Revert compile-cache device index regression on CPU (#53304)
Kevin H. Luu committed
Aug 21, 2026 at 13:08 UTC
c0ff33404b93d82f2f2f30a362f372fc99eca8f7
2 files changed
+2
-6
vllm/compilation/backends.py
+1
-4
@@ -1074,10 +1074,7 @@ class VllmBackend:
1074
self.compilation_config.cache_dir = cache_dir
1075
rank = vllm_config.parallel_config.rank
1076
dp_rank = vllm_config.parallel_config.data_parallel_index
1077
- dev = torch.accelerator.current_device_index()
1078
- local_cache_dir = os.path.join(
1079
- cache_dir, f"rank_{rank}_{dp_rank}_dev{dev}", self.prefix
1080
- )
1077
+ local_cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}", self.prefix)
1078
os.makedirs(local_cache_dir, exist_ok=True)
1079
self.compilation_config.local_cache_dir = local_cache_dir
1080
vllm/compilation/decorators.py
+1
-2
@@ -559,8 +559,7 @@ def _support_torch_compile(
559
560
rank = self.vllm_config.parallel_config.rank
561
dp_rank = self.vllm_config.parallel_config.data_parallel_index
562
- dev = torch.accelerator.current_device_index()
563
- cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}_dev{dev}")
562
+ cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}")
563
aot_compilation_path = os.path.join(cache_dir, "model")
564
if not envs.VLLM_DISABLE_COMPILE_CACHE:
565
loaded_fn = _try_load_aot_compiled_fn(self, aot_compilation_path)