rlx-ir 0.2.17

Tensor IR for the RLX ML compiler — standalone, serializable, optimizable
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
// SPDX-License-Identifier: MIT OR Apache-2.0

//! Fusion policy for HIR → MIR lowering.

/// How HIR block ops lower to MIR.
///
/// Fusion is a first-class concern at the HIR layer: model builders
/// choose whether to express **intent** (`Fusable`) or emit **fused
/// MIR ops directly** (`Direct`).
#[cfg_attr(feature = "serialize", derive(serde::Serialize, serde::Deserialize))]
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum FusionPolicy {
    /// Lower to primitive chains the optimizer recognizes (`MatMul →
    /// Add → Silu`, shared-input matmul pairs, …). Fusion passes in
    /// `rlx_opt` collapse them to fused ops.
    #[default]
    Fusable,
    /// Lower straight to fused MIR ops (`FusedMatMulBiasAct`,
    /// `FusedSwiGLU`, `FusedResidualRmsNorm`, …). Passes still run for
    /// escape-hatch `HirOp::Mir` nodes and opportunistic cleanup.
    Direct,
}

impl FusionPolicy {
    pub fn is_direct(self) -> bool {
        matches!(self, Self::Direct)
    }

    /// Lower HIR to **primitive MIR chains** so autodiff can skip
    /// tier-2 fused-op unfuse when you control lowering explicitly.
    ///
    /// Same as [`Self::Fusable`]. Fusion passes may still fuse later
    /// during [`CompilePipeline`](../../rlx-opt/src/compiler.rs) for
    /// inference; run `rlx_opt::prepare_graph_for_ad` before AD when
    /// the graph already contains fused ops from `Direct` lowering.
    pub const fn for_autodiff() -> Self {
        Self::Fusable
    }
}