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
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185

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

use anyhow::Result;
use rlx_ir::HirGraphExt;
use rlx_ir::Shape;
use rlx_ir::hir::HirMut;

use super::BlockStage;
use crate::context::FlowCtx;
use crate::value::FlowValue;

/// QKV weight layout for BERT-family encoders.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum BertQkvStyle {
    /// `attention.self.{query,key,value}` (BERT / MiniLM).
    Bert,
    /// `attention.attn.{q,k,v}` (mpnet-style).
    Mpnet,
}

#[derive(Debug, Clone)]
pub struct BertEncoderLayerSpec {
    pub layer_prefix: String,
    pub qkv_style: BertQkvStyle,
    pub hidden_size: usize,
    pub num_heads: usize,
    pub head_dim: usize,
    pub eps: f32,
    pub attention_mask_input: String,
}

impl BertEncoderLayerSpec {
    pub fn hf(
        layer_prefix: impl Into<String>,
        qkv_style: BertQkvStyle,
        hidden_size: usize,
        num_heads: usize,
        eps: f32,
    ) -> Self {
        Self {
            layer_prefix: layer_prefix.into(),
            qkv_style,
            hidden_size,
            num_heads,
            head_dim: hidden_size / num_heads,
            eps,
            attention_mask_input: "attention_mask".into(),
        }
    }
}

#[derive(Debug, Clone)]
pub struct BertEncoderLayerStage {
    pub spec: BertEncoderLayerSpec,
}

impl BertEncoderLayerStage {
    pub fn new(spec: BertEncoderLayerSpec) -> Self {
        Self { spec }
    }
}

impl BlockStage for BertEncoderLayerStage {
    fn emit(&self, ctx: &mut FlowCtx<'_>, input: FlowValue) -> Result<Option<FlowValue>> {
        let spec = &self.spec;
        let h = spec.hidden_size;
        let nh = spec.num_heads;
        let dh = spec.head_dim;
        let lp = &spec.layer_prefix;

        let (qkv_w, qkv_b) = load_fused_qkv(ctx, lp, h, spec.qkv_style)?;

        let out_w = ctx.load_param(&format!("{lp}.attention.output.dense.weight"), true)?;
        let out_b = ctx.load_param(&format!("{lp}.attention.output.dense.bias"), false)?;
        let ln1_g = ctx.load_param(&format!("{lp}.attention.output.LayerNorm.weight"), false)?;
        let ln1_b = ctx.load_param(&format!("{lp}.attention.output.LayerNorm.bias"), false)?;
        let ln2_g = ctx.load_param(&format!("{lp}.output.LayerNorm.weight"), false)?;
        let ln2_b = ctx.load_param(&format!("{lp}.output.LayerNorm.bias"), false)?;
        let int_w = ctx.load_param(&format!("{lp}.intermediate.dense.weight"), true)?;
        let int_b = ctx.load_param(&format!("{lp}.intermediate.dense.bias"), false)?;
        let out2_w = ctx.load_param(&format!("{lp}.output.dense.weight"), true)?;
        let out2_b = ctx.load_param(&format!("{lp}.output.dense.bias"), false)?;

        let mask_id = ctx
            .state
            .inputs
            .get(&spec.attention_mask_input)
            .map(|(id, _)| *id)
            .ok_or_else(|| {
                anyhow::anyhow!(
                    "BertEncoderLayer requires input `{}`",
                    spec.attention_mask_input
                )
            })?;

        let mut gb = HirMut::new(ctx.hir());
        let skip = input.id;

        let qkv_mm = gb.mm(skip, qkv_w);
        let qkv = gb.add(qkv_mm, qkv_b);
        let last_ax = gb.shape(qkv).rank() - 1;
        let q = gb.narrow_(qkv, last_ax, 0, h);
        let k = gb.narrow_(qkv, last_ax, h, h);
        let v = gb.narrow_(qkv, last_ax, 2 * h, h);
        let attn = gb.attention_(q, k, v, mask_id, nh, dh);

