github/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2025-09-06 11:00:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix chat-style prompt templates (#1970)

Browse Source 
Also use a new version of Mistral OpenOrca.

Signed-off-by: Jared Van Bortel <jared@nomic.ai>
This commit is contained in:
Jared Van Bortel
2024-02-21 15:45:32 -05:00
committed by GitHub
parent b8f5c74f40
commit 4fc4d94be4
22 changed files with 429 additions and 307 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
gpt4all-backend/llmodel.h
View File

@@ -29,23 +29,23 @@ public:
class Implementation {
public:
Implementation(Dlhandle&&);
Implementation(const Implementation&) = delete;
Implementation(Implementation&&);
Implementation(Dlhandle &&);
Implementation(const Implementation &) = delete;
Implementation(Implementation &&);
~Implementation();
std::string_view modelType() const { return m_modelType; }
std::string_view buildVariant() const { return m_buildVariant; }
static bool isImplementation(const Dlhandle&);
static const std::vector<Implementation>& implementationList();
static const Implementation *implementation(const char *fname, const std::string& buildVariant);
static bool isImplementation(const Dlhandle &dl);
static const std::vector<Implementation> &implementationList();
static const Implementation *implementation(const char *fname, const std::string &buildVariant);
static LLModel *construct(const std::string &modelPath, std::string buildVariant = "auto", int n_ctx = 2048);
static std::vector<GPUDevice> availableGPUDevices();
static int32_t maxContextLength(const std::string &modelPath);
static int32_t layerCount(const std::string &modelPath);
static void setImplementationsSearchPath(const std::string& path);
static const std::string& implementationsSearchPath();
static void setImplementationsSearchPath(const std::string &path);
static const std::string &implementationsSearchPath();
private:
static LLModel *constructDefaultLlama();
@@ -82,26 +82,30 @@ public:
virtual bool supportsEmbedding() const = 0;
virtual bool supportsCompletion() const = 0;
virtual bool loadModel(const std::string &modelPath, int n_ctx, int ngl) = 0;
virtual bool isModelBlacklisted(const std::string &modelPath) { (void)modelPath; return false; };
virtual bool isModelLoaded() const = 0;
virtual size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) = 0;
virtual size_t stateSize() const { return 0; }
virtual size_t saveState(uint8_t */*dest*/) const { return 0; }
virtual size_t restoreState(const uint8_t */*src*/) { return 0; }
virtual size_t saveState(uint8_t *dest) const { (void)dest; return 0; }
virtual size_t restoreState(const uint8_t *src) { (void)src; return 0; }
// This method requires the model to return true from supportsCompletion otherwise it will throw
// an error
virtual void prompt(const std::string &prompt,
const std::string &promptTemplate,
std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &ctx);
PromptContext &ctx,
bool special = false,
std::string *fakeReply = nullptr);
virtual std::vector<float> embedding(const std::string &text);
virtual void setThreadCount(int32_t /*n_threads*/) {}
virtual void setThreadCount(int32_t n_threads) { (void)n_threads; }
virtual int32_t threadCount() const { return 1; }
const Implementation& implementation() const {
const Implementation &implementation() const {
return *m_implementation;
}
@@ -110,7 +114,7 @@ public:
return {};
}
virtual bool initializeGPUDevice(size_t memoryRequired, const std::string& name) const {
virtual bool initializeGPUDevice(size_t memoryRequired, const std::string &name) const {
(void)memoryRequired;
(void)name;
return false;
@@ -132,12 +136,13 @@ public:
protected:
// These are pure virtual because subclasses need to implement as the default implementation of
// 'prompt' above calls these functions
virtual std::vector<Token> tokenize(PromptContext &, const std::string&) const = 0;
virtual std::string tokenToString(Token) const = 0;
virtual std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special = false) const = 0;
virtual std::string tokenToString(Token id) const = 0;
virtual Token sampleToken(PromptContext &ctx) const = 0;
virtual bool evalTokens(PromptContext &/*ctx*/, const std::vector<int32_t>& /*tokens*/) 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;
virtual const std::vector<Token> &endTokens() const = 0;
virtual bool shouldAddBOS() const = 0;
virtual int32_t maxContextLength(std::string const &modelPath) const
{
@@ -166,6 +171,15 @@ protected:
return true;
}
void decodePrompt(std::function<bool(int32_t)> promptCallback,
std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &promptCtx,
std::vector<Token> embd_inp);
void generateResponse(std::function<bool(int32_t, const std::string&)> responseCallback,
std::function<bool(bool)> recalculateCallback,
PromptContext &promptCtx);
private:
friend class LLMImplementation;
};