Repository navigation

SFT learner fixes: training mode for gradient checkpointing, Ascend OOM detection, real-loss autobatch probe - #331

Merged

michaellchung merged 5 commits into

mainfrom

sft-night-fixes

Nov 3, 2026

Merged

SFT learner fixes: training mode for gradient checkpointing, Ascend OOM detection, real-loss autobatch probe#331
michaellchung merged 5 commits into
mainfrom
sft-night-fixes

Commits on Nov 3, 2026
- accel: recognize an allocator OOM that carries no vendor exception class
  
  authored and
  MichaelChung
  committed
  d39b586
- Autobatch: probe the native loss instead of a stand-in for it
  
  authored and
  MichaelChung
  committed
  1f3afac
- Learner: name the chunked loss kernel in the causal kernel recipe log
  
  authored and
  MichaelChung
  committed
  a7ef6b8
- Learner: enter training mode before enabling gradient checkpointing
  
  authored and
  MichaelChung
  committed
  b8dbd24
- Learner: keep first-step warm-up out of the s/step telemetry
  
  authored and
  MichaelChung
  committed
  0d8b68c