Configuration

There is no config.yaml. A run is one .env config file passed to one of three runners:

bash scripts/train/train.sh  train/configs/my_run.env
bash scripts/eval/eval.sh    eval/configs/my_eval.env
bash scripts/infer/infer.sh  infer/configs/my_infer.env

The runner sources the config, applies the model preset, resolves the scaffold and backend, runs preflight, then execs verl (fully_async_main or main_ppo) with the Hydra overrides built from those values. Every omitted key falls back to a default, so a config file is ~20 lines rather than ~100.

Config layers

Every value belongs to exactly one layer:

lib/common_env.sh sources site.env if it exists. No cluster-specific value is hardcoded in the runners or libs: with no site.env the stack falls back to a portable path — in-pod inline build, no mounts, no image rewrite. That path is slower but runs anywhere.

Parameter tiers

Templates carry T0–T1 plus a few T2 examples. Setting a T2/T3 knob is one KEY=VALUE line in the config. Multi-word values must be quoted (EVAL_VLLM_EXTRA_ARGS="--enforce-eager --dtype bfloat16").

Axes

Only these four require a different template. Everything else is a parameter.

Node counts (NNODES, N_NODES_TRAIN, N_NODES_ROLLOUT) are parameters, not axes: the same template scales from one node to four.

Model presets

A preset fills a slot of correlated values, so switching models is a one-line change. It only fills what the config left unset — an explicit MODEL_PATH always wins.

† tool_parser is not a function of the model alone. Qwen3.5/3.6 chat templates emit XML → always qwen3_coder. The 30B needs qwen3_coder under cc but hermes under ohsdk/oh. The preset resolves this from SCAFFOLD and preflight re-checks it; a wrong value means a 100% tool-parse failure rate on step 0 (see Tool-call parser mismatch).

For a model outside the table, leave MODEL_PRESET empty and set MODEL_PATH, TOOL_PARSER, ENGINE, SP_SIZE, MAX_PROMPT, and MAX_RESP explicitly.

Required fields

train — PROJECT_NAME · EXP_TAG · TRAIN_INDEX · VAL_INDEX · topology:

eval replaces the data inputs with DATASET_PATH (a directory of harbor task dirs) or DATASET_NAME + N_TASKS, and adds serving knobs: GEN_TP, EVAL_TEMPERATURE, MAX_INPUT_TOKENS / MAX_OUTPUT_TOKENS / MAX_MODEL_LEN, N_CONCURRENT, and EVAL_ENABLE_EXPERT_PARALLEL (dense → false, MoE → true).

infer takes a base TRAIN_INDEX plus RESULTS_DIR + OUTPUT_INDEX, an optional per-node INSTANCES_FILE shard, N_TRIALS / N_CONCURRENT, and the vLLM topology (GEN_TP, GEN_DP, VLLM_NNODES, VLLM_HEAD_HOST). A multi-node DP run uses the same config on every node; the node whose IP matches VLLM_HEAD_HOST becomes rank 0.

Common edits

One line each, in the run config (scripts/{train,eval,infer}/configs/*.env):

MODEL_PRESET=qwen36-27b        # switch policy model
# MODEL_PATH=/models/my-model TOOL_PARSER=qwen3_coder SP_SIZE=1   # model outside the preset table

SCAFFOLD=cc                    # ohsdk | oh | cc | oc
TRAIN_MODE=async               # sync ↔ async — also copy the matching template (entrypoint + YAML change)
BACKEND=docker                 # Docker sandboxes — see /docs/run-training/backends

NNODES=3                       # add nodes: NNODES must equal the sum
N_NODES_TRAIN=2
N_NODES_ROLLOUT=1

adv_estimator=grpo             # algorithm: ppo | grpo
policy_loss_mode=gspo

N_CONCURRENT=32                # rollout parallelism
TRAIN_BSZ=8 N_RESP=4           # smoke run (plumbing check, not a learning signal)

MODEL_PATH=/ckpt/global_step_30_hf   # eval a checkpoint (in an eval config)
WANDB_MODE=disabled            # opt out of wandb

Override a single run without editing any file:

K8S_KUBECONFIG=/tmp/other.yaml bash scripts/train/train.sh <config>

VENV_PATH reuses a prebuilt venv. Verify its editable installs resolve to the trees you expect — see Venv.

Site variables

scripts/lib/site.env, set once per cluster:

K8S_KUBECONFIG=/path/to/kubeconfig.yaml        # which K8s cluster
MODEL_ROOT=/models                             # where presets resolve model paths
NEW_VERL_DIR=/repos/verl-swe_agent_opd_dev     # required with USE_NEW_VERL=1
HARBOR_OPENSWE_IMAGE_REGISTRY=reg:5001/openswe # prebuilt task images
HARBOR_NYDUS_MIRROR=...                        # mirror official val images
DOCKER_HOST=tcp://10.0.0.2:2376                # remote docker daemon

The full surface. An empty value is always legal; it turns the corresponding acceleration off.

Hyperparameters

The default profile is GRPO advantages with a GSPO policy loss, lr 1e-6, and 64 × 8 = 512 trials per step. The context window comes from the model preset. To move off the defaults:

adv_estimator=grpo            # ppo | grpo
policy_loss_mode=gspo         # sequence-level policy loss
learning_rate=1.0e-06
LR_SCHEDULER=constant         # keep constant on fully-async (see preflight)
TRAIN_BSZ=64
N_RESP=8                      # 64 × 8 = 512 trials/step
MAX_PROMPT=40000
MAX_RESP=91072                # window = prompt + response
SAVE_FREQ=5
TOTAL_EPOCHS=3
TEMPERATURE=1.0

The meaning of each symbol is defined in Core Concepts.

wandb

The API key never belongs in a config file:

export WANDB_API_KEY=...       # in your shell, before launch
# or: WANDB_MODE=disabled

Outputs

A run writes, relative to the repo root:

infer additionally writes RESULTS_DIR (per-trial results) and OUTPUT_INDEX (the selected-instances parquet); eval writes harbor trial dirs plus the score summary in its log. Layout details are in Results & Artifacts.