Merge pull request #65 from jmorganca/bindings

call llama.cpp directly from go

.gitignore

@@ -3,8 +3,5 @@
 .env
 .venv
 *.spec
-build
 dist
-__pycache__
 ollama
-ggml-metal.metal

Makefile

@@ -1,19 +0,0 @@
-default: ollama
-
-.PHONY: llama
-llama:
-	cmake -S llama -B llama/build -DLLAMA_METAL=on
-	cmake --build llama/build
-
-.PHONY: ollama
-ollama: llama
-	go build .
-
-.PHONY: app
-app: ollama
-	npm install --prefix app
-	npm run --prefix app make:sign
-
-clean:
-	go clean
-	rm -rf llama/build

README.md

@@ -75,7 +75,7 @@ ollama run ~/Downloads/vicuna-7b-v1.3.ggmlv3.q4_1.bin
 ## Building
 
 ```
-make
+go build .
 ```
 
 To run it start the server:

api/types.go

@@ -1,5 +1,7 @@
 package api
 
+import "runtime"
+
 type PullRequest struct {
 	Model string `json:"model"`
 }
@@ -14,93 +16,76 @@ type GenerateRequest struct {
 	Model  string `json:"model"`
 	Prompt string `json:"prompt"`
 
-	ModelOptions   *ModelOptions   `json:"model_opts,omitempty"`
+	Options `json:"options"`
-	PredictOptions *PredictOptions `json:"predict_opts,omitempty"`
-}
-
-type ModelOptions struct {
-	ContextSize int    `json:"context_size,omitempty"`
-	Seed        int    `json:"seed,omitempty"`
-	NBatch      int    `json:"n_batch,omitempty"`
-	F16Memory   bool   `json:"memory_f16,omitempty"`
-	MLock       bool   `json:"mlock,omitempty"`
-	MMap        bool   `json:"mmap,omitempty"`
-	VocabOnly   bool   `json:"vocab_only,omitempty"`
-	LowVRAM     bool   `json:"low_vram,omitempty"`
-	Embeddings  bool   `json:"embeddings,omitempty"`
-	NUMA        bool   `json:"numa,omitempty"`
-	NGPULayers  int    `json:"gpu_layers,omitempty"`
-	MainGPU     string `json:"main_gpu,omitempty"`
-	TensorSplit string `json:"tensor_split,omitempty"`
-}
-
-type PredictOptions struct {
-	Seed        int     `json:"seed,omitempty"`
-	Threads     int     `json:"threads,omitempty"`
-	Tokens      int     `json:"tokens,omitempty"`
-	TopK        int     `json:"top_k,omitempty"`
-	Repeat      int     `json:"repeat,omitempty"`
-	Batch       int     `json:"batch,omitempty"`
-	NKeep       int     `json:"nkeep,omitempty"`
-	TopP        float64 `json:"top_p,omitempty"`
-	Temperature float64 `json:"temp,omitempty"`
-	Penalty     float64 `json:"penalty,omitempty"`
-	F16KV       bool
-	DebugMode   bool
-	StopPrompts []string
-	IgnoreEOS   bool `json:"ignore_eos,omitempty"`
-
-	TailFreeSamplingZ float64 `json:"tfs_z,omitempty"`
-	TypicalP          float64 `json:"typical_p,omitempty"`
-	FrequencyPenalty  float64 `json:"freq_penalty,omitempty"`
-	PresencePenalty   float64 `json:"pres_penalty,omitempty"`
-	Mirostat          int     `json:"mirostat,omitempty"`
-	MirostatETA       float64 `json:"mirostat_lr,omitempty"`
-	MirostatTAU       float64 `json:"mirostat_ent,omitempty"`
-	PenalizeNL        bool    `json:"penalize_nl,omitempty"`
-	LogitBias         string  `json:"logit_bias,omitempty"`
-
-	PathPromptCache string
-	MLock           bool `json:"mlock,omitempty"`
-	MMap            bool `json:"mmap,omitempty"`
-	PromptCacheAll  bool
-	PromptCacheRO   bool
-	MainGPU         string
-	TensorSplit     string
-}
-
-var DefaultModelOptions ModelOptions = ModelOptions{
-	ContextSize: 512,
-	Seed:        0,
-	F16Memory:   true,
-	MLock:       false,
-	Embeddings:  true,
-	MMap:        true,
-	LowVRAM:     false,
-}
-
-var DefaultPredictOptions PredictOptions = PredictOptions{
-	Seed:              -1,
-	Threads:           -1,
-	Tokens:            512,
-	Penalty:           1.1,
-	Repeat:            64,
-	Batch:             512,
-	NKeep:             64,
-	TopK:              90,
-	TopP:              0.86,
-	TailFreeSamplingZ: 1.0,
-	TypicalP:          1.0,
-	Temperature:       0.8,
-	FrequencyPenalty:  0.0,
-	PresencePenalty:   0.0,
-	Mirostat:          0,
-	MirostatTAU:       5.0,
-	MirostatETA:       0.1,
-	MMap:              true,
-	StopPrompts:       []string{"llama"},
 }
 
 type GenerateResponse struct {
 	Response string `json:"response"`
 }
+
+type Options struct {
+	Seed int `json:"seed,omitempty"`
+
+	// Backend options
+	UseNUMA bool `json:"numa,omitempty"`
+
+	// Model options
+	NumCtx        int  `json:"num_ctx,omitempty"`
+	NumBatch      int  `json:"num_batch,omitempty"`
+	NumGPU        int  `json:"num_gpu,omitempty"`
+	MainGPU       int  `json:"main_gpu,omitempty"`
+	LowVRAM       bool `json:"low_vram,omitempty"`
+	F16KV         bool `json:"f16_kv,omitempty"`
+	LogitsAll     bool `json:"logits_all,omitempty"`
+	VocabOnly     bool `json:"vocab_only,omitempty"`
+	UseMMap       bool `json:"use_mmap,omitempty"`
+	UseMLock      bool `json:"use_mlock,omitempty"`
+	EmbeddingOnly bool `json:"embedding_only,omitempty"`
+
+	// Predict options
+	RepeatLastN      int     `json:"repeat_last_n,omitempty"`
+	RepeatPenalty    float32 `json:"repeat_penalty,omitempty"`
+	FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
+	PresencePenalty  float32 `json:"presence_penalty,omitempty"`
+	Temperature      float32 `json:"temperature,omitempty"`
+	TopK             int     `json:"top_k,omitempty"`
+	TopP             float32 `json:"top_p,omitempty"`
+	TFSZ             float32 `json:"tfs_z,omitempty"`
+	TypicalP         float32 `json:"typical_p,omitempty"`
+	Mirostat         int     `json:"mirostat,omitempty"`
+	MirostatTau      float32 `json:"mirostat_tau,omitempty"`
+	MirostatEta      float32 `json:"mirostat_eta,omitempty"`
+
+	NumThread int `json:"num_thread,omitempty"`
+}
+
+func DefaultOptions() Options {
+	return Options{
+		Seed: -1,
+
+		UseNUMA: false,
+
+		NumCtx:   512,
+		NumBatch: 512,
+		NumGPU:   1,
+		LowVRAM:  false,
+		F16KV:    true,
+		UseMMap:  true,
+		UseMLock: false,
+
+		RepeatLastN:      512,
+		RepeatPenalty:    1.1,
+		FrequencyPenalty: 0.0,
+		PresencePenalty:  0.0,
+		Temperature:      0.8,
+		TopK:             40,
+		TopP:             0.9,
+		TFSZ:             1.0,
+		TypicalP:         1.0,
+		Mirostat:         0,
+		MirostatTau:      5.0,
+		MirostatEta:      0.1,
+
+		NumThread: runtime.NumCPU(),
+	}
+}

go.mod

@@ -39,6 +39,7 @@ require (
 	golang.org/x/arch v0.3.0 // indirect
 	golang.org/x/crypto v0.10.0 // indirect
 	golang.org/x/net v0.10.0 // indirect
+	golang.org/x/sync v0.3.0
 	golang.org/x/sys v0.10.0 // indirect
 	golang.org/x/term v0.10.0
 	golang.org/x/text v0.10.0 // indirect

go.sum

@@ -99,6 +99,8 @@ golang.org/x/net v0.10.0/go.mod h1:0qNGK6F8kojg2nk9dLZ2mShWaEBan6FAoqfSigmmuDg=
 golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20220722155255-886fb9371eb4/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.1.0/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sync v0.3.0 h1:ftCYgMx6zT/asHUrPw8BLLscYtGznsLAnjq5RH9P66E=
+golang.org/x/sync v0.3.0/go.mod h1:FU7BRWz2tNW+3quACPkgCx/L+uEAv1htQ0V83Z9Rj+Y=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20210615035016-665e8c7367d1/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

llama/.gitignore

@@ -1 +0,0 @@
-build

llama/CMakeLists.txt

@@ -1,23 +0,0 @@
-cmake_minimum_required(VERSION 3.12)
-project(binding)
-
-include(FetchContent)
-
-FetchContent_Declare(
-    llama_cpp
-    GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-    GIT_TAG        55dbb91
-)
-
-FetchContent_MakeAvailable(llama_cpp)
-
-add_library(binding ${CMAKE_CURRENT_SOURCE_DIR}/binding/binding.cpp ${llama_cpp_SOURCE_DIR}/examples/common.cpp)
-target_include_directories(binding PRIVATE ${llama_cpp_SOURCE_DIR}/examples)
-target_link_libraries(binding llama ggml_static)
-
-if (LLAMA_METAL)
-    configure_file(${llama_cpp_SOURCE_DIR}/ggml-metal.metal ${CMAKE_CURRENT_BINARY_DIR}/../../ggml-metal.metal COPYONLY)
-endif()
-
-add_custom_target(copy_libllama ALL COMMAND ${CMAKE_COMMAND} -E copy_if_different $<TARGET_FILE:llama> ${CMAKE_CURRENT_BINARY_DIR})
-add_custom_target(copy_libggml_static ALL COMMAND ${CMAKE_COMMAND} -E copy_if_different $<TARGET_FILE:ggml_static> ${CMAKE_CURRENT_BINARY_DIR})

