fix: vllm version compatibility

Signed-off-by: wxiwnd <[email protected]>

xinference/model/llm/vllm/core.py

@@ -499,7 +499,7 @@ async def async_generate(
         request_id: Optional[str] = None,
     ) -> Union[Completion, AsyncGenerator[CompletionChunk, None]]:
         try:
-            from vllm.sampling_params import GuidedDecodingParams, SamplingParams
+            from vllm.sampling_params import SamplingParams
         except ImportError:
             error_message = "Failed to import module 'vllm'"
             installation_guide = [
@@ -533,6 +533,8 @@ async def async_generate(
 
         if VLLM_INSTALLED and vllm.__version__ >= "0.6.3":
             # guided decoding only available for vllm >= 0.6.3
+            from vllm.sampling_params import GuidedDecodingParams
+
             guided_options = GuidedDecodingParams.from_optional(
                 json=sanitized_generate_config.pop("guided_json", None),
                 regex=sanitized_generate_config.pop("guided_regex", None),
@@ -549,6 +551,7 @@ async def async_generate(
                 guided_decoding=guided_options, **sanitized_generate_config
             )
         else:
+            # ignore generate configs
             sanitized_generate_config.pop("guided_json", None)
             sanitized_generate_config.pop("guided_regex", None)
             sanitized_generate_config.pop("guided_choice", None)