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 8, 00:00 UTC
[compile] Move torch_aot_compile directory under torch_compile_cache (#34831)
Signed-off-by: zhxchen17 <zhxchen17@fb.com>
Zhengxu Chen committed
Feb 20, 2026 at 11:46 UTC
e4a5d8c653fc00adb06922bddcb7fec14b01a62b
1 file changed
+5
-4
vllm/compilation/decorators.py
+5
-4
@@ -407,10 +407,10 @@ def _support_torch_compile(
407
if envs.VLLM_USE_AOT_COMPILE:
408
"""
409
When using torch.compile in AOT mode, we store the cache artifacts
410
- under VLLM_CACHE_ROOT/torch_aot_compile/{hash}/rank_i_j. The {hash}
411
- contains all of the factors except for the source files being
412
- traced through, because we don't actually know which source files
413
- to check at this point (before dynamo runs).
410
+ under VLLM_CACHE_ROOT/torch_compile_cache/torch_aot_compile/{hash}
411
+ The {hash} contains all of the factors except for the source files
412
+ being traced through, because we don't actually know which source
413
+ files to check at this point (before dynamo runs).
414
On loading we will actually look at the source files being traced
415
through. If any source file have changed (compared with the
416
serialized backend artifacts), then we need to generate a new AOT
@@ -424,6 +424,7 @@ def _support_torch_compile(
424
hash_key = hashlib.sha256(str(factors).encode()).hexdigest()
425
cache_dir = os.path.join(
426
envs.VLLM_CACHE_ROOT,
427
+ "torch_compile_cache",
428
"torch_aot_compile",
429
hash_key,
430
)