[PyTorch] Reduce NVFP4 quantization work for the fused GroupedMLP fc2 input with activation recompute - #3837

[PyTorch] Reduce NVFP4 quantization work for the fused GroupedMLP fc2 input with activation recompute#3837

xrennvidia wants to merge 16 commits into

NVIDIA:mainNVIDIA/TransformerEngine:mainfrom

xrennvidia:xren/fused_grouped_mlp_rowwise_fc2x_recomputexrennvidia/TransformerEngine:xren/fused_grouped_mlp_rowwise_fc2x_recomputeCopy head branch name to clipboard