tt-x10-fsdp2-fa2-117-30B

tt-x10-fsdp2-fa2 — step 117 (X10b training backend: FSDP2 vs Megatron)

GRPO checkpoint from the TaskTrove X10b backend experiment, trained on Iris / CoreWeave (cw-rno2a, 4x8 H100). Base model Qwen/Qwen3-Coder-30B-A3B-Instruct, trained on DCAgent/exp_rpt_multifile with SkyRL + Terminus-2. Single variable vs the X8 longclip base: trainer.strategy=fsdp2 (+ flash-attention-2, async). Campaign verifier is pass_ratio shaping.

Checkpoint selection

global_step_117 was selected by the trailing-5 reward EMA (alpha = 1/3) over the stitched 6-generation curve (steps 1-131, no gaps) — the highest-EMA checkpoint (EMA 0.2535; step reward 0.2793; pass@8 0.3906; entropy 0.0215), inside the reward plateau. Export performed on-Iris from the sharded FSDP checkpoint bank via the checkpoint-export job (hf_upload_mode=latest).

Run status — question ANSWERED; stopped at 131/400 (plateau)

X10b's question — does FSDP2 survive the X8 schedule that collapsed every Megatron arm — is answered yes: 131 stable steps with entropy falling 0.284 -> 0.0129 (~22x, exploration exhausted) and response length held (11.9k -> 13.5k tokens; no length collapse). Reward plateaued from ~step 40 (20-step windows 40-131 flat within noise; peak raw 0.332 at step 116); the remaining horizon buys nothing. Stopped by owner decision 2026-08-15 at logged step 131, banked 129.

See training_logs/ for metrics.csv (stitched per-step curve + trailing-5 EMA), report.md, reward_plot.png, and the launch rl_config.yaml. W&B (dogml/OpenThoughts-Agent): base jjjhduya, final gen r7 0s0kinwg.

Training Traces

open-athena/tt-x10-fsdp2-fa2 — 1/8 systematic subsample (every 8th trial of 89,425 across 5 run generations, uniform coverage of steps 1-131). Full-set transfer from CoreWeave object storage was ~430 GB; the subsample is a documented deviation, approved by the owner, for an arm whose question was answered before horizon.