rlx_flow/blocks/
llama_decode_layer.rs

1// RLX — versatile ML compiler + runtime.
2// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
3//
4// This program is free software: you can redistribute it and/or modify
5// it under the terms of the GNU General Public License as published by
6// the Free Software Foundation, version 3.
7//
8// This program is distributed in the hope that it will be useful,
9// but WITHOUT ANY WARRANTY; without even the implied warranty of
10// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11// GNU General Public License for more details.
12//
13// You should have received a copy of the GNU General Public License
14// along with this program. If not, see <https://www.gnu.org/licenses/>.
15
16use anyhow::Result;
17use rlx_ir::HirGraphExt;
18use rlx_ir::hir::HirMut;
19use rlx_ir::op::MaskKind;
20use rlx_ir::shape;
21
22use super::BlockStage;
23use crate::context::FlowCtx;
24use crate::value::FlowValue;
25#[derive(Debug, Clone)]
26pub struct LlamaDecodeLayerSpec {
27    pub num_heads: usize,
28    pub head_dim: usize,
29    pub num_kv_heads: usize,
30    pub kv_group_size: usize,
31    pub eps: f32,
32    pub use_custom_mask: bool,
33    pub hidden_shape: rlx_ir::Shape,
34    /// RoPE pairing flavor. GGUF Llama weights need [`rlx_ir::RopeStyle::GptJ`];
35    /// HF-safetensors checkpoints use [`rlx_ir::RopeStyle::NeoX`] (default).
36    pub rope_style: rlx_ir::RopeStyle,
37}
38
39#[derive(Debug, Clone)]
40pub struct LlamaDecodeLayerStage {
41    pub layer_prefix: String,
42    pub spec: LlamaDecodeLayerSpec,
43    pub layer_idx: usize,
44    pub kv_out: std::sync::Arc<std::sync::Mutex<Vec<rlx_ir::HirNodeId>>>,
45    /// Optional EAGLE3-style tap for the pre-attention-norm layer
46    /// input. Mirrors the field on
47    /// [`crate::blocks::GemmaDecodeLayerStage`]; see that doc for
48    /// semantics and push-order guarantees.
49    pub aux_in_out: Option<std::sync::Arc<std::sync::Mutex<Vec<rlx_ir::HirNodeId>>>>,
50}
51
52impl LlamaDecodeLayerStage {
53    pub fn layer(
54        layer_idx: usize,
55        spec: LlamaDecodeLayerSpec,
56        kv_out: std::sync::Arc<std::sync::Mutex<Vec<rlx_ir::HirNodeId>>>,
57    ) -> Self {
58        Self {
59            layer_prefix: format!("model.layers.{layer_idx}"),
60            spec,
61            layer_idx,
62            kv_out,
63            aux_in_out: None,
64        }
65    }
66
67    pub fn with_aux_input_tap(
68        mut self,
69        sink: std::sync::Arc<std::sync::Mutex<Vec<rlx_ir::HirNodeId>>>,
70    ) -> Self {
71        self.aux_in_out = Some(sink);
72        self
73    }
74}
75
76impl BlockStage for LlamaDecodeLayerStage {
77    fn emit(&self, ctx: &mut FlowCtx<'_>, input: FlowValue) -> Result<Option<FlowValue>> {
78        if let Some(sink) = self.aux_in_out.as_ref() {
79            sink.lock().expect("aux in out").push(input.id);
80        }
81
82        let decode = ctx
83            .state
84            .decode
85            .clone()
86            .ok_or_else(|| anyhow::anyhow!("LlamaDecodeLayer requires BindDecodeInputs"))?;
87        let zero_beta = ctx
88            .state
89            .zero_beta
90            .ok_or_else(|| anyhow::anyhow!("LlamaDecodeLayer requires ZeroBeta"))?;
91
92        let lp = &self.layer_prefix;
93        let spec = &self.spec;
94        let in_ln_g = ctx.load_param(&format!("{lp}.input_layernorm.weight"), false)?;
95        let q_w = ctx.load_param(&format!("{lp}.self_attn.q_proj.weight"), true)?;
