All pull requests

Pull requests

• rl/syncer: one syncer with async multi-island sync for LoRA and full-parameter RL

  #292·michaellchung opened Oct 22, 2026Contributor·

  ·

• Move fork and image references from michaellchung to agentenv

  #291·michaellchung opened Oct 22, 2026Contributor·

  ·

• S26 E6: FP8 quantized rollout and full-parameter dedicated-engine island

  #290·michaellchung opened Oct 22, 2026Contributor·

  ·

• S26 E4/E5: phase timing, analysis tools, release writer Volume optimization

  #289·michaellchung opened Oct 22, 2026Contributor·

  ·

• S26 E0.2/E11: cluster comparison tools and full-parameter no-sync single island

  #288·michaellchung opened Oct 22, 2026Contributor·

  ·

• S26 E1: single-island overhead (background round cuts, publish fast path, profilers)

  #287·michaellchung opened Oct 22, 2026Contributor·

  ·

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

  #286·michaellchung opened Oct 21, 2026Contributor·

  ·

• S26 E2: async buffer depth defaults to the policy-age limit; opt-in adaptive TIS

  #285·michaellchung opened Oct 21, 2026Contributor·

  ·

• S26 E3: faster multi-island weight sync (pipelined sweeps, WAN stream robustness)

  #284·michaellchung opened Oct 21, 2026Contributor·

  ·

• S26-E7: reject verl with expert parallelism; require fsdp2 strategy

  #283·michaellchung opened Oct 21, 2026Contributor·

  ·

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

  #282·michaellchung opened Oct 21, 2026Contributor·

  ·

• S26-E0.1: efficiency evaluation (bare vs yeto, verl and Miles)

  #281·michaellchung opened Oct 21, 2026Contributor·

  ·

• S26-E0.5: remove the self-written Codex harness

  #280·michaellchung opened Oct 21, 2026Contributor·

  ·

• S25 final integration: sampler islands on spot, synced trainers to samplers, round cut, stage-1 rehearsal, agentic RL, eval suite

  #279·michaellchung opened Oct 20, 2026Contributor·

  ·

• S25-J: round-3 integration — sharded requant, images, verl cuts, device families, stage-1 launcher

  #278·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-I: verl Harbor agent loop on boat sandboxes

  #277·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-H: Miles Harbor agent on boat sandboxes (N10.8b)

  #276·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-G: boat sandbox provider for Harbor (N10.8)

  #275·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-F: reuse Harbor for agentic RL on Miles and verl (N10.1-10.4)

  #274·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-E: recording proxy and chat-template token alignment (N10.5-10.6)

  #273·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-D: SamplerEngine port, SGLang/vllm-ascend adapters, post-load checks (N2)

  #272·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-C: resume manifest assertions for cut restore (N7.6)

  #271·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-B: MXFP4 routed-expert RTN requantization (N8.5)

  #270·michaellchung opened Oct 19, 2026Contributor·

  ·

• S25-A: SSH node pool, OSS store, W4A8 requant, full-param trainer island

  #269·michaellchung opened Oct 19, 2026Contributor·

  ·

• S24 wrap-up: convertible pool follows releases and paces rounds, stop-run ends the run, syncer push timing

  #268·michaellchung opened Oct 19, 2026Contributor·

  ·