A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 22:45 UTC
[compile] Clean up AOT compile bypass on evaluate_guards. (#33578)
Signed-off-by: zhxchen17 <zhxchen17@fb.com>
Zhengxu Chen committed
Feb 4, 2026 at 05:12 UTC
bcd2f74c0d1e85a2da4dcb41849ad75a7e3fdaf4
1 file changed
-8
vllm/compilation/wrapper.py
-8
@@ -151,14 +151,6 @@ class TorchCompileWithNoGuardsWrapper:
151
"requires VLLM_USE_BYTECODE_HOOK=0. "
152
)
153
154
- if envs.VLLM_USE_AOT_COMPILE:
155
- # disabled until https://github.com/pytorch/pytorch/pull/169239
156
- # is picked up.
157
- assert ds_type != DynamicShapesType.BACKED, (
158
- "evaluate_guards for backed shapes requires "
159
- "VLLM_USE_AOT_COMPILE=False. "
160
- )
161
-
154
options["guard_filter_fn"] = lambda x: [
155
entry.guard_type == "SHAPE_ENV" for entry in x
156
]