+}
+
+llama_pos llama_kv_cache_seq_pos_max(struct llama_kv_cache & cache, llama_seq_id seq_id) {
+    llama_pos result = 0;
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id)) {
+            result = std::max(result, cache.cells[i].pos);
+        }
+    }
+
+    return result;
+}
+
+void llama_kv_cache_defrag(struct llama_kv_cache & cache) {
+    if (!cache.recurrent) {
+        cache.do_defrag = true;
+    }
+}
+
+int32_t llama_get_kv_cache_token_count(const struct llama_kv_cache & kv) {
+    int result = 0;
+
+    for (uint32_t i = 0; i < kv.size; i++) {
+        result += kv.cells[i].seq_id.size();
+    }
+
+    return result;
+}
+
+int32_t llama_get_kv_cache_used_cells(const struct llama_kv_cache & kv) {
+    return kv.used;
+}
+
+bool llama_kv_cache_can_shift(const struct llama_kv_cache & kv) {
+    return kv.can_shift;
+}
+
+//
+// kv cache view
+//
+
+struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache & kv, int32_t n_seq_max) {
+    struct llama_kv_cache_view result = {
+        /*.n_cells            = */ 0,
+        /*.n_seq_max          = */ n_seq_max,
+        /*.token_count        = */ 0,
+        /*.used_cells         = */ llama_get_kv_cache_used_cells(kv),
+        /*.max_contiguous     = */ 0,
+        /*.max_contiguous_idx = */ -1,
+        /*.cells              = */ nullptr,
+        /*.cells_sequences    = */ nullptr,
+    };
+
+    return result;
+}
+
+void llama_kv_cache_view_free(struct llama_kv_cache_view * view) {
+    if (view->cells != nullptr) {
+        free(view->cells);
+        view->cells = nullptr;
+    }
+    if (view->cells_sequences != nullptr) {
+        free(view->cells_sequences);
+        view->cells_sequences = nullptr;
+    }
+}
+
+void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_kv_cache & kv) {
+    if (uint32_t(view->n_cells) < kv.size || view->cells == nullptr) {
+        view->n_cells = int32_t(kv.size);
+        void * p = realloc(view->cells, sizeof(struct llama_kv_cache_view_cell) * view->n_cells);
+        GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells");
+        view->cells = (struct llama_kv_cache_view_cell *)p;
+        p = realloc(view->cells_sequences, sizeof(llama_seq_id) * view->n_seq_max * view->n_cells);
+        GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells sequences");
+        view->cells_sequences = (llama_seq_id *)p;
+    }
+
+    const std::vector<llama_kv_cell> & kv_cells = kv.cells;
+    llama_kv_cache_view_cell * c_curr = view->cells;
+    llama_seq_id * cs_curr = view->cells_sequences;
+    int32_t used_cells = 0;
+    int32_t token_count = 0;
+    int32_t curr_contig_idx = -1;
+    uint32_t max_contig = 0;
+    int32_t max_contig_idx = -1;
+
+    for (int32_t i = 0; i < int32_t(kv.size); i++, c_curr++, cs_curr += view->n_seq_max) {
+        const size_t curr_size = kv_cells[i].seq_id.size();
+        token_count += curr_size;
+        c_curr->pos = kv_cells[i].pos + kv_cells[i].delta;
+
+        if (curr_size > 0) {
+            if (curr_contig_idx >= 0 && uint32_t(i - curr_contig_idx) > max_contig) {
+                max_contig = i - curr_contig_idx;
+                max_contig_idx = curr_contig_idx;
+            }
+            curr_contig_idx = -1;
+        } else if (curr_contig_idx < 0) {
+            curr_contig_idx = i;
+        }
+
+        int seq_idx = 0;
+        for (const llama_seq_id it : kv_cells[i].seq_id) {
+            if (seq_idx >= view->n_seq_max) {
+                break;
+            }
+            cs_curr[seq_idx] = it;
+            seq_idx++;
+        }
+        if (seq_idx != 0) {
+            used_cells++;
+        }
+        for (; seq_idx < view->n_seq_max; seq_idx++) {
+            cs_curr[seq_idx] = -1;
+        }
+    }
+    if (curr_contig_idx >= 0 && kv_cells.size() - curr_contig_idx > max_contig) {
+        max_contig_idx = curr_contig_idx;
+        max_contig = kv_cells.size() - curr_contig_idx;
+    }
+    view->max_contiguous = max_contig;
+    view->max_contiguous_idx = max_contig_idx;
+    view->token_count = token_count;
+    view->used_cells = used_cells;
+    if (uint32_t(used_cells) != kv.used) {
+        LLAMA_LOG_ERROR("%s: used cells mismatch. kv_cache says %d but we calculated %d\n",
+            __func__, kv.used, used_cells);
+    }
+}

src/llama-kv-cache.h

@@ -0,0 +1,218 @@
+#pragma once
+
+#include "llama.h"
+
+#include "ggml-cpp.h"
+
+#include <set>
+#include <vector>
+
+struct llama_kv_cell {
+    llama_pos pos   = -1;
+    llama_pos delta = 0;
+    int32_t   src   = -1; // used by recurrent state models to copy states
+    int32_t   tail  = -1;
+
+    std::set<llama_seq_id> seq_id;
+
+    bool has_seq_id(const llama_seq_id & id) const {
+        return seq_id.find(id) != seq_id.end();
+    }
+
+    bool is_empty() const {
+        return seq_id.empty();
+    }
+
+    bool is_same_seq(const llama_kv_cell & other) const {
+        return seq_id == other.seq_id;
+    }
+};
+
+// ring-buffer of cached KV data
+struct llama_kv_cache {
+    bool has_shift = false;
+    bool do_defrag = false;
+    bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token
+    bool v_trans   = true;  // the value tensor is transposed
+    bool can_shift = false;
+
+    // Note: The value of head isn't only used to optimize searching
+    // for a free KV slot. llama_decode_internal also uses it, so it
+    // cannot be freely changed after a slot has been allocated.
+    uint32_t head = 0;
+    uint32_t size = 0;
+    uint32_t used = 0; // used cells (i.e. at least one seq_id)
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    ggml_type type_k = GGML_TYPE_F16;
+    ggml_type type_v = GGML_TYPE_F16;
+
+    std::vector<llama_kv_cell> cells;
+
+    std::vector<struct ggml_tensor *> k_l; // per layer
+    std::vector<struct ggml_tensor *> v_l;
+
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    size_t total_size() const {
+        size_t size = 0;
+        for (const auto & buf : bufs) {
+            size += ggml_backend_buffer_get_size(buf.get());
+        }
+
+        return size;
+    }
+
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos max_pos() const {
+        llama_pos max_pos = -1;
+        for (const auto & cell : cells) {
+            max_pos = std::max(max_pos, cell.pos);
+        }
+
+        return max_pos;
+    }
+};
+
+// a structure holds information about the slot found in llama_kv_cache_find_slot
+struct llama_kv_cache_slot_info {
+    std::pair<uint32_t, uint32_t> boundaries; // slot boundaries [begin, end)
+    bool found = false;                       // the slot was found
+
+    explicit llama_kv_cache_slot_info(bool found_) : found{found_} {}
+    llama_kv_cache_slot_info(uint32_t begin, uint32_t end) : boundaries{begin, end}, found{true} {}
+
+    operator bool() const { return found; }
+};
+
+// TODO: maybe not needed
+uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams);
+
+bool llama_kv_cache_init(
+        struct llama_kv_cache & cache,
+            const llama_model & model,
+          const llama_cparams & cparams,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                     uint32_t   kv_size,
+                         bool   offload);
+
+// find an empty slot of size "n_tokens" in the cache
+// updates the cache head
+// returns a structure holding information about the slot found
+// Note: On success, it's important that cache.head points
+// to the first cell of the slot.
+struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
+           struct llama_kv_cache & cache,
+       const struct llama_ubatch & batch);
+
+// find how many cells are currently in use
+uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache);
+
+void llama_kv_cache_clear(struct llama_kv_cache & cache);
+
+bool llama_kv_cache_seq_rm(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1);
+
+void llama_kv_cache_seq_cp(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id_src,
+                 llama_seq_id   seq_id_dst,
+                    llama_pos   p0,
+                    llama_pos   p1);
+
+void llama_kv_cache_seq_keep(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id);
+
+void llama_kv_cache_seq_add(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1,
+                    llama_pos   delta);
+
+void llama_kv_cache_seq_div(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1,
+                          int   d);
+
+llama_pos llama_kv_cache_seq_pos_max(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id);
+
+void llama_kv_cache_defrag(struct llama_kv_cache & cache);
+
+int32_t llama_get_kv_cache_token_count(const struct llama_kv_cache & kv);
+
+int32_t llama_get_kv_cache_used_cells(const struct llama_kv_cache & kv);
+
+bool llama_kv_cache_can_shift(const struct llama_kv_cache & kv);
+
+//
+// kv cache view
+//
+
+struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache & kv, int32_t n_seq_max);
+
+void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_kv_cache & kv);
+
+//
+// kv cache restore
+//
+
+// saves the kv_cache state for future recovery.
+// used to rollback llama_kv_cache_find_slot changes.
+struct llama_kv_slot_restorer {
+    struct llama_kv_cache_state {
+        uint32_t head = 0;
+        uint32_t n    = 0;
+    } old_state;
+
+    // for non-recurrent models only
+    // list of slots to restore
+    std::vector<std::pair<uint32_t, uint32_t>> slot_boundaries;
+
+    bool do_restore = false;
+
+    explicit llama_kv_slot_restorer(const struct llama_kv_cache & cache) {
+        old_state.head = cache.head;
+        old_state.n    = cache.n;
+    }
+
+    // saves a slot information for future restoration
+    void save(const struct llama_kv_cache_slot_info & slot) {
+        if (slot) {
+            do_restore = true;
+            if (slot.boundaries.first != slot.boundaries.second) {
+                slot_boundaries.push_back(slot.boundaries);
+            }
+        }
+    }
+
+    // must be explicitly called to restore the kv_cache state
+    // and rollback changes from all llama_kv_cache_find_slot calls
+    void restore(struct llama_kv_cache & cache) {
+        if (do_restore) {
+            cache.head = old_state.head;
+            cache.n    = old_state.n;
+
+            if (cache.recurrent) { // recurrent models like Mamba or RWKV can't have a state partially erased
+                llama_kv_cache_seq_rm(cache, -1, -1, -1);
+            } else {
+                for (auto & slot : slot_boundaries) {
+                    llama_kv_cache_seq_rm(cache, -1, slot.first, slot.second);
+                }
+            }
+        }
+    }
+};
+

