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

convert/convert_qwen25vl.go

@@ -39,10 +39,6 @@ func (q *qwen25VLModel) KV(t *Tokenizer) KV {
 		}
 	}
 
-	if q.VisionModel.FullAttentionBlocks == nil {
-		kv["qwen25vl.vision.fullatt_block_indexes"] = []int32{7, 15, 23, 31}
-	}
-
 	kv["qwen25vl.vision.block_count"] = cmp.Or(q.VisionModel.Depth, 32)
 	kv["qwen25vl.vision.embedding_length"] = q.VisionModel.HiddenSize
 	kv["qwen25vl.vision.attention.head_count"] = cmp.Or(q.VisionModel.NumHeads, 16)
@@ -53,12 +49,19 @@ func (q *qwen25VLModel) KV(t *Tokenizer) KV {
 	kv["qwen25vl.vision.window_size"] = cmp.Or(q.VisionModel.WindowSize, 112)
 	kv["qwen25vl.vision.attention.layer_norm_epsilon"] = cmp.Or(q.VisionModel.RMSNormEps, 1e-6)
 	kv["qwen25vl.vision.rope.freq_base"] = cmp.Or(q.VisionModel.RopeTheta, 1e4)
-	kv["qwen25vl.vision.fullatt_block_indexes"] = q.VisionModel.FullAttentionBlocks
+	kv["qwen25vl.vision.fullatt_block_indexes"] = q.fullAttentionBlocks()
 	kv["qwen25vl.vision.temporal_patch_size"] = cmp.Or(q.VisionModel.TemporalPatchSize, 2)
 
 	return kv
 }
 
+func (q *qwen25VLModel) fullAttentionBlocks() []int32 {
+	if len(q.VisionModel.FullAttentionBlocks) > 0 {
+		return q.VisionModel.FullAttentionBlocks
+	}
+	return []int32{7, 15, 23, 31}
+}
+
 func (q *qwen25VLModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	var out []*ggml.Tensor
 

convert/convert_qwen25vl_test.go

@@ -0,0 +1,44 @@
+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)
+			}
+		})
+	}
+}

llama/compat/llama-ollama-compat.cpp

@@ -2958,6 +2958,7 @@ void handle_qwen25vl_clip(gguf_context * meta, ggml_context * ctx) {
             gguf_set_val_u32(meta, "clip.vision.n_wa_pattern", (uint32_t)(arr[0] + 1));
         }
     }
+    inject_u32_if_missing(meta, "clip.vision.n_wa_pattern", 8);
 
     // Default image_size = 560 (Qwen2VLVisionModel default, no image_size in HF config).
     inject_u32_if_missing(meta, "clip.vision.image_size", 560);