t2ance/mat-03-gpu-task-headroom-and-timeout

03 GPU task headroom and timeout

Does GRPO on the GPU image task dogs-vs-cats-redux-kernels-edition improve the Qwen3.5-27B policy's solutions in a way the run artifacts can attribute to the trained weights rather than to the tree search alone, and what governs the answer: the headroom above what the untrained policy already writes in round 1, and how the execution timeout is stated to the policy and priced in the reward? The report's answer: the untrained 27B already writes an ImageNet-pretrained torchvision ResNet, the recipe behind all 31 gold nodes in the window, present among round 1's valid nodes in six of the eight training runs other than ray21 and reaching gold (0.0315 against the 0.03882 threshold) in 175s of script time, so the task has no headroom left for training to be credited with once that recipe is found; the run read at the time as a learning curve (ray21, best log loss 0.258 to 0.014 over five rounds) is not attributable to the weights without a frozen-actor control, because the same recipe appears in round 1 of the other runs; printing the 600s timeout in the prompt did not lower the timeout fraction (0.06 to 0.31 over eight steps), while restating it as a 1800s hard ceiling removed timeouts at step 1 (0 of 64 nodes) and raised the step-1 valid rate to 0.36; and GPU declaration (6 to 81 percent) never returned to the pre-fix near-zero level. This repository is the data root of that question: everything its runs wrote, minus the exclusion list in section 5.

1. Question and links

• Question and configs: personal/experiments/rl-training/03-gpu-task-headroom-and-timeout/ in t2ance/ml-agent-training-personal. This question has no card.md: each of its five leaves holds only one launcher script under <leaf>/configs/run_<leaf>.sh; the leaf-to-run mapping in section 2 comes from those scripts and from the report's own run-script citations, not from a card.
• Report: personal/experiments/rl-training/03-gpu-task-headroom-and-timeout/report/main.tex in t2ance/ml-agent-training-personal.
• Code: task registration commit 37305c7; prompt de-bias commits 515ee50 and 27a2f3b; the timeout wording, per-node telemetry and pod CPU change in 44044a6; packing in 3312da5, bf0b9da, 1cce549 — all in ruz048/ml-agent-training, training_verl/tree_grpo_trainer.py and training_verl/tree_rollout.py. No pull request is named in the report.
• Issue: none is named in the report or in the leaf configs.
• W&B project: pqin/tree-grpo. Runs, grouped by leaf (section 2 maps each to its rollout directory): tree_grpo_27b_dogscats_gpu_diag: gpu_diag_run2, gpu_diag_run3; tree_grpo_27b_dogscats: 1e4_0703, 1e4_0703_g8, 1e4_20260707_16node_vllm06, 1e4_20260707_16node_k8s2_ray21_vllm03 (ray21); tree_grpo_27b_dogscats_16node_t600surf: 1e4_20260708_16node_t600surf_save5 (t600surf); tree_grpo_27b_dogscats_scale64_t1800: scale64_t1800_ceil_0708, scale64_t1800_cpu4_0709; tree_grpo_27b_dogscats_pack_k4, seven launches: 20260709_073241, 20260709_165143, 20260709_212635, 20260710_050152, 20260710_065744, 20260710_152640, 20260710_201830. Sixteen runs in all; every one shows W&B state crashed (W&B's label for any exit without a clean finish, per the report's Experiments section), including the runs the report reads as finished steps.
• Status: finished (closed 2026-07-11). The window runs from the first probe (gpu_diag_run3, round 1 logged 2026-07-03 05:47 UTC, computed from its rollout file name) through the pack_k4 lineage's kept checkpoint global_step_10, saved 2026-07-11 02:20 UTC per the report. The report's header states "Opened: 2026-07-03, Status: Closed, Related: 01, 02, 04"; the task was set aside on 2026-07-11 when the campaign pivoted to the spooky task (report 04).

2. Directory tree

Nine run directories, one per rollout run id ($ROLLOUT_RUN_ID, the timestamp each launch's harness process generated at start). Every run directory holds round_N/ subdirectories and one timing.jsonl; every round_N/ holds up to four files sharing one Unix-epoch id: <task>__<epoch>.jsonl (the round's node records, section 4), <task>__<epoch>__tree.txt (a plain-text search tree), <task>__<epoch>__transcript.md (the same tree with every node's full prompt/response/thinking as Markdown), and <task>__<epoch>__summary_<epoch2>.json (the round's own manifest, section 4) — the summary file is written only once that round's rollout has fully finished, so a round interrupted mid-rollout has the first three files but no summary. Every file is listed with its size in FILES.txt.

