revert to origina llama_sample_token_mirostat

guinmoon · Nov 30, 2023 · b937dfd · b937dfd
1 parent 01de340
commit b937dfd
Show file tree

Hide file tree

Showing 5 changed files with 8 additions and 9 deletions.
diff --git a/Sources/llmfarm_core/LLMBase.swift b/Sources/llmfarm_core/LLMBase.swift
@@ -222,7 +222,7 @@ public class LLMBase {
                 if class_name != "llmfarm_core.LLaMa"{
                     res_token =  llama_dadbed9_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, Int32(mirostat_m), &mirostat_mu, vocabSize);
                 }else{
-                    res_token =  llama_sample_token_mirostat_for_dadbed9(ctx, &candidates_p, mirostat_tau, mirostat_eta, Int32(mirostat_m), &mirostat_mu, vocabSize);
+                    res_token =  llama_sample_token_mirostat_for_dadbed9(ctx, &candidates_p, mirostat_tau, mirostat_eta, Int32(mirostat_m), &mirostat_mu);
                 }
             } else if(mirostat == 2) {
                 var mirostat_mu: Float = 2.0 * mirostat_tau

diff --git a/Sources/llmfarm_core_cpp/gpt_spm.cpp b/Sources/llmfarm_core_cpp/gpt_spm.cpp
@@ -349,8 +349,8 @@ llama_token llama_sample_token_for_dadbed9(struct llama_context * ctx, llama_dad
 }
 
 
-llama_token llama_sample_token_mirostat_for_dadbed9(struct llama_context * ctx, llama_dadbed9_token_data_array * candidates,float tau,float   eta,int   m,float * mu,int vocabSize ) {
-    return llama_sample_token_mirostat(ctx, (llama_token_data_array *)candidates,tau,eta,m,mu,vocabSize );
+llama_token llama_sample_token_mirostat_for_dadbed9(struct llama_context * ctx, llama_dadbed9_token_data_array * candidates,float tau,float   eta,int   m,float * mu) {
+    return llama_sample_token_mirostat(ctx, (llama_token_data_array *)candidates,tau,eta,m,mu );
 }
 
 llama_token llama_sample_token_mirostat_v2_for_dadbed9(struct llama_context * ctx, llama_dadbed9_token_data_array * candidates,float tau,float   eta, float * mu ) {

diff --git a/Sources/llmfarm_core_cpp/llama/llama.cpp b/Sources/llmfarm_core_cpp/llama/llama.cpp
@@ -7552,11 +7552,11 @@ void llama_sample_classifier_free_guidance(
     }
 }
 
-llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu, int vocabSize) {
+llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu) {
     GGML_ASSERT(ctx);
 
-//    auto N = float(llama_n_vocab(llama_get_model(ctx)));
-    auto N = float(vocabSize);
+    auto N = float(llama_n_vocab(llama_get_model(ctx)));
+//    auto N = float(vocabSize);
     int64_t t_start_sample_us;
     t_start_sample_us = ggml_time_us();
 

diff --git a/Sources/llmfarm_core_cpp/spm-headers/gpt_spm.h b/Sources/llmfarm_core_cpp/spm-headers/gpt_spm.h
@@ -154,7 +154,7 @@ struct llama_grammar* llama_load_grammar(const char* grammar_path);
 
 void llama_sample_grammar_for_dadbed9(struct llama_context * ctx, llama_dadbed9_token_data_array * candidates, const struct llama_grammar * grammar );
 llama_token llama_sample_token_for_dadbed9(struct llama_context * ctx, llama_dadbed9_token_data_array * candidates );
-llama_token llama_sample_token_mirostat_for_dadbed9(struct llama_context * ctx, llama_dadbed9_token_data_array * candidates,float tau,float   eta,int   m,float * mu,int vocabSize );
+llama_token llama_sample_token_mirostat_for_dadbed9(struct llama_context * ctx, llama_dadbed9_token_data_array * candidates,float tau,float   eta,int   m,float * mu );
 llama_token llama_sample_token_mirostat_v2_for_dadbed9(struct llama_context * ctx, llama_dadbed9_token_data_array * candidates,float tau,float   eta, float * mu ) ;
 
 #ifdef __cplusplus

diff --git a/Sources/llmfarm_core_cpp/spm-headers/llama.h b/Sources/llmfarm_core_cpp/spm-headers/llama.h
@@ -645,8 +645,7 @@ extern "C" {
                            float   tau,
                            float   eta,
                              int   m,
-                           float * mu,
-                           int vocabSize);
+                           float * mu);
 
     /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.