mirrors/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-05-10 18:06:33 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

ollamarunner: Fix memory leak when processing images

Browse source 
The context (and therefore associated input tensors) was not being
properly closed when images were being processed. We were trying to
close them but in reality we were closing over an empty list, preventing
anything from actually being freed.

Fixes #10434
This commit is contained in:
Jesse Gross 2025-05-01 11:34:02 -07:00 • committed by Jesse Gross
parent 938e8447e8
commit 8e8f2c6d67
1 changed files with 7 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

22
runner/ollamarunner/runner.go
View file

@ -34,14 +34,10 @@ import (
_ "github.com/ollama/ollama/model/models"
)
type contextList struct {
list []ml.Context
}
type Sequence struct {
// ctxs are used for allocating tensors that last the lifetime of the sequence, such as
// multimodal embeddings
ctxs *contextList
ctxs []ml.Context
// batch index
iBatch int
@ -177,8 +173,10 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
// inputs processes the prompt and images into a list of inputs
// by splitting the prompt on [img-<n>] tags, tokenizing text and
// decoding images
func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, *contextList, error) {
func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, []ml.Context, error) {
var inputs []input.Input
var ctxs []ml.Context
var parts []string
var matches [][]string
@ -192,13 +190,6 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, *
parts = []string{prompt}
}
var contexts contextList
runtime.AddCleanup(&contexts, func(ctxs []ml.Context) {
for _, ctx := range ctxs {
ctx.Close()
}
}, contexts.list)
postTokenize := false
for i, part := range parts {
// text - tokenize
@ -228,7 +219,8 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, *
}
ctx := s.model.Backend().NewContext()
contexts.list = append(contexts.list, ctx)
runtime.SetFinalizer(ctx, func(c ml.Context) { c.Close() })
ctxs = append(ctxs, ctx)
imageEmbeddings, err := multimodalProcessor.EncodeMultimodal(ctx, images[imageIndex].Data)
if err != nil {
return nil, nil, err
@ -251,7 +243,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, *
}
}
return inputs, &contexts, nil
return inputs, ctxs, nil
}
type Server struct {