add min_p sampling parameter (nomic-ai#2014)

Signed-off-by: Christopher Barrera <[email protected]>
Co-authored-by: Jared Van Bortel <[email protected]>

gpt4all-backend/llamamodel.cpp

@@ -64,6 +64,7 @@ static int llama_sample_top_p_top_k(
         int last_n_tokens_size,
         int top_k,
         float top_p,
+        float min_p,
         float temp,
         float repeat_penalty,
         int32_t pos) {
@@ -83,6 +84,7 @@ static int llama_sample_top_p_top_k(
     llama_sample_tail_free(ctx, &candidates_p, 1.0f, 1);
     llama_sample_typical(ctx, &candidates_p, 1.0f, 1);
     llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+    llama_sample_min_p(ctx, &candidates_p, min_p, 1);
     llama_sample_temp(ctx, &candidates_p, temp);
     return llama_sample_token(ctx, &candidates_p);
 }
@@ -392,7 +394,7 @@ LLModel::Token LLamaModel::sampleToken(PromptContext &promptCtx) const
     const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
     return llama_sample_top_p_top_k(d_ptr->ctx,
         promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
-        n_prev_toks, promptCtx.top_k, promptCtx.top_p, promptCtx.temp,
+        n_prev_toks, promptCtx.top_k, promptCtx.top_p, promptCtx.min_p, promptCtx.temp,
         promptCtx.repeat_penalty, promptCtx.n_last_batch_tokens - 1);
 }
 

gpt4all-backend/llmodel.h

@@ -66,6 +66,7 @@ class LLModel {
         int32_t n_predict = 200;
         int32_t top_k = 40;
         float   top_p = 0.9f;
+        float   min_p = 0.0f;
         float   temp = 0.9f;
         int32_t n_batch = 9;
         float   repeat_penalty = 1.10f;

gpt4all-backend/llmodel_c.cpp

@@ -134,6 +134,7 @@ void llmodel_prompt(llmodel_model model, const char *prompt,
     wrapper->promptContext.n_predict = ctx->n_predict;
     wrapper->promptContext.top_k = ctx->top_k;
     wrapper->promptContext.top_p = ctx->top_p;
+    wrapper->promptContext.min_p = ctx->min_p;
     wrapper->promptContext.temp = ctx->temp;
     wrapper->promptContext.n_batch = ctx->n_batch;
     wrapper->promptContext.repeat_penalty = ctx->repeat_penalty;
@@ -156,6 +157,7 @@ void llmodel_prompt(llmodel_model model, const char *prompt,
     ctx->n_predict = wrapper->promptContext.n_predict;
     ctx->top_k = wrapper->promptContext.top_k;
     ctx->top_p = wrapper->promptContext.top_p;
+    ctx->min_p = wrapper->promptContext.min_p;
     ctx->temp = wrapper->promptContext.temp;
     ctx->n_batch = wrapper->promptContext.n_batch;
     ctx->repeat_penalty = wrapper->promptContext.repeat_penalty;

gpt4all-backend/llmodel_c.h

@@ -39,6 +39,7 @@ struct llmodel_prompt_context {
     int32_t n_predict;      // number of tokens to predict
     int32_t top_k;          // top k logits to sample from
     float top_p;            // nucleus sampling probability threshold
+    float min_p;            // Min P sampling
     float temp;             // temperature to adjust model's output distribution
     int32_t n_batch;        // number of predictions to generate in parallel
     float repeat_penalty;   // penalty factor for repeated tokens

gpt4all-bindings/cli/app.py

@@ -120,6 +120,7 @@ def _old_loop(gpt4all_instance):
             n_predict=200,
             top_k=40,
             top_p=0.9,
+            min_p=0.0,
             temp=0.9,
             n_batch=9,
             repeat_penalty=1.1,
@@ -156,6 +157,7 @@ def _new_loop(gpt4all_instance):
                 temp=0.9,
                 top_k=40,
                 top_p=0.9,
+                min_p=0.0,
                 repeat_penalty=1.1,
                 repeat_last_n=64,
                 n_batch=9,

gpt4all-bindings/csharp/Gpt4All/Bindings/LLPromptContext.cs

@@ -64,6 +64,15 @@ public float TopP
         set => _ctx.top_p = value;
     }
 
+    /// <summary>
+    /// min p sampling probability threshold
+    /// </summary>
+    public float MinP
+    {
+        get => _ctx.min_p;
+        set => _ctx.min_p = value;
+    }
+
     /// <summary>
     /// temperature to adjust model's output distribution
     /// </summary>

gpt4all-bindings/csharp/Gpt4All/Bindings/NativeMethods.cs

@@ -29,6 +29,8 @@ public unsafe partial struct llmodel_prompt_context
 
     public float top_p;
 
