matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-14 01:37:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

runner.go: Better abstract vision model integration

Browse Source 
-Update mllama to take the cross attention state as embeddings in
a batch, more similar to how Llava handles it. This improves
integration with the input cache.
-Pass locations in a prompt for embeddings using tags similar to Llava.
-Abstract interface to vision models so the main runner accesses Clip
and Mllama similarly

Co-authored-by: Michael Yang <mxyng@pm.me>
This commit is contained in:
Jesse Gross
2024-10-11 15:34:01 -07:00
committed by Jesse Gross
parent 712e99d477
commit c826e57475
13 changed files with 534 additions and 454 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
server/prompt.go
View File

@@ -75,11 +75,16 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
currMsgIdx := n
if isMllama {
lastMsgIdx := len(msgs) - 1
for i := lastMsgIdx; i >= currMsgIdx; i-- {
if len(msgs[i].Images) > 0 {
data, aspectRatioID, err := imageproc.Preprocess(msgs[i].Images[0])
for cnt, msg := range msgs[currMsgIdx:] {
prefix := ""
imgPrompt := ""
prompt := msg.Content
for _, i := range msg.Images {
var imgData llm.ImageData
if isMllama {
data, aspectRatioID, err := imageproc.Preprocess(i)
if err != nil {
return "", nil, err
}
@@ -90,37 +95,30 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
return "", nil, err
}
imgData := llm.ImageData{
imgData = llm.ImageData{
ID: len(images),
Data: buf.Bytes(),
AspectRatioID: aspectRatioID,
}
msgs[i].Content = strings.TrimSpace("<|image|>" + msgs[i].Content)
images = append(images, imgData)
break
}
}
} else {
for cnt, msg := range msgs[currMsgIdx:] {
prefix := ""
prompt := msg.Content
for _, i := range msg.Images {
imgData := llm.ImageData{
imgPrompt = "<|image|>"
} else {
imgData = llm.ImageData{
ID: len(images),
Data: i,
}
imgTag := fmt.Sprintf("[img-%d]", imgData.ID)
if !strings.Contains(prompt, "[img]") {
prefix += imgTag
} else {
prompt = strings.Replace(prompt, "[img]", imgTag, 1)
}
images = append(images, imgData)
imgPrompt = " "
}
msgs[currMsgIdx+cnt].Content = strings.TrimSpace(prefix + " " + prompt)
imgTag := fmt.Sprintf("[img-%d]", imgData.ID)
if !strings.Contains(prompt, "[img]") {
prefix += imgTag
} else {
prompt = strings.Replace(prompt, "[img]", imgTag, 1)
}
images = append(images, imgData)
}
msgs[currMsgIdx+cnt].Content = strings.TrimSpace(prefix + imgPrompt + prompt)
}
// truncate any messages that do not fit into the context window