perf: Reduce MXFP8 refit weight-transfer overheads - #3294

Open

seonjinn wants to merge 81 commits into

NVIDIA-NeMo:mainfrom

seonjinn:sna/pr-mxfp8-refit-optimization

Commits on Jul 21, 2026

perf(vllm): optimize quantized refit paths
seonjinn
committed
b88aa2c
Merge branch 'main' into sna/pr-mxfp8-refit-optimization
seonjinn
authored
438e20d

Commits on Jul 27, 2026

Merge branch 'main' into sna/pr-mxfp8-refit-optimization
seonjinn
committed
96ba299
fix(trtllm): align refit metadata interface
seonjinn
committed
0fb59b6
test(vllm): include MXFP8 refit tests in L0
seonjinn
committed
c4a5e14
fix(refit): complete prequant metadata handshake
seonjinn
committed
5c53bd4
test(refit): cover MXFP8 optimization paths
seonjinn
committed
98fc197
test(refit): fix vLLM test import order
seonjinn
committed
b237fd2

Commits on Jul 28, 2026

Merge upstream/main into sna/pr-mxfp8-refit-optimization
seonjinn
committed
d1f9fc9
test(fp8): cover MXFP8 MoE padding path
seonjinn
committed
ab030c3
Merge upstream/main into sna/pr-mxfp8-refit-optimization
seonjinn
committed
2c147ac

Commits on Jul 29, 2026

Commits on Jul 30, 2026

Commits on Jul 31, 2026

Merge remote-tracking branch 'upstream/main' into sna/pr-mxfp8-refit-optimization-squashed
seonjinn
committed
2339686
fix(vllm): preserve MXFP8 refit on vLLM 0.25
seonjinn
committed
5880499
style(vllm): sort MXFP8 backend import
seonjinn
committed
123cc07
fix(vllm): allow partial configs in quant validation
seonjinn
committed
ef3fa08
Merge remote-tracking branch 'upstream/main' into sna/pr-mxfp8-refit-optimization-squashed
seonjinn
committed
7989113
feat(recipe): add async Qwen3 30B MXFP8 rollout
seonjinn
committed
c4a7b06
feat(recipe): add async Qwen3 235B MXFP8 rollout
seonjinn
committed
69f7995
ci: allowlist async recipe result path
seonjinn
committed
d53bfdf
fix(vllm): initialize async driver FP8 config
seonjinn
committed
2405bf5
fix(vllm): propagate FP8 config to refit workers
seonjinn
committed
0a11e94
fix(recipe): keep Qwen router gate in BF16
seonjinn
committed
81ce860
fix(recipe): exclude Qwen MoE routers from MXFP8
seonjinn
committed
9e43378

Commits on Aug 1, 2026

fix(vllm): patch FP8 in RayExecutorV2 workers
seonjinn
committed
6e582b8
fix(vllm): serialize FP8 worker pre-init hook
seonjinn
committed
93394c2
test(vllm): share RayExecutorV2 patch recorder
seonjinn
committed
f3c9196
feat(vllm): configure refit runtime optimizations
seonjinn
committed
e3384d2
fix(refit): propagate runtime configuration
seonjinn
committed
c4d453e
Merge remote-tracking branch 'upstream/main' into sna/pr-mxfp8-refit-optimization-squashed
seonjinn
committed
d43372e
fix(refit): cover standalone algorithm paths
seonjinn
committed
2b80a85
fix(vllm): scope refit optimization config
seonjinn
committed
0b6b5cc
fix(refit): allow matching fp8 reshard storage
seonjinn
committed
cd7732c

Commits on Aug 13, 2026

Commits on Aug 18, 2026

Merge remote-tracking branch 'origin/main' into review/pr3294-latest-20260817
seonjinn
committed
88721ce

Commits on Aug 25, 2026

Commits on Aug 26, 2026

Merge branch 'main' into pr-3294-fixes
seonjinn
committed
1d6a10a
fix(dynamo): align prepare_refit_info override with interface
seonjinn
committed
8dc4c36
Merge branch 'main' into sna/pr-mxfp8-refit-optimization
seonjinn
committed
db442ef
Merge main into sna/pr-mxfp8-refit-optimization
seonjinn
committed
4427f4c
fix: resolve CI unit-test failures after main merge
seonjinn
committed
c162cfd
Merge remote branch update, keeping local merge resolution and CI fixes
seonjinn
committed
382fa8b
Merge main into sna/pr-mxfp8-refit-optimization
seonjinn
committed
125f178
fix(refit): derive prequantize metadata without re-exporting weights
seonjinn
committed
e28239c
Fix pyrefly no-matching-overload false positive on masked_fill
seonjinn
committed
232da17

Commits on Aug 27, 2026

test: update refit cleanup fixture after main merge
seonjinn
committed
6b2ab6e
refactor(refit): limit optimized recipes to sync RL
seonjinn
committed
06b6fbb

Commits on Sep 4, 2026

Merge remote-tracking branch 'origin/main' into codex/pr3294-push-20260904
seonjinn
committed
7c9b881
fix(mxfp8): handle mixed grouped expert refit
seonjinn
committed
bf03429
test(mxfp8): align refit fixtures with vllm
seonjinn
committed
d9246a9
test(mxfp8): model routed expert config
seonjinn
committed
443a9f9
fix(vllm): reroute grouped MXFP8 scale sidecars
seonjinn
committed
b8159de
test(mxfp8): patch modular kernel factory
seonjinn
committed
8d38b28
test(mxfp8): use block-aligned hidden size
seonjinn
committed
658d385
test(mxfp8): exercise unpadded kernel reuse
seonjinn
committed
172d346
test(mxfp8): use current refit loader contract
seonjinn
committed
546d346
test(mxfp8): update remaining loader fixtures
seonjinn
committed
db15416

Commits on Sep 7, 2026

test(refit): reject native loader cache wrapping
seonjinn
committed
1d4d013
fix(refit): preserve vLLM online loaders
seonjinn
committed
1b5e642
Merge remote-tracking branch 'origin/main' into codex/pr3294-push-20260904
seonjinn
committed
f377f0a
test(refit): avoid duplicate kv scale finalization
seonjinn
committed
b7d4ce6
test(refit): avoid duplicate fp8 kv finalization
seonjinn
committed
201ad9e
test(refit): use a realized module fixture
seonjinn
committed
7438153
test(refit): expect serialized FP8 config
seonjinn
committed
08e7d10
test(refit): follow collective sender contract
seonjinn
committed
d003483
fix(refit): preserve MXFP8 kernel scale storage
seonjinn
committed
238f241
test(refit): update offload call expectation
seonjinn
committed
b72f4d4

Commits on Sep 12, 2026

Merge branch 'main' into sna/pr-mxfp8-refit-optimization
seonjinn
committed
bc024e8
style: apply lint fixes
seonjinn
committed
0122151