penfever/tasktrove-repro-x10-x2

TaskTrove reproducibility package — X10 fsdp2-fa2 and X2 clip 0.2

Artifacts for two arms of the TaskTrove hyperparameter campaign (marin-community/marin#7785). The two arms run on different clusters with different launchers, so each has its own section.

Both arms train Qwen/Qwen3-Coder-30B-A3B-Instruct on DCAgent/exp_rpt_multifile through the SkyRL terminal-bench entrypoint with the Harbor agent harness.

Everything here is a record of what ran. Where a fact could not be recovered, this document says so instead of substituting a plausible value.

Contents

Path	What it is
x10-fsdp2-fa2/tasktrove_x10_fsdp2.yaml	the launch config, byte-identical to the one that ran
x10-fsdp2-fa2/generate_x10_configs.py	the generator that emitted it
x10-fsdp2-fa2/manifest-r4.json	Iris job manifest for the current generation
x10-fsdp2-fa2/task-entrypoint-r4.md	the container setup and run command, verbatim
x10-fsdp2-fa2/finelog-r4.log	the training log, 35,523 lines
x2-clip-hi0p2/tasktrove_x2_clip_hi0p2.yaml	the launch config
x2-clip-hi0p2/jupiter-rendered-sbatch-EXAMPLE.sbatch	a rendered sbatch, see the caveat below
x2-clip-hi0p2/jupiter-launcher-config-EXAMPLE.json	the launcher's resolved job JSON, same caveat
launcher/	the untracked scripts that generate configs and submit the Jupiter arms
curves/reward_raw.csv	per-step reward/avg_raw_reward for every arm in the campaign
curves/reward_curve_tt-x10-fsdp2-fa2.png	the X10 curve

---

X10 fsdp2-fa2 — CoreWeave Iris, cluster cw-rno2a

Tests one variable against x8_longclip base: trainer.strategy = fsdp2 instead of megatron. Async, disaggregated, 4 nodes x 8 H100 = 32 GPUs, policy 2x8 and 4 engines at TP4. Horizon 400 steps.

Config pin, verifiable

The launcher names the resolved config by the first 16 hex characters of the SHA-256 of the config file bytes (cloud/iris/iris_backend.py:1896). All four generations ran /tmp/marin-rl-configs/66ed71d78b69a082.yaml, and the config in this package hashes to the same value:

shasum -a 256 x10-fsdp2-fa2/tasktrove_x10_fsdp2.yaml | cut -c1-16
# 66ed71d78b69a082

That is a byte-identity proof, not a claim.

Generations and runtime commits

The arm was resumed three times. Every generation used the same config; only the MarinSkyRL runtime commit changed. The checkpoint bank is shared, so ckpt_path points at the original tt-x10-fsdp2-fa2 prefix in every generation.

Job	MarinSkyRL runtime commit	Outcome
/benjaminfeuer/tt-x10-fsdp2-fa2	a5eac52fb0e1	wedged at step 75, no application output for about 8.8 h
/benjaminfeuer/tt-x10-fsdp2-fa2-r2	88b8bd60be12	resumed at step 75, reached 86, trainer died at 22:44 UTC while the controller still read running
/benjaminfeuer/tt-x10-fsdp2-fa2-r3	e1c2dff6a963	failed at startup, dataset ... got size 0, because the relaunch omitted --train_data
/benjaminfeuer/tt-x10-fsdp2-fa2-r4	e1c2dff6a963	current, resumed at step 84

Full commit for the current generation: e1c2dff6a9633a9ac71fdd93a6e8890346cfcf0e, repository https://github.com/marin-community/MarinSkyRL.git.

Codebase

There is no cluster-local checkout. The task image clones MarinSkyRL at the pinned commit and asserts the revision before doing anything else:

checkout=/app/marinskyrl
git init -q "$checkout"
git -C "$checkout" remote add origin https://github.com/marin-community/MarinSkyRL.git
git -C "$checkout" fetch --depth=1 origin e1c2dff6a9633a9ac71fdd93a6e8890346cfcf0e
git -C "$checkout" checkout -q --detach FETCH_HEAD
test "$(git -C "$checkout" rev-parse HEAD)" = e1c2dff6a9633a9ac71fdd93a6e8890346cfcf0e
bash /app/marinskyrl/cloud/iris/bootstrap_runtime.sh "$checkout" "$IRIS_VENV" \
     /app/marinskyrl/.iris-runtime-env fsdp production

The runtime profile is fsdp, resolved from uv.lock at that commit. See task-entrypoint-r4.md for the complete setup and run command.

