simrig 0.2.2__tar.gz → 0.3.0__tar.gz

{simrig-0.2.2 → simrig-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: simrig
-Version: 0.2.2
+Version: 0.3.0
 Summary: Agent- and human-friendly workflows for MuJoCo Playground training and evaluation.
 Author: SimRig contributors
 License-Expression: MIT
@@ -16,11 +16,13 @@ Requires-Dist: mujoco; extra == "mujoco"
 Provides-Extra: hf
 Requires-Dist: huggingface_hub; extra == "hf"
 Provides-Extra: playground
-Requires-Dist: gymnasium[mujoco]; extra == "playground"
-Requires-Dist: mujoco; extra == "playground"
-Requires-Dist: jax; extra == "playground"
-Requires-Dist: brax; extra == "playground"
-Requires-Dist: playground; extra == "playground"
+Requires-Dist: gymnasium[mujoco]==1.3.0; extra == "playground"
+Requires-Dist: mujoco==3.10.0; extra == "playground"
+Requires-Dist: mujoco-mjx==3.10.0; extra == "playground"
+Requires-Dist: jax==0.10.2; extra == "playground"
+Requires-Dist: brax==0.14.2; extra == "playground"
+Requires-Dist: playground==0.2.0; extra == "playground"
+Requires-Dist: warp-lang==1.15.0; extra == "playground"
 Provides-Extra: dev
 Requires-Dist: build; extra == "dev"
 Requires-Dist: numpy; extra == "dev"
@@ -216,10 +218,38 @@ transforms update the browser scene. The camera follows the robot without
 streaming rendered image frames. Use `--render-mode mujoco` for the older local
 image-stream preview or `--render-mode topdown` for the schematic fallback.
 
+### Train on a Lambda Cloud GPU
+
+After launching a Lambda On-Demand instance with an SSH key, SimRig can connect,
+sync this checkout, verify JAX GPU visibility, train, monitor a detached run,
+and download its artifacts:
+
+```bash
+simrig cloud lambda connect INSTANCE_IP --identity ~/Downloads/lambda-key.pem
+simrig cloud lambda prepare INSTANCE_IP --identity ~/Downloads/lambda-key.pem
+simrig cloud lambda smoke INSTANCE_IP Go1JoystickFlatTerrain \
+  --identity ~/Downloads/lambda-key.pem
+simrig cloud lambda train INSTANCE_IP Go1JoystickFlatTerrain \
+  --identity ~/Downloads/lambda-key.pem \
+  --preset smoke
+```
+
+Only after the environment and PPO smoke gates pass, start a detached large
+run with `--preset cloud --detach`. SimRig operates on an instance you already
+provisioned; it never launches or terminates billable Lambda resources. See the
+complete [Lambda Cloud GPU guide](docs/lambda-cloud.md), including persistent
+storage, monitoring, artifact download, and shutdown reminders.
+
+Lambda preparation requires Python 3.11+ and installs a pinned Playground
+training stack. Every run records its Python and package versions; checkpoint
+eval, demo, and preview reject a different recorded runtime unless
+`--allow-runtime-mismatch` is explicitly selected for qualitative review.
+
 ## Documentation
 
 - [Examples](examples/README.md)
 - [Agent workflow](docs/agent_workflow.md)
+- [Lambda Cloud GPU training](docs/lambda-cloud.md)
 - [Agent skill installation](docs/skill-installation.md)
 - [Contributing](CONTRIBUTING.md)
 - [SimRig skill source](skills/simrig/SKILL.md)

