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 16, 08:15 UTC
[V1] Use custom ops for piecewise CUDA graphs (#10227)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Woosuk Kwon committed
Nov 11, 2024 at 11:26 UTC
fe15729a2b77d760fcf99da76f15806c5eab33df
1 file changed
-2
vllm/v1/worker/gpu_model_runner.py
-2
@@ -1,4 +1,3 @@
1
-import os
1
import time
2
from dataclasses import dataclass
3
from typing import TYPE_CHECKING, Dict, List, Optional, Set
@@ -406,7 +405,6 @@ class GPUModelRunner:
405
if self.use_cuda_graph:
406
# FIXME(woosuk): Currently, we do not use inductor to reduce the
407
# compilation time and any potential issues with the inductor.
409
- os.environ["VLLM_CUSTOM_OPS"] = "none"
408
set_compilation_config(
409
CompilationConfig(
410
use_cudagraph=True,