2 files changed, 16 insertions, 13 deletions
diff --git a/scripts/eval.py b/scripts/eval.py
index bc471dc..33314cf 100644
--- a/scripts/eval.py
+++ b/scripts/eval.py
@@ -88,8 +88,8 @@ def main():
             A_soft = predictor(raw_texts, tau=tau, mode="eval_soft")
             logits_soft = olmo_wrapper(olmo_ids, A_soft)
             nll_soft = F.cross_entropy(
-                logits_soft[:, :-1].contiguous().view(-1, vocab_size),
-                olmo_labels[:, 1:].contiguous().view(-1),
+                logits_soft.contiguous().view(-1, vocab_size),
+                olmo_labels.contiguous().view(-1),
             )
             nll_soft_sum += nll_soft.item()
 
@@ -97,8 +97,8 @@ def main():
             A_hard = predictor(raw_texts, tau=tau, mode="eval_hard")
             logits_hard = olmo_wrapper(olmo_ids, A_hard)
             nll_hard = F.cross_entropy(
-                logits_hard[:, :-1].contiguous().view(-1, vocab_size),
-                olmo_labels[:, 1:].contiguous().view(-1),
+                logits_hard.contiguous().view(-1, vocab_size),
+                olmo_labels.contiguous().view(-1),
             )
             nll_hard_sum += nll_hard.item()
 
@@ -106,8 +106,8 @@ def main():
             A_ones = create_all_ones_A(olmo_ids.shape[0]).to(device)
             logits_base = olmo_wrapper(olmo_ids, A_ones)
             nll_base = F.cross_entropy(
-                logits_base[:, :-1].contiguous().view(-1, vocab_size),
-                olmo_labels[:, 1:].contiguous().view(-1),
+                logits_base.contiguous().view(-1, vocab_size),
+                olmo_labels.contiguous().view(-1),
             )
             nll_baseline_sum += nll_base.item()
 
diff --git a/scripts/sanity_check.py b/scripts/sanity_check.py
index f30bd58..9745482 100644
--- a/scripts/sanity_check.py
+++ b/scripts/sanity_check.py
@@ -48,11 +48,14 @@ def get_test_batch(tokenizer, seq_len: int = 64, device: str = "cpu"):
 
 def compute_dagformer_nll(wrapper: DAGFormerOLMo, input_ids: torch.Tensor,
                           labels: torch.Tensor, A: torch.Tensor) -> torch.Tensor:
-    """Compute NLL using DAGFormer modified forward."""
+    """Compute NLL using DAGFormer modified forward.
+
+    labels is already shifted (chunk[1:seq_len+1]), no additional shift needed.
+    """
     logits = wrapper.forward(input_ids, A)
     nll = F.cross_entropy(
-        logits[:, :-1].contiguous().view(-1, logits.size(-1)),
-        labels[:, 1:].contiguous().view(-1),
+        logits.contiguous().view(-1, logits.size(-1)),
+        labels.contiguous().view(-1),
     )
     return nll
 
@@ -132,8 +135,8 @@ def check_4_gradient_flow(wrapper, tokenizer, device):
 
     logits = wrapper.forward(input_ids, A)
     nll = F.cross_entropy(
-        logits[:, :-1].contiguous().view(-1, logits.size(-1)),
-        labels[:, 1:].contiguous().view(-1),
+        logits.contiguous().view(-1, logits.size(-1)),
+        labels.contiguous().view(-1),
     )
     nll.backward()
 
@@ -176,8 +179,8 @@ def check_5_normalization_smoke(wrapper_factory, tokenizer, device):
                 logits = wrapper.forward(input_ids, A)
             is_finite = torch.isfinite(logits).all().item()
             nll = F.cross_entropy(
-                logits[:, :-1].contiguous().view(-1, logits.size(-1)),
-                labels[:, 1:].contiguous().view(-1),
+                logits.contiguous().view(-1, logits.size(-1)),
+                labels.contiguous().view(-1),
             ).item()
             print(f"  {method:12s}: NLL={nll:.4f}, finite={is_finite}")
             if not is_finite: