Improve observability during LLM inference (#3536)

ludwig-ai · Aug 15, 2023 · 756ddf5 · 756ddf5
1 parent 5ba3494
commit 756ddf5
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/ludwig/models/llm.py b/ludwig/models/llm.py
@@ -398,7 +398,7 @@ def generate(
                 input_ids_sample_no_padding = remove_left_padding(input_ids_sample, self.tokenizer)
                 logger.info(
                     "Decoded text inputs for the first example in batch: "
-                    f"{self.tokenizer.decode(input_ids_sample_no_padding[0])}"
+                    f"{self.tokenizer.decode(input_ids_sample_no_padding[0], skip_special_tokens=True)}"
                 )
 
                 if input_ids_sample_no_padding.shape[1] > self.max_input_length:
@@ -422,6 +422,10 @@ def generate(
                         return_dict_in_generate=True,
                         output_scores=True,
                     )
+                    logger.info(
+                        "Decoded generated output for the first example in batch: "
+                        f"{self.tokenizer.batch_decode(model_outputs.sequences, skip_special_tokens=True)[0]}"
+                    )
 
                 sequences_list.append(model_outputs.sequences[0])