+    public float min_p;
+
     public float temp;
 
     [NativeTypeName("int32_t")]

gpt4all-bindings/csharp/Gpt4All/Extensions/LLPromptContextExtensions.cs

@@ -16,6 +16,7 @@ public static string Dump(this LLModelPromptContext context)
             n_predict = {ctx.n_predict}
             top_k = {ctx.top_k}
             top_p = {ctx.top_p}
+            min_p = {ctx.min_p}
             temp = {ctx.temp}
             n_batch = {ctx.n_batch}
             repeat_penalty = {ctx.repeat_penalty}

gpt4all-bindings/csharp/Gpt4All/Extensions/PredictRequestOptionsExtensions.cs

@@ -12,6 +12,7 @@ public static LLModelPromptContext ToPromptContext(this PredictRequestOptions op
             TokensSize = opts.TokensSize,
             TopK = opts.TopK,
             TopP = opts.TopP,
+            MinP = opts.MinP,
             PastNum = opts.PastConversationTokensNum,
             RepeatPenalty = opts.RepeatPenalty,
             Temperature = opts.Temperature,

gpt4all-bindings/csharp/Gpt4All/Prediction/PredictRequestOptions.cs

@@ -16,6 +16,8 @@ public record PredictRequestOptions
 
     public float TopP { get; init; } = 0.9f;
 
+    public float MinP { get; init; } = 0.0f;
+
     public float Temperature { get; init; } = 0.1f;
 
     public int Batches { get; init; } = 8;

gpt4all-bindings/golang/binding.cpp

@@ -36,7 +36,7 @@ std::string res = "";
 void * mm;
 
 void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n, float repeat_penalty, int n_ctx, int tokens, int top_k,
-                            float top_p, float temp, int n_batch,float ctx_erase)
+                            float top_p, float min_p, float temp, int n_batch,float ctx_erase)
 {
     llmodel_model* model = (llmodel_model*) m;
 
@@ -69,6 +69,7 @@ void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n,
         .n_predict = 50,
         .top_k = 10,
         .top_p = 0.9,
+        .min_p = 0.0,
         .temp = 1.0,
         .n_batch = 1,
         .repeat_penalty = 1.2,
@@ -83,6 +84,7 @@ void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n,
     prompt_context->top_k = top_k;
     prompt_context->context_erase = ctx_erase;
     prompt_context->top_p = top_p;
+    prompt_context->min_p = min_p;
     prompt_context->temp = temp;
     prompt_context->n_batch = n_batch;    
 

gpt4all-bindings/golang/binding.h

@@ -7,12 +7,12 @@ extern "C" {
 void* load_model(const char *fname, int n_threads);
 
 void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n, float repeat_penalty, int n_ctx, int tokens, int top_k,
-                            float top_p, float temp, int n_batch,float ctx_erase);
+                            float top_p, float min_p, float temp, int n_batch,float ctx_erase);
 
 void free_model(void *state_ptr);
 
 extern unsigned char getTokenCallback(void *, char *);
 
 #ifdef __cplusplus
 }
-#endif
+#endif

gpt4all-bindings/golang/gpt4all.go

@@ -7,7 +7,7 @@ package gpt4all
 // #cgo LDFLAGS: -lgpt4all -lm -lstdc++ -ldl
 // void* load_model(const char *fname, int n_threads);
 // void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n, float repeat_penalty, int n_ctx, int tokens, int top_k,
-//                            float top_p, float temp, int n_batch,float ctx_erase);
+//                            float top_p, float min_p, float temp, int n_batch,float ctx_erase);
 // void free_model(void *state_ptr);
 // extern unsigned char getTokenCallback(void *, char *);
 // void llmodel_set_implementation_search_path(const char *path);
@@ -58,7 +58,7 @@ func (l *Model) Predict(text string, opts ...PredictOption) (string, error) {
 	out := make([]byte, po.Tokens)
 
