OSSForks/ollama
1
0
Fork 0
mirror of https://github.com/zebrajr/ollama.git synced 2025-12-06 00:19:51 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

Revert "add truncate and shift parameters (#12519)" (#12545)

Browse Source 
This reverts commit 6a62b894c7.
This commit is contained in:
Jeffrey Morgan 2025-10-08 17:57:57 -07:00 committed by GitHub
parent 6a62b894c7
commit 7d965258ce
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
8 changed files with 67 additions and 272 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
api/types.go
View File

@ -106,14 +106,6 @@ type GenerateRequest struct {
// before this option was introduced)
Think *ThinkValue `json:"think,omitempty"`
// Truncate is a boolean that, when set to true, truncates the chat history messages
// if the rendered prompt exceeds the context length limit.
Truncate *bool `json:"truncate,omitempty"`
// Shift is a boolean that, when set to true, shifts the chat history
// when hitting the context length limit instead of erroring.
Shift *bool `json:"shift,omitempty"`
// DebugRenderOnly is a debug option that, when set to true, returns the rendered
// template instead of calling the model.
DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
@ -148,14 +140,6 @@ type ChatRequest struct {
// for supported models.
Think *ThinkValue `json:"think,omitempty"`
// Truncate is a boolean that, when set to true, truncates the chat history messages
// if the rendered prompt exceeds the context length limit.
Truncate *bool `json:"truncate,omitempty"`
// Shift is a boolean that, when set to true, shifts the chat history
// when hitting the context length limit instead of erroring.
Shift *bool `json:"shift,omitempty"`
// DebugRenderOnly is a debug option that, when set to true, returns the rendered
// template instead of calling the model.
DebugRenderOnly bool `json:"_debug_render_only,omitempty"`

6
llm/server.go
View File

@ -1379,9 +1379,7 @@ type CompletionRequest struct {
Images []ImageData
Options *api.Options
Grammar string // set before sending the request to the subprocess
Shift bool
Truncate bool
Grammar string // set before sending the request to the subprocess
}
// DoneReason represents the reason why a completion response is done
@ -1500,7 +1498,7 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
return fmt.Errorf("failed reading llm error response: %w", err)
}
log.Printf("llm predict error: %s", bodyBytes)
return api.StatusError{StatusCode: res.StatusCode, Status: res.Status, ErrorMessage: strings.TrimSpace(string(bodyBytes))}
return fmt.Errorf("%s", bodyBytes)
}
scanner := bufio.NewScanner(res.Body)

22
runner/llamarunner/runner.go
View File

@ -79,9 +79,6 @@ type Sequence struct {
// true if an embedding are to be returned instead of text generation
embeddingOnly bool
// shift if context window is exceeded
shift bool
doneReason llm.DoneReason
// Metrics
@ -97,12 +94,8 @@ type NewSequenceParams struct {
numKeep int
samplingParams *llama.SamplingParams
embedding bool
shift bool
truncate bool
}
var errorInputTooLong = errors.New("the input length exceeds the context length")
func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSequenceParams) (*Sequence, error) {
s.ready.Wait()
@ -128,10 +121,6 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
if len(inputs) > s.cache.numCtx {
discard := len(inputs) - s.cache.numCtx
if !params.truncate {
return nil, errorInputTooLong
}
newInputs := inputs[:params.numKeep]
newInputs = append(newInputs, inputs[params.numKeep+discard:]...)
@ -399,11 +388,6 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
for i, input := range seq.inputs {
if len(seq.cache.Inputs)+len(seq.pendingInputs)+1 > s.cache.numCtx {
if len(seq.pendingInputs) == 0 {
if !seq.shift {
s.removeSequence(seqIdx, llm.DoneReasonLength)
break
}
err := s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
if err != nil {
var reprocess *ErrReprocessInputs
@ -599,14 +583,8 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
numKeep: req.Options.NumKeep,
samplingParams: &samplingParams,
embedding: false,
shift: req.Shift,
truncate: req.Truncate,
})
if err != nil {
if errors.Is(err, errorInputTooLong) {
http.Error(w, err.Error(), http.StatusBadRequest)
return
}
http.Error(w, fmt.Sprintf("Failed to create new sequence: %v", err), http.StatusInternalServerError)
return
}

