llama_cpp 0.14.4 → 0.14.5

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 92ebd411f54255b05074ef79ed3e220c9ff4332164cfc831122d766226322515
-  data.tar.gz: 571f0ec65776945d40a54e31bba26cc4194b19965cc4841ce40b9ad3d94605df
+  metadata.gz: 7d80abb57b135ff04718e34099accaaabf3358553b0f061d79b195a99386739d
+  data.tar.gz: 5b24a9b7846b962f4063a0e50f15c6d9a9c874d1931ed32c200f3383869a2fd9
 SHA512:
-  metadata.gz: 34ca9567b7eb96add562e977f22f8b2be087c026c85bf92cd5e31f9a96ea5f02a841bdf05f745c4079740a4bb01476fb9bab313317d66dbf8870fa829a269c86
-  data.tar.gz: 64b19ef010bb52800cd3710c1ec70bcb7b747e53b6ea7d8f13d84d336b1ee67868153f5bbdfee0b4131dfddaf1c656c49bd774084ee2d14f191d22d215a47737
+  metadata.gz: dfb20e108a57b65ff624db1e2ee37034ffca406d906268d89ff441099a02c00fd67743a786a0353df2368614003604a4bf5982089024f14aee2e0f95e210e297
+  data.tar.gz: 0a0bbd93dfe57e033f25e5c3e3d61fb568362aa2d317851dbb69fe620e5e30bc8b08c27272579e7841c50b87984abf70ade4a9e7e34fb2615e106a5c2474b79e

data/CHANGELOG.md

@@ -1,10 +1,18 @@
+## [[0.14.5](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.4...v0.14.5)] - 2024-04-13
+
+- Bump llama.cpp from b2608 to b2658.
+  - Add magic number constants.
+  - Add `token_cls` and `token_sep` methods to `Model`.
+
+Implementation bindings for llama_state_get_size, llama_state_get_data, llama_state_set_data, llama_state_load_file, llama_state_save_file, llama_state_seq_get_size, llama_state_seq_get_data, llama_state_seq_set_data, llama_state_seq_save_file, and llama_state_seq_load_file has been skipped.
+
 ## [[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.
+Implementation bindings for 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
 

data/examples/chat.rb

@@ -127,8 +127,7 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
       end
 
       if input_echo
-        output = []
-        embd.each { |token| output << context.model.token_to_piece(token) }
+        output = embd.map { |token| context.model.token_to_piece(token) }
         output_str = output.join
         output_str.chomp!(antiprompt) if first_input
         print(output_str)
@@ -136,8 +135,7 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
 
       if embd_input.size <= n_consumed
         if antiprompt.size.positive?
-          last_output = []
-          last_n_tokens.each { |token| last_output << context.model.token_to_piece(token) }
+          last_output = last_n_tokens.map { |token| context.model.token_to_piece(token) }
           last_output_str = last_output.join
 
           search_start_pos = last_output_str.size > antiprompt.size ? last_output_str.size - antiprompt.size : 0

data/ext/llama_cpp/extconf.rb

@@ -37,6 +37,7 @@ if RUBY_PLATFORM.match?(/darwin/)
     abort('Failed to set installation path for libllama.dylib.') unless mkstatus.success?
   end
   FileUtils.cp("#{LLAMA_CPP_DIR}/ggml-metal-embed.metal", VENDOR_LIB_DIR)
+  FileUtils.cp("#{LLAMA_CPP_DIR}/ggml-metal.metal", VENDOR_LIB_DIR)
 end
 
 abort('libstdc++ is not found.') unless have_library('stdc++')

data/ext/llama_cpp/llama_cpp.cpp

@@ -1478,6 +1478,8 @@ public:
     rb_define_method(rb_cLLaMAModel, "type", RUBY_METHOD_FUNC(_llama_model_get_type), 1);
     rb_define_method(rb_cLLaMAModel, "token_bos", RUBY_METHOD_FUNC(_llama_model_token_bos), 0);
     rb_define_method(rb_cLLaMAModel, "token_eos", RUBY_METHOD_FUNC(_llama_model_token_eos), 0);
+    rb_define_method(rb_cLLaMAModel, "token_cls", RUBY_METHOD_FUNC(_llama_model_token_cls), 0);
+    rb_define_method(rb_cLLaMAModel, "token_sep", RUBY_METHOD_FUNC(_llama_model_token_sep), 0);
     rb_define_method(rb_cLLaMAModel, "token_nl", RUBY_METHOD_FUNC(_llama_model_token_nl), 0);
     rb_define_method(rb_cLLaMAModel, "add_bos_token?", RUBY_METHOD_FUNC(_llama_model_add_bos_token), 0);
     rb_define_method(rb_cLLaMAModel, "add_eos_token?", RUBY_METHOD_FUNC(_llama_model_add_eos_token), 0);
@@ -1743,6 +1745,16 @@ private:
     return INT2NUM(llama_token_eos(ptr->model));
   }
 
+  static VALUE _llama_model_token_cls(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_cls(ptr->model));
+  }
+
+  static VALUE _llama_model_token_sep(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_sep(ptr->model));
+  }
+
   static VALUE _llama_model_token_nl(VALUE self) {
     LLaMAModelWrapper* ptr = get_llama_model(self);
     return INT2NUM(llama_token_nl(ptr->model));
@@ -3414,15 +3426,26 @@ extern "C" void Init_llama_cpp(void) {
   ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGSN;
   rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGSN", rb_str_new2(ss_magic.str().c_str()));
 