 	C.model_prompt(input, l.state, (*C.char)(unsafe.Pointer(&out[0])), C.int(po.RepeatLastN), C.float(po.RepeatPenalty), C.int(po.ContextSize),
-		C.int(po.Tokens), C.int(po.TopK), C.float(po.TopP), C.float(po.Temperature), C.int(po.Batch), C.float(po.ContextErase))
+		C.int(po.Tokens), C.int(po.TopK), C.float(po.TopP), C.float(po.MinP), C.float(po.Temperature), C.int(po.Batch), C.float(po.ContextErase))
 
 	res := C.GoString((*C.char)(unsafe.Pointer(&out[0])))
 	res = strings.TrimPrefix(res, " ")

gpt4all-bindings/golang/options.go

@@ -2,7 +2,7 @@ package gpt4all
 
 type PredictOptions struct {
 	ContextSize, RepeatLastN, Tokens, TopK, Batch  int
-	TopP, Temperature, ContextErase, RepeatPenalty float64
+	TopP, MinP, Temperature, ContextErase, RepeatPenalty float64
 }
 
 type PredictOption func(p *PredictOptions)
@@ -11,6 +11,7 @@ var DefaultOptions PredictOptions = PredictOptions{
 	Tokens:        200,
 	TopK:          10,
 	TopP:          0.90,
+        MinP:          0.0,
 	Temperature:   0.96,
 	Batch:         1,
 	ContextErase:  0.55,
@@ -50,6 +51,13 @@ func SetTopP(topp float64) PredictOption {
 	}
 }
 
+// SetMinP sets the value for min p sampling
+func SetMinP(minp float64) PredictOption {
+	return func(p *PredictOptions) {
+		p.MinP = minp
+	}
+}
+
 // SetRepeatPenalty sets the repeat penalty.
 func SetRepeatPenalty(ce float64) PredictOption {
 	return func(p *PredictOptions) {

gpt4all-bindings/java/src/main/java/com/hexadevlabs/gpt4all/LLModel.java

@@ -32,6 +32,7 @@ private GenerationConfig() {
             n_predict.set(128);
             top_k.set(40);
             top_p.set(0.95);
+            min_p.set(0.0);
             temp.set(0.28);
             n_batch.set(8);
             repeat_penalty.set(1.1);
@@ -71,6 +72,11 @@ public Builder withTopP(float top_p) {
                 return this;
             }
 
+            public Builder withMinP(float min_p) {
+                configToBuild.min_p.set(min_p);
+                return this;
+            }
+
             public Builder withTemp(float temp) {
                 configToBuild.temp.set(temp);
                 return this;

gpt4all-bindings/java/src/main/java/com/hexadevlabs/gpt4all/LLModelLibrary.java

@@ -48,6 +48,7 @@ class LLModelPromptContext extends Struct {
         public final int32_t n_predict = new int32_t();
         public final int32_t top_k = new int32_t();
         public final Float top_p = new Float();
+        public final Float min_p = new Float();
         public final Float temp = new Float();
         public final int32_t n_batch = new int32_t();
         public final Float repeat_penalty = new Float();

gpt4all-bindings/python/gpt4all/_pyllmodel.py

@@ -49,6 +49,7 @@ class LLModelPromptContext(ctypes.Structure):
         ("n_predict", ctypes.c_int32),
         ("top_k", ctypes.c_int32),
         ("top_p", ctypes.c_float),
+        ("min_p", ctypes.c_float),
         ("temp", ctypes.c_float),
         ("n_batch", ctypes.c_int32),
         ("repeat_penalty", ctypes.c_float),
@@ -241,6 +242,7 @@ def _set_context(
         n_predict: int = 4096,
         top_k: int = 40,
         top_p: float = 0.9,
+        min_p: float = 0.0,
         temp: float = 0.1,
         n_batch: int = 8,
         repeat_penalty: float = 1.2,
@@ -257,6 +259,7 @@ def _set_context(
                 n_predict=n_predict,
                 top_k=top_k,
                 top_p=top_p,
+                min_p=min_p,
                 temp=temp,
                 n_batch=n_batch,
                 repeat_penalty=repeat_penalty,
@@ -272,6 +275,7 @@ def _set_context(
         self.context.n_predict = n_predict
         self.context.top_k = top_k
         self.context.top_p = top_p
+        self.context.min_p = min_p
         self.context.temp = temp
         self.context.n_batch = n_batch
         self.context.repeat_penalty = repeat_penalty
@@ -297,6 +301,7 @@ def prompt_model(
         n_predict: int = 4096,
         top_k: int = 40,
         top_p: float = 0.9,
+        min_p: float = 0.0,
         temp: float = 0.1,
         n_batch: int = 8,
         repeat_penalty: float = 1.2,
@@ -334,6 +339,7 @@ def prompt_model(
             n_predict=n_predict,
             top_k=top_k,
             top_p=top_p,
+            min_p=min_p,
             temp=temp,
             n_batch=n_batch,
             repeat_penalty=repeat_penalty,

