[Kernel] Turn off CUTLASS scaled_mm for Ada Lovelace (#6384)

This commit is contained in:
Tyler Michael Smith
2024-07-14 13:37:19 +00:00
committed by GitHub
parent 6ef3bf912c
commit 9dad5cc859
3 changed files with 15 additions and 9 deletions
@@ -38,7 +38,13 @@ bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability) {
if (cuda_device_capability >= 90) {
return CUDA_VERSION >= 12000;
} else if (cuda_device_capability >= 89) {
return CUDA_VERSION >= 12040;
// CUTLASS Kernels have not been tuned for Ada Lovelace systems
// and are slower than torch.mm. Return false unconditionally in this case.
return false;
// Once the CUTLASS kernels have been optimized for Lovelace systems,
// use the following check:
// return CUDA_VERSION >= 12040;
}
#endif
@@ -98,4 +104,4 @@ void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
TORCH_CHECK(version_num >= 75);
cutlass_scaled_mm_sm75(c, a, b, a_scales, b_scales, bias);
}
}
}