src/llama-mmap.cpp

@@ -0,0 +1,585 @@
+#include "llama-mmap.h"
+
+#include "llama-impl.h"
+
+#include "ggml.h"
+
+#include <cstring>
+#include <climits>
+#include <stdexcept>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.h>
+            #include <fcntl.h>
+        #endif
+        #if defined(_POSIX_MEMLOCK_RANGE)
+            #include <sys/resource.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #ifndef PATH_MAX
+        #define PATH_MAX MAX_PATH
+    #endif
+    #include <io.h>
+#endif
+
+// TODO: consider moving to llama-impl.h if needed in more places
+#if defined(_WIN32)
+std::string llama_format_win_err(DWORD err) {
+    LPSTR buf;
+    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
+    if (!size) {
+        return "FormatMessageA failed";
+    }
+    std::string ret(buf, size);
+    LocalFree(buf);
+    return ret;
+}
+#endif
+
+// llama_file
+
+struct llama_file::impl {
+#if defined(_WIN32)
+    HANDLE fp_win32;
+    std::string GetErrorMessageWin32(DWORD error_code) const {
+        std::string ret;
+        LPSTR lpMsgBuf = NULL;
+        DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                    NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
+        if (!bufLen) {
+            ret = format("Win32 error code: %lx", error_code);
+        } else {
+            ret = lpMsgBuf;
+            LocalFree(lpMsgBuf);
+        }
+
+        return ret;
+    }
+
+    impl(const char * fname, const char * mode) {
+        fp = ggml_fopen(fname, mode);
+        if (fp == NULL) {
+            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
+        }
+        fp_win32 = (HANDLE) _get_osfhandle(_fileno(fp));
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+        LARGE_INTEGER li;
+        li.QuadPart = 0;
+        BOOL ret = SetFilePointerEx(fp_win32, li, &li, FILE_CURRENT);
+        if (!ret) {
+            throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
+        }
+
+        return li.QuadPart;
+    }
+
+    void seek(size_t offset, int whence) const {
+        static_assert(SEEK_SET == FILE_BEGIN, "SEEK_SET != FILE_BEGIN");
+        static_assert(SEEK_CUR == FILE_CURRENT, "SEEK_CUR != FILE_CURRENT");
+        static_assert(SEEK_END == FILE_END, "SEEK_END != FILE_END");
+
+        LARGE_INTEGER li;
+        li.QuadPart = offset;
+        BOOL ret = SetFilePointerEx(fp_win32, li, NULL, whence);
+        if (!ret) {
+            throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
+        }
+    }
+
+    void read_raw(void * ptr, size_t len) const {
+        size_t bytes_read = 0;
+        while (bytes_read < len) {
+            size_t chunk_size = std::min<size_t>(len - bytes_read, 64*1024*1024);
+            DWORD chunk_read = 0;
+            BOOL result = ReadFile(fp_win32, reinterpret_cast<char*>(ptr) + bytes_read, chunk_size, &chunk_read, NULL);
+            if (!result) {
+                throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
+            }
+            if (chunk_read < chunk_size || chunk_read == 0) {
+                throw std::runtime_error("unexpectedly reached end of file");
+            }
+
+            bytes_read += chunk_read;
+        }
+    }
+
+    uint32_t read_u32() const {
+        uint32_t val;
+        read_raw(&val, sizeof(val));
+        return val;
+    }
+
+    void write_raw(const void * ptr, size_t len) const {
+        size_t bytes_written = 0;
+        while (bytes_written < len) {
+            size_t chunk_size = std::min<size_t>(len - bytes_written, 64*1024*1024);
+            DWORD chunk_written = 0;
+            BOOL result = WriteFile(fp_win32, reinterpret_cast<char const*>(ptr) + bytes_written, chunk_size, &chunk_written, NULL);
+            if (!result) {
+                throw std::runtime_error(format("write error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
+            }
+            if (chunk_written < chunk_size || chunk_written == 0) {
+                throw std::runtime_error("unexpectedly failed to write bytes");
+            }
+
+            bytes_written += chunk_written;
+        }
+    }
+
+    void write_u32(uint32_t val) const {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~impl() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+#else
+    impl(const char * fname, const char * mode) {
+        fp = ggml_fopen(fname, mode);
+        if (fp == NULL) {
+            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
+        }
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+// TODO: this ifdef is never true?
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        if (ret == -1) {
+            throw std::runtime_error(format("ftell error: %s", strerror(errno)));
+        }
+
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) const {
+// TODO: this ifdef is never true?
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        if (ret != 0) {
+            throw std::runtime_error(format("seek error: %s", strerror(errno)));
+        }
+    }
+
+    void read_raw(void * ptr, size_t len) const {
+        if (len == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, len, 1, fp);
+        if (ferror(fp)) {
+            throw std::runtime_error(format("read error: %s", strerror(errno)));
+        }
+        if (ret != 1) {
+            throw std::runtime_error("unexpectedly reached end of file");
+        }
+    }
+
+    uint32_t read_u32() const {
+        uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    void write_raw(const void * ptr, size_t len) const {
+        if (len == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, len, 1, fp);
+        if (ret != 1) {
+            throw std::runtime_error(format("write error: %s", strerror(errno)));
+        }
+    }
+
+    void write_u32(uint32_t val) const {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~impl() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+#endif
+
+    FILE * fp;
+    size_t size;
+};
+
+llama_file::llama_file(const char * fname, const char * mode) : pimpl(std::make_unique<impl>(fname, mode)) {}
+llama_file::~llama_file() = default;
+
+size_t llama_file::tell() const { return pimpl->tell(); }
+size_t llama_file::size() const { return pimpl->size; }
+
+int llama_file::fileno() const {
+#ifdef _WIN32
+    return _fileno(pimpl->fp);
+#else
+    return ::fileno(pimpl->fp);
+#endif
+}
+
+void llama_file::seek(size_t offset, int whence) const { pimpl->seek(offset, whence); }
+void llama_file::read_raw(void * ptr, size_t len) const { pimpl->read_raw(ptr, len); }
+
+uint32_t llama_file::read_u32() const { return pimpl->read_u32(); }
+
+void llama_file::write_raw(const void * ptr, size_t len) const { pimpl->write_raw(ptr, len); }
+void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); }
+
+// llama_mmap
+
+struct llama_mmap::impl {
+#ifdef _POSIX_MAPPED_FILES
+    std::vector<std::pair<size_t, size_t>> mapped_fragments;
+
+    impl(struct llama_file * file, size_t prefetch, bool numa) {
+        size = file->size();
+        int fd = file->fileno();
+        int flags = MAP_SHARED;
+        if (numa) { prefetch = 0; }
+#ifdef __linux__
+        if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) {
+            LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n",
+                    strerror(errno));
+        }
+        if (prefetch) { flags |= MAP_POPULATE; }
+#endif
+        addr = mmap(NULL, file->size(), PROT_READ, flags, fd, 0);
+        if (addr == MAP_FAILED) {
+            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
+        }
+
+        if (prefetch > 0) {
+            if (posix_madvise(addr, std::min(file->size(), prefetch), POSIX_MADV_WILLNEED)) {
+                LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+        if (numa) {
+            if (posix_madvise(addr, file->size(), POSIX_MADV_RANDOM)) {
+                LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+
+        mapped_fragments.emplace_back(0, file->size());
+    }
+
+    static void align_range(size_t * first, size_t * last, size_t page_size) {
+        size_t offset_in_page = *first & (page_size - 1);
+        size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page;
+        *first += offset_to_page;
+
+        *last = *last & ~(page_size - 1);
+
+        if (*last <= *first) {
+            *last = *first;
+        }
+    }
+
+    void unmap_fragment(size_t first, size_t last) {
+        int page_size = sysconf(_SC_PAGESIZE);
+        align_range(&first, &last, page_size);
+        size_t len = last - first;
+
+        if (len == 0) {
+            return;
+        }
+
+        GGML_ASSERT(first % page_size == 0);
+        GGML_ASSERT(last % page_size == 0);
+        GGML_ASSERT(last > first);
+
+        void * next_page_start = (uint8_t *) addr + first;
+
+        if (munmap(next_page_start, len)) {
+            LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
+        }
+
+        std::vector<std::pair<size_t, size_t>> new_mapped_fragments;
+        for (const auto & frag : mapped_fragments) {
+            if (frag.first < first && frag.second > last) {
+                new_mapped_fragments.emplace_back(frag.first, first);
+                new_mapped_fragments.emplace_back(last, frag.second);
+            } else if (frag.first < first && frag.second > first) {
+                new_mapped_fragments.emplace_back(frag.first, first);
+            } else if (frag.first < last && frag.second > last) {
+                new_mapped_fragments.emplace_back(last, frag.second);
+            } else if (frag.first >= first && frag.second <= last) {
+            } else {
+                new_mapped_fragments.push_back(frag);
+            }
+        }
+        mapped_fragments = std::move(new_mapped_fragments);
+    }
+
+    ~impl() {
+        for (const auto & frag : mapped_fragments) {
+            if (munmap((char *) addr + frag.first, frag.second - frag.first)) {
+                LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
+            }
+        }
+    }
+#elif defined(_WIN32)
+    impl(struct llama_file * file, size_t prefetch, bool numa) {
+        GGML_UNUSED(numa);
+
+        size = file->size();
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(file->fileno());
+
+        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+
+        if (hMapping == NULL) {
+            DWORD error = GetLastError();
+            throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+        DWORD error = GetLastError();
+        CloseHandle(hMapping);
+
+        if (addr == NULL) {
+            throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        if (prefetch > 0) {
+#if _WIN32_WINNT >= 0x602
+            BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
+            HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
+
+            pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
+
+            if (pPrefetchVirtualMemory) {
+                WIN32_MEMORY_RANGE_ENTRY range;
+                range.VirtualAddress = addr;
+                range.NumberOfBytes = (SIZE_T) std::min(size, prefetch);
+                if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                    LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n",
+                            llama_format_win_err(GetLastError()).c_str());
+                }
+            }
+#else
+            throw std::runtime_error("PrefetchVirtualMemory unavailable");
+#endif
+        }
+    }
+
+    void unmap_fragment(size_t first, size_t last) {
+        GGML_UNUSED(first);
+        GGML_UNUSED(last);
+    }
+
+    ~impl() {
+        if (!UnmapViewOfFile(addr)) {
+            LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    impl(struct llama_file * file, size_t prefetch, bool numa) {
+        GGML_UNUSED(file);
+        GGML_UNUSED(prefetch);
+        GGML_UNUSED(numa);
+
+        throw std::runtime_error("mmap not supported");
+    }
+
+    void unmap_fragment(size_t first, size_t last) {
+        GGML_UNUSED(first);
+        GGML_UNUSED(last);
+
+        throw std::runtime_error("mmap not supported");
+    }
+#endif
+
+    void * addr;
+    size_t size;
+};
+
+llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa) : pimpl(std::make_unique<impl>(file, prefetch, numa)) {}
+llama_mmap::~llama_mmap() = default;
+
+size_t llama_mmap::size() const { return pimpl->size; }
+void * llama_mmap::addr() const { return pimpl->addr; }
+
+void llama_mmap::unmap_fragment(size_t first, size_t last) { pimpl->unmap_fragment(first, last); }
+
+#if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
+const bool llama_mmap::SUPPORTED  = true;
+#else
+const bool llama_mmap::SUPPORTED  = false;
+#endif
+
+// llama_mlock
+
+struct llama_mlock::impl {
+#ifdef _POSIX_MEMLOCK_RANGE
+    static size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    bool raw_lock(const void * addr, size_t size) const {
+        if (!mlock(addr, size)) {
+            return true;
+        }
+
+#ifdef __APPLE__
+#define MLOCK_SUGGESTION \
+        "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
+        "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MEMLOCK (ulimit -l).\n"
+#else
+#define MLOCK_SUGGESTION \
+        "Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).\n"
+#endif
+
+        char* errmsg = std::strerror(errno);
+        bool suggest = (errno == ENOMEM);
+
+        struct rlimit lock_limit;
+        if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
+            suggest = false;
+        }
+        if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) {
+            suggest = false;
+        }
+
+        LLAMA_LOG_WARN("warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
+                size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
+        return false;
+    }
+
+    static void raw_unlock(void * addr, size_t size) {
+        if (munlock(addr, size)) {
+            LLAMA_LOG_WARN("warning: failed to munlock buffer: %s\n", std::strerror(errno));
+        }
+    }
+#elif defined(_WIN32)
+    static size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool raw_lock(void * ptr, size_t len) const {
+        for (int tries = 1; ; tries++) {
+            if (VirtualLock(ptr, len)) {
+                return true;
+            }
+            if (tries == 2) {
+                LLAMA_LOG_WARN("warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
+                    len, size, llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+
+            SIZE_T min_ws_size, max_ws_size;
+            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
+                LLAMA_LOG_WARN("warning: GetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+            size_t increment = len + 1048576;
+            min_ws_size += increment;
+            max_ws_size += increment;
+            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
+                LLAMA_LOG_WARN("warning: SetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+        }
+    }
+
+    static void raw_unlock(void * ptr, size_t len) {
+        if (!VirtualUnlock(ptr, len)) {
+            LLAMA_LOG_WARN("warning: failed to VirtualUnlock buffer: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static size_t lock_granularity() {
+        return (size_t) 65536;
+    }
+
+    bool raw_lock(const void * addr, size_t len) const {
+        LLAMA_LOG_WARN("warning: mlock not supported on this system\n");
+        return false;
+    }
+
+    static void raw_unlock(const void * addr, size_t len) {}
+#endif
+
+    impl() : addr(NULL), size(0), failed_already(false) {}
+
+    void init(void * ptr) {
+        GGML_ASSERT(addr == NULL && size == 0);
+        addr = ptr;
+    }
+
+    void grow_to(size_t target_size) {
+        GGML_ASSERT(addr);
+        if (failed_already) {
+            return;
+        }
+        size_t granularity = lock_granularity();
+        target_size = (target_size + granularity - 1) & ~(granularity - 1);
+        if (target_size > size) {
+            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
+                size = target_size;
+            } else {
+                failed_already = true;
+            }
+        }
+    }
+
+    void * addr;
+    size_t size;
+
+    bool failed_already;
+};
+
+llama_mlock::llama_mlock() : pimpl(std::make_unique<impl>()) {}
+llama_mlock::~llama_mlock() = default;
+
+void llama_mlock::init(void * ptr) { pimpl->init(ptr); }
+void llama_mlock::grow_to(size_t target_size) { pimpl->grow_to(target_size); }
+
+#if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
+const bool llama_mlock::SUPPORTED = true;
+#else
+const bool llama_mlock::SUPPORTED = false;
+#endif
+
+size_t llama_path_max() {
+    return PATH_MAX;
+}

