github/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2025-07-05 11:36:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

Backend prompt dedup (#822)

Browse Source 
* Deduplicated prompt() function code
This commit is contained in:
AT 2023-06-04 08:59:24 -04:00 committed by GitHub
parent e53195a002
commit b5971b0d41
10 changed files with 286 additions and 457 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
gpt4all-backend/CMakeLists.txt
View File

@ -103,7 +103,7 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
endforeach()
add_library(llmodel
llmodel.h llmodel.cpp
llmodel.h llmodel.cpp llmodel_shared.cpp
llmodel_c.h llmodel_c.cpp
dlhandle.h
)

169
gpt4all-backend/gptj.cpp
View File

@ -890,91 +890,15 @@ size_t GPTJ::restoreState(const uint8_t *src)
return gptj_set_state_data(d_ptr->model, &d_ptr->rng, src);
}
void GPTJ::prompt(const std::string &prompt,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &promptCtx) {
if (!isModelLoaded()) {
std::cerr << "GPT-J ERROR: prompt won't work with an unloaded model!\n";
return;
std::vector<LLModel::Token> GPTJ::tokenize(const std::string &str) const
{
return ::gpt_tokenize(d_ptr->vocab, str);
}
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(d_ptr->vocab, prompt);
// save the context size
promptCtx.n_ctx = d_ptr->model->hparams.n_ctx;
if ((int) embd_inp.size() > promptCtx.n_ctx - 4) {
responseCallback(-1, "ERROR: The prompt size exceeds the context window size and cannot be processed.");
std::cerr << "GPT-J ERROR: The prompt is" << embd_inp.size() <<
"tokens and the context window is" << promptCtx.n_ctx << "!\n";
return;
}
promptCtx.n_predict = std::min(promptCtx.n_predict, promptCtx.n_ctx - (int) embd_inp.size());
promptCtx.n_past = std::min(promptCtx.n_past, promptCtx.n_ctx);
// determine the required inference memory per token:
static bool initialized = false;
static std::vector<gpt_vocab::id> p_instruct;
static std::vector<gpt_vocab::id> r_instruct;
if (!initialized) {
gptj_eval(*d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, promptCtx.logits,
d_ptr->mem_per_token);
initialized = true;
}
// process the prompt in batches
size_t i = 0;
while (i < embd_inp.size()) {
size_t batch_end = std::min(i + promptCtx.n_batch, embd_inp.size());
std::vector<gpt_vocab::id> batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
// Check if the context has run out...
if (promptCtx.n_past + int32_t(batch.size()) > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << "GPTJ: reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, batch)) {
std::cerr << "GPT-J ERROR: Failed to process prompt\n";
return;
}
size_t tokens = batch_end - i;
for (size_t t = 0; t < tokens; ++t) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(batch.at(t));
if (!promptCallback(batch.at(t)))
return;
}
promptCtx.n_past += batch.size();
i = batch_end;
}
std::string cachedResponse;
std::vector<gpt_vocab::id> cachedTokens;
std::unordered_set<std::string> reversePrompts
= { "### Instruction", "### Prompt", "### Response", "### Human", "### Assistant", "### Context" };
// predict next tokens
for (int i = 0; i < promptCtx.n_predict; i++) {
// sample next token
const int n_vocab = d_ptr->model->hparams.n_vocab;
gpt_vocab::id id = 0;
LLModel::Token GPTJ::sampleToken(PromptContext &promptCtx) const
{
const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
id = gpt_sample_top_k_top_p(n_vocab,
return gpt_sample_top_k_top_p(d_ptr->model->hparams.n_vocab,
promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
n_prev_toks,
promptCtx.logits,
@ -983,68 +907,35 @@ void GPTJ::prompt(const std::string &prompt,
d_ptr->rng);
}
// Check if the context has run out...
if (promptCtx.n_past + 1 > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << "GPTJ: reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + 1 <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, { id })) {
std::cerr << "GPT-J ERROR: Failed to predict next token\n";
return;
}
promptCtx.n_past += 1;
// display text
if (id == 50256 /*end of text*/)
return;
const std::string str = d_ptr->vocab.id_to_token[id];
// Check if the provided str is part of our reverse prompts
bool foundPartialReversePrompt = false;
const std::string completed = cachedResponse + str;
if (reversePrompts.find(completed) != reversePrompts.end())
return;
// Check if it partially matches our reverse prompts and if so, cache
for (const auto &s : reversePrompts) {
if (s.compare(0, completed.size(), completed) == 0) {
foundPartialReversePrompt = true;
cachedResponse = completed;
break;
}
}
// Regardless the token gets added to our cache
cachedTokens.push_back(id);
// Continue if we have found a partial match
if (foundPartialReversePrompt)
continue;
// Empty the cache
for (auto t : cachedTokens) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(t);
if (!responseCallback(t, d_ptr->vocab.id_to_token[t]))
return;
}
cachedTokens.clear();
}
}
bool GPTJ::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens)
std::string_view GPTJ::tokenToString(Token id) const
{
return d_ptr->vocab.id_to_token[id];
}
bool GPTJ::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
{
// determine the required inference memory per token:
static bool initialized = false;
if (!initialized) {
gptj_eval(*d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, ctx.logits,
d_ptr->mem_per_token);
initialized = true;
}
return gptj_eval(*d_ptr->model, d_ptr->n_threads, ctx.n_past, tokens, ctx.logits, d_ptr->mem_per_token);
}
int32_t GPTJ::contextLength() const
{
return d_ptr->model->hparams.n_ctx;
}
const std::vector<LLModel::Token> &GPTJ::endTokens() const
{
static const std::vector<LLModel::Token> fres = {50256};
return fres;
}
#if defined(_WIN32)
#define DLL_EXPORT __declspec(dllexport)
#else

