Skip to content

CollectiveX: give b200-nscale DeepEP V2 its RDMA rate from the registry / CollectiveX：从注册表为 b200-nscale 的 DeepEP V2 提供 RDMA 速率 #4639

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

Merged
Oseltamivir merged 1 commit into from
Oct 29, 2026
Merged

CollectiveX: give b200-nscale DeepEP V2 its RDMA rate from the registry / CollectiveX：从注册表为 b200-nscale 的 DeepEP V2 提供 RDMA 速率 #4639

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
4 changes: 2 additions & 2 deletions collectivex/bench/ep_deepep_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,7 +136,7 @@ class DeepEPV2Backend(LegacyBufferLL, EPBackend):
# vLLM's low-latency send/receive split (see LegacyBufferLL).
ll_recv_hook = True
# The num_sms the timed dispatch passes: 0 hands the count to DeepEP's model, as vLLM does;
# create_buffer overrides it only on EFA.
# create_buffer overrides it where the registry sets rdma_gbs.
_dispatch_num_sms = 0

@staticmethod
Expand Down Expand Up @@ -380,7 +380,7 @@ def dispatch(self, p):
num_experts=self.args.experts,
num_max_tokens_per_rank=self._dispatch_capacity(p.T),
expert_alignment=1,
num_sms=self._dispatch_num_sms, # 0 (DeepEP's model) except on EFA; see create_buffer
num_sms=self._dispatch_num_sms, # 0 (DeepEP's model) unless rdma_gbs is set
async_with_compute_stream=False,
do_handle_copy=True,
do_cpu_sync=False,
Expand Down
3 changes: 2 additions & 1 deletion collectivex/configs/platform_config.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,7 +65,8 @@
},
"network": {
"socket_ifname": "bond0",
"rdma_devices": "mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_5,mlx5_10,mlx5_11"
"rdma_devices": "mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_5,mlx5_10,mlx5_11",
"rdma_gbs": 50
}
},
"b300": {
Expand Down
4 changes: 2 additions & 2 deletions collectivex/docs/methodology.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,8 +182,8 @@ when the LSA team is smaller than the world. x86 EP16 is that hybrid path, GIN a
domains of eight scale-up ranks each. GB EP16 stays inside MNNVL and uses LSA. Hybrid EP16 allocates
33 GIN QPs per peer (a fixed 512-QP budget) instead of the automatic 65/129, which exhausts the
per-NIC QP budget on CX-7. DeepEP's SM model reads the scale-out rate from an `ibstat` probe that cannot
see EFA and that the vLLM image does not ship, so on the EFA pools and on h200-dgxc the registry's
`network.rdma_gbs` supplies the per-GPU rate and the resulting count is passed explicitly
see EFA and that the vLLM image does not ship, so on the EFA pools, h200-dgxc and b200-nscale the
registry's `network.rdma_gbs` supplies the per-GPU rate and the resulting count is passed explicitly
(`library_config.sm_model_rdma_gbs`).

**A virtualized pool can make a scale-out row measure the hypervisor.** ElasticBuffer registers its GIN
Expand Down
Loading