dearwolf/LocalAI
1
0
Fork 0
mirror of https://github.com/mudler/LocalAI.git synced 2025-05-28 14:35:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 12

feat(template): read jinja templates from gguf files (#4332)

Browse source 
* Read jinja templates as fallback

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Move templating out of model loader

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Test TemplateMessages

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Set role and content from transformers

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Tests: be more flexible

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* More jinja

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Small refactoring and adaptations

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
This commit is contained in:
Ettore Di Giacinto 2024-12-08 13:50:33 +01:00 • committed by GitHub
parent f5e1527a5a
commit cea5a0ea42
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
23 changed files with 971 additions and 785 deletions
Download patch file Download diff file Expand all files Collapse all files

146
core/http/endpoints/openai/chat.go
View file

@ -14,6 +14,8 @@ import (
"github.com/mudler/LocalAI/core/config"
"github.com/mudler/LocalAI/core/schema"
"github.com/mudler/LocalAI/pkg/functions"
"github.com/mudler/LocalAI/pkg/templates"
model "github.com/mudler/LocalAI/pkg/model"
"github.com/rs/zerolog/log"
"github.com/valyala/fasthttp"
@ -24,7 +26,7 @@ import (
// @Param request body schema.OpenAIRequest true "query params"
// @Success 200 {object} schema.OpenAIResponse "Response"
// @Router /v1/chat/completions [post]
func ChatEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, startupOptions *config.ApplicationConfig) func(c *fiber.Ctx) error {
func ChatEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, evaluator *templates.Evaluator, startupOptions *config.ApplicationConfig) func(c *fiber.Ctx) error {
var id, textContentToReturn string
var created int
@ -298,148 +300,10 @@ func ChatEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, startup
// If we are using the tokenizer template, we don't need to process the messages
// unless we are processing functions
if !config.TemplateConfig.UseTokenizerTemplate || shouldUseFn {
suppressConfigSystemPrompt := false
mess := []string{}
for messageIndex, i := range input.Messages {
var content string
role := i.Role
// if function call, we might want to customize the role so we can display better that the "assistant called a json action"
// if an "assistant_function_call" role is defined, we use it, otherwise we use the role that is passed by in the request
if (i.FunctionCall != nil || i.ToolCalls != nil) && i.Role == "assistant" {
roleFn := "assistant_function_call"
r := config.Roles[roleFn]
if r != "" {
role = roleFn
}
}
r := config.Roles[role]
contentExists := i.Content != nil && i.StringContent != ""
fcall := i.FunctionCall
if len(i.ToolCalls) > 0 {
fcall = i.ToolCalls
}
// First attempt to populate content via a chat message specific template
if config.TemplateConfig.ChatMessage != "" {
chatMessageData := model.ChatMessageTemplateData{
SystemPrompt: config.SystemPrompt,
Role: r,
RoleName: role,
Content: i.StringContent,
FunctionCall: fcall,
FunctionName: i.Name,
