EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Fix] RuntimeError: get_cfg Unsupported input_type:Float4_e2m1fn_x2 in using aiter-mxfp4-moe (#10981)

Browse Source 
Co-authored-by: wunhuang <wunhuang@amd.com>
This commit is contained in:
kk
2025-09-27 13:12:22 +08:00
committed by GitHub
parent 8260574729
commit 8ebf72fef3
2 changed files with 19 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
python/sglang/srt/layers/quantization/mxfp4.py
View File

@@ -843,10 +843,18 @@ class Mxfp4DynamicQuantMoEMethod(FusedMoEMethodBase):
topk_weights = topk_weights.to( topk_weights = topk_weights.to(
torch.float32 torch.float32
) # aiter's moe_sorting requires topk_weights to be FP32 ) # aiter's moe_sorting requires topk_weights to be FP32
if hasattr(torch, "float4_e2m1fn_x2"):
w13_weight = layer.w13_weight.view(torch.float4_e2m1fn_x2)
w2_weight = layer.w2_weight.view(torch.float4_e2m1fn_x2)
else:
w13_weight = layer.w13_weight
w2_weight = layer.w2_weight
output = fused_moe( output = fused_moe(
x, x,
layer.w13_weight, w13_weight,
layer.w2_weight, w2_weight,
topk_weights, topk_weights,
topk_ids, topk_ids,
quant_type=QuantType.per_1x32, quant_type=QuantType.per_1x32,

11
python/sglang/srt/layers/quantization/quark/quark_moe.py
View File

@@ -183,10 +183,17 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
moe_runner_config = self.moe_runner_config moe_runner_config = self.moe_runner_config
topk_weights, topk_ids, _ = topk_output topk_weights, topk_ids, _ = topk_output
if hasattr(torch, "float4_e2m1fn_x2"):
w13_weight = layer.w13_weight.view(torch.float4_e2m1fn_x2)
w2_weight = layer.w2_weight.view(torch.float4_e2m1fn_x2)
else:
w13_weight = layer.w13_weight
w2_weight = layer.w2_weight
output = fused_moe( output = fused_moe(
x, x,
layer.w13_weight, w13_weight,
layer.w2_weight, w2_weight,
topk_weights, topk_weights,
topk_ids, topk_ids,
quant_type=QuantType.per_1x32, quant_type=QuantType.per_1x32,