CollectiveX: wire-basis payload bandwidth + nccl-ep LL hold in the support matrix - #953
Merged
Merged
Conversation
…nccl-ep LL cells - reader.ts prefers wire_byte_provenance (per-(token,expert) for the LL layouts that do not rank-deduplicate; InferenceX#2786) over the deduplicated byte_provenance, whose rates were a lower bound presented as wire bandwidth (34% low on nccl-ep LL EP8 at T=128). Pre-wire artifacts fall back and only ever understate. - Known-support matrix: nccl-ep low-latency EP8 cells flip to broken with an honest note — the T<=128 clamp only reduced exposure to the un-fenced combine race (DeepEP #642), it was never a safety boundary; rows are held until a fenced wheel ships (registry hold in InferenceX#2786). - MoRI LL note updated: the benchmark now measures AsyncLL split-phase, the kernel SGLang deploys (validated on mi355x 2026-08-30).
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
…opk-6 assert) Bisected on-metal with a pure mori.ops probe on mia1-p01-g18: AsyncLL passes at top-k 8 with both 256 and 384 experts and dies at top-k 6 with both, on the device assert (pe >= 0) && (pe < worldSize) in EpDispatchLowLatencyAsyncRecvCopyMultiBlock. Known upstream: ROCm/mori#505 ('AsyncLL slot assignment double-allocates when top-k does not divide warpSize', merged 2026-07-31); every shipped mi35x image predates it (newest mori-0706). Cells flip green with the image bump; no new upstream issue needed.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Companion to SemiAnalysisAI/InferenceX#2786 (measurement-side).
packages/db/src/collectivex/reader.tsmapped rates only frombyte_provenance, which is rank-deduplicated — for DeepEP/UCCL/NCCL low-latency layouts (one copy per token–expert assignment) that published a lower bound as the wire rate, 34.1% low on nccl-ep LL EP8 at T=128 (77.4 vs 117.4 MB per direction), and made LL GB/s incomparable across backends. The reader now prefers the newwire_byte_provenancefield; artifacts written before it fall back and only ever understate.Tests: reader/known-support/data suites 72 pass (new test pins the wire-basis preference and payload_bytes); tsc clean on db+app.
Note
Low Risk
Changes affect CollectiveX metric interpretation and curated support labels only; no auth, persistence, or runtime execution paths.
Overview
Payload bandwidth now derives from
wire_byte_provenancewhen artifacts include it, instead of rank-deduplicatedbyte_provenancealone—so low-latency token–expert layouts no longer show understated wire GB/s; older shards still fall back and only understate. A reader test and fixture hook (wireBytesFactor) lock in that preference forpayload_bytesand payload rates.The known-support matrix reclassifies nccl-ep low-latency from “works with T≤128 clamp” to held/broken (
nccl-ll-fence-race) across NVIDIA SKUs, and marks MoRI LL on mi3xx as broken for AsyncLL top-k 6 (mori-asyncll-topk6) while updating the scale-up-only footnote to describe AsyncLL split-phase.Reviewed by Cursor Bugbot for commit 1728eb7. Bugbot is set up for automated code reviews on this repo. Configure here.