[BugFix] Fix get tokenizer when using ray (vllm-project#3301)

Author: esmeetu

tests/async_engine/test_async_llm_engine.py

@@ -89,3 +89,6 @@ async def test_new_requests_event():
     await asyncio.sleep(0.01)
     assert engine.engine.add_request_calls == 3
     assert engine.engine.step_calls == old_step_calls + 1
+
+    engine = MockAsyncLLMEngine(worker_use_ray=True, engine_use_ray=True)
+    assert engine.get_tokenizer() is not None

vllm/engine/async_llm_engine.py

@@ -5,6 +5,8 @@
 from typing import (Any, Dict, Iterable, List, Optional, Set, Tuple, Type,
                     Union, AsyncIterator, Callable)
 
+from transformers import PreTrainedTokenizer
+
 from vllm.lora.request import LoRARequest
 from vllm.config import ModelConfig
 from vllm.engine.arg_utils import AsyncEngineArgs
@@ -372,8 +374,11 @@ def _error_callback(self, exc: Exception) -> None:
         self.set_errored(exc)
         self._request_tracker.propagate_exception(exc)
 
-    def get_tokenizer(self):
-        return self.engine.tokenizer.tokenizer
+    async def get_tokenizer(self) -> "PreTrainedTokenizer":
+        if self.engine_use_ray:
+            return await self.engine.get_tokenizer.remote()
+        else:
+            return self.engine.get_tokenizer()
 
     def start_background_loop(self) -> None:
         """Start the background loop."""

vllm/engine/llm_engine.py

@@ -7,6 +7,8 @@
 from typing import (TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Tuple,
                     Union)
 
+from transformers import PreTrainedTokenizer
+
 import vllm
 from vllm.lora.request import LoRARequest
 from vllm.config import (CacheConfig, DeviceConfig, ModelConfig,
@@ -163,7 +165,11 @@ def __reduce__(self):
         # the closure used to initialize Ray worker actors
         raise RuntimeError("LLMEngine should not be pickled!")
 
-    def get_tokenizer_for_seq(self, sequence: Sequence):
+    def get_tokenizer(self) -> "PreTrainedTokenizer":
+        return self.tokenizer.get_lora_tokenizer()
+
+    def get_tokenizer_for_seq(self,
+                              sequence: Sequence) -> "PreTrainedTokenizer":
         return self.tokenizer.get_lora_tokenizer(sequence.lora_request)
 
     def _dispatch_worker(self):

vllm/entrypoints/openai/serving_chat.py

@@ -65,7 +65,7 @@ async def create_chat_completion(
             lora_request = self._maybe_get_lora(request)
             guided_decode_logits_processor = (
                 await get_guided_decoding_logits_processor(
-                    request, self.engine.get_tokenizer()))
+                    request, await self.engine.get_tokenizer()))
             if guided_decode_logits_processor:
                 if sampling_params.logits_processors is None:
                     sampling_params.logits_processors = []

vllm/entrypoints/openai/serving_completion.py

@@ -126,7 +126,7 @@ async def create_completion(self, request: CompletionRequest,
             lora_request = self._maybe_get_lora(request)
             guided_decode_logit_processor = (
                 await get_guided_decoding_logits_processor(
-                    request, self.engine.get_tokenizer()))
+                    request, await self.engine.get_tokenizer()))
             if guided_decode_logit_processor is not None:
                 if sampling_params.logits_processors is None:
                     sampling_params.logits_processors = []

vllm/transformers_utils/tokenizer.py

@@ -120,7 +120,8 @@ async def encode_async(
 
     def get_lora_tokenizer(
             self,
-            lora_request: Optional[LoRARequest]) -> "PreTrainedTokenizer":
+            lora_request: Optional[LoRARequest] = None
+    ) -> "PreTrainedTokenizer":
         if not lora_request or not self.enable_lora:
             return self.tokenizer
         if lora_request.lora_int_id not in self.lora_tokenizers:
@@ -133,7 +134,8 @@ def get_lora_tokenizer(
 
     async def get_lora_tokenizer_async(
             self,
-            lora_request: Optional[LoRARequest]) -> "PreTrainedTokenizer":
+            lora_request: Optional[LoRARequest] = None
+    ) -> "PreTrainedTokenizer":
         if not lora_request or not self.enable_lora:
             return self.tokenizer
         if lora_request.lora_int_id not in self.lora_tokenizers: