[Oseltamivir]

Summary

On b200-nscale, DeepEP V2 normal-mode EP16 (b200-nscale-deepep-v2-{bf16,fp8}-n2) fails in create_buffer:

deep_ep/buffers/elastic.py:809 get_theoretical_num_sms
    if self.num_scaleout_ranks > 1 and (rdma_traffic / rdma_gbs) > (nvlink_traffic / nvlink_gbs):
ZeroDivisionError: float division by zero
Failed to get RDMA connection speed: [Errno 2] No such file or directory: 'ibstat'

The vLLM v0.32.0 image has no ibstat, so DeepEP's SM model reads a scale-out rate of 0. This is the same failure #4615 fixed on h200-dgxc.

• platform_config.json: b200-nscale network.rdma_gbs = 50, the per-GPU rate of its 400 Gb/s InfiniBand rails (what the ibstat probe would derive).
• methodology.md lists b200-nscale with the pools that take the rate from the registry.
• Two comments in ep_deepep_v2.py that said the override is EFA-only now say it applies wherever rdma_gbs is set.

EP8 is single-node, so the model never uses the scale-out rate there and its SM count is unchanged. Low-latency mode does not use the model.

python -m pytest collectivex/tests: 249 passed, 12 skipped. The two shards will be rerun on b200-nscale from this branch.

中文

在 b200-nscale 上，DeepEP V2 normal 模式 EP16（b200-nscale-deepep-v2-{bf16,fp8}-n2）在 create_buffer 中失败：DeepEP 的 SM 模型因 vLLM v0.32.0 镜像不含 ibstat 而读到 scale-out 速率为 0，在 get_theoretical_num_sms 中触发 ZeroDivisionError。这与 #4615 在 h200-dgxc 上修复的问题相同。

• platform_config.json：b200-nscale 设置 network.rdma_gbs = 50，即其 400 Gb/s InfiniBand 链路的每 GPU 速率（也是 ibstat 探测会得出的值）。
• methodology.md 将 b200-nscale 列入从注册表获取速率的资源池。
• ep_deepep_v2.py 中两处称该覆盖仅适用于 EFA 的注释，改为适用于所有设置了 rdma_gbs 的情况。

EP8 为单节点，模型不使用 scale-out 速率，SM 数量不变；low-latency 模式不使用该模型。

python -m pytest collectivex/tests：249 通过，12 跳过。两个分片将在 b200-nscale 上基于此分支重新运行。

AI disclosure

This change was diagnosed and written with an AI coding assistant, and reviewed by the author.