Repository navigation

S26 E0.3: bias-stacking ablation tooling for Qwen3-0.6B on Miles - #286

Closed

michaellchung wants to merge 85 commits into

s25-finalfrom

s26-e03-ablation

Closed

S26 E0.3: bias-stacking ablation tooling for Qwen3-0.6B on Miles#286
michaellchung wants to merge 85 commits into
s25-finalfrom
s26-e03-ablation

Commits on Oct 20, 2026
- tools/eval: add efficiency evaluation (bare vs yeto, verl and Miles)
  
  chowhsu
  and
  claude
  committed
  5a01732
- syncer: pipeline policy-sweep fragments and async per-round checkpoints
  
  chowhsu
  and
  claude
  committed
  8358c1b
- tools/eval: bare Miles trains with unfused attention (match yeto generic recipe; TE flash fails on sm86)
  
  chowhsu
  and
  claude
  committed
  261ba2e
- syncer: bf16 sweep pushes and sweep-aligned checkpoint intervals
  
  chowhsu
  and
  claude
  committed
  eb9e229
- tools/eval: yeto+verl async arm passes the staleness-2 TIS algorithm spec
  
  chowhsu
  and
  claude
  committed
  c3c9d09
- syncer: int8 sweep pushes, bf16 downlink deltas, sync/train overlap
  
  chowhsu
  and
  claude
  committed
  1a783d0
- E6: quantized rollout tiers with an FFN/routed-expert scope rule
  
  chowhsu
  and
  claude
  committed
  ce3822d
- E6: FP8 card allowlist from the GPU spec; keep tier-0 argv golden
  
  chowhsu
  and
  claude
  committed
  20f603e
- tools/eval: bare verl arms run the unpatched fork; async uses merged LoRA and /dev/shm NCCL host buffers
  
  chowhsu
  and
  claude
  committed
  63dc064
- overlap: keep rollout policy age 0, relabel rebased cuts, push whole sweep
  
  chowhsu
  and
  claude
  committed
  8af78fa
- E6: engine sampler tier settings, dense expert count, CPU proofs
  
  chowhsu
  and
  claude
  committed
  6b74178
- E6: load Miles process_utils from its file in the env test
  
  chowhsu
  and
  claude
  committed
  a215af7
- tools/eval: parse bare verl fully_async rows by training step; rates over rounds that report the value
  
  chowhsu
  and
  claude
  committed
  052c31e
- tools/eval: bare Miles arms do not shuffle prompts (yeto's Miles argv does not)
  
  chowhsu
  and
  claude
  committed
  c9245e5
- verl adapter: fully_async keeps NCCL host buffers on /dev/shm in every Ray worker
  
  chowhsu
  and
  claude
  committed
  73442ad
- tools/eval: read the yeto+Miles launcher exit code into the parsed run
  
  chowhsu
  and
  claude
  committed
  0122f4e
- test: overlap island with bf16 downlink deltas against the real syncer
  
  chowhsu
  and
  claude
  committed
  152374b
- miles adapter: opt-in gradient-accumulation fusion; generic recipe trains without dropout
  
  chowhsu
  and
  claude
  committed
  e0f833d
- tools/eval: yeto+Miles arm takes the bare Miles recipe settings
  
  chowhsu
  and
  claude
  committed
  6a4263e
- tools/eval: read only the run's own tapes (skip the repo copy and the store)
  
  chowhsu
  and
  claude
  committed
  ff893bc
- tools/eval: verl fully_async generation time from the rollouter, else missing
  
  chowhsu
  and
  claude
  committed
  ba17953
Commits on Oct 21, 2026
- eff eval: final round cut only for yeto+Miles; verl async generation from the rollouter
  
  chowhsu
  and
  claude
  committed
  a4b1104
- verl: hash the LoRA on the trainer for the single-island publish fast path
  
  chowhsu
  and
  claude
  committed
  77ce9cc
- round cut: write before the trainer offloads; commit the store in the background
  
  chowhsu
  and
  claude
  committed
  74016eb
- fully_async cut: record untrained prompts without stopping the worker
  
