xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-17 07:05:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[cold start time] add envs.VLLM_COMPILE_DEPYF to guard decompile (#20790)

Browse Source 
Signed-off-by: Boyuan Feng <boyuan@meta.com>
This commit is contained in:
Boyuan Feng 2025-07-11 23:06:13 -07:00 committed by GitHub
parent a3a5a47e48
commit c1c8ca57ff
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 19 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/compilation/wrapper.py
View File

@ -95,16 +95,26 @@ class TorchCompileWrapperWithCustomDispatcher:
self.compiled_codes.append(new_code)
local_cache_dir = self.vllm_config.compilation_config.local_cache_dir
if isinstance(local_cache_dir, str):
decompiled_file_name = ("transformed_code.py"
if envs.VLLM_COMPILE_DEPYF else
"transformed_code_README.txt")
decompiled_file = os.path.join(local_cache_dir,
"transformed_code.py")
decompiled_file_name)
if not os.path.exists(decompiled_file):
try:
# usually the decompilation will succeed for most models,
# as we guarantee a full-graph compilation in Dynamo.
# but there's no 100% guarantee, since decompliation is
# not a reversible process.
if envs.VLLM_COMPILE_DEPYF:
import depyf
src = depyf.decompile(new_code)
else:
src = (
"To get a transformed_code.py file, re-run with "
"VLLM_COMPILE_DEPYF=1")
with open(decompiled_file, "w") as f:
f.write(src)

6
vllm/envs.py
View File

@ -97,6 +97,7 @@ if TYPE_CHECKING:
VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
VLLM_DISABLE_COMPILE_CACHE: bool = False
VLLM_COMPILE_DEPYF: bool = False
Q_SCALE_CONSTANT: int = 200
K_SCALE_CONSTANT: int = 200
V_SCALE_CONSTANT: int = 100
@ -741,6 +742,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_DISABLE_COMPILE_CACHE":
lambda: bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))),
# If set, vllm will decompile the torch compiled code and dump to
# transformed_code.py. This is useful for debugging.
"VLLM_COMPILE_DEPYF":
lambda: bool(int(os.getenv("VLLM_COMPILE_DEPYF", "0"))),
# If set, vllm will run in development mode, which will enable
# some additional endpoints for developing and debugging,
# e.g. `/reset_prefix_cache`