[Feature] batch_size, reward, done, attention_key in LLMEnv

ghstack-source-id: 90e0ff9
Pull Request resolved: #2824

Author: vmoens

test/test_env.py

@@ -4861,6 +4861,81 @@ def policy(td):
                     r_reset[0, 0]["observation"] != r_reset[0, 1]["observation"]
                 ).any()
 
+    @pytest.mark.parametrize(
+        "str2str,stack_method",
+        [
+            [True, None],
+            [False, "as_padded_tensor"],
+        ],
+    )
+    @pytest.mark.parametrize("batched", [True])
+    @pytest.mark.parametrize("device", [None])
+    @pytest.mark.parametrize("batch_size", [4])
+    @pytest.mark.parametrize("repeats", [3])
+    @pytest.mark.parametrize(
+        "assign_reward,assign_done", [[True, False], [True, True], [False, True]]
+    )
+    def test_done_and_reward(
+        self,
+        str2str,
+        batched,
+        stack_method,
+        device,
+        batch_size,
+        repeats,
+        assign_reward,
+        assign_done,
+    ):
+        with pytest.raises(
+            ValueError, match="str2str"
+        ) if str2str else contextlib.nullcontext():
+            if str2str:
+                kwargs = {
+                    "dataloader": self.DummyDataLoader(batch_size=batch_size),
+                    "data_keys": ["observation"],
+                    "example_data": "a string!",
+                    "repeats": repeats,
+                    "assign_reward": assign_reward,
+                    "assign_done": assign_done,
+                }
+            else:
+                if stack_method is None:
+                    stack_method = as_padded_tensor
+                kwargs = {
+                    "dataloader": self.DummyTensorDataLoader(
+                        padding=True, batch_size=batch_size
+                    ),
+                    "data_keys": ["observation"],
+                    "data_specs": [Unbounded(shape=(-1,), dtype=torch.int64)],
+                    "stack_method": stack_method,
+                    "repeats": repeats,
+                    "assign_reward": assign_reward,
+                    "assign_done": assign_done,
+                }
+            kwargs.update({"str2str": str2str, "device": device})
+            env = LLMEnv.from_dataloader(**kwargs)
+            # We want to make sure that transforms that rely on the done state work appropriately
+            env.append_transform(StepCounter(max_steps=10))
+
+            def policy(td):
+                td["action"] = torch.ones(
+                    td.shape + (torch.randint(10, (1,)).item(),), dtype=torch.int64
+                )
+                return td
+
+            if batched:
+                r = env.rollout(
+                    100,
+                    policy,
+                    tensordict=TensorDict(batch_size=[3]),
+                    break_when_any_done=False,
+                )
+            else:
+                r = env.rollout(100, policy, break_when_any_done=False)
+            if assign_done:
+                assert "terminated" in r
+                assert "done" in r
+
 
 if __name__ == "__main__":
     args, unknown = argparse.ArgumentParser().parse_known_args()

torchrl/data/map/tdstorage.py

@@ -128,7 +128,6 @@ def __init__(
         self.in_keys = query_module.in_keys
         if out_keys is not None:
             self.out_keys = out_keys
-            assert not self._has_lazy_out_keys()
 
         self.query_module = query_module
         self.index_key = query_module.index_key

torchrl/data/postprocs/postprocs.py

@@ -11,8 +11,6 @@
 from tensordict.utils import expand_right
 from torch import nn
 
-from torchrl.objectives.value.functional import reward2go
-
 
 def _get_reward(
     gamma: float,
@@ -367,13 +365,16 @@ def __init__(
         time_dim: int = 2,
         discount: float = 1.0,
     ):
+        from torchrl.objectives.value.functional import reward2go
+
         super().__init__()
         self.in_keys = [unravel_key(reward_key), unravel_key(done_key)]
         if reward_key_out is None:
             reward_key_out = reward_key
         self.out_keys = [unravel_key(reward_key_out)]
         self.time_dim = time_dim
         self.discount = discount
+        self.reward2go = reward2go
 
     def forward(self, tensordict):
         # Get done
@@ -385,6 +386,6 @@ def forward(self, tensordict):
                 f"reward and done state are expected to have the same shape. Got reard.shape={reward.shape} "
                 f"and done.shape={done.shape}."
             )
-        reward = reward2go(reward, done, time_dim=-2, gamma=self.discount)
+        reward = self.reward2go(reward, done, time_dim=-2, gamma=self.discount)
         tensordict.set(("next", self.out_keys[0]), reward)
         return tensordict

torchrl/envs/common.py

@@ -2788,7 +2788,11 @@ def _reset_check_done(self, tensordict, tensordict_reset):
             if reset_value is not None:
                 for done_key in done_key_group:
                     done_val = tensordict_reset.get(done_key)
-                    if done_val[reset_value].any() and not self._allow_done_after_reset:
+                    if (
+                        done_val.any()
+                        and done_val[reset_value].any()
+                        and not self._allow_done_after_reset
+                    ):
                         raise RuntimeError(
                             f"Env done entry '{done_key}' was (partially) True after reset on specified '_reset' dimensions. This is not allowed."
                         )
@@ -3588,7 +3592,7 @@ def maybe_reset(self, tensordict: TensorDictBase) -> TensorDictBase:
         """
         any_done = self.any_done(tensordict)
         if any_done:
-            return self.reset(tensordict, select_reset_only=True)
+            tensordict = self.reset(tensordict, select_reset_only=True)
         return tensordict
 
     def empty_cache(self):