LastMessage: messageIndex == (len(input.Messages) - 1),
Function: config.Grammar != "" && (messageIndex == (len(input.Messages) - 1)),
MessageIndex: messageIndex,
}
templatedChatMessage, err := ml.EvaluateTemplateForChatMessage(config.TemplateConfig.ChatMessage, chatMessageData)
if err != nil {
log.Error().Err(err).Interface("message", chatMessageData).Str("template", config.TemplateConfig.ChatMessage).Msg("error processing message with template, skipping")
} else {
if templatedChatMessage == "" {
log.Warn().Msgf("template \"%s\" produced blank output for %+v. Skipping!", config.TemplateConfig.ChatMessage, chatMessageData)
continue // TODO: This continue is here intentionally to skip over the line `mess = append(mess, content)` below, and to prevent the sprintf
}
log.Debug().Msgf("templated message for chat: %s", templatedChatMessage)
content = templatedChatMessage
}
}
marshalAnyRole := func(f any) {
j, err := json.Marshal(f)
if err == nil {
if contentExists {
content += "\n" + fmt.Sprint(r, " ", string(j))
} else {
content = fmt.Sprint(r, " ", string(j))
}
}
}
marshalAny := func(f any) {
j, err := json.Marshal(f)
if err == nil {
if contentExists {
content += "\n" + string(j)
} else {
content = string(j)
}
}
}
// If this model doesn't have such a template, or if that template fails to return a value, template at the message level.
if content == "" {
if r != "" {
if contentExists {
content = fmt.Sprint(r, i.StringContent)
}
if i.FunctionCall != nil {
marshalAnyRole(i.FunctionCall)
}
if i.ToolCalls != nil {
marshalAnyRole(i.ToolCalls)
}
} else {
if contentExists {
content = fmt.Sprint(i.StringContent)
}
if i.FunctionCall != nil {
marshalAny(i.FunctionCall)
}
if i.ToolCalls != nil {
marshalAny(i.ToolCalls)
}
}
// Special Handling: System. We care if it was printed at all, not the r branch, so check seperately
if contentExists && role == "system" {
suppressConfigSystemPrompt = true
}
}
mess = append(mess, content)
}
joinCharacter := "\n"
if config.TemplateConfig.JoinChatMessagesByCharacter != nil {
joinCharacter = *config.TemplateConfig.JoinChatMessagesByCharacter
}
predInput = strings.Join(mess, joinCharacter)
log.Debug().Msgf("Prompt (before templating): %s", predInput)
templateFile := ""
// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
if ml.ExistsInModelPath(fmt.Sprintf("%s.tmpl", config.Model)) {
templateFile = config.Model
}
if config.TemplateConfig.Chat != "" && !shouldUseFn {
templateFile = config.TemplateConfig.Chat
}
if config.TemplateConfig.Functions != "" && shouldUseFn {
templateFile = config.TemplateConfig.Functions
}
if templateFile != "" {
templatedInput, err := ml.EvaluateTemplateForPrompt(model.ChatPromptTemplate, templateFile, model.PromptTemplateData{
SystemPrompt: config.SystemPrompt,
SuppressSystemPrompt: suppressConfigSystemPrompt,
Input: predInput,
Functions: funcs,
})
if err == nil {
predInput = templatedInput
log.Debug().Msgf("Template found, input modified to: %s", predInput)
} else {
log.Debug().Msgf("Template failed loading: %s", err.Error())
}
}
predInput = evaluator.TemplateMessages(input.Messages, config, funcs, shouldUseFn)
log.Debug().Msgf("Prompt (after templating): %s", predInput)
if shouldUseFn && config.Grammar != "" {
if config.Grammar != "" {
log.Debug().Msgf("Grammar: %+v", config.Grammar)
}
}