src/llama-mmap.h

@@ -0,0 +1,67 @@
+#pragma once
+
+#include <memory>
+#include <vector>
+
+struct llama_file;
+struct llama_mmap;
+struct llama_mlock;
+
+using llama_files  = std::vector<std::unique_ptr<llama_file>>;
+using llama_mmaps  = std::vector<std::unique_ptr<llama_mmap>>;
+using llama_mlocks = std::vector<std::unique_ptr<llama_mlock>>;
+
+struct llama_file {
+    llama_file(const char * fname, const char * mode);
+    ~llama_file();
+
+    size_t tell() const;
+    size_t size() const;
+
+    int fileno() const;
+
+    void seek(size_t offset, int whence) const;
+
+    void read_raw(void * ptr, size_t len) const;
+    uint32_t read_u32() const;
+
+    void write_raw(const void * ptr, size_t len) const;
+    void write_u32(uint32_t val) const;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+struct llama_mmap {
+    llama_mmap(const llama_mmap &) = delete;
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false);
+    ~llama_mmap();
+
+    size_t size() const;
+    void * addr() const;
+
+    void unmap_fragment(size_t first, size_t last);
+
+    static const bool SUPPORTED;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+struct llama_mlock {
+    llama_mlock();
+    ~llama_mlock();
+
+    void init(void * ptr);
+    void grow_to(size_t target_size);
+
+    static const bool SUPPORTED;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+size_t llama_path_max();

src/llama-model-loader.h

@@ -0,0 +1,158 @@
+#pragma once
+
+#include "llama.h"
+
+#include "llama-impl.h"
+#include "llama-arch.h"
+#include "llama-mmap.h"
+
+#include "ggml-cpp.h"
+
+#include <cstddef>
+#include <map>
+#include <stdexcept>
+#include <unordered_map>
+
+using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
+
+enum llama_fver {
+    GGUF_FILE_VERSION_V1 = 1,
+    GGUF_FILE_VERSION_V2 = 2,
+    GGUF_FILE_VERSION_V3 = 3,
+};
+
+const char * llama_file_version_name(llama_fver version);
+
+struct llama_model_loader {
+    // Holds information on a model weight
+    struct llama_tensor_weight {
+        uint16_t  idx; // source file index
+        size_t   offs; // tensor data offset in the original file
+
+        ggml_tensor * tensor;
+
+        llama_tensor_weight(const llama_file * file, uint16_t idx, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) {
+            const int tensor_idx = gguf_find_tensor(gguf_ctx,  ggml_get_name(tensor));
+            if (tensor_idx < 0) {
+                throw std::runtime_error(format("tensor '%s' not found in the model", ggml_get_name(tensor)));
+            }
+
+            offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx);
+            if (offs + ggml_nbytes(tensor) < offs || offs + ggml_nbytes(tensor) > file->size()) {
+                throw std::runtime_error(format("tensor '%s' data is not within the file bounds, model is corrupted or incomplete", ggml_get_name(tensor)));
+            }
+        }
+    };
+
+    // custom comparator to sort weights more nicely by layer
+    struct weight_name_comparer {
+        bool operator()(const std::string & a, const std::string & b) const {
+            int a_layer = -1;
+            int b_layer = -1;
+            sscanf(a.c_str(), "blk.%d.", &a_layer);
+            sscanf(b.c_str(), "blk.%d.", &b_layer);
+            if (a_layer != b_layer) {
+                return a_layer < b_layer;
+            }
+            return a < b;
+        }
+    };
+
+    static const int TENSOR_NOT_REQUIRED = 1;
+    static const int TENSOR_DUPLICATED   = 2;
+
+    int n_kv      = 0;
+    int n_tensors = 0;
+    int n_created = 0;
+
+    uint64_t n_elements = 0;
+    size_t   n_bytes    = 0;
+
+    bool use_mmap = false;
+    bool check_tensors;
+
+    llama_files files;
+    llama_ftype ftype;
+    llama_fver  fver;
+
+    llama_mmaps mappings;
+
+    std::map<std::string, struct llama_tensor_weight, weight_name_comparer> weights_map;
+    std::unordered_map<std::string, struct llama_model_kv_override> kv_overrides;
+
+    gguf_context_ptr meta;
+    std::vector<ggml_context_ptr> contexts;
+
+    std::string arch_name;
+    LLM_KV      llm_kv    = LLM_KV(LLM_ARCH_UNKNOWN);
+
+    size_t size_done = 0;
+    size_t size_data = 0;
+    std::vector<std::pair<size_t, size_t>> mmaps_used;
+
+    llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p);
+
+    template<typename T>
+    typename std::enable_if<std::is_integral<T>::value, bool>::type
+    get_arr_n(const std::string & key, T & result, bool required = true);
+
+    template<typename T>
+    typename std::enable_if<std::is_integral<T>::value, bool>::type
+    get_arr_n(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T>
+    bool get_arr(const std::string & key, std::vector<T> & result, bool required = true);
+
+    template<typename T, size_t N_MAX>
+    bool get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required = true);
+
+    template<typename T>
+    bool get_arr(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T>
+    bool get_key(const std::string & key, T & result, bool required = true);
+
+    template<typename T>
+    bool get_key(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T, size_t N_MAX>
+    bool get_key_or_arr(const std::string & key, std::array<T, N_MAX> & result, uint32_t n, bool required = true);
+
+    template<typename T>
+    bool get_key_or_arr(enum llm_kv kid, T & result, uint32_t n, bool required = true);
+
+    std::string get_arch_name() const;
+
+    enum llm_arch get_arch() const;
+
+    const llama_tensor_weight * get_weight(const char * name) const;
+
+    const llama_tensor_weight & require_weight(const char * name) const;
+
+    struct ggml_tensor * get_tensor_meta(const char * name) const;
+
+    struct ggml_tensor * require_tensor_meta(const std::string & name) const;
+
+    const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const;
+
+    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::initializer_list<int64_t> & ne, int flags = 0);
+
+    struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::initializer_list<int64_t> & ne, size_t offset, bool required = true);
+
+    void done_getting_tensors() const;
+
+    void init_mappings(bool prefetch = true, llama_mlocks * mlock_mmaps = nullptr);
+
+    void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const;
+
+    // for backwards compatibility, does not support ggml-backend
+    void load_data_for(struct ggml_tensor * cur) const;
+
+    // Returns false if cancelled by progress_callback
+    bool load_all_data(
+            struct ggml_context * ctx,
+            llama_buf_map & bufs,
+            llama_mlocks * lmlocks,
+            llama_progress_callback progress_callback,
+            void * progress_callback_user_data);
+};

