Wow Qwen......

generate_imatrix.py

@@ -811,10 +811,22 @@ class TransformerRunner:
                 qkv_for_ssm = normed @ qkv_w.T
                 self._record(f'{pfx}.ssm_conv1d.weight', qkv_for_ssm)
         if ssm_out_w is not None:
-            # SSM output projection — use normed as proxy for recurrent output
-            ssm_proxy = normed[:, :ssm_out_w.shape[1]] if normed.shape[-1] >= ssm_out_w.shape[1] else normed
+            # SSM output projection — use qkv output as proxy for recurrent output
+            # (qkv is 10240, ssm_out expects 6144)
+            if 'qkv' in locals() and qkv.shape[-1] >= ssm_out_w.shape[1]:
+                ssm_proxy = qkv[:, :ssm_out_w.shape[1]]
+            else:
+                # Fallback zero pad
+                ssm_proxy = np.zeros((seq_len, ssm_out_w.shape[1]), dtype=np.float32)
+                
             self._record(f'{pfx}.ssm_out.weight', ssm_proxy)
-            ssm_out_vec = ssm_proxy @ ssm_out_w.T
+            
+            # Note: We do NOT need to actually add the SSM output vector to hidden 
+            # for importance matrix calculation. We just need to record the inputs 
+            # to all quantized layers. The actual output isn't critical since we 
+            # aren't doing loss backprop. But if we do, it must match hidden's dimension.
+            if ssm_out_w.shape[0] == hidden.shape[-1]:
+                ssm_out_vec = ssm_proxy @ ssm_out_w.T
 
         # Combine attention + SSM
         hidden = hidden + attn_out_vec + ssm_out_vec