llama/binding/binding.cpp

@@ -1,705 +0,0 @@
-// MIT License
-
-// Copyright (c) 2023 go-skynet authors
-
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-
-// The above copyright notice and this permission notice shall be included in all
-// copies or substantial portions of the Software.
-
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-// SOFTWARE.
-
-#include "common.h"
-#include "llama.h"
-
-#include "binding.h"
-
-#include <cassert>
-#include <cinttypes>
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <fstream>
-#include <iostream>
-#include <regex>
-#include <sstream>
-#include <string>
-#include <vector>
-#if defined(__unix__) || (defined(__APPLE__) && defined(__MACH__))
-#include <signal.h>
-#include <unistd.h>
-#elif defined(_WIN32)
-#define WIN32_LEAN_AND_MEAN
-#define NOMINMAX
-#include <signal.h>
-#include <windows.h>
-#endif
-
-#if defined(__unix__) || (defined(__APPLE__) && defined(__MACH__)) || \
-    defined(_WIN32)
-void sigint_handler(int signo) {
-  if (signo == SIGINT) {
-    _exit(130);
-  }
-}
-#endif
-
-int get_embeddings(void *params_ptr, void *state_pr, float *res_embeddings) {
-  gpt_params *params_p = (gpt_params *)params_ptr;
-  llama_context *ctx = (llama_context *)state_pr;
-  gpt_params params = *params_p;
-
-  if (params.seed <= 0) {
-    params.seed = time(NULL);
-  }
-
-  std::mt19937 rng(params.seed);
-
-  llama_init_backend(params.numa);
-
-  int n_past = 0;
-
-  // Add a space in front of the first character to match OG llama tokenizer
-  // behavior
-  params.prompt.insert(0, 1, ' ');
-
-  // tokenize the prompt
-  auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
-
-  // determine newline token
-  auto llama_token_newline = ::llama_tokenize(ctx, "\n", false);
-
-  if (embd_inp.size() > 0) {
-    if (llama_eval(ctx, embd_inp.data(), embd_inp.size(), n_past,
-                   params.n_threads)) {
-      fprintf(stderr, "%s : failed to eval\n", __func__);
-      return 1;
-    }
-  }
-
-  const int n_embd = llama_n_embd(ctx);
-
-  const auto embeddings = llama_get_embeddings(ctx);
-
-  for (int i = 0; i < n_embd; i++) {
-    res_embeddings[i] = embeddings[i];
-  }
-
-  return 0;
-}
-
-int get_token_embeddings(void *params_ptr, void *state_pr, int *tokens,
-                         int tokenSize, float *res_embeddings) {
-  gpt_params *params_p = (gpt_params *)params_ptr;
-  llama_context *ctx = (llama_context *)state_pr;
-  gpt_params params = *params_p;
-
-  for (int i = 0; i < tokenSize; i++) {
-    auto token_str = llama_token_to_str(ctx, tokens[i]);
-    if (token_str == nullptr) {
-      continue;
-    }
-    std::vector<std::string> my_vector;
-    std::string str_token(token_str); // create a new std::string from the char*
-    params_p->prompt += str_token;
-  }
-
-  return get_embeddings(params_ptr, state_pr, res_embeddings);
-}
-
-int eval(void *params_ptr, void *state_pr, char *text) {
-  gpt_params *params_p = (gpt_params *)params_ptr;
-  llama_context *ctx = (llama_context *)state_pr;
-
-  auto n_past = 0;
-  auto last_n_tokens_data =
-      std::vector<llama_token>(params_p->repeat_last_n, 0);
-
-  auto tokens = std::vector<llama_token>(params_p->n_ctx);
-  auto n_prompt_tokens =
-      llama_tokenize(ctx, text, tokens.data(), tokens.size(), true);
-
-  if (n_prompt_tokens < 1) {
-    fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);
-    return 1;
-  }
-
-  // evaluate prompt
-  return llama_eval(ctx, tokens.data(), n_prompt_tokens, n_past,
-                    params_p->n_threads);
-}
-
-int llama_predict(void *params_ptr, void *state_pr, char *result, bool debug) {
-  gpt_params *params_p = (gpt_params *)params_ptr;
-  llama_context *ctx = (llama_context *)state_pr;
-
-  gpt_params params = *params_p;
-
-  const int n_ctx = llama_n_ctx(ctx);
-
-  if (params.seed <= 0) {
-    params.seed = time(NULL);
-  }
-
-  std::mt19937 rng(params.seed);
-
-  std::string path_session = params.path_prompt_cache;
-  std::vector<llama_token> session_tokens;
-
-  if (!path_session.empty()) {
-    if (debug) {
-      fprintf(stderr, "%s: attempting to load saved session from '%s'\n",
-              __func__, path_session.c_str());
-    }
-    // fopen to check for existing session
-    FILE *fp = std::fopen(path_session.c_str(), "rb");
-    if (fp != NULL) {
-      std::fclose(fp);
-
-      session_tokens.resize(n_ctx);
-      size_t n_token_count_out = 0;
-      if (!llama_load_session_file(
-              ctx, path_session.c_str(), session_tokens.data(),
-              session_tokens.capacity(), &n_token_count_out)) {
-        fprintf(stderr, "%s: error: failed to load session file '%s'\n",
-                __func__, path_session.c_str());
-        return 1;
-      }
-      session_tokens.resize(n_token_count_out);
-      llama_set_rng_seed(ctx, params.seed);
-      if (debug) {
-        fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n",
-                __func__, (int)session_tokens.size());
-      }
-    } else {
-      if (debug) {
-        fprintf(stderr, "%s: session file does not exist, will create\n",
-                __func__);
-      }
-    }
-  }
-
-  std::vector<llama_token> embd_inp;
-  if (!params.prompt.empty() || session_tokens.empty()) {
-    // Add a space in front of the first character to match OG llama tokenizer
-    // behavior
-    params.prompt.insert(0, 1, ' ');
-
-    embd_inp = ::llama_tokenize(ctx, params.prompt, true);
-  } else {
-    embd_inp = session_tokens;
-  }
-
-  // debug message about similarity of saved session, if applicable
-  size_t n_matching_session_tokens = 0;
-  if (session_tokens.size()) {
-    for (llama_token id : session_tokens) {
-      if (n_matching_session_tokens >= embd_inp.size() ||
-          id != embd_inp[n_matching_session_tokens]) {
-        break;
-      }
-      n_matching_session_tokens++;
-    }
-    if (debug) {
-      if (params.prompt.empty() &&
-          n_matching_session_tokens == embd_inp.size()) {
-        fprintf(stderr, "%s: using full prompt from session file\n", __func__);
-      } else if (n_matching_session_tokens >= embd_inp.size()) {
-        fprintf(stderr, "%s: session file has exact match for prompt!\n",
-                __func__);
-      } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
-        fprintf(stderr,
-                "%s: warning: session file has low similarity to prompt (%zu / "
-                "%zu tokens); will mostly be reevaluated\n",
-                __func__, n_matching_session_tokens, embd_inp.size());
-      } else {
-        fprintf(stderr, "%s: session file matches %zu / %zu tokens of prompt\n",
-                __func__, n_matching_session_tokens, embd_inp.size());
-      }
-    }
-  }
-  // if we will use the cache for the full prompt without reaching the end of
-  // the cache, force reevaluation of the last token token to recalculate the
-  // cached logits
-  if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() &&
-      session_tokens.size() > embd_inp.size()) {
-    session_tokens.resize(embd_inp.size() - 1);
-  }
-  // number of tokens to keep when resetting context
-  if (params.n_keep < 0 || params.n_keep > (int)embd_inp.size()) {
-    params.n_keep = (int)embd_inp.size();
-  }
-
-  // determine newline token
-  auto llama_token_newline = ::llama_tokenize(ctx, "\n", false);
-
-  // TODO: replace with ring-buffer
-  std::vector<llama_token> last_n_tokens(n_ctx);
-  std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
-
-  bool need_to_save_session =
-      !path_session.empty() && n_matching_session_tokens < embd_inp.size();
-  int n_past = 0;
-  int n_remain = params.n_predict;
-  int n_consumed = 0;
-  int n_session_consumed = 0;
-
-  std::vector<llama_token> embd;
-  std::string res = "";
-
-  // do one empty run to warm up the model
-  {
-    const std::vector<llama_token> tmp = {
-        llama_token_bos(),
-    };
-    llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
-    llama_reset_timings(ctx);
-  }
-
-  while (n_remain != 0) {
-    // predict
-    if (embd.size() > 0) {
-      // infinite text generation via context swapping
-      // if we run out of context:
-      // - take the n_keep first tokens from the original prompt (via n_past)
-      // - take half of the last (n_ctx - n_keep) tokens and recompute the
-      // logits in batches
-      if (n_past + (int)embd.size() > n_ctx) {
-        const int n_left = n_past - params.n_keep;
-
-        // always keep the first token - BOS
-        n_past = std::max(1, params.n_keep);
-
-        // insert n_left/2 tokens at the start of embd from last_n_tokens
-        embd.insert(embd.begin(),
-                    last_n_tokens.begin() + n_ctx - n_left / 2 - embd.size(),
-                    last_n_tokens.end() - embd.size());
-
-        // stop saving session if we run out of context
-        path_session.clear();
-
-        // printf("\n---\n");
-        // printf("resetting: '");
-        // for (int i = 0; i < (int) embd.size(); i++) {
-        //     printf("%s", llama_token_to_str(ctx, embd[i]));
-        // }
-        // printf("'\n");
-        // printf("\n---\n");
-      }
-
-      // try to reuse a matching prefix from the loaded session instead of
-      // re-eval (via n_past)
-      if (n_session_consumed < (int)session_tokens.size()) {
-        size_t i = 0;
-        for (; i < embd.size(); i++) {
-          if (embd[i] != session_tokens[n_session_consumed]) {
-            session_tokens.resize(n_session_consumed);
-            break;
-          }
-
-          n_past++;
-          n_session_consumed++;
-
-          if (n_session_consumed >= (int)session_tokens.size()) {
-            ++i;
-            break;
-          }
-        }
-        if (i > 0) {
-          embd.erase(embd.begin(), embd.begin() + i);
-        }
-      }
-
-      // evaluate tokens in batches
-      // embd is typically prepared beforehand to fit within a batch, but not
-      // always
-      for (int i = 0; i < (int)embd.size(); i += params.n_batch) {
-        int n_eval = (int)embd.size() - i;
-        if (n_eval > params.n_batch) {
-          n_eval = params.n_batch;
-        }
-        if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
-          fprintf(stderr, "%s : failed to eval\n", __func__);
-          return 1;
-        }
-        n_past += n_eval;
-      }
-
-      if (embd.size() > 0 && !path_session.empty()) {
-        session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
-        n_session_consumed = session_tokens.size();
-      }
-    }
-
-    embd.clear();
-
-    if ((int)embd_inp.size() <= n_consumed) {
-      // out of user input, sample next token
-      const float temp = params.temp;
-      const int32_t top_k =
-          params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
-      const float top_p = params.top_p;
-      const float tfs_z = params.tfs_z;
-      const float typical_p = params.typical_p;
-      const int32_t repeat_last_n =
-          params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
-      const float repeat_penalty = params.repeat_penalty;
-      const float alpha_presence = params.presence_penalty;
-      const float alpha_frequency = params.frequency_penalty;
-      const int mirostat = params.mirostat;
-      const float mirostat_tau = params.mirostat_tau;
-      const float mirostat_eta = params.mirostat_eta;
