biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix][EP+DP] Use pplx-kernel internode instead of intranode (#19034)

Browse Source 
Signed-off-by: Tyler Michael Smith <tysmith@redhat.com>
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
Tyler Michael Smith
2025-06-02 23:36:51 -04:00
committed by GitHub
parent 5bc1ad6cee
commit 8a57872b2a
2 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/layers/fused_moe/layer.py
View File

@@ -269,9 +269,13 @@ class FusedMoEMethodBase(QuantizeMethodBase):
hidden_dim_scale_bytes=(0 if moe.in_dtype.itemsize != 1 else (
(moe.hidden_dim + moe.block_size - 1) // moe.block_size *
torch.float32.itemsize)),
group_name=all2all_manager.cpu_group.group_name,
)
# Intranode pplx a2a takes a group name while internode does not.
if not all2all_manager.internode:
all_to_all_args[
"group_name"] = all2all_manager.cpu_group.group_name
handle = all2all_manager.get_handle(all_to_all_args)
prepare_finalize = PplxPrepareAndFinalize(