src/llama-model.h

@@ -0,0 +1,389 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-arch.h"
+#include "llama-hparams.h"
+#include "llama-vocab.h"
+#include "llama-mmap.h"
+
+#include "ggml-cpp.h"
+
+#include <vector>
+
+// available models
+// TODO: this enum does not follow the enum naming convention
+enum llm_type {
+    MODEL_UNKNOWN,
+    MODEL_14M,
+    MODEL_17M,
+    MODEL_22M,
+    MODEL_33M,
+    MODEL_60M,
+    MODEL_70M,
+    MODEL_80M,
+    MODEL_109M,
+    MODEL_137M,
+    MODEL_160M,
+    MODEL_220M,
+    MODEL_250M,
+    MODEL_270M,
+    MODEL_335M,
+    MODEL_410M,
+    MODEL_450M,
+    MODEL_770M,
+    MODEL_780M,
+    MODEL_0_5B,
+    MODEL_1B,
+    MODEL_1_3B,
+    MODEL_1_4B,
+    MODEL_1_5B,
+    MODEL_1_6B,
+    MODEL_2B,
+    MODEL_2_8B,
+    MODEL_3B,
+    MODEL_4B,
+    MODEL_6B,
+    MODEL_6_9B,
+    MODEL_7B,
+    MODEL_8B,
+    MODEL_9B,
+    MODEL_11B,
+    MODEL_12B,
+    MODEL_13B,
+    MODEL_14B,
+    MODEL_15B,
+    MODEL_16B,
+    MODEL_20B,
+    MODEL_30B,
+    MODEL_32B,
+    MODEL_34B,
+    MODEL_35B,
+    MODEL_40B,
+    MODEL_65B,
+    MODEL_70B,
+    MODEL_236B,
+    MODEL_314B,
+    MODEL_SMALL,
+    MODEL_MEDIUM,
+    MODEL_LARGE,
+    MODEL_XL,
+    MODEL_A1_7B,
+    MODEL_A2_7B,
+    MODEL_8x7B,
+    MODEL_8x22B,
+    MODEL_16x12B,
+    MODEL_10B_128x3_66B,
+    MODEL_57B_A14B,
+    MODEL_27B,
+};
+
+struct llama_layer_posnet {
+    // resnet
+    struct ggml_tensor * norm1   = nullptr;
+    struct ggml_tensor * norm1_b = nullptr;
+
+    struct ggml_tensor * conv1   = nullptr;
+    struct ggml_tensor * conv1_b = nullptr;
+
+    struct ggml_tensor * norm2   = nullptr;
+    struct ggml_tensor * norm2_b = nullptr;
+
+    struct ggml_tensor * conv2   = nullptr;
+    struct ggml_tensor * conv2_b = nullptr;
+
+    // attention
+    struct ggml_tensor * attn_norm   = nullptr;
+    struct ggml_tensor * attn_norm_b = nullptr;
+
+    struct ggml_tensor * attn_q   = nullptr;
+    struct ggml_tensor * attn_q_b = nullptr;
+
+    struct ggml_tensor * attn_k   = nullptr;
+    struct ggml_tensor * attn_k_b = nullptr;
+
+    struct ggml_tensor * attn_v   = nullptr;
+    struct ggml_tensor * attn_v_b = nullptr;
+
+    struct ggml_tensor * attn_o   = nullptr;
+    struct ggml_tensor * attn_o_b = nullptr;
+
+    // normalize
+    struct ggml_tensor * norm   = nullptr;
+    struct ggml_tensor * norm_b = nullptr;
+};
+
+struct llama_layer_convnext {
+    struct ggml_tensor * dw   = nullptr;
+    struct ggml_tensor * dw_b = nullptr;
+
+    struct ggml_tensor * norm   = nullptr;
+    struct ggml_tensor * norm_b = nullptr;
+
+    struct ggml_tensor * pw1   = nullptr;
+    struct ggml_tensor * pw1_b = nullptr;
+
+    struct ggml_tensor * pw2   = nullptr;
+    struct ggml_tensor * pw2_b = nullptr;
+
+    struct ggml_tensor * gamma = nullptr;
+};
+
+struct llama_layer {
+    // normalization
+    struct ggml_tensor * attn_norm       = nullptr;
+    struct ggml_tensor * attn_norm_b     = nullptr;
+    struct ggml_tensor * attn_norm_2     = nullptr;
+    struct ggml_tensor * attn_norm_2_b   = nullptr;
+    struct ggml_tensor * attn_q_norm     = nullptr;
+    struct ggml_tensor * attn_q_norm_b   = nullptr;
+    struct ggml_tensor * attn_k_norm     = nullptr;
+    struct ggml_tensor * attn_k_norm_b   = nullptr;
+    struct ggml_tensor * attn_out_norm   = nullptr;
+    struct ggml_tensor * attn_out_norm_b = nullptr;
+    struct ggml_tensor * attn_q_a_norm   = nullptr;
+    struct ggml_tensor * attn_kv_a_norm  = nullptr;
+    struct ggml_tensor * attn_sub_norm   = nullptr;
+    struct ggml_tensor * attn_post_norm  = nullptr;
+    struct ggml_tensor * ffn_sub_norm    = nullptr;
+    struct ggml_tensor * attn_norm_cross = nullptr;
+    struct ggml_tensor * attn_norm_enc   = nullptr;
+
+    // attention
+    struct ggml_tensor * wq        = nullptr;
+    struct ggml_tensor * wk        = nullptr;
+    struct ggml_tensor * wv        = nullptr;
+    struct ggml_tensor * wo        = nullptr;
+    struct ggml_tensor * wqkv      = nullptr;
+    struct ggml_tensor * wq_a      = nullptr;
+    struct ggml_tensor * wq_b      = nullptr;
+    struct ggml_tensor * wkv_a_mqa = nullptr;
+    struct ggml_tensor * wkv_b     = nullptr;
+    struct ggml_tensor * wq_cross  = nullptr;
+    struct ggml_tensor * wk_cross  = nullptr;
+    struct ggml_tensor * wv_cross  = nullptr;
+    struct ggml_tensor * wo_cross  = nullptr;
+    struct ggml_tensor * wq_enc    = nullptr;
+    struct ggml_tensor * wk_enc    = nullptr;
+    struct ggml_tensor * wv_enc    = nullptr;
+    struct ggml_tensor * wo_enc    = nullptr;
+
+    // attention bias
+    struct ggml_tensor * bq   = nullptr;
+    struct ggml_tensor * bk   = nullptr;
+    struct ggml_tensor * bv   = nullptr;
+    struct ggml_tensor * bo   = nullptr;
+    struct ggml_tensor * bqkv = nullptr;
+
+    // relative position bias
+    struct ggml_tensor * attn_rel_b       = nullptr;
+    struct ggml_tensor * attn_rel_b_enc   = nullptr;
+    struct ggml_tensor * attn_rel_b_cross = nullptr;
+
+    // normalization
+    struct ggml_tensor * ffn_norm         = nullptr;
+    struct ggml_tensor * ffn_norm_b       = nullptr;
+    struct ggml_tensor * ffn_post_norm    = nullptr;
+    struct ggml_tensor * layer_out_norm   = nullptr;
+    struct ggml_tensor * layer_out_norm_b = nullptr;
+    struct ggml_tensor * ffn_norm_exps    = nullptr;
+    struct ggml_tensor * ffn_norm_enc     = nullptr;
+
+    // ff
+    struct ggml_tensor * ffn_gate     = nullptr; // w1
+    struct ggml_tensor * ffn_down     = nullptr; // w2
+    struct ggml_tensor * ffn_up       = nullptr; // w3
+    struct ggml_tensor * ffn_gate_enc = nullptr;
+    struct ggml_tensor * ffn_down_enc = nullptr;
+    struct ggml_tensor * ffn_up_enc   = nullptr;
+
+    // ff MoE
+    struct ggml_tensor * ffn_gate_inp  = nullptr;
+    struct ggml_tensor * ffn_gate_exps = nullptr;
+    struct ggml_tensor * ffn_down_exps = nullptr;
+    struct ggml_tensor * ffn_up_exps   = nullptr;
+
+    // ff shared expert (shexp)
+    struct ggml_tensor * ffn_gate_inp_shexp = nullptr;
+    struct ggml_tensor * ffn_gate_shexp     = nullptr;
+    struct ggml_tensor * ffn_down_shexp     = nullptr;
+    struct ggml_tensor * ffn_up_shexp       = nullptr;
+
+    // ff bias
+    struct ggml_tensor * ffn_gate_b = nullptr;
+    struct ggml_tensor * ffn_down_b = nullptr; // b2
+    struct ggml_tensor * ffn_up_b   = nullptr; // b3
+    struct ggml_tensor * ffn_act    = nullptr;
+
+    // mamba proj
+    struct ggml_tensor * ssm_in  = nullptr;
+    struct ggml_tensor * ssm_x   = nullptr;
+    struct ggml_tensor * ssm_dt  = nullptr;
+    struct ggml_tensor * ssm_out = nullptr;
+
+    // mamba
+    struct ggml_tensor * ssm_conv1d = nullptr;
+    struct ggml_tensor * ssm_a      = nullptr;
+    struct ggml_tensor * ssm_d      = nullptr;
+
+    // mamba bias
+    struct ggml_tensor * ssm_conv1d_b = nullptr;
+    struct ggml_tensor * ssm_dt_b     = nullptr;
+
+    // rwkv
+    struct ggml_tensor * time_mix_w1         = nullptr;
+    struct ggml_tensor * time_mix_w2         = nullptr;
+    struct ggml_tensor * time_mix_lerp_x     = nullptr;
+    struct ggml_tensor * time_mix_lerp_w     = nullptr;
+    struct ggml_tensor * time_mix_lerp_k     = nullptr;
+    struct ggml_tensor * time_mix_lerp_v     = nullptr;
+    struct ggml_tensor * time_mix_lerp_r     = nullptr;
+    struct ggml_tensor * time_mix_lerp_g     = nullptr;
+
+    struct ggml_tensor * time_mix_first      = nullptr;
+    struct ggml_tensor * time_mix_decay      = nullptr;
+    struct ggml_tensor * time_mix_decay_w1   = nullptr;
+    struct ggml_tensor * time_mix_decay_w2   = nullptr;
+    struct ggml_tensor * time_mix_key        = nullptr;
+    struct ggml_tensor * time_mix_value      = nullptr;
+    struct ggml_tensor * time_mix_receptance = nullptr;
+    struct ggml_tensor * time_mix_gate       = nullptr;
+
+    struct ggml_tensor * time_mix_ln     = nullptr;
+    struct ggml_tensor * time_mix_ln_b   = nullptr;
+    struct ggml_tensor * time_mix_output = nullptr;
+
+    struct ggml_tensor * channel_mix_lerp_k = nullptr;
+    struct ggml_tensor * channel_mix_lerp_r = nullptr;
+
+    struct ggml_tensor * channel_mix_key        = nullptr;
+    struct ggml_tensor * channel_mix_receptance = nullptr;
+    struct ggml_tensor * channel_mix_value      = nullptr;
+
+    // long rope factors
+    struct ggml_tensor * rope_long  = nullptr;
+    struct ggml_tensor * rope_short = nullptr;
+    struct ggml_tensor * rope_freqs = nullptr;
+
+    // bitnet scale
+    struct ggml_tensor * wq_scale       = nullptr;
+    struct ggml_tensor * wk_scale       = nullptr;
+    struct ggml_tensor * wv_scale       = nullptr;
+    struct ggml_tensor * wo_scale       = nullptr;
+    struct ggml_tensor * ffn_gate_scale = nullptr;
+    struct ggml_tensor * ffn_up_scale   = nullptr;
+    struct ggml_tensor * ffn_down_scale = nullptr;
+
+    struct llama_layer_posnet posnet;
+
+    struct llama_layer_convnext convnext;
+};
+
+struct llama_model {
+    llm_type type = MODEL_UNKNOWN;
+    llm_arch arch = LLM_ARCH_UNKNOWN;
+
+    llama_ftype ftype = LLAMA_FTYPE_ALL_F32;
+
+    std::string name = "n/a";
+
+    llama_hparams hparams = {};
+    llama_vocab   vocab;
+
+    struct ggml_tensor * tok_embd   = nullptr;
+    struct ggml_tensor * type_embd  = nullptr;
+    struct ggml_tensor * pos_embd   = nullptr;
+    struct ggml_tensor * tok_norm   = nullptr;
+    struct ggml_tensor * tok_norm_b = nullptr;
+
+    struct ggml_tensor * output_norm     = nullptr;
+    struct ggml_tensor * output_norm_b   = nullptr;
+    struct ggml_tensor * output          = nullptr;
+    struct ggml_tensor * output_b        = nullptr;
+    struct ggml_tensor * output_norm_enc = nullptr;
+
+    // classifier
+    struct ggml_tensor * cls       = nullptr;
+    struct ggml_tensor * cls_b     = nullptr;
+    struct ggml_tensor * cls_out   = nullptr;
+    struct ggml_tensor * cls_out_b = nullptr;
+
+    struct ggml_tensor * conv1d   = nullptr;
+    struct ggml_tensor * conv1d_b = nullptr;
+
+    std::vector<llama_layer> layers;
+
+    // gguf metadata
+    std::unordered_map<std::string, std::string> gguf_kv;
+
+    llama_split_mode split_mode;
+    int main_gpu;
+    int n_gpu_layers;
+
+    std::vector<std::string> rpc_servers;
+
+    // list of devices used in this model
+    std::vector<ggml_backend_dev_t> devices;
+
+
+    // lists of buffer types used for each layer
+    using buft_list_t = std::vector<std::pair<ggml_backend_dev_t, ggml_backend_buffer_type_t>>;
+    buft_list_t cpu_buft_list;
+    std::map<ggml_backend_dev_t, buft_list_t> gpu_buft_list;
+
+    struct layer_dev {
+        ggml_backend_dev_t dev;
+        buft_list_t * buft_list;
+    };
+
+    layer_dev dev_input = {};
+    layer_dev dev_output = {};
+    std::vector<layer_dev> dev_layer;
+
+    // contexts where the model tensors metadata is stored
+    std::vector<ggml_context_ptr> ctxs;
+
+    // the model memory buffers for the tensor data
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    // model memory mapped files
+    llama_mmaps mappings;
+
+    // objects representing data potentially being locked in memory
+    llama_mlocks mlock_bufs;
+    llama_mlocks mlock_mmaps;
+
+    // for quantize-stats only
+    std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
+
+    int64_t t_load_us  = 0;
+    int64_t t_start_us = 0;
+
+    // total number of parameters in the model
+    uint64_t n_elements = 0;
+
+    // total size of all the tensors in the model in bytes
+    size_t  n_bytes     = 0;
+};
+
+const char * llm_type_name(llm_type type);
+
+std::string llama_model_arch_name (const llama_model & model);
+std::string llama_model_type_name (const llama_model & model);
+std::string llama_model_ftype_name(const llama_model & model);
+
+// used by llama_adapter_cvec
+ggml_backend_buffer_type_t llama_model_select_buft(const llama_model & model, int il);
+
+// used by llama_adapter_lora
+struct ggml_tensor * llama_model_get_tensor(const struct llama_model & model, const char * name);
+
+size_t llama_model_max_nodes(const llama_model & model);
+
+struct llama_model_loader;
+
+// TODO: become llama_model methods
+void llm_load_stats     (llama_model_loader & ml, llama_model & model);
+void llm_load_arch      (llama_model_loader & ml, llama_model & model);
+void llm_load_hparams   (llama_model_loader & ml, llama_model & model);
+void llm_load_vocab     (llama_model_loader & ml, llama_model & model);
+void llm_load_print_meta(llama_model_loader & ml, llama_model & model);

