21 lines
851 B
Markdown
21 lines
851 B
Markdown
# Blackwell Notes
|
|
|
|
This repository targets a Blackwell-style workflow, but keeps the build configuration explicit because local toolchain support may differ across systems.
|
|
|
|
## Build Guidance
|
|
|
|
- Prefer explicit architecture selection over hidden defaults.
|
|
- Use `KERNEL_LAB_CUDA_ARCH=120` for Python-side build helpers when your local environment supports it.
|
|
- Use `-DCMAKE_CUDA_ARCHITECTURES=120` with CMake for direct native builds.
|
|
- If your toolkit does not yet accept the exact architecture value you want, adjust the build flag rather than editing the kernels.
|
|
|
|
## What To Watch On A New GPU Generation
|
|
|
|
- compiler support for the target architecture
|
|
- PyTorch wheel compatibility
|
|
- Triton support level
|
|
- driver/toolkit mismatch
|
|
- profiler tool compatibility
|
|
|
|
Treat environment validation as part of the lab, not as a one-time setup nuisance.
|