CompressedGemma commited on
Commit
f67ea3a
·
verified ·
1 Parent(s): 965a465

Some tensors are transposed lmao

Browse files
Files changed (1) hide show
  1. generate_imatrix.py +5 -1
generate_imatrix.py CHANGED
@@ -792,7 +792,11 @@ class TransformerRunner:
792
  # Gated attention output
793
  if gate_w is not None:
794
  self._record(f'{pfx}.attn_gate.weight', attn_result)
795
- attn_out_vec = attn_result @ gate_w.T
 
 
 
 
796
 
797
  # ── SSM path ──
798
  ssm_alpha_w = self._get_weight(f'{pfx}.ssm_alpha.weight')
 
792
  # Gated attention output
793
  if gate_w is not None:
794
  self._record(f'{pfx}.attn_gate.weight', attn_result)
795
+ # Some GGUF tensors are transposed. Ensure output matches hidden dim.
796
+ if gate_w.shape[1] == hidden.shape[-1]:
797
+ attn_out_vec = attn_result @ gate_w
798
+ else:
799
+ attn_out_vec = attn_result @ gate_w.T
800
 
801
  # ── SSM path ──
802
  ssm_alpha_w = self._get_weight(f'{pfx}.ssm_alpha.weight')