pruning

Modules

`modelopt.torch.puzzletron.pruning.attention_ffn_surgery`	Model-agnostic surgery primitives for FFN-channel and attention-head pruning.
`modelopt.torch.puzzletron.pruning.compact_runtime`	Reversible compact forwards for physically sliced attention and GDN candidates.
`modelopt.torch.puzzletron.pruning.dynamic_block_prune`	Generic (AnyModel) dynamic single-block pruning — no realized weights.
`modelopt.torch.puzzletron.pruning.elastic_sampling`	Prefix-size sampling for nested (Matryoshka) elastic bypass distillation.
`modelopt.torch.puzzletron.pruning.embedding_pruning`	Descriptor-owned residual hidden-width ranking, sorting, and slicing.
`modelopt.torch.puzzletron.pruning.expert_removal_pruning_mixin`
`modelopt.torch.puzzletron.pruning.ffn_intermediate_pruning_mixin`
`modelopt.torch.puzzletron.pruning.gated_delta_net`	Qwen GatedDeltaNet geometry and coupled permutation helpers.
`modelopt.torch.puzzletron.pruning.gated_delta_net_pruning_mixin`
`modelopt.torch.puzzletron.pruning.kv_heads_pruning_mixin`
`modelopt.torch.puzzletron.pruning.latent_moe_surgery`	Weight-only surgery for latent-projected mixture-of-experts layers.
`modelopt.torch.puzzletron.pruning.mamba2_surgery`	Generic packed-tensor surgery for Mamba2-style mixers.
`modelopt.torch.puzzletron.pruning.materialize`	Materialize a pruned variant from the sorted teacher — slice/merge, no realized checkpoints.
`modelopt.torch.puzzletron.pruning.moe_mamba_pruning_mixin`
`modelopt.torch.puzzletron.pruning.ple_pruning`	Global per-layer-embedding (PLE) channel ranking and tensor surgery.
`modelopt.torch.puzzletron.pruning.pruning_mixin`
`modelopt.torch.puzzletron.pruning.pruning_utils`
`modelopt.torch.puzzletron.pruning.runtime_candidate`	Shared reversible application of one complete typed block candidate.
`modelopt.torch.puzzletron.pruning.runtime_hidden_width`	Differentiable full-envelope execution for a nested residual width.
`modelopt.torch.puzzletron.pruning.runtime_ple`	Static-envelope runtime slicing for descriptor-owned PLE channels.
`modelopt.torch.puzzletron.pruning.sorted_teacher`	Build a "sorted teacher": the teacher with each layer's FFN channels and attention heads reordered by importance (from activation scoring).

Structured pruning mixins and checkpoint utilities for Puzzletron.