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 4, 03:45 UTC
fix: default TILELANG_CLEANUP_TEMP_FILES=1 to avoid shared /tmp conflicts (#41486)
Signed-off-by: Samaresh Kumar Singh <ssam3003@gmail.com>
SamareshSingh committed
May 7, 2026 at 21:59 UTC
01b0f3adab12f883bf8a4f2460134fa8839b54be
1 file changed
+7
vllm/env_override.py
+7
@@ -112,6 +112,13 @@ os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1"
112
# in the environment.
113
os.environ.setdefault("TRITON_CACHE_AUTOTUNING", "1")
114
115
+# When unset, TileLang routes JIT temp dirs through a world-shared
116
+# /tmp/tvm-debug-mode-tempdirs/ whose ownership is pinned to whichever
117
+# user compiled first, breaking every other user on a shared host.
118
+# Opt into per-process tempdirs unless the user explicitly chose the
119
+# debug layout (see https://github.com/vllm-project/vllm/issues/41410).
120
+os.environ.setdefault("TILELANG_CLEANUP_TEMP_FILES", "1")
121
+
122
# ===================================================
123
# torch 2.9 Inductor PythonWrapperCodegen monkeypatch
124
# ===================================================