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
// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.
use anyhow::Result;
use rlx_ir::HirGraphExt;
use rlx_ir::hir::HirMut;
use rlx_ir::{DType, Shape};
use super::BlockStage;
use crate::context::FlowCtx;
use crate::value::FlowValue;
/// Multiply token embeddings by `sqrt(hidden_size)` (Gemma / Gemma 2).
#[derive(Debug, Clone)]
pub struct EmbedScaleStage {
pub hidden_size: usize,
}
impl EmbedScaleStage {
pub fn new(hidden_size: usize) -> Self {
Self { hidden_size }
}
}
impl BlockStage for EmbedScaleStage {
fn emit(&self, ctx: &mut FlowCtx<'_>, input: FlowValue) -> Result<Option<FlowValue>> {
let scale = (self.hidden_size as f32).sqrt();
let name = format!("gemma.embed_scale.{}", self.hidden_size);
let scale_id = ctx.synth_param(&name, vec![scale], Shape::new(&[1], DType::F32));
let mut gb = HirMut::new(ctx.hir());
let out = gb.mul(input.id, scale_id);
Ok(Some(ctx.wrap(out, input.shape.clone())))
}
}