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 6, 14:30 UTC
[ROCm] Fix AttributeError for torch.compiler.skip_all_guards_unsafe on older PyTorch (#37219)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Mar 16, 2026 at 22:34 UTC
54a62a79f70982742a227c845b96148e6401d0e7
1 file changed
+6
-1
vllm/compilation/wrapper.py
+6
-1
@@ -112,7 +112,12 @@ class TorchCompileWithNoGuardsWrapper:
112
entry.guard_type == "SHAPE_ENV" for entry in x
113
]
114
else:
115
- options["guard_filter_fn"] = torch.compiler.skip_all_guards_unsafe
115
+ if hasattr(torch.compiler, "skip_all_guards_unsafe"):
116
+ # Torch 2.10+ provides skip_all_guards_unsafe
117
+ options["guard_filter_fn"] = torch.compiler.skip_all_guards_unsafe
118
+ else:
119
+ # Equivalent fallback for older PyTorch: skip all guards
120
+ options["guard_filter_fn"] = lambda x: [False for _ in x]
121
122
compiled_ptr: Any = self.forward
123
# Validate that unbacked dynamic shapes require VLLM_USE_BYTECODE_HOOK=False