Use FP32 master LoRA weights for stable L4 distillation

Harden v7 distillation with BF16 and non-finite guards

Place injected Geometry MoE experts on AniGen device and dtype

Make Geometry MoE sparse-aware for AniGen SLat FFNs

Add exact 2:4 shared-teacher recovery with final re-prune

Add selective FP8 QAT shared-teacher recovery

Keep MoE router probabilities differentiable for load balancing loss

Add Geometry-Inductive MoE shared-teacher training loop

Make Geometry MoE trainable with shared-base teacher mode

Backprop rig-aware flow symmetry prior during v7 distillation

Integrate flow equivariance symmetry loss for v7 distillation

Support ordered adapter chains across v7 distillation stages

Make v7 LoRA sparse-aware for AniGen SparseLinear layers

Fix sparse batch-size handling in v7 few-step distillation

Map cached AniGen condition dicts to real SS and SLat model signatures

Fix v7 ModelOpt FP8 and AutoQuantize API for 2026 release

Add function-preserving Geometry MoE injection for AniGen FFNs

Backprop KD and MeanFlow sequentially to fit L4 memory

Materialize teacher inference targets for v7 backward

Materialize teacher inference targets for v7 backward

Store progressive few-step students as chained compact LoRA adapters

Store v7 few-step students as compact chained LoRA adapters

Download selective AniGen flow teacher checkpoints for v7 training

Update v7 selective FP8 to current ModelOpt ordered quant_cfg API

Support combined SS/SLat teacher cache records

Use lightweight v7 teacher cache with regenerated SS/SLat noise

Make v7 flow distillation L4 memory efficient with shared teacher base

Make v7 flow distillation L4 memory efficient with shared teacher base

Fix Geometry MoE top-k weight indexing

Add exact FastGen rectified-flow bridge for AniGen SS

Add v7 distillation FP8 sparsity DINOv3 multiview symmetry MoE stack