Repository navigation

feat(vllm): support refit-safe MXFP8 linear backends - #3638

Open

seonjinn wants to merge 51 commits into

NVIDIA-NeMo:mainfrom

seonjinn:sna/mxfp8-linear-backend-refit

Open

feat(vllm): support refit-safe MXFP8 linear backends#3638
seonjinn wants to merge 51 commits into
NVIDIA-NeMo:mainfrom
seonjinn:sna/mxfp8-linear-backend-refit

Commits on Aug 13, 2026
- fix(vllm): support MXFP8 MoE padding
  seonjinn
  committed
  9e96522
- test(vllm): cover Nano MXFP8 padding lifecycle
  seonjinn
  committed
  cf30aac
- fix(vllm): preserve gated MXFP8 shard layout
  seonjinn
  committed
  bfca5e5
- test(vllm): tighten MXFP8 padding coverage
  seonjinn
  committed
  8915578
- fix(vllm): finalize checkpoint engine refits
  seonjinn
  committed
  f25dacf
- test(vllm): initialize TP metadata for padding test
  seonjinn
  committed
  b50af2a
- test(vllm): centralize TP metadata fixture
  seonjinn
  committed
  4357e59
- test(vllm): validate MXFP8 MoE padding contract
  seonjinn
  committed
  f537781
- test(vllm): cover grouped MXFP8 refit shapes
  seonjinn
  committed
  969e3fb
- fix(vllm): preserve grouped MXFP8 refit shapes
  seonjinn
  committed
  f1c78bc
- test(vllm): cover grouped MXFP8 refit loading
  seonjinn
  committed
  3f6b636
- fix(vllm): finalize checkpoint refits once
  seonjinn
  committed
  80422f0
- test(vllm): verify MXFP8 padding parity
  seonjinn
  committed
  1c32863
- fix(vllm): preserve MXFP8 zero scale exponents
  seonjinn
  committed
  14ddf7d
- fix(vllm): patch MXFP8 workers under Ray executor v2
  seonjinn
  committed
  6a215f3
- fix(vllm): pass FP8 config to Ray v2 actors
  seonjinn
  committed
  f62bd5f
- fix(vllm): isolate Ray v2 FP8 worker configs
  seonjinn
  committed
  31390d4
- test(vllm): serialize Ray v2 FP8 workers
  seonjinn
  committed
  aabb197
Commits on Aug 14, 2026
- test(vllm): cover MXFP8 linear backend refit
  seonjinn
  committed
  fe8811d
- feat(vllm): preserve MXFP8 linear backend refit
  seonjinn
  committed
  504b99f
- fix(vllm): preserve non-module refit test doubles
  seonjinn
  committed
  6bbb623
- fix(vllm): abort incomplete MXFP8 reloads
  seonjinn
  committed
  9a3a2ca
- test(vllm): initialize TP rank for MXFP8 backend test
  seonjinn
  committed
  1e74704
- refactor(vllm): identify MXFP8 kernels by type
  seonjinn
  committed
  eb05783
Commits on Aug 18, 2026
- Merge branch 'main' into sna/mxfp8-linear-backend-refit
  seonjinn
  committed
  d073cda
- Merge branch 'main' into sna/mxfp8-moe-padding
  seonjinn
  committed
  0440e4f
- style: format MXFP8 backend test
  seonjinn
  committed
  2c04e11
- Merge branch 'sna/mxfp8-moe-padding' into sna/mxfp8-linear-backend-refit
  seonjinn
  committed
  7eb5ef0
- Merge remote-tracking branch 'origin/main' into review/pr3638-merge-only-20260817
  seonjinn
  committed
  a19ae31
Commits on Sep 12, 2026
- Merge commit '87a0aa606d6941f265baa64be2d6351236152081' into HEAD
  seonjinn
  committed
  800d3d8
Commits on Sep 28, 2026
- Merge latest main into PR #3638
  seonjinn
  committed
  78a49b6
- fix(vllm): unify mixed native refit targets
  seonjinn
  committed
  4d6f05a
- Merge latest main into PR #3638
  seonjinn
  committed
  5df5c4a
Commits on Sep 30, 2026
- Merge latest main into PR #3638
  seonjinn
  committed
  eadd135
- Merge latest main into PR #3638
  seonjinn
  committed
  f52a8de
Commits on Oct 1, 2026
- Merge remote-tracking branch 'upstream/main' into codex/refresh-pr3638-vllm029-clean-20261001
  seonjinn
  committed
  ebbef7f
- fix(vllm): align native MXFP8 refit with 0.29
  seonjinn
  committed
  1dca77b
Commits on Oct 3, 2026
- chore: merge latest main for vLLM 0.29 compatibility
  seonjinn
  committed
  c79116d
Commits on Oct 4, 2026
- chore: merge latest main for vLLM 0.29 (PR3638)
  seonjinn
  committed
  32c8288
Commits on Oct 12, 2026
- Merge latest main and cover MXFP8 collective reload context
  seonjinn
  committed
  8a39ac8
Commits on Oct 15, 2026
- Merge upstream/main into sna/mxfp8-linear-backend-refit
  seonjinn
  committed
  5f7e957
Commits on Oct 20, 2026
- Merge latest main into sna/mxfp8-linear-backend-refit
  seonjinn
  committed
  3e79c96
Commits on Oct 21, 2026
- Address #3638 review: refit guards, import scope, tests
  seonjinn
  committed
  91a7da2
- Reject DSV4 FP8 on nccl-reshard and harden qkv refit test
  seonjinn
  committed
  96105c3
- Tolerate refit test doubles in MXFP8 reload root scan
  seonjinn
  committed
  cf947cc
Commits on Oct 22, 2026
- Add GB200 functional legs for native MXFP8 linear refit
  seonjinn
  committed
  893a738
Commits on Oct 23, 2026
- test: enable CUDA graphs in all GB200 MXFP8 refit legs
  seonjinn
  committed
  a7eeef4
Commits on Oct 25, 2026
- Give NVFP4 IPC reload test stub a vLLM config
  seonjinn
  committed
  88f6357
- Merge remote-tracking branch 'origin/main' into sna/mxfp8-linear-backend-refit-review-fixes
  seonjinn
  committed
  2870907
Commits on Oct 27, 2026
- Merge latest main into sna/mxfp8-linear-backend-refit
  seonjinn
  committed
  83c655c
Commits on Oct 30, 2026
- Merge upstream/main into sna/mxfp8-linear-backend-refit
  seonjinn
  committed
  4d1fb42

Commits

Commits on Aug 13, 2026

Commits on Aug 14, 2026

Commits on Aug 18, 2026

Commits on Sep 12, 2026

Commits on Sep 28, 2026

Commits on Sep 30, 2026

Commits on Oct 1, 2026

Commits on Oct 3, 2026

Commits on Oct 4, 2026

Commits on Oct 12, 2026

Commits on Oct 15, 2026

Commits on Oct 20, 2026

Commits on Oct 21, 2026

Commits on Oct 22, 2026

Commits on Oct 23, 2026

Commits on Oct 25, 2026

Commits on Oct 27, 2026

Commits on Oct 30, 2026