-      const bool penalize_nl = params.penalize_nl;
-
-      // optionally save the session on first sample (for faster prompt loading
-      // next time)
-      if (!path_session.empty() && need_to_save_session &&
-          !params.prompt_cache_ro) {
-        need_to_save_session = false;
-        llama_save_session_file(ctx, path_session.c_str(),
-                                session_tokens.data(), session_tokens.size());
-      }
-
-      llama_token id = 0;
-
-      {
-        auto logits = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
-
-        // Apply params.logit_bias map
-        for (auto it = params.logit_bias.begin(); it != params.logit_bias.end();
-             it++) {
-          logits[it->first] += it->second;
-        }
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-          candidates.emplace_back(
-              llama_token_data{token_id, logits[token_id], 0.0f});
-        }
-
-        llama_token_data_array candidates_p = {candidates.data(),
-                                               candidates.size(), false};
-
-        // Apply penalties
-        float nl_logit = logits[llama_token_nl()];
-        auto last_n_repeat =
-            std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
-        llama_sample_repetition_penalty(
-            ctx, &candidates_p,
-            last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-            last_n_repeat, repeat_penalty);
-        llama_sample_frequency_and_presence_penalties(
-            ctx, &candidates_p,
-            last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-            last_n_repeat, alpha_frequency, alpha_presence);
-        if (!penalize_nl) {
-          logits[llama_token_nl()] = nl_logit;
-        }
-
-        if (temp <= 0) {
-          // Greedy sampling
-          id = llama_sample_token_greedy(ctx, &candidates_p);
-        } else {
-          if (mirostat == 1) {
-            static float mirostat_mu = 2.0f * mirostat_tau;
-            const int mirostat_m = 100;
-            llama_sample_temperature(ctx, &candidates_p, temp);
-            id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau,
-                                             mirostat_eta, mirostat_m,
-                                             &mirostat_mu);
-          } else if (mirostat == 2) {
-            static float mirostat_mu = 2.0f * mirostat_tau;
-            llama_sample_temperature(ctx, &candidates_p, temp);
-            id = llama_sample_token_mirostat_v2(
-                ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-          } else {
-            // Temperature sampling
-            llama_sample_top_k(ctx, &candidates_p, top_k, 1);
-            llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
-            llama_sample_typical(ctx, &candidates_p, typical_p, 1);
-            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-            llama_sample_temperature(ctx, &candidates_p, temp);
-            id = llama_sample_token(ctx, &candidates_p);
-          }
-        }
-        // printf("`%d`", candidates_p.size);
-
-        last_n_tokens.erase(last_n_tokens.begin());
-        last_n_tokens.push_back(id);
-      }
-
-      // add it to the context
-      embd.push_back(id);
-
-      // decrement remaining sampling budget
-      --n_remain;
-
-      // call the token callback, no need to check if one is actually
-      // registered, that will be handled on the Go side.
-      auto token_str = llama_token_to_str(ctx, id);
-      if (!tokenCallback(state_pr, (char *)token_str)) {
-        break;
-      }
-    } else {
-      // some user input remains from prompt or interaction, forward it to
-      // processing
-      while ((int)embd_inp.size() > n_consumed) {
-        embd.push_back(embd_inp[n_consumed]);
-        last_n_tokens.erase(last_n_tokens.begin());
-        last_n_tokens.push_back(embd_inp[n_consumed]);
-        ++n_consumed;
-        if ((int)embd.size() >= params.n_batch) {
-          break;
-        }
-      }
-    }
-
-    for (auto id : embd) {
-      res += llama_token_to_str(ctx, id);
-    }
-
-    // check for stop prompt
-    if (params.antiprompt.size()) {
-      std::string last_output;
-      for (auto id : last_n_tokens) {
-        last_output += llama_token_to_str(ctx, id);
-      }
-      // Check if each of the reverse prompts appears at the end of the output.
-      for (std::string &antiprompt : params.antiprompt) {
-        // size_t extra_padding = params.interactive ? 0 : 2;
-        size_t extra_padding = 2;
-        size_t search_start_pos =
-            last_output.length() >
-                    static_cast<size_t>(antiprompt.length() + extra_padding)
-                ? last_output.length() -
-                      static_cast<size_t>(antiprompt.length() + extra_padding)
-                : 0;
-
-        if (last_output.find(antiprompt.c_str(), search_start_pos) !=
-            std::string::npos) {
-          goto end;
-        }
-      }
-    }
-
-    // end of text token
-    if (!embd.empty() && embd.back() == llama_token_eos()) {
-      break;
-    }
-  }
-
-  if (!path_session.empty() && params.prompt_cache_all &&
-      !params.prompt_cache_ro) {
-    if (debug) {
-      fprintf(stderr, "\n%s: saving final output to session file '%s'\n",
-              __func__, path_session.c_str());
-    }
-    llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(),
-                            session_tokens.size());
-  }
-
-end:
-#if defined(_WIN32)
-  signal(SIGINT, SIG_DFL);
-#endif
-
-  if (debug) {
-    llama_print_timings(ctx);
-    llama_reset_timings(ctx);
-  }
-
-  strcpy(result, res.c_str());
-  return 0;
-}
-
-void llama_binding_free_model(void *state_ptr) {
-  llama_context *ctx = (llama_context *)state_ptr;
-  llama_free(ctx);
-}
-
-void llama_free_params(void *params_ptr) {
-  gpt_params *params = (gpt_params *)params_ptr;
-  delete params;
-}
-
-int load_state(void *ctx, char *statefile, char *modes) {
-  llama_context *state = (llama_context *)ctx;
-  const llama_context *constState = static_cast<const llama_context *>(state);
-  const size_t state_size = llama_get_state_size(state);
-  uint8_t *state_mem = new uint8_t[state_size];
-
-  {
-    FILE *fp_read = fopen(statefile, modes);
-    if (state_size != llama_get_state_size(constState)) {
-      fprintf(stderr, "\n%s : failed to validate state size\n", __func__);
-      return 1;
-    }
-
-    const size_t ret = fread(state_mem, 1, state_size, fp_read);
-    if (ret != state_size) {
-      fprintf(stderr, "\n%s : failed to read state\n", __func__);
-      return 1;
-    }
-
-    llama_set_state_data(
-        state, state_mem); // could also read directly from memory mapped file
-    fclose(fp_read);
-  }
-
-  return 0;
-}
-
-void save_state(void *ctx, char *dst, char *modes) {
-  llama_context *state = (llama_context *)ctx;
-
-  const size_t state_size = llama_get_state_size(state);
-  uint8_t *state_mem = new uint8_t[state_size];
-
-  // Save state (rng, logits, embedding and kv_cache) to file
-  {
-    FILE *fp_write = fopen(dst, modes);
-    llama_copy_state_data(
-        state, state_mem); // could also copy directly to memory mapped file
-    fwrite(state_mem, 1, state_size, fp_write);
-    fclose(fp_write);
-  }
-}
-
-void *llama_allocate_params(
-    const char *prompt, int seed, int threads, int tokens, int top_k,
-    float top_p, float temp, float repeat_penalty, int repeat_last_n,
-    bool ignore_eos, bool memory_f16, int n_batch, int n_keep,
-    const char **antiprompt, int antiprompt_count, float tfs_z, float typical_p,
-    float frequency_penalty, float presence_penalty, int mirostat,
-    float mirostat_eta, float mirostat_tau, bool penalize_nl,
-    const char *logit_bias, bool mlock, bool mmap, const char *maingpu,
-    const char *tensorsplit) {
-  gpt_params *params = new gpt_params;
-  params->seed = seed;
-  params->n_threads = threads;
-  params->n_predict = tokens;
-  params->repeat_last_n = repeat_last_n;
-  params->top_k = top_k;
-  params->top_p = top_p;
-  params->memory_f16 = memory_f16;
-  params->temp = temp;
-  params->use_mmap = mmap;
-  params->use_mlock = mlock;
-  params->repeat_penalty = repeat_penalty;
-  params->n_batch = n_batch;
-  params->n_keep = n_keep;
-  if (maingpu[0] != '\0') {
-    params->main_gpu = std::stoi(maingpu);
-  }
-
-  if (tensorsplit[0] != '\0') {
-    std::string arg_next = tensorsplit;
-    // split string by , and /
-    const std::regex regex{R"([,/]+)"};
-    std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
-    std::vector<std::string> split_arg{it, {}};
-    GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
-
-    for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
-      if (i < split_arg.size()) {
-        params->tensor_split[i] = std::stof(split_arg[i]);
-      } else {
-        params->tensor_split[i] = 0.0f;
-      }
-    }
-  }
-
-  if (ignore_eos) {
-    params->logit_bias[llama_token_eos()] = -INFINITY;
-  }
-
-  for (int i = 0; i < antiprompt_count; i++) {
-    params->antiprompt.push_back(antiprompt[i]);
-  }
-
-  params->tfs_z = tfs_z;
-  params->typical_p = typical_p;
-  params->presence_penalty = presence_penalty;
-  params->mirostat = mirostat;
-  params->mirostat_eta = mirostat_eta;
-  params->mirostat_tau = mirostat_tau;
-  params->penalize_nl = penalize_nl;
-  std::stringstream ss(logit_bias);
-  llama_token key;
-  char sign;
-  std::string value_str;
-  if (ss >> key && ss >> sign && std::getline(ss, value_str) &&
-      (sign == '+' || sign == '-')) {
-    params->logit_bias[key] =
-        std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
-  }
-  params->frequency_penalty = frequency_penalty;
-  params->prompt = prompt;
-
-  return params;
-}
-
-void *load_model(const char *fname, int n_ctx, int n_seed, bool memory_f16,
-                 bool mlock, bool embeddings, bool mmap, bool low_vram,
-                 bool vocab_only, int n_gpu_layers, int n_batch,
-                 const char *maingpu, const char *tensorsplit, bool numa) {
-  // load the model
-  auto lparams = llama_context_default_params();
-
-  lparams.n_ctx = n_ctx;
-  lparams.seed = n_seed;
-  lparams.f16_kv = memory_f16;
-  lparams.embedding = embeddings;
-  lparams.use_mlock = mlock;
-  lparams.n_gpu_layers = n_gpu_layers;
-  lparams.use_mmap = mmap;
-  lparams.low_vram = low_vram;
-  lparams.vocab_only = vocab_only;
-
-  if (maingpu[0] != '\0') {
-    lparams.main_gpu = std::stoi(maingpu);
-  }
-
-  if (tensorsplit[0] != '\0') {
-    std::string arg_next = tensorsplit;
-    // split string by , and /
-    const std::regex regex{R"([,/]+)"};
-    std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
-    std::vector<std::string> split_arg{it, {}};
-    GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
-
-    for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
-      if (i < split_arg.size()) {
-        lparams.tensor_split[i] = std::stof(split_arg[i]);
-      } else {
-        lparams.tensor_split[i] = 0.0f;
-      }
-    }
-  }
-
-  lparams.n_batch = n_batch;
-
-  llama_init_backend(numa);
-  void *res = nullptr;
-  try {
-    res = llama_init_from_file(fname, lparams);
-  } catch (std::runtime_error &e) {
-    fprintf(stderr, "failed %s", e.what());
-    return res;
-  }
-
-  return res;
-}