47
core/http/endpoints/openai/completion.go
View file

@ -16,6 +16,7 @@ import (
"github.com/mudler/LocalAI/core/schema"
"github.com/mudler/LocalAI/pkg/functions"
model "github.com/mudler/LocalAI/pkg/model"
"github.com/mudler/LocalAI/pkg/templates"
"github.com/rs/zerolog/log"
"github.com/valyala/fasthttp"
)
@ -25,7 +26,7 @@ import (
// @Param request body schema.OpenAIRequest true "query params"
// @Success 200 {object} schema.OpenAIResponse "Response"
// @Router /v1/completions [post]
func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, evaluator *templates.Evaluator, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
id := uuid.New().String()
created := int(time.Now().Unix())
@ -94,17 +95,6 @@ func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, a
c.Set("Transfer-Encoding", "chunked")
}
templateFile := ""
// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
if ml.ExistsInModelPath(fmt.Sprintf("%s.tmpl", config.Model)) {
templateFile = config.Model
}
if config.TemplateConfig.Completion != "" {
templateFile = config.TemplateConfig.Completion
}
if input.Stream {
if len(config.PromptStrings) > 1 {
return errors.New("cannot handle more than 1 `PromptStrings` when Streaming")
@ -112,15 +102,13 @@ func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, a
predInput := config.PromptStrings[0]
if templateFile != "" {
templatedInput, err := ml.EvaluateTemplateForPrompt(model.CompletionPromptTemplate, templateFile, model.PromptTemplateData{
Input: predInput,
SystemPrompt: config.SystemPrompt,
})
if err == nil {
predInput = templatedInput
log.Debug().Msgf("Template found, input modified to: %s", predInput)
}
templatedInput, err := evaluator.EvaluateTemplateForPrompt(templates.CompletionPromptTemplate, *config, templates.PromptTemplateData{
Input: predInput,
SystemPrompt: config.SystemPrompt,
})
if err == nil {
predInput = templatedInput
log.Debug().Msgf("Template found, input modified to: %s", predInput)
}
responses := make(chan schema.OpenAIResponse)
@ -165,16 +153,13 @@ func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, a
totalTokenUsage := backend.TokenUsage{}
for k, i := range config.PromptStrings {
if templateFile != "" {
// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
templatedInput, err := ml.EvaluateTemplateForPrompt(model.CompletionPromptTemplate, templateFile, model.PromptTemplateData{
SystemPrompt: config.SystemPrompt,
Input: i,
})
if err == nil {
i = templatedInput
log.Debug().Msgf("Template found, input modified to: %s", i)
}
templatedInput, err := evaluator.EvaluateTemplateForPrompt(templates.CompletionPromptTemplate, *config, templates.PromptTemplateData{
SystemPrompt: config.SystemPrompt,
Input: i,
})
if err == nil {
i = templatedInput
log.Debug().Msgf("Template found, input modified to: %s", i)
}
r, tokenUsage, err := ComputeChoices(

33
core/http/endpoints/openai/edit.go
View file

@ -12,6 +12,7 @@ import (
"github.com/google/uuid"
"github.com/mudler/LocalAI/core/schema"
model "github.com/mudler/LocalAI/pkg/model"
"github.com/mudler/LocalAI/pkg/templates"
"github.com/rs/zerolog/log"
)
@ -21,7 +22,8 @@ import (
// @Param request body schema.OpenAIRequest true "query params"
// @Success 200 {object} schema.OpenAIResponse "Response"
// @Router /v1/edits [post]
func EditEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
func EditEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, evaluator *templates.Evaluator, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
return func(c *fiber.Ctx) error {
modelFile, input, err := readRequest(c, cl, ml, appConfig, true)
if err != nil {
@ -35,31 +37,18 @@ func EditEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConf
log.Debug().Msgf("Parameter Config: %+v", config)
templateFile := ""
// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
if ml.ExistsInModelPath(fmt.Sprintf("%s.tmpl", config.Model)) {
templateFile = config.Model
}
if config.TemplateConfig.Edit != "" {
templateFile = config.TemplateConfig.Edit
}
var result []schema.Choice
totalTokenUsage := backend.TokenUsage{}
for _, i := range config.InputStrings {
if templateFile != "" {
templatedInput, err := ml.EvaluateTemplateForPrompt(model.EditPromptTemplate, templateFile, model.PromptTemplateData{
Input: i,
Instruction: input.Instruction,
SystemPrompt: config.SystemPrompt,
})
if err == nil {
i = templatedInput
log.Debug().Msgf("Template found, input modified to: %s", i)
}
templatedInput, err := evaluator.EvaluateTemplateForPrompt(templates.EditPromptTemplate, *config, templates.PromptTemplateData{
Input: i,
Instruction: input.Instruction,
SystemPrompt: config.SystemPrompt,
})
if err == nil {
i = templatedInput
log.Debug().Msgf("Template found, input modified to: %s", i)
}
r, tokenUsage, err := ComputeChoices(input, i, config, appConfig, ml, func(s string, c *[]schema.Choice) {