chore: update mllama to use ollama engine (#10637)

This commit is contained in:
Michael Yang
2025-05-13 17:36:02 -07:00
committed by GitHub
parent 0478d440f0
commit 23125648b8
67 changed files with 785 additions and 4354 deletions

View File

@@ -2,8 +2,6 @@ package server
import (
"bytes"
"image"
"image/png"
"testing"
"github.com/google/go-cmp/cmp"
@@ -14,10 +12,9 @@ import (
func TestChatPrompt(t *testing.T) {
type expect struct {
prompt string
images [][]byte
aspectRatioID int
error error
prompt string
images [][]byte
error error
}
tmpl, err := template.Parse(`
@@ -28,28 +25,6 @@ func TestChatPrompt(t *testing.T) {
t.Fatal(err)
}
visionModel := Model{Template: tmpl, ProjectorPaths: []string{"vision"}}
mllamaModel := Model{Template: tmpl, ProjectorPaths: []string{"vision"}, Config: ConfigV2{ModelFamilies: []string{"mllama"}}}
createImg := func(width, height int) ([]byte, error) {
img := image.NewRGBA(image.Rect(0, 0, width, height))
var buf bytes.Buffer
if err := png.Encode(&buf, img); err != nil {
return nil, err
}
return buf.Bytes(), nil
}
imgBuf, err := createImg(5, 5)
if err != nil {
t.Fatal(err)
}
imgBuf2, err := createImg(6, 6)
if err != nil {
t.Fatal(err)
}
cases := []struct {
name string
@@ -227,90 +202,6 @@ func TestChatPrompt(t *testing.T) {
images: [][]byte{[]byte("one hotdog"), []byte("two hotdogs")},
},
},
{
name: "messages with mllama (no images)",
model: mllamaModel,
limit: 2048,
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. ",
},
},
{
name: "messages with mllama single prompt",
model: mllamaModel,
limit: 2048,
msgs: []api.Message{
{Role: "user", Content: "How many hotdogs are in this image?", Images: []api.ImageData{imgBuf}},
},
expect: expect{
prompt: "[img-0]<|image|>How many hotdogs are in this image? ",
images: [][]byte{imgBuf},
aspectRatioID: 1,
},
},
{
name: "messages with mllama",
model: mllamaModel,
limit: 2048,
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.", Images: []api.ImageData{imgBuf}},
},
expect: expect{
prompt: "You're a test, Harry! I-I'm a what? [img-0]<|image|>A test. And a thumping good one at that, I'd wager. ",
images: [][]byte{imgBuf},
aspectRatioID: 1,
},
},
{
name: "multiple messages with mllama",
model: mllamaModel,
limit: 2048,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{imgBuf}},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{imgBuf2}},
},
expect: expect{
prompt: "[img-0]<|image|>You're a test, Harry! I-I'm a what? [img-1]<|image|>A test. And a thumping good one at that, I'd wager. ",
images: [][]byte{imgBuf, imgBuf2},
aspectRatioID: 1,
},
},
{
name: "earlier image with mllama",
model: mllamaModel,
limit: 2048,
msgs: []api.Message{
{Role: "user", Content: "How many hotdogs are in this image?", Images: []api.ImageData{imgBuf}},
{Role: "assistant", Content: "There are four hotdogs."},
{Role: "user", Content: "Which ones have mustard?"},
},
expect: expect{
prompt: "[img-0]<|image|>How many hotdogs are in this image? There are four hotdogs. Which ones have mustard? ",
images: [][]byte{imgBuf},
aspectRatioID: 1,
},
},
{
name: "too many images with mllama",
model: mllamaModel,
limit: 2048,
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.", Images: []api.ImageData{imgBuf, imgBuf}},
},
expect: expect{
error: errTooManyImages,
},
},
}
for _, tt := range cases {
@@ -341,10 +232,6 @@ func TestChatPrompt(t *testing.T) {
if !bytes.Equal(images[i].Data, tt.images[i]) {
t.Errorf("expected %q, got %q", tt.images[i], images[i].Data)
}
} else {
if images[i].AspectRatioID != tt.aspectRatioID {
t.Errorf("expected aspect ratio %d, got %d", tt.aspectRatioID, images[i].AspectRatioID)
}
}
}
})