llama/binding/binding.h

@@ -1,69 +0,0 @@
-// MIT License
-
-// Copyright (c) 2023 go-skynet authors
-
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-
-// The above copyright notice and this permission notice shall be included in all
-// copies or substantial portions of the Software.
-
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-// SOFTWARE.
-
-#ifdef __cplusplus
-
-extern "C" {
-
-#endif
-
-#include <stdbool.h>
-
-extern unsigned char tokenCallback(void *, char *);
-
-int load_state(void *ctx, char *statefile, char *modes);
-
-int eval(void *params_ptr, void *ctx, char *text);
-
-void save_state(void *ctx, char *dst, char *modes);
-
-void *load_model(const char *fname, int n_ctx, int n_seed, bool memory_f16,
-                 bool mlock, bool embeddings, bool mmap, bool low_vram,
-                 bool vocab_only, int n_gpu, int n_batch, const char *maingpu,
-                 const char *tensorsplit, bool numa);
-
-int get_embeddings(void *params_ptr, void *state_pr, float *res_embeddings);
-
-int get_token_embeddings(void *params_ptr, void *state_pr, int *tokens,
-                         int tokenSize, float *res_embeddings);
-
-void *llama_allocate_params(
-    const char *prompt, int seed, int threads, int tokens, int top_k,
-    float top_p, float temp, float repeat_penalty, int repeat_last_n,
-    bool ignore_eos, bool memory_f16, int n_batch, int n_keep,
-    const char **antiprompt, int antiprompt_count, float tfs_z, float typical_p,
-    float frequency_penalty, float presence_penalty, int mirostat,
-    float mirostat_eta, float mirostat_tau, bool penalize_nl,
-    const char *logit_bias, bool mlock, bool mmap, const char *maingpu,
-    const char *tensorsplit);
-
-void llama_free_params(void *params_ptr);
-
-void llama_binding_free_model(void *state);
-
-int llama_predict(void *params_ptr, void *state_pr, char *result, bool debug);
-
-#ifdef __cplusplus
-
-}
-
-#endif

llama/ggml-cuda.h

@@ -0,0 +1,62 @@
+/**
+ * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_CUDA_MAX_DEVICES       16
+
+void   ggml_init_cublas(void);
+void   ggml_cuda_set_tensor_split(const float * tensor_split);
+
+void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+
+// TODO: export these with GGML_API
+void * ggml_cuda_host_malloc(size_t size);
+void   ggml_cuda_host_free(void * ptr);
+
+void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
+void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_scratch_size(size_t scratch_size);
+void   ggml_cuda_free_scratch(void);
+bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+
+#ifdef  __cplusplus
+}
+#endif

llama/ggml-metal.h

@@ -0,0 +1,97 @@
+/**
+ * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include <stddef.h>
+#include <stdbool.h>
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 16
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_metal_context;
+
+// number of command buffers to use
+struct ggml_metal_context * ggml_metal_init(int n_cb);
+void ggml_metal_free(struct ggml_metal_context * ctx);
+
+// set the number of command buffers to use
+void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
+
+// creates a mapping between a host memory buffer and a device memory buffer
+// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
+// - the mapping is used during computation to determine the arguments of the compute kernels
+// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
+// - max_size specifies the maximum size of a tensor and is used to create shared views such
+//   that it is guaranteed that the tensor will fit in at least one of the views
+//
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                       const char * name,
+                             void * data,
+                           size_t   size,
+                           size_t   max_size);
+
+// set data from host memory into the device
+void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// get data from the device into host memory
+void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// same as ggml_graph_compute but uses Metal
+// creates gf->n_threads command buffers in parallel
+void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+#ifdef __cplusplus
+}
+#endif
+

llama/k_quants.h

@@ -0,0 +1,183 @@
+/**
+ * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#include <stdint.h>
+#include <assert.h>
+#include <stddef.h>
+
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elemenets each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
+// Quantization
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+// Quantization with histogram collection
+size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+

