Skip to content

fix(ep): support 12-rank HT combine #10

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
fergusfinn wants to merge 9 commits into from
Closed

fix(ep): support 12-rank HT combine #10

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
05a7d4d
fix(ep): drain CXI FIFO rings before quiet
fergusfinn Jun 17, 2026
a953b7b
ep: make HT internode dispatch+combine CUDA-graph replay-safe
fergusfinn Jun 10, 2026
d19ddf1
ep: expose per-expert recv counts as a device tensor in worst-tokens …
fergusfinn Jun 10, 2026
f52c2ac
ep: add CUDA-graph capture/replay correctness test for HT internode
fergusfinn Jun 10, 2026
efa5a9f
ep: graph-mode notify must not write the host-mapped counters
fergusfinn Jun 10, 2026
c63447f
ep: fix bench test teardown ordering (cudaErrorContextIsDestroyed abort)
fergusfinn Jun 10, 2026
5e65b25
ep/bench: accept+validate device recv-counts tensor in worst-tokens d…
Jun 10, 2026
398b4de
ep: add dispatch timeout counter diagnostics
fergusfinn Jun 11, 2026
2df5384
fix(ep): add 12-rank HT configs
fergusfinn Jun 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 20 additions & 2 deletions ep/bench/buffer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -754,6 +754,7 @@ def get_dispatch_config(num_ranks: int) -> Config:
2: Config(Buffer.num_sms, 24, 256, 6, 128),
4: Config(Buffer.num_sms, 6, 256, 6, 128),
8: Config(Buffer.num_sms, 6, 256, 6, 128),
12: Config(Buffer.num_sms, 6, 288, 6, 256),
16: Config(Buffer.num_sms, 36, 288, 20, 512 if Buffer._is_efa() else 128),
24: Config(Buffer.num_sms, 8, 288, 32, 128),
32: Config(Buffer.num_sms, 32, 288, 32, 512 if Buffer._is_efa() else 128),
Expand Down Expand Up @@ -781,7 +782,8 @@ def get_combine_config(num_ranks: int) -> Config:
config_map = {
2: Config(Buffer.num_sms, 10, 256, 6, 128),
4: Config(Buffer.num_sms, 9, 256, 6, 128),
8: Config(Buffer.num_sms, 4, 256, 6, 128),
8: Config(Buffer.num_sms, 4, 256, 8, 128),
12: Config(Buffer.num_sms, 4, 288, 12, 128),
16: Config(Buffer.num_sms, 4, 288, 12, 512 if Buffer._is_efa() else 128),
24: Config(Buffer.num_sms, 1, 288, 8, 128),
32: Config(Buffer.num_sms, 1, 288, 8, 512 if Buffer._is_efa() else 128),
Expand Down Expand Up @@ -1577,6 +1579,18 @@ def internode_dispatch(
dtype=torch.int32,
device=x.device,
)
# Device-visible per-expert counts for CUDA-graph mode: the
# host list is unavailable without a sync when num_worst_tokens
# is set, so the notify kernel also writes a device tensor.
recv_expert_counts = (
torch.zeros(
(int(num_tokens_per_expert.size(0)) // self.group_size,),
dtype=torch.int32,
device=x.device,
)
if num_worst_tokens > 0
else None
)
(
num_recv_tokens,
num_rdma_recv_tokens,
Expand All @@ -1595,6 +1609,7 @@ def internode_dispatch(
int(num_tokens_per_expert.size(0)),
int(expert_alignment),
int(num_worst_tokens),
0 if recv_expert_counts is None else recv_expert_counts.data_ptr(),
config,
rdma_channel_prefix_matrix.data_ptr(),
recv_rdma_rank_prefix_sum.data_ptr(),
Expand Down Expand Up @@ -1759,7 +1774,9 @@ def internode_dispatch(
(recv_x, recv_x_scales) if x_scales is not None else recv_x,
recv_topk_idx,
recv_topk_weights,
num_recv_tokens_per_expert_list,
recv_expert_counts
if recv_expert_counts is not None
else num_recv_tokens_per_expert_list,
handle,
EventOverlap(
event,
Expand Down Expand Up @@ -1848,6 +1865,7 @@ def internode_combine(
rdma_channel_prefix_matrix.data_ptr(),
rdma_rank_prefix_sum.data_ptr(),
gbl_channel_prefix_matrix.data_ptr(),
gbl_rank_prefix_sum.data_ptr(),
send_rdma_head.data_ptr(),
send_nvl_head.data_ptr(),
config,
Expand Down
Loading