README.md                    this front page
FILES.txt                    every published file with its byte size
20260703_052111/              gpu_diag probe "run3" (round_1 2026-07-03 05:47 UTC): round_1 (16 nodes, 0/16 valid, tier none), round_2
20260703_205538/              dogscats fit "1e4_0703" (round_1 2026-07-03 21:18 UTC, lr 1e-4, 16 nodes): round_1..3 complete, round_4 mid-rollout (no summary)
20260703_232538/              dogscats fit "1e4_0703_g8" (round_1 2026-07-03 23:46 UTC, 32 nodes): round_1 complete, round_2 mid-rollout (no summary)
20260707_223956/              "ray21" (round_1 2026-07-07 22:57 UTC, 16 nodes, the run read at the time as the learning curve): round_1..5 complete, round_6 stopped mid-rollout at 11/16 node records (no summary)
20260708_060256/              "t600surf" (round_1 2026-07-08 06:27 UTC, 16 nodes, the 600s timeout printed in the prompt): round_1..8 complete, round_9 mid-rollout (no summary)
20260708_160814/              "scale64_ceil" (round_1 2026-07-08 16:21 UTC, 64 nodes, 1800s hard-ceiling wording): round_1..4, all four with a summary; round_4's rollout finished (26/64 InfraError) but its training update was killed with no checkpoint
20260709_000235/              "cpu4" (round_1 2026-07-09 00:23 UTC, same 64-node/1800s config, GPU pod CPU request 8 to 4): round_1..3, all three with a summary; round_3's update died of a foreign-process CUDA OOM, no checkpoint
20260709_073241/              pack_k4 lineage, launch 1 of 7 (2026-07-09 07:41 UTC): round_1 only, mid-rollout (no summary)
20260710_201830/              pack_k4 lineage, launch 7 of 7, the last (round_9 2026-07-10 20:57 UTC, round_10 summary 2026-07-11 01:57 UTC): round_9 and round_10, both complete; round_10 is the lineage's final saved state, checkpoint global_step_10

Leaf and report-row map (leaves are under personal/experiments/rl-training/03-gpu-task-headroom-and-timeout/, table rows are in report/main.tex): tree_grpo_27b_dogscats_gpu_diag owns 20260703_052111 (report table row "diag_run3, 1" in "Wiring probes and the first fits"; the sibling probe run2 logged one W&B step but its own rollout directory is not identified — its console log was overwritten by run3's, per the report's Evidence paragraph in the same section). tree_grpo_27b_dogscats owns 20260703_205538 (rows "1e4_0703, 1/2/3"), 20260703_232538 (row "1e4_0703_g8, 1"), and 20260707_223956 (ray21, table in "ray21: did the best score climb?"); the report's Baseline subsection groups these three as sharing one prompt and timeout, and its Implementation subsection cites one launcher, run_tree_grpo_27b_dogscats.sh, for the leaf. tree_grpo_27b_dogscats_16node_t600surf owns 20260708_060256 (table in "t600surf: the budget printed in the prompt"); this run is also cited as an input by rl-training/04-noise-floor-and-snr-fixes — see t2ance/mat-04-noise-floor-and-snr-fixes for how that question uses it. tree_grpo_27b_dogscats_scale64_t1800 owns 20260708_160814 (ceil) and 20260709_000235 (cpu4), both in the report's "scale64_ceil and cpu4" section; the launcher run_tree_grpo_27b_dogscats_scale64_t1800.sh on disk was edited in place between the two launches and today reflects only cpu4. tree_grpo_27b_dogscats_pack_k4 owns 20260709_073241 and 20260710_201830; the report counts seven W&B launches in this lineage (section 1's pack_k4 group lists all seven), but only these two kept a rollout directory under this root, and 20260709_073241's own round_1 never finished (no summary, 0 W&B history rows for that launch). The report's per-step best-log-loss table for this lineage (0.037, 0.044, 0.131, 0.029, 0.163, 0.058, 0.044, 0.021, 0.038, 0.048 for steps 1 to 10) is reconstructed from all seven launches' W&B history, so only steps 9 and 10 of that table can be checked against a local round record in this root (20260710_201830/round_9, round_10); the other eight steps' round records were written by the five middle launches, whose rollout directories are not present here — their W&B history (section 1) is the only surviving record of those rounds.

Top-level files: README.md (this file) and FILES.txt (the complete file-and-size index the publish program regenerates before every upload; the tree above is a guide to it, not a substitute).