simrig-0.2.2/simrig.egg-info/PKG-INFO → simrig-0.3.0/README.md

@@ -1,33 +1,3 @@
-Metadata-Version: 2.4
-Name: simrig
-Version: 0.2.2
-Summary: Agent- and human-friendly workflows for MuJoCo Playground training and evaluation.
-Author: SimRig contributors
-License-Expression: MIT
-Project-URL: Homepage, https://github.com/Su1eym4n/simrig
-Project-URL: Repository, https://github.com/Su1eym4n/simrig
-Project-URL: Issues, https://github.com/Su1eym4n/simrig/issues
-Project-URL: Changelog, https://github.com/Su1eym4n/simrig/blob/main/CHANGELOG.md
-Requires-Python: >=3.10
-Description-Content-Type: text/markdown
-License-File: LICENSE
-Provides-Extra: mujoco
-Requires-Dist: mujoco; extra == "mujoco"
-Provides-Extra: hf
-Requires-Dist: huggingface_hub; extra == "hf"
-Provides-Extra: playground
-Requires-Dist: gymnasium[mujoco]; extra == "playground"
-Requires-Dist: mujoco; extra == "playground"
-Requires-Dist: jax; extra == "playground"
-Requires-Dist: brax; extra == "playground"
-Requires-Dist: playground; extra == "playground"
-Provides-Extra: dev
-Requires-Dist: build; extra == "dev"
-Requires-Dist: numpy; extra == "dev"
-Requires-Dist: pytest; extra == "dev"
-Requires-Dist: twine; extra == "dev"
-Dynamic: license-file
-
 # SimRig
 
 Turn MuJoCo robots into trained policies with agent-guided task design, PPO
@@ -216,10 +186,38 @@ transforms update the browser scene. The camera follows the robot without
 streaming rendered image frames. Use `--render-mode mujoco` for the older local
 image-stream preview or `--render-mode topdown` for the schematic fallback.
 
+### Train on a Lambda Cloud GPU
+
+After launching a Lambda On-Demand instance with an SSH key, SimRig can connect,
+sync this checkout, verify JAX GPU visibility, train, monitor a detached run,
+and download its artifacts:
+
+```bash
+simrig cloud lambda connect INSTANCE_IP --identity ~/Downloads/lambda-key.pem
+simrig cloud lambda prepare INSTANCE_IP --identity ~/Downloads/lambda-key.pem
+simrig cloud lambda smoke INSTANCE_IP Go1JoystickFlatTerrain \
+  --identity ~/Downloads/lambda-key.pem
+simrig cloud lambda train INSTANCE_IP Go1JoystickFlatTerrain \
+  --identity ~/Downloads/lambda-key.pem \
+  --preset smoke
+```
+
+Only after the environment and PPO smoke gates pass, start a detached large
+run with `--preset cloud --detach`. SimRig operates on an instance you already
+provisioned; it never launches or terminates billable Lambda resources. See the
+complete [Lambda Cloud GPU guide](docs/lambda-cloud.md), including persistent
+storage, monitoring, artifact download, and shutdown reminders.
+
+Lambda preparation requires Python 3.11+ and installs a pinned Playground
+training stack. Every run records its Python and package versions; checkpoint
+eval, demo, and preview reject a different recorded runtime unless
+`--allow-runtime-mismatch` is explicitly selected for qualitative review.
+
 ## Documentation
 
 - [Examples](examples/README.md)
 - [Agent workflow](docs/agent_workflow.md)
+- [Lambda Cloud GPU training](docs/lambda-cloud.md)
 - [Agent skill installation](docs/skill-installation.md)
 - [Contributing](CONTRIBUTING.md)
 - [SimRig skill source](skills/simrig/SKILL.md)

{simrig-0.2.2 → simrig-0.3.0}/pyproject.toml

@@ -23,11 +23,13 @@ Changelog = "https://github.com/Su1eym4n/simrig/blob/main/CHANGELOG.md"
 mujoco = ["mujoco"]
 hf = ["huggingface_hub"]
 playground = [
-  "gymnasium[mujoco]",
-  "mujoco",
-  "jax",
-  "brax",
-  "playground",
+  "gymnasium[mujoco]==1.3.0",
+  "mujoco==3.10.0",
+  "mujoco-mjx==3.10.0",
+  "jax==0.10.2",
+  "brax==0.14.2",
+  "playground==0.2.0",
+  "warp-lang==1.15.0",
 ]
 dev = ["build", "numpy", "pytest", "twine"]
 

{simrig-0.2.2 → simrig-0.3.0}/simrig/_version.py

@@ -1,3 +1,3 @@
 """Single source of truth for the SimRig package version."""
 
