Skip to content

Fix Python ep_test.py HT mode (under-sized recv buffers → ncclInvalidUsage) - #2252

Open
ybenvidia wants to merge 1 commit into
NVIDIA:devfrom
ybenvidia:dev
Open

Fix Python ep_test.py HT mode (under-sized recv buffers → ncclInvalidUsage)#2252
ybenvidia wants to merge 1 commit into
NVIDIA:devfrom
ybenvidia:dev

Conversation

@ybenvidia

Copy link
Copy Markdown

In HT mode, ep_test.py set max_recv_tokens_per_rank = num_tokens * n_ranks (missing the * top_k factor) and sized the recv tensor as max_dispatch_tokens_per_rank * num_local_experts. Both under-size the HT receive buffer, so ncclEpDispatch fails with ncclInvalidUsage. ep_test.cu uses nRanks * num_tokens * top_k and num_recv_tokens = max_recv_tokens_per_rank. This aligns the Python reference with the C one; HT then passes 16/16 ranks (dispatch + combine verified). LL unchanged.

Signed-off-by: ybenabou <ybenabou@nvidia.com>
@artpol84

Copy link
Copy Markdown
Contributor

/mirror

@artpol84

Copy link
Copy Markdown
Contributor

/mirror dev

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants