[Feature] Make QDP CUDA kernel build targets configurable and future-compatible

[viiccwen]

Related Issues

Closes #1282

Changes

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Why

QDP's CUDA kernel build currently relies on a small hardcoded set of nvcc -gencode targets. That makes it awkward to keep existing GPUs working while also supporting newer architectures such as Blackwell, and it forces follow-up source edits whenever the desired target mix changes.

This change makes architecture targeting configurable and toolchain-aware so one build configuration can remain usable across current and newer NVIDIA GPU generations without changing QDP runtime behavior.

How

• derive the default cubin/PTX target set from a project shortlist filtered by the local nvcc supported architecture lists
• add QDP_CUDA_ARCH_LIST as an explicit override for local builds, CI, and packaging workflows
• preserve a legacy fallback when nvcc does not expose architecture listing flags
• validate the change locally on both Ada and Blackwell GPUs and rerun targeted QDP GPU tests

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[viiccwen]

cc @ryankert01, @rich7420, need testing on ur local machine.
Already testing in both RTX 4090 and newer architectures Pro 6000 Blackwell. It's all work well.

[Copilot]

Pull request overview

This PR updates qdp-kernels’ CUDA build script to make nvcc -gencode architecture targeting configurable and to better accommodate newer NVIDIA GPU architectures without requiring source edits.

Changes:

• Introduces an architecture-target model for nvcc -gencode flags (SM and optional PTX targets).
• Adds QDP_CUDA_ARCH_LIST to explicitly override the CUDA target set at build time.
• Attempts to derive default targets from nvcc’s supported architecture lists, with a legacy fallback when listing flags are unavailable.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[viiccwen]

@ryankert01 sounds like getting-started.md. I'll woke on it.

[ryankert01]

sorry @viiccwen, I thought we have a section say we only support some Nvidia gpus (30, 40), but we turns out don't have this section. So, I think it's not needed.

Also, just a heads up, the 30s gpu server that I have access to is currently compromised. So I might not be able to test it anywhere soon.

[400Ping]

Agree on writing a section on the current supporting GPUs(Including Nvidia and AMD).

[viiccwen]

Goti it, I think a backend-level section is more accurate than listing specific GPU models.

For NVIDIA, QDP does not maintain a fixed supported-SKU whitelist. The actual CUDA targets generated by the build depend on the installed CUDA toolkit and the local nvcc supported architectures, with QDP selecting from its default architecture shortlist.

For AMD, support similarly depends on the local ROCm environment and the Triton backend used by QDP, rather than a hardcoded model list in the repo.

Because of that, I plan to document this as a “Supported GPU Backends” section instead of list a set of supported GPU models.

[ryankert01]

tested!