-__version__ = "0.2.2"
+__version__ = "0.3.0"

{simrig-0.2.2 → simrig-0.3.0}/simrig/cli.py

@@ -12,6 +12,16 @@ from simrig._version import __version__
 from simrig.core import report_markdown, to_dict
 from simrig.huggingface import resolve_policy_checkpoint
 from simrig.io import save_json, save_report_pair, slugify
+from simrig.lambda_cloud import (
+    LambdaSSHConfig,
+    check_lambda,
+    connect_lambda,
+    fetch_lambda,
+    prepare_lambda,
+    smoke_lambda,
+    status_lambda,
+    train_lambda,
+)
 from simrig.mujoco_backend import inspect_model, list_models
 from simrig.paths import ensure_project_dirs
 from simrig.playground_backend import (
@@ -128,6 +138,11 @@ def build_parser() -> argparse.ArgumentParser:
     eval_parser.add_argument("--small-network", action=argparse.BooleanOptionalAction, default=None)
     eval_parser.add_argument("--hf-revision", help="Revision for hf:// policy checkpoints.")
     eval_parser.add_argument("--hf-token", help="Hugging Face token for private policy repos.")
+    eval_parser.add_argument(
+        "--allow-runtime-mismatch",
+        action="store_true",
+        help="Allow an explicitly qualitative rollout when recorded runtime versions differ.",
+    )
     eval_parser.add_argument("--json", action="store_true")
     eval_parser.set_defaults(func=_cmd_eval)
 
@@ -142,6 +157,11 @@ def build_parser() -> argparse.ArgumentParser:
     demo_parser.add_argument("--command", type=float, nargs="+")
     demo_parser.add_argument("--speed", type=float, default=1.0)
     demo_parser.add_argument("--camera-distance", type=float)
+    demo_parser.add_argument(
+        "--allow-runtime-mismatch",
+        action="store_true",
+        help="Allow an explicitly qualitative demo when recorded runtime versions differ.",
+    )
     demo_parser.add_argument("--json", action="store_true")
     demo_parser.set_defaults(func=_cmd_demo)
 
@@ -161,6 +181,11 @@ def build_parser() -> argparse.ArgumentParser:
     preview_parser.add_argument("--command", type=float, nargs="+")
     preview_parser.add_argument("--camera")
     preview_parser.add_argument("--paused", action="store_true", help="Start the browser preview paused.")
+    preview_parser.add_argument(
+        "--allow-runtime-mismatch",
+        action="store_true",
+        help="Allow an explicitly qualitative preview when recorded runtime versions differ.",
+    )
     preview_parser.add_argument(
         "--render-mode",
         choices=("threejs", "mujoco", "topdown"),
@@ -215,6 +240,119 @@ def build_parser() -> argparse.ArgumentParser:
     validate_env_parser.add_argument("--json", action="store_true")
     validate_env_parser.set_defaults(func=_cmd_validate_env)
 