        let attn_mm = gb.mm(attn, out_w);
        let attn_out = gb.add(attn_mm, out_b);
        let res1 = gb.add(attn_out, skip);
        let normed1 = gb.ln(res1, ln1_g, ln1_b, spec.eps);

        let int_mm = gb.mm(normed1, int_w);
        let int_add = gb.add(int_mm, int_b);
        let ffn_int = gb.gelu(int_add);
        let out2_mm = gb.mm(ffn_int, out2_w);
        let ffn_out = gb.add(out2_mm, out2_b);
        let res2 = gb.add(ffn_out, normed1);
        let out = gb.ln(res2, ln2_g, ln2_b, spec.eps);

        Ok(Some(ctx.wrap(out, input.shape.clone())))
    }
}

fn load_fused_qkv(
    ctx: &mut FlowCtx<'_>,
    layer_prefix: &str,
    h: usize,
    style: BertQkvStyle,
) -> Result<(rlx_ir::HirNodeId, rlx_ir::HirNodeId)> {
    let (wq_key, wk_key, wv_key, bq_key, bk_key, bv_key) = match style {
        BertQkvStyle::Bert => (
            format!("{layer_prefix}.attention.self.query.weight"),
            format!("{layer_prefix}.attention.self.key.weight"),
            format!("{layer_prefix}.attention.self.value.weight"),
            format!("{layer_prefix}.attention.self.query.bias"),
            format!("{layer_prefix}.attention.self.key.bias"),
            format!("{layer_prefix}.attention.self.value.bias"),
        ),
        BertQkvStyle::Mpnet => (
            format!("{layer_prefix}.attention.attn.q.weight"),
            format!("{layer_prefix}.attention.attn.k.weight"),
            format!("{layer_prefix}.attention.attn.v.weight"),
            format!("{layer_prefix}.attention.attn.q.bias"),
            format!("{layer_prefix}.attention.attn.k.bias"),
            format!("{layer_prefix}.attention.attn.v.bias"),
        ),
    };

    let wq_data = ctx.weights.take(&wq_key, true)?;
    let wk_data = ctx.weights.take(&wk_key, true)?;
    let wv_data = ctx.weights.take(&wv_key, true)?;
    let (bq_data, _) = ctx.weights.take(&bq_key, false)?;
    let (bk_data, _) = ctx.weights.take(&bk_key, false)?;
    let (bv_data, _) = ctx.weights.take(&bv_key, false)?;

    let w_name = format!("{layer_prefix}.attention.qkv.weight");
    let b_name = format!("{layer_prefix}.attention.qkv.bias");

    let (wq, _) = wq_data;
    let (wk, _) = wk_data;
    let (wv, _) = wv_data;

    let mut fused_w = vec![0f32; h * 3 * h];
    let mut fused_b = vec![0f32; 3 * h];
    for row in 0..h {
        fused_w[row * 3 * h..row * 3 * h + h].copy_from_slice(&wq[row * h..(row + 1) * h]);
        fused_w[row * 3 * h + h..row * 3 * h + 2 * h].copy_from_slice(&wk[row * h..(row + 1) * h]);
        fused_w[row * 3 * h + 2 * h..row * 3 * h + 3 * h]
            .copy_from_slice(&wv[row * h..(row + 1) * h]);
    }
    fused_b[..h].copy_from_slice(&bq_data);
    fused_b[h..2 * h].copy_from_slice(&bk_data);
    fused_b[2 * h..].copy_from_slice(&bv_data);

    let w_id = ctx
        .hir()
        .param(&w_name, Shape::new(&[h, 3 * h], rlx_ir::DType::F32));
    let b_id = ctx
        .hir()
        .param(&b_name, Shape::new(&[3 * h], rlx_ir::DType::F32));
    ctx.params.insert(w_name, fused_w);
    ctx.params.insert(b_name, fused_b);
    Ok((w_id, b_id))
}