Deploy BYOD-Llama-3.1-8B full-precision demo

app.py

@@ -75,7 +75,7 @@ def generate(
         permanent_unmask=True,
         confidence_guided=True,
         proportional_unmask=False,
-        early_stopping=False,
+        early_stopping=True,
         # This delays retention of predicted endings; it does not alter their
         # sampling probability. Keep it optional so answers can end naturally.
         confidence_eos_eot_inf=bool(delay_eos_eot),

src/diffusion_lm/inference.py

@@ -985,11 +985,14 @@ def denoise_stream(session: InferenceSession, question: str, system_prompt: str,
             show_eos_tokens=include_pre_remask_prediction,
         )
         last_confidence = float(confidence[block_start:block_end].mean().cpu())
-        # Match the legacy application's criterion: compare complete sampled
-        # answer token sequences before the next iteration's re-masking.  This
-        # includes EOS/padding tokens, so a changing invisible tail does not
-        # count as convergence.
-        last_predictions.append(tuple(ids[answer_start + block_start : answer_start + block_end]))
+        # Compare the visible sampled answer before the next iteration's
+        # re-masking. Tokens after the first EOS are not part of the answer and
+        # must not prevent convergence. Excluding EOS itself still preserves
+        # its position through the tuple length: moving EOS changes the prefix.
+        prediction = ids[answer_start:]
+        if session.tokenizer.eos_token_id in prediction:
+            prediction = prediction[:prediction.index(session.tokenizer.eos_token_id)]
+        last_predictions.append(tuple(prediction))
         if len(last_predictions) > 3:
             last_predictions.pop(0)
         stopped_early = early_stopping and len(last_predictions) == 3 and len(set(last_predictions)) == 1
@@ -1079,7 +1082,7 @@ def denoise_stream(session: InferenceSession, question: str, system_prompt: str,
         if permanent_unmask:
             status += f" · retained {len(retained)} tokens"
         if stopped_early:
-            status += " · block stopped early (same prediction for 3 iterations)"
+            status += " · block stopped early (same answer for 3 iterations)"
         html = render_denoising_step(
             ids,
             confidence.tolist(),