minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion] fix: per-shard FP8 scale shape for single-GPU fused linears (#32157)

Browse Source
This commit is contained in:
Yihao Wang
2026-07-28 20:28:01 -07:00
committed by GitHub
parent b21cc8eb9b
commit 9ea964a535
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/multimodal_gen/runtime/layers/linear.py
+3 -3
View File
@@ -490,8 +490,9 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
tp_group: dist.ProcessGroup = None, tp_group: dist.ProcessGroup = None,
): ):
tp_group = tp_group or get_tp_group() tp_group = tp_group or get_tp_group()
if get_group_size(tp_group) > 1: # Set output_sizes BEFORE super().__init__() so ColumnParallelLinear derives
self.output_sizes = output_sizes # per-shard output_partition_sizes.
self.output_sizes = output_sizes
super().__init__( super().__init__(
input_size=input_size, input_size=input_size,
output_size=sum(output_sizes), output_size=sum(output_sizes),
@@ -503,7 +504,6 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
prefix=prefix, prefix=prefix,
tp_group=tp_group, tp_group=tp_group,
) )
self.output_sizes = output_sizes
assert all(output_size % self.tp_size == 0 for output_size in output_sizes) assert all(output_size % self.tp_size == 0 for output_size in output_sizes)
def weight_loader( def weight_loader(