src/llama-quant.cpp

@@ -0,0 +1,929 @@
+#include "llama-quant.h"
+
+#include "llama-impl.h"
+#include "llama-model.h"
+#include "llama-model-loader.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <fstream>
+#include <mutex>
+#include <thread>
+#include <unordered_map>
+
+// TODO: replace with ggml API call
+#define QK_K 256
+
+static void zeros(std::ofstream & file, size_t n) {
+    char zero = 0;
+    for (size_t i = 0; i < n; ++i) {
+        file.write(&zero, 1);
+    }
+}
+
+struct quantize_state_internal {
+    const llama_model                 & model;
+    const llama_model_quantize_params * params;
+
+    int n_attention_wv = 0;
+    int n_ffn_down     = 0;
+    int n_ffn_gate     = 0;
+    int n_ffn_up       = 0;
+    int i_attention_wv = 0;
+    int i_ffn_down     = 0;
+    int i_ffn_gate     = 0;
+    int i_ffn_up       = 0;
+
+    int n_k_quantized = 0;
+    int n_fallback    = 0;
+
+    bool has_imatrix = false;
+
+    // used to figure out if a model shares tok_embd with the output weight
+    bool has_output = false;
+
+    quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params)
+        : model(model)
+        , params(params)
+        {}
+};
+
+static void llama_tensor_dequantize_internal(
+    struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
+    const size_t nelements, const int nthread
+) {
+    if (output.size() < nelements) {
+        output.resize(nelements);
+    }
+    float * f32_output = (float *) output.data();
+
+    const ggml_type_traits * qtype = ggml_get_type_traits(tensor->type);
+    if (ggml_is_quantized(tensor->type)) {
+        if (qtype->to_float == NULL) {
+            throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
+        }
+    } else if (tensor->type != GGML_TYPE_F16 &&
+               tensor->type != GGML_TYPE_BF16) {
+        throw std::runtime_error(format("cannot dequantize/convert tensor type %s", ggml_type_name(tensor->type)));
+    }
+
+    if (nthread < 2) {
+        if (tensor->type == GGML_TYPE_F16) {
+            ggml_fp16_to_fp32_row((ggml_fp16_t *)tensor->data, f32_output, nelements);
+        } else if (tensor->type == GGML_TYPE_BF16) {
+            ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements);
+        } else if (ggml_is_quantized(tensor->type)) {
+            qtype->to_float(tensor->data, f32_output, nelements);
+        } else {
+            GGML_ABORT("fatal error"); // unreachable
+        }
+        return;
+    }
+
+    size_t block_size;
+    if (tensor->type == GGML_TYPE_F16 ||
+        tensor->type == GGML_TYPE_BF16) {
+        block_size = 1;
+    } else {
+        block_size = (size_t)ggml_blck_size(tensor->type);
+    }
+
+    size_t block_size_bytes = ggml_type_size(tensor->type);
+
+    GGML_ASSERT(nelements % block_size == 0);
+    size_t nblocks = nelements / block_size;
+    size_t blocks_per_thread = nblocks / nthread;
+    size_t spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count
+
+    size_t in_buff_offs = 0;
+    size_t out_buff_offs = 0;
+
+    for (int tnum = 0; tnum < nthread; tnum++) {
+        size_t thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread
+        size_t thr_elems = thr_blocks * block_size; // number of elements for this thread
+        size_t thr_block_bytes = thr_blocks * block_size_bytes; // number of input bytes for this thread
+
+        auto compute = [qtype] (ggml_type typ, uint8_t * inbuf, float * outbuf, int nels) {
+            if (typ == GGML_TYPE_F16) {
+                ggml_fp16_to_fp32_row((ggml_fp16_t *)inbuf, outbuf, nels);
+            } else if (typ == GGML_TYPE_BF16) {
+                ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels);
+            } else {
+                qtype->to_float(inbuf, outbuf, nels);
+            }
+        };
+        workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
+        in_buff_offs += thr_block_bytes;
+        out_buff_offs += thr_elems;
+    }
+    for (auto & w : workers) { w.join(); }
+    workers.clear();
+}
+
+static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
+    const std::string name = ggml_get_name(tensor);
+
+    // TODO: avoid hardcoded tensor names - use the TN_* constants
+    const llm_arch arch = qs.model.arch;
+    const auto       tn = LLM_TN(arch);
+
+    auto use_more_bits = [](int i_layer, int n_layers) -> bool {
+        return i_layer < n_layers/8 || i_layer >= 7*n_layers/8 || (i_layer - n_layers/8)%3 == 2;
+    };
+    const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
+    auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) {
+        if (n_expert > 1) {
+            // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but occasionally randomly
+            // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
+            // for getting the current layer as I initially thought, and we need to resort to parsing the
+            // tensor name.
+            if (sscanf(name, "blk.%d.", &i_layer) != 1) {
+                throw std::runtime_error(format("Failed to determine layer for tensor %s", name));
+            }
+            if (i_layer < 0 || i_layer >= n_layer) {
+                throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name, n_layer));
+            }
+        }
+        return std::make_pair(i_layer, n_layer);
+    };
+
+    // for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings
+    // with the quantization of the output tensor
+    if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) {
+        if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
+            new_type = qs.params->output_tensor_type;
+        } else {
+            int nx = tensor->ne[0];
+            if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
+                new_type = GGML_TYPE_Q8_0;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
+                     ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M   ||
+                     ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
+                new_type = GGML_TYPE_Q5_K;
+            }
+            else if (new_type != GGML_TYPE_Q8_0) {
+                new_type = GGML_TYPE_Q6_K;
+            }
+        }
+    } else if (name == "token_embd.weight") {
+        if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
+            new_type = qs.params->token_embedding_type;
+        } else {
+            if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
+                ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
+                new_type = GGML_TYPE_Q2_K;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
+                new_type = GGML_TYPE_IQ3_S;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+                new_type = GGML_TYPE_IQ3_S;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) {
+                new_type = GGML_TYPE_Q4_K;
+            }
+        }
+    } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
+               ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M    || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
+        if (name.find("attn_v.weight") != std::string::npos) {
+            if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K;
+            else new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            ++qs.i_attention_wv;
+        }
+        else if (qs.model.hparams.n_expert == 8 && name.find("attn_k.weight") != std::string::npos) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (name.find("ffn_down") != std::string::npos) {
+            if (qs.i_ffn_down < qs.n_ffn_down/8) {
+                new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            }
+            ++qs.i_ffn_down;
+        }
+        else if (name.find("attn_output.weight") != std::string::npos) {
+            if (qs.model.hparams.n_expert == 8) {
+                new_type = GGML_TYPE_Q5_K;
+            } else {
+                if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) new_type = GGML_TYPE_IQ2_XXS;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ3_S;
+            }
+        }
+    } else if (name.find("attn_v.weight") != std::string::npos) {
+        if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
+            new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS;
+        }
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+            new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q5_K;
+        }
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
+                use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
+        if (qs.model.type == MODEL_70B) {
+            // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
+            // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
+            // nearly negligible increase in model size by quantizing this tensor with more bits:
+            if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
+        }
+        if (qs.model.hparams.n_expert == 8) {
+            // for the 8-expert model, bumping this to Q8_0 trades just ~128MB
+            // TODO: explore better strategies
+            new_type = GGML_TYPE_Q8_0;
+        }
+        ++qs.i_attention_wv;
+    } else if (name.find("attn_k.weight") != std::string::npos) {
+        if (qs.model.hparams.n_expert == 8) {
+            // for the 8-expert model, bumping this to Q8_0 trades just ~128MB
+            // TODO: explore better strategies
+            new_type = GGML_TYPE_Q8_0;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
+            new_type = GGML_TYPE_IQ3_XXS;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            new_type = GGML_TYPE_IQ2_S;
+        }
+    } else if (name.find("attn_q.weight") != std::string::npos) {
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
+            new_type = GGML_TYPE_IQ3_XXS;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            new_type = GGML_TYPE_IQ2_S;
+        }
+    } else if (name.find("ffn_down") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
+        if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
+            if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && !qs.has_imatrix) {
+            new_type = i_layer < n_layer/8 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+            new_type = i_layer < n_layer/16 ? GGML_TYPE_Q5_K
+                     : arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
+                     : GGML_TYPE_Q3_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (i_layer < n_layer/8 ||
+                    (qs.model.hparams.n_expert == 8 && use_more_bits(i_layer, n_layer)))) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
+            new_type = arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
+            if (arch == LLM_ARCH_FALCON) {
+                new_type = i_layer < n_layer/16 ? GGML_TYPE_Q6_K :
+                           use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+            } else {
+                if (use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
+            }
+        }
+        else if (i_layer < n_layer/8 && (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && !qs.has_imatrix) {
+            new_type = GGML_TYPE_Q5_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
+            new_type = GGML_TYPE_Q5_K;
+        }
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0)
+                && qs.has_imatrix && i_layer < n_layer/8) {
+            // Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0.
+            // We only do it when an imatrix is provided because a) we want to make sure that one can always get the
+            // same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
+            new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
+        }
+        ++qs.i_ffn_down;
+    } else if (name.find("attn_output.weight") != std::string::npos) {
+        if (arch != LLM_ARCH_FALCON) {
+            if (qs.model.hparams.n_expert == 8) {
+                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
+                    ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL  ||
+                    ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S  ||
+                    ftype == LLAMA_FTYPE_MOSTLY_IQ3_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) {
+                    new_type = GGML_TYPE_Q5_K;
+                }
+            } else {
+                if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   ) new_type = GGML_TYPE_Q3_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ) new_type = GGML_TYPE_Q4_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ) new_type = GGML_TYPE_Q5_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M  ) new_type = GGML_TYPE_Q4_K;
+            }