3. How to read each kind of file

• Text, JSON, JSONL, and log files: https://huggingface.co/datasets/t2ance/mat-03-gpu-task-headroom-and-timeout/raw/main/<path>, for example raw/main/20260708_060256/timing.jsonl or raw/main/20260708_060256/round_1/dogs-vs-cats-redux-kernels-edition__summary_1783494100.json. A round's node-record .jsonl carries the full prompt, the raw completion, the thinking text, and (when the node executed) the token ids and logprobs, so it runs a few MB per round (raw/main/20260708_060256/round_1/dogs-vs-cats-redux-kernels-edition__1783492030.jsonl is 3.3 MB) — fetch those, and the matching __transcript.md, at resolve/main/<path> instead (raw/ on a Git-LFS-tracked file returns only the pointer).
• No adapter, LoRA checkpoint, or full-model shard exists anywhere in this root: this question's runs wrote only the harness's tree-search records (round records, tree/transcript text, round summaries, per-step timing) and the console-logged phases_s timing; the verl checkpoints those runs produced (fp32 shards, optimizer state) live outside this tree under /data2/peijia/ml-agent-checkpoints/tree-grpo/<run_name>/ on the workspace host and were never copied here (section 5).
• The per-round curve: a round's own <task>__summary_<epoch>.json (section 4) gives that round's best_test_metric and valid count; the medal tiers the report cites are computed from the leaderboard thresholds in ml-agent-training/harness/tasks.py for dogs-vs-cats-redux-kernels-edition (N=1315 entrants): gold ≤0.03882, silver ≤0.05038, bronze ≤0.06127, median ≤0.12216, else tier "none". The finer per-step metrics the report tables quote (critic/score/mean, rollout/valid_rate, rollout/frac_timeout, rollout/frac_gpu_declared, rollout_corr/ppl_ratio, actor/kl_loss, medal/best_tier, medal/frac_gold) are verl's W&B history for each run named in section 1, not a file in this root; they follow the naming families visible in a node record's sibling fields (rollout/*, rollout_corr/*, actor/*, medal/*, critic/*, timing_s/*, timing_per_token_ms/*, global_seqlen/*, response_length/*, prompt_length/*, training/*, perf/*).
• Two invariants the report checks against these files: a killed node's exec_timings.exec (script time alone, inside the node record) should sit just past the run's TREE_EXEC_TIMEOUT — the report confirms 603 to 629s under the 600s limit and reports the 1800s ceiling's kills separately; and rollout_corr/ppl_ratio (not in this root; W&B only) should stay near 1, with values past about 1.3 read as the learning rate diverging (the pack_k4 lineage's own history reaches 12.51 by step 10, cited in the report's "pack_k4 lineage" section).
• round_N/<task>__<epoch>__tree.txt: one line per node in expansion order, indented under its parent, showing mode (draft/improve/debug), BUGGY/ok/test=, the sibling-group id, and (for an executed node) its k8s pod id, declared resources, GPU utilization, script time, and exception type if any.
• round_N/<task>__<epoch>__transcript.md: the same tree rendered as Markdown, followed by one subsection per node with its full system/user prompt and its response, thinking, and code, collapsed under <details>.
• timing.jsonl (run root, one line per completed training step, section 4): the trainer's own phase timer; phases_s.gen is generation plus the round's remote execution wall clock, phases_s.save_checkpoint appears only on a step that actually checkpointed.

4. Row fields

round_N/<task>__<epoch>.jsonl, one row per tree-search node: id (short hex node id, also the tree/transcript key); round; step (an internal per-node index the harness stamps at generation time; it is not always round - 1, so read it as an implementation counter rather than the training step); exp_id (the MLE-bench task slug); parent_id (null for a draft/root node); sibling_group (the GRPO group id the node's reward is rank-normalised within); mode (draft, improve, or debug); plan, thinking, raw_completion, code (the model's stated plan, chain-of-thought, full sampled text, and the extracted script; code is empty when no runnable script was extracted); prompt_messages (the full chat prompt sent to the policy); response_token_ids, response_logprobs (present only when the engine returned them); status (ok, crashed, timeout, or a no-code/no-submission state); is_buggy (true unless status is ok); exc_type (the Python exception class name, or null); test_metric (the task's own log loss, null when invalid); oriented (the metric oriented so that higher is better — for this log-loss task, the negative of test_metric, matching rollout/valid_auc; null when invalid); reward (the GRPO reward: exp(-test_metric) when valid, -0.5 for a crash, timeout, no-code, no-submission, or invalid-submission failure, -0.2 for an idle-GPU kill, per the report's Reward paragraph); resources (the node's declared {cpu, mem, gpu} resource block, null if none was parsed); gpu_util (percent GPU utilization measured during execution, null if the node never ran on a GPU pod); exec_time (total wall time the executor measured for the node); exec_timings (pending, stage, exec, retrieve — exec is the script's own time, the field the report's timeout analysis uses); decl_deps_ignored, decl_malformed, decl_miss_reason (flags and a reason for a malformed or ignored resource declaration); term_tail (the tail of the script's captured stdout/stderr); where (the exact execution location, e.g. a k8s pod id built from the rollout run id, node id, and attempt number); source (none for a draft node; the origin of any reference text an improve/debug node's prompt drew from, otherwise); n_children (how many child nodes this node was later expanded into); grade_error (the MLE-bench grader's error message for an invalid submission, otherwise null).

round_N/<task>__summary_<epoch2>.json, one object per round, the round's own manifest: exp_id; round; model (the served policy path, e.g. /data2/peijia/models/qwen3_5_27b); exec_backend (k8s for every run in this window); adapter_path (empty for round 1's base-policy rollout, the previous round's trained LoRA path afterward); nodes (nodes attempted this round); valid (nodes that produced a valid, gradeable submission); expansions (GRPO groups this round); phase (a round-purpose tag, train for every round in this window); best_test_metric, best_node_id (the round's best log loss and the node that produced it); honest_test_metric (a held-out/honest-evaluation metric field; null in every round of this window, since this campaign ran no held-out evaluation); rollout_file, workspace (the harness's own recorded, run-relative paths to the round's rollout file and exec-sandbox workspace at write time — map them onto this root via the run's rollout-run-id directory name, e.g. rollout_file: "data/rollouts/20260708_060256/round_1/..." is this root's 20260708_060256/round_1/...); gen_llm_seconds, gen_exec_seconds (the round's total time spent generating versus executing, summed across nodes).

timing.jsonl, one row per completed training step: step; total_s (wall time for the step); phases_s — update_weights, gen (rollout: generation plus remote execution), old_log_prob, ref, adv, update_actor, and save_checkpoint (present only on a step that checkpointed).

The W&B config/history/summary fields named in section 3 are not files in this root — they live in each run's W&B export under personal/experiments/rl-training/03-gpu-task-headroom-and-timeout/results/wandb_exports/<wandb_run_id>/ in the personal repository (config.json, history.jsonl, summary.json), which is outside this data root and was used only to find the run URLs and confirm the mapping above.

5. Exclusions and inputs

None of the rule's five excluded kinds is physically present in this root, because this question's runs never wrote verl checkpoints into it in the first place: default_local_dir/CKPT_DIR in every launcher points at /data2/peijia/ml-agent-checkpoints/tree-grpo/<run_name>/ on the workspace host, outside this tree, per this project's own workspace layout. Stated against the rule's list anyway: optimizer state, RNG state, and other per-step extra state — not present, would live only under the external checkpoint directory; full-model shards of a LoRA step — not present, same reason; the critic or value model — not applicable, this campaign uses GRPO (grpo_valid_only), which has no critic network (the critic/* W&B keys are reward statistics, not a critic model); resume pointers such as a latest-step file — not present, same external-checkpoint reason (every run in this window used resume_mode=disable except the pack_k4 lineage, which used resume_mode=auto against its own external checkpoint directory, also outside this root); an intermediate full-model dump next to a saved adapter — not applicable, no adapter was ever saved into this root.

Inputs (not copied into this root):

• The base model: Qwen3_5ForConditionalGeneration at /data2/peijia/models/qwen3_5_27b on the workspace host (MODEL_PATH in every leaf's launcher script, and the model field of every round summary), the public Qwen3.5-27B weights.
• The task: dogs-vs-cats-redux-kernels-edition (MLE-bench-lite: predict P(dog) for 12,500 photos, scored by binary log loss), as the harness's packed task cache and the MLE-bench private answers used for grading; both are benchmark data and stay out of this repository, matching this project's own workspace rule for MLE-bench data.
• No teacher model: this is a plain GRPO training question, not distillation.
• No other-question run is an input: this question's registry entry lists no inputs (inputs: {}). The 0.048 log-loss retrain-noise figure this report cites against this task's within-gold movements is a number carried from rl-training/04-noise-floor-and-snr-fixes's own retrain-noise measurement (recorded in personal/docs/lessons/qwen3_5_27b.md; report 04 itself flags the figure as "unverified" there because its per-rerun output is not retained, under personal/experiments/rl-training/04-noise-floor-and-snr-fixes/results/noise_measurement/noise_results/) — it is cited here as a number, not copied in as a run.