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 11, 08:45 UTC
[compile] Disable aot when eager backend is used. (#30810)
Signed-off-by: zhxchen17 <zhxchen17@fb.com>
Zhengxu Chen committed
Dec 17, 2025 at 04:55 UTC
177c391db2ad8dfc05906473525d4ae0a55549e0
1 file changed
+9
-2
vllm/compilation/decorators.py
+9
-2
@@ -435,7 +435,10 @@ def _support_torch_compile(
435
return self.aot_compiled_fn(self, *args, **kwargs)
436
437
if self.compiled:
438
- assert not envs.VLLM_USE_AOT_COMPILE
438
+ assert (
439
+ not envs.VLLM_USE_AOT_COMPILE
440
+ or self.vllm_config.compilation_config.backend == "eager"
441
+ )
442
return TorchCompileWithNoGuardsWrapper.__call__(self, *args, **kwargs)
443
444
# This is the path for the first compilation.
@@ -508,7 +511,11 @@ def _support_torch_compile(
511
_torch27_patch_tensor_subclasses(),
512
torch._inductor.config.patch(**inductor_config_patches),
513
):
511
- if envs.VLLM_USE_AOT_COMPILE:
514
+ use_aot_compile = envs.VLLM_USE_AOT_COMPILE
515
+ if self.vllm_config.compilation_config.backend == "eager":
516
+ logger.warning("Detected eager backend, disabling AOT compile.")
517
+ use_aot_compile = False
518
+ if use_aot_compile:
519
self.aot_compiled_fn = self.aot_compile(*args, **kwargs)
520
output = self.aot_compiled_fn(self, *args, **kwargs)
521
assert aot_compilation_path is not None