tugot17 commented on PR #644: URL: https://github.com/apache/tvm-ffi/pull/644#issuecomment-4833062687
Right, and that lines up — the change only touches the JIT fallback at the bottom of `load_torch_c_dlpack_extension`. The AOT path (`import torch_c_dlpack_ext`) and the newer-torch builtin (`__dlpack_c_exchange_api__`, which returns early) both run before we ever reach the cache filename, so neither is affected. CUDA AOT wheels keep resolving exactly as they do today; the version-specific name only kicks in once we've already decided to JIT-compile into `~/.cache/tvm-ffi`. So it's really just narrowing the local JIT cache, which is per-machine and recompilable anyway — it doesn't constrain the broadly-compatible prebuilt wheel. And agreed, with newer torch shipping the extension and short-circuiting up top, this path is increasingly legacy; it mainly matters for the older-torch / no-prebuilt case, which on ROCm is where we hit it. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
