Struct CompileOptions 

pub struct CompileOptions {

    pub precision: Precision,
    pub policy: Option<PrecisionPolicy>,
    pub scaled_quant: Option<ScaledQuantConfig>,
    pub rng: RngOptions,
    pub dce: bool,
    pub constant_folding: bool,
    pub verbose: bool,
    pub fusion_target: Option<FusionTarget>,
    pub fusion_opts: FusionOptions,
    pub arena_alignment: usize,
    pub assert_fusion_clean: bool,
    pub supported_ops: Option<&'static [OpKind]>,
    pub dim_binding: Option<DimBinding>,
    pub param_bindings: Option<HashMap<String, Vec<f32>>>,
    pub quant_param_bindings: Option<HashMap<String, Vec<u8>>>,
    pub kernel_dispatch: KernelDispatchConfig,
    pub disable_mpsgraph: bool,
    pub scan_unroll_max_length: u32,
    pub cache_param_invariant: bool,
    pub lint_numerics: bool,
    pub fusion_report: bool,
    pub disable_conv_bias_act_fusion: bool,
    pub no_io_peaks_output: bool,
}

All knobs the compile pipeline understands. Add new fields here rather than introducing new compile entry points.

Fields

precision: Precision

Target numeric precision for execution. Default: F32.

policy: Option<PrecisionPolicy>

Optional per-op precision policy (mixed precision rewrite).

scaled_quant: Option<ScaledQuantConfig>

Opt-in native low-precision GEMM. When set, every 2-D Op::MatMul is rewritten to a ScaledMatMul whose operands are dynamically quantized to this element format + scale layout — any ScaledFormat, including a parameterized Custom (e.g. f4e3m0). Off by default; changes numerics.

rng: RngOptions

RNG policy for in-graph Op::RngNormal / Op::RngUniform nodes.

dce: bool

Run dead-code elimination as part of compile. Default: true.

constant_folding: bool

Run constant folding. Default: true (cheap, only helps).

verbose: bool

Verbose pass logging. Equivalent to RLX_VERBOSE=1 or [rlx_ir::env::set("RLX_VERBOSE", "1")].

fusion_target: Option<FusionTarget>

Override fusion pipeline target (default: inferred from device).

fusion_opts: FusionOptions

Per-target fusion toggles (Metal env overrides, skip fusion, …).

arena_alignment: usize

Arena alignment for buffer planning. Default: 64.

assert_fusion_clean: bool

Panic at compile time if fusion diagnostics report missed patterns.

supported_ops: Option<&'static [OpKind]>

Backend op claim set for backend-aware fusion + post-fusion legalization. Set by Backend::compile implementations.

dim_binding: Option<DimBinding>

When set, specialize symbolic dims before backend lowering.

param_bindings: Option<HashMap<String, Vec<f32>>>

Bake fixed param tensors into constants before DCE / constant folding.

quant_param_bindings: Option<HashMap<String, Vec<u8>>>

Packed U8 GGUF weights for Op::Param nodes in DequantMatMul (TPU HLO bake).

kernel_dispatch: KernelDispatchConfig

Native vs common IR lowering (KernelDispatchConfig, RLX_KERNEL_DISPATCH=common).

disable_mpsgraph: bool

Prevent Metal from lowering the graph through MPSGraph.

This keeps the regular Metal thunk schedule while retaining all other compile options. Off by default.

scan_unroll_max_length: u32

On backends that list OpKind::Scan (host-fallback), Scans with length <= scan_unroll_max_length are IR-unrolled into body replicas so they run as ordinary device ops. Longer Scans keep ScanHost. Default: 64. 0 disables unroll on claiming backends; u32::MAX prefers unroll for every eligible Scan.

cache_param_invariant: bool

Hoist the param-invariant subgraph (compute that depends only on weights, never on Op::Input) into a separate “prepare” graph run ONCE, feeding its outputs into the main graph across forwards. Complements param_bindings (which folds such compute away at compile when weight VALUES are known); this handles run-time-only weights. Opt-in; default off (also settable via RLX_CACHE_PARAM_INVARIANT=1).

lint_numerics: bool

Static provable NaN/Inf lint during compile (RLX_LINT_NUMERICS).

fusion_report: bool

Print fusion before/after report (RLX_FUSION_REPORT).

disable_conv_bias_act_fusion: bool

Skip Conv+Bias+Act fusion (RLX_DISABLE_CONV_BIAS_ACT_FUSION).

no_io_peaks_output: bool

Disable IO-gated peaks-only fusion (RLX_NO_IO_PEAKS_OUTPUT).

Implementations

impl CompileOptions

pub fn baseline() -> Self

Defaults with no process-env overlay (tests / explicit builders).

pub fn from_env() -> Self

baseline plus Public/compile-layer RLX_* overlays from the registry.

pub fn apply_env(&mut self)

Overlay compile-layer env keys (idempotent). Uses rlx_ir::env_registry so aliases resolve.

pub fn new() -> Self

pub fn precision(self, p: Precision) -> Self

pub fn policy(self, p: PrecisionPolicy) -> Self

pub fn scaled_quant(self, cfg: ScaledQuantConfig) -> Self

Enable native low-precision GEMM for every 2-D matmul in the graph, in the given element format + scale layout. Accepts any ScaledFormat via ScaledQuantConfig — e.g. ScaledQuantConfig::fp8_e4m3(), or a parameterized ScaledFormat::custom(3, 0) (f4e3m0).

pub fn rng(self, rng: RngOptions) -> Self

pub fn rng_backend(self, backend: RngBackend) -> Self

pub fn rng_seed(self, seed: u64) -> Self

pub fn no_policy(self) -> Self

pub fn with_dce(self, on: bool) -> Self

pub fn with_constant_folding(self, on: bool) -> Self

pub fn with_verbose(self, on: bool) -> Self

pub fn fusion_target(self, target: FusionTarget) -> Self

pub fn fusion_opts(self, opts: FusionOptions) -> Self

pub fn arena_alignment(self, bytes: usize) -> Self

pub fn supported_ops(self, ops: &'static [OpKind]) -> Self

pub fn assert_fusion_clean(self, on: bool) -> Self

pub fn dim_binding(self, binding: DimBinding) -> Self

pub fn param_bindings(self, bindings: HashMap<String, Vec<f32>>) -> Self

pub fn quant_param_bindings(self, bindings: HashMap<String, Vec<u8>>) -> Self

Packed U8 weights for TPU GGUF DequantMatMul param bake at compile time.

pub fn kernel_dispatch(self, policy: KernelDispatchPolicy) -> Self

pub fn kernel_dispatch_config(self, config: KernelDispatchConfig) -> Self

pub fn scan_unroll_max_length(self, max: u32) -> Self

Prefer IR-unrolling Scans with length <= max on Scan-claiming backends.

pub fn force_common_kinds(self, kinds: &'static [OpKind]) -> Self

Force listed logical kernels to use common IR even when native is in supported_ops.

pub fn force_native_kinds(self, kinds: &'static [OpKind]) -> Self

Keep listed logical kernels native even under ForceCommon / missing from supported_ops.

pub fn with_lint_numerics(self, on: bool) -> Self

pub fn with_fusion_report(self, on: bool) -> Self

pub fn with_disable_mpsgraph(self, on: bool) -> Self

pub fn with_cache_param_invariant(self, on: bool) -> Self

pub fn with_no_io_peaks_output(self, on: bool) -> Self

pub fn with_disable_conv_bias_act_fusion(self, on: bool) -> Self

Trait Implementations

impl Clone for CompileOptions

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for CompileOptions

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for CompileOptions

fn default() -> Self

Returns the “default value” for a type.