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

Reintroduce memory usage fix (#9535)

Browse Source
This commit is contained in:
fzyzcjy
2025-08-25 15:02:31 +08:00
committed by GitHub
parent fda4792620
commit 433266c125
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/layers/quantization/modelopt_quant.py
View File

@@ -1212,11 +1212,13 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
# Process w13 weights # Process w13 weights
w13_blockscale_swizzled = self.swizzle_blockscale(layer.w13_weight_scale) w13_blockscale_swizzled = self.swizzle_blockscale(layer.w13_weight_scale)
del layer.w13_weight_scale
layer.w13_blockscale_swizzled.data.copy_(w13_blockscale_swizzled) layer.w13_blockscale_swizzled.data.copy_(w13_blockscale_swizzled)
layer.w13_weight = Parameter(layer.w13_weight.data, requires_grad=False) layer.w13_weight = Parameter(layer.w13_weight.data, requires_grad=False)
# Process w2 weights # Process w2 weights
w2_blockscale_swizzled = self.swizzle_blockscale(layer.w2_weight_scale) w2_blockscale_swizzled = self.swizzle_blockscale(layer.w2_weight_scale)
del layer.w2_weight_scale
layer.w2_blockscale_swizzled.data.copy_(w2_blockscale_swizzled) layer.w2_blockscale_swizzled.data.copy_(w2_blockscale_swizzled)
layer.w2_weight = Parameter(layer.w2_weight.data, requires_grad=False) layer.w2_weight = Parameter(layer.w2_weight.data, requires_grad=False)