libmir-cuda 0.1.0

CUDA inference backend for libmir
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

use mircuda::{DeviceBuffer, Stream, bf16};
use runtime::kv::{BlockTable, KvWritePlan};

use super::{
    DecodeDenseSwiGlu, DenseScratch, DenseSwiGluConfig, DenseSwiGluWeights, DownProjection,
    GateUpBuffers, GateUpProjection, validate,
};
use crate::{
    CudaBackend, Error, PrefillAttentionBf16, Result, RmsNormBf16,
    kernels::{ElementwiseBf16, PackedGatedBf16},
};

pub struct PrefillDenseSwiGlu {
    attention: PrefillAttentionBf16,
    post_attention_norm: RmsNormBf16,
    gate_up: GateUpProjection,
    down: DownProjection,
    activation: PackedGatedBf16,
    hidden_ops: ElementwiseBf16,
    scratch: DenseScratch,
    stream: Stream,
    config: DenseSwiGluConfig,
    tokens: usize,
}

impl PrefillDenseSwiGlu {
    pub(in crate::backend) fn new(
        backend: &CudaBackend,
        config: DenseSwiGluConfig,
        tokens: usize,
        weights: DenseSwiGluWeights<'_>,
    ) -> Result<Self> {
        validate(config)?;
        if tokens == 0 {
            return Err(Error::InvalidDecoderKernel("dense prefill chunk is empty"));
        }
        let hidden = config.attention.hidden_size;
        let intermediate = config.intermediate_size;
        Ok(Self {
            attention: PrefillAttentionBf16::new(
                backend,
                config.attention,
                tokens,
                Some(weights.attention),
            )?,
            post_attention_norm: RmsNormBf16::new(
                backend,
                tokens,
                hidden,
                config.attention.rms_norm_epsilon,
            )?,
            gate_up: GateUpProjection::new(backend, config, tokens, Some(weights.gate_up))?,
            down: DownProjection::new(backend, config, tokens, Some(weights.down))?,
            activation: PackedGatedBf16::compile(&backend.inner.compiler, tokens, intermediate)?,
            hidden_ops: ElementwiseBf16::compile(
                &backend.inner.compiler,
                super::scratch::elements(tokens, hidden)?,
            )?,
            scratch: DenseScratch::new(backend, config, tokens)?,
            stream: backend.inner.stream.clone(),
            config,
            tokens,
        })
    }

    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &mut self,
        state: &mut DecodeDenseSwiGlu,
        input: &DeviceBuffer<bf16>,
        weights: DenseSwiGluWeights<'_>,
        write_plan: &KvWritePlan,
        table: &BlockTable,
        start_position: usize,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.attention.execute(
            &mut state.attention,
            input,
            weights.attention,
            write_plan,
            table,
            start_position,
            &mut self.scratch.attention,
        )?;
        self.hidden_ops.add(
            &self.stream,
            input,
            &self.scratch.attention,
            &mut self.scratch.residual,
        )?;
        let separate = self.gate_up.execute(
            &self.scratch.residual,
            &self.post_attention_norm,
            weights.post_attention_norm,
            weights.gate_up,
            &mut GateUpBuffers {
                normalized: &mut self.scratch.normalized,
                packed: &mut self.scratch.gate_up,
                separate: &mut self.scratch.gate_up_separate,
            },
        )?;
        let fused_down = separate
            && self.down.execute_gated(
                &self.scratch.gate_up_separate[0],
                &self.scratch.gate_up_separate[1],
                self.config.activation,
                weights.down,
                &mut self.scratch.mlp,
            )?;
        if separate && !fused_down {
            self.activation.execute_separate(
                &self.stream,
                &self.scratch.gate_up_separate[0],
                &self.scratch.gate_up_separate[1],
                &mut self.scratch.activated,
                self.config.activation.into(),
            )?;
        } else if !fused_down {
            self.activation.execute(
                &self.stream,
                &self.scratch.gate_up,
                &mut self.scratch.activated,
                self.config.activation.into(),
            )?;
        }
        if !fused_down {
            self.down
                .execute(&self.scratch.activated, weights.down, &mut self.scratch.mlp)?;
        }
        self.hidden_ops
            .add(&self.stream, &self.scratch.residual, &self.scratch.mlp, output)
    }

    #[must_use]
    pub const fn tokens(&self) -> usize {
        self.tokens
    }
}