gpt4all-bindings/python/gpt4all/gpt4all.py

@@ -289,6 +289,7 @@ def generate(
         temp: float = 0.7,
         top_k: int = 40,
         top_p: float = 0.4,
+        min_p: float = 0.0,
         repeat_penalty: float = 1.18,
         repeat_last_n: int = 64,
         n_batch: int = 8,
@@ -305,6 +306,7 @@ def generate(
             temp: The model temperature. Larger values increase creativity but decrease factuality.
             top_k: Randomly sample from the top_k most likely tokens at each generation step. Set this to 1 for greedy decoding.
             top_p: Randomly sample at each generation step from the top most likely tokens whose probabilities add up to top_p.
+            min_p: Randomly sample at each generation step from the top most likely tokens whose probabilities are at least min_p.
             repeat_penalty: Penalize the model for repetition. Higher values result in less repetition.
             repeat_last_n: How far in the models generation history to apply the repeat penalty.
             n_batch: Number of prompt tokens processed in parallel. Larger values decrease latency but increase resource requirements.
@@ -325,6 +327,7 @@ def generate(
             temp=temp,
             top_k=top_k,
             top_p=top_p,
+            min_p=min_p,
             repeat_penalty=repeat_penalty,
             repeat_last_n=repeat_last_n,
             n_batch=n_batch,

gpt4all-bindings/typescript/index.cc

@@ -248,6 +248,7 @@ Napi::Value NodeModelWrapper::GetRequiredMemory(const Napi::CallbackInfo& info)
              .n_predict = 128,
              .top_k = 40,
              .top_p = 0.9f,
+             .min_p = 0.0f,
              .temp = 0.72f,
              .n_batch = 8,
              .repeat_penalty = 1.0f,
@@ -277,6 +278,8 @@ Napi::Value NodeModelWrapper::GetRequiredMemory(const Napi::CallbackInfo& info)
             promptContext.top_k = inputObject.Get("top_k").As<Napi::Number>().Int32Value();
        if(inputObject.Has("top_p")) 
             promptContext.top_p = inputObject.Get("top_p").As<Napi::Number>().FloatValue();
+       if(inputObject.Has("min_p")) 
+            promptContext.min_p = inputObject.Get("min_p").As<Napi::Number>().FloatValue();
        if(inputObject.Has("temp")) 
             promptContext.temp = inputObject.Get("temp").As<Napi::Number>().FloatValue();
        if(inputObject.Has("n_batch")) 

gpt4all-chat/chatllm.cpp

@@ -568,16 +568,17 @@ bool ChatLLM::prompt(const QList<QString> &collectionList, const QString &prompt
     const int32_t n_predict = MySettings::globalInstance()->modelMaxLength(m_modelInfo);
     const int32_t top_k = MySettings::globalInstance()->modelTopK(m_modelInfo);
     const float top_p = MySettings::globalInstance()->modelTopP(m_modelInfo);
+    const float min_p = MySettings::globalInstance()->modelMinP(m_modelInfo);
     const float temp = MySettings::globalInstance()->modelTemperature(m_modelInfo);
     const int32_t n_batch = MySettings::globalInstance()->modelPromptBatchSize(m_modelInfo);
     const float repeat_penalty = MySettings::globalInstance()->modelRepeatPenalty(m_modelInfo);
     const int32_t repeat_penalty_tokens = MySettings::globalInstance()->modelRepeatPenaltyTokens(m_modelInfo);
-    return promptInternal(collectionList, prompt, promptTemplate, n_predict, top_k, top_p, temp, n_batch,
+    return promptInternal(collectionList, prompt, promptTemplate, n_predict, top_k, top_p, min_p, temp, n_batch,
         repeat_penalty, repeat_penalty_tokens);
 }
 
