llama_cpp 0.14.2 → 0.14.4

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: d720138462a39f3fd9853befa19e55543a794eb4d1c379f7d9db516a4569db68
-  data.tar.gz: 9779852b62cf57ab208275b746ca2492921cf1f92d4ebf6be26a668f1a7bbb66
+  metadata.gz: 92ebd411f54255b05074ef79ed3e220c9ff4332164cfc831122d766226322515
+  data.tar.gz: 571f0ec65776945d40a54e31bba26cc4194b19965cc4841ce40b9ad3d94605df
 SHA512:
-  metadata.gz: 7fa80468abc917099b58009a7a821c704989b8086026e92f8e71a1310ea7ec0449276aeb4653bdb4ddf499183c785b0513ab75b1fa6a94659fe15be2cf05190c
-  data.tar.gz: '091155784ead62d3ef4ec68f3b4c9f6b1fc97ef87db45327266712912501746b08df983e0d0b81b518d229d4d31f1a0d77ad36f2d7156c26141b8116049d3206'
+  metadata.gz: 34ca9567b7eb96add562e977f22f8b2be087c026c85bf92cd5e31f9a96ea5f02a841bdf05f745c4079740a4bb01476fb9bab313317d66dbf8870fa829a269c86
+  data.tar.gz: 64b19ef010bb52800cd3710c1ec70bcb7b747e53b6ea7d8f13d84d336b1ee67868153f5bbdfee0b4131dfddaf1c656c49bd774084ee2d14f191d22d215a47737

data/CHANGELOG.md

@@ -1,3 +1,17 @@
+## [[0.14.4](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.3...v0.14.4)] - 2024-04-06
+
+- Bump llama.cpp from b2496 to b2573.
+  - Add file type constants.
+- Bump llama.cpp from b2573 to b2608.
+
+Implementation of llama_split_path, llama_split_prefix binding, llama_grammar_accept, and decode_utf8 has been skipped.
+
+## [[0.14.3](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.2...v0.14.3)] - 2024-03-23
+
+- Bump llama.cpp from b2435 to b2496.
+  - Add `n_layer` method to `Model`.
+  - Add `apply_control_vector` method to `Context`.
+
 ## [[0.14.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.1...v0.14.2)] - 2024-03-16
 
 - Fix to use metal embed library on macOS.

data/ext/llama_cpp/llama_cpp.cpp

@@ -1466,6 +1466,7 @@ public:
     rb_define_method(rb_cLLaMAModel, "n_vocab", RUBY_METHOD_FUNC(_llama_model_get_model_n_vocab), 0);
     rb_define_method(rb_cLLaMAModel, "n_ctx_train", RUBY_METHOD_FUNC(_llama_model_get_model_n_ctx_train), 0);
     rb_define_method(rb_cLLaMAModel, "n_embd", RUBY_METHOD_FUNC(_llama_model_get_model_n_embd), 0);
+    rb_define_method(rb_cLLaMAModel, "n_layer", RUBY_METHOD_FUNC(_llama_model_get_model_n_layer), 0);
     rb_define_method(rb_cLLaMAModel, "rope_freq_scale_train", RUBY_METHOD_FUNC(_llama_model_rope_freq_scale_train), 0);
     rb_define_method(rb_cLLaMAModel, "token_to_piece", RUBY_METHOD_FUNC(_llama_model_token_to_piece), 1);
     rb_define_method(rb_cLLaMAModel, "tokenize", RUBY_METHOD_FUNC(_llama_model_tokenize), -1);
@@ -1613,6 +1614,11 @@ private:
     return INT2NUM(llama_n_embd(ptr->model));
   }
 
