transformers

35b8576e - NVFP4: MoE experts and ModelOpt exports on finegrained-kernels

NVFP4: MoE experts and ModelOpt exports on finegrained-kernels - NVFP4 linears and experts run on kernels-community/finegrained-kernels. - Experts get their own grouped and batched implementations. - A config reader turns a ModelOpt export into an NVFP4Config. - FP8 layers of a ModelOpt mixed-precision export load dequantized. - NVFP4Config gains activation_format and dequantize. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>