refactor(config): Move device & amp args to PreTrainedConfig (#812)

Co-authored-by: Simon Alibert <75076266+aliberts@users.noreply.github.com>

lerobot/common/policies/factory.py

@@ -16,7 +16,6 @@
 
 import logging
 
-import torch
 from torch import nn
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
@@ -76,7 +75,6 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
 
 def make_policy(
     cfg: PreTrainedConfig,
-    device: str | torch.device,
     ds_meta: LeRobotDatasetMetadata | None = None,
     env_cfg: EnvConfig | None = None,
 ) -> PreTrainedPolicy:
@@ -88,7 +86,6 @@ def make_policy(
     Args:
         cfg (PreTrainedConfig): The config of the policy to make. If `pretrained_path` is set, the policy will
             be loaded with the weights from that path.
-        device (str): the device to load the policy onto.
         ds_meta (LeRobotDatasetMetadata | None, optional): Dataset metadata to take input/output shapes and
             statistics to use for (un)normalization of inputs/outputs in the policy. Defaults to None.
         env_cfg (EnvConfig | None, optional): The config of a gym environment to parse features from. Must be
@@ -96,7 +93,7 @@ def make_policy(
 
     Raises:
         ValueError: Either ds_meta or env and env_cfg must be provided.
-        NotImplementedError: if the policy.type is 'vqbet' and the device 'mps' (due to an incompatibility)
+        NotImplementedError: if the policy.type is 'vqbet' and the policy device 'mps' (due to an incompatibility)
 
     Returns:
         PreTrainedPolicy: _description_
@@ -111,7 +108,7 @@ def make_policy(
     # https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment
     # variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be
     # slower than running natively on MPS.
-    if cfg.type == "vqbet" and str(device) == "mps":
+    if cfg.type == "vqbet" and cfg.device == "mps":
         raise NotImplementedError(
             "Current implementation of VQBeT does not support `mps` backend. "
             "Please use `cpu` or `cuda` backend."
@@ -145,7 +142,7 @@ def make_policy(
         # Make a fresh policy.
         policy = policy_cls(**kwargs)
 
-    policy.to(device)
+    policy.to(cfg.device)
     assert isinstance(policy, nn.Module)
 
     # policy = torch.compile(policy, mode="reduce-overhead")

lerobot/common/policies/pi0/configuration_pi0.py

@@ -90,6 +90,7 @@ class PI0Config(PreTrainedConfig):
     def __post_init__(self):
         super().__post_init__()
 
+        # TODO(Steven): Validate device and amp? in all policy configs?
         """Input validation (not exhaustive)."""
         if self.n_action_steps > self.chunk_size:
             raise ValueError(

lerobot/common/policies/pi0/conversion_scripts/benchmark.py

@@ -45,7 +45,7 @@ def main():
 
     cfg = PreTrainedConfig.from_pretrained(ckpt_torch_dir)
     cfg.pretrained_path = ckpt_torch_dir
-    policy = make_policy(cfg, device, ds_meta=dataset.meta)
+    policy = make_policy(cfg, ds_meta=dataset.meta)
 
     # policy = torch.compile(policy, mode="reduce-overhead")
 

lerobot/common/policies/pi0/conversion_scripts/compare_with_jax.py

@@ -101,7 +101,7 @@ def main():
 
     cfg = PreTrainedConfig.from_pretrained(ckpt_torch_dir)
     cfg.pretrained_path = ckpt_torch_dir
-    policy = make_policy(cfg, device, dataset_meta)
+    policy = make_policy(cfg, dataset_meta)
 
     # loss_dict = policy.forward(batch, noise=noise, time=time_beta)
     # loss_dict["loss"].backward()

lerobot/common/policies/pretrained.py

@@ -86,7 +86,6 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
         cache_dir: str | Path | None = None,
         local_files_only: bool = False,
         revision: str | None = None,
-        map_location: str = "cpu",
         strict: bool = False,
         **kwargs,
     ) -> T:
@@ -111,7 +110,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
         if os.path.isdir(model_id):
             print("Loading weights from local directory")
             model_file = os.path.join(model_id, SAFETENSORS_SINGLE_FILE)
-            policy = cls._load_as_safetensor(instance, model_file, map_location, strict)
+            policy = cls._load_as_safetensor(instance, model_file, config.device, strict)
         else:
             try:
                 model_file = hf_hub_download(
@@ -125,13 +124,13 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
                     token=token,
                     local_files_only=local_files_only,
                 )
-                policy = cls._load_as_safetensor(instance, model_file, map_location, strict)
+                policy = cls._load_as_safetensor(instance, model_file, config.device, strict)
             except HfHubHTTPError as e:
                 raise FileNotFoundError(
                     f"{SAFETENSORS_SINGLE_FILE} not found on the HuggingFace Hub in {model_id}"
                 ) from e
 
-        policy.to(map_location)
+        policy.to(config.device)
         policy.eval()
         return policy