fix(g05): match author inference precision

src/lerobot/policies/g05/configuration_g05.py

@@ -168,6 +168,7 @@ class G05Config(PreTrainedConfig):
     discrete_action: bool = True
     continuous_action: bool = False
     return_continuous_action: bool = False
+    model_weights_to_bf16: bool = True
 
     policy_action_dim: int = 20
     policy_state_dim: int = 20

src/lerobot/policies/g05/modeling_g05.py

@@ -165,10 +165,25 @@ class G05Policy(PreTrainedPolicy):
         if callable(reset):
             reset()
 
+    def _apply_author_inference_precision(self) -> None:
+        """Match the released serving path's BF16 weights with declared FP32 islands."""
+
+        self.backend.to(dtype=torch.bfloat16)
+        apply_fp32_params = getattr(self.backend, "apply_fp32_params", None)
+        if callable(apply_fp32_params):
+            apply_fp32_params()
+
     def to(self, *args, **kwargs) -> G05Policy:
-        """Move the author ActionCodec sidecar along with the policy module."""
+        """Apply author inference precision and move the ActionCodec sidecar."""
 
         result = super().to(*args, **kwargs)
+        explicit_dtype = "dtype" in kwargs or any(isinstance(arg, torch.dtype | Tensor) for arg in args)
+        if (
+            self.config.model_weights_to_bf16
+            and not explicit_dtype
+            and next(self.backend.parameters()).device.type == "cuda"
+        ):
+            self._apply_author_inference_precision()
         action_tokenizer = getattr(self.backend, "action_tokenizer", None)
         move_tokenizer = getattr(action_tokenizer, "to", None)
         if callable(move_tokenizer):

tests/policies/g05/test_g05.py

@@ -685,6 +685,25 @@ def test_policy_to_moves_non_module_action_tokenizer_sidecar():
     assert backend.action_tokenizer.device == next(policy.parameters()).device
 
 
+def test_author_inference_precision_preserves_declared_fp32_parameters():
+    class MixedPrecisionBackend(TinyG05Backend):
+        def __init__(self):
+            super().__init__()
+            self.bulk_weight = nn.Parameter(torch.ones(2))
+            self.precision_weight = nn.Parameter(torch.ones(2))
+
+        def apply_fp32_params(self):
+            self.precision_weight.data = self.precision_weight.data.float()
+
+    backend = MixedPrecisionBackend()
+    policy = G05Policy(_config(), backend=backend)
+
+    policy._apply_author_inference_precision()
+
+    assert backend.bulk_weight.dtype is torch.bfloat16
+    assert backend.precision_weight.dtype is torch.float32
+
+
 def test_batch_two_preserves_each_raw_task_and_every_camera_slot():
     backend = TinyG05Backend()
     policy = G05Policy(_config(), backend=backend)