Add support for full CUDA GPU offloading (#105)

Signed-off-by: mudler <mudler@mocaccino.org>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

.github/workflows/test.yaml

@@ -27,14 +27,9 @@ jobs:
         with: 
           submodules: true
 
-      - name: Dependencies
-        run: |
-          brew update
-          brew install sdl2
-
       - name: Test
         run: |
-          make test
+          CMAKE_ARGS="-DLLAMA_F16C=OFF -DLLAMA_AVX512=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF" make test
 
   macOS-metal-latest:
     runs-on: macOS-latest
@@ -45,12 +40,7 @@ jobs:
         with: 
           submodules: true
 
-      - name: Dependencies
-        run: |
-          brew update
-          brew install sdl2
-
       - name: Test
         run: |
-          make BUILD_TYPE=metal test
+          CMAKE_ARGS="-DLLAMA_F16C=OFF -DLLAMA_AVX512=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF" make BUILD_TYPE=metal test
           CGO_LDFLAGS="-framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders" LIBRARY_PATH=$PWD C_INCLUDE_PATH=$PWD go build -o testbuild ./examples

1902.patch

@@ -0,0 +1,120 @@
+From 200892a3a54323eb65ca9c8d8afb6043ca2d8944 Mon Sep 17 00:00:00 2001
+From: mudler <mudler@mocaccino.org>
+Date: Fri, 16 Jun 2023 23:43:36 +0200
+Subject: [PATCH] Pass pointer to params in llama_init_from_file
+
+Especially with golang bindings, calling by value has the side-effect of
+values not being copied correctly. This has been observed with the
+bindings in https://github.com/go-skynet/go-llama.cpp/pull/105.
+---
+ examples/common.cpp                                          | 2 +-
+ examples/quantize-stats/quantize-stats.cpp                   | 2 +-
+ examples/save-load-state/save-load-state.cpp                 | 4 ++--
+ examples/train-text-from-scratch/train-text-from-scratch.cpp | 2 +-
+ llama.cpp                                                    | 3 ++-
+ llama.h                                                      | 2 +-
+ tests/test-tokenizer-0.cpp                                   | 2 +-
+ 7 files changed, 9 insertions(+), 8 deletions(-)
+
+diff --git a/examples/common.cpp b/examples/common.cpp
+index 055383beff9..7cf48e82158 100644
+--- a/examples/common.cpp
++++ b/examples/common.cpp
+@@ -555,7 +555,7 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
+     lparams.logits_all   = params.perplexity;
+     lparams.embedding    = params.embedding;
+
+-    llama_context * lctx = llama_init_from_file(params.model.c_str(), lparams);
++    llama_context * lctx = llama_init_from_file(params.model.c_str(), &lparams);
+
+     if (lctx == NULL) {
+         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp
+index 6b8018ee284..a7c1e873a92 100644
+--- a/examples/quantize-stats/quantize-stats.cpp
++++ b/examples/quantize-stats/quantize-stats.cpp
+@@ -330,7 +330,7 @@ int main(int argc, char ** argv) {
+         lparams.f16_kv     = false;
+         lparams.use_mlock  = false;
+
+-        ctx = llama_init_from_file(params.model.c_str(), lparams);
++        ctx = llama_init_from_file(params.model.c_str(), &lparams);
+
+         if (ctx == NULL) {
+             fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp
+index da4d37ad03d..07ee6750d4c 100644
+--- a/examples/save-load-state/save-load-state.cpp
++++ b/examples/save-load-state/save-load-state.cpp
+@@ -35,7 +35,7 @@ int main(int argc, char ** argv) {
+     auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);
+
+     // init
+-    auto ctx = llama_init_from_file(params.model.c_str(), lparams);
++    auto ctx = llama_init_from_file(params.model.c_str(), &lparams);
+     auto tokens = std::vector<llama_token>(params.n_ctx);
+     auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), int(tokens.size()), true);
+
+@@ -95,7 +95,7 @@ int main(int argc, char ** argv) {
+     llama_free(ctx);
+
+     // load new model
+-    auto ctx2 = llama_init_from_file(params.model.c_str(), lparams);
++    auto ctx2 = llama_init_from_file(params.model.c_str(), &lparams);
+
+     // Load state (rng, logits, embedding and kv_cache) from file
+     {
+diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp
+index 7ec85951adc..1c7a06c21be 100644
+--- a/examples/train-text-from-scratch/train-text-from-scratch.cpp
++++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp
+@@ -3054,7 +3054,7 @@ int main(int argc, char ** argv) {
+     struct llama_context_params llama_params = llama_context_default_params();
+     llama_params.vocab_only = true;
+
+-    struct llama_context * lctx = llama_init_from_file(params.fn_vocab_model, llama_params);
++    struct llama_context * lctx = llama_init_from_file(params.fn_vocab_model, &llama_params);
+
+     struct llama_vocab vocab;
+     {
+diff --git a/llama.cpp b/llama.cpp
+index 81f047ed298..0629e873886 100644
+--- a/llama.cpp
++++ b/llama.cpp
+@@ -2618,8 +2618,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
+
+ struct llama_context * llama_init_from_file(
+                              const char * path_model,
+-            struct llama_context_params   params) {
++         const struct llama_context_params *  params_ptr) {
+     ggml_time_init();
++    struct llama_context_params params = *params_ptr;
+
+     llama_context * ctx = new llama_context;
+
+diff --git a/llama.h b/llama.h
+index 1241ba6c0ec..faf2675f125 100644
+--- a/llama.h
++++ b/llama.h
+@@ -142,7 +142,7 @@ extern "C" {
+     // Return NULL on failure
+     LLAMA_API struct llama_context * llama_init_from_file(
+                              const char * path_model,
+-            struct llama_context_params   params);
++            const struct llama_context_params * params);
+
+     // Frees all allocated memory
+     LLAMA_API void llama_free(struct llama_context * ctx);
+diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp
+index ab1538a0cf3..b405df8e687 100644
+--- a/tests/test-tokenizer-0.cpp
++++ b/tests/test-tokenizer-0.cpp
+@@ -36,7 +36,7 @@ int main(int argc, char **argv) {
+
+         lparams.vocab_only = true;
+
+-        ctx = llama_init_from_file(fname.c_str(), lparams);
++        ctx = llama_init_from_file(fname.c_str(), &lparams);
+
+         if (ctx == NULL) {
+             fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());

Makefile

@@ -176,7 +176,7 @@ $(info )
 # Use this if you want to set the default behavior
 
 
-llama.cpp/ggml.o:
+llama.cpp/ggml.o: prepare
 	mkdir -p build
 	cd build && cmake ../llama.cpp $(CMAKE_ARGS) && VERBOSE=1 cmake --build . --config Release && cp -rf CMakeFiles/ggml.dir/ggml.c.o ../llama.cpp/ggml.o
 
@@ -193,16 +193,22 @@ llama.cpp/k_quants.o: llama.cpp/ggml.o
 	cd build && cp -rf CMakeFiles/ggml.dir/k_quants.c.o ../llama.cpp/k_quants.o
 
 llama.cpp/llama.o:
-	cd build && make llama.o && cp -rf CMakeFiles/llama.dir/llama.cpp.o ../llama.cpp/llama.o
+	cd build && cp -rf CMakeFiles/llama.dir/llama.cpp.o ../llama.cpp/llama.o
 
 llama.cpp/common.o:
-	cd build && make common && cp -rf examples/CMakeFiles/common.dir/common.cpp.o ../llama.cpp/common.o
+	cd build && cp -rf examples/CMakeFiles/common.dir/common.cpp.o ../llama.cpp/common.o
 
-binding.o: llama.cpp/ggml.o llama.cpp/llama.o llama.cpp/common.o
+binding.o: prepare llama.cpp/ggml.o llama.cpp/llama.o llama.cpp/common.o
 	$(CXX) $(CXXFLAGS) -I./llama.cpp -I./llama.cpp/examples binding.cpp -o binding.o -c $(LDFLAGS)
 
+## https://github.com/ggerganov/llama.cpp/pull/1902
+prepare:
+	cd llama.cpp && patch -p1 < ../1902.patch
+	touch $@
+
 libbinding.a: binding.o llama.cpp/k_quants.o $(EXTRA_TARGETS)
 	ar src libbinding.a llama.cpp/ggml.o llama.cpp/k_quants.o $(EXTRA_TARGETS) llama.cpp/common.o llama.cpp/llama.o binding.o
+
 clean:
 	rm -rf *.o
 	rm -rf *.a

binding.cpp

@@ -1,5 +1,6 @@
 #include "common.h"
 #include "llama.h"
+
 #include "binding.h"
 
 #include <cassert>
@@ -125,7 +126,7 @@ int llama_predict(void* params_ptr, void* state_pr, char* result, bool debug) {
 
     std::mt19937 rng(params.seed);
 
-    llama_init_backend();
+
 
     std::string path_session = params.path_prompt_cache;
     std::vector<llama_token> session_tokens;
@@ -590,7 +591,7 @@ void* llama_allocate_params(const char *prompt, int seed, int threads, int token
 }
 
 
-void* load_model(const char *fname, int n_ctx, int n_seed, bool memory_f16, bool mlock, bool embeddings, bool mmap, int n_gpu_layers, int n_batch, const char *maingpu, const char *tensorsplit) {
+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) {
     // load the model
     auto lparams = llama_context_default_params();
 
@@ -601,6 +602,8 @@ void* load_model(const char *fname, int n_ctx, int n_seed, bool memory_f16, bool
     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);
@@ -625,13 +628,14 @@ void* load_model(const char *fname, int n_ctx, int n_seed, bool memory_f16, bool
 
     lparams.n_batch      = n_batch;
 
+    llama_init_backend();
     void* res = nullptr;
     try {
-        res = llama_init_from_file(fname, lparams);
+        res = llama_init_from_file(fname, &lparams);
     } catch(std::runtime_error& e) {   
         fprintf(stderr, "failed %s",e.what());
         return res;
     }
 
     return res;
-}
+}

binding.h

@@ -14,7 +14,7 @@ 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, int n_gpu, int n_batch, const char *maingpu, const char *tensorsplit);
+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);
 
 int get_embeddings(void* params_ptr, void* state_pr, float * res_embeddings);
 

llama.go

@@ -23,7 +23,7 @@ type LLama struct {
 func New(model string, opts ...ModelOption) (*LLama, error) {
 	mo := NewModelOptions(opts...)
 	modelPath := C.CString(model)
-	result := 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.int(mo.NGPULayers), C.int(mo.NBatch), C.CString(mo.MainGPU), C.CString(mo.TensorSplit))
+	result := 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))
 	if result == nil {
 		return nil, fmt.Errorf("failed loading model")
 	}

options.go

@@ -7,6 +7,8 @@ type ModelOptions struct {
 	F16Memory   bool
 	MLock       bool
 	MMap        bool
+	VocabOnly   bool
+	LowVRAM     bool
 	Embeddings  bool
 	NGPULayers  int
 	MainGPU     string
@@ -50,6 +52,7 @@ var DefaultModelOptions ModelOptions = ModelOptions{
 	MLock:       false,
 	Embeddings:  false,
 	MMap:        true,
+	LowVRAM:     false,
 }
 
 var DefaultOptions PredictOptions = PredictOptions{
@@ -58,7 +61,7 @@ var DefaultOptions PredictOptions = PredictOptions{
 	Tokens:            128,
 	Penalty:           1.1,
 	Repeat:            64,
-	Batch:             8,
+	Batch:             512,
 	NKeep:             64,
 	TopK:              40,
 	TopP:              0.95,
@@ -128,6 +131,14 @@ func SetPredictionMainGPU(maingpu string) PredictOption {
 	}
 }
 
+var VocabOnly ModelOption = func(p *ModelOptions) {
+	p.VocabOnly = true
+}
+
+var EnabelLowVRAM ModelOption = func(p *ModelOptions) {
+	p.LowVRAM = true
+}
+
 var EnableEmbeddings ModelOption = func(p *ModelOptions) {
 	p.Embeddings = true
 }