Files
ollama_ollama/convert/convert_qwen25vl_test.go
T
Daniel Hiltgen 33878e671a llama: default qwen2.5vl window attention metadata (#16868)
Existing qwen2.5vl GGUFs can contain an empty qwen25vl.vision.fullatt_block_indexes array. The compat layer translated the projector metadata but left clip.vision.n_wa_pattern unset, causing llama-server to fail loading the CLIP model.

Default the runtime compat value to the standard Qwen2.5-VL pattern when the key cannot be derived, and make the converter emit the same default for nil or empty fullatt block metadata.

Fixes #16540
2026-06-24 10:35:29 -07:00

45 lines
887 B
Go

package convert
import (
"slices"
"testing"
)
func TestQwen25VLFullAttentionBlockDefaults(t *testing.T) {
tokenizer := &Tokenizer{Vocabulary: &Vocabulary{}}
for _, tt := range []struct {
name string
blocks []int32
want []int32
}{
{
name: "nil",
want: []int32{7, 15, 23, 31},
},
{
name: "empty",
blocks: []int32{},
want: []int32{7, 15, 23, 31},
},
{
name: "custom",
blocks: []int32{5, 17},
want: []int32{5, 17},
},
} {
t.Run(tt.name, func(t *testing.T) {
model := &qwen25VLModel{}
model.VisionModel.FullAttentionBlocks = tt.blocks
got, ok := model.KV(tokenizer)["qwen25vl.vision.fullatt_block_indexes"].([]int32)
if !ok {
t.Fatalf("fullatt_block_indexes has unexpected type %T", got)
}
if !slices.Equal(got, tt.want) {
t.Fatalf("fullatt_block_indexes = %v, want %v", got, tt.want)
}
})
}
}