fix generation bug #532

src/llmtuner/tuner/ppo/trainer.py

@@ -10,7 +10,7 @@ from trl import PPOTrainer
 from trl.core import LengthSampler
 
 from llmtuner.extras.logging import get_logger
-from llmtuner.extras.misc import AverageMeter, count_parameters, get_logits_processor, get_stopping_criteria
+from llmtuner.extras.misc import AverageMeter, count_parameters, get_logits_processor
 from llmtuner.tuner.core.trainer import PeftTrainer
 from llmtuner.tuner.ppo.utils import cast_layernorm_dtype, replace_model
 
@@ -74,10 +74,9 @@ class PPOPeftTrainer(PPOTrainer, PeftTrainer):
 
         # Keyword arguments for `model.generate`
         gen_kwargs = self.generating_args.to_dict()
-        gen_kwargs["eos_token_id"] = self.tokenizer.eos_token_id
+        gen_kwargs["eos_token_id"] = [self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids
         gen_kwargs["pad_token_id"] = self.tokenizer.pad_token_id
         gen_kwargs["logits_processor"] = get_logits_processor()
-        gen_kwargs["stopping_criteria"] = get_stopping_criteria(self.tokenizer.additional_special_tokens_ids)
 
         length_sampler = LengthSampler(max_target_length // 2, max_target_length)
         unwrapped_model: "AutoModelForCausalLMWithValueHead" = self.accelerator.unwrap_model(self.model)

src/llmtuner/tuner/sft/trainer.py

@@ -50,9 +50,10 @@ class Seq2SeqPeftTrainer(PeftTrainer):
         loss, generated_tokens, labels = super().prediction_step(
             model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys
         )
-        generated_tokens = (
-            generated_tokens[:, max(prompt_len, label_len):] if generated_tokens is not None else None
-        )
+        if generated_tokens is not None:
+            generated_tokens[:, :max(prompt_len, label_len)] = (
+                self.tokenizer.pad_token_id * torch.ones_like(generated_tokens[:, :max(prompt_len, label_len)])
+            )
 
         return (loss, generated_tokens, labels)
 
@@ -72,10 +73,7 @@ class Seq2SeqPeftTrainer(PeftTrainer):
                 assert self.tokenizer.padding_side == "left", "This method only accepts left-padded tensor."
                 pad_token_id = self.tokenizer.pad_token_id
             else:
-                if self.model.config.pad_token_id is not None:
-                    pad_token_id = self.model.config.pad_token_id
-                else:
-                    raise ValueError("Pad_token_id must be set in the configuration of the model.")
+                raise ValueError("PAD token is required.")
 
         padded_tensor = pad_token_id * torch.ones_like(tgt_tensor)
         padded_tensor[:, -src_tensor.shape[-1]:] = src_tensor # adopt left-padding

src/llmtuner/tuner/sft/workflow.py

@@ -5,7 +5,7 @@ from transformers import DataCollatorForSeq2Seq
 
 from llmtuner.dsets import get_dataset, preprocess_dataset, split_dataset
 from llmtuner.extras.constants import IGNORE_INDEX
-from llmtuner.extras.misc import get_logits_processor, get_stopping_criteria
+from llmtuner.extras.misc import get_logits_processor
 from llmtuner.extras.ploting import plot_loss
 from llmtuner.tuner.core import load_model_and_tokenizer
 from llmtuner.tuner.sft.metric import ComputeMetrics
@@ -52,10 +52,9 @@ def run_sft(
 
     # Keyword arguments for `model.generate`
     gen_kwargs = generating_args.to_dict()
-    gen_kwargs["eos_token_id"] = tokenizer.eos_token_id
+    gen_kwargs["eos_token_id"] = [tokenizer.eos_token_id] + tokenizer.additional_special_tokens_ids
     gen_kwargs["pad_token_id"] = tokenizer.pad_token_id
     gen_kwargs["logits_processor"] = get_logits_processor()
-    gen_kwargs["stopping_criteria"] = get_stopping_criteria(tokenizer.additional_special_tokens_ids)
 
     # Training
     if training_args.do_train: