rlx-opt 0.2.6

RLX optimizer facade — fusion, autodiff, and compile pipelines
Documentation

rlx-opt

Facade crate that re-exports rlx-fusion, rlx-autodiff, and rlx-compile with the historical rlx_opt:: module paths. Most callers use rlx and import rlx::opt::*; depend on rlx-opt directly only when you want the optimizer stack without the runtime.

Crate split (0.2)

Crate Role
rlx-fusion MIR fusion passes, Pass trait, unfuse_fused_for_autodiff
rlx-autodiff grad_with_loss, jvp, hvp, vmap, prepare_graph_for_ad
rlx-compile CompilePipeline, legalization, memory plan, precision / PTQ

Implementation lives in those crates; rlx-opt only wires backward-compatible rlx_opt::fusion, rlx_opt::autodiff, rlx_opt::legalize, etc.

Features

Feature Enables
compile (default) rlx-compile — HIR → MIR → LIR pipeline
training (default) rlx-autodiff — reverse / forward AD, vmap
full both

Fusion pipeline (via rlx-compile)

Fusion is backend-aware: fusion_passes_for_supported selects passes from a backend's OpKind claim set so the optimizer never emits fused ops the target cannot lower (e.g. Metal may skip FuseAttentionBlock).

Typical order:

  1. Constant folding — fold compile-time-known subgraphs.
  2. Fusion passes — gated pattern fusions + elementwise regions.
  3. Legalize — broadcast materialization, backend-specific rewrites.
  4. Memory plan — liveness → arena buffer assignment.
  5. DCE — dead-code elimination (last in the fusion pipeline).

See rlx-fusion/README.md and rlx-compile/README.md for crate-local detail.

Install

[dependencies]
rlx-opt = "0.2"

Build / test

cargo build -p rlx-opt
cargo test  -p rlx-fusion -p rlx-autodiff -p rlx-compile

Gotchas

  • Pass order matters: const-fold before fusion; memory plan after fusion.
  • New fused ops need Op + infer + verifier + every backend thunk you target.
  • precision inserts Cast nodes; some are eliminated by fusion peepholes.

License

GPL-3.0-only.