openenv/tbench2: DeepSeek-V4-Flash launcher variant

examples/experimental/openenv/README.md

@@ -0,0 +1,90 @@
+# OpenEnv Terminal-Bench-2 GRPO (GLM-4.7-Flash, single node)
+
+Train GLM-4.7-Flash with GRPO on the HuggingFace [OpenEnv](https://github.com/huggingface/openenv)
+**Terminal-Bench-2 (tbench2)** environment. A miles-side adapter runs the multi-turn
+agentic loop (`reset(task_id)` → { policy emits one shell command → `step(exec)` →
+feed output back } → `evaluate`) against an unmodified OpenEnv env server; the reward
+is the binary pytest result (1.0 = all tests pass, else 0.0).
+
+This guide targets a **single H200 node with 8 GPUs**. The run is colocated
+(training + rollout on the same 8 GPUs): TP=4, EP=2, one SGLang engine per GPU.
+
+## Prerequisites
+
+- The node has Docker available (the env server launches one container per task).
+- miles is installed and GLM-4.7-Flash weights are reachable (the launcher pulls
+  `zai-org/GLM-4.7-Flash` from HF and converts it to `torch_dist` on first run).
+- Install the OpenEnv tbench2 env client (isolate it if its deps clash with the
+  miles image):
+
+  ```bash
+  pip install -e <OpenEnv>/envs/tbench2_env
+  ```
+
+## 1. Build the prompt data
+
+Clone the TB2 suite and emit one prompt row per `task_id`:
+
+```bash
+git clone --depth 1 https://github.com/laude-institute/terminal-bench-2.git /workspace/terminal-bench-2
+python make_tbench2_data.py --tasks_dir /workspace/terminal-bench-2 --output /root/tbench2_train.jsonl
+# add --n 8 for a small smoke subset
+```
+
+## 2. Start the env server
+
+Run it in a separate shell (or off-node — see note). Docker mode gives real TB2
+fidelity; it needs the Docker socket and pulls the per-task images on first use:
+
+```bash
+# Raise the open-file limit first (see Notes): the WebSocket env server holds an
+# FD per live session + Docker connection and leaks sockets on unclean
+# disconnects, so the default 1024 soft limit is exhausted on a long run.
+ulimit -n 1048576
+TB2_MODE=docker TB2_TASKS_DIR=/workspace/terminal-bench-2 MAX_CONCURRENT_ENVS=32 \
+    python -m tbench2_env.server.app --port 8003
+```
+
+`MAX_CONCURRENT_ENVS` caps live sandboxes; keep it at or below the rollout batch
+concurrency. Per-task containers are heavy on disk — if you'd rather not colocate
+them with the GPU workload, run the env server on a separate Docker host and point
+the launcher at it via `--openenv-env-url http://<env-host>:8003`.
+
+## 3. Launch training
+
+```bash
+python run-openenv-tbench2.py --openenv-env-url http://localhost:8003
+```
+
+Common overrides:
+
+| Flag / env var | Default | Purpose |
+| --- | --- | --- |
+| `--openenv-env-url` | `http://localhost:8003` | Env server URL |
+| `--prompt-data` | `/root/tbench2_train.jsonl` | Prompt set from step 1 |
+| `--num-rollout` | (launcher) | Number of GRPO steps |
+| `OPENENV_MAX_TURNS` | `30` | Max agent turns per episode |
+| `OPENENV_MAX_ROLLOUT_TIME_SECONDS` | `3600` | Per-episode wall-clock cap; a straggler that exceeds it is terminated and scored 0 |
+| `--dump-details <dir>` | off | Dump per-episode tokens/logprobs/masks/reward for inspection |
+| `WANDB_KEY`, `--wandb-project`, `--wandb-team` | — | W&B logging |
+
+## Notes
+
+- **Reward signal.** The binary sparse reward needs a task subset where the base
+  policy *sometimes* succeeds (advantage variance). On the full TB2 suite,
+  GLM-4.7-Flash's low base solve-rate yields a near-flat GRPO signal — use a
+  variance-band subset (or a stronger base) to see a learning climb.
+- **`_step` vs. rollout.** W&B `_step` is an internal log-call index that advances
+  several times per rollout; it is **not** the training step. Read the driver log's
+  `rollout N:` counter for true progress.
+- **Sandbox leakage.** Upstream OpenEnv creates task containers with `remove=False`
+  and only tears them down on a clean session close (the idle reaper is off by
+  default), so an unclean disconnect (trainer crash) can orphan containers. Sweep
+  stale TB2 containers between runs, e.g. `docker rm -f` of any older than the
+  episode wall-cap.
+- **Open-file limit.** The same unclean disconnects also leak socket FDs in the
+  env server process. On a long run under the default 1024 soft limit the accept
+  loop eventually fails every connection with `OSError: [Errno 24] Too many open
+  files`, silently throttling rollouts. Start the server with a raised limit
+  (`ulimit -n 1048576`, as in step 2); if a running server is already saturated,
+  restart it with the higher limit.

