rlx-flow 0.2.1

Block assembly-line API for RLX model builders — fusion-first, config-driven
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
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.

//! Gemma / Gemma 2 decoder blocks for tier-0 [`ModelFlow`] recipes.

use std::sync::{Arc, Mutex};

use super::{
    GeGluStage, GemmaDecodeLayerSpec, GemmaDecodeLayerStage, GemmaKvTapStage, GemmaRmsNormStage,
    SelfAttnPrefillSpec,
};
use crate::layer::LayerStack;
use crate::stage::FlowStage;
use rlx_ir::op::MaskKind;

/// Per-architecture layer recipe (norm placement + FFN style).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum GemmaLayerStyle {
    Gemma,
    Gemma2,
    Gemma3,
    Gemma4,
}

/// Build prefill self-attention spec for one layer.
pub fn gemma_attn_spec(
    layer: usize,
    num_heads: usize,
    head_dim: usize,
    num_kv_heads: usize,
    mask: MaskKind,
    score_scale: Option<f32>,
    attn_logit_softcap: Option<f32>,
) -> SelfAttnPrefillSpec {
    let prefix = format!("model.layers.{layer}");
    SelfAttnPrefillSpec {
        q_key: format!("{prefix}.self_attn.q_proj.weight"),
        k_key: format!("{prefix}.self_attn.k_proj.weight"),
        v_key: format!("{prefix}.self_attn.v_proj.weight"),
        num_heads,
        head_dim,
        num_kv_heads,
        mask,
        score_scale,
        attn_logit_softcap,
    }
}

/// Sliding-window mask for Gemma 2 local-attention layers.
pub fn gemma2_layer_mask(_layer: usize, window: usize) -> MaskKind {
    MaskKind::SlidingWindow(window)
}

/// Gemma 3 / 4 strided pattern: `stride-1` layers use full causal, others sliding.
pub fn gemma_strided_layer_mask(layer: usize, window: usize, stride: usize) -> MaskKind {
    if stride > 1 && (layer + 1).is_multiple_of(stride) {
        MaskKind::Causal
    } else {
        MaskKind::SlidingWindow(window)
    }
}

/// Composed Gemma prefill decoder block.
pub fn gemma_prefill_layer_composed(
    layer_idx: usize,
    style: GemmaLayerStyle,
    attn: SelfAttnPrefillSpec,
    eps: f32,
    kv_sink: Option<Arc<Mutex<Vec<rlx_ir::HirNodeId>>>>,
) -> FlowStage {
    let prefix = format!("model.layers.{layer_idx}");
    let mut stack = LayerStack::named(format!("layer{layer_idx}"))
        .residual_save()
        .stage(FlowStage::GemmaRmsNorm(GemmaRmsNormStage::hf_layer(
            format!("{prefix}.input_layernorm"),
            eps,
        )));

    if let Some(sink) = kv_sink {
        stack = stack.stage(FlowStage::GemmaKvTap(GemmaKvTapStage::layer(
            layer_idx,
            attn.head_dim,
            eps,
            sink,
        )));
    }

    stack = stack
        .self_attn_prefill(attn)
        .linear(format!("{prefix}.self_attn.o_proj.weight"), true)
        .residual_add()
        .residual_save();

    stack = if matches!(
        style,
        GemmaLayerStyle::Gemma2 | GemmaLayerStyle::Gemma3 | GemmaLayerStyle::Gemma4
    ) {
        stack.stage(FlowStage::GemmaRmsNorm(GemmaRmsNormStage::hf_layer(
            format!("{prefix}.pre_feedforward_layernorm"),
            eps,
        )))
    } else {
        stack.stage(FlowStage::GemmaRmsNorm(GemmaRmsNormStage::hf_layer(
            format!("{prefix}.post_attention_layernorm"),
            eps,
        )))
    };

    stack = stack.stage(FlowStage::GeGlu(GeGluStage::hf_mlp(&prefix)));

    if matches!(
        style,
        GemmaLayerStyle::Gemma2 | GemmaLayerStyle::Gemma3 | GemmaLayerStyle::Gemma4
    ) {
        stack = stack.stage(FlowStage::GemmaRmsNorm(GemmaRmsNormStage::hf_layer(
            format!("{prefix}.post_feedforward_layernorm"),
            eps,
        )));
    }

    stack.residual_add().build()
}

/// MoE placeholder — dense Gemma paths use [`gemma_prefill_layer_composed`].
pub fn gemma_moe_prefill_layer_composed(
    layer_idx: usize,
    style: GemmaLayerStyle,
    attn: SelfAttnPrefillSpec,
    eps: f32,
    kv_sink: Option<Arc<Mutex<Vec<rlx_ir::HirNodeId>>>>,
    _moe: super::MoeFfnStage,
) -> FlowStage {
    gemma_prefill_layer_composed(layer_idx, style, attn, eps, kv_sink)
}

pub fn gemma_moe_decode_layer_composed(
    layer_idx: usize,
    spec: GemmaDecodeLayerSpec,
    kv_out: Arc<Mutex<Vec<rlx_ir::HirNodeId>>>,
    _moe: super::MoeFfnStage,
) -> FlowStage {
    FlowStage::Named {
        name: format!("layer{layer_idx}"),
        inner: Arc::new(FlowStage::GemmaDecodeLayer(GemmaDecodeLayerStage::layer(
            layer_idx, spec, kv_out,
        ))),
    }
}