llama/llama-util.h

@@ -0,0 +1,530 @@
+/**
+ * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// Internal header to be included only by llama.cpp.
+// Contains wrappers around OS interfaces.
+
+#ifndef LLAMA_UTIL_H
+#define LLAMA_UTIL_H
+
+#include <cstdio>
+#include <cstdint>
+#include <cerrno>
+#include <cstring>
+#include <cstdarg>
+#include <cstdlib>
+#include <climits>
+
+#include <string>
+#include <vector>
+#include <stdexcept>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.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>
+    #include <io.h>
+    #include <stdio.h> // for _fseeki64
+#endif
+
+#define LLAMA_ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            abort(); \
+        } \
+    } while (0)
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
+__attribute__((format(printf, 1, 2)))
+#endif
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    LLAMA_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+struct 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) {
+        fp = std::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 {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        LLAMA_ASSERT(ret == 0); // same
+    }
+
+    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(std::string("unexpectedly reached end of file"));
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    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(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+#if defined(_WIN32)
+static 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
+
+struct llama_mmap {
+    void * addr;
+    size_t size;
+
+    llama_mmap(const llama_mmap &) = delete;
+
+#ifdef _POSIX_MAPPED_FILES
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
+        size = file->size;
+        int fd = fileno(file->fp);
+        int flags = MAP_PRIVATE;
+        // prefetch/readahead impairs performance on NUMA systems
+        if (numa) { prefetch = 0; }
+#ifdef __linux__
+        if (prefetch) { flags |= MAP_POPULATE; }
+#endif
+        addr = mmap(NULL, file->size, PROT_READ | PROT_WRITE, flags, fd, 0);
+        if (addr == MAP_FAILED) {
+            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
+        }
+
+        if (prefetch > 0) {
+            // Advise the kernel to preload the mapped memory
+            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
+                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+        if (numa) {
+            // advise the kernel not to use readahead
+            // (because the next page might not belong on the same node)
+            if (madvise(addr, file->size, MADV_RANDOM)) {
+                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+    }
+
+    ~llama_mmap() {
+        munmap(addr, size);
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
+        (void) numa;
+
+        size = file->size;
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
+
+        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+        DWORD error = GetLastError();
+
+        if (hMapping == NULL) {
+            throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        addr = MapViewOfFile(hMapping, FILE_MAP_COPY, 0, 0, 0);
+        error = GetLastError();
+        CloseHandle(hMapping);
+
+        if (addr == NULL) {
+            throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            WIN32_MEMORY_RANGE_ENTRY range;
+            range.VirtualAddress = addr;
+            range.NumberOfBytes = (SIZE_T)size;
+            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+            }
+        }
+        #else
+        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
+        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
+    }
+
+    ~llama_mmap() {
+        if (!UnmapViewOfFile(addr)) {
+            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    llama_mmap(struct llama_file *, bool prefetch = true, bool numa = false) {
+        (void) prefetch;
+        (void) numa;
+
+        throw std::runtime_error(std::string("mmap not supported"));
+    }
+#endif
+};
+
+// Represents some region of memory being locked using mlock or VirtualLock;
+// will automatically unlock on destruction.
+struct llama_mlock {
+    void * addr = NULL;
+    size_t size = 0;
+    bool failed_already = false;
+
+    llama_mlock() {}
+    llama_mlock(const llama_mlock &) = delete;
+
+    ~llama_mlock() {
+        if (size) {
+            raw_unlock(addr, size);
+        }
+    }
+
+    void init(void * ptr) {
+        LLAMA_ASSERT(addr == NULL && size == 0);
+        addr = ptr;
+    }
+
+    void grow_to(size_t target_size) {
+        LLAMA_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;
+            }
+        }
+    }
+
+#ifdef _POSIX_MEMLOCK_RANGE
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    #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_MLOCK (ulimit -l).\n"
+    #else
+        #define MLOCK_SUGGESTION \
+            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+    #endif
+
+    bool raw_lock(const void * addr, size_t size) {
+        if (!mlock(addr, size)) {
+            return true;
+        } else {
+            char* errmsg = std::strerror(errno);
+            bool suggest = (errno == ENOMEM);
+
+            // Check if the resource limit is fine after all
+            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;
+
+            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
+                    size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
+            return false;
+        }
+    }
+
+    #undef MLOCK_SUGGESTION
+
+    void raw_unlock(void * addr, size_t size) {
+        if (munlock(addr, size)) {
+            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
+        }
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool raw_lock(void * ptr, size_t len) {
+        for (int tries = 1; ; tries++) {
+            if (VirtualLock(ptr, len)) {
+                return true;
+            }
+            if (tries == 2) {
+                fprintf(stderr, "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;
+            }
+
+            // It failed but this was only the first try; increase the working
+            // set size and try again.
+            SIZE_T min_ws_size, max_ws_size;
+            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
+                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+            // Per MSDN: "The maximum number of pages that a process can lock
+            // is equal to the number of pages in its minimum working set minus
+            // a small overhead."
+            // Hopefully a megabyte is enough overhead:
+            size_t increment = len + 1048576;
+            // The minimum must be <= the maximum, so we need to increase both:
+            min_ws_size += increment;
+            max_ws_size += increment;
+            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
+                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+        }
+    }
+
+    void raw_unlock(void * ptr, size_t len) {
+        if (!VirtualUnlock(ptr, len)) {
+            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    size_t lock_granularity() {
+        return (size_t) 65536;
+    }
+
+    bool raw_lock(const void * addr, size_t len) {
+        fprintf(stderr, "warning: mlock not supported on this system\n");
+        return false;
+    }
+
+    void raw_unlock(const void * addr, size_t len) {}
+#endif
+};
+
+// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
+struct llama_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    llama_buffer() = default;
+
+    void resize(size_t len) {
+#ifdef GGML_USE_METAL
+        free(addr);
+        int result = posix_memalign((void **) &addr, getpagesize(), len);
+        if (result == 0) {
+            memset(addr, 0, len);
+        }
+        else {
+            addr = NULL;
+        }
+#else
+        delete[] addr;
+        addr = new uint8_t[len];
+#endif
+        size = len;
+    }
+
+    ~llama_buffer() {
+#ifdef GGML_USE_METAL
+        free(addr);
+#else
+        delete[] addr;
+#endif
+        addr = NULL;
+    }
+
+    // disable copy and move
+    llama_buffer(const llama_buffer&) = delete;
+    llama_buffer(llama_buffer&&) = delete;
+    llama_buffer& operator=(const llama_buffer&) = delete;
+    llama_buffer& operator=(llama_buffer&&) = delete;
+};
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+struct llama_ctx_buffer {
+    uint8_t * addr = NULL;
+    bool is_cuda;
+    size_t size = 0;
+
+    llama_ctx_buffer() = default;
+
+    void resize(size_t size) {
+        free();
+
+        addr = (uint8_t *) ggml_cuda_host_malloc(size);
+        if (addr) {
+            is_cuda = true;
+        }
+        else {
+            // fall back to pageable memory
+            addr = new uint8_t[size];
+            is_cuda = false;
+        }
+        this->size = size;
+    }
+
+    void free() {
+        if (addr) {
+            if (is_cuda) {
+                ggml_cuda_host_free(addr);
+            }
+            else {
+                delete[] addr;
+            }
+        }
+        addr = NULL;
+    }
+
+    ~llama_ctx_buffer() {
+        free();
+    }
+
+    // disable copy and move
+    llama_ctx_buffer(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer(llama_ctx_buffer&&) = delete;
+    llama_ctx_buffer& operator=(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer& operator=(llama_ctx_buffer&&) = delete;
+};
+#else
+typedef llama_buffer llama_ctx_buffer;
+#endif
+
+#endif

llama/llama.go

@@ -1,215 +1,234 @@
-// MIT License
-
-// Copyright (c) 2023 go-skynet authors
-
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-
-// The above copyright notice and this permission notice shall be included in all
-// copies or substantial portions of the Software.
-
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-// SOFTWARE.
-
 package llama
 
-// #cgo LDFLAGS: -Lbuild -lbinding -lllama -lm -lggml_static -lstdc++
+/*
-// #cgo CXXFLAGS: -std=c++11
+#cgo CPPFLAGS: -O3 -DNDEBUG=1
-// #cgo darwin LDFLAGS: -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+#cgo CXXFLAGS: -std=c++11
-// #include "binding/binding.h"
+#cgo darwin CPPFLAGS: -DGGML_USE_METAL=1 -DGGML_METAL_NDEBUG=1
-// #include <stdlib.h>
+#cgo darwin LDFLAGS: -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
-import "C"
+#include <stdlib.h>
+#include "llama.h"
 
+struct llama_sample_options
+{
+	float repeat_penalty;
+	float frequency_penalty;
+	float presence_penalty;
+	float temperature;
+	int32_t top_k;
+	float top_p;
+	float tfs_z;
+	float typical_p;
+	int mirostat;
+	float mirostat_tau;
+	float mirostat_eta;
+};
+
+llama_token llama_sample(
+		struct llama_context *ctx,
+		struct llama_token_data *candidates,
+		size_t n_candidates,
+		const llama_token *last_tokens,
+		size_t n_last_tokens,
+		struct llama_sample_options *opts)
+{
+	llama_token_data_array candidates_p = {
+		candidates,
+		n_candidates,
+		false,
+	};
+
+	llama_sample_repetition_penalty(
+		ctx, &candidates_p,
+		last_tokens, n_last_tokens,
+		opts->repeat_penalty);
+
+	llama_sample_frequency_and_presence_penalties(
+		ctx, &candidates_p,
+		last_tokens, n_last_tokens,
+		opts->frequency_penalty, opts->presence_penalty);
+
+	if (opts->temperature <= 0) {
+		return llama_sample_token_greedy(ctx, &candidates_p);
+	}
+
+	if (opts->mirostat == 1) {
+		int mirostat_m = 100;
+		float mirostat_mu = 2.0f * opts->mirostat_tau;
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token_mirostat(
+			ctx, &candidates_p,
+			opts->mirostat_tau, opts->mirostat_eta,
+			mirostat_m, &mirostat_mu);
+	} else if (opts->mirostat == 2) {
+		float mirostat_mu = 2.0f * opts->mirostat_tau;
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token_mirostat_v2(
+			ctx, &candidates_p,
+			opts->mirostat_tau, opts->mirostat_eta,
+			&mirostat_mu);
+	} else {
+		llama_sample_top_k(ctx, &candidates_p, opts->top_k, 1);
+		llama_sample_tail_free(ctx, &candidates_p, opts->tfs_z, 1);
+		llama_sample_typical(ctx, &candidates_p, opts->typical_p, 1);
+		llama_sample_top_p(ctx, &candidates_p, opts->top_p, 1);
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token(ctx, &candidates_p);
+	}
+}
+*/
+import "C"
 import (
-	"fmt"
+	"errors"
+	"io"
+	"os"
 	"strings"
-	"sync"
 	"unsafe"
+
+	"github.com/jmorganca/ollama/api"
 )
 
-type LLama struct {
+type llama struct {
-	ctx         unsafe.Pointer
+	params *C.struct_llama_context_params
-	embeddings  bool
+	model  *C.struct_llama_model
-	contextSize int
+	ctx    *C.struct_llama_context
+
+	api.Options
 }
 
-func New(model string, mo ModelOptions) (*LLama, error) {
+func New(model string, opts api.Options) (*llama, error) {
-	modelPath := C.CString(model)
+	if _, err := os.Stat(model); err != nil {
-	defer C.free(unsafe.Pointer(modelPath))
+		return nil, err
-
-	ctx := C.load_model(modelPath, C.int(mo.ContextSize), C.int(mo.Seed), C.bool(mo.F16Memory), C.bool(mo.MLock), C.bool(mo.Embeddings), C.bool(mo.MMap), C.bool(mo.LowVRAM), C.bool(mo.VocabOnly), C.int(mo.NGPULayers), C.int(mo.NBatch), C.CString(mo.MainGPU), C.CString(mo.TensorSplit), C.bool(mo.NUMA))
-	if ctx == nil {
-		return nil, fmt.Errorf("failed loading model")
 	}
 
-	ll := &LLama{ctx: ctx, contextSize: mo.ContextSize, embeddings: mo.Embeddings}
+	llm := llama{Options: opts}
 
-	return ll, nil
+	C.llama_backend_init(C.bool(llm.UseNUMA))
+
+	params := C.llama_context_default_params()
+	params.seed = C.uint(llm.Seed)
+	params.n_ctx = C.int(llm.NumCtx)
+	params.n_batch = C.int(llm.NumBatch)
+	params.n_gpu_layers = C.int(llm.NumGPU)
+	params.main_gpu = C.int(llm.MainGPU)
+	params.low_vram = C.bool(llm.LowVRAM)
+	params.f16_kv = C.bool(llm.F16KV)
+	params.logits_all = C.bool(llm.LogitsAll)
+	params.vocab_only = C.bool(llm.VocabOnly)
+	params.use_mmap = C.bool(llm.UseMMap)
+	params.use_mlock = C.bool(llm.UseMLock)
+	params.embedding = C.bool(llm.EmbeddingOnly)
+	llm.params = &params
+
+	cModel := C.CString(model)
+	defer C.free(unsafe.Pointer(cModel))
+
+	llm.model = C.llama_load_model_from_file(cModel, params)
+	llm.ctx = C.llama_new_context_with_model(llm.model, params)
+
+	// warm up the model
+	bos := []C.llama_token{C.llama_token_bos()}
+	C.llama_eval(llm.ctx, unsafe.SliceData(bos), C.int(len(bos)), 0, C.int(opts.NumThread))
+	C.llama_reset_timings(llm.ctx)
+
+	return &llm, nil
 }
 
