Module op 

Operation types — every tensor op in the RLX IR.

Designed for pattern-matching fusion: ops are grouped by category so fusion passes can reason about them structurally.

Enums

Activation

Unary element-wise activation functions.

AttentionBwdWrt

Which forward input an Op::AttentionBackward node differentiates.

BinaryOp

Binary element-wise operations.

ChainOperand

An operand inside a fused ChainStep — either a graph-level input to the Op::ElementwiseRegion (by index 0..num_inputs) or the result of a previous step in the chain (by index 0..step_position).

ChainStep

One step in a fused element-wise chain. Each step produces exactly one scalar result (per element); later steps can refer to it via ChainOperand::Step. The whole chain runs per element in registers.

CmpOp

Comparison operations (return Bool tensor).

MaskKind

What kind of attention mask the kernel should apply.

Op

An operation in the RLX IR graph.

OpKind

PLAN L4: discriminant for each Op variant. Used by Op::kind + the Backend::supported_ops trait method to declare which ops a backend can lower; the LegalizeForBackend pass in rlx-opt checks the graph against this set and fails the compile when an unsupported op is present (instead of silent fallback).

ReduceOp

Reduction operations along specified axes.

ScaleMode

Scale-tracking strategy for Op::FakeQuantize. Determines how the per-channel s[c] is computed each forward pass.

SteKind

Straight-through estimator variants for Op::FakeQuantize’s backward. The forward is the same regardless: discrete clamp(round(x/s)) * s. The choice here affects only the gradient w.r.t. x during training.