14
gpt4all-backend/gptj_impl.h
View File

@ -20,17 +20,19 @@ public:
size_t stateSize() const override;
size_t saveState(uint8_t *dest) const override;
size_t restoreState(const uint8_t *src) override;
void prompt(const std::string &prompt,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &ctx) override;
bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) override;
void setThreadCount(int32_t n_threads) override;
int32_t threadCount() const override;
private:
GPTJPrivate *d_ptr;
protected:
std::vector<Token> tokenize(const std::string&) const override;
Token sampleToken(PromptContext &ctx) const override;
std::string_view tokenToString(Token) const override;
bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
int32_t contextLength() const override;
const std::vector<Token>& endTokens() const override;
};
#endif // GPTJ_H

162
gpt4all-backend/llamamodel.cpp
View File

@ -163,157 +163,45 @@ size_t LLamaModel::restoreState(const uint8_t *src)
return llama_set_state_data(d_ptr->ctx, const_cast<uint8_t*>(src));
}
void LLamaModel::prompt(const std::string &prompt,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &promptCtx) {
if (!isModelLoaded()) {
std::cerr << "LLAMA ERROR: prompt won't work with an unloaded model!\n";
return;
std::vector<LLModel::Token> LLamaModel::tokenize(const std::string &str) const
{
std::vector<LLModel::Token> fres(str.size()+4);
auto fres_len = llama_tokenize(d_ptr->ctx, str.c_str(), fres.data(), fres.size(), d_ptr->empty);
fres.resize(fres_len);
return fres;
}
gpt_params params;
params.prompt = prompt;
// Add a space in front of the first character to match OG llama tokenizer behavior
params.prompt.insert(0, 1, ' ');
// tokenize the prompt
std::vector<llama_token> embd_inp(params.prompt.size() + 4);
int n = llama_tokenize(d_ptr->ctx, params.prompt.c_str(), embd_inp.data(), embd_inp.size(), d_ptr->empty);
assert(n >= 0);
embd_inp.resize(n);
d_ptr->empty = false;
// save the context size
promptCtx.n_ctx = llama_n_ctx(d_ptr->ctx);
if ((int) embd_inp.size() > promptCtx.n_ctx - 4) {
responseCallback(-1, "The prompt size exceeds the context window size and cannot be processed.");
std::cerr << "LLAMA ERROR: The prompt is" << embd_inp.size() <<
"tokens and the context window is" << promptCtx.n_ctx << "!\n";
return;
std::string_view LLamaModel::tokenToString(Token id) const
{
return llama_token_to_str(d_ptr->ctx, id);
}
promptCtx.n_predict = std::min(promptCtx.n_predict, promptCtx.n_ctx - (int) embd_inp.size());
promptCtx.n_past = std::min(promptCtx.n_past, promptCtx.n_ctx);
// number of tokens to keep when resetting context
params.n_keep = (int)embd_inp.size();
// process the prompt in batches
size_t i = 0;
while (i < embd_inp.size()) {
size_t batch_end = std::min(i + promptCtx.n_batch, embd_inp.size());
std::vector<llama_token> batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
// Check if the context has run out...
if (promptCtx.n_past + int32_t(batch.size()) > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << "LLAMA: reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, batch)) {
std::cerr << "LLAMA ERROR: Failed to process prompt\n";
return;
}
size_t tokens = batch_end - i;
for (size_t t = 0; t < tokens; ++t) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(batch.at(t));
if (!promptCallback(batch.at(t)))
return;