-func (l *LLama) Free() {
+func (llm *llama) Close() {
-	C.llama_binding_free_model(l.ctx)
+	defer C.llama_free_model(llm.model)
+	defer C.llama_free(llm.ctx)
+
+	C.llama_print_timings(llm.ctx)
 }
 
-func (l *LLama) Eval(text string, po PredictOptions) error {
+func (llm *llama) Predict(prompt string, fn func(string)) error {
-	input := C.CString(text)
+	if tokens := llm.tokenize(prompt); tokens != nil {
-	if po.Tokens == 0 {
+		return llm.generate(tokens, fn)
-		po.Tokens = 99999999
-	}
-	defer C.free(unsafe.Pointer(input))
-
-	reverseCount := len(po.StopPrompts)
-	reversePrompt := make([]*C.char, reverseCount)
-	var pass **C.char
-	for i, s := range po.StopPrompts {
-		cs := C.CString(s)
-		reversePrompt[i] = cs
-		pass = &reversePrompt[0]
-		defer C.free(unsafe.Pointer(cs))
 	}
 
-	cLogitBias := C.CString(po.LogitBias)
+	return errors.New("llama: tokenize")
-	defer C.free(unsafe.Pointer(cLogitBias))
+}
 
-	cMainGPU := C.CString(po.MainGPU)
+func (llm *llama) tokenize(prompt string) []C.llama_token {
-	defer C.free(unsafe.Pointer(cMainGPU))
+	cPrompt := C.CString(prompt)
+	defer C.free(unsafe.Pointer(cPrompt))
 
-	cTensorSplit := C.CString(po.TensorSplit)
+	tokens := make([]C.llama_token, llm.NumCtx)
-	defer C.free(unsafe.Pointer(cTensorSplit))
+	if n := C.llama_tokenize(llm.ctx, cPrompt, unsafe.SliceData(tokens), C.int(len(tokens)), true); n > 0 {
-
+		return tokens[:n]
-	params := C.llama_allocate_params(input, C.int(po.Seed), C.int(po.Threads), C.int(po.Tokens), C.int(po.TopK),
-		C.float(po.TopP), C.float(po.Temperature), C.float(po.Penalty), C.int(po.Repeat),
-		C.bool(po.IgnoreEOS), C.bool(po.F16KV),
-		C.int(po.Batch), C.int(po.NKeep), pass, C.int(reverseCount),
-		C.float(po.TailFreeSamplingZ), C.float(po.TypicalP), C.float(po.FrequencyPenalty), C.float(po.PresencePenalty),
-		C.int(po.Mirostat), C.float(po.MirostatETA), C.float(po.MirostatTAU), C.bool(po.PenalizeNL), cLogitBias,
-		C.bool(po.MLock), C.bool(po.MMap), cMainGPU, cTensorSplit,
-	)
-	defer C.llama_free_params(params)
-
-	ret := C.eval(params, l.ctx, input)
-	if ret != 0 {
-		return fmt.Errorf("inference failed")
 	}
 
 	return nil
 }
 
-func (l *LLama) Predict(text string, po PredictOptions) (string, error) {
+func (llm *llama) detokenize(tokens ...C.llama_token) string {
-	if po.TokenCallback != nil {
+	var sb strings.Builder
-		setCallback(l.ctx, po.TokenCallback)
+	for _, token := range tokens {
+		sb.WriteString(C.GoString(C.llama_token_to_str(llm.ctx, token)))
 	}
 
-	input := C.CString(text)
+	return sb.String()
-	if po.Tokens == 0 {
-		po.Tokens = 99999999
-	}
-	defer C.free(unsafe.Pointer(input))
-
-	out := make([]byte, po.Tokens)
-
-	reverseCount := len(po.StopPrompts)
-	reversePrompt := make([]*C.char, reverseCount)
-	var pass **C.char
-	for i, s := range po.StopPrompts {
-		cs := C.CString(s)
-		reversePrompt[i] = cs
-		pass = &reversePrompt[0]
-		defer C.free(unsafe.Pointer(cs))
-	}
-
-	cLogitBias := C.CString(po.LogitBias)
-	defer C.free(unsafe.Pointer(cLogitBias))
-
-	cMainGPU := C.CString(po.MainGPU)
-	defer C.free(unsafe.Pointer(cMainGPU))
-
-	cTensorSplit := C.CString(po.TensorSplit)
-	defer C.free(unsafe.Pointer(cTensorSplit))
-
-	params := C.llama_allocate_params(input, C.int(po.Seed), C.int(po.Threads), C.int(po.Tokens), C.int(po.TopK),
-		C.float(po.TopP), C.float(po.Temperature), C.float(po.Penalty), C.int(po.Repeat),
-		C.bool(po.IgnoreEOS), C.bool(po.F16KV),
-		C.int(po.Batch), C.int(po.NKeep), pass, C.int(reverseCount),
-		C.float(po.TailFreeSamplingZ), C.float(po.TypicalP), C.float(po.FrequencyPenalty), C.float(po.PresencePenalty),
-		C.int(po.Mirostat), C.float(po.MirostatETA), C.float(po.MirostatTAU), C.bool(po.PenalizeNL), cLogitBias,
-		C.bool(po.MLock), C.bool(po.MMap), cMainGPU, cTensorSplit,
-	)
-	defer C.llama_free_params(params)
-
-	ret := C.llama_predict(params, l.ctx, (*C.char)(unsafe.Pointer(&out[0])), C.bool(po.DebugMode))
-	if ret != 0 {
-		return "", fmt.Errorf("inference failed")
-	}
-	res := C.GoString((*C.char)(unsafe.Pointer(&out[0])))
-
-	res = strings.TrimPrefix(res, " ")
-	res = strings.TrimPrefix(res, text)
-	res = strings.TrimPrefix(res, "\n")
-
-	for _, s := range po.StopPrompts {
-		res = strings.TrimRight(res, s)
-	}
-
-	if po.TokenCallback != nil {
-		setCallback(l.ctx, nil)
-	}
-
-	return res, nil
 }
 
-// CGo only allows us to use static calls from C to Go, we can't just dynamically pass in func's.
+func (llm *llama) generate(tokens []C.llama_token, fn func(string)) error {
-// This is the next best thing, we register the callbacks in this map and call tokenCallback from
+	var opts C.struct_llama_sample_options
-// the C code. We also attach a finalizer to LLama, so it will unregister the callback when the
+	opts.repeat_penalty = C.float(llm.RepeatPenalty)
-// garbage collection frees it.
+	opts.frequency_penalty = C.float(llm.FrequencyPenalty)
+	opts.presence_penalty = C.float(llm.PresencePenalty)
+	opts.temperature = C.float(llm.Temperature)
+	opts.top_k = C.int(llm.TopK)
+	opts.top_p = C.float(llm.TopP)
+	opts.tfs_z = C.float(llm.TFSZ)
+	opts.typical_p = C.float(llm.TypicalP)
+	opts.mirostat = C.int(llm.Mirostat)
+	opts.mirostat_tau = C.float(llm.MirostatTau)
+	opts.mirostat_eta = C.float(llm.MirostatEta)
 
-// SetTokenCallback registers a callback for the individual tokens created when running Predict. It
+	pastTokens := deque[C.llama_token]{capacity: llm.RepeatLastN}
-// will be called once for each token. The callback shall return true as long as the model should
-// continue predicting the next token. When the callback returns false the predictor will return.
-// The tokens are just converted into Go strings, they are not trimmed or otherwise changed. Also
-// the tokens may not be valid UTF-8.
-// Pass in nil to remove a callback.
-//
-// It is save to call this method while a prediction is running.
-func (l *LLama) SetTokenCallback(callback func(token string) bool) {
-	setCallback(l.ctx, callback)
-}
 
-var (
+	for C.llama_get_kv_cache_token_count(llm.ctx) < C.int(llm.NumCtx) {
-	m         sync.Mutex
+		if retval := C.llama_eval(llm.ctx, unsafe.SliceData(tokens), C.int(len(tokens)), C.llama_get_kv_cache_token_count(llm.ctx), C.int(llm.NumThread)); retval != 0 {
-	callbacks = map[uintptr]func(string) bool{}
+			return errors.New("llama: eval")
-)
+		}
 
-//export tokenCallback
+		token, err := llm.sample(pastTokens, &opts)
-func tokenCallback(statePtr unsafe.Pointer, token *C.char) bool {
+		switch {
-	m.Lock()
+		case err != nil:
-	defer m.Unlock()
+			return err
+		case errors.Is(err, io.EOF):
+			return nil
+		}
 
-	if callback, ok := callbacks[uintptr(statePtr)]; ok {
+		fn(llm.detokenize(token))
-		return callback(C.GoString(token))
+
+		tokens = []C.llama_token{token}
+
+		pastTokens.PushLeft(token)
 	}
 
-	return true
+	return nil
 }
 
-// setCallback can be used to register a token callback for LLama. Pass in a nil callback to
+func (llm *llama) sample(pastTokens deque[C.llama_token], opts *C.struct_llama_sample_options) (C.llama_token, error) {
-// remove the callback.
+	numVocab := int(C.llama_n_vocab(llm.ctx))
-func setCallback(statePtr unsafe.Pointer, callback func(string) bool) {
+	logits := unsafe.Slice(C.llama_get_logits(llm.ctx), numVocab)
-	m.Lock()
-	defer m.Unlock()
 
