mvanhorn opened a new pull request, #605: URL: https://github.com/apache/tvm-ffi/pull/605
## Summary Users whose CUDA driver is too old for automatic compute-capability detection can now find the `TVM_FFI_CUDA_ARCH_LIST` override without reading the source, both in the JIT build guide and in the error message they hit. ## Why this matters In issue [#430](https://github.com/apache/tvm-ffi/issues/430) a user on driver 470.82.01 hit a failure in `_get_cuda_target()` in `python/tvm_ffi/cpp/extension.py` because compute-capability auto-detection fails on stale driver metadata, and asked how to set the arch manually. Maintainer tqchen confirmed `TVM_FFI_CUDA_ARCH_LIST` already covers this and suggested surfacing it first before changing the API. The auto-detection gap itself was handled by merged #440; what was left was discoverability of the manual escape hatch. ## Testing The `_get_cuda_target()` docstring in `python/tvm_ffi/cpp/extension.py` now documents the env-var override and its `"8.9 9.0a"` space-separated format, and the "Could not detect CUDA compute_cap automatically" branch now names `TVM_FFI_CUDA_ARCH_LIST` as the remedy with the format example. A matching note was added to `docs/guides/kernel_library_guide.rst` for old-driver and undetectable-GPU cases. The override parsing path is unchanged, so existing `TVM_FFI_CUDA_ARCH_LIST` behavior is preserved. Documentation and message change only; doc build runs in CI. Fixes #430 -- 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]
