Repository navigation

S24: sampler island, trainer start/stop, decoupled PPO, identity tiering - #261

Closed

michaellchung wants to merge 64 commits into

mainfrom

s24-impl

Closed

S24: sampler island, trainer start/stop, decoupled PPO, identity tiering#261
michaellchung wants to merge 64 commits into
mainfrom
s24-impl

Commits on Oct 14, 2026
- Add the SCHEDULED consistent-cut trigger for scheduler-ordered stops
  
  chowhsu
  and
  claude
  committed
  198bf71
- rl-identity-tiering L9.1/L9.2: semantic hash dual write, tolerant PluginRef
  
  chowhsu
  and
  claude
  committed
  17b98c3
- Add trainer lifecycle, buffer and scheduler ports over IslandDriver
  
  chowhsu
  and
  claude
  committed
  eeaa187
- Record trainer start phases; setup reads the image manifest; fix model prefetch
  
  chowhsu
  and
  claude
  committed
  4a39de3
- Record the sample-buffer watermark in the cut; wait after publications
  
  chowhsu
  and
  claude
  committed
  a51fb3a
- Add the durable file sample buffer with arrival polling
  
  chowhsu
  and
  claude
  committed
  81b49f8
- Add merge metrics, island sync timing and syncer tape clock (L5.5-L5.7)
  
  chowhsu
  and
  claude
  committed
  766c927
- Drop Megatron-Bridge's eager modelopt import in the ports image layer
  
  chowhsu
  and
  claude
  committed
  2838755
- Add neutral CorrectionSpec fields and keep the correction group out of the phase-1 hash
  
  chowhsu
  and
  claude
  committed
  b8eb808
- Map correction settings to Miles, refuse decoupled-PPO breakers, record and compare per island
  
  chowhsu
  and
  claude
  committed
  25d07c1
- Raise policy-age cap to 4, emit offpolicy_gate, add off-policy correction tests
  
  chowhsu
  and
  claude
  committed
  5976f1b
- Make the submitted-phase clock a seam; update tests for the new prefetch and envs
  
  chowhsu
  and
  claude
  committed
  dc96be6
- Load fake lifecycle cuts with weights_only=True
  
  chowhsu
  and
  claude
  committed
  619016a
- Add scheduler S1: backlog start, ski-rental stop, journaled idle time
  
  chowhsu
  and
  claude
  committed
  2352deb
- Keep plugin sources and default tapes unchanged: build offpolicy_gate in the driver
  
  chowhsu
  and
  claude
  committed
  3ffea04
- Add gated opt-in SGLang snapshot mode for sampler engines (L8.2)
  
  chowhsu
  and
  claude
  committed
  b5e5328
- rl-identity-tiering L9.4: yeto/codepack.py run code pack (git-patch mode)
  
  chowhsu
  and
  claude
  committed
  98d88d1
- Merge branch 's24-miles-import' into s24-image
  
  chowhsu
  and
  claude
  committed
  0ffd52a
- Record trainer_start T_s for non-echo sky islands
  
  chowhsu
  and
  claude
  committed
  8c9ff33
- rl-identity-tiering L9.4-L9.6: launcher uses the run code pack
  
  chowhsu
  and
  claude
  committed
  50cb213
- Run the trainer scheduler in the head loop behind --rl-trainer-scheduler
  
  chowhsu
  and
  claude
  committed
  6b444d9
- Regenerate decoupling golden samples for the phase-1 correction hash (D8)
  
  chowhsu
  and
  claude
  committed
  69a2722
- Wire the Miles rollout through the durable sample buffer (L4.5)
  
  chowhsu
  and
  claude
  committed
  c34b0da
- Map off-policy correction to verl fully_async and add CPU parity checks
  
  chowhsu
  and
  claude
  committed
  da796bb
- Keep the sampler snapshot gate on the island side only
  
  chowhsu
  and
  claude
  committed
  05c359c
- Pin Miles 17aeb5cb5 and ports image 17aeb5c-2fa8801
  
  chowhsu
  and
  claude
  committed
  47e3052
- Record the L8.2 NCCL-path snapshot compatibility pass
  
  chowhsu
  and
  claude
  committed
  0a5241f
- Add convertible nodes: node classes, conversion port, cut rule, scheduler
  
  chowhsu
  and
  claude
  committed
  c8a437e
- Merge branch 's24-l4-wiring' into s24-l5-remote
  chowhsu
  committed
  95ba2e8
- Label offpolicy_gate events with the batch's engine backend
  