+        } else {
+            if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
+        }
+    }
+    else if (name.find("attn_qkv.weight") != std::string::npos) {
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
+    }
+    else if (name.find("ffn_gate") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
+            new_type = GGML_TYPE_IQ3_XXS;
+        }
+        ++qs.i_ffn_gate;
+    }
+    else if (name.find("ffn_up") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
+            new_type = GGML_TYPE_IQ3_XXS;
+        }
+        ++qs.i_ffn_up;
+    }
+
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    //}
+    // IK: let's remove this, else Q2_K is almost the same as Q3_K_S
+    //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    //}
+    // This can be used to reduce the size of the Q5_K_S model.
+    // The associated PPL increase is fully in line with the size reduction
+    //else {
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
+    //}
+    bool convert_incompatible_tensor = false;
+    if (new_type == GGML_TYPE_Q2_K    || new_type == GGML_TYPE_Q3_K    || new_type == GGML_TYPE_Q4_K   ||
+        new_type == GGML_TYPE_Q5_K    || new_type == GGML_TYPE_Q6_K    || new_type == GGML_TYPE_IQ4_XS ||
+        new_type == GGML_TYPE_IQ2_XS  || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S  ||
+        new_type == GGML_TYPE_IQ3_XXS || new_type == GGML_TYPE_IQ1_S   || new_type == GGML_TYPE_IQ3_S  ||
+        new_type == GGML_TYPE_IQ1_M) {
+        int nx = tensor->ne[0];
+        int ny = tensor->ne[1];
+        if (nx % QK_K != 0) {
+            LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type));
+            convert_incompatible_tensor = true;
+        } else {
+            ++qs.n_k_quantized;
+        }
+    }
+    if (convert_incompatible_tensor) {
+        switch (new_type) {
+            case GGML_TYPE_TQ1_0:
+            case GGML_TYPE_TQ2_0:  new_type = GGML_TYPE_Q4_0; break;  // TODO: use a symmetric type instead
+            case GGML_TYPE_IQ2_XXS:
+            case GGML_TYPE_IQ2_XS:
+            case GGML_TYPE_IQ2_S:
+            case GGML_TYPE_IQ3_XXS:
+            case GGML_TYPE_IQ3_S:
+            case GGML_TYPE_IQ1_S:
+            case GGML_TYPE_IQ1_M:
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break;
+            case GGML_TYPE_Q4_K:   new_type = GGML_TYPE_Q5_0;   break;
+            case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q5_1;   break;
+            case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q8_0;   break;
+            default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
+        }
+        if (tensor->ne[0] % ggml_blck_size(new_type) != 0) {
+            new_type = GGML_TYPE_F16;
+        }
+        LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
+        ++qs.n_fallback;
+    }
+
+    return new_type;
+}
+
+static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int64_t chunk_size, int64_t nrows, int64_t n_per_row, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
+    if (nthread < 2) {
+        // single-thread
+        size_t new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, imatrix);
+        if (!ggml_validate_row_data(new_type, new_data, new_size)) {
+            throw std::runtime_error("quantized data validation failed");
+        }
+        return new_size;
+    }
+
+    std::mutex mutex;
+    int64_t counter = 0;
+    size_t new_size = 0;
+    bool valid = true;
+    auto compute = [&mutex, &counter, &new_size, &valid, new_type, f32_data, new_data, chunk_size,
+            nrows, n_per_row, imatrix]() {
+        const int64_t nrows_per_chunk = chunk_size / n_per_row;
+        size_t local_size = 0;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int64_t first_row = counter; counter += nrows_per_chunk;
+            if (first_row >= nrows) {
+                if (local_size > 0) {
+                    new_size += local_size;
+                }
+                break;
+            }
+            lock.unlock();
+            const int64_t this_nrow = std::min(nrows - first_row, nrows_per_chunk);
+            size_t this_size = ggml_quantize_chunk(new_type, f32_data, new_data, first_row * n_per_row, this_nrow, n_per_row, imatrix);
+            local_size += this_size;
+
+            // validate the quantized data
+            const size_t row_size  = ggml_row_size(new_type, n_per_row);
+            void * this_data = (char *) new_data + first_row * row_size;
+            if (!ggml_validate_row_data(new_type, this_data, this_size)) {
+                std::unique_lock<std::mutex> lock(mutex);
+                valid = false;
+                break;
+            }
+        }
+    };
+    for (int it = 0; it < nthread - 1; ++it) {
+        workers.emplace_back(compute);
+    }
+    compute();
+    for (auto & w : workers) { w.join(); }
+    workers.clear();
+    if (!valid) {
+        throw std::runtime_error("quantized data validation failed");
+    }
+    return new_size;
+}
+
+static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
+    ggml_type default_type;
+    llama_ftype ftype = params->ftype;
+
+    switch (params->ftype) {
+        case LLAMA_FTYPE_MOSTLY_Q4_0: default_type = GGML_TYPE_Q4_0; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_1: default_type = GGML_TYPE_Q4_1; break;
+        case LLAMA_FTYPE_MOSTLY_Q5_0: default_type = GGML_TYPE_Q5_0; break;
+        case LLAMA_FTYPE_MOSTLY_Q5_1: default_type = GGML_TYPE_Q5_1; break;
+        case LLAMA_FTYPE_MOSTLY_Q8_0: default_type = GGML_TYPE_Q8_0; break;
+        case LLAMA_FTYPE_MOSTLY_F16:  default_type = GGML_TYPE_F16;  break;
+        case LLAMA_FTYPE_MOSTLY_BF16: default_type = GGML_TYPE_BF16; break;
+        case LLAMA_FTYPE_ALL_F32:     default_type = GGML_TYPE_F32;  break;
+
+        // K-quants
+        case LLAMA_FTYPE_MOSTLY_Q2_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q2_K:    default_type = GGML_TYPE_Q2_K;    break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_XS:  default_type = GGML_TYPE_IQ3_S;   break;
+        case LLAMA_FTYPE_MOSTLY_Q3_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q3_K_M:
+        case LLAMA_FTYPE_MOSTLY_Q3_K_L:  default_type = GGML_TYPE_Q3_K;    break;
+        case LLAMA_FTYPE_MOSTLY_Q4_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q4_K_M:  default_type = GGML_TYPE_Q4_K;    break;
+        case LLAMA_FTYPE_MOSTLY_Q5_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q5_K_M:  default_type = GGML_TYPE_Q5_K;    break;
+        case LLAMA_FTYPE_MOSTLY_Q6_K:    default_type = GGML_TYPE_Q6_K;    break;
+        case LLAMA_FTYPE_MOSTLY_TQ1_0:   default_type = GGML_TYPE_TQ1_0;   break;
+        case LLAMA_FTYPE_MOSTLY_TQ2_0:   default_type = GGML_TYPE_TQ2_0;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS: default_type = GGML_TYPE_IQ2_XXS; break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XS:  default_type = GGML_TYPE_IQ2_XS;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_S:   default_type = GGML_TYPE_IQ2_XS;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_M:   default_type = GGML_TYPE_IQ2_S;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_XXS: default_type = GGML_TYPE_IQ3_XXS; break;
+        case LLAMA_FTYPE_MOSTLY_IQ1_S:   default_type = GGML_TYPE_IQ1_S;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ1_M:   default_type = GGML_TYPE_IQ1_M;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ4_NL:  default_type = GGML_TYPE_IQ4_NL;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ4_XS:  default_type = GGML_TYPE_IQ4_XS;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_S:   default_type = GGML_TYPE_IQ3_S;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_M:   default_type = GGML_TYPE_IQ3_S;   break;
+
+        default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
+    }
+
+    int nthread = params->nthread;
+
+    if (nthread <= 0) {
+        nthread = std::thread::hardware_concurrency();
+    }
+
+    // mmap consistently increases speed Linux, and also increases speed on Windows with
+    // hot cache. It may cause a slowdown on macOS, possibly related to free memory.
+#if defined(__linux__) || defined(_WIN32)
+    constexpr bool use_mmap = true;
+#else
+    constexpr bool use_mmap = false;
+#endif
+
+    llama_model_kv_override * kv_overrides = nullptr;
+    if (params->kv_overrides) {
+        auto v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
+        kv_overrides = v->data();
+    }
+    llama_model_loader ml(fname_inp, use_mmap, /*check_tensors*/ true, kv_overrides);
+    ml.init_mappings(false); // no prefetching
+
+    llama_model model;
+    llm_load_arch   (ml, model);
+    llm_load_hparams(ml, model);
+    llm_load_stats  (ml, model);
+
+    struct quantize_state_internal qs(model, params);
+
+    if (params->only_copy) {
+        ftype = model.ftype;
+    }
+    const std::unordered_map<std::string, std::vector<float>> * imatrix_data = nullptr;
+    if (params->imatrix) {
+        imatrix_data = static_cast<const std::unordered_map<std::string, std::vector<float>>*>(params->imatrix);
+        if (imatrix_data) {
+            LLAMA_LOG_INFO("================================ Have weights data with %d entries\n",int(imatrix_data->size()));
+            qs.has_imatrix = true;
+            // check imatrix for nans or infs
+            for (const auto & kv : *imatrix_data) {
+                for (float f : kv.second) {
+                    if (!std::isfinite(f)) {
+                        throw std::runtime_error(format("imatrix contains non-finite value %f\n", f));
+                    }
+                }
+            }
+        }
+    }
+
+    const size_t align = GGUF_DEFAULT_ALIGNMENT;
+    gguf_context_ptr ctx_out { gguf_init_empty() };
+
+    // copy the KV pairs from the input file
+    gguf_set_kv     (ctx_out.get(), ml.meta.get());
+    gguf_set_val_u32(ctx_out.get(), "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV
+    gguf_set_val_u32(ctx_out.get(), "general.file_type", ftype); // TODO: use LLM_KV
+
+    // Remove split metadata
+    gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str());
+    gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str());
+    gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str());
+
+    if (params->kv_overrides) {
+        const std::vector<llama_model_kv_override> & overrides = *(const std::vector<llama_model_kv_override> *)params->kv_overrides;
+        for (const auto & o : overrides) {
+            if (o.key[0] == 0) break;
+            if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) {
+                gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
+            } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
+                gguf_set_val_i32(ctx_out.get(), o.key, o.val_i64);
+            } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
+                gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
+            } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
+                gguf_set_val_str(ctx_out.get(), o.key, o.val_str);
+            } else {
+                LLAMA_LOG_WARN("%s: unknown KV override type for key %s\n", __func__, o.key);
+            }
+        }
+    }
+
+    // make a list of weights
+    std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
+    tensors.reserve(ml.weights_map.size());
+    for (const auto & it : ml.weights_map) {
+        tensors.push_back(&it.second);
+    }
+
+    // keep_split requires that the weights are sorted by split index
+    if (params->keep_split) {
+        std::sort(tensors.begin(), tensors.end(), [](const llama_model_loader::llama_tensor_weight * a, const llama_model_loader::llama_tensor_weight * b) {
+            if (a->idx == b->idx) {
+                return a->offs < b->offs;
+            }
+            return a->idx < b->idx;
+        });
+    }
+
+    for (const auto * it : tensors) {
+        const struct ggml_tensor * tensor = it->tensor;
+
+        const std::string name = ggml_get_name(tensor);
+
+        // TODO: avoid hardcoded tensor names - use the TN_* constants
+        if (name.find("attn_v.weight")   != std::string::npos ||
+            name.find("attn_qkv.weight") != std::string::npos ||
+            name.find("attn_kv_b.weight")!= std::string::npos) {
+            ++qs.n_attention_wv;
+        } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
+            qs.has_output = true;
+        }
+    }
+
+    qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
+
+    // sanity checks
+    {
+        const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
+        // attention layers have a non-zero number of kv heads
+        int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