+    cloud_parser = sub.add_parser(
+        "cloud",
+        help="Connect to and train on an already-provisioned cloud GPU.",
+    )
+    cloud_providers = cloud_parser.add_subparsers(dest="cloud_provider", required=True)
+    lambda_parser = cloud_providers.add_parser(
+        "lambda",
+        help="Use a Lambda On-Demand Cloud instance over SSH.",
+    )
+    lambda_actions = lambda_parser.add_subparsers(dest="cloud_action", required=True)
+
+    lambda_connect = lambda_actions.add_parser(
+        "connect",
+        help="Open an interactive SSH connection to the instance.",
+    )
+    _add_lambda_connection_args(lambda_connect)
+    lambda_connect.add_argument(
+        "--tunnel-port",
+        type=int,
+        help="Forward this localhost port to the same port on the instance.",
+    )
+    lambda_connect.set_defaults(func=_cmd_lambda_connect)
+
+    lambda_check = lambda_actions.add_parser(
+        "check",
+        help="Check SSH plus NVIDIA and JAX device visibility.",
+    )
+    _add_lambda_connection_args(lambda_check)
+    lambda_check.set_defaults(func=_cmd_lambda_check)
+
+    lambda_prepare = lambda_actions.add_parser(
+        "prepare",
+        help="Sync this checkout and prepare a GPU-enabled remote virtualenv.",
+    )
+    _add_lambda_connection_args(lambda_prepare)
+    lambda_prepare.add_argument("--project", type=Path, default=Path("."))
+    lambda_prepare.add_argument("--remote-dir")
+    lambda_prepare.add_argument(
+        "--jax-cuda",
+        choices=("preinstalled", "cuda12", "cuda13"),
+        default="preinstalled",
+        help="Use Lambda's preinstalled JAX or install a pip CUDA wheel.",
+    )
+    lambda_prepare.add_argument(
+        "--python",
+        dest="python_command",
+        default="python3",
+        help="Remote Python 3.11+ executable used to create the training virtualenv.",
+    )
+    lambda_prepare.set_defaults(func=_cmd_lambda_prepare)
+
+    lambda_smoke = lambda_actions.add_parser(
+        "smoke",
+        help="Run the reset/step environment smoke gate on the remote GPU.",
+    )
+    _add_lambda_connection_args(lambda_smoke)
+    lambda_smoke.add_argument(
+        "env_name",
+        help="Playground env name or synced custom *.py env module path.",
+    )
+    lambda_smoke.add_argument("--remote-dir")
+    lambda_smoke.add_argument("--steps", type=int, default=10)
+    lambda_smoke.set_defaults(func=_cmd_lambda_smoke)
+
+    lambda_train = lambda_actions.add_parser(
+        "train",
+        help="Run training on a prepared Lambda instance (smoke preset by default).",
+    )
+    _add_lambda_connection_args(lambda_train)
+    lambda_train.add_argument(
+        "env_name",
+        help="Playground env name or synced custom *.py env module path.",
+    )
+    lambda_train.add_argument(
+        "--preset",
+        choices=("smoke", "local", "cloud"),
+        default="smoke",
+    )
+    lambda_train.add_argument("--remote-dir")
+    lambda_train.add_argument(
+        "--output",
+        help="Remote run directory, relative to --remote-dir unless absolute.",
+    )
+    lambda_train.add_argument(
+        "--detach",
+        action="store_true",
+        help="Keep training after SSH disconnects and write train.log/train.pid.",
+    )
+    lambda_train.add_argument("--timesteps", type=int)
+    lambda_train.add_argument("--num-envs", type=int)
+    lambda_train.add_argument("--batch-size", type=int)
+    lambda_train.set_defaults(func=_cmd_lambda_train)
+
+    lambda_status = lambda_actions.add_parser(
+        "status",
+        help="Show detached process state and recent training log lines.",
+    )
+    _add_lambda_connection_args(lambda_status)
+    lambda_status.add_argument("output", help="Remote output returned by cloud lambda train.")
+    lambda_status.add_argument("--remote-dir")
+    lambda_status.add_argument("--lines", type=int, default=30)
+    lambda_status.set_defaults(func=_cmd_lambda_status)
+
+    lambda_fetch = lambda_actions.add_parser(
+        "fetch",
+        help="Download a remote run directory into local runs/.",
+    )
+    _add_lambda_connection_args(lambda_fetch)
+    lambda_fetch.add_argument("output", help="Remote output returned by cloud lambda train.")
+    lambda_fetch.add_argument("--remote-dir")
+    lambda_fetch.add_argument("--local-output", type=Path)
+    lambda_fetch.set_defaults(func=_cmd_lambda_fetch)
+
     return parser
 
 
@@ -292,6 +430,7 @@ def _cmd_eval(args: argparse.Namespace) -> None:
         small_network=args.small_network,
         seed=args.seed,
         command=command,
+        allow_runtime_mismatch=args.allow_runtime_mismatch,
     )
     save_json(Path("reports") / f"{slugify(args.env_name)}_eval.json", result)
     _print(result, as_json=args.json)
@@ -313,6 +452,7 @@ def _cmd_demo(args: argparse.Namespace) -> None:
         command=command,
         speed=args.speed,
         camera_distance=args.camera_distance,
+        allow_runtime_mismatch=args.allow_runtime_mismatch,
     )
     _print(result, as_json=args.json)
 
@@ -357,6 +497,7 @@ def _cmd_preview(args: argparse.Namespace) -> None:
         render_mode=args.render_mode,
         paused=args.paused,
         fps=args.fps,
+        allow_runtime_mismatch=args.allow_runtime_mismatch,
     )
 
 
@@ -374,6 +515,96 @@ def _cmd_validate_env(args: argparse.Namespace) -> int:
     return 0 if result.passed else 1
 
 
+def _add_lambda_connection_args(parser: argparse.ArgumentParser) -> None:
+    parser.add_argument("host", help="Public IP address or DNS name from Lambda Cloud.")
+    parser.add_argument(
+        "--identity",
+        "-i",
+        type=Path,
+        help="Path to the SSH private key selected when the instance was launched.",
+    )
+    parser.add_argument("--user", default="ubuntu")
+    parser.add_argument("--port", type=int, default=22)
+
+
+def _lambda_config(args: argparse.Namespace) -> LambdaSSHConfig:
+    return LambdaSSHConfig(
+        host=args.host,
+        user=args.user,
+        identity=args.identity,
+        port=args.port,
+    )
+
+
+def _cmd_lambda_connect(args: argparse.Namespace) -> int:
+    return connect_lambda(_lambda_config(args), tunnel_port=args.tunnel_port)
+
+
+def _cmd_lambda_check(args: argparse.Namespace) -> int:
+    return check_lambda(_lambda_config(args))
+
+
+def _cmd_lambda_prepare(args: argparse.Namespace) -> None:
+    prepare_lambda(
+        _lambda_config(args),
+        project_dir=args.project,
+        remote_dir=args.remote_dir,
+        jax_cuda=args.jax_cuda,
+        python_command=args.python_command,
+    )
+    print(f"prepared Lambda project: {args.remote_dir or f'/home/{args.user}/simrig'}")
+
+
+def _cmd_lambda_train(args: argparse.Namespace) -> int:
+    result = train_lambda(
+        _lambda_config(args),
+        args.env_name,
+        preset_name=args.preset,
+        remote_dir=args.remote_dir,
+        output=args.output,
+        detach=args.detach,
+        timesteps=args.timesteps,
+        num_envs=args.num_envs,
+        batch_size=args.batch_size,
+    )
+    if result.returncode == 0:
+        mode = "detached run" if result.detached else "run"
+        print(f"Lambda {mode}: {result.output_dir}")
+    return result.returncode
+
+
+def _cmd_lambda_smoke(args: argparse.Namespace) -> int:
+    return smoke_lambda(
+        _lambda_config(args),
+        args.env_name,
+        remote_dir=args.remote_dir,
+        steps=args.steps,
+    )
+
+
+def _cmd_lambda_status(args: argparse.Namespace) -> int:
+    return status_lambda(
+        _lambda_config(args),
+        args.output,
+        remote_dir=args.remote_dir,
+        lines=args.lines,
+    )
+
+
+def _cmd_lambda_fetch(args: argparse.Namespace) -> None:
+    destination = fetch_lambda(
+        _lambda_config(args),
+        args.output,
+        remote_dir=args.remote_dir,
+        local_output=args.local_output,
+    )
+    print(f"downloaded run: {destination}")
+    print(
+        "After verifying these artifacts or persistent storage, terminate the "
+        "Lambda instance in the cloud console to stop compute charges."
+    )
+
+
 def _training_overrides(args: argparse.Namespace) -> dict[str, Any]:
     overrides: dict[str, Any] = {}
     for cli_name, key in (