Repository navigation
CollectiveX: give h200-dgxc DeepEP V2 its RDMA rate from the registry / CollectiveX：由注册表为 h200-dgxc 的 DeepEP V2 提供 RDMA 速率 - #4615
Merged
Conversation
The vLLM v0.32.0 image does not ship ibstat, so DeepEP V2's SM model read a scale-out rate of 0 on h200-dgxc and normal-mode EP16 died with ZeroDivisionError in get_theoretical_num_sms. The registry now sets rdma_gbs to 50 GB/s, the value the ibstat probe derives from a 400 Gb/s ConnectX-7 port, as the EFA pools already do. vLLM v0.32.0 镜像不包含 ibstat，导致 h200-dgxc 上 DeepEP V2 的 SM 模型读到的 scale-out 速率为 0， normal 模式 EP16 在 get_theoretical_num_sms 中因 ZeroDivisionError 失败。注册表现在将 rdma_gbs 设为 50 GB/s（即 ibstat 探测从 400 Gb/s ConnectX-7 端口得出的值），与 EFA 资源池的做法一致。
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
DeepEP V2 normal-mode EP16 on h200-dgxc died with
ZeroDivisionErroringet_theoretical_num_sms. The SM model reads the scale-out rate fromibstat, which the vLLM v0.32.0 image does not ship, so the rate came back 0 and the hybrid branch divided by it.platform_config.json: h200-dgxc's network block setsrdma_gbs: 50, the value theibstatprobe derives from a 400 Gb/s ConnectX-7 port (rate / 8). The EFA pools already supply their rate this way, and the harness passes it to DeepEP and records it aslibrary_config.sm_model_rdma_gbs.Validated on h200-dgxc:
h200-dgxc-deepep-v2-bf16-n2now succeeds (decode and prefill EP16), where it failed on both cases before.cases.jsonis unchanged;python -m pytest collectivex/testspasses.中文
h200-dgxc 上 DeepEP V2 normal 模式 EP16 在
get_theoretical_num_sms中因ZeroDivisionError失败。SM 模型通过ibstat读取 scale-out 速率，而 vLLM v0.32.0 镜像不包含ibstat，速率返回 0，hybrid 分支随之除零。platform_config.json：h200-dgxc 的 network 块设置rdma_gbs: 50，即ibstat探测从 400 Gb/s ConnectX-7 端口得出的值（速率 / 8）。EFA 资源池已用同样方式提供速率，harness 会将其传给 DeepEP，并记录为library_config.sm_model_rdma_gbs。已在 h200-dgxc 上验证：
h200-dgxc-deepep-v2-bf16-n2现已成功（decode 与 prefill EP16），此前两个用例均失败。cases.json未变；python -m pytest collectivex/tests通过。AI disclosure
Drafted and tested with an AI coding assistant; reviewed by the author.