25
runner/ollamarunner/runner.go
View File

@ -88,9 +88,6 @@ type Sequence struct {
// true if an embedding are to be returned instead of text generation
embeddingOnly bool
// shift if context window is exceeded
shift bool
doneReason llm.DoneReason
// Metrics
@ -106,12 +103,8 @@ type NewSequenceParams struct {
numKeep int32
sampler sample.Sampler
embedding bool
shift bool
truncate bool
}
var errorInputTooLong = errors.New("the input length exceeds the context length")
func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSequenceParams) (*Sequence, error) {
s.ready.Wait()
@ -133,11 +126,6 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
if int32(len(inputs)) > s.cache.numCtx {
discard := int32(len(inputs)) - s.cache.numCtx
if !params.truncate {
return nil, errorInputTooLong
}
promptStart := params.numKeep + discard
// If we need to truncate in the middle of a unbreakable batch, remove the entire batch
@ -190,7 +178,6 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
embeddingOnly: params.embedding,
stop: params.stop,
numKeep: params.numKeep,
shift: params.shift,
}, nil
}
@ -535,12 +522,6 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
break
}
if !seq.shift {
s.removeSequence(seqIdx, llm.DoneReasonLength)
nextBatch.seqs[seqIdx] = nil
break
}
err = s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
if err != nil {
var reprocess *ErrReprocessInputs
@ -843,14 +824,8 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
numKeep: int32(req.Options.NumKeep),
sampler: sampler,
embedding: false,
shift: req.Shift,
truncate: req.Truncate,
})
if err != nil {
if errors.Is(err, errorInputTooLong) {
http.Error(w, err.Error(), http.StatusBadRequest)
return
}
http.Error(w, fmt.Sprintf("Failed to create new sequence: %v", err), http.StatusInternalServerError)
return
}

4
server/prompt.go
View File

@ -20,7 +20,7 @@ type tokenizeFunc func(context.Context, string) ([]int, error)
// chatPrompt accepts a list of messages and returns the prompt and images that should be used for the next chat turn.
// chatPrompt truncates any messages that exceed the context window of the model, making sure to always include 1) the
// latest message and 2) system messages
func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message, tools []api.Tool, think *api.ThinkValue, truncate bool) (prompt string, images []llm.ImageData, _ error) {
func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message, tools []api.Tool, think *api.ThinkValue) (prompt string, images []llm.ImageData, _ error) {
var system []api.Message
// TODO: Ideally we would compute this from the projector metadata but some pieces are implementation dependent
@ -59,7 +59,7 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
}
}
if truncate && ctxLen > opts.NumCtx {
if ctxLen > opts.NumCtx {
slog.Debug("truncating input messages which exceed context length", "truncated", len(msgs[i:]))
break
} else {

102
server/prompt_test.go
View File

@ -27,18 +27,16 @@ func TestChatPrompt(t *testing.T) {
visionModel := Model{Template: tmpl, ProjectorPaths: []string{"vision"}}
cases := []struct {
name string
model Model
limit int
truncate bool
msgs []api.Message
name string
model Model
limit int
msgs []api.Message
expect
}{
{
name: "messages",
model: visionModel,
limit: 64,
truncate: true,
name: "messages",
model: visionModel,
limit: 64,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
@ -49,10 +47,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "truncate messages",
model: visionModel,
limit: 1,
truncate: true,
name: "truncate messages",
model: visionModel,
limit: 1,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
@ -63,10 +60,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "truncate messages with image",
model: visionModel,
limit: 64,
truncate: true,
name: "truncate messages with image",
model: visionModel,
limit: 64,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
@ -80,10 +76,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "truncate messages with images",
model: visionModel,
limit: 64,
truncate: true,
name: "truncate messages with images",
model: visionModel,
limit: 64,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{[]byte("something")}},
{Role: "assistant", Content: "I-I'm a what?"},
@ -97,10 +92,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "messages with images",
model: visionModel,
limit: 2048,
truncate: true,
name: "messages with images",
model: visionModel,
limit: 2048,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{[]byte("something")}},
{Role: "assistant", Content: "I-I'm a what?"},
@ -115,10 +109,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "message with image tag",
model: visionModel,
limit: 2048,
truncate: true,
name: "message with image tag",
model: visionModel,
limit: 2048,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry! [img]", Images: []api.ImageData{[]byte("something")}},
{Role: "assistant", Content: "I-I'm a what?"},
@ -133,10 +126,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "messages with interleaved images",
model: visionModel,
limit: 2048,
truncate: true,
name: "messages with interleaved images",
model: visionModel,
limit: 2048,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "user", Images: []api.ImageData{[]byte("something")}},
@ -153,10 +145,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "truncate message with interleaved images",
model: visionModel,
limit: 1024,
truncate: true,
name: "truncate message with interleaved images",
model: visionModel,
limit: 1024,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "user", Images: []api.ImageData{[]byte("something")}},
@ -172,10 +163,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "message with system prompt",
model: visionModel,
limit: 2048,
truncate: true,
name: "message with system prompt",
model: visionModel,
limit: 2048,
msgs: []api.Message{
{Role: "system", Content: "You are the Test Who Lived."},
{Role: "user", Content: "You're a test, Harry!"},
@ -187,10 +177,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "out of order system",
model: visionModel,
limit: 2048,
truncate: true,
name: "out of order system",
model: visionModel,
limit: 2048,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
@ -202,10 +191,9 @@ func TestChatPrompt(t *testing.T) {
},
},
{
name: "multiple images same prompt",
model: visionModel,
limit: 2048,
truncate: true,
name: "multiple images same prompt",
model: visionModel,
limit: 2048,
msgs: []api.Message{
{Role: "user", Content: "Compare these two pictures of hotdogs", Images: []api.ImageData{[]byte("one hotdog"), []byte("two hotdogs")}},
},
@ -214,20 +202,6 @@ func TestChatPrompt(t *testing.T) {
images: [][]byte{[]byte("one hotdog"), []byte("two hotdogs")},
},
},
{
name: "no truncate with limit exceeded",
model: visionModel,
limit: 10,
truncate: false,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
},
expect: expect{
prompt: "You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
},
},
}
for _, tt := range cases {
@ -235,7 +209,7 @@ func TestChatPrompt(t *testing.T) {
model := tt.model
opts := api.Options{Runner: api.Runner{NumCtx: tt.limit}}
think := false
prompt, images, err := chatPrompt(t.Context(), &model, mockRunner{}.Tokenize, &opts, tt.msgs, nil, &api.ThinkValue{Value: think}, tt.truncate)
prompt, images, err := chatPrompt(t.Context(), &model, mockRunner{}.Tokenize, &opts, tt.msgs, nil, &api.ThinkValue{Value: think})
if tt.error == nil && err != nil {
t.Fatal(err)
} else if tt.error != nil && err != tt.error {

