llama_cpp 0.3.2 → 0.3.4

data/ext/llama_cpp/src/llama.h

@@ -88,7 +88,13 @@ extern "C" {
         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
+
+        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
+
+        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+        float    rope_freq_base;  // RoPE base frequency
+        float    rope_freq_scale; // RoPE frequency scaling factor
+
         // 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
@@ -148,6 +154,8 @@ extern "C" {
         int32_t n_eval;
     };
 
+    LLAMA_API int llama_max_devices();
+
     LLAMA_API struct llama_context_params llama_context_default_params();
     LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
 
@@ -158,7 +166,9 @@ extern "C" {
     // Initialize the llama + ggml backend
     // If numa is true, use NUMA optimizations
     // Call once at the start of the program
-    LLAMA_API void llama_init_backend(bool numa);
+    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();
 
@@ -268,10 +278,21 @@ extern "C" {
                              int   n_max_tokens,
                             bool   add_bos);
 
+    LLAMA_API int llama_tokenize_with_model(
+        const struct llama_model * model,
+                      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);
 
+    LLAMA_API int llama_n_vocab_from_model(const struct llama_model * model);
+    LLAMA_API int llama_n_ctx_from_model  (const struct llama_model * model);
+    LLAMA_API int llama_n_embd_from_model (const struct llama_model * model);
+
     // Get the vocabulary as output parameters.
     // Returns number of results.
     LLAMA_API int llama_get_vocab(
@@ -280,6 +301,12 @@ extern "C" {
                                  float * scores,
                                    int   capacity);
 
+    LLAMA_API int llama_get_vocab_from_model(
+              const struct llama_model * model,
+                          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
@@ -292,7 +319,13 @@ extern "C" {
     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);
+    LLAMA_API const char * llama_token_to_str(
+            const struct llama_context * ctx,
+                           llama_token   token);
+
+    LLAMA_API const char * llama_token_to_str_with_model(
+              const struct llama_model * model,
+                           llama_token   token);
 
     // Special tokens
     LLAMA_API llama_token llama_token_bos();  // beginning-of-sentence
@@ -307,6 +340,16 @@ extern "C" {
     /// @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.
+    LLAMA_API void llama_sample_classifier_free_guidance(
+              struct llama_context * ctx,
+            llama_token_data_array * candidates,
+              struct llama_context * guidance_ctx,
+                             float   scale);
+
     /// @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);
 

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 = '0.3.2'
+  VERSION = '0.3.4'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'master-481f793'
+  LLAMA_CPP_VERSION = 'master-d924522'
 end

data/lib/llama_cpp.rb

@@ -108,4 +108,5 @@ module LLaMACpp
   end
 end
 
-LLaMACpp.init_backend
+LLaMACpp.backend_init
+at_exit { LLaMACpp.backend_free }

data/sig/llama_cpp.rbs

@@ -26,7 +26,8 @@ module LLaMACpp
   LLAMA_FTYPE_MOSTLY_Q5_K_M: Integer
   LLAMA_FTYPE_MOSTLY_Q6_K: Integer
 
-  def self?.init_backend: (?numa: bool) -> void
+  def self?.backend_init: (?numa: bool) -> void
+  def self?.backend_free: () -> void
   def self?.model_quantize: (input_path: String, output_path: String, params: ModelQuantizeParams) -> void
   def self?.generate: (::LLaMACpp::Context, String,
     ?n_predict: Integer, ?n_threads: Integer, ?n_keep: Integer, ?n_batch: Integer,
@@ -38,6 +39,7 @@ module LLaMACpp
   def self?.token_nl: () -> Integer
   def self?.mmap_supported?: () -> bool
   def self?.mlock_supported?: () -> bool
+  def self?.max_devices: () -> Integer
 
   class TokenData
     public
@@ -68,6 +70,12 @@ module LLaMACpp
     def free: () -> void
     def load: (model_path: String, params: ::LLaMACpp::ContextParams) -> void
     def apply_lora_from_file: (lora_path: String, ?base_model_path: String, ?n_threads: Integer) -> void
+    def n_vocab: () -> Integer
+    def n_ctx: () -> Integer
+    def n_embd: () -> Integer
+    def vocab: (capacity: Integer) -> [Array[String], Array[Float]]
+    def token_to_str: (Integer) -> String
+    def tokenize: (text: String, ?n_max_tokens: Integer, ?add_bos: bool) -> Array[Integer]
   end
 
   class Timings
@@ -108,6 +116,7 @@ module LLaMACpp
     def save_session_file: (session_path: String, session_tokens: Array[Integer]) -> void
     def sample_repetition_penalty: (::LLaMACpp::TokenDataArray, Array[Integer], penalty: Float) -> void
     def sample_frequency_and_presence_penalties: (::LLaMACpp::TokenDataArray, Array[Integer], frequency: Float, presence: Float) -> void
+    def sample_classifier_free_guidance: (::LLaMACpp::TokenDataArray, guidance: ::LLaMACpp::Context, scale: Float) -> void
     def sample_softmax: (::LLaMACpp::TokenDataArray) -> void
     def sample_top_k: (::LLaMACpp::TokenDataArray, k: Integer, ?min_keep: Integer) -> void
     def sample_top_p: (::LLaMACpp::TokenDataArray, prob: Float, ?min_keep: Integer) -> void
@@ -138,6 +147,10 @@ module LLaMACpp
     def main_gpu: () -> Integer
     def main_gpu=: (Integer) -> Integer
     def tensor_split: () -> Array[Float]
+    def rope_freq_base=: (Float) -> Float
+    def rope_freq_base: () -> Float
+    def rope_freq_scale=: (Float) -> Float
+    def rope_freq_scale: () -> Float
     def low_vram: () -> bool
     def low_vram=: (bool) -> bool
     def seed: () -> Integer

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.3.2
+  version: 0.3.4
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2023-07-08 00:00:00.000000000 Z
+date: 2023-07-22 00:00:00.000000000 Z
 dependencies: []
 description: llama_cpp.rb provides Ruby bindings for the llama.cpp.
 email:
@@ -35,6 +35,8 @@ files:
 - ext/llama_cpp/src/ggml-metal.h
 - ext/llama_cpp/src/ggml-metal.m
 - ext/llama_cpp/src/ggml-metal.metal
+- ext/llama_cpp/src/ggml-mpi.c
+- ext/llama_cpp/src/ggml-mpi.h
 - ext/llama_cpp/src/ggml-opencl.cpp
 - ext/llama_cpp/src/ggml-opencl.h
 - ext/llama_cpp/src/ggml.c