llama_cpp 0.12.0 → 0.12.2

data/vendor/tmp/llama.cpp/llama.h

@@ -43,7 +43,7 @@
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 3
+#define LLAMA_SESSION_VERSION 4
 
 #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.
@@ -103,6 +103,9 @@ extern "C" {
         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
+        LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -115,6 +118,12 @@ extern "C" {
         LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
     };
 
+    enum llama_split_mode {
+        LLAMA_SPLIT_NONE    = 0, // single GPU
+        LLAMA_SPLIT_LAYER   = 1, // split layers and KV across GPUs
+        LLAMA_SPLIT_ROW     = 2, // split rows across GPUs
+    };
+
     typedef struct llama_token_data {
         llama_token id; // token id
         float logit;    // log-odds of the token
@@ -177,8 +186,16 @@ extern "C" {
 
     struct llama_model_params {
         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
-        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
+        enum llama_split_mode split_mode; // how to split the model across multiple GPUs
+
+        // main_gpu interpretation depends on split_mode:
+        // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
+        // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
+        // LLAMA_SPLIT_LAYER: ignored
+        int32_t main_gpu;
+
+        // proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES
+        const float * tensor_split;
 
         // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
         // If the provided progress_callback returns true, model loading continues.
@@ -232,6 +249,7 @@ extern "C" {
         bool quantize_output_tensor; // quantize output.weight
         bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
         bool pure;                   // disable k-quant mixtures and quantize all tensors to the same type
+        void * imatrix;              // pointer to importance matrix data
     } llama_model_quantize_params;
 
     // grammar types
@@ -484,6 +502,17 @@ extern "C" {
                        llama_pos   p1,
                        llama_pos   delta);
 
+    // Integer division of the positions by factor of `d > 1`
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d);
+
     //
     // State / sessions
     //
@@ -685,14 +714,21 @@ extern "C" {
                            float   penalty_present);
 
     /// @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(
+    /// @param logits Logits extracted from the original generation context.
+    /// @param logits_guidance Logits extracted from 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.
+    /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
+    LLAMA_API void llama_sample_apply_guidance(
+              struct llama_context * ctx,
+                             float * logits,
+                             float * logits_guidance,
+                             float   scale);
+
+    LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance(
               struct llama_context * ctx,
             llama_token_data_array * candidates,
               struct llama_context * guidance_ctx,
-                             float   scale);
+                             float   scale),
+              "use llama_sample_apply_guidance() instead");
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     LLAMA_API void llama_sample_softmax(

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.12.0
+  version: 0.12.2
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2024-01-11 00:00:00.000000000 Z
+date: 2024-01-20 00:00:00.000000000 Z
 dependencies: []
 description: llama_cpp.rb provides Ruby bindings for the llama.cpp.
 email: