Add test case for compiling multiple graphs (#21044)

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
+6
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))
+34 -1
View File
@@ -20,9 +20,38 @@ from .monitor import start_monitoring_torch_compile
logger = init_logger(__name__)
IGNORE_COMPILE_KEY = "_ignore_compile_vllm"
_T = TypeVar("_T", bound=type[nn.Module])
def ignore_torch_compile(cls: _T) -> _T:
"""
A decorator to ignore support_torch_compile decorator
on the class. This is useful when a parent class has
a support_torch_compile decorator, but we don't want to
compile the class `cls` that inherits the parent class.
This only ignores compiling the forward of the class the
decorator is applied to.
If the parent has ignore_torch_compile but the child has
support_torch_compile, the child will still be compiled.
If the class has one or more submodules
that have support_torch_compile decorator applied, compile will
not be ignored for those submodules.
"""
setattr(cls, IGNORE_COMPILE_KEY, True)
return cls
def _should_ignore_torch_compile(cls) -> bool:
"""
Check if the class should be ignored for torch.compile.
"""
return getattr(cls, IGNORE_COMPILE_KEY, False)
@overload
def support_torch_compile(
*,
@@ -148,6 +177,8 @@ def _support_torch_compile(
old_init = cls.__init__
setattr(cls, IGNORE_COMPILE_KEY, False)
def __init__(self, *, vllm_config: VllmConfig, prefix: str = '', **kwargs):
old_init(self, vllm_config=vllm_config, prefix=prefix, **kwargs)
self.vllm_config = vllm_config
@@ -156,9 +187,11 @@ def _support_torch_compile(
self.do_not_compile = \
vllm_config.compilation_config.level in [
CompilationLevel.NO_COMPILATION, CompilationLevel.DYNAMO_AS_IS
] or not supports_dynamo()
] or not supports_dynamo() or _should_ignore_torch_compile(
self.__class__)
if self.do_not_compile:
return
compilation_counter.num_models_seen += 1
TorchCompileWrapperWithCustomDispatcher.__init__(
self, compilation_level=vllm_config.compilation_config.level)