96        let k_w = ctx.load_param(&format!("{lp}.self_attn.k_proj.weight"), true)?;
97        let v_w = ctx.load_param(&format!("{lp}.self_attn.v_proj.weight"), true)?;
98        let o_w = ctx.load_param(&format!("{lp}.self_attn.o_proj.weight"), true)?;
99        let post_ln_g = ctx.load_param(&format!("{lp}.post_attention_layernorm.weight"), false)?;
100        let gate_w = ctx.load_param(&format!("{lp}.mlp.gate_proj.weight"), true)?;
101        let up_w = ctx.load_param(&format!("{lp}.mlp.up_proj.weight"), true)?;
102        let down_w = ctx.load_param(&format!("{lp}.mlp.down_proj.weight"), true)?;
103
104        let past_k = decode.past_k.get(self.layer_idx);
105        let past_v = decode.past_v.get(self.layer_idx);
106
107        let mut gb = HirMut::new(ctx.hir());
108        let normed_in = gb.rms_norm(input.id, in_ln_g, zero_beta, spec.eps);
109        let q = gb.mm(normed_in, q_w);
110        let k = gb.mm(normed_in, k_w);
111        let v = gb.mm(normed_in, v_w);
112
113        let q_rope = gb.rope_styled(q, decode.cos, decode.sin, spec.head_dim, spec.rope_style);
114        let k_rope = gb.rope_styled(k, decode.cos, decode.sin, spec.head_dim, spec.rope_style);
115
116        let (new_k, new_v) = match (past_k, past_v) {
117            (Some(past_k), Some(past_v)) => (
118                gb.concat_(vec![*past_k, k_rope], 1),
119                gb.concat_(vec![*past_v, v], 1),
120            ),
121            _ => (k_rope, v),
122        };
123        self.kv_out.lock().expect("kv out").push(new_k);
124        self.kv_out.lock().expect("kv out").push(new_v);
125
126        let k_rep = super::self_attn::repeat_kv(
127            &mut gb,
128            new_k,
129            spec.num_kv_heads,
130            spec.head_dim,
131            spec.kv_group_size,
132        );
133        let v_rep = super::self_attn::repeat_kv(
134            &mut gb,
135            new_v,
136            spec.num_kv_heads,
137            spec.head_dim,
138            spec.kv_group_size,
139        );
140
141        let attn_shape = shape::attention_shape(gb.shape(q_rope));
142        let attn = if spec.use_custom_mask {
143            let mask = decode
144                .mask
145                .ok_or_else(|| anyhow::anyhow!("custom mask requested but not bound"))?;
146            gb.attention(
147                q_rope,
148                k_rep,
149                v_rep,
150                mask,
151                spec.num_heads,
152                spec.head_dim,
153                attn_shape,
154            )
155        } else {
156            gb.attention_kind(
157                q_rope,
158                k_rep,
159                v_rep,
160                spec.num_heads,
161                spec.head_dim,
162                MaskKind::Causal,
163                attn_shape,
164            )
165        };
166
167        let attn_out = gb.mm(attn, o_w);
168        let post_attn = gb.add(input.id, attn_out);
169        let normed_post = gb.rms_norm(post_attn, post_ln_g, zero_beta, spec.eps);
170        let gate = gb.mm(normed_post, gate_w);
171        let up = gb.mm(normed_post, up_w);
172        let gate_act = gb.silu(gate);
173        let swiglu = gb.mul(gate_act, up);
174        let ffn_out = gb.mm(swiglu, down_w);
175        let h_id = gb.add(post_attn, ffn_out);
176
177        Ok(Some(ctx.wrap(h_id, spec.hidden_shape.clone())))
178    }
179}
rlx_flow/blocks/llama_decode_layer.rs

rlx_flow/blocks/
llama_decode_layer.rs
