biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 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
2 changed files with 19 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
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.
import depyf
src = depyf.decompile(new_code)
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)