}
promptCtx.n_past += batch.size();
i = batch_end;
}
std::string cachedResponse;
std::vector<llama_token> cachedTokens;
std::unordered_set<std::string> reversePrompts
= { "### Instruction", "### Prompt", "### Response", "### Human", "### Assistant" };
// predict next tokens
for (int i = 0; i < promptCtx.n_predict; i++) {
// sample next token
LLModel::Token LLamaModel::sampleToken(PromptContext &promptCtx) const
{
const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
llama_token id = llama_sample_top_p_top_k(d_ptr->ctx,
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,
promptCtx.repeat_penalty);
// Check if the context has run out...
if (promptCtx.n_past + 1 > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << "LLAMA: reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + 1 <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, { id })) {
std::cerr << "LLAMA ERROR: Failed to predict next token\n";
return;
}
promptCtx.n_past += 1;
// display text
if (id == llama_token_eos())
return;
const std::string str = llama_token_to_str(d_ptr->ctx, id);
// Check if the provided str is part of our reverse prompts
bool foundPartialReversePrompt = false;
const std::string completed = cachedResponse + str;
if (reversePrompts.find(completed) != reversePrompts.end()) {
return;
}
// Check if it partially matches our reverse prompts and if so, cache
for (const auto &s : reversePrompts) {
if (s.compare(0, completed.size(), completed) == 0) {
foundPartialReversePrompt = true;
cachedResponse = completed;
break;
}
}
// Regardless the token gets added to our cache
cachedTokens.push_back(id);
// Continue if we have found a partial match
if (foundPartialReversePrompt)
continue;
// Empty the cache
for (auto t : cachedTokens) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(t);
if (!responseCallback(t, llama_token_to_str(d_ptr->ctx, t)))
return;
}
cachedTokens.clear();
}
}
bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens)
bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
{
d_ptr->empty = false;
return llama_eval(d_ptr->ctx, tokens.data(), tokens.size(), ctx.n_past, d_ptr->n_threads) == 0;
}
int32_t LLamaModel::contextLength() const
{
return llama_n_ctx(d_ptr->ctx);
}
const std::vector<LLModel::Token> &LLamaModel::endTokens() const
{
static const std::vector<LLModel::Token> fres = {llama_token_eos()};
return fres;
}
#if defined(_WIN32)
#define DLL_EXPORT __declspec(dllexport)
#else

14
gpt4all-backend/llamamodel_impl.h
View File

@ -20,17 +20,19 @@ public:
size_t stateSize() const override;
size_t saveState(uint8_t *dest) const override;
size_t restoreState(const uint8_t *src) override;
void prompt(const std::string &prompt,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &ctx) override;
bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) override;
void setThreadCount(int32_t n_threads) override;
int32_t threadCount() const override;
private:
LLamaPrivate *d_ptr;
protected:
std::vector<Token> tokenize(const std::string&) const override;
std::string_view tokenToString(Token) const override;
Token sampleToken(PromptContext& ctx) const override;
bool evalTokens(PromptContext& ctx, const std::vector<int32_t> &tokens) const override;
int32_t contextLength() const override;
const std::vector<Token>& endTokens() const override;
};
#endif // LLAMAMODEL_H

