xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-25 00:05:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

enable unbacked with aot_compile (#30462)

Browse Source 
Signed-off-by: Laith Sakka <lsakka@meta.com>
This commit is contained in:
Laith Sakka 2025-12-14 11:14:06 +03:00 committed by GitHub
parent 97f2f160fd
commit f569c654e1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 8 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
tests/compile/test_dynamic_shapes_compilation.py
View File

@ -36,7 +36,7 @@ def get_test_models():
DynamicShapesType.BACKED_SIZE_OBLIVIOUS,
],
)
@pytest.mark.parametrize("use_aot_compile", ["0"])
@pytest.mark.parametrize("use_aot_compile", ["0", "1"])
@pytest.mark.parametrize("use_bytecode_hook", [True, False])
@pytest.mark.parametrize("evaluate_guards", [False, True])
@pytest.mark.skipif(
@ -54,6 +54,12 @@ def test_dynamic_shapes_compilation(
if use_bytecode_hook and shapes_type == DynamicShapesType.UNBACKED:
pytest.skip("UNBACKED dynamic shapes require VLLM_USE_BYTECODE_HOOK=0")
if evaluate_guards and shapes_type == DynamicShapesType.UNBACKED:
pytest.skip("unbacked dynamic shapes do not add guards")
if evaluate_guards and use_aot_compile:
pytest.skip("evaluate_guards requires use_aot_compile=0")
monkeypatch.setenv("VLLM_USE_AOT_COMPILE", use_aot_compile)
monkeypatch.setenv("VLLM_USE_BYTECODE_HOOK", "1" if use_bytecode_hook else "0")
@ -120,7 +126,7 @@ def test_model_specialization_with_evaluate_guards(
and dynamic_shapes_type == DynamicShapesType.BACKED
and evaluate_guards
):
pytest.skip("evaluate_guards for backed does not work with aot_compile =1")
pytest.skip("evaluate_guards for backed does not work with aot_compile=1")
@support_torch_compile
class ModelWithSizeCheck(torch.nn.Module):

8
vllm/compilation/decorators.py
View File

@ -390,14 +390,6 @@ def _support_torch_compile(
serialized backend artifacts), then we need to generate a new AOT
compile artifact from scratch.
"""
# Validate that AOT compile is not used with unbacked dynamic
# shapes. aot_compile re-allocates backed symbols post dynamo!
if ds_type == DynamicShapesType.UNBACKED:
raise ValueError(
"AOT compilation is not compatible with UNBACKED dynamic shapes. "
"Please use BACKED or BACKED_SIZE_OBLIVIOUS dynamic shapes type "
"when VLLM_USE_AOT_COMPILE is enabled."
)
from .caching import compilation_config_hash_factors
factors: list[str] = compilation_config_hash_factors(self.vllm_config)