 bool ChatLLM::promptInternal(const QList<QString> &collectionList, const QString &prompt, const QString &promptTemplate,
-    int32_t n_predict, int32_t top_k, float top_p, float temp, int32_t n_batch, float repeat_penalty,
+    int32_t n_predict, int32_t top_k, float top_p, float min_p, float temp, int32_t n_batch, float repeat_penalty,
     int32_t repeat_penalty_tokens)
 {
     if (!isModelLoaded())
@@ -608,6 +609,7 @@ bool ChatLLM::promptInternal(const QList<QString> &collectionList, const QString
     m_ctx.n_predict = n_predict;
     m_ctx.top_k = top_k;
     m_ctx.top_p = top_p;
+    m_ctx.min_p = min_p;
     m_ctx.temp = temp;
     m_ctx.n_batch = n_batch;
     m_ctx.repeat_penalty = repeat_penalty;
@@ -1020,6 +1022,7 @@ void ChatLLM::processSystemPrompt()
     const int32_t n_predict = MySettings::globalInstance()->modelMaxLength(m_modelInfo);
     const int32_t top_k = MySettings::globalInstance()->modelTopK(m_modelInfo);
     const float top_p = MySettings::globalInstance()->modelTopP(m_modelInfo);
+    const float min_p = MySettings::globalInstance()->modelMinP(m_modelInfo);
     const float temp = MySettings::globalInstance()->modelTemperature(m_modelInfo);
     const int32_t n_batch = MySettings::globalInstance()->modelPromptBatchSize(m_modelInfo);
     const float repeat_penalty = MySettings::globalInstance()->modelRepeatPenalty(m_modelInfo);
@@ -1028,6 +1031,7 @@ void ChatLLM::processSystemPrompt()
     m_ctx.n_predict = n_predict;
     m_ctx.top_k = top_k;
     m_ctx.top_p = top_p;
+    m_ctx.min_p = min_p;
     m_ctx.temp = temp;
     m_ctx.n_batch = n_batch;
     m_ctx.repeat_penalty = repeat_penalty;
@@ -1067,6 +1071,7 @@ void ChatLLM::processRestoreStateFromText()
     const int32_t n_predict = MySettings::globalInstance()->modelMaxLength(m_modelInfo);
     const int32_t top_k = MySettings::globalInstance()->modelTopK(m_modelInfo);
     const float top_p = MySettings::globalInstance()->modelTopP(m_modelInfo);
+    const float min_p = MySettings::globalInstance()->modelMinP(m_modelInfo);
     const float temp = MySettings::globalInstance()->modelTemperature(m_modelInfo);
     const int32_t n_batch = MySettings::globalInstance()->modelPromptBatchSize(m_modelInfo);
     const float repeat_penalty = MySettings::globalInstance()->modelRepeatPenalty(m_modelInfo);
@@ -1075,6 +1080,7 @@ void ChatLLM::processRestoreStateFromText()
     m_ctx.n_predict = n_predict;
     m_ctx.top_k = top_k;
     m_ctx.top_p = top_p;
+    m_ctx.min_p = min_p;
     m_ctx.temp = temp;
     m_ctx.n_batch = n_batch;
     m_ctx.repeat_penalty = repeat_penalty;

gpt4all-chat/chatllm.h

@@ -139,7 +139,7 @@ public Q_SLOTS:
 
 protected:
     bool promptInternal(const QList<QString> &collectionList, const QString &prompt, const QString &promptTemplate,
-        int32_t n_predict, int32_t top_k, float top_p, float temp, int32_t n_batch, float repeat_penalty,
+        int32_t n_predict, int32_t top_k, float top_p, float min_p, float temp, int32_t n_batch, float repeat_penalty,
         int32_t repeat_penalty_tokens);
     bool handlePrompt(int32_t token);
     bool handleResponse(int32_t token, const std::string &response);