[CI/Build] vLLM cache directory for images (#6444)
This commit is contained in:
@@ -98,8 +98,7 @@ class TPUWorker(LoraNotSupportedWorkerBase):
|
||||
# Use persistent cache to avoid XLA recompilation.
|
||||
# NOTE(woosuk): This does not completely eliminate the recompilation
|
||||
# overhead because dynamo does not cache the compiled results.
|
||||
xr.initialize_cache(os.path.expanduser(envs.VLLM_XLA_CACHE_PATH),
|
||||
readonly=False)
|
||||
xr.initialize_cache(envs.VLLM_XLA_CACHE_PATH, readonly=False)
|
||||
|
||||
def load_model(self):
|
||||
self.model_runner.load_model()
|
||||
|
||||
Reference in New Issue
Block a user