Closed vkuzo closed 1 month ago
Is there a plan to support AMP?
Is there a plan to support AMP?
Sorry for late reply! We don't have a plan to support AMP in the near future because the eng cost to support delayed scaling in an AMP-like API would be too high, because delayed scaling is stateful. For now we would like to have a consistent API between dynamic and delayed scaling for easy ablation studies. If the community converges on dynamic scaling in the future (which is stateless), we could adjust.
configurability
Float8Linear
to individual modulesperformance
torch._scaled_mm
support for per-tensor scaled float8 gemmtorch._scaled_mm
support for rowwise scaled float8 gemmdistributed
other
use_fast_accum
(float8 accumulation of gemm) option to UX - https://github.com/pytorch-labs/float8_experimental/pull/144