[Bugfix] Include device index in compile cache paths (#38962)
Signed-off-by: Neil Schemenauer <nas@arctrix.com> Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com> Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Neil Schemenauer committed
Aug 21, 2026 at 09:07 UTC
6dcc5d7caeffa7f69239153b171249e867ca6050
2 files changed
+6
-2
vllm/compilation/backends.py
+4
-1
@@ -1074,7 +1074,10 @@ 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
- local_cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}", self.prefix)
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
+ )
1081
os.makedirs(local_cache_dir, exist_ok=True)
1082
self.compilation_config.local_cache_dir = local_cache_dir
1083
vllm/compilation/decorators.py
+2
-1
@@ -559,7 +559,8 @@ 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
- cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}")
562
+ dev = torch.accelerator.current_device_index()
563
+ cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}_dev{dev}")
564
aot_compilation_path = os.path.join(cache_dir, "model")
565
if not envs.VLLM_DISABLE_COMPILE_CACHE:
566
loaded_fn = _try_load_aot_compiled_fn(self, aot_compilation_path)