21
gpt4all-backend/llmodel.h
View File

@ -7,11 +7,14 @@
#include <string_view>
#include <fstream>
#include <cstdint>
#include <limits>
class Dlhandle;
class LLModel {
public:
using Token = int32_t;
class Implementation {
LLModel *(*construct_)();
@ -63,8 +66,8 @@ public:
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &ctx) = 0;
virtual bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) = 0;
PromptContext &ctx);
virtual void setThreadCount(int32_t /*n_threads*/) {}
virtual int32_t threadCount() const { return 1; }
@ -84,10 +87,20 @@ public:
}
protected:
const Implementation *m_implementation = nullptr;
// These are pure virtual because subclasses need to implement as the default implementation of
// 'prompt' above calls these functions
virtual std::vector<Token> tokenize(const std::string&) const = 0;
virtual std::string_view tokenToString(Token) const = 0;
virtual Token sampleToken(PromptContext &ctx) const = 0;
virtual bool evalTokens(PromptContext &/*ctx*/, const std::vector<int32_t>& /*tokens*/) const = 0;
virtual int32_t contextLength() const = 0;
virtual const std::vector<Token>& endTokens() const = 0;
// This is a helper function called from the default implementation of 'prompt' but it can be
// shared by all base classes so it isn't virtual
void recalculateContext(PromptContext &promptCtx, std::function<bool(bool)> recalculate);
static std::string m_implementations_search_path;
const Implementation *m_implementation = nullptr;
static std::string m_implementations_search_path;
};
#endif // LLMODEL_H

133
gpt4all-backend/llmodel_shared.cpp
View File

@ -2,6 +2,7 @@
#include <cassert>
#include <iostream>
#include <unordered_set>
void LLModel::recalculateContext(PromptContext &promptCtx, std::function<bool(bool)> recalculate) {
size_t i = 0;
@ -24,3 +25,135 @@ void LLModel::recalculateContext(PromptContext &promptCtx, std::function<bool(bo
stop_generating:
recalculate(false);
}
void LLModel::prompt(const std::string &prompt,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &promptCtx)
{
if (!isModelLoaded()) {
std::cerr << implementation().modelType << " ERROR: prompt won't work with an unloaded model!\n";
return;
}
// tokenize the prompt
std::vector<Token> embd_inp = tokenize(prompt);
// save the context size
promptCtx.n_ctx = contextLength();
if ((int) embd_inp.size() > promptCtx.n_ctx - 4) {
responseCallback(-1, "ERROR: The prompt size exceeds the context window size and cannot be processed.");
std::cerr << implementation().modelType << " ERROR: The prompt is" << embd_inp.size() <<
"tokens and the context window is" << promptCtx.n_ctx << "!\n";
return;
}
promptCtx.n_predict = std::min(promptCtx.n_predict, promptCtx.n_ctx - (int) embd_inp.size());
promptCtx.n_past = std::min(promptCtx.n_past, promptCtx.n_ctx);
// process the prompt in batches
size_t i = 0;
while (i < embd_inp.size()) {
size_t batch_end = std::min(i + promptCtx.n_batch, embd_inp.size());
std::vector<Token> batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
// Check if the context has run out...
if (promptCtx.n_past + int32_t(batch.size()) > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << implementation().modelType << ": reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, batch)) {
std::cerr << implementation().modelType << " ERROR: Failed to process prompt\n";
return;
}
size_t tokens = batch_end - i;
for (size_t t = 0; t < tokens; ++t) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(batch.at(t));
if (!promptCallback(batch.at(t)))
return;
}
promptCtx.n_past += batch.size();
i = batch_end;
}
std::string cachedResponse;
std::vector<Token> cachedTokens;
std::unordered_set<std::string> reversePrompts
= { "### Instruction", "### Prompt", "### Response", "### Human", "### Assistant", "### Context" };
// predict next tokens
for (int i = 0; i < promptCtx.n_predict; i++) {
// sample next token
auto id = sampleToken(promptCtx);
// Check if the context has run out...
if (promptCtx.n_past + 1 > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << implementation().modelType << ": reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + 1 <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, { id })) {
std::cerr << implementation().modelType << " ERROR: Failed to predict next token\n";
return;
}
promptCtx.n_past += 1;
// display text
for (const auto token : endTokens()) {
if (id == token) return;
}
const std::string_view str = tokenToString(id);
// Check if the provided str is part of our reverse prompts
bool foundPartialReversePrompt = false;
const std::string completed = cachedResponse + std::string(str);
if (reversePrompts.find(completed) != reversePrompts.end())
return;
// Check if it partially matches our reverse prompts and if so, cache
for (const auto& s : reversePrompts) {
if (s.compare(0, completed.size(), completed) == 0) {
foundPartialReversePrompt = true;
cachedResponse = completed;
break;
}
}
// Regardless the token gets added to our cache
cachedTokens.push_back(id);
// Continue if we have found a partial match
if (foundPartialReversePrompt)
continue;
// Empty the cache
for (auto t : cachedTokens) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(t);
//TODO: Conversion to std::string can be avoided here...
if (!responseCallback(t, std::string(tokenToString(t))))
return;
}
cachedTokens.clear();
}
}

