Revert compile-cache device index regression on CPU (#53304)
This commit is contained in:
@@ -1074,10 +1074,7 @@ class VllmBackend:
|
|||||||
self.compilation_config.cache_dir = cache_dir
|
self.compilation_config.cache_dir = cache_dir
|
||||||
rank = vllm_config.parallel_config.rank
|
rank = vllm_config.parallel_config.rank
|
||||||
dp_rank = vllm_config.parallel_config.data_parallel_index
|
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}", self.prefix)
|
||||||
local_cache_dir = os.path.join(
|
|
||||||
cache_dir, f"rank_{rank}_{dp_rank}_dev{dev}", self.prefix
|
|
||||||
)
|
|
||||||
os.makedirs(local_cache_dir, exist_ok=True)
|
os.makedirs(local_cache_dir, exist_ok=True)
|
||||||
self.compilation_config.local_cache_dir = local_cache_dir
|
self.compilation_config.local_cache_dir = local_cache_dir
|
||||||
|
|
||||||
|
|||||||
@@ -559,8 +559,7 @@ def _support_torch_compile(
|
|||||||
|
|
||||||
rank = self.vllm_config.parallel_config.rank
|
rank = self.vllm_config.parallel_config.rank
|
||||||
dp_rank = self.vllm_config.parallel_config.data_parallel_index
|
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}")
|
||||||
cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}_dev{dev}")
|
|
||||||
aot_compilation_path = os.path.join(cache_dir, "model")
|
aot_compilation_path = os.path.join(cache_dir, "model")
|
||||||
if not envs.VLLM_DISABLE_COMPILE_CACHE:
|
if not envs.VLLM_DISABLE_COMPILE_CACHE:
|
||||||
loaded_fn = _try_load_aot_compiled_fn(self, aot_compilation_path)
|
loaded_fn = _try_load_aot_compiled_fn(self, aot_compilation_path)
|
||||||
|
|||||||
Reference in New Issue
Block a user