From aa0b65c7c79ca67ba96f0484de814f22b980cc6e Mon Sep 17 00:00:00 2001 From: "Wang, FangYuan" <39615225+At1a8@users.noreply.github.com> Date: Wed, 23 Sep 2026 16:35:56 +0800 Subject: [PATCH] [AMD] Fix DeepSeek-V4 accuracy by not passing num_token_non_padded to MoE topk (#39804) Co-authored-by: Thomas Wang Co-authored-by: HAI --- python/sglang/srt/models/deepseek_v2.py | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-) diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py index 47361dd5ba..6c30cd3908 100644 --- a/python/sglang/srt/models/deepseek_v2.py +++ b/python/sglang/srt/models/deepseek_v2.py @@ -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()