[ROCm][CI][MoE] Fix double-transpose of fused w3 expert weights (#47874)
Signed-off-by: stefankoncarevic <stefan.koncarevic@amd.com> Signed-off-by: Andreas Karatzas <akaratza@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
stefankoncarevic committed
Jul 9, 2026 at 01:59 UTC
bc44f9feb7ab8e0c0be68a8c6e2656e9be053236
1 file changed
+10
-6
vllm/model_executor/layers/fused_moe/routed_experts.py
+10
-6
@@ -886,17 +886,21 @@ class RoutedExperts(PluggableLayer):
886
param_name = weight_name.removeprefix(f"{self.layer_name}.")
887
param = getattr(self, param_name)
888
if is_fused:
889
+ # w1 and w3 share one fused tensor; use a local copy so the
890
+ # transpose below doesn't mutate loaded_weight across
891
+ # iterations (else w3 is transposed twice and wrongly chunked)
892
+ fused_weight = loaded_weight
893
if shard_id in {"w1", "w3"}:
890
- if loaded_weight.shape[-1] != unpadded_hidden:
894
+ if fused_weight.shape[-1] != unpadded_hidden:
895
# [..., hidden, intermediate] -> [..., intermediate, hidden]
892
- loaded_weight = loaded_weight.transpose(-1, -2)
896
+ fused_weight = fused_weight.transpose(-1, -2)
897
# Repurpose expert_id for deconcatenating w1 and w3
894
- experts_shard = loaded_weight.chunk(2, dim=1)[expert_id]
898
+ experts_shard = fused_weight.chunk(2, dim=1)[expert_id]
899
else:
896
- if loaded_weight.shape[-2] != unpadded_hidden:
900
+ if fused_weight.shape[-2] != unpadded_hidden:
901
# [..., intermediate, hidden] -> [..., hidden, intermediate]
898
- loaded_weight = loaded_weight.transpose(-1, -2)
899
- experts_shard = loaded_weight
902
+ fused_weight = fused_weight.transpose(-1, -2)
903
+ experts_shard = fused_weight
904
start = 0
905
else:
906
# loaded_weight is a single expert weight, so we add a dummy expert