Repository navigation

fix(inference): make online FP8 weight updates safe and unbiased - #3628

Closed

mikasenghaas wants to merge 10 commits into

fix/muon-uneven-fsdp-paddingfrom

fix/abort-stale-vllm-requests

Closed

fix(inference): make online FP8 weight updates safe and unbiased#3628
mikasenghaas wants to merge 10 commits into
fix/muon-uneven-fsdp-paddingfrom
fix/abort-stale-vllm-requests

Commits on Sep 22, 2026
- fix(inference): abort stale requests before weight updates
  mikasenghaas
  committed
  5748803
- feat(inference): add fp8_ue8m0_weight_scales for power-of-two online FP8 weight scales
  
  authored and
  mikasenghaas
  committed
  1a228d1
- fix(inference): honour the swiglu clamp on vLLM's TritonExperts fp8 fast path
  
  authored and
  mikasenghaas
  committed
  6f6c3e6
- feat(inference): add fp8_stochastic_weight_rounding for unbiased online FP8 weights
  
  authored and
  mikasenghaas
  committed
  e31bc3c
- refactor(inference): keep FP8 backport focused
  mikasenghaas
  committed
  e09e92d
- test(inference): cover stochastic FP8 rounding
  mikasenghaas
  committed
  8164b5b
- style(inference): format FP8 patch
  mikasenghaas
  committed
  a210be6
- fix(trainer): match vLLM's per-token FP8 activation quantization bit for bit
  
  authored and
  mikasenghaas
  committed
  147b503
Commits on Sep 23, 2026
- fix(inference): align DSV4 request tool placement
  mikasenghaas
  committed
  a659f82
- fix(sft): export env vars in multi-node launcher
  mikasenghaas
  committed
  a3ffa1a