From c0ff33404b93d82f2f2f30a362f372fc99eca8f7 Mon Sep 17 00:00:00 2001 From: "Kevin H. Luu" Date: Fri, 21 Aug 2026 13:08:13 -0700 Subject: [PATCH] Revert compile-cache device index regression on CPU (#53304) --- vllm/compilation/backends.py | 5 +---- vllm/compilation/decorators.py | 3 +-- 2 files changed, 2 insertions(+), 6 deletions(-) diff --git a/vllm/compilation/backends.py b/vllm/compilation/backends.py index 8117be69f1..439a87e152 100644 --- a/vllm/compilation/backends.py +++ b/vllm/compilation/backends.py @@ -1074,10 +1074,7 @@ class VllmBackend: self.compilation_config.cache_dir = cache_dir rank = vllm_config.parallel_config.rank dp_rank = vllm_config.parallel_config.data_parallel_index - dev = torch.accelerator.current_device_index() - local_cache_dir = os.path.join( - cache_dir, f"rank_{rank}_{dp_rank}_dev{dev}", self.prefix - ) + local_cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}", self.prefix) os.makedirs(local_cache_dir, exist_ok=True) self.compilation_config.local_cache_dir = local_cache_dir diff --git a/vllm/compilation/decorators.py b/vllm/compilation/decorators.py index eb9d73a2d3..ad710a6fe3 100644 --- a/vllm/compilation/decorators.py +++ b/vllm/compilation/decorators.py @@ -559,8 +559,7 @@ def _support_torch_compile( rank = self.vllm_config.parallel_config.rank dp_rank = self.vllm_config.parallel_config.data_parallel_index - dev = torch.accelerator.current_device_index() - cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}_dev{dev}") + cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}") aot_compilation_path = os.path.join(cache_dir, "model") if not envs.VLLM_DISABLE_COMPILE_CACHE: loaded_fn = _try_load_aot_compiled_fn(self, aot_compilation_path)