fix(pi052): avoid dense CE over padded tokens

Select only supervised text and FAST action-code positions before cross-entropy to avoid full-vocabulary loss tensors over padded sequences.

Co-authored-by: Cursor <cursoragent@cursor.com>

src/lerobot/policies/pi052/modeling_pi052.py

@@ -77,13 +77,13 @@ def _shifted_ce(logits: Tensor, labels: Tensor) -> Tensor:
     shift_logits = logits[:, :-1, :].contiguous()
     shift_labels = labels[:, 1:].contiguous().long()
     valid = shift_labels != -100
-    loss = F.cross_entropy(
-        shift_logits.reshape(-1, shift_logits.shape[-1]),
-        shift_labels.reshape(-1),
-        ignore_index=-100,
-        reduction="sum",
+    if not bool(valid.any().item()):
+        return shift_logits.sum() * 0.0
+    return F.cross_entropy(
+        shift_logits[valid],
+        shift_labels[valid],
+        reduction="mean",
     )
-    return loss / valid.sum().clamp(min=1)
 
 
 def _mark_target_span_causal(
@@ -140,13 +140,13 @@ def _fast_ce(
     if predict_actions_t is not None:
         sample_mask = predict_actions_t[:, None].expand_as(shift_valid)
         shift_valid = shift_valid & sample_mask
-    shift_targets = shift_targets.masked_fill(~shift_valid, -100)
+    if not bool(shift_valid.any().item()):
+        return shift_logits.sum() * 0.0
     return F.cross_entropy(
-        shift_logits.reshape(-1, shift_logits.shape[-1]),
-        shift_targets.reshape(-1),
-        ignore_index=-100,
-        reduction="sum",
-    ) / shift_valid.sum().clamp(min=1)
+        shift_logits[shift_valid],
+        shift_targets[shift_valid],
+        reduction="mean",
+    )
 
 
 # ----------------------------------------------------------------------

tests/policies/pi052/test_pi052_attention_masking.py

@@ -37,7 +37,7 @@ import torch
 pytest.importorskip("transformers")
 
 from lerobot.policies.pi05.modeling_pi05 import make_att_2d_masks  # noqa: E402
-from lerobot.policies.pi052.modeling_pi052 import _mark_target_span_causal  # noqa: E402
+from lerobot.policies.pi052.modeling_pi052 import _mark_target_span_causal, _shifted_ce  # noqa: E402
 
 # ---------------------------------------------------------------------------
 # A synthetic PI052 prefix layout: [images, prompt-lang, target-lang]
@@ -136,3 +136,14 @@ def test_unmarked_mask_is_bidirectional_the_bug():
         "raw embed_prefix mask is bidirectional over language — the first "
         "target token can see the last, which is the collapse bug"
     )
+
+
+def test_shifted_ce_returns_zero_when_no_text_positions_are_supervised():
+    logits = torch.randn(2, 4, 8, requires_grad=True)
+    labels = torch.full((2, 4), -100, dtype=torch.long)
+
+    loss = _shifted_ce(logits, labels)
+
+    assert loss.item() == 0
+    loss.backward()
+    assert logits.grad is not None

tests/policies/pi052/test_pi052_fast_action_loss.py

@@ -73,3 +73,15 @@ def test_fast_ce_masks_non_action_samples():
     )
 
     assert torch.allclose(loss, expected)
+
+
+def test_fast_ce_returns_zero_when_no_action_code_positions_are_valid():
+    logits = torch.randn(2, 4, 8, requires_grad=True)
+    action_tokens = torch.tensor([[1, 2, 3, 4], [1, 2, 5, 6]])
+    action_code_mask = torch.zeros_like(action_tokens, dtype=torch.bool)
+
+    loss = _fast_ce(logits, action_tokens, action_code_mask, predict_actions_t=None)
+
+    assert loss.item() == 0
+    loss.backward()
+    assert logits.grad is not None