[Performance] Avoid parent env construction in ParallelEnv startup

benchmarks/test_envs_benchmark.py

@@ -3,6 +3,8 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 import argparse
+import time
+from functools import partial
 
 import pytest
 import torch
@@ -12,6 +14,30 @@
 from torchrl.envs.libs.dm_control import DMControlEnv
 from torchrl.envs.libs.libero import _has_libero, LiberoEnv
 from torchrl.envs.transforms.functional import cat_frames
+from torchrl.testing.mocking_classes import CountingEnv
+
+
+class _SlowStartEnv(CountingEnv):
+    def __init__(self, delay: float = 0.1):
+        time.sleep(delay)
+        super().__init__()
+
+
+def _make_slow_start_env(delay: float) -> _SlowStartEnv:
+    return _SlowStartEnv(delay)
+
+
+def _run_parallel_cold_start(metadata_from_workers: bool) -> None:
+    env = ParallelEnv(
+        4,
+        partial(_make_slow_start_env, 0.1),
+        mp_start_method="spawn",
+        metadata_from_workers=metadata_from_workers,
+    )
+    try:
+        env.reset()
+    finally:
+        env.close(raise_if_closed=False)
 
 
 def make_simple_env():
@@ -111,6 +137,16 @@ def test_parallel(benchmark):
     benchmark(execute_env, c)
 
 
+@pytest.mark.parametrize("metadata_from_workers", [False, True])
+def test_parallel_cold_start(benchmark, metadata_from_workers):
+    benchmark.pedantic(
+        _run_parallel_cold_start,
+        args=(metadata_from_workers,),
+        rounds=3,
+        iterations=1,
+    )
+
+
 @pytest.mark.skipif(not _has_libero, reason="libero not found")
 def test_libero(benchmark):
     # raw simulation + render throughput of the LIBERO adapter (steps/s)

docs/source/reference/envs_vectorized.rst

@@ -24,6 +24,28 @@ It is important that your environment specs match the input and output that it s
 :class:`ParallelEnv` will create buffers from these specs to communicate with the spawn processes.
 Check the :func:`~torchrl.envs.check_env_specs` method for a sanity check.
 
+By default, :class:`ParallelEnv` obtains those specs by creating temporary
+environments in the parent process. For environments that are expensive or
+unsafe to construct in the parent, ``metadata_from_workers=True`` obtains the
+metadata from the real, long-lived worker environments instead. The workers
+start during :class:`ParallelEnv` construction, their schemas are validated,
+and no parent-side environment is created. This opt-in mode currently requires
+``use_buffers=False`` (which is selected automatically when the argument is
+unset); passing ``use_buffers=True`` raises an error. A common factory can still
+receive per-worker settings through ``create_env_kwargs``:
+
+.. code-block:: python
+
+    from functools import partial
+
+    make_env = partial(GymEnv, "Pendulum-v1")
+    env = ParallelEnv(
+        4,
+        make_env,
+        create_env_kwargs=[{"frame_skip": i + 1} for i in range(4)],
+        metadata_from_workers=True,
+    )
+
 .. code-block::
    :caption: Parallel environment
 

sota-implementations/vla_grpo/README.md

@@ -243,6 +243,30 @@ workers plus the evaluator share one process policy server.
 python sota-implementations/vla_grpo/vla-grpo.py --config-name vla_grpo_libero
 ```
 
+The LIBERO recipe enables worker-originated environment metadata. This keeps
+MuJoCo/EGL construction inside the long-lived rollout workers and avoids a
+temporary parent-side environment wave. The nested startup path can be measured
+without loading the VLA policy:
+
+```bash
+python sota-implementations/vla_grpo/bench_libero_startup.py \
+  --mode legacy-parent --inner-start-method spawn \
+  --output-dir /root/artifacts/libero-startup/legacy-spawn
+python sota-implementations/vla_grpo/bench_libero_startup.py \
+  --mode worker-metadata --inner-start-method spawn \
+  --output-dir /root/artifacts/libero-startup/worker-spawn
+python sota-implementations/vla_grpo/bench_libero_startup.py \
+  --mode worker-metadata --inner-start-method forkserver \
+  --output-dir /root/artifacts/libero-startup/worker-forkserver
+```
+
+Outer subcollector processes always use ``spawn``. The script permits
+``forkserver`` or diagnostic ``fork`` only in worker-metadata mode, after the
+subcollector parent has been kept free of EGL contexts. Each output directory
+contains the exact command, per-construction marker files, and a JSON summary
+covering nested environment readiness, first-step latency, peak process count,
+and shutdown cleanup.
+
 Rollout clients request random sampling; eval and video clients request
 deterministic decoding from the same server, so rollout and eval are synced by
 one explicit TensorDict weight update after each optimizer step. The replay