Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

feat: include tokens usage for streamed output #4282

Merged
merged 3 commits into from
Nov 28, 2024
Merged

feat: include tokens usage for streamed output #4282

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
2931ea4
Use pb.Reply instead of []byte with Reply.GetMessage() in llama grpc …
mintyleaf Nov 27, 2024
8e8e05d
Merge branch 'master' into fix/stream_tokens_usage
mintyleaf Nov 27, 2024
e459118
Merge branch 'master' into fix/stream_tokens_usage
mudler Nov 28, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 10 additions & 2 deletions core/backend/llm.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,8 +117,12 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
ss := ""

var partialRune []byte
err := inferenceModel.PredictStream(ctx, opts, func(chars []byte) {
partialRune = append(partialRune, chars...)
err := inferenceModel.PredictStream(ctx, opts, func(reply *proto.Reply) {
msg := reply.GetMessage()
partialRune = append(partialRune, msg...)

tokenUsage.Prompt = int(reply.PromptTokens)
tokenUsage.Completion = int(reply.Tokens)

for len(partialRune) > 0 {
r, size := utf8.DecodeRune(partialRune)
Expand All @@ -132,6 +136,10 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im

partialRune = partialRune[size:]
}

if len(msg) == 0 {
tokenCallback("", tokenUsage)
}
})
return LLMResponse{
Response: ss,
Expand Down
9 changes: 8 additions & 1 deletion core/http/endpoints/openai/chat.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,11 +39,15 @@ func ChatEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, startup
responses <- initialMessage

ComputeChoices(req, s, config, startupOptions, loader, func(s string, c *[]schema.Choice) {}, func(s string, usage backend.TokenUsage) bool {
choices := []schema.Choice{}
if s != "" {
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not entirely sure about this as I didn't checked what OpenAI does in this case (and hard to replicate).

If the token is empty does it send back an empty choice, or just skips it? Probably doesn't matter either, just making sure we keep an eye to not break any user flow

Copy link
Contributor Author

@mintyleaf mintyleaf Nov 28, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

i don't really know, but that token callback function is used only for streaming reply mode

and in that mode we chop runes of the token, and if there is some - we calling that function, so before that commit in that specific case the tokenCallback never will be called with empty first argument which i check here

the problem is the real usage is sent by llama as final reply where message field bytes is empty

and in that specific case i explicitly call tokenCallback just to capture the usage part

next, about choices - they are always was strictly set to a slice with exactly one element before
so i just used it as an opportunity to not rewrite much of the code and somehow flag out that everything except usage part is need to be ignored, so the real final reply can be constructed after receiving from backend depending on the tools stuff e.t.c.

choices = append(choices, schema.Choice{Delta: &schema.Message{Content: &s}, Index: 0})
}
resp := schema.OpenAIResponse{
ID: id,
Created: created,
Model: req.Model, // we have to return what the user sent here, due to OpenAI spec.
Choices: []schema.Choice{{Delta: &schema.Message{Content: &s}, Index: 0}},
Choices: choices,
Object: "chat.completion.chunk",
Usage: schema.OpenAIUsage{
PromptTokens: usage.Prompt,
Expand Down Expand Up @@ -465,6 +469,9 @@ func ChatEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, startup
toolsCalled := false
for ev := range responses {
usage = &ev.Usage // Copy a pointer to the latest usage chunk so that the stop message can reference it
if len(ev.Choices) == 0 {
break
}
if len(ev.Choices[0].Delta.ToolCalls) > 0 {
toolsCalled = true
}
Expand Down
2 changes: 1 addition & 1 deletion pkg/grpc/backend.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,7 @@ type Backend interface {
Embeddings(ctx context.Context, in *pb.PredictOptions, opts ...grpc.CallOption) (*pb.EmbeddingResult, error)
Predict(ctx context.Context, in *pb.PredictOptions, opts ...grpc.CallOption) (*pb.Reply, error)
LoadModel(ctx context.Context, in *pb.ModelOptions, opts ...grpc.CallOption) (*pb.Result, error)
PredictStream(ctx context.Context, in *pb.PredictOptions, f func(s []byte), opts ...grpc.CallOption) error
PredictStream(ctx context.Context, in *pb.PredictOptions, f func(reply *pb.Reply), opts ...grpc.CallOption) error
GenerateImage(ctx context.Context, in *pb.GenerateImageRequest, opts ...grpc.CallOption) (*pb.Result, error)
TTS(ctx context.Context, in *pb.TTSRequest, opts ...grpc.CallOption) (*pb.Result, error)
SoundGeneration(ctx context.Context, in *pb.SoundGenerationRequest, opts ...grpc.CallOption) (*pb.Result, error)
Expand Down
6 changes: 3 additions & 3 deletions pkg/grpc/client.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,7 +136,7 @@ func (c *Client) LoadModel(ctx context.Context, in *pb.ModelOptions, opts ...grp
return client.LoadModel(ctx, in, opts...)
}

func (c *Client) PredictStream(ctx context.Context, in *pb.PredictOptions, f func(s []byte), opts ...grpc.CallOption) error {
func (c *Client) PredictStream(ctx context.Context, in *pb.PredictOptions, f func(reply *pb.Reply), opts ...grpc.CallOption) error {
if !c.parallel {
c.opMutex.Lock()
defer c.opMutex.Unlock()
Expand All @@ -158,7 +158,7 @@ func (c *Client) PredictStream(ctx context.Context, in *pb.PredictOptions, f fun
}

for {
feature, err := stream.Recv()
reply, err := stream.Recv()
if err == io.EOF {
break
}
Expand All @@ -167,7 +167,7 @@ func (c *Client) PredictStream(ctx context.Context, in *pb.PredictOptions, f fun

return err
}
f(feature.GetMessage())
f(reply)
}

return nil
Expand Down
6 changes: 3 additions & 3 deletions pkg/grpc/embed.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,7 +35,7 @@ func (e *embedBackend) LoadModel(ctx context.Context, in *pb.ModelOptions, opts
return e.s.LoadModel(ctx, in)
}

func (e *embedBackend) PredictStream(ctx context.Context, in *pb.PredictOptions, f func(s []byte), opts ...grpc.CallOption) error {
func (e *embedBackend) PredictStream(ctx context.Context, in *pb.PredictOptions, f func(reply *pb.Reply), opts ...grpc.CallOption) error {
bs := &embedBackendServerStream{
ctx: ctx,
fn: f,
Expand Down Expand Up @@ -97,11 +97,11 @@ func (e *embedBackend) GetTokenMetrics(ctx context.Context, in *pb.MetricsReques

type embedBackendServerStream struct {
ctx context.Context
fn func(s []byte)
fn func(reply *pb.Reply)
}

func (e *embedBackendServerStream) Send(reply *pb.Reply) error {
e.fn(reply.GetMessage())
e.fn(reply)
return nil
}

Expand Down
Loading