[fix] Fix mxfp4 triton MoE tp bug (#9473)

Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>
This commit is contained in:
hlu1
2025-08-23 01:48:40 -07:00
committed by GitHub
parent c9dd70fbde
commit ccd3fb946e
3 changed files with 14 additions and 12 deletions
+5 -6
View File
@@ -793,12 +793,11 @@ class GptOssForCausalLM(nn.Module):
intermediate_size % mxfp4_block == 0
), f"{intermediate_size=} must be divisible by {mxfp4_block=}"
intermediate_size_block = intermediate_size // mxfp4_block
if _is_sm100_supported:
per_rank_intermediate_size_block = math.ceil(
intermediate_size_block / moe_tp_size
)
else:
per_rank_intermediate_size_block = intermediate_size_block // moe_tp_size
per_rank_intermediate_size_block = math.ceil(
intermediate_size_block / moe_tp_size
)
per_rank_intermediate_size = per_rank_intermediate_size_block * mxfp4_block
# Calculate common slicing bounds for current rank