From 14c4dda24e1cf623935331dd77831e842e301f93 Mon Sep 17 00:00:00 2001 From: Khoa Pham Date: Fri, 16 Oct 2026 23:47:41 -0700 Subject: [PATCH] [Perf] deepseek_v4: fuse block-FP8 shared experts into MegaMoE for any 32-aligned block size (#44837) Co-authored-by: Claude Fable 5.1 --- python/sglang/srt/layers/moe/mega_moe.py | 20 ++++++++++++++++---- 1 file changed, 16 insertions(+), 4 deletions(-) diff --git a/python/sglang/srt/layers/moe/mega_moe.py b/python/sglang/srt/layers/moe/mega_moe.py index 02a7f91606..cedb344db0 100644 --- a/python/sglang/srt/layers/moe/mega_moe.py +++ b/python/sglang/srt/layers/moe/mega_moe.py @@ -200,7 +200,8 @@ def should_fuse_mega_moe_shared_experts(moe: DeepseekV2MoE) -> bool: and is_sm100_supported() and _mega_moe_mma_type(moe.experts) == "fp8xfp4" and moe.shared_experts_is_fp8 - and moe.shared_experts_weight_block_size == [128, 128] + and moe.shared_experts_weight_block_size is not None + and moe.shared_experts_weight_block_size[1] % 32 == 0 ) @@ -617,10 +618,21 @@ def _transform_mega_moe_shared_sf(linear) -> torch.Tensor: if sf.dtype == torch.int32: # Requantized for the DeepGEMM linear runner. sf = inverse_transform_scale_ue8m0(sf, mn=n)[:, : k // 128] + block_m, block_k = 128, 128 + else: + block_m, block_k = linear.quant_method.quant_config.weight_block_size + sf = sf.repeat_interleave(block_k // 32, dim=-1) + if block_m == 128: + return transform_sf_into_required_layout( + sf, mn=n, k=k, recipe=(128, 32), disable_ue8m0_cast=False + ) + if block_m != 1: + sf = sf.repeat_interleave(block_m, dim=0) return transform_sf_into_required_layout( - sf.repeat_interleave(4, dim=-1), + sf.unsqueeze(0).contiguous(), mn=n, k=k, - recipe=(128, 32), + recipe=(1, 32), + num_groups=1, disable_ue8m0_cast=False, - ) + ).squeeze(0)