  chowhsu
  and
  claude
  committed
  8e5884a
- Record the correction group and prox_anchor on the verl island tape
  
  chowhsu
  and
  claude
  committed
  d3a423a
- Start and stop a remote trainer island from the head scheduler (L5.3c)
  
  chowhsu
  and
  claude
  committed
  5ff3b06
- Wire the Miles conversion port callbacks: load_version, engine in-flight requests, admission
  
  chowhsu
  and
  claude
  committed
  d944da3
- Stamp the version and warm the LoRA adapter in MilesPublisher.load_version
  
  chowhsu
  and
  claude
  committed
  2a21396
- Add sampler-only and trainer islands over the sample buffer (L4.6)
  
  chowhsu
  and
  claude
  committed
  5fd3f3b
- Keep the trainer island binding free of an import cycle
  
  chowhsu
  and
  claude
  committed
  adc2417
- Read the convertible node's weight checksum after the KV onload
  
  chowhsu
  and
  claude
  committed
  faef230
- Merge branch 's24-l5-measure' into s24-integ
  chowhsu
  committed
  34102b2
- Merge branch 's24-l9-identity' into s24-integ
  chowhsu
  committed
  108e66c
- Merge branch 's24-l8-snapshot' into s24-integ
  chowhsu
  committed
  83bff47
- Merge branch 's24-l5-remote' into s24-integ
  chowhsu
  committed
  3643236
- Merge branch 's24-sampler-mode' into s24-integ
  chowhsu
  committed
  9bdcef9
- Merge branch 's24-l7-convert' into s24-integ
  chowhsu
  committed
  e89bcf7
- Compose the L5.4 sampler/trainer pair in one launch
  
  chowhsu
  and
  claude
  committed
  e12c6f7
- Create the Modal volume when the code pack is first stored
  
  chowhsu
  and
  claude
  committed
  06d617b
- Pin Miles 44d41efd0 and ports image 44d41ef-2fa8801 (lazy modelopt)
  
  chowhsu
  and
  claude
  committed
  f013058
- Add check-run levers for the L5.4 start/stop GPU check
  
  chowhsu
  and
  claude
  committed
  4649cc4
- Create the Modal Volume when the code pack is stored
  
  chowhsu
  and
  claude
  committed
  0d473e7
- Let a no-sync buffer island train every parameter in run_miles
  
  chowhsu
  and
  claude
  committed
  bd81194
- Keep the off-policy gate inputs on the Miles rollout handle
  
  chowhsu
  and
  claude
  committed
  8dec16e
- Accept --debug-rollout-only on a fixed-partition sampler island
  
  chowhsu
  and
  claude
  committed
  7174b9f
- Give the sampler island its rollout metadata sink
  
  chowhsu
  and
  claude
  committed
  3e6eca4
- Set the sampler island's train parallel config for Miles post-processing
  
  chowhsu
  and
  claude
  committed
  183b724
- Merge s24-image into s24-integ
  
  chowhsu
  and
  claude
  committed
  555286f
- Merge branch 's24-gpu-l610' into s24-integ
  chowhsu
  committed
  e50885d
- Put RayMetadataSink.discard on the Ray sink, not the dir source
  
  chowhsu
  and
  claude
  committed
  79280be
- Count Miles' own stale buffer drops in offpolicy_gate
  
  chowhsu
  and
  claude
  committed
  439705e
- Pin the clock in the compaction byte-identity test
  
  chowhsu
  and
  claude
  committed
  dce70b3
- Publish a no-sync full-parameter island through a full-parameter digest
  
  chowhsu
  and
  claude
  committed
  db1315b
- Stop the trainer scheduler after repeated failed incarnations
  
  chowhsu
  and
  claude
  committed
  387d05b
- Run Miles' end-of-round hooks in the sample buffer reader
  
  chowhsu
  and
  claude
  committed
  212a9be
- Merge s24-integ (dce70b3c) into s24-gpu-l54
  
  chowhsu
  and
  claude
  committed
  4345e7b
- Time a held trainer island's start from its spawn
  
  chowhsu
  and
  claude
  committed
  6160480
- Cut a full-parameter trainer: carry every trainable parameter
  
  chowhsu
  and
  claude
  committed
  2748b9b

Commits

Commits on Oct 14, 2026