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
+  ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGSQ;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGSQ", rb_str_new2(ss_magic.str().c_str()));
+
   ss_magic.str("");
   ss_magic.clear(std::stringstream::goodbit);
   ss_magic << std::showbase << std::hex << LLAMA_SESSION_MAGIC;
   rb_define_const(rb_mLLaMACpp, "LLAMA_SESSION_MAGIC", rb_str_new2(ss_magic.str().c_str()));
 
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
+  ss_magic << std::showbase << std::hex << LLAMA_STATE_SEQ_MAGIC;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_STATE_SEQ_MAGIC", rb_str_new2(ss_magic.str().c_str()));
+
   ss_magic.str("");
   ss_magic.clear(std::stringstream::goodbit);
   ss_magic << std::showbase << std::hex << LLAMA_DEFAULT_SEED;
   rb_define_const(rb_mLLaMACpp, "LLAMA_DEFAULT_SEED", rb_str_new2(ss_magic.str().c_str()));
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_SESSION_VERSION", rb_str_new2(std::to_string(LLAMA_SESSION_VERSION).c_str()));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_STATE_SEQ_VERSION", rb_str_new2(std::to_string(LLAMA_STATE_SEQ_VERSION).c_str()));
 }

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 = '0.14.4'
+  VERSION = '0.14.5'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b2608'
+  LLAMA_CPP_VERSION = 'b2658'
 end

data/sig/llama_cpp.rbs

@@ -3,6 +3,14 @@ module LLaMACpp
   LLAMA_CPP_VERSION: String
   LLAMA_DEFALUT_SEED: String
 
+  LLAMA_FILE_MAGIC_GGLA: String
+  LLAMA_FILE_MAGIC_GGSN: String
+  LLAMA_FILE_MAGIC_GGSQ: String
+  LLAMA_SESSION_MAGIC: String
+  LLAMA_SESSION_VERSION: String
+  LLAMA_STATE_SEQ_MAGIC: String
+  LLAMA_STATE_SEQ_VERSION: String
+
   LLAMA_VOCAB_TYPE_NONE: Integer
   LLAMA_VOCAB_TYPE_SPM: Integer
   LLAMA_VOCAB_TYPE_BPE: Integer
@@ -124,6 +132,8 @@ module LLaMACpp
     def type: (Integer) -> Integer
     def token_bos: () -> Integer
     def token_eos: () -> Integer
+    def token_cls: () -> Integer
+    def token_sep: () -> Integer
     def token_nl: () -> Integer
     def add_bos_token?: () -> bool
     def add_eos_token?: () -> bool

data/vendor/tmp/llama.cpp/LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023 Georgi Gerganov
+Copyright (c) 2023-2024 The ggml 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

data/vendor/tmp/llama.cpp/Makefile

@@ -1,7 +1,7 @@
 # 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 gguf-split llama-bench libllava.a llava-cli baby-llama beam-search  \
+	simple batched batched-bench save-load-state server gguf gguf-split eval-callback 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
@@ -10,7 +10,7 @@ TEST_TARGETS = \
 	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-json-schema-to-grammar
+	tests/test-json-schema-to-grammar tests/test-grammar-integration
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -648,7 +648,7 @@ CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])'
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 ifndef CUDA_DOCKER_ARCH
 ifndef CUDA_POWER_ARCH
-$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via environment variable CUDA_DOCKER_ARCH, e.g. by running "export CUDA_DOCKER_ARCH=compute_XX" on Unix-like systems, where XX is the minimum compute capability that the code needs to run on. A list with compute capabilities can be found here: https://developer.nvidia.com/cuda-gpus )
 endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
@@ -805,6 +805,10 @@ gguf-split: examples/gguf-split/gguf-split.cpp ggml.o llama.o $(COMMON_DEPS) $(O
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+eval-callback: examples/eval-callback/eval-callback.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)
@@ -923,6 +927,10 @@ tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+tests/test-grammar-integration: tests/test-grammar-integration.cpp ggml.o llama.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 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)

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

@@ -137,7 +137,7 @@ extern "C" {
     /*
       Example usage:
 
-        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be asigned
+        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
         // preferrably to run on the same backend as the buffer
         ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 

data/vendor/tmp/llama.cpp/ggml-cuda.cu

@@ -1225,7 +1225,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
-    int ldc = id == ctx.device ? ne0 : row_diff;
+    int64_t ldc = id == ctx.device ? ne0 : row_diff;
 
     const int compute_capability = ggml_cuda_info().devices[id].cc;
 
@@ -1377,8 +1377,8 @@ static void ggml_cuda_op_mul_mat(
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
 
-    const int nb2 = dst->nb[2];
-    const int nb3 = dst->nb[3];
+    const int64_t nb2 = dst->nb[2];
+    const int64_t nb3 = dst->nb[3];
 
     GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
     GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
@@ -2617,6 +2617,7 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size
         return false;
     }
 
+#if CUDART_VERSION >= 11100
     cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
     if (err != cudaSuccess) {
         // clear the error
@@ -2627,6 +2628,9 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size
         return false;
     }
     return true;
+#else
+    return false;
+#endif
 }
 
 GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer) {