+        if (llama_model_has_encoder(&model)) {
+            n_attn_layer *= 3;
+        }
+        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+    }
+
+    size_t total_size_org = 0;
+    size_t total_size_new = 0;
+
+    std::vector<std::thread> workers;
+    workers.reserve(nthread);
+
+    int idx = 0;
+
+    std::vector<no_init<uint8_t>> read_data;
+    std::vector<no_init<uint8_t>> work;
+    std::vector<no_init<float>> f32_conv_buf;
+
+    uint16_t n_split = 1;
+
+    // Assume split index is continuous
+    if (params->keep_split) {
+        for (const auto * it : tensors) {
+            n_split = std::max(uint16_t(it->idx + 1), n_split);
+        }
+    }
+    std::vector<gguf_context_ptr> ctx_outs(n_split);
+    ctx_outs[0] = std::move(ctx_out);
+
+    // populate the original tensors so we get an initial meta data
+    for (const auto * it : tensors) {
+        uint16_t i_split = params->keep_split ? it->idx : 0;
+        struct ggml_tensor * tensor = it->tensor;
+        if (!ctx_outs[i_split]) {
+            ctx_outs[i_split].reset(gguf_init_empty());
+        }
+        gguf_add_tensor(ctx_outs[i_split].get(), tensor);
+    }
+
+    // Set split info if needed
+    if (n_split > 1) {
+        for (size_t i = 0; i < ctx_outs.size(); ++i) {
+            gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i);
+            gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split);
+            gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors);
+        }
+    }
+
+    int cur_split = -1;
+    std::ofstream fout;
+    auto close_ofstream = [&]() {
+        // Write metadata and close file handler
+        if (fout.is_open()) {
+            fout.seekp(0);
+            std::vector<uint8_t> data(gguf_get_meta_size(ctx_outs[cur_split].get()));
+            gguf_get_meta_data(ctx_outs[cur_split].get(), data.data());
+            fout.write((const char *) data.data(), data.size());
+            fout.close();
+        }
+    };
+    auto new_ofstream = [&](int index) {
+        cur_split = index;
+        GGML_ASSERT(ctx_outs[cur_split] && "Find uninitialized gguf_context");
+        std::string fname = fname_out;
+        if (params->keep_split) {
+            std::vector<char> split_path(llama_path_max(), 0);
+            llama_split_path(split_path.data(), split_path.size(), fname_out.c_str(), cur_split, n_split);
+            fname = std::string(split_path.data());
+        }
+
+        fout = std::ofstream(fname, std::ios::binary);
+        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+        const size_t meta_size = gguf_get_meta_size(ctx_outs[cur_split].get());
+        // placeholder for the meta data
+        ::zeros(fout, meta_size);
+    };
+
+    const auto tn = LLM_TN(model.arch);
+    new_ofstream(0);
+    for (const auto * it : tensors) {
+        const auto & weight = *it;
+        struct ggml_tensor * tensor = weight.tensor;
+        if (weight.idx != cur_split && params->keep_split) {
+            close_ofstream();
+            new_ofstream(weight.idx);
+        }
+
+        const std::string name = ggml_get_name(tensor);
+
+        if (!ml.use_mmap) {
+            if (read_data.size() < ggml_nbytes(tensor)) {
+                read_data.resize(ggml_nbytes(tensor));
+            }
+            tensor->data = read_data.data();
+        }
+        ml.load_data_for(tensor);
+
+        LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ",
+               ++idx, ml.n_tensors,
+               ggml_get_name(tensor),
+               llama_format_tensor_shape(tensor).c_str(),
+               ggml_type_name(tensor->type));
+
+        // This used to be a regex, but <regex> has an extreme cost to compile times.
+        bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'?
+
+        // quantize only 2D and 3D tensors (experts)
+        quantize &= (ggml_n_dims(tensor) >= 2);
+
+        // do not quantize norm tensors
+        quantize &= name.find("_norm.weight") == std::string::npos;
+
+        quantize &= params->quantize_output_tensor || name != "output.weight";
+        quantize &= !params->only_copy;
+
+        // do not quantize expert gating tensors
+        // NOTE: can't use LLM_TN here because the layer number is not known
+        quantize &= name.find("ffn_gate_inp.weight") == std::string::npos;
+
+        // do not quantize positional embeddings and token types (BERT)
+        quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_POS_EMBD,    "weight");
+        quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_TOKEN_TYPES, "weight");
+
+        // do not quantize Mamba's small yet 2D weights
+        // NOTE: can't use LLM_TN here because the layer number is not known
+        quantize &= name.find("ssm_conv1d.weight") == std::string::npos;
+
+        // do not quantize RWKV's time_mix_first tensors
+        quantize &= name.find("time_mix_first.weight") == std::string::npos;
+        quantize &= name.find("time_mix_w1.weight") == std::string::npos;
+        quantize &= name.find("time_mix_w2.weight") == std::string::npos;
+        quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos;
+        quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos;
+
+        // do not quantize relative position bias (T5)
+        quantize &= name.find("attn_rel_b.weight") == std::string::npos;
+
+        enum ggml_type new_type;
+        void * new_data;
+        size_t new_size;
+
+        if (quantize) {
+            new_type = default_type;
+
+            // get more optimal quantization type based on the tensor shape, layer, etc.
+            if (!params->pure && ggml_is_quantized(default_type)) {
+                new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
+            }
+            if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
+                new_type = params->token_embedding_type;
+            }
+            if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) {
+                new_type = params->output_tensor_type;
+            }
+
+            // If we've decided to quantize to the same type the tensor is already
+            // in then there's nothing to do.
+            quantize = tensor->type != new_type;
+        }
+
+        if (!quantize) {
+            new_type = tensor->type;
+            new_data = tensor->data;
+            new_size = ggml_nbytes(tensor);
+            LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
+        } else {
+            const int64_t nelements = ggml_nelements(tensor);
+
+            const float * imatrix = nullptr;
+            if (imatrix_data) {
+                auto it = imatrix_data->find(tensor->name);
+                if (it == imatrix_data->end()) {
+                    LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
+                } else {
+                    if (it->second.size() == (size_t)tensor->ne[0]*tensor->ne[2]) {
+                        imatrix = it->second.data();
+                    } else {
+                        LLAMA_LOG_INFO("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
+                                int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name);
+
+                        // this can happen when quantizing an old mixtral model with split tensors with a new incompatible imatrix
+                        // this is a significant error and it may be good idea to abort the process if this happens,
+                        // since many people will miss the error and not realize that most of the model is being quantized without an imatrix
+                        // tok_embd should be ignored in this case, since it always causes this warning
+                        if (name != tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
+                            throw std::runtime_error(format("imatrix size %d is different from tensor size %d for %s",
+                                    int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name));
+                        }
+                    }
+                }
+            }
+            if ((new_type == GGML_TYPE_IQ2_XXS ||
+                 new_type == GGML_TYPE_IQ2_XS  ||
+                 new_type == GGML_TYPE_IQ2_S   ||
+                 new_type == GGML_TYPE_IQ1_S   ||
+                (new_type == GGML_TYPE_IQ1_M && strcmp(tensor->name, "token_embd.weight") && strcmp(tensor->name, "output.weight"))  ||
+                (new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) {
+                LLAMA_LOG_ERROR("\n\n============================================================\n");
+                LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
+                LLAMA_LOG_ERROR("The result will be garbage, so bailing out\n");
+                LLAMA_LOG_ERROR("============================================================\n\n");
+                throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
+            }
+
+            float * f32_data;
+
+            if (tensor->type == GGML_TYPE_F32) {
+                f32_data = (float *) tensor->data;
+            } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
+                throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
+            } else {
+                llama_tensor_dequantize_internal(tensor, f32_conv_buf, workers, nelements, nthread);
+                f32_data = (float *) f32_conv_buf.data();
+            }
+
+            LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
+            fflush(stdout);
+
+            if (work.size() < (size_t)nelements * 4) {
+                work.resize(nelements * 4); // upper bound on size
+            }
+            new_data = work.data();
+
+            const int64_t n_per_row = tensor->ne[0];
+            const int64_t nrows = tensor->ne[1];
+
+            static const int64_t min_chunk_size = 32 * 512;
+            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
+
+            const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
+            const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
+            const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1;
+
+            // quantize each expert separately since they have different importance matrices
+            new_size = 0;
+            for (int64_t i03 = 0; i03 < tensor->ne[2]; ++i03) {
+                const float * f32_data_03 = f32_data + i03 * nelements_matrix;
+                void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
+                const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
+
+                new_size += llama_tensor_quantize_internal(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
+            }
+            LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
+        }
+        total_size_org += ggml_nbytes(tensor);
+        total_size_new += new_size;
+
+        // update the gguf meta data as we go
+        gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
+        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data, new_size);
+
+        // write tensor data + padding
+        fout.write((const char *) new_data, new_size);
+        zeros(fout, GGML_PAD(new_size, align) - new_size);
+    }
+    close_ofstream();
+
+    LLAMA_LOG_INFO("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+
+    if (qs.n_fallback > 0) {
+        LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) required fallback quantization\n",
+                __func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback);
+    }
+}
+
+//
+// interface implementation
+//
+
+struct llama_model_quantize_params llama_model_quantize_default_params() {
+    struct llama_model_quantize_params result = {
+        /*.nthread                     =*/ 0,
+        /*.ftype                       =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
+        /*.output_tensor_type          =*/ GGML_TYPE_COUNT,
+        /*.token_embedding_type        =*/ GGML_TYPE_COUNT,
+        /*.allow_requantize            =*/ false,
+        /*.quantize_output_tensor      =*/ true,
+        /*.only_copy                   =*/ false,
+        /*.pure                        =*/ false,
+        /*.keep_split                  =*/ false,
+        /*.imatrix                     =*/ nullptr,
+        /*.kv_overrides                =*/ nullptr,
+    };
+
+    return result;
+}
+
+uint32_t llama_model_quantize(
+        const char * fname_inp,
+        const char * fname_out,
+        const llama_model_quantize_params * params) {
+    try {
+        llama_model_quantize_internal(fname_inp, fname_out, params);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: failed to quantize: %s\n", __func__, err.what());
+        return 1;
+    }
+
+    return 0;
+}

