CollectiveX: add FlashInfer one-sided NVLink EP for the GB SKUs / CollectiveX：为 GB SKU 新增 FlashInfer 单边 NVLink EP 后端 by Oseltamivir · Pull Request #2450 · SemiAnalysisAI/InferenceX

CollectiveX: label backend maturity and correct the low-latency docs

Oseltamivir

committed

98ef7f6

CollectiveX: add FlashInfer one-sided NVLink EP for the GB SKUs

Oseltamivir

committed

9423456

CollectiveX: give the FlashInfer communicator the whole CommBackend contract

Oseltamivir

committed

893e304

CollectiveX: dispatch routing metadata so the FlashInfer oracle views are real

Oseltamivir

committed

81efeae

CollectiveX: complete the FlashInfer oracle view (local masking + expert counts)

Oseltamivir

committed

3ff418a

CollectiveX: shape the FlashInfer combine scatter from the actual receive

Oseltamivir

committed

2f53134

CollectiveX: model FlashInfer's BF16 combine accumulation instead of widening the gate

Oseltamivir

committed

4237e23

CollectiveX: model FlashInfer's truncating BF16 conversion; revert bf16-rank-sum

Oseltamivir

committed

fbadba6

CollectiveX: truncate only where the kernel converts, not where the adapter does

Oseltamivir

committed

2319638

CollectiveX: sum the combine expectation in the same canonical order as the staged input

Oseltamivir

committed

8799ba0

CollectiveX: revert the canonical-ordering change (measured no-op)

Oseltamivir

committed

33cd01b

CollectiveX: reproduce FlashInfer's per-slot tree combine from the kernel source

Oseltamivir

committed

f4454ce

CollectiveX: revert the per-slot tree combine model (measured no-op)

Oseltamivir

committed

27dfba1

CollectiveX: build the expected per-rank message with the same function that stages it

Oseltamivir

committed

78e6ad3

CollectiveX: revert the shared _expert_transform expectation (measured no-op)

Oseltamivir

committed

5e06578

CollectiveX: drop the truncating output model for flashinfer-ep

Oseltamivir

committed

badb8a3

CollectiveX: stage the flashinfer combine in place, not through a fresh buffer

Oseltamivir

committed

0d9d66f

CollectiveX: submit the flashinfer combine payload from the workspace region

Oseltamivir

committed

19f0049

CollectiveX: model FlashInfer's payload-dtype combine reduction

Oseltamivir

committed

4a34895

CollectiveX: tighten the flashinfer wheel probe and stage docstring

Oseltamivir

committed

6e7d2fa

Merge remote-tracking branch 'origin/main' into collectivex-flashinfer-ep

Oseltamivir

committed

92c7d0c

CollectiveX: document FlashInfer EP and its payload-dtype combine

Oseltamivir

committed

c12fdeb

CollectiveX: drop the disproven truncating-store model

Oseltamivir

committed

0bef5e9

CollectiveX: run the unit tests on every PR

Oseltamivir

committed

6d05435

CollectiveX: note the slot-tree oracle's memory scaling

Oseltamivir

committed

49d25ec

Navigation Menu

Uh oh!

CollectiveX: add FlashInfer one-sided NVLink EP for the GB SKUs / CollectiveX：为 GB SKU 新增 FlashInfer 单边 NVLink EP 后端 - #2450

CollectiveX: add FlashInfer one-sided NVLink EP for the GB SKUs / CollectiveX：为 GB SKU 新增 FlashInfer 单边 NVLink EP 后端#2450
Oseltamivir merged 29 commits into
mainfrom
collectivex-flashinfer-ep

Commits on Aug 2, 2026

CollectiveX: label backend maturity and correct the low-latency docs

CollectiveX: add FlashInfer one-sided NVLink EP for the GB SKUs

CollectiveX: give the FlashInfer communicator the whole CommBackend contract

CollectiveX: dispatch routing metadata so the FlashInfer oracle views are real

CollectiveX: complete the FlashInfer oracle view (local masking + expert counts)

CollectiveX: shape the FlashInfer combine scatter from the actual receive

CollectiveX: model FlashInfer's BF16 combine accumulation instead of widening the gate

CollectiveX: model FlashInfer's truncating BF16 conversion; revert bf16-rank-sum

CollectiveX: truncate only where the kernel converts, not where the adapter does

CollectiveX: sum the combine expectation in the same canonical order as the staged input

CollectiveX: revert the canonical-ordering change (measured no-op)

CollectiveX: reproduce FlashInfer's per-slot tree combine from the kernel source

CollectiveX: revert the per-slot tree combine model (measured no-op)

CollectiveX: build the expected per-rank message with the same function that stages it

CollectiveX: revert the shared _expert_transform expectation (measured no-op)

CollectiveX: drop the truncating output model for flashinfer-ep

CollectiveX: stage the flashinfer combine in place, not through a fresh buffer

CollectiveX: submit the flashinfer combine payload from the workspace region

CollectiveX: model FlashInfer's payload-dtype combine reduction

CollectiveX: tighten the flashinfer wheel probe and stage docstring

Merge remote-tracking branch 'origin/main' into collectivex-flashinfer-ep

CollectiveX: document FlashInfer EP and its payload-dtype combine

CollectiveX: drop the disproven truncating-store model

CollectiveX: run the unit tests on every PR

CollectiveX: note the slot-tree oracle's memory scaling

Commits on Aug 3, 2026

CollectiveX: resolve review — untimed routing cast, ready_for_review trigger

Merge remote-tracking branch 'origin/main' into collectivex-flashinfer-ep

CollectiveX: re-sync the run_ep parser mirror and guard it against drift

Merge remote-tracking branch 'origin/main' into collectivex-flashinfer-ep