Struct FusionOptions 

pub struct FusionOptions {
    pub skip_fusion: bool,
    pub unfuse_elementwise_regions: bool,
    pub keep_elementwise_regions: bool,
    pub decompose_fusion_regions: bool,
    pub fk_fusion: bool,
    pub fuse_region_prologue: bool,
    pub fuse_batch_preprocess: bool,
    pub native_fk_regions: bool,
    pub fusion_limits: FusionLimits,
    pub disable_conv_bias_act_fusion: bool,
    pub no_io_peaks_output: bool,
}

Per-target fusion toggles (env-driven on Metal today).

Fields

skip_fusion: bool

Skip all pattern fusions (Metal: RLX_METAL_NO_FUSION).

unfuse_elementwise_regions: bool

Break ElementwiseRegion back into primitives after marking.

keep_elementwise_regions: bool

Keep fused ElementwiseRegion through lowering (env: RLX_KEEP_ELEMENTWISE_REGIONS).

decompose_fusion_regions: bool

Decompose FKL-style transform / batch regions before backend lowering.

fk_fusion: bool

Run FKL passes (MarkTransformRegions, prologue, batch). Env: RLX_NO_FK_FUSION=1 disables.

fuse_region_prologue: bool

Fold ResizeNearest2x into ElementwiseRegion prologue. Env: RLX_FUSE_REGION_PROLOGUE=0 disables.

fuse_batch_preprocess: bool

Merge parallel region slices into BatchElementwiseRegion. Env: RLX_FUSE_BATCH_PREPROCESS=0 disables.

native_fk_regions: bool

Keep TransformRegion / BatchElementwiseRegion in MIR for native lowering. Env: RLX_NATIVE_FK_REGIONS=1.

fusion_limits: FusionLimits

Caps for fused elementwise chains (encoder / scratch limits).

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 output selection (RLX_NO_IO_PEAKS_OUTPUT).

Implementations

impl FusionOptions

pub fn from_metal_env() -> FusionOptions

Read Metal-specific env overrides.

pub fn merge_env(self) -> FusionOptions

Merge session options with compile-time env overrides.

pub fn apply_native_fk_defaults(self, target: FusionTarget) -> FusionOptions

GPU-class targets keep native FKL regions unless opted out.

pub fn for_cpu() -> FusionOptions

CPU executes element-wise chains as per-op thunks — mark then unfuse.

pub fn for_metal() -> FusionOptions

Metal keeps RMSNorm / matmul fusions but unfuses ElementwiseRegion by default (fused MSL mis-lowers long chains on deep transformer graphs) — UNLESS RLX_KEEP_ELEMENTWISE_REGIONS asks to keep them (short chains lower fine and stay fused → fewer dispatches on the dispatch-bound backward).

pub fn for_wgpu() -> FusionOptions

wgpu region kernel only supports trailing/scalar broadcast via modulus — unfuse so LegalizeBroadcast Expand + Binary run separately.

impl FusionOptions

pub fn for_target(target: FusionTarget) -> FusionOptions

Canonical option set for a target, matching what each backend’s compile path uses: CPU unfuses element-wise regions (it executes them as per-op thunks), Metal/MLX keep matmul/norm fusions but unfuse long element-wise chains, wgpu unfuses for its modulus-broadcast kernel, and the remaining GPU-class targets take the defaults (native FKL regions are then enabled by apply_native_fk_defaults). This is the option set the one-call fuse / Fuse entries pick.

Trait Implementations

impl Clone for FusionOptions

fn clone(&self) -> FusionOptions

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for FusionOptions

impl Debug for FusionOptions

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

Formats the value using the given formatter.

impl Default for FusionOptions

fn default() -> FusionOptions

Returns the “default value” for a type.

impl Eq for FusionOptions

impl PartialEq for FusionOptions

fn eq(&self, other: &FusionOptions) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for FusionOptions