Repository navigation

S26-E9: robustness fixes (preemption marker, verl cuts, data slots, one-command verl pairs) - #282

Closed

michaellchung wants to merge 23 commits into

s25-finalfrom

s26-e9-robust

Closed

S26-E9: robustness fixes (preemption marker, verl cuts, data slots, one-command verl pairs)#282
michaellchung wants to merge 23 commits into
s25-finalfrom
s26-e9-robust

Commits on Oct 20, 2026
- Thread preflight on by default for eval launches; log and record every loosened value
  
  chowhsu
  and
  claude
  committed
  fbefc71
- Samplers draw disjoint data slots; trainer ends FINISHED on an exhausted buffer; launcher follows rescheduled Modal containers
  
  chowhsu
  and
  claude
  committed
  59ab2b4
- Forward the Modal reclaim signal to a driver inside a Ray actor through the marker file
  
  chowhsu
  and
  claude
  committed
  9592db2
- verl cuts: save only optimizer-held parameters; round cuts on the fully_async path
  
  chowhsu
  and
  claude
  committed
  e20373e
- Set the reclaim marker path from --rl-checkpoint-store (the launcher's resume store), not the island flag
  
  chowhsu
  and
  claude
  committed
  e7cae5d
- One verl launch starts its sampler islands: the verl island entry runs the engine sampler for role sampler
  
  chowhsu
  and
  claude
  committed
  fa7413f
- Test: a resumable trainer island on Modal also gets the reclaim marker path (E9.1)
  
  chowhsu
  and
  claude
  committed
  ef34293
- verl sampler entry names the engine factory by path (no adapter-to-adapter import)
  
  chowhsu
  and
  claude
  committed
  4079f75
- Relay a rank>0 Modal reclaim to the rank-0 driver; one launch for synced trainers and their samplers
  
  chowhsu
  and
  claude
  committed
  cc8eba8
- Reclaim marker keeps its single-node format; only a relayed marker names the rank
  
  chowhsu
  and
  claude
  committed
  3469d0c
- A merging trainer creates the sample-buffer index (verl pair with several samplers)
  
  chowhsu
  and
  claude
  committed
  d1c91c1
- verl buffer trainer releases the per-shard read positions
  
  chowhsu
  and
  claude
  committed
  18f2c94
- Test: pass the shard trainer buffer by identity (an empty buffer is falsy)
  
  chowhsu
  and
  claude
  committed
  2e125ed
- verl buffer trainer: the buffer's end ends the run FINISHED
  
  chowhsu
  and
  claude
  committed
  e3e94c1
- Engine sampler island: a finished run ends its tape with rl_learner_finalized
  
  chowhsu
  and
  claude
  committed
  621cac6
- Engine sampler tape records carry the launcher island id
  
  chowhsu
  and
  claude
  committed
  1c23a60
- fully_async round cuts: the rollout pool exports no in-flight work
  
  chowhsu
  and
  claude
  committed
  7ee7aab
Commits on Oct 21, 2026
- verl sync island trains on every card of its container; multi-card Modal verl islands default NCCL_SHM_DISABLE=1
  
  chowhsu
  and
  claude
  committed
  af35378
- Miles sampler: forward the data-slot env to the Ray rollout workers; samples record their data row
  
  chowhsu
  and
  claude
  committed
  f6bdc9f
- verl vLLM readback: one tmp file per replica process
  
  chowhsu
  and
  claude
  committed
  0b64895
- Head syncer waits only for a synced group's trainers, not its sampler islands
  
  chowhsu
  and
  claude
  committed
  0752eb4
- Syncer quorum counts a synced group's trainers only
  
  chowhsu
  and
  claude
  committed
  c711302
- Dry-run plan sizes the role syncers like the real launch (synced group)
  
  chowhsu
  and
  claude
  committed
  b1739c1

Commits

Commits on Oct 20, 2026

Commits on Oct 21, 2026