src/llama-quant.h

@@ -0,0 +1 @@
+#pragma once

src/llama-sampling.cpp

@@ -1,5 +1,6 @@
 #include "llama-sampling.h"
 
+#include "llama-impl.h"
 #include "llama-vocab.h"
 #include "llama-grammar.h"
 
@@ -14,6 +15,118 @@
 #include <numeric>
 #include <random>
 #include <unordered_map>
+#include <stdexcept>
+
+// the ring buffer works similarly to std::deque, but with a fixed capacity
+template<typename T>
+struct ring_buffer {
+    ring_buffer(size_t cap) : capacity(cap), data(cap) {}
+
+    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 (capacity == 0) {
+            throw std::runtime_error("ring buffer: capacity is zero");
+        }
+
+        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;
+    }
+
+    //T & operator[](size_t i) {
+    //    if (i >= sz) {
+    //        throw std::runtime_error("ring buffer: index out of bounds");
+    //    }
+    //    return data[(first + i) % capacity];
+    //}
+
+    //const T & at(size_t i) const {
+    //    if (i >= sz) {
+    //        throw std::runtime_error("ring buffer: index out of bounds");
+    //    }
+    //    return data[(first + i) % capacity];
+    //}
+
+    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;
+};
 
 static int llama_sample_dist(llama_token_data_array * cur_p, std::mt19937 & rng) {
     // iterator for the probabilities

src/llama-vocab.cpp

@@ -1,5 +1,7 @@
 #include "llama-vocab.h"
 
+#include "llama-impl.h"
+
 #include "unicode.h"
 
 #include <algorithm>
@@ -16,22 +18,6 @@
 // helpers
 //
 
-LLAMA_ATTRIBUTE_FORMAT(1, 2)
-static std::string format(const char * fmt, ...) {
-    va_list ap;
-    va_list ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
 struct naive_trie {
     naive_trie() : has_value(false), value(0) {
     }

src/llama-vocab.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "llama-impl.h"
+#include "llama.h"
 
 #include <string>
 #include <vector>
@@ -8,6 +8,18 @@
 #include <map>
 #include <set>
 
+static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
+    switch (type) {
+        case LLAMA_VOCAB_TYPE_NONE: return "no vocab";
+        case LLAMA_VOCAB_TYPE_SPM:  return "SPM";
+        case LLAMA_VOCAB_TYPE_BPE:  return "BPE";
+        case LLAMA_VOCAB_TYPE_WPM:  return "WPM";
+        case LLAMA_VOCAB_TYPE_UGM:  return "UGM";
+        case LLAMA_VOCAB_TYPE_RWKV: return "RWKV";
+        default:                    return "unknown";
+    }
+}
+
 struct llm_tokenizer;
 
 struct llama_vocab {