llama_cpp 0.9.0 → 0.9.1

data/ext/llama_cpp/src/llama.h

@@ -106,6 +106,14 @@ extern "C" {
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
 
+    enum llama_rope_scaling_type {
+        LLAMA_ROPE_SCALING_UNSPECIFIED = -1,
+        LLAMA_ROPE_SCALING_NONE        = 0,
+        LLAMA_ROPE_SCALING_LINEAR      = 1,
+        LLAMA_ROPE_SCALING_YARN        = 2,
+        LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
+    };
+
     typedef struct llama_token_data {
         llama_token id; // token id
         float logit;    // log-odds of the token
@@ -172,13 +180,19 @@ extern "C" {
         uint32_t n_batch;         // prompt processing maximum batch size
         uint32_t n_threads;       // number of threads to use for generation
         uint32_t n_threads_batch; // number of threads to use for batch processing
+        int8_t   rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
 
         // ref: https://github.com/ggerganov/llama.cpp/pull/2054
-        float rope_freq_base;  // RoPE base frequency, 0 = from model
-        float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
+        float    rope_freq_base;   // RoPE base frequency, 0 = from model
+        float    rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
+        float    yarn_ext_factor;  // YaRN extrapolation mix factor, NaN = from model
+        float    yarn_attn_factor; // YaRN magnitude scaling factor
+        float    yarn_beta_fast;   // YaRN low correction dim
+        float    yarn_beta_slow;   // YaRN high correction dim
+        uint32_t yarn_orig_ctx;    // YaRN original context size
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
         bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool embedding;  // embedding mode only
@@ -191,6 +205,7 @@ extern "C" {
         bool allow_requantize;       // allow quantizing non-f32/f16 tensors
         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
     } llama_model_quantize_params;
 
     // grammar types
@@ -333,17 +348,14 @@ extern "C" {
     LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
             "avoid using this, it will be removed in the future, instead - count the tokens in user code");
 
-    // Remove all tokens data of cells in [c0, c1)
-    // c0 < 0 : [0,  c1]
-    // c1 < 0 : [c0, inf)
-    LLAMA_API void llama_kv_cache_tokens_rm(
-            struct llama_context * ctx,
-                         int32_t   c0,
-                         int32_t   c1);
+    // Clear the KV cache
+    LLAMA_API void llama_kv_cache_clear(
+            struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
     LLAMA_API void llama_kv_cache_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
@@ -600,6 +612,13 @@ extern "C" {
                            float   p,
                           size_t   min_keep);
 
+    /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
+    LLAMA_API void llama_sample_min_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
     /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
     LLAMA_API void llama_sample_tail_free(
             struct llama_context * ctx,
@@ -658,6 +677,7 @@ extern "C" {
                            float * mu);
 
     /// @details Selects the token with the highest probability.
+    ///          Does not compute the token probabilities. Use llama_sample_softmax() instead.
     LLAMA_API llama_token llama_sample_token_greedy(
             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 = '0.9.0'
+  VERSION = '0.9.1'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b1429'
+  LLAMA_CPP_VERSION = 'b1472'
 end

data/sig/llama_cpp.rbs

@@ -31,6 +31,12 @@ module LLaMACpp
   LLAMA_GRETYPE_CHAR_RNG_UPPER: Integer
   LLAMA_GRETYPE_CHAR_ALT: Integer
 
+  LLAMA_ROPE_SCALING_UNSPECIFIED: Integer
+  LLAMA_ROPE_SCALING_NONE: Integer
+  LLAMA_ROPE_SCALING_LINEAR: Integer
+  LLAMA_ROPE_SCALING_YARN: Integer
+  LLAMA_ROPE_SCALING_MAX_VALUE: Integer
+
   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
@@ -162,11 +168,11 @@ module LLaMACpp
     def print_timings: () -> void
     def reset_timings: () -> void
     def kv_cache_token_count: () -> Integer
-    def kv_cache_tokens_rm: (Integer, Integer) -> void
+    def kv_cache_clear: () -> void
     def kv_cache_seq_rm: (Integer, Integer,Integer) -> void
     def kv_cache_seq_cp: (Integer, Integer,Integer, Integer) -> void
     def kv_cache_seq_keep: (Integer) -> void
-    def kv_cache_seq_shift: (Integer, Integer, Ingteger, Integer) -> void
+    def kv_cache_seq_shift: (Integer, Integer, Integer, Integer) -> void
     def set_rng_seed: (Integer) -> void
     def load_session_file: (session_path: String) -> void
     def save_session_file: (session_path: String, session_tokens: Array[Integer]) -> void
@@ -175,6 +181,7 @@ module LLaMACpp
     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
+    def sample_min_p: (::LLaMACpp::TokenDataArray, prob: Float, ?min_keep: Integer) -> void
     def sample_tail_free: (::LLaMACpp::TokenDataArray, z: Float, ?min_keep: Integer) -> void
     def sample_typical: (::LLaMACpp::TokenDataArray, prob: Float, ?min_keep: Integer) -> void
     def sample_temp: (::LLaMACpp::TokenDataArray, temp: Float) -> void
@@ -200,10 +207,22 @@ module LLaMACpp
     def n_threads=: (Integer) -> Integer
     def n_threads_batch: () -> Integer
     def n_threads_batch=: (Integer) -> Integer
+    def rope_scaling_type=: (Integer) -> Integer
+    def rope_scaling_type: () -> Integer
     def rope_freq_base=: (Float) -> Float
     def rope_freq_base: () -> Float
     def rope_freq_scale=: (Float) -> Float
     def rope_freq_scale: () -> Float
+    def yarn_ext_factor=: (Float) -> Float
+    def yarn_ext_factor: () -> Float
+    def yarn_attn_factor=: (Float) -> Float
+    def yarn_attn_factor: () -> Float
+    def yarn_beta_fast=: (Float) -> Float
+    def yarn_beta_fast: () -> Float
+    def yarn_beta_slow=: (Float) -> Float
+    def yarn_beta_slow: () -> Float
+    def yarn_orig_ctx=: (Integer) -> Integer
+    def yarn_orig_ctx: () -> Integer
     def mul_mat_q: () -> bool
     def mul_mat_q=: (bool) -> bool
     def f16_kv: () -> bool
@@ -227,6 +246,8 @@ module LLaMACpp
     def quantize_output_tensor=: (bool) -> bool
     def only_copy: () -> bool
     def only_copy=: (bool) -> bool
+    def pure: () -> bool
+    def pure=: (bool) -> bool
   end
 
   class Params = ContextParams

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.9.0
+  version: 0.9.1
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2023-10-28 00:00:00.000000000 Z
+date: 2023-11-03 00:00:00.000000000 Z
 dependencies: []
 description: llama_cpp.rb provides Ruby bindings for the llama.cpp.
 email:
@@ -36,6 +36,7 @@ files:
 - ext/llama_cpp/src/ggml-backend.h
 - ext/llama_cpp/src/ggml-cuda.cu
 - ext/llama_cpp/src/ggml-cuda.h
+- ext/llama_cpp/src/ggml-impl.h
 - ext/llama_cpp/src/ggml-metal.h
 - ext/llama_cpp/src/ggml-metal.m
 - ext/llama_cpp/src/ggml-metal.metal
@@ -43,10 +44,10 @@ files:
 - 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-quants.c
+- ext/llama_cpp/src/ggml-quants.h
 - ext/llama_cpp/src/ggml.c
 - ext/llama_cpp/src/ggml.h
-- ext/llama_cpp/src/k_quants.c
-- ext/llama_cpp/src/k_quants.h
 - ext/llama_cpp/src/llama-util.h
 - ext/llama_cpp/src/llama.cpp
 - ext/llama_cpp/src/llama.h