A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 01:45 UTC
[Core] Move function tracing setup to util function (#4352)
Nick Hill committed
Apr 25, 2024 at 16:45 UTC
efffb63f584c1ce4fdcf4e7b7fd0bfc8b33a733a
2 files changed
+24
-15
vllm/utils.py
+20
-1
@@ -1,10 +1,13 @@
1
import asyncio
2
+import datetime
3
import enum
4
import gc
5
import glob
6
import os
7
import socket
8
import subprocess
9
+import tempfile
10
+import threading
11
import uuid
12
import warnings
13
from collections import defaultdict
@@ -18,7 +21,7 @@ import psutil
21
import torch
22
from packaging.version import Version, parse
23
21
-from vllm.logger import init_logger
24
+from vllm.logger import enable_trace_function_call, init_logger
25
26
T = TypeVar("T")
27
logger = init_logger(__name__)
@@ -607,3 +610,19 @@ def find_nccl_library():
610
raise ValueError("NCCL only supports CUDA and ROCm backends.")
611
logger.info(f"Found nccl from library {so_file}")
612
return so_file
613
+
614
+
615
+def enable_trace_function_call_for_thread() -> None:
616
+ """Set up function tracing for the current thread,
617
+ if enabled via the VLLM_TRACE_FUNCTION environment variable
618
+ """
619
+
620
+ if int(os.getenv("VLLM_TRACE_FUNCTION", "0")):
621
+ tmp_dir = tempfile.gettempdir()
622
+ filename = (f"VLLM_TRACE_FUNCTION_for_process_{os.getpid()}"
623
+ f"_thread_{threading.get_ident()}_"
624
+ f"at_{datetime.datetime.now()}.log").replace(" ", "_")
625
+ log_path = os.path.join(tmp_dir, "vllm", get_vllm_instance_id(),
626
+ filename)
627
+ os.makedirs(os.path.dirname(log_path), exist_ok=True)
628
+ enable_trace_function_call(log_path)
vllm/worker/worker_base.py
+4
-14
@@ -1,15 +1,13 @@
1
-import datetime
1
import importlib
2
import os
4
-import tempfile
5
-import threading
3
from abc import ABC, abstractmethod
4
from typing import Dict, List, Set, Tuple
5
9
-from vllm.logger import enable_trace_function_call, init_logger
6
+from vllm.logger import init_logger
7
from vllm.lora.request import LoRARequest
8
from vllm.sequence import SamplerOutput, SequenceGroupMetadata
12
-from vllm.utils import get_vllm_instance_id, update_environment_variables
9
+from vllm.utils import (enable_trace_function_call_for_thread,
10
+ update_environment_variables)
11
12
logger = init_logger(__name__)
13
@@ -128,15 +126,7 @@ class WorkerWrapperBase:
126
function tracing if required.
127
Arguments are passed to the worker class constructor.
128
"""
131
- if int(os.getenv("VLLM_TRACE_FUNCTION", "0")):
132
- tmp_dir = tempfile.gettempdir()
133
- filename = (f"VLLM_TRACE_FUNCTION_for_process_{os.getpid()}"
134
- f"_thread_{threading.get_ident()}_"
135
- f"at_{datetime.datetime.now()}.log").replace(" ", "_")
136
- log_path = os.path.join(tmp_dir, "vllm", get_vllm_instance_id(),
137
- filename)
138
- os.makedirs(os.path.dirname(log_path), exist_ok=True)
139
- enable_trace_function_call(log_path)
129
+ enable_trace_function_call_for_thread()
130
131
mod = importlib.import_module(self.worker_module_name)
132
worker_class = getattr(mod, self.worker_class_name)