Fix CI (#9012)

python/sglang/srt/entrypoints/engine.py

@@ -67,6 +67,7 @@ from sglang.srt.utils import (
     MultiprocessingSerializer,
     assert_pkg_version,
     configure_logger,
+    get_bool_env_var,
     get_zmq_socket,
     is_cuda,
     kill_process_tree,
@@ -627,7 +628,6 @@ def _set_envs_and_config(server_args: ServerArgs):
     os.environ["NCCL_CUMEM_ENABLE"] = str(int(server_args.enable_symm_mem))
     if not server_args.enable_symm_mem:
         os.environ["NCCL_NVLS_ENABLE"] = str(int(server_args.enable_nccl_nvls))
-    os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
     os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "4"
     os.environ["CUDA_MODULE_LOADING"] = "AUTO"
 
@@ -647,7 +647,7 @@ def _set_envs_and_config(server_args: ServerArgs):
             "reinstall the latest version by following the instructions "
             "at https://docs.flashinfer.ai/installation.html.",
         )
-    if _is_cuda:
+    if _is_cuda and not get_bool_env_var("SGLANG_SKIP_SGL_KERNEL_VERSION_CHECK"):
         assert_pkg_version(
             "sgl-kernel",
             "0.3.3",