+  static VALUE _llama_model_get_model_n_layer(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_n_layer(ptr->model));
+  }
+
   static VALUE _llama_model_rope_freq_scale_train(VALUE self) {
     LLaMAModelWrapper* ptr = get_llama_model(self);
     return DBL2NUM(llama_rope_freq_scale_train(ptr->model));
@@ -2083,6 +2089,7 @@ public:
     rb_define_method(rb_cLLaMAContext, "sample_token", RUBY_METHOD_FUNC(_llama_context_sample_token), 1);
     rb_define_method(rb_cLLaMAContext, "sample_grammar", RUBY_METHOD_FUNC(_llama_context_sample_grammar), -1);
     rb_define_method(rb_cLLaMAContext, "grammar_accept_token", RUBY_METHOD_FUNC(_llama_context_grammar_accept_token), -1);
+    rb_define_method(rb_cLLaMAContext, "apply_control_vector", RUBY_METHOD_FUNC(_llama_context_apply_control_vector), -1);
   }
 
 private:
@@ -3153,6 +3160,59 @@ private:
 
     return Qnil;
   }
+
+  static VALUE _llama_context_apply_control_vector(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[4] = { rb_intern("data"), rb_intern("n_embd"), rb_intern("il_start"), rb_intern("il_end") };
+    VALUE kw_values[4] = { Qundef, Qundef, Qundef, Qundef };
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 4, 0, kw_values);
+
+    if (!RB_TYPE_P(kw_values[0], T_ARRAY) && !NIL_P(kw_values[0])) {
+      rb_raise(rb_eArgError, "data must be an Array or nil");
+      return Qnil;
+    }
+    if (!RB_INTEGER_TYPE_P(kw_values[1])) {
+      rb_raise(rb_eArgError, "n_embd must be an Integer");
+      return Qnil;
+    }
+    if (!RB_INTEGER_TYPE_P(kw_values[2])) {
+      rb_raise(rb_eArgError, "il_start must be an Integer");
+      return Qnil;
+    }
+    if (!RB_INTEGER_TYPE_P(kw_values[3])) {
+      rb_raise(rb_eArgError, "il_end must be an Integer");
+      return Qnil;
+    }
+
+    LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
+    if (ctx_ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+
+    std::vector<float> data(RARRAY_LEN(kw_values[0]));
+    for (size_t i = 0; i < data.size(); i++) {
+      data[i] = NUM2DBL(rb_ary_entry(kw_values[0], i));
+    }
+    const int32_t n_embd = NUM2INT(kw_values[1]);
+    const int32_t il_start = NUM2INT(kw_values[2]);
+    const int32_t il_end = NUM2INT(kw_values[3]);
+
+    int32_t err = 0;
+    if (NIL_P(kw_values[0])) {
+      err = llama_control_vector_apply(ctx_ptr->ctx, NULL, 0, n_embd, il_start, il_end);
+    } else {
+      err = llama_control_vector_apply(ctx_ptr->ctx, data.data(), data.size(), n_embd, il_start, il_end);
+    }
+
+    if (err) {
+      rb_raise(rb_eRuntimeError, "Failed to apply control vector");
+      return Qnil;
+    }
+
+    return Qnil;
+  }
 };
 
 const rb_data_type_t RbLLaMAContext::llama_context_type = {
@@ -3311,6 +3371,10 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ3_XXS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ3_XXS));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ1_S", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ1_S));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ4_NL", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ4_NL));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ3_S", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ3_S));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ3_M", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ3_M));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ4_XS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ4_XS));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ1_M", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ1_M));
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_GUESSED", INT2NUM(LLAMA_FTYPE_GUESSED));
 

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.14.2'
+  VERSION = '0.14.4'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b2435'
+  LLAMA_CPP_VERSION = 'b2608'
 end

data/sig/llama_cpp.rbs

@@ -32,6 +32,10 @@ module LLaMACpp
   LLAMA_FTYPE_MOSTLY_IQ3_XXS: Integer
   LLAMA_FTYPE_MOSTLY_IQ1_S: Integer
   LLAMA_FTYPE_MOSTLY_IQ4_NL: Integer
+  LLAMA_FTYPE_MOSTLY_IQ3_S: Integer
+  LLAMA_FTYPE_MOSTLY_IQ3_M: Integer
+  LLAMA_FTYPE_MOSTLY_IQ4_XS: Integer
+  LLAMA_FTYPE_MOSTLY_IQ1_M: Integer
 
   LLAMA_KV_OVERRIDE_TYPE_INT: Integer
   LLAMA_KV_OVERRIDE_TYPE_FLOAT: Integer
@@ -108,6 +112,7 @@ module LLaMACpp
     def n_vocab: () -> Integer
     def n_ctx_train: () -> Integer
     def n_embd: () -> Integer
+    def n_layer: () -> Integer
     def rope_freq_scale_train: () -> Float
     def token_to_piece: (Integer) -> String
     def tokenize: (text: String, ?n_max_tokens: Integer, ?add_bos: bool, ?special: bool) -> Array[Integer]
@@ -244,6 +249,7 @@ module LLaMACpp
     def sample_token: (::LLaMACpp::TokenDataArray) -> Integer
     def sample_grammar: (::LLaMACpp::TokenDataArray, grammar: ::LLaMACpp::Grammar) -> void
     def grammar_accept_token: (grammar: ::LLaMACpp::Grammar, token: Integer) -> void
+    def apply_control_vector: (data: Array[Float], n_embd: Integer, il_start: Integer, il_end: Integer) -> void
   end
 
   class ContextParams

data/vendor/tmp/llama.cpp/Makefile

@@ -1,15 +1,16 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
+	simple batched batched-bench save-load-state server gguf gguf-split llama-bench libllava.a llava-cli baby-llama beam-search  \
+	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
 	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease
+	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease                                 \
+	tests/test-json-schema-to-grammar
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -391,14 +392,20 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
+# LLAMA_CUBLAS is deprecated and will be removed in the future
+	LLAMA_CUDA := 1
+endif
+
+ifdef LLAMA_CUDA
 	ifneq ('', '$(wildcard /opt/cuda)')
 		CUDA_PATH ?= /opt/cuda
 	else
 		CUDA_PATH ?= /usr/local/cuda
 	endif
-	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
 	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
+	OBJS         += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
 	MK_NVCCFLAGS += -use_fast_math
 ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
@@ -453,19 +460,30 @@ ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 else
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-#ifdef LLAMA_CUDA_CUBLAS
-#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
-#endif # LLAMA_CUDA_CUBLAS
+ifdef LLAMA_CUDA_NO_PEER_COPY
+	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
+endif # LLAMA_CUDA_NO_PEER_COPY
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.h
+
 ifdef JETSON_EOL_MODULE_DETECT
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+define NVCC_COMPILE
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endef # NVCC_COMPILE
 else
+define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endef # NVCC_COMPILE
 endif # JETSON_EOL_MODULE_DETECT
-endif # LLAMA_CUBLAS
+
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+	$(NVCC_COMPILE)
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+	$(NVCC_COMPILE)
+
+endif # LLAMA_CUDA
 
 ifdef LLAMA_CLBLAST
 
@@ -511,7 +529,6 @@ ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
 endif # LLAMA_VULKAN
 
 ifdef LLAMA_HIPBLAS
-
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH	?= /usr
 		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
@@ -523,7 +540,7 @@ ifdef LLAMA_HIPBLAS
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
-	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
 ifdef LLAMA_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
 endif # LLAMA_HIP_UMA
@@ -536,9 +553,18 @@ endif # LLAMA_HIP_UMA
 ifdef LLAMA_CUDA_FORCE_DMMV
 	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
+ifdef LLAMA_CUDA_NO_PEER_COPY
+	HIPFLAGS 	+= -DGGML_CUDA_NO_PEER_COPY
+endif # LLAMA_CUDA_NO_PEER_COPY
 	OBJS        += ggml-cuda.o
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	OBJS        += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+
 endif # LLAMA_HIPBLAS
 
 ifdef LLAMA_METAL
@@ -555,7 +581,7 @@ endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
-ggml-metal.o: ggml-metal.m ggml-metal.h
+ggml-metal.o: ggml-metal.m ggml-metal.h ggml.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
 ifdef LLAMA_METAL_EMBED_LIBRARY
@@ -591,12 +617,17 @@ override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
 # identify CUDA host compiler
-ifdef LLAMA_CUBLAS
+ifdef LLAMA_CUDA
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
 CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
 endif
 
