biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Add test case for compiling multiple graphs (#21044)

Browse Source 
Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
This commit is contained in:
Yong Hoon Shin
2025-07-23 11:00:47 -07:00
committed by GitHub
parent 8560a5b258
commit 4ac7713e32
3 changed files with 390 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/compilation/compiler_interface.py
View File

@@ -423,6 +423,12 @@ class InductorAdaptor(CompilerInterface):
if is_torch_equal_or_newer("2.6"):
stack.enter_context(
torch._inductor.config.patch(fx_graph_remote_cache=False))
# InductorAdaptor (unfortunately) requires AOTAutogradCache
# to be turned off to run. It will fail to acquire the hash_str
# and error if not.
# StandaloneInductorAdaptor (PyTorch 2.8+) fixes this problem.
stack.enter_context(
torch._functorch.config.patch(enable_autograd_cache=False))
stack.enter_context(
torch._functorch.config.patch(
enable_remote_autograd_cache=False))