Nemo Gym Recipe Integration

Integrating a NeMo Gym recipe

Build a new recipe by following the proven Calendar flow and the accumulated PITFAIL records. Keep the user-facing workflow to exactly three steps:

1. prepare data;
2. start the local NeMo Gym service;
3. start remote Relax training.

Non-negotiable rules

• Deliver executable scripts, not a prose-only procedure.
• Run steps 1 and 2 locally. Fix and retry until data, readiness, and the verifier all pass.
• Put only deployment-specific values in env.sh: image, model path, Gym host/port, shared data path, and optionally remote Ray address.
• Write training hyperparameters directly in the training script. Do not create NEMO_GYM_NUM_ROLLOUT, NEMO_GYM_N_SAMPLES_PER_PROMPT, context-length, batch-size, or parser environment variables.
• The training script must contain the complete parameter arrays and final ray job submit. It may use scripts/run_training.sh as the remote job entry and scripts/run_agent_app.sh as the thin agent client, but must not delegate to another model training recipe.
• Never enable --no-wait by default. Add it only when the caller explicitly sets RAY_NO_WAIT=1.
• Data preparation must create the exact shared file checked by training. A file left only in a Docker volume is not prepared data.
• Training consumes the raw NeMo Gym JSONL. Do not pass an already converted *_relax.jsonl back through convert_dataset.py.
• Do not claim success from /readyz, a Ray Job SUCCEEDED state, or a one-sample rollout alone.
• A NeMo HTTP 500 must leave a traceback in docker logs; do not accept access-log-only failures.
• Preserve user changes and replace only the named recipe container after verifying its ownership label.
• Never run broad ray stop, pkill, Docker prune, or sandbox cleanup on a shared system.

Read before editing

Read these files completely because they evolve with real failures:

rg --files examples/nemo_gym_agentic/recipes | rg '/PITFAIL\.md$' | sort

Then read:

• every returned PITFAIL.md;
• the target environment config in the pinned NeMo Gym checkout;
• the target agent, resource server, dataset declaration, verifier, and cleanup behavior;
• examples/nemo_gym_agentic/service/Dockerfile;
• examples/nemo_gym_agentic/scripts/convert_dataset.py;
• examples/nemo_gym_agentic/scripts/run_training.sh;
• examples/nemo_gym_agentic/scripts/run_agent_app.sh;
• the Calendar recipe as the simple-agent reference;
• Workplace Assistant for stateful tool/resource cleanup;
• R2E-Gym for sandbox, artifact, and multi-process callback propagation.

Do not infer graph names from directory names. Record the exact contract before implementing:

config path:
agent name and type:
resource server name:
dataset repo/artifact/split:
raw verifier fields:
stateful cleanup requirement:
tool-call parser:
reasoning parser:
maximum agent steps:

Required recipe files

Create the following under examples/nemo_gym_agentic/recipes/<recipe>/:

env.sh
prepare_<recipe>.sh
start_<recipe>_gym.sh
run-<model>-nemo-gym-<recipe>.sh
verify_<recipe>.py
README.md
PITFAIL.md

Use the exact graph names from the pinned Gym config. If the agent/resource needs a dedicated venv, patch, or runtime asset, wire it into the shared Dockerfile and build it into the image.

Step 1: prepare data

The host entry script must source its sibling env.sh and require an absolute NEMO_GYM_SOURCE_DATA ending in the expected raw filename.

Implementation requirements:

1. Download or materialize the selected split inside the pinned Gym image.
2. Write the raw JSONL directly to a filesystem visible at the same absolute path from every remote Ray node.
3. Do not assume that a path visible in the development container is bind-mountable by an outer Docker daemon. Probe the mount from a temporary container when Docker-in-Docker is possible.
4. Preserve all verifier and agent fields. In particular, keep responses_create_params, tools, ground truth, expected answers, environment-specific metadata, repository identity, and unknown passthrough fields.
5. Validate that every non-empty line is a JSON object and that raw/conversion checks retain the expected row count.
6. Do not hardcode a dataset row count or tool count from a README; validate the downloaded artifact.
7. Use the correct split. A held-out benchmark/test split is smoke/evaluation data, not an RL training set.

Conversion has one source of truth:

shared raw Gym JSONL
    -> Ray job calls convert_dataset.py once
    -> ${EXP_DIR}/data/<recipe>_train.jsonl

An optional converted file produced during preparation is only a schema check. Document that clearly and never use it as NEMO_GYM_SOURCE_DATA.

Before continuing, run the prepare script and assert the exact raw path is non-empty.

Step 2: start the local Gym service

Prefer a local Docker service with a private Ray cluster. The remote Relax cluster communicates with it only over HTTP.

Keep the two Ray systems separate:

Gym private Ray: owned by the local container
Relax Ray:       RAY_ADDRESS on the remote training cluster

The Gym launcher must unset inherited RAY_ADDRESS and RAY_JOB_SUBMISSION_ID before starting its private Ray.

Networking requirements:

• GYM_HOST must be assigned to the Docker daemon host and reachable from every Relax worker.
• The Gym host must reach the Relax head's Agentic Chat API on port 8000.
• The callback base URL must include /agentic_api.
• The callback allowlist contains the bare host/IP used by the Relax callback URL. A CIDR is an allowlist, not a Ray address.
• Put Gym and Relax internal addresses in NO_PROXY; do not rely on an HTTP proxy to bridge internal traffic.
• Choose a unique service port block and a non-overlapping private Ray port block. Avoid the host ephemeral port range.
• The service port base and the training NEMO_GYM_GATEWAY_PORT must resolve to the same Gateway port.