examples/experimental/openenv/make_tbench2_data.py

@@ -0,0 +1,78 @@
+"""Generate a prompt dataset for the OpenEnv Terminal-Bench-2 (tbench2) run.
+
+The *tasks* are not hand-written: they are the
+Terminal-Bench-2 suite shipped in the laude-institute/terminal-bench-2 repo. The
+env serves the per-task instruction at reset(), so each prompt-data row only
+needs the system prompt (how the agent should behave) plus the ``task_id`` in
+metadata. ``openenv_agent_function`` drives the multi-turn loop and reads
+metadata["task_id"].
+
+task_ids are the top-level task directory names in the TB2 repo checkout
+(e.g. "chess-best-move"); a valid task dir contains a ``task.toml``.
+
+    # all 89 tasks
+    python make_tbench2_data.py --tasks_dir /workspace/terminal-bench-2 \
+        --output /root/tbench2_train.jsonl
+    # a small smoke subset
+    python make_tbench2_data.py --tasks_dir /workspace/terminal-bench-2 \
+        --output /root/tbench2_smoke.jsonl --n 8
+    # an explicit subset
+    python make_tbench2_data.py --tasks_dir /workspace/terminal-bench-2 \
+        --tasks chess-best-move,circuit-fibsqrt
+"""
+
+import json
+from pathlib import Path
+
+from tap import Tap
+
+# The agent contract must match openenv_agent_function._multi_turn: one shell
+# command per turn inside a single ```bash block; TASK_COMPLETE to stop.
+_SYSTEM = (
+    "You are an autonomous terminal agent solving a Terminal-Bench task. You will "
+    "be given the task instruction, then interact with a real Linux shell. On each "
+    "turn respond with EXACTLY ONE shell command inside a single ```bash code block "
+    "and nothing else. Inspect the environment, make the required changes, and "
+    "verify your work. When you are confident the task is fully complete, reply with "
+    "TASK_COMPLETE (with no code block)."
+)
+
+
+class Args(Tap):
+    tasks_dir: str = "/workspace/terminal-bench-2"  # TB2 repo checkout
+    output: str = "/root/tbench2_train.jsonl"
+    n: int = 0  # 0 = all discovered tasks
+    tasks: str = ""  # optional comma-separated explicit task_ids
+
+
+def _discover_task_ids(tasks_dir: Path) -> list[str]:
+    return sorted(p.name for p in tasks_dir.iterdir() if (p / "task.toml").is_file())
+
+
+def main() -> None:
+    args = Args().parse_args()
+    tasks_dir = Path(args.tasks_dir).expanduser().resolve()
+
+    if args.tasks:
+        task_ids = [t.strip() for t in args.tasks.split(",") if t.strip()]
+    else:
+        task_ids = _discover_task_ids(tasks_dir)
+        if args.n > 0:
+            task_ids = task_ids[: args.n]
+
+    missing = [t for t in task_ids if not (tasks_dir / t / "task.toml").is_file()]
+    if missing:
+        raise SystemExit(f"task_ids without a task.toml in {tasks_dir}: {missing}")
+
+    with open(args.output, "w") as f:
+        for tid in task_ids:
+            row = {
+                "prompt": [{"role": "system", "content": _SYSTEM}],
+                "metadata": {"task_id": tid},
+            }
+            f.write(json.dumps(row) + "\n")
+    print(f"Wrote {len(task_ids)} TB2 tasks to {args.output}")
+
+
+if __name__ == "__main__":
+    main()