169
gpt4all-backend/mpt.cpp
View File

@ -815,91 +815,20 @@ size_t MPT::restoreState(const uint8_t *src)
return mpt_set_state_data(d_ptr->model, &d_ptr->rng, src);
}
void MPT::prompt(const std::string &prompt,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &promptCtx) {
if (!isModelLoaded()) {
std::cerr << "GPT-J ERROR: prompt won't work with an unloaded model!\n";
return;
std::vector<LLModel::Token> MPT::tokenize(const std::string &str) const
{
return ::gpt_tokenize(d_ptr->vocab, str);
}
// tokenize the prompt
std::vector<int> embd_inp = gpt_tokenize(d_ptr->vocab, prompt);
// save the context size
promptCtx.n_ctx = d_ptr->model->hparams.n_ctx;
if ((int) embd_inp.size() > promptCtx.n_ctx - 4) {
responseCallback(-1, "ERROR: The prompt size exceeds the context window size and cannot be processed.");
std::cerr << "GPT-J ERROR: The prompt is" << embd_inp.size() <<
"tokens and the context window is" << promptCtx.n_ctx << "!\n";
return;
std::string_view MPT::tokenToString(Token id) const
{
return d_ptr->vocab.id_to_token[id];
}
promptCtx.n_predict = std::min(promptCtx.n_predict, promptCtx.n_ctx - (int) embd_inp.size());
promptCtx.n_past = std::min(promptCtx.n_past, promptCtx.n_ctx);
// determine the required inference memory per token:
static bool initialized = false;
static std::vector<int> p_instruct;
static std::vector<int> r_instruct;
if (!initialized) {
mpt_eval(*d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, promptCtx.logits,
d_ptr->mem_per_token);
initialized = true;
}
// process the prompt in batches
size_t i = 0;
while (i < embd_inp.size()) {
size_t batch_end = std::min(i + promptCtx.n_batch, embd_inp.size());
std::vector<int> batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
// Check if the context has run out...
if (promptCtx.n_past + int32_t(batch.size()) > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << "MPT: reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, batch)) {
std::cerr << "GPT-J ERROR: Failed to process prompt\n";
return;
}
size_t tokens = batch_end - i;
for (size_t t = 0; t < tokens; ++t) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(batch.at(t));
if (!promptCallback(batch.at(t)))
return;
}
promptCtx.n_past += batch.size();
i = batch_end;
}
std::string cachedResponse;
std::vector<int> cachedTokens;
std::unordered_set<std::string> reversePrompts
= { "### Instruction", "### Prompt", "### Response", "### Human", "### Assistant", "### Context" };
// predict next tokens
for (int i = 0; i < promptCtx.n_predict; i++) {
// sample next token
const int n_vocab = d_ptr->model->hparams.n_vocab;
int id = 0;
LLModel::Token MPT::sampleToken(PromptContext &promptCtx) const
{
const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
id = gpt_sample_top_k_top_p(n_vocab,
return gpt_sample_top_k_top_p(d_ptr->model->hparams.n_vocab,
promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
n_prev_toks,
promptCtx.logits,
@ -908,72 +837,30 @@ void MPT::prompt(const std::string &prompt,
d_ptr->rng);
}
// Check if the context has run out...
if (promptCtx.n_past + 1 > promptCtx.n_ctx) {
const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
// Erase the first percentage of context from the tokens...
std::cerr << "MPT: reached the end of the context window so resizing\n";
promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
promptCtx.n_past = promptCtx.tokens.size();
recalculateContext(promptCtx, recalculateCallback);
assert(promptCtx.n_past + 1 <= promptCtx.n_ctx);
}
if (!evalTokens(promptCtx, { id })) {
std::cerr << "GPT-J ERROR: Failed to predict next token\n";
return;
}
promptCtx.n_past += 1;
// display tex
// mpt-7b-chat has special token for end
if (d_ptr->has_im_end && id == d_ptr->vocab.token_to_id["<|im_end|>"])
return;
if (id == 0 /*end of text*/)
return;
const std::string str = d_ptr->vocab.id_to_token[id];
// Check if the provided str is part of our reverse prompts
bool foundPartialReversePrompt = false;
const std::string completed = cachedResponse + str;
if (reversePrompts.find(completed) != reversePrompts.end())
return;
// Check if it partially matches our reverse prompts and if so, cache
for (const auto &s : reversePrompts) {
if (s.compare(0, completed.size(), completed) == 0) {
foundPartialReversePrompt = true;
cachedResponse = completed;
break;
}
}
// Regardless the token gets added to our cache
cachedTokens.push_back(id);
// Continue if we have found a partial match
if (foundPartialReversePrompt)
continue;
// Empty the cache
for (auto t : cachedTokens) {
if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
promptCtx.tokens.erase(promptCtx.tokens.begin());
promptCtx.tokens.push_back(t);
if (!responseCallback(t, d_ptr->vocab.id_to_token[t]))
return;
}
cachedTokens.clear();
}
}
bool MPT::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens)
bool MPT::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
{
// determine the required inference memory per token:
static bool initialized = false;
if (!initialized) {
mpt_eval(*d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, ctx.logits,
d_ptr->mem_per_token);
initialized = true;
}
return mpt_eval(*d_ptr->model, d_ptr->n_threads, ctx.n_past, tokens, ctx.logits, d_ptr->mem_per_token);
}
int32_t MPT::contextLength() const
{
return d_ptr->model->hparams.n_ctx;
}
const std::vector<LLModel::Token> &MPT::endTokens() const
{
static const std::vector<LLModel::Token> fres = {0, d_ptr->vocab.token_to_id["<|im_end|>"]};
return fres;
}
#if defined(_WIN32)
#define DLL_EXPORT __declspec(dllexport)
#else