63
server/routes.go
View File

@ -468,12 +468,10 @@ func (s *Server) GenerateHandler(c *gin.Context) {
var sb strings.Builder
defer close(ch)
if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
Prompt: prompt,
Images: images,
Format: req.Format,
Options: opts,
Shift: req.Shift == nil || *req.Shift,
Truncate: req.Truncate == nil || *req.Truncate,
Prompt: prompt,
Images: images,
Format: req.Format,
Options: opts,
}, func(cr llm.CompletionResponse) {
res := api.GenerateResponse{
Model: req.Model,
@ -535,7 +533,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
ch <- res
}); err != nil {
ch <- err
ch <- gin.H{"error": err.Error()}
}
}()
@ -549,11 +547,13 @@ func (s *Server) GenerateHandler(c *gin.Context) {
sbThinking.WriteString(t.Thinking)
sbContent.WriteString(t.Response)
r = t
case api.StatusError:
c.JSON(t.StatusCode, gin.H{"error": t.ErrorMessage})
return
case error:
c.JSON(http.StatusInternalServerError, gin.H{"error": t.Error()})
case gin.H:
msg, ok := t["error"].(string)
if !ok {
msg = "unexpected error format in response"
}
c.JSON(http.StatusInternalServerError, gin.H{"error": msg})
return
default:
c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected response"})
@ -1618,18 +1618,6 @@ func streamResponse(c *gin.Context, ch chan any) {
return false
}
if statusError, ok := val.(api.StatusError); ok {
c.Header("Content-Type", "application/json")
c.AbortWithStatusJSON(statusError.StatusCode, gin.H{"error": statusError.ErrorMessage})
return false
}
if err, ok := val.(error); ok {
c.Header("Content-Type", "application/json")
c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
return false
}
bts, err := json.Marshal(val)
if err != nil {
slog.Info(fmt.Sprintf("streamResponse: json.Marshal failed with %s", err))
@ -1947,8 +1935,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
}
}
truncate := req.Truncate == nil || *req.Truncate
prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, processedTools, req.Think, truncate)
prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, processedTools, req.Think)
if err != nil {
slog.Error("chat prompt error", "error", err)
c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
@ -1997,12 +1984,10 @@ func (s *Server) ChatHandler(c *gin.Context) {
defer close(ch)
if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
Prompt: prompt,
Images: images,
Format: req.Format,
Options: opts,
Shift: req.Shift == nil || *req.Shift,
Truncate: truncate,
Prompt: prompt,
Images: images,
Format: req.Format,
Options: opts,
}, func(r llm.CompletionResponse) {
res := api.ChatResponse{
Model: req.Model,
@ -2075,7 +2060,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
ch <- res
}); err != nil {
ch <- err
ch <- gin.H{"error": err.Error()}
}
}()
@ -2093,11 +2078,13 @@ func (s *Server) ChatHandler(c *gin.Context) {
if len(req.Tools) > 0 {
toolCalls = append(toolCalls, t.Message.ToolCalls...)
}
case api.StatusError:
c.JSON(t.StatusCode, gin.H{"error": t.ErrorMessage})
return
case error:
c.JSON(http.StatusInternalServerError, gin.H{"error": t.Error()})
case gin.H:
msg, ok := t["error"].(string)
if !ok {
msg = "unexpected error format in response"
}
c.JSON(http.StatusInternalServerError, gin.H{"error": msg})
return
default:
c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected response"})

101
server/routes_generate_test.go
View File

@ -594,58 +594,6 @@ func TestGenerateChat(t *testing.T) {
t.Errorf("final tool call mismatch (-got +want):\n%s", diff)
}
})
t.Run("status error non-streaming", func(t *testing.T) {
mock.CompletionFn = func(ctx context.Context, r llm.CompletionRequest, fn func(r llm.CompletionResponse)) error {
return api.StatusError{
StatusCode: http.StatusServiceUnavailable,
Status: "Service Unavailable",
ErrorMessage: "model is overloaded",
}
}
stream := false
w := createRequest(t, s.ChatHandler, api.ChatRequest{
Model: "test",
Messages: []api.Message{
{Role: "user", Content: "Hello!"},
},
Stream: &stream,
})
if w.Code != http.StatusServiceUnavailable {
t.Errorf("expected status 503, got %d", w.Code)
}
if diff := cmp.Diff(w.Body.String(), `{"error":"model is overloaded"}`); diff != "" {
t.Errorf("mismatch (-got +want):\n%s", diff)
}
})
t.Run("status error streaming", func(t *testing.T) {
mock.CompletionFn = func(ctx context.Context, r llm.CompletionRequest, fn func(r llm.CompletionResponse)) error {
return api.StatusError{
StatusCode: http.StatusTooManyRequests,
Status: "Too Many Requests",
ErrorMessage: "rate limit exceeded",
}
}
w := createRequest(t, s.ChatHandler, api.ChatRequest{
Model: "test",
Messages: []api.Message{
{Role: "user", Content: "Hello!"},
},
})
if w.Code != http.StatusTooManyRequests {
t.Errorf("expected status 429, got %d", w.Code)
}
if diff := cmp.Diff(w.Body.String(), `{"error":"rate limit exceeded"}`); diff != "" {
t.Errorf("mismatch (-got +want):\n%s", diff)
}
})
}
func TestGenerate(t *testing.T) {
@ -1020,55 +968,6 @@ func TestGenerate(t *testing.T) {
t.Errorf("mismatch (-got +want):\n%s", diff)
}
})
t.Run("status error non-streaming", func(t *testing.T) {
mock.CompletionFn = func(ctx context.Context, r llm.CompletionRequest, fn func(r llm.CompletionResponse)) error {
return api.StatusError{
StatusCode: http.StatusServiceUnavailable,
Status: "Service Unavailable",
ErrorMessage: "model is overloaded",
}
}
streamRequest := false
w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
Model: "test",
Prompt: "Hello!",
Stream: &streamRequest,
})
if w.Code != http.StatusServiceUnavailable {
t.Errorf("expected status 503, got %d", w.Code)
}
if diff := cmp.Diff(w.Body.String(), `{"error":"model is overloaded"}`); diff != "" {
t.Errorf("mismatch (-got +want):\n%s", diff)
}
})
t.Run("status error streaming", func(t *testing.T) {
mock.CompletionFn = func(ctx context.Context, r llm.CompletionRequest, fn func(r llm.CompletionResponse)) error {
return api.StatusError{
StatusCode: http.StatusTooManyRequests,
Status: "Too Many Requests",
ErrorMessage: "rate limit exceeded",
}
}
w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
Model: "test",
Prompt: "Hello!",
Stream: &stream,
})
if w.Code != http.StatusTooManyRequests {
t.Errorf("expected status 429, got %d", w.Code)
}
if diff := cmp.Diff(w.Body.String(), `{"error":"rate limit exceeded"}`); diff != "" {
t.Errorf("mismatch (-got +want):\n%s", diff)
}
})
}
func TestChatWithPromptEndingInThinkTag(t *testing.T) {