fix max seq len (#489)

Author: LiuXiaoxuanPKU

vllm/config.py

@@ -204,10 +204,10 @@ class SchedulerConfig:
     """
 
     def __init__(self, max_num_batched_tokens: int, max_num_seqs: int,
-                 max_seq_len: int) -> None:
+                 max_model_len: int) -> None:
         self.max_num_batched_tokens = max_num_batched_tokens
         self.max_num_seqs = max_num_seqs
-        self.max_seq_len = max_seq_len
+        self.max_model_len = max_model_len
 
 
 _STR_DTYPE_TO_TORCH_DTYPE = {

vllm/core/scheduler.py

@@ -190,7 +190,9 @@ def _schedule(
                     break
 
                 num_prompt_tokens = seq_group.get_seqs()[0].get_len()
-                if num_prompt_tokens > self.scheduler_config.max_seq_len:
+                if num_prompt_tokens > min(
+                        self.scheduler_config.max_model_len,
+                        self.scheduler_config.max_num_batched_tokens):
                     logger.warning(
                         f"Input prompt ({num_prompt_tokens} tokens) is too long"
                         " and exceeds limit of "

vllm/engine/arg_utils.py

@@ -155,11 +155,10 @@ def create_engine_configs(
         parallel_config = ParallelConfig(self.pipeline_parallel_size,
                                          self.tensor_parallel_size,
                                          self.worker_use_ray)
-        model_max_len = getattr(model_config.hf_config,
+        max_model_len = getattr(model_config.hf_config,
                                 'max_position_embeddings', float('inf'))
-        max_seq_len = min(self.max_num_batched_tokens, model_max_len)
         scheduler_config = SchedulerConfig(self.max_num_batched_tokens,
-                                           self.max_num_seqs, max_seq_len)
+                                           self.max_num_seqs, max_model_len)
         return model_config, cache_config, parallel_config, scheduler_config
 
 

vllm/engine/llm_engine.py

@@ -300,8 +300,7 @@ def _stop_sequences(self, seq_groups: List[SequenceGroup]) -> None:
                     continue
 
                 # Check if the sequence has reached max_seq_len.
-                if (seq.get_len() >
-                        self.scheduler.scheduler_config.max_seq_len):
+                if seq.get_len() > self.scheduler_config.max_model_len:
                     self.scheduler.free_seq(
                         seq, SequenceStatus.FINISHED_LENGTH_CAPPED)
                     continue