-	if callback == nil {
+	candidates := make([]C.struct_llama_token_data, 0, numVocab)
-		delete(callbacks, uintptr(statePtr))
+	for i := 0; i < numVocab; i++ {
-	} else {
+		candidates = append(candidates, C.llama_token_data{
-		callbacks[uintptr(statePtr)] = callback
+			id:    C.int(i),
+			logit: logits[i],
+			p:     0,
+		})
 	}
+
+	token := C.llama_sample(
+		llm.ctx,
+		unsafe.SliceData(candidates), C.ulong(len(candidates)),
+		unsafe.SliceData(pastTokens.Data()), C.ulong(pastTokens.Len()),
+		opts)
+	if token != C.llama_token_eos() {
+		return token, nil
+	}
+
+	return 0, io.EOF
 }

llama/llama.h

@@ -0,0 +1,410 @@
+/**
+ * llama.cpp - git 5bf2a2771886ee86137e01dbc7492f78fb392066
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef LLAMA_H
+#define LLAMA_H
+
+#include "ggml.h"
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
+#else
+#define LLAMA_MAX_DEVICES 1
+#endif // GGML_USE_CUBLAS
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAMA_API __declspec(dllexport)
+#        else
+#            define LLAMA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAMA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAMA_API
+#endif
+
+#ifdef __GNUC__
+#    define DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define DEPRECATED(func, hint) func
+#endif
+
+#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
+#define LLAMA_FILE_MAGIC_GGLA        0x67676c61u // 'ggla'
+#define LLAMA_FILE_MAGIC_GGMF        0x67676d66u // 'ggmf'
+#define LLAMA_FILE_MAGIC_GGML        0x67676d6cu // 'ggml'
+#define LLAMA_FILE_MAGIC_GGSN        0x6767736eu // 'ggsn'
+
+#define LLAMA_FILE_VERSION           3
+#define LLAMA_FILE_MAGIC             LLAMA_FILE_MAGIC_GGJT
+#define LLAMA_FILE_MAGIC_UNVERSIONED LLAMA_FILE_MAGIC_GGML
+#define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
+#define LLAMA_SESSION_VERSION        1
+
+#define LLAMA_DEFAULT_SEED           0xFFFFFFFF
+
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
+// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
+#define LLAMA_SUPPORTS_GPU_OFFLOAD
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    //
+    // C interface
+    //
+    // TODO: show sample usage
+    //
+
+    struct llama_model;
+    struct llama_context;
+
+    typedef int llama_token;
+
+    typedef struct llama_token_data {
+        llama_token id; // token id
+        float logit;    // log-odds of the token
+        float p;        // probability of the token
+    } llama_token_data;
+
+    typedef struct llama_token_data_array {
+        llama_token_data * data;
+        size_t size;
+        bool sorted;
+    } llama_token_data_array;
+
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
+
+   struct llama_context_params {
+        uint32_t seed;                         // RNG seed, -1 for random
+        int32_t  n_ctx;                        // text context
+        int32_t  n_batch;                      // prompt processing batch size
+        int32_t  n_gpu_layers;                 // number of layers to store in VRAM
+        int32_t  main_gpu;                     // the GPU that is used for scratch and small tensors
+        float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
+
+        // Keep the booleans together to avoid misalignment during copy-by-value.
+        bool low_vram;   // if true, reduce VRAM usage at the cost of performance
+        bool f16_kv;     // use fp16 for KV cache
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mmap;   // use mmap if possible
+        bool use_mlock;  // force system to keep model in RAM
+        bool embedding;  // embedding mode only
+    };
+    // model file types
+    enum llama_ftype {
+        LLAMA_FTYPE_ALL_F32              = 0,
+        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
+    };
+
+    // model quantization parameters
+    typedef struct llama_model_quantize_params {
+        int nthread;                 // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype   ftype;    // quantize to this llama_ftype
+        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor; // quantize output.weight
+    } llama_model_quantize_params;
+
+    // performance timing information
+    struct llama_timings {
+        double t_start_ms;
+        double t_end_ms;
+        double t_load_ms;
+        double t_sample_ms;
+        double t_p_eval_ms;
+        double t_eval_ms;
+
+        int32_t n_sample;
+        int32_t n_p_eval;
+        int32_t n_eval;
+    };
+
+    LLAMA_API struct llama_context_params llama_context_default_params();
+    LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
+
+    LLAMA_API bool llama_mmap_supported();
+    LLAMA_API bool llama_mlock_supported();
+
+    // TODO: not great API - very likely to change
+    // Initialize the llama + ggml backend
+    // If numa is true, use NUMA optimizations
+    // Call once at the start of the program
+    LLAMA_API void llama_backend_init(bool numa);
+    // Call once at the end of the program - currently only used for MPI
+    LLAMA_API void llama_backend_free();
+
+    LLAMA_API int64_t llama_time_us();
+
+    LLAMA_API struct llama_model * llama_load_model_from_file(
+                             const char * path_model,
+            struct llama_context_params   params);
+
+    LLAMA_API void llama_free_model(struct llama_model * model);
+
+    LLAMA_API struct llama_context * llama_new_context_with_model(
+                     struct llama_model * model,
+            struct llama_context_params   params);
+
+    // Various functions for loading a ggml llama model.
+    // Allocate (almost) all memory needed for the model.
+    // Return NULL on failure
+    LLAMA_API DEPRECATED(struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params),
+            "please use llama_load_model_from_file combined with llama_new_context_with_model instead");
+
+    // Frees all allocated memory
+    LLAMA_API void llama_free(struct llama_context * ctx);
+
+    // Returns 0 on success
+    LLAMA_API int llama_model_quantize(
+            const char * fname_inp,
+            const char * fname_out,
+            const llama_model_quantize_params * params);
+
+    // Apply a LoRA adapter to a loaded model
+    // path_base_model is the path to a higher quality model to use as a base for
+    // the layers modified by the adapter. Can be NULL to use the current loaded model.
+    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
+    // will be applied on top of the previous one
+    // Returns 0 on success
+    LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
+            struct llama_context * ctx,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads),
+            "please use llama_model_apply_lora_from_file instead");
+
+    LLAMA_API int llama_model_apply_lora_from_file(
+            const struct llama_model * model,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads);
+
+    // Returns the number of tokens in the KV cache
+    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
+
+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+
+    // Returns the maximum size in bytes of the state (rng, logits, embedding
+    // and kv_cache) - will often be smaller after compacting tokens
+    LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
+
+    // Copies the state to the specified destination address.
+    // Destination needs to have allocated enough memory.
+    // Returns the number of bytes copied
+    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
+
+    // Set the state reading from the specified address
+    // Returns the number of bytes read
+    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
+
+    // Save/load session file
+    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
+    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+
+    // Run the llama inference to obtain the logits and probabilities for the next token.
+    // tokens + n_tokens is the provided batch of new tokens to process
+    // n_past is the number of tokens to use from previous eval calls
+    // Returns 0 on success
+    LLAMA_API int llama_eval(
+            struct llama_context * ctx,
+               const llama_token * tokens,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads);
+
+    // Same as llama_eval, but use float matrix input directly.
+    LLAMA_API int llama_eval_embd(
+            struct llama_context * ctx,
+                     const float * embd,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads);
+
+    // Export a static computation graph for context of 511 and batch size of 1
+    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
+    //       parameters here to keep things simple
+    // IMPORTANT: do not use for anything else other than debugging and testing!
+    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+
+    // Convert the provided text into tokens.
+    // The tokens pointer must be large enough to hold the resulting tokens.
+    // Returns the number of tokens on success, no more than n_max_tokens
+    // Returns a negative number on failure - the number of tokens that would have been returned
+    // TODO: not sure if correct
+    LLAMA_API int llama_tokenize(
+            struct llama_context * ctx,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
+    LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (const struct llama_context * ctx);
+
+    // Get the vocabulary as output parameters.
+    // Returns number of results.
+    LLAMA_API int llama_get_vocab(
+            const struct llama_context * ctx,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
+    // Token logits obtained from the last call to llama_eval()
+    // The logits for the last token are stored in the last row
+    // Can be mutated in order to change the probabilities of the next token
+    // Rows: n_tokens
+    // Cols: n_vocab
+    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
+
+    // Get the embeddings for the input
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
+
+    // Token Id -> String. Uses the vocabulary in the provided context
+    LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
+
+    // Special tokens
+    LLAMA_API llama_token llama_token_bos();  // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos();  // end-of-sentence
+    LLAMA_API llama_token llama_token_nl();   // next-line
+
+    // Sampling functions
+
+    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
+    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
+
+    /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
+    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+
+    /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
+    /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
+    /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
+    /// @params smooth_factor Smooth factor between guidance logits and original logits. 1.0f means only use guidance logits. 0.0f means only original logits.
+    LLAMA_API void llama_sample_classifier_free_guidance(
+              struct llama_context * ctx,
+            llama_token_data_array * candidates,
+              struct llama_context * guidance_ctx,
+                             float   scale,
+                             float   smooth_factor);
+
+    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
+    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
+
+    /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+
+    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
+    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
+
+    /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
+    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+
+    /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
+    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
+    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
+    /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
+    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
+    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+
+    /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
+    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
+    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
+    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+
+    /// @details Selects the token with the highest probability.
+    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Randomly selects a token from the candidates based on their probabilities.
+    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    // Performance information
+    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
+    LLAMA_API void llama_print_timings(struct llama_context * ctx);
+    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
+
+    // Print system information
+    LLAMA_API const char * llama_print_system_info(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
+#ifdef LLAMA_API_INTERNAL
+
+#include <vector>
+#include <string>
+struct ggml_tensor;
+
+const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+
+#endif
+
+#endif // LLAMA_H

llama/llama_cublas.go

@@ -1,9 +0,0 @@
-//go:build cublas
-// +build cublas
-
-package llama
-
-/*
-#cgo LDFLAGS: -lcublas -lcudart -L/usr/local/cuda/lib64/
-*/
-import "C"

llama/llama_metal.go

@@ -1,2 +0,0 @@
-//go:build metal
-package llama

llama/llama_openblas.go

@@ -1,9 +0,0 @@
-//go:build openblas
-// +build openblas
-
-package llama
-
-/*
-#cgo LDFLAGS: -lopenblas
-*/
-import "C"