  chowhsu
  and
  claude
  committed
  3b00fe5
- golden: regenerate decoupling samples for the explicit Miles dropout 0.0
  
  chowhsu
  and
  claude
  committed
  84d43ce
- driver: time the rollout offload after generation as its own phase
  
  chowhsu
  and
  claude
  committed
  36eb805
- eff eval: --checkpoint final|every-round for the yeto+Miles arm
  
  chowhsu
  and
  claude
  committed
  dfe8b91
- syncer: send each fragment's FINAL version as an exact FULL broadcast
  
  chowhsu
  and
  claude
  committed
  bc8f53b
- round cut: write the trainer shards in the background for periodic cuts
  
  chowhsu
  and
  claude
  committed
  8ef1018
- round cut: skip a periodic cut while the previous commit runs; time the sample filter hook
  
  chowhsu
  and
  claude
  committed
  de275ef
- cut digests: hash tensor buffers without a bytes copy
  
  chowhsu
  and
  claude
  committed
  77b460c
- round cut: a periodic cut does not wait for the store commit it starts
  
  chowhsu
  and
  claude
  committed
  86d8872
- rollout metadata hook: time its steps (E1 diagnosis)
  
  chowhsu
  and
  claude
  committed
  0f615ab
- fully_async: estimate the cross-version IS truncation off the critical path
  
  chowhsu
  and
  claude
  committed
  427d142
- test: the fully_async round report fake takes build_metadata's new keyword
  
  chowhsu
  and
  claude
  committed
  4015cc5
- round cuts: commit the deferred cut fully off the critical path
  
  chowhsu
  and
  claude
  committed
  1502e7d
- driver: take the terminal load probe off the round's critical path
  
  chowhsu
  and
  claude
  committed
  ae82993
- round cuts: collect a skipped prepared cut's writer at the next cut
  
  chowhsu
  and
  claude
  committed
  31a951a
- round cuts: do not wake an offloaded trainer for a declined periodic cut
  
  chowhsu
  and
  claude
  committed
  0a11aa1
- eval: give the yeto+Miles arm the bare arms' container resources
  
  chowhsu
  and
  claude
  committed
  f19a055
- E6: fixes from the H100 gate runs (fused ignore aliases, token scoring, BF16 KV)
  
  chowhsu
  and
  claude
  committed
  4854dda
- E6: refuse tier 1 on verl, set verl max_model_len, native-FP8 gate probes
  
  chowhsu
  and
  claude
  committed
  ffe86d0
- round cuts: sync a deferred cut's state dir as it was at the cut
  
  chowhsu
  and
  claude
  committed
  19e595a
- E6: pin the FP8-fixed ports image; tier 1 trainer FP8 source (Miles native + ignore list)
  
  chowhsu
  and
  claude
  committed
  d9a2703
- E6: Miles RL drivers for the tier groups (Modal H100!)
  
  chowhsu
  and
  claude
  committed
  7f100f4
- E6: full-parameter pair needs a fixed-partition sampler island
  
  chowhsu
  and
  claude
  committed
  8d1df81
- E6: Dockerfile and tests follow the new image pins; neutral flag help
  
  chowhsu
  and
  claude
  committed
  e3da116
- E6: pair islands are 2xH100 each (fixed partition applies to every island)
  
  chowhsu
  and
  claude
  committed
  2011906
- E6: tier 2 also runs as one launch per role (--rl-sample-role); pair driver uses two launches
  
  chowhsu
  and
  claude
  committed
  aa3b97c
- E6: per-round parser of the Miles RL runs
  
  chowhsu
  and
  claude
  committed
  d7c79be
- wan: per-stream stall timeout, rate-aware chunk striping, sync diagnostics
  
  chowhsu
  and
  claude
  committed
  3dcbec8
- Modal store: no reload while a writer holds uncommitted files
  
  chowhsu
  and
  claude
  committed
  be553cf
- E6: pair trainer keeps the Adam state on the host (OOM on one H100)
  
  chowhsu
  and
  claude
  committed
  beaf4e0
- round cuts: copy a background cut's state on the device first
  