Submit command

Run from the MarinSkyRL repository root:

uv run --frozen --with daytona python -m cloud.iris.iris_backend \
  --rl_config <abs path>/tasktrove_x10_fsdp2.yaml \
  --model_path Qwen/Qwen3-Coder-30B-A3B-Instruct \
  --train_data '["DCAgent/exp_rpt_multifile"]' \
  --num-nodes 4 \
  --cluster cw-rno2a \
  --priority interactive \
  --job-name tt-x10-fsdp2-fa2-r4 \
  --runtime-commit e1c2dff6a9633a9ac71fdd93a6e8890346cfcf0e \
  --no-wait

Two things that bite:

• --with daytona is required on macOS. The frozen lock installs harbor[daytona] only on linux, and the launcher's mandatory pre-launch Daytona snapshot purge imports the SDK on the launch host. Without it the submission dies at ModuleNotFoundError: No module named 'daytona', and it dies after a clean dry run, because the dry run skips the purge.
• Omitting --train_data is what killed r3. The flag is not inherited from the config.

Durable outputs

s3://marin-us-east-02a/iris/tt-x10-fsdp2-fa2/checkpoints          # shared across generations
s3://marin-us-east-02a/iris/tt-x10-fsdp2-fa2-r4/exports
s3://marin-us-east-02a/iris/tt-x10-fsdp2-fa2-r4/trace_jobs
s3://marin-us-east-02a/iris/cw-rno2a/rendezvous/tt-x10-fsdp2-fa2-r4

Cluster access used kubeconfig ~/.kube/coreweave-iris, context marin-rn02a_RNO2A.

Known caveat on the bank

Before the r2 resume, generation_buffer_state.pt inside global_step_75 was renamed to .pre364 on S3. It used the pre-#364 list layout and is unreadable at the current runtime. The data is preserved under the new name; a reproducer starting from that checkpoint will not find a generation buffer there.

---

X2 clip 0.2 — Jupiter (JSC) GH200, aarch64, 4 GPUs per node

Sweeps eps_clip_high with eps_clip_low held at 0.2. This arm is the 0.2 endpoint. 6 nodes, 11:59:00 wall limit per link, 80-step horizon, --max_restarts 5.

Cluster paths

What	Path
OpenThoughts-Agent repo (WORKDIR)	/e/scratch/jureap59/feuer1/OpenThoughts-Agent
SkyRL subcheckout	/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL
Launcher python (conda otagent)	/e/scratch/jureap59/feuer1/miniforge3/envs/otagent/bin/python
Container SIF	/e/scratch/jureap59/feuer1/containers/skyrl_megatron_vllm0202rc0_r5.sif
Apptainer overlay	/e/scratch/jureap59/feuer1/containers/skyrl_titan_overlay.img
Container PYTHONPATH deps	/e/scratch/jureap59/feuer1/sif_pydeps_titan_a1fdd7e:/e/scratch/jureap59/feuer1/sif_pydeps
Container binds	/e/scratch, /e/data1
Sweep configs	/e/scratch/jureap59/feuer1/tasktrove_sweep/artifacts
Extracted training data	/e/scratch/jureap59/feuer1/tasks/exp_rpt_multifile
Model snapshot	/e/home/jusers/feuer1/jupiter/.cache/huggingface/hub/models--Qwen--Qwen3-Coder-30B-A3B-Instruct/snapshots/b2cff646eb4bb1d68355c01b18ae02e7cf42d120
Experiments root	/e/data1/datasets/playground/ot-baf

Slurm: partition booster, account reformo, --gres=gpu:4, --cpus-per-task=288. There is no reservation; passing one makes sbatch reject the job.

