Crate pf_model 

pf-model

Captures model-weight diffs (LoRA, IA³, full-finetune, in-place TTT) and implements the TIES + DARE task-vector merge from agent_docs/model-layer.md.

What ships in Phase 5 (this commit)

• diff::ModelDiff tagged enum + per-variant payloads.
• serialize::store_diff / serialize::load_diff: round-trip every variant through any pf_core::cas::BlobStore.
• merge::dare / merge::ties_merge: the two task-arithmetic primitives, tested on small synthetic tensors.

Wire format

For Phase 5 we use a single JSON-typed wire format (model.diff.v1) for every variant. Parameter tensors are stored as Vec<f32> with shape metadata; safetensors interop lands in Phase 10’s vLLM adapter where it’s actually needed (no point pulling 50 kloc of safetensors deps now).

Re-exports

pub use diff::DiffKind;

pub use diff::FullDelta;

pub use diff::IA3Delta;

pub use diff::InPlaceTttDelta;

pub use diff::LoraAdapter;

pub use diff::LoraDelta;

pub use diff::ModelDiff;

pub use diff::TttStep;

pub use merge::TiesParams;

pub use merge::dare;

pub use merge::ties_merge;

pub use serialize::load_diff;

pub use serialize::store_diff;

Modules

diff

Typed weight-diff payloads for the four supported diff kinds.

merge

TIES + DARE task-arithmetic merge primitives.

serialize

Round-trip every ModelDiff variant through a pf_core::cas::BlobStore.