Fix torch profiler bugs for bench_offline_throughput.py (#6557)

python/sglang/srt/managers/tokenizer_manager.py

@@ -116,6 +116,7 @@ from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.utils import (
     dataclass_to_string_truncated,
+    get_bool_env_var,
     get_zmq_socket,
     kill_process_tree,
 )
@@ -805,6 +806,8 @@ class TokenizerManager:
         profile_by_stage: bool = False,
     ):
         self.auto_create_handle_loop()
+        env_with_stack: bool = get_bool_env_var("SGLANG_PROFILE_WITH_STACK", "true")
+        with_stack = False if with_stack is False or env_with_stack is False else True
         req = ProfileReq(
             type=ProfileReqType.START_PROFILE,
             output_dir=output_dir,