LIT: Fix assertion error when generation stops short of max_length

PiperOrigin-RevId: 681096247

lit_nlp/examples/prompt_debugging/transformers_lms.py

@@ -284,7 +284,7 @@ def _get_batched_outputs(
         attention_mask=encoded_inputs["attention_mask"],
         max_length=self.max_length,
     )
-    ntok_out = self.max_length - encoded_inputs["input_ids"].shape[1]
+    ntok_out = outputs.shape[1] - encoded_inputs["input_ids"].shape[1]
 
     responses = self.tokenizer.batch_decode(
         outputs[:, -ntok_out:], skip_special_tokens=True
@@ -297,11 +297,13 @@ def _get_batched_outputs(
 
       batched_outputs = {
           "embs": embeddings.cpu().to(torch.float),
+          # Input tokens: <int>[batch_size]
           "ntok_in": (
               torch.sum(encoded_inputs["attention_mask"], axis=1)
               .cpu()
               .to(torch.int)
           ),
+          # Output tokens: <int>[batch_size]
           "ntok_out": torch.full(
               (encoded_inputs["input_ids"].shape[0],), ntok_out
           ),
@@ -310,7 +312,9 @@ def _get_batched_outputs(
       embeddings = self.embedding_table(outputs)
       batched_outputs = {
           "embs": embeddings,
+          # Input tokens: <int>[batch_size]
           "ntok_in": tf.reduce_sum(encoded_inputs["attention_mask"], axis=1),
+          # Output tokens: <int>[batch_size]
           "ntok_out": tf.fill(
               [
                   encoded_inputs["input_ids"].shape[0],