14
gpt4all-backend/mpt_impl.h
View File

@ -20,17 +20,19 @@ public:
size_t stateSize() const override;
size_t saveState(uint8_t *dest) const override;
size_t restoreState(const uint8_t *src) override;
void prompt(const std::string &prompt,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &ctx) override;
bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) override;
void setThreadCount(int32_t n_threads) override;
int32_t threadCount() const override;
private:
MPTPrivate *d_ptr;
protected:
std::vector<Token> tokenize(const std::string&) const override;
std::string_view tokenToString(Token) const override;
Token sampleToken(PromptContext &ctx) const override;
bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
int32_t contextLength() const override;
const std::vector<Token>& endTokens() const override;
};
#endif // MPT_H

13
gpt4all-chat/chatgpt.h
View File

@ -24,7 +24,7 @@ public:
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &ctx) override;
bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) override { return true; }
void setThreadCount(int32_t n_threads) override;
int32_t threadCount() const override;
@ -34,6 +34,17 @@ public:
QList<QString> context() const { return m_context; }
void setContext(const QList<QString> &context) { m_context = context; }
protected:
// We have to implement these as they are pure virtual in base class, but we don't actually use
// them as they are only called from the default implementation of 'prompt' which we override and
// completely replace
std::vector<Token> tokenize(const std::string&) const override { return std::vector<Token>(); }
std::string_view tokenToString(Token) const override { return std::string_view(); }
Token sampleToken(PromptContext &ctx) const override { return -1; }
bool evalTokens(PromptContext &/*ctx*/, const std::vector<int32_t>& /*tokens*/) const override { return false; }
int32_t contextLength() const override { return -1; }
const std::vector<Token>& endTokens() const override { static const std::vector<Token> fres; return fres; }
private Q_SLOTS:
void handleFinished();
void handleReadyRead();