llama/options.go

@@ -1,98 +0,0 @@
-// MIT License
-
-// Copyright (c) 2023 go-skynet authors
-
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-
-// The above copyright notice and this permission notice shall be included in all
-// copies or substantial portions of the Software.
-
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-// SOFTWARE.
-
-package llama
-
-type ModelOptions struct {
-	ContextSize int
-	Seed        int
-	NBatch      int
-	F16Memory   bool
-	MLock       bool
-	MMap        bool
-	VocabOnly   bool
-	LowVRAM     bool
-	Embeddings  bool
-	NUMA        bool
-	NGPULayers  int
-	MainGPU     string
-	TensorSplit string
-}
-
-type PredictOptions struct {
-	Seed, Threads, Tokens, TopK, Repeat, Batch, NKeep int
-	TopP, Temperature, Penalty                        float64
-	F16KV                                             bool
-	DebugMode                                         bool
-	StopPrompts                                       []string
-	IgnoreEOS                                         bool
-
-	TailFreeSamplingZ float64
-	TypicalP          float64
-	FrequencyPenalty  float64
-	PresencePenalty   float64
-	Mirostat          int
-	MirostatETA       float64
-	MirostatTAU       float64
-	PenalizeNL        bool
-	LogitBias         string
-	TokenCallback     func(string) bool
-
-	MLock, MMap bool
-	MainGPU     string
-	TensorSplit string
-}
-
-type PredictOption func(p *PredictOptions)
-
-type ModelOption func(p *ModelOptions)
-
-var DefaultModelOptions ModelOptions = ModelOptions{
-	ContextSize: 512,
-	Seed:        0,
-	F16Memory:   false,
-	MLock:       false,
-	Embeddings:  false,
-	MMap:        true,
-	LowVRAM:     false,
-}
-
-var DefaultOptions PredictOptions = PredictOptions{
-	Seed:              -1,
-	Threads:           4,
-	Tokens:            128,
-	Penalty:           1.1,
-	Repeat:            64,
-	Batch:             512,
-	NKeep:             64,
-	TopK:              40,
-	TopP:              0.95,
-	TailFreeSamplingZ: 1.0,
-	TypicalP:          1.0,
-	Temperature:       0.8,
-	FrequencyPenalty:  0.0,
-	PresencePenalty:   0.0,
-	Mirostat:          0,
-	MirostatTAU:       5.0,
-	MirostatETA:       0.1,
-	MMap:              true,
-}

llama/utils.go

@@ -0,0 +1,104 @@
+package llama
+
+type node[T any] struct {
+	t    T
+	next *node[T]
+	prev *node[T]
+}
+
+type deque[T any] struct {
+	head     *node[T]
+	tail     *node[T]
+	size     int
+	capacity int
+}
+
+func (d *deque[T]) Empty() bool {
+	return d.size == 0
+}
+
+func (d *deque[T]) Len() int {
+	return d.size
+}
+
+func (d *deque[T]) Cap() int {
+	return d.capacity
+}
+
+func (d *deque[T]) Push(t T) {
+	if d.capacity > 0 && d.size >= d.capacity {
+		d.PopLeft()
+	}
+
+	n := node[T]{t: t}
+	if d.head != nil {
+		n.next = d.head
+		d.head.prev = &n
+		d.head = &n
+	} else {
+		d.head = &n
+		d.tail = &n
+	}
+
+	d.size++
+}
+
+func (d *deque[T]) PushLeft(t T) {
+	if d.capacity > 0 && d.size >= d.capacity {
+		d.Pop()
+	}
+
+	n := node[T]{t: t}
+	if d.tail != nil {
+		n.prev = d.tail
+		d.tail.next = &n
+		d.tail = &n
+	} else {
+		d.head = &n
+		d.tail = &n
+	}
+
+	d.size++
+}
+
+func (d *deque[T]) Pop() *T {
+	if d.Empty() {
+		return nil
+	}
+
+	head := d.head
+	d.head = head.next
+	if d.head != nil {
+		d.head.prev = nil
+	} else {
+		d.tail = nil
+	}
+
+	d.size--
+	return &head.t
+}
+
+func (d *deque[T]) PopLeft() *T {
+	if d.Empty() {
+		return nil
+	}
+
+	tail := d.tail
+	d.tail = tail.prev
+	if d.tail != nil {
+		d.tail.next = nil
+	} else {
+		d.head = nil
+	}
+
+	d.size--
+	return &tail.t
+}
+
+func (d *deque[T]) Data() (data []T) {
+	for n := d.head; n != nil; n = n.next {
+		data = append(data, n.t)
+	}
+
+	return data
+}

server/routes.go

@@ -11,12 +11,12 @@ import (
 	"net/http"
 	"os"
 	"path"
-	"runtime"
 	"strings"
 	"text/template"
 
 	"github.com/gin-gonic/gin"
 	"github.com/lithammer/fuzzysearch/fuzzy"
+	"golang.org/x/sync/errgroup"
 
 	"github.com/jmorganca/ollama/api"
 	"github.com/jmorganca/ollama/llama"
@@ -36,14 +36,10 @@ func cacheDir() string {
 }
 
 func generate(c *gin.Context) {
-	var req api.GenerateRequest
+	req := api.GenerateRequest{
-	if req.ModelOptions == nil {
+		Options: api.DefaultOptions(),
-		req.ModelOptions = &api.DefaultModelOptions
 	}
 
-	if req.PredictOptions == nil {
-		req.PredictOptions = &api.DefaultPredictOptions
-	}
 	if err := c.ShouldBindJSON(&req); err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
@@ -60,15 +56,12 @@ func generate(c *gin.Context) {
 		req.Model = path.Join(cacheDir(), "models", req.Model+".bin")
 	}
 
-	modelOpts := getModelOpts(req)
+	llm, err := llama.New(req.Model, req.Options)
-	modelOpts.NGPULayers = 1 // hard-code this for now
-
-	model, err := llama.New(req.Model, modelOpts)
 	if err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
-	defer model.Free()
+	defer llm.Close()
 
 	templateNames := make([]string, 0, len(templates.Templates()))
 	for _, template := range templates.Templates() {
@@ -87,43 +80,41 @@ func generate(c *gin.Context) {
 	}
 
 	ch := make(chan string)
-	model.SetTokenCallback(func(token string) bool {
+	g, _ := errgroup.WithContext(c.Request.Context())
-		ch <- token
+	g.Go(func() error {
-		return true
-	})
-
-	predictOpts := getPredictOpts(req)
-
-	go func() {
 		defer close(ch)
-		_, err := model.Predict(req.Prompt, predictOpts)
+		return llm.Predict(req.Prompt, func(s string) {
-		if err != nil {
+			ch <- s
-			panic(err)
+		})
-		}
-	}()
-
-	c.Stream(func(w io.Writer) bool {
-		token, ok := <-ch
-		if !ok {
-			return false
-		}
-
-		resp := api.GenerateResponse{
-			Response: token,
-		}
-
-		bts, err := json.Marshal(resp)
-		if err != nil {
-			return false
-		}
-
-		bts = append(bts, '\n')
-		if _, err := w.Write(bts); err != nil {
-			return false
-		}
-
-		return true
 	})
+
+	g.Go(func() error {
+		c.Stream(func(w io.Writer) bool {
+			s, ok := <-ch
+			if !ok {
+				return false
+			}
+
+			bts, err := json.Marshal(api.GenerateResponse{Response: s})
+			if err != nil {
+				return false
+			}
+
+			bts = append(bts, '\n')
+			if _, err := w.Write(bts); err != nil {
+				return false
+			}
+
+			return true
+		})
+
+		return nil
+	})
+
+	if err := g.Wait(); err != nil && !errors.Is(err, io.EOF) {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
 }
 
 func Serve(ln net.Listener) error {
@@ -195,53 +186,3 @@ func matchRankOne(source string, targets []string) (bestMatch string, bestRank i
 
 	return
 }
-
-func getModelOpts(req api.GenerateRequest) llama.ModelOptions {
-	var opts llama.ModelOptions
-	opts.ContextSize = req.ModelOptions.ContextSize
-	opts.Seed = req.ModelOptions.Seed
-	opts.F16Memory = req.ModelOptions.F16Memory
-	opts.MLock = req.ModelOptions.MLock
-	opts.Embeddings = req.ModelOptions.Embeddings
-	opts.MMap = req.ModelOptions.MMap
-	opts.LowVRAM = req.ModelOptions.LowVRAM
-
-	opts.NBatch = req.ModelOptions.NBatch
-	opts.VocabOnly = req.ModelOptions.VocabOnly
-	opts.NUMA = req.ModelOptions.NUMA
-	opts.NGPULayers = req.ModelOptions.NGPULayers
-	opts.MainGPU = req.ModelOptions.MainGPU
-	opts.TensorSplit = req.ModelOptions.TensorSplit
-
-	return opts
-}
-
-func getPredictOpts(req api.GenerateRequest) llama.PredictOptions {
-	var opts llama.PredictOptions
-
-	if req.PredictOptions.Threads == -1 {
-		opts.Threads = runtime.NumCPU()
-	} else {
-		opts.Threads = req.PredictOptions.Threads
-	}
-
-	opts.Seed = req.PredictOptions.Seed
-	opts.Tokens = req.PredictOptions.Tokens
-	opts.Penalty = req.PredictOptions.Penalty
-	opts.Repeat = req.PredictOptions.Repeat
-	opts.Batch = req.PredictOptions.Batch
-	opts.NKeep = req.PredictOptions.NKeep
-	opts.TopK = req.PredictOptions.TopK
-	opts.TopP = req.PredictOptions.TopP
-	opts.TailFreeSamplingZ = req.PredictOptions.TailFreeSamplingZ
-	opts.TypicalP = req.PredictOptions.TypicalP
-	opts.Temperature = req.PredictOptions.Temperature
-	opts.FrequencyPenalty = req.PredictOptions.FrequencyPenalty
-	opts.PresencePenalty = req.PredictOptions.PresencePenalty
-	opts.Mirostat = req.PredictOptions.Mirostat
-	opts.MirostatTAU = req.PredictOptions.MirostatTAU
-	opts.MirostatETA = req.PredictOptions.MirostatETA
-	opts.MMap = req.PredictOptions.MMap
-
-	return opts
-}