PyTorch enables FP8 tests on AMD ROCm GPUs
PyTorch's ROCm backend now supports FP8 matrix multiplication tests on AMD GPUs like MI300 and RDNA4.
The PyTorch team has enabled FP8 OpInfo tests for matrix multiplication operations on AMD ROCm-capable GPUs. This update replaces blanket ROCm test skips with `PLATFORM_SUPPORTS_FP8` checks, allowing FP8-capable hardware like AMD's MI300 (gfx942) and RDNA4 (gfx120x) to properly test FP8 matrix operations.
The change ensures consistent test IDs across platforms by using `float8_e4m3fn` as a stable label, even when underlying hardware uses different FP8 variants like `float8_e4m3fnuz` on MI300. Tests will now run on NVIDIA SM8.9+, MI300 (gfx942), MI350 (gfx950 with ROCm 6.5+), RDNA4 (gfx120x with ROCm 6.5+), and gfx1250 (ROCm 7.14+), while being skipped on older or unsupported configurations.
- PyTorch's ROCm backend now enables FP8 OpInfo tests for matrix multiplication using `PLATFORM_SUPPORTS_FP8` instead of blanket skips
- FP8-capable GPUs like AMD MI300 (gfx942) and RDNA4 (gfx120x) can now properly test FP8 operations with platform-native dtypes
- Tests use consistent IDs across platforms while adapting to hardware-specific FP8 variants
Why It Matters
Enables proper FP8 testing on AMD GPUs, improving PyTorch's ROCm support for AI/ML workloads that rely on FP8 matrix operations.