github/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2025-09-06 19:10:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

add min_p sampling parameter (#2014)

Browse Source 
Signed-off-by: Christopher Barrera <cb@arda.tx.rr.com>
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
This commit is contained in:
chrisbarrera
2024-02-24 16:51:34 -06:00
committed by GitHub
parent a153cc5b25
commit f8b1069a1c
28 changed files with 176 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
gpt4all-bindings/cli/app.py
View File

@@ -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,

9
gpt4all-bindings/csharp/Gpt4All/Bindings/LLPromptContext.cs
View File

@@ -64,6 +64,15 @@ public unsafe class LLModelPromptContext
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>

2
gpt4all-bindings/csharp/Gpt4All/Bindings/NativeMethods.cs
View File

@@ -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")]

1
gpt4all-bindings/csharp/Gpt4All/Extensions/LLPromptContextExtensions.cs
View File

@@ -16,6 +16,7 @@ internal static class LLPromptContextExtensions
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}

1
gpt4all-bindings/csharp/Gpt4All/Extensions/PredictRequestOptionsExtensions.cs
View File

@@ -12,6 +12,7 @@ public static class PredictRequestOptionsExtensions
TokensSize = opts.TokensSize,
TopK = opts.TopK,
TopP = opts.TopP,
MinP = opts.MinP,
PastNum = opts.PastConversationTokensNum,
RepeatPenalty = opts.RepeatPenalty,
Temperature = opts.Temperature,

2
gpt4all-bindings/csharp/Gpt4All/Prediction/PredictRequestOptions.cs
View File

@@ -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;

4
gpt4all-bindings/golang/binding.cpp
View File

@@ -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;

4
gpt4all-bindings/golang/binding.h
View File

@@ -7,7 +7,7 @@ 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);
@@ -15,4 +15,4 @@ extern unsigned char getTokenCallback(void *, char *);
#ifdef __cplusplus
}
#endif
#endif

4
gpt4all-bindings/golang/gpt4all.go
View File

@@ -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, " ")

10
gpt4all-bindings/golang/options.go
View File

@@ -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) {

6
gpt4all-bindings/java/src/main/java/com/hexadevlabs/gpt4all/LLModel.java
View File

@@ -32,6 +32,7 @@ public class LLModel implements AutoCloseable {
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 class LLModel implements AutoCloseable {
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;

1
gpt4all-bindings/java/src/main/java/com/hexadevlabs/gpt4all/LLModelLibrary.java
View File

@@ -48,6 +48,7 @@ public interface LLModelLibrary {
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();

6
gpt4all-bindings/python/gpt4all/_pyllmodel.py
View File

@@ -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 @@ class LLModel:
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 @@ class LLModel:
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 @@ class LLModel:
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 @@ class LLModel:
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 @@ class LLModel:
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,

3
gpt4all-bindings/python/gpt4all/gpt4all.py
View File

@@ -289,6 +289,7 @@ class GPT4All:
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 @@ class GPT4All:
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 @@ class GPT4All:
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,

3
gpt4all-bindings/typescript/index.cc
View File

@@ -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"))