Fix early CUDA init via get_architecture_class_name import (vllm-proj…

…ect#3770)

Signed-off-by: Lei Wen <[email protected]>
Co-authored-by: Lei Wen <[email protected]>

vllm/engine/llm_engine.py

@@ -13,7 +13,6 @@
 from vllm.executor.executor_base import ExecutorBase
 from vllm.logger import init_logger
 from vllm.lora.request import LoRARequest
-from vllm.model_executor.model_loader import get_architecture_class_name
 from vllm.outputs import RequestOutput
 from vllm.sampling_params import SamplingParams
 from vllm.sequence import (MultiModalData, SamplerOutput, Sequence,
@@ -115,6 +114,8 @@ def __init__(
 
         # If usage stat is enabled, collect relevant info.
         if is_usage_stats_enabled():
+            from vllm.model_executor.model_loader import (
+                get_architecture_class_name)
             usage_message.report_usage(
                 get_architecture_class_name(model_config),
                 usage_context,