rlx-peft 0.2.16

Parameter-efficient adaptation on rlx graphs: LoRA, IA3, AdaLoRA, DoRA and OFT, matching HuggingFace PEFT
Documentation

rlx-peft

Parameter-efficient adaptation on RLX graphs — LoRA, IA3, AdaLoRA, DoRA and OFT as operations on an rlx_ir::Graph rather than as wrappers around another framework. Downstream package: builds on rlx-ir / rlx-runtime without touching framework core.

What's here

Each adapter reproduces HuggingFace PEFT's definition, which is what published results are measured under:

adapter update trainable
lora_delta W + (α/r)·B A A ∈ ℝ^{r×in}, B ∈ ℝ^{out×r}
ia3_apply (x Wᵀ) ⊙ l l ∈ ℝ^{out}
adalora_delta W + (α/r)·P diag(Λ) Q P, Λ, Q, with rank annealing
dora_weight m ⊙ (W + BA) / ‖W + BA‖_col m ∈ ℝ^{out} plus LoRA's A, B
oft_weight R W, R block-diagonal orthogonal skew-symmetric Q per block
  • adapters — the host-arithmetic definitions above, plus LoraConfig (r, alpha) and dora_init_magnitude.
  • oft — skew and cayley_orthogonal (R = (I − Q)(I + Q)⁻¹), the Cayley transform that makes the block rotation orthogonal by construction.
  • graph — the same updates as RLX graphs: lora_delta_node, lora_delta_graph, ia3_graph, adalora_delta_graph, dora_graph. These run on any backend the runtime has compiled in.
  • ParamBudget / lora_param_count / lora_is_economical — trainable-vs-total accounting. "LoRA matches full fine-tuning" means nothing without the denominator, and r(in + out) stops saving anything as r approaches min(in, out).

Initialisation conventions

Two are reproduced rather than chosen, because getting them wrong is a common source of "PEFT hurt my model" reports:

  • B starts at zero, so ΔW = 0 and the adapted model is identical to the base model at step 0. A nonzero B perturbs a pretrained network before any training happens.
  • IA3's l starts at one, for the same reason.

Features

  • cpu (default) — CPU execution for the graph API.
  • metal / mlx / coreml / cuda / gpu — forward the corresponding rlx-runtime backend feature.

Install

[dependencies]
rlx-peft = "0.2"

Quickstart

use rlx_peft::{lora_delta, LoraConfig};

// ΔW for one [out, in] layer, given the trained factors.
let delta = lora_delta(&a, &b, in_features, out_features, LoraConfig { r: 8, alpha: 16.0 })?;

License

MIT OR Apache-2.0.