+ifdef LLAMA_CURL
+override CXXFLAGS := $(CXXFLAGS) -DLLAMA_USE_CURL
+override LDFLAGS  := $(LDFLAGS) -lcurl
+endif
+
 #
 # Print build information
 #
@@ -611,7 +642,7 @@ $(info I NVCCFLAGS: $(NVCCFLAGS))
 $(info I LDFLAGS:   $(LDFLAGS))
 $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
-ifdef LLAMA_CUBLAS
+ifdef LLAMA_CUDA
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
 CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
@@ -621,8 +652,15 @@ $(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be exp
 endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
-endif # LLAMA_CUBLAS
+endif # LLAMA_CUDA
+$(info )
+
+ifdef LLAMA_CUBLAS
+$(info !!!!)
+$(info LLAMA_CUBLAS is deprecated and will be removed in the future. Use LLAMA_CUDA instead.)
+$(info !!!!)
 $(info )
+endif
 
 #
 # Build library
@@ -643,7 +681,10 @@ ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
 unicode.o: unicode.cpp unicode.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o
+unicode-data.o: unicode-data.cpp unicode-data.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
 
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -663,9 +704,15 @@ console.o: common/console.cpp common/console.h
 grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+json-schema-to-grammar.o: common/json-schema-to-grammar.cpp common/json-schema-to-grammar.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 train.o: common/train.cpp common/train.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+ngram-cache.o: common/ngram-cache.cpp common/ngram-cache.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
@@ -677,7 +724,8 @@ lib: llama.o ggml.o $(OBJS)
 	ar rcs libllama.a $^
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.a *.dll *.dylib benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.a *.dll *.dylib benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf ggml-cuda/*.o
 
 #
 # Examples
@@ -745,7 +793,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp json-schema-to-grammar.o common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
@@ -753,6 +801,10 @@ gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+gguf-split: examples/gguf-split/gguf-split.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -790,6 +842,10 @@ export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+retrieval: examples/retrieval/retrieval.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -802,14 +858,24 @@ lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-create.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) -o lookup-create $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-merge.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) -o lookup-merge $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-stats.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) -o lookup-stats $(LDFLAGS)
 
 passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
@@ -861,6 +927,10 @@ tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp json-schema-to-grammar.o ggml.o llama.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

data/vendor/tmp/llama.cpp/ggml-alloc.c

@@ -548,7 +548,11 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
 
-        if (ggml_is_view(node)) {
+        // TODO: better way to add external dependencies
+        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
             struct ggml_tensor * view_src = node->view_src;
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
@@ -565,8 +569,8 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
 
             ggml_gallocr_hash_get(galloc, src)->n_children += 1;
 
-            // allocate explicit inputs and leafs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
+            // allocate explicit inputs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
                 ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
             }
         }
@@ -701,8 +705,13 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         struct ggml_tensor * leaf = graph->leafs[i];
         struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
         galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
-        galloc->leaf_allocs[i].leaf.offset = hn->offset;
-        galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        if (leaf->view_src || leaf->data) {
+            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
+            galloc->leaf_allocs[i].leaf.size_max = 0;
+        } else {
+            galloc->leaf_allocs[i].leaf.offset = hn->offset;
+            galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        }
     }
 
     // reallocate buffers if needed

data/vendor/tmp/llama.cpp/ggml-backend-impl.h

@@ -103,6 +103,11 @@ extern "C" {
         // check if the backend supports an operation
         bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
 
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
         // (optional) event synchronization
         ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
         void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);