[yugt]

Follow-up to #27, as agreed there.

What it does

On a card older than Volta the CUDA backend now stops before compiling anything:

Error: CUDA init

Caused by:
    NVIDIA GeForce GTX 1080 Ti is sm_61: the CUDA backend requires sm_70 or newer (Volta, Turing, Ampere and later); Pascal and older cards have no tensor cores and are not supported

Before, the same command printed cuda device 0: NVIDIA GeForce GTX 1080 Ti, sm_61 (Volta, m8n8k4) and failed in NVRTC with identifier "__nanosleep" is undefined.

• min_arch_error(name, cc) in src/cuda/mod.rs, called in Gpu::new right after compute_capability().
• The error carries the CUDA init context, like a missing driver or a missing NVRTC. So without --backend, backend::open goes on to Vulkan and, when that fails too (on this card it does, missing shaderFloat16), prints both reasons. The tests skip the CUDA backend on such a card instead of failing. Say if you would rather have it fatal.
• device_log no longer labels cc < 7.0 as Volta. Gpu::new does not reach it for such a card any more, so this is only a guard.

No change for sm_70 and newer.

Tested

• GTX 1080 Ti (sm_61, driver 580.95.05, NVRTC 12.6): bench --backend cuda gives the error above; bench without --backend gives no GPU backend loads: CUDA: CUDA init: NVIDIA GeForce GTX 1080 Ti is sm_61: ...; Vulkan: Vulkan init: no usable Vulkan device 0 (0 usable; NVIDIA GeForce GTX 1080 Ti: missing shaderFloat16); cargo test --test kernels skips both backends (26 passed).
• V100 16 GB (sm_70, NVRTC 12.6): bench with --backend cuda, without --backend, and with --backend vulkan run as before.
• The 9 unit tests on min_arch_error, arch_support_error and device_log pass on both hosts (3 are new).
• Not run: the full suite on this branch. On the V100, --test kernels has the 4 Vulkan f16 failures of main that docs: RTX 2080 Ti, Quadro RTX 6000, V100, A40, RTX A6000, RTX 4500 Ada and A100 40 GB results; GTX 1080 Ti does not run #27 fixes; nothing else fails there.

🤖 Generated with Claude Code

https://claude.ai/code/session_014eNPMupWL2tifApEBfcCgi

[Vasyl-Trefilov]

Looks good, and I agree with keeping it non-fatal — the CUDA init context makes the auto path try Vulkan and lets the tests skip. Two small things: maybe drop the "no tensor cores" justification and just say sm_70+ is required, and update the 1080 Ti note in #27/SUPPORTED_GPUS so it says it's rejected at init instead of failing in NVRTC. Unit tests pass here (27/27). Mergeable from my side.