Skip to content

CollectiveX: footnote the H200 uccl-ep EP16 receive ring and correct the H100 cells / CollectiveX:为 H200 uccl-ep EP16 接收环添加脚注并修正 H100 单元格 - #1574

Merged
Oseltamivir merged 3 commits into
masterfrom
cx-uccl-h200-ring
Oct 30, 2026
Merged

Oseltamivir merged 3 commits into
masterfrom
cx-uccl-h200-ring

Conversation

@Oseltamivir

@Oseltamivir Oseltamivir commented Oct 29, 2026 •

Copy link
Copy Markdown
Contributor

Summary

H200 uccl-ep EP16 now has a measured prefill row. The cell's footnote says how it runs.

  • New footnote uccl-h200-rdma-recv-ring, replacing uccl-hseries-topend:
  • The old footnote was stale. It cited an August audit and compared against a B200 pool that has since been retired.

H100 cells, corrected for the EFA pool. Measured results now back each of these (run 1793321755):

Cell Before Now
uccl-ep normal EP16 broken (August init crash) works; decode and prefill rows are published
uccl-ep low-latency EP16 not enabled works over EFA (~3.4 ms floor at T=1); decode rows are published
DeepEP V2 normal EP16 works works, decode only: ~25 ms at T=512 through the NCCL GIN proxy, so the prefill ladder can't finish within the 90-minute case budget
nccl-ep low-latency EP16 not enabled works, decode only (~1.2 ms floor at T=1)
DeepEP V2 low-latency EP16 broken (virtualized-pod GDR) broken: needs NVSHMEM IBGDA, which EFA doesn't have; every rank segfaults at setup

Low-latency EP16 notes match the registry. The registry enables low-latency EP16 for uccl-ep and nccl-ep on H100 and H200, but the notes said no pool did:

  • H200's DeepEP V2, uccl-ep and nccl-ep low-latency EP16 cells, all with published rows, now show as working. The virtualized-GDR "held" note is removed.
  • ll-ep16-not-enabled now refers to the pool; only B200 uses it.
  • The nccl-ep low-latency notes list the real EP16 pools and include FP8.

The unused uccl-h100-init-crash and hseries-ll-gdr footnotes are removed. (The AMD cell corrections follow in #1580.) The CollectiveX component tests (73) pass, and oxfmt and oxlint are clean.

中文

H200 uccl-ep EP16 现已有实测的 prefill 结果。该单元格的脚注说明了其运行方式。

  • 新脚注 uccl-h200-rdma-recv-ring,取代 uccl-hseries-topend:
  • 旧脚注已过时。 它引用的是 8 月的审计,并与一个已退役的 B200 集群比较。

H100 单元格按 EFA 集群修正。 以下各项现均有实测结果支撑(运行 1793321755):

单元格 修正前 修正后
uccl-ep normal EP16 不可用(8 月初始化崩溃) 可用;decode 与 prefill 结果均已发布
uccl-ep 低延迟 EP16 未启用 可在 EFA 上运行(T=1 下限约 3.4 ms);decode 结果已发布
DeepEP V2 normal EP16 可用 可用,仅有 decode:经 NCCL GIN 代理 T=512 约 25 ms,prefill 梯级无法在 90 分钟单用例时限内完成
nccl-ep 低延迟 EP16 未启用 可用,仅有 decode(T=1 下限约 1.2 ms)
DeepEP V2 低延迟 EP16 不可用(虚拟化节点 GDR) 不可用:需要 NVSHMEM IBGDA,EFA 不支持;所有 rank 在初始化时段错误

低延迟 EP16 脚注与注册表一致。 注册表已在 H100 和 H200 上为 uccl-ep 与 nccl-ep 启用低延迟 EP16,但脚注称没有集群启用:

  • H200 的 DeepEP V2、uccl-ep 与 nccl-ep 低延迟 EP16 单元格均已有发布结果,现改为可用;移除虚拟化 GDR“暂不发布”脚注。
  • ll-ep16-not-enabled 改为指本集群,仅 B200 使用。
  • nccl-ep 低延迟脚注列出实际启用 EP16 的集群,并包含 FP8。

已移除未使用的 uccl-h100-init-crash 与 hseries-ll-gdr 脚注。(AMD 单元格修正见 #1580。)CollectiveX 组件测试(73 项)通过,oxfmt 与 oxlint 检查通过。

AI disclosure

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

The cell now runs with a 1040-token RDMA receive ring (InferenceX#4651); at
the default depth the two-node BF16 prefill dispatch deadlocks
(uccl-project/uccl#1110). The footnote it replaces described an August audit
against a since-retired B200 pool.

中文:为 H200 uccl-ep EP16 单元格添加接收环脚注。该单元格现使用 1040 token
的 RDMA 接收环运行(InferenceX#4651);默认深度下双节点 BF16 prefill 的
dispatch 会死锁(uccl-project/uccl#1110)。被取代的脚注描述的是 8 月针对一个
已退役 B200 集群的审计。
@vercel

vercel Bot commented Oct 29, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
inferencemax-app Ready Ready Preview Oct 30, 2026 2:36am UTC

Request Review

The H100 pool is AWS EFA. Measured results now back these cells:
- uccl-ep EP16 works in both modes; the August init-crash note is gone.
- DeepEP V2 normal EP16 is decode only: ~25 ms at T=512 through the NCCL
  GIN proxy, so the prefill ladder can't finish within the 90-minute
  case budget.
- nccl-ep low-latency EP16 is decode only, for the same reason (~1.2 ms
  floor at T=1).
- DeepEP V2 low-latency EP16 is broken: it needs NVSHMEM IBGDA, which EFA
  doesn't have, and every rank segfaults at setup.

中文:按 EFA 集群修正 H100 支持矩阵单元格。H100 集群为 AWS EFA,以下单元格现均有实测结果支撑:
- uccl-ep EP16 两种模式均可用,移除 8 月的初始化崩溃脚注。
- DeepEP V2 normal EP16 仅有 decode:经 NCCL GIN 代理 T=512 约 25 ms,
  prefill 梯级无法在 90 分钟单用例时限内完成。
- nccl-ep 低延迟 EP16 同理仅有 decode(T=1 下限约 1.2 ms)。
- DeepEP V2 低延迟 EP16 不可用:其需要 NVSHMEM IBGDA,EFA 不支持,
  所有 rank 在初始化时段错误。
@Oseltamivir Oseltamivir changed the title CollectiveX: footnote the H200 uccl-ep EP16 receive ring / CollectiveX:为 H200 uccl-ep EP16 接收环添加脚注 CollectiveX: footnote the H200 uccl-ep EP16 receive ring and correct the H100 cells / CollectiveX:为 H200 uccl-ep EP16 接收环添加脚注并修正 H100 单元格 Oct 30, 2026

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

Fix All in Cursor

❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.

Reviewed by Cursor Bugbot for commit 7264b5c. Configure here.

The registry enables low-latency EP16 for uccl-ep and nccl-ep on H100 and
H200, and those pools have measured, published rows. The notes still said
no pool enabled it.
- H100 uccl-ep low-latency EP16 gets its own EFA footnote (~3.4 ms at T=1).
- H200's DeepEP V2, uccl-ep and nccl-ep low-latency EP16 cells now show as
  working; the stale virtualized-GDR note is removed.
- ll-ep16-not-enabled now refers to the pool (only B200 uses it).
- The nccl-ep low-latency notes list the real EP16 pools and include FP8.

中文:使低延迟 EP16 单元格与注册表一致。注册表已在 H100 和 H200 上为 uccl-ep 与
nccl-ep 启用低延迟 EP16,且这些集群已有实测并发布的结果,但脚注仍称没有集群启用。
- H100 uccl-ep 低延迟 EP16 新增 EFA 脚注(T=1 约 3.4 ms)。
- H200 的 DeepEP V2、uccl-ep 与 nccl-ep 低延迟 EP16 单元格改为可用;移除过时的虚拟化 GDR 脚注。
- ll-ep16-not-enabled 改为指本集群(仅 B200 使用)。
- nccl-ep 低延迟脚注列出实际启用 EP16 的集群,并包含 FP8。
@Oseltamivir
Oseltamivir merged commit 0d52662 into master Oct 30, 2026
33 checks passed
@Oseltamivir
Oseltamivir deleted the cx-uccl-h200-ring branch October 30, 2026 02:46
@Oseltamivir Oseltamivir changed the title CollectiveX: footnote the H200 uccl-ep EP16 receive ring and correct the H100 cells / CollectiveX:为 H200 uccl-ep EP16 接收环添加脚注并修正 H100 单元格 CollectiveX: match the support matrix to measured results / CollectiveX:使支持矩阵与实测结果一致 Oct 30, 2026
@Oseltamivir Oseltamivir changed the title CollectiveX: match the support matrix to measured results / CollectiveX:使支持矩阵与实测结果一致 CollectiveX: footnote the H200 uccl-ep EP16 receive ring and correct the H100 cells / CollectiveX:为 H200 uccl-ep EP16 接收环添加脚注并修正 H100 单元格 Oct 30, 2026

This branch was successfully deployed

1 active deployment
Preview — cb3485c0 Deployed Oct 30, 2026 by vercel[bot]
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.

1 participant