The launcher must build NEMO_GYM_GATEWAY_ENVIRONMENTS_JSON with the exact environment, config, agent URL, readiness URLs, interrupt policy, concurrency, queue, and deadline.

Stateful environments require a real cleanup contract keyed by opaque rollout ID. Add cleanup and probe endpoints if the resource server retains sessions, databases, sandboxes, or containers. Do not report cancellation as clean merely because the local HTTP task was cancelled.

Container lifecycle:

1. build a fresh image containing the current checkout;
2. verify the existing named container has the expected recipe label;
3. replace only that container;
4. wait with a bounded readiness loop and show logs on timeout;
5. verify /readyz reports the pinned Gym commit, healthy janitor, and zero active trials.

Run verify_<recipe>.py against the real resource server. It must prove that a known-correct response receives the success reward and a known-wrong response does not.

For multi-turn or tool environments, also run a deterministic full trial and verify callback history, tool-result feedback, final reward, artifacts, and cleanup. A direct resource verifier alone does not validate callback routing.

Step 3: write the direct training script

Follow the Calendar/R2E layout:

1. resolve SCRIPT_DIR, EXAMPLE_DIR, RELAX_ROOT, run_training.sh, and run_agent_app.sh;
2. source the Relax entrypoint and model config;
3. validate model and raw data paths;
4. prepend the current RELAX_ROOT to RUNTIME_ENV_JSON.env_vars.PYTHONPATH and set py_executable=/usr/bin/python3;
5. derive the Ray Jobs dashboard from RAY_ADDRESS, unless RAY_DASHBOARD_ADDRESS is explicit;
6. generate a unique submission ID;
7. define complete checkpoint, rollout, GRPO, optimizer, performance, SGLang, tracking, and misc argument arrays in this script;
8. submit run_training.sh with the Gateway URL, raw source data, converted prompt output, and actual data count.

Hard-code recipe training choices in the script, for example:

--num-rollout 3
--rollout-batch-size 1
--n-samples-per-prompt 8
--global-batch-size 8
--rollout-max-prompt-len 6144
--rollout-max-response-len 2048
--rollout-max-context-len 8192
--agentic-reasoning-parser qwen3

Change these values by editing the recipe script, not by adding environment-variable wrappers.

Set --agent-env with the exact Gateway URL, environment, config, model alias, interrupt policy, deadline, and lease. Preserve the opaque rollout prefix across every subprocess, Ray boundary, sandbox, and model callback. Model callbacks must use:

/ng-rollout/<opaque-rollout-id>/v1/responses

or the corresponding /v1/chat/completions route.

Select parsers from actual model output:

• Qwen reasoning output needs --agentic-reasoning-parser qwen3 when the verifier rejects inline <think>.
• Tool-call parsers must match the model/chat-template format and only matter when tools are present.

For GRPO, one sample only validates plumbing. Use at least four samples per prompt for a meaningful group advantage and confirm reward variance. Keep prompt, response, context, and per-GPU token budgets compatible.

Do not use Ray WORKING_DIR=./ when every cluster node already sees the same shared checkout. Resolve agent commands and working directories to identical absolute shared paths.

Validation ladder

Do not skip levels or substitute one level for another.

1. Static: bash -n, focused unit tests, converter preservation tests, and Docker patch apply checks.
2. Data: raw path exists, rows parse, required verifier fields survive, and train does not accidentally use a test split.
3. Service: fresh image/container, /readyz, expected commit, zero initial trials, and deterministic verifier pass/fail behavior.
4. Network: every remote Ray node can reach the Gateway; the Gym container can reach the active Relax /agentic_api route.
5. Trial: one real trial completes with expected turns/tools/artifacts/reward and returns active_trials to zero.
6. Training: multi-sample rollout produces reward variance, non-zero advantage where expected, an optimizer step, no Actor OOM/traceback, and a checkpoint.

Ray Job SUCCEEDED is not sufficient. Inspect rollout JSONL, Actor logs, optimizer metrics, checkpoint output, Gateway trial counts, and environment-specific sandbox cleanup.

Failure triage

Reward zero is not automatically an integration failure. Separate these cases:

• verifier never ran or could not parse the response;
• environment/evaluator failed;
• callback/lifecycle failed;
• the model completed the environment but produced a genuinely wrong answer.

Definition of done

The recipe is complete only when all of the following are true:

• The seven required recipe files exist and use the pinned graph names.
• One data command creates the exact shared raw source path.
• One service command starts a fresh labeled container and reaches ready.
• The verifier distinguishes correct and incorrect responses.
• One direct training script contains the actual hyperparameters and waits by default.
• Remote nodes reach Gym and Gym reaches the active Relax callback route.
• A real multi-sample training run reaches an optimizer step and writes a checkpoint.
• Actor logs contain no hidden OOM/traceback.
• Gateway active trials and environment resources return to a clean state.
• New environment-specific failures are recorded concisely in that recipe's PITFAIL.md.

When handing off, show only the three commands the user needs, the resulting data/checkpoint paths, and any remaining unverified item.