transformers

fb79484a - Move NVFP4Config onto the fine-grained machinery, with the producers' NVFP4 exports

Move NVFP4Config onto the fine-grained machinery, with the producers' NVFP4 exports - NVFP4Config / NVFP4HfQuantizer become a block-scaled format: E2M1 weights with E4M3 group-16 scales and an fp32 global (two-level), in both modelopt layouts (per expert, stacked per layer) and compressed-tensors'; the gate|up halves' globals fold through FineGrainedWeightGlobals; a weight-only checkpoint refuses quantized activations; integrations/nvfp4 is frozen. - shared_gate_up_activation_global: share the experts' gate_up activation global (max over all experts or over the rank's). - The modelopt and compressed-tensors readers learn NVFP4 (incl. modelopt MIXED_PRECISION plans across formats). - MuseSpark: its input-scaled RMS norm fuses into the kernels' chain, and its TP plan keeps experts with a post-expert norm whole. - Tests: the NVFP4 suite, the producer readers and skip-lists, the NVFP4 marshalling and conversion tests, mixed FP8/MXFP4/NVFP4 plans. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>