Qwen......

generate_imatrix.py

@@ -141,16 +141,16 @@ class GGUFModel:
         n_head = self.kv.get(f'{arch}.attention.head_count', 0)
         n_head_kv = self.kv.get(f'{arch}.attention.head_count_kv', 0)
 
-        # Auto-detect head_dim: prefer fused QKV tensor dims over n_embd/n_head
-        # (Qwen 3.6 has n_embd=5120, n_head=24 → non-integer ratio;
-        #  real head_dim=320, computed from QKV: 10240 / (24+2*4) = 320)
+        # Auto-detect head_dim: prefer derived from attn_gate over n_embd/n_head
+        # (Qwen 3.6 has hybrid 10240 QKV output but attn_gate requires 6144. 
+        #  6144 / 24 heads = 256 real head_dim).
         head_dim = 0
-        qkv_name = 'blk.0.attn_qkv.weight'
-        if qkv_name in self.tensor_infos:
-            qkv_cols = self.tensor_infos[qkv_name]['dims'][1]  # output dim
-            n_head_total = n_head + 2 * n_head_kv
-            if n_head_total > 0:
-                head_dim = qkv_cols // n_head_total
+        gate_name = 'blk.0.attn_gate.weight'
+        if gate_name in self.tensor_infos:
+            # attn_gate is [n_embd, n_head * head_dim]
+            gate_cols = self.tensor_infos[gate_name]['dims'][1]  # input dim
+            if n_head > 0:
+                head_dim = gate_cols // n_head
         if head_dim == 0 and n_head > 0:
             head_dim = n_embd // n_head