vLLM vllm/compilation/backends.py - where the torch.compile cache directory is resolved

Read at v0.30.0. Lines ~1062-1086: when compilation_config.cache_dir is empty, vLLM hashes [env_hash, config_hash, code_hash, compiler_hash] with SHA-256, truncates to 10 hex chars, and joins it onto envs.VLLM_CACHE_ROOT/torch_compile_cache. vllm/envs.py line 34 sets VLLM_CACHE_ROOT = os.path.expanduser('~/.cache/vllm'), so with no override the cache root is the invoking user's home directory - inside the container filesystem for a containerised deployment.