[Pytorch] Add QuantizedTensor support in FusedAdam.step for MXFP8BlockScaling and Float8BlockScaling quantized model init.#2753

Merged

vthumbe1503 merged 15 commits intoNVIDIA:mainfrom

jomitchellnv:jm/fix-fused-adam-quantized-weights-fsdp2

Mar 13, 2026

Commits on Mar 11, 2026

Updates FusedAdam with FSDP2 and MXFP8
Jonathan Mitchell
committed
removes xfailing unit test for MXFPr MXFP8
Jonathan Mitchell
committed
addresses comments related to reset parameters and guard against self.capturable
Jonathan Mitchell
committed
adds e2e unit test
Jonathan Mitchell
committed
adds test to non meta device init
Jonathan Mitchell
committed
attempts to add float8block scaling fsdp hooks
Jonathan Mitchell
committed
adds e2e test for Float8BlockScaling
Jonathan Mitchell
committed
addresses review comments and code cleanup
Jonathan Mitchell
committed
more review comments addressed
Jonathan Mitchell
committed
removes unused block_len param
Jonathan Mitchell
committed