[LoRA] Change lora_tokenizers capacity

vllm/engine/llm_engine.py

@@ -619,7 +619,7 @@ def _init_tokenizer(self) -> BaseTokenizerGroup:
             model_config=self.model_config,
             scheduler_config=self.scheduler_config,
             parallel_config=self.parallel_config,
-            enable_lora=bool(self.lora_config))
+            lora_config=self.lora_config)
 
     def _verify_args(self) -> None:
         self.model_config.verify_with_parallel_config(self.parallel_config)

vllm/engine/multiprocessing/client.py

@@ -93,8 +93,7 @@ def __init__(self, ipc_path: str, engine_config: VllmConfig,
             model_config=self.model_config,
             scheduler_config=engine_config.scheduler_config,
             parallel_config=engine_config.parallel_config,
-            enable_lora=bool(engine_config.lora_config),
-        )
+            lora_config=engine_config.lora_config)
         self.input_preprocessor = InputPreprocessor(self.model_config,
                                                     self.tokenizer)
 

vllm/transformers_utils/tokenizer_group/__init__.py

@@ -1,7 +1,7 @@
 from typing import Optional, Type
 
-from vllm.config import (ModelConfig, ParallelConfig, SchedulerConfig,
-                         TokenizerPoolConfig)
+from vllm.config import (LoRAConfig, ModelConfig, ParallelConfig,
+                         SchedulerConfig, TokenizerPoolConfig)
 from vllm.executor.ray_utils import ray
 
 from .base_tokenizer_group import AnyTokenizer, BaseTokenizerGroup
@@ -16,10 +16,11 @@
 def init_tokenizer_from_configs(model_config: ModelConfig,
                                 scheduler_config: SchedulerConfig,
                                 parallel_config: ParallelConfig,
-                                enable_lora: bool):
+                                lora_config: LoRAConfig):
     init_kwargs = dict(tokenizer_id=model_config.tokenizer,
-                       enable_lora=enable_lora,
+                       enable_lora=bool(lora_config),
                        max_num_seqs=scheduler_config.max_num_seqs,
+                       max_loras=lora_config.max_loras if lora_config else 0,
                        max_input_length=None,
                        tokenizer_mode=model_config.tokenizer_mode,
                        trust_remote_code=model_config.trust_remote_code,

vllm/transformers_utils/tokenizer_group/tokenizer_group.py

@@ -21,8 +21,9 @@ def __init__(self, tokenizer_id: str, enable_lora: bool, max_num_seqs: int,
         self.enable_lora = enable_lora
         self.max_input_length = max_input_length
         self.tokenizer = get_tokenizer(self.tokenizer_id, **tokenizer_config)
+        max_loras = tokenizer_config.get("max_loras", 0)
         self.lora_tokenizers = LRUCache[AnyTokenizer](
-            capacity=max_num_seqs if enable_lora else 0)
+            capacity=max(max_loras, max_num_seqs) if enable_lora else 0)
 
     @classmethod
     def from_config(cls, tokenizer_pool_config: Optional[TokenizerPoolConfig],

vllm/v1/engine/async_llm.py

@@ -51,7 +51,7 @@ def __init__(
             model_config=vllm_config.model_config,
             scheduler_config=vllm_config.scheduler_config,
             parallel_config=vllm_config.parallel_config,
-            enable_lora=bool(vllm_config.lora_config))
+            lora_config=vllm_config.lora_config)
         self.tokenizer.ping()
 
         # Request streams (map of request_id -> AsyncStream).

vllm/v1/engine/llm_engine.py

@@ -46,7 +46,7 @@ def __init__(
             model_config=vllm_config.model_config,
             scheduler_config=vllm_config.scheduler_config,
             parallel_config=vllm_config.parallel_config,
-            enable_lora=bool(vllm_config.lora_config))
+            lora_config=vllm_config.lora_config)
         self.tokenizer.ping()
 
         # Processor (convert Inputs --> EngineCoreRequests)