[Test] Add PEnv tests for devices

ghstack-source-id: ad6dcce87bcde2a22f619b9742324073eb67cd5b
Pull Request resolved: #2843

Author: vmoens

test/_utils_internal.py

@@ -267,21 +267,23 @@ def _make_envs(
     transformed_in,
     transformed_out,
     N,
-    device="cpu",
+    p_env_device=None,
+    env_device=None,
+    # device="cpu",
     kwargs=None,
     local_mp_ctx=mp_ctx,
 ):
     torch.manual_seed(0)
     if not transformed_in:
 
         def create_env_fn():
-            return GymEnv(env_name, frame_skip=frame_skip, device=device)
+            return GymEnv(env_name, frame_skip=frame_skip, device=env_device)
 
     else:
         if env_name == PONG_VERSIONED():
 
             def create_env_fn():
-                base_env = GymEnv(env_name, frame_skip=frame_skip, device=device)
+                base_env = GymEnv(env_name, frame_skip=frame_skip, device=env_device)
                 in_keys = list(base_env.observation_spec.keys(True, True))[:1]
                 return TransformedEnv(
                     base_env,
@@ -292,7 +294,7 @@ def create_env_fn():
 
             def create_env_fn():
 
-                base_env = GymEnv(env_name, frame_skip=frame_skip, device=device)
+                base_env = GymEnv(env_name, frame_skip=frame_skip, device=env_device)
                 in_keys = list(base_env.observation_spec.keys(True, True))[:1]
 
                 return TransformedEnv(
@@ -305,9 +307,15 @@ def create_env_fn():
 
     env0 = create_env_fn()
     env_parallel = ParallelEnv(
-        N, create_env_fn, create_env_kwargs=kwargs, mp_start_method=local_mp_ctx
+        N,
+        create_env_fn,
+        create_env_kwargs=kwargs,
+        mp_start_method=local_mp_ctx,
+        device=p_env_device,
+    )
+    env_serial = SerialEnv(
+        N, create_env_fn, create_env_kwargs=kwargs, device=p_env_device
     )
-    env_serial = SerialEnv(N, create_env_fn, create_env_kwargs=kwargs)
 
     for key in env0.observation_spec.keys(True, True):
         obs_key = key

test/test_env.py

@@ -1464,12 +1464,29 @@ def make_env():
         "transformed_in,transformed_out", [[True, True], [False, False]]
     )  # 1226: effociency
     @pytest.mark.parametrize("static_seed", [False, True])
+    @pytest.mark.parametrize("penv_device", ["cpu", None])
+    @pytest.mark.parametrize("env_device", ["cpu", None])
+    @pytest.mark.parametrize("bwad", [True, False])
     def test_parallel_env_seed(
-        self, env_name, frame_skip, transformed_in, transformed_out, static_seed
+        self,
+        env_name,
+        frame_skip,
+        transformed_in,
+        transformed_out,
+        static_seed,
+        penv_device,
+        env_device,
+        bwad,
     ):
         env_name = env_name()
         env_parallel, env_serial, _, _ = _make_envs(
-            env_name, frame_skip, transformed_in, transformed_out, 5
+            env_name,
+            frame_skip,
+            transformed_in,
+            transformed_out,
+            5,
+            p_env_device=penv_device,
+            env_device=env_device,
         )
         try:
             out_seed_serial = env_serial.set_seed(0, static_seed=static_seed)
@@ -1479,7 +1496,10 @@ def test_parallel_env_seed(
             torch.manual_seed(0)
 
             td_serial = env_serial.rollout(
-                max_steps=10, auto_reset=False, tensordict=td0_serial
+                max_steps=10,
+                auto_reset=False,
+                tensordict=td0_serial,
+                break_when_any_done=bwad,
             ).contiguous()
             key = "pixels" if "pixels" in td_serial.keys() else "observation"
             torch.testing.assert_close(
@@ -1494,7 +1514,10 @@ def test_parallel_env_seed(
             torch.manual_seed(0)
             assert out_seed_parallel == out_seed_serial
             td_parallel = env_parallel.rollout(
-                max_steps=10, auto_reset=False, tensordict=td0_parallel
+                max_steps=10,
+                auto_reset=False,
+                tensordict=td0_parallel,
+                break_when_any_done=bwad,
             ).contiguous()
             torch.testing.assert_close(
                 td_parallel[:, :-1].get(("next", key)), td_parallel[:, 1:].get(key)
@@ -1670,7 +1693,7 @@ def test_parallel_env_device(
             frame_skip,
             transformed_in=transformed_in,
             transformed_out=transformed_out,
-            device=device,
+            env_device=device,
             N=N,
             local_mp_ctx="spawn",
         )