  chowhsu
  and
  claude
  committed
  8154509
- publish fast path: stage the digest's tensors with one GPU wait
  
  chowhsu
  and
  claude
  committed
  13d53ab
- rollout metadata hook: read each sink value once per hook run
  
  chowhsu
  and
  claude
  committed
  19f2648
- eval: tell an NVIDIA A10 from an A10G
  
  chowhsu
  and
  claude
  committed
  4a1c245
- Modal store: the write guard is a container-wide flock
  
  chowhsu
  and
  claude
  committed
  1fe2c76
- test: compare staged digests without their timing fields
  
  chowhsu
  and
  claude
  committed
  1fde82f
- round cuts: keep a sleeping trainer's cut on the host path
  
  chowhsu
  and
  claude
  committed
  d583c59
- E6: tier 1 on Miles defaults to the trainer FP8 source
  
  chowhsu
  and
  claude
  committed
  f4cd14e
- RL: full parameters on one no-sync island with dedicated engine cards
  
  chowhsu
  and
  claude
  committed
  bce94af
- E6: pin the ports image with the SGLang compressed-tensors FP8 reload
  
  chowhsu
  and
  claude
  committed
  8fe48c9
- round cuts: serialize and hash background shards in a writer process
  
  chowhsu
  and
  claude
  committed
  2f3d79c
- E6: neutral fp8-source help; full-parameter dedicated-engine group is 1 train + 1 engine card
  
  chowhsu
  and
  claude
  committed
  f74ac99
- launcher: pin a Modal island to one GPU model with --modal-gpu-name
  
  chowhsu
  and
  claude
  committed
  27f66dd
- publish fastpath: hash the policy while a colocated trainer offloads
  
  chowhsu
  and
  claude
  committed
  d8d159c
- E6: full-parameter dedicated-engine group runs without round cuts
  
  chowhsu
  and
  claude
  committed
  dfa4a58
- wan: replay INIT after a mid-init redial, link probe with container relaunch
  
  chowhsu
  and
  claude
  committed
  2509fde
- Merge s26-e01-effeval into E0.3 ablation branch
  
  chowhsu
  and
  claude
  committed
  156b5ec
- Merge s26-e3-sync into E0.3 ablation branch
  
  chowhsu
  and
  claude
  committed
  ca78ca0
- Merge s26-e6-quant into E0.3 ablation branch
  
  chowhsu
  and
  claude
  committed
  755375e
- E0.3: bias-stacking ablation drivers (Qwen3-0.6B, Miles, layers L0-L3)
  
  chowhsu
  and
  claude
  committed
  2e1d8d3
- E0.3: head driver plan line without STEPS
  
  chowhsu
  and
  claude
  committed
  a78e73a
- eval: opt-in trainer step profile (--rl-profile-trainer)
  
  chowhsu
  and
  claude
  committed
  1c0df83
- E0.3: per-round parser; multi-island layers replace slow containers with the link probe
  
  chowhsu
  and
  claude
  committed
  91adbef
- round cuts: stage a background cut's device copies on the host at the step
  
  chowhsu
  and
  claude
  committed
  fe5e612
- E0.3 v2: 256 samples per round, no checkpoint store so fully_async keeps its background groups
  
  chowhsu
  and
  claude
  committed
  3c9ee84
- Revert "round cuts: stage a background cut's device copies on the host at the step"
  
  chowhsu
  and
  claude
  committed
  40cf73a
- E0.3: parser groups v2 run ids by layer
  
  chowhsu
  and
  claude
  committed
  92cb83c
- round cuts: hand a background cut to its writer in one shared buffer
  
  chowhsu
  and
  claude
  committed
  74c2f6f
- Merge s26-e1-overhead into E0.3 ablation branch
  
  chowhsu
  and
  claude
  committed
  9079c7e
- E0.3: keep the GSM8K task inputs under data/ (inputs, not run logs)
  
  chowhsu
  and
  claude
  committed
  b799606
- E0.3: pass the task file as an explicit local path
  
  chowhsu
  and
  claude
  committed
  1a3014f

Commits

Commits on Oct 20, 2026

Commits on Oct 21, 2026