[AMD] Fix DeepSeek-V4 accuracy by not passing num_token_non_padded to MoE topk (#39804)

Co-authored-by: Thomas Wang <thomawan@amd.com>
Co-authored-by: HAI <hixiao@gmail.com>
This commit is contained in:
Wang, FangYuan
2026-09-23 01:35:56 -07:00
committed by GitHub
co-authored by Thomas Wang HAI
parent 48c3854620
commit aa0b65c7c7
+8 -5
View File
@@ -924,11 +924,14 @@ class DeepseekV2MoE(nn.Module):
input_ids_global=input_ids_global,
)
num_token_non_padded = (
forward_batch.moe_num_token_non_padded()
if forward_batch is not None
else None
)
if _is_hip:
num_token_non_padded = None
else:
num_token_non_padded = (
forward_batch.moe_num_token_non_padded()
if forward_batch is not None
else None
)
if not self._enable_a2a_moe:
if self._can_dual_stream_graph(hidden_states):
fwd = get_forward()