Submit command

From the OpenThoughts-Agent repo root, with DCFT set to it:

CFG=/e/scratch/jureap59/feuer1/tasktrove_sweep/artifacts/x2_clip/configs/tasktrove_x2_clip_hi0p2.yaml
/e/scratch/jureap59/feuer1/miniforge3/envs/otagent/bin/python -m hpc.launch --job_type rl \
  --rl_config "$CFG" \
  --model_path Qwen/Qwen3-Coder-30B-A3B-Instruct \
  --train_data '["DCAgent/exp_rpt_multifile"]' \
  --num_nodes 6 \
  --time_limit 11:59:00 \
  --max_restarts 5 \
  --experiments_dir /e/data1/datasets/playground/ot-baf \
  --job_name tt-x2_clip-hi0p2

In practice this was driven by launcher/launch.sh, which reads the node count from the config header rather than duplicating the geometry rule:

ARM_FILTER=hi0p2 ./launch.sh x2_clip            # add --dry-run to print the command only

launch.sh globs *.yaml, so ARM_FILTER is required to submit a single arm without disturbing the others. It also staggers 6 minutes between arms, because simultaneous Hugging Face pulls have produced 429 deaths at high pod counts.

The launcher must run from the otagent conda environment, not the RL venv: task extraction imports google.cloud.storage, which the RL venv lacks. The launcher then selects the SIF for the training itself.

Commit pins — read this before trusting a single revision

This arm did not run at one pinned commit. It ran as a restart chain from 2026-07-31T20:42:22 to 2026-08-09T20:27:09, and the shared Jupiter checkouts moved underneath it several times. Both checkouts are shared across the whole campaign, so a git pull for one arm changes the code the next link of every other arm starts with.

Revisions at the two ends of the chain:

Repository	At the first link, 2026-07-31	At the last link, 2026-08-09
OpenThoughts-Agent	23387aa8	bcf9d268
SkyRL subcheckout	04ac9c11	08f9619d

The SkyRL checkout moved about fourteen times inside that window and OpenThoughts-Agent about five. Reflog timestamps are Jupiter local, which is UTC+2. If you need the exact revision for a specific link, take the link's start time from sacct --name=tt-x2_clip-hi0p2 and match it against the reflog.

For a single-revision reproduction, use the last-link pair above. That is the code that produced the final banked steps, and the campaign scores this arm on a trailing-five EMA at step 20, which is inside the earlier revisions.

The chain

Root and links, from sacct. States are the raw Slurm states; on this campaign a FAILED link is normally a crash that the next link resumes from, not a lost arm.

1147591 1147592                                   2026-07-31
1169962..1169967  1169976..1169981                2026-08-01
1170689 1170691 1170693 1170694 1170695 1170696   2026-08-01..08-03
1219766 1219767 1219768 1219769 1219770 1219771   2026-08-03..08-04
1234215                                           2026-08-04
1270775..1270777 1270779 1270781 1270782          2026-08-07..08-08
1287826..1287829                                  2026-08-08..08-09

Caveat on the two EXAMPLE files

jupiter-rendered-sbatch-EXAMPLE.sbatch and jupiter-launcher-config-EXAMPLE.json are not X2's own renders. X2's run tree under /e/data1/datasets/playground/ot-baf was removed during cleanup after the arm completed, and those files went with it. The two files here were rendered by the same launcher for a currently running arm (tt-x3_kl-kl0b), and they are included because they show the container wiring that no other artifact records: the SIF, the overlay, the pydeps PYTHONPATH, the module loads, and the python -m hpc.rl_launch_utils --config <json> entrypoint. Treat the geometry, the job name, and the hydra overrides in them as belonging to that other arm.

Regenerate X2's own sbatch by running launch.sh with --dry-run and then submitting; the launcher writes the render into the job's sbatch/ directory.

---

Data and permissions

Training data is the public DCAgent/exp_rpt_multifile. The Jupiter launcher extracts it to /e/scratch/jureap59/feuer1/tasks/exp_rpt_multifile and the hydra override points at the local path, not the Hub id.

All Jupiter paths listed above have been set world-readable, with executables and shell scripts world-executable, using OpenThoughts-Agent/scripts/permissions/fix_permissions.sh. That script excludes keys.env and secrets.env from every read pass and locks them to 600 last, so no credential is exposed by the change.

Scoring

The campaign scores an arm on the trailing-five-step EMA of reward/avg_raw_reward at the 20-step probe horizon, tie-broken by the slope over the final ten steps. curves/reward_raw.csv holds the per-step series for every arm, so the number is recomputable without the logs.