| 12345678910111213141516 |
- diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/fused_marlin_moe.py b/python/sglang/srt/layers/moe/fused_moe_triton/fused_marlin_moe.py
- index 7f76b2616e..9e8b7df1c1 100644
- --- a/python/sglang/srt/layers/moe/fused_moe_triton/fused_marlin_moe.py
- +++ b/python/sglang/srt/layers/moe/fused_moe_triton/fused_marlin_moe.py
- @@ -287,7 +287,10 @@ def fused_marlin_moe(
- dtype=hidden_states.dtype,
- )
- # Marlin skips masked expert rows, so their shared cache must start at zero.
- - intermediate_cache13 = torch.zeros(
- + # Without an expert map every (token, expert) row is local and written by
- + # both GEMMs, so the ~M*topk*K zero-fill (0.8 GB per MoE layer for an
- + # 8K-token MiMo prefill chunk) is skipped.
- + intermediate_cache13 = (torch.zeros if expert_map is not None else torch.empty)(
- (M * topk_ids.shape[1] * max(gemm1_n, K),),
- device=hidden_states.device,
- dtype=hidden_states.dtype,
|
