libmir-cuda 0.2.0

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, bf16};
use runtime::kv::{BlockTable, KvWritePlan};

use super::{AffineGatedFullAttentionMoeLayerConfig, scratch::FullAttentionLayerScratch};
use crate::{
    CudaAffineGatedFullAttention, CudaAffineGatedFullAttentionExecution,
    CudaAffineGatedFullAttentionState, CudaAffineSharedExpertMoe,
    CudaAffineSharedExpertMoeExecution, CudaBackend, CudaTensor, Error, Result,
    kernels::{ElementwiseBf16, ShiftedRmsNorm},
};

#[derive(Debug)]
pub struct CudaAffineGatedFullAttentionMoeExecution {
    backend: CudaBackend,
    attention: CudaAffineGatedFullAttentionExecution,
    moe: CudaAffineSharedExpertMoeExecution,
    input_norm: ShiftedRmsNorm,
    post_attention_norm: ShiftedRmsNorm,
    residual: ElementwiseBf16,
    input_norm_weight: CudaTensor,
    post_attention_norm_weight: CudaTensor,
    scratch: FullAttentionLayerScratch,
}

impl CudaAffineGatedFullAttentionMoeExecution {
    #[allow(clippy::too_many_arguments)]
    pub(super) fn new(
        backend: &CudaBackend,
        config: AffineGatedFullAttentionMoeLayerConfig,
        attention: &CudaAffineGatedFullAttention,
        moe: &CudaAffineSharedExpertMoe,
        input_norm_weight: &CudaTensor,
        post_attention_norm_weight: &CudaTensor,
        tokens: usize,
    ) -> Result<Self> {
        let hidden = config.attention.hidden_size;
        let norm = || {
            ShiftedRmsNorm::compile(
                &backend.inner.compiler,
                tokens,
                hidden,
                config.rms_norm_epsilon,
                config.norm_weight_shift,
            )
        };
        let elements = tokens
            .checked_mul(hidden)
            .ok_or(Error::InvalidDecoderKernel("gated full-attention residual shape overflow"))?;
        Ok(Self {
            backend: backend.clone(),
            attention: attention.prepare(tokens)?,
            moe: moe.prepare(tokens)?,
            input_norm: norm()?,
            post_attention_norm: norm()?,
            residual: ElementwiseBf16::compile(&backend.inner.compiler, elements)?,
            input_norm_weight: input_norm_weight.clone(),
            post_attention_norm_weight: post_attention_norm_weight.clone(),
            scratch: FullAttentionLayerScratch::new(backend, tokens, hidden)?,
        })
    }

    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &mut self,
        input: &DeviceBuffer<bf16>,
        positions: &DeviceBuffer<u32>,
        state: &mut CudaAffineGatedFullAttentionState,
        write_plan: &KvWritePlan,
        table: &BlockTable,
        start_position: usize,
        window: Option<usize>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.execute_with_image_span(
            input, positions, state, write_plan, table, start_position, window, None, output,
        )
    }

    #[allow(clippy::too_many_arguments)]
    pub fn execute_with_image_span(
        &mut self,
        input: &DeviceBuffer<bf16>,
        positions: &DeviceBuffer<u32>,
        state: &mut CudaAffineGatedFullAttentionState,
        write_plan: &KvWritePlan,
        table: &BlockTable,
        start_position: usize,
        window: Option<usize>,
        image_span: Option<(usize, usize)>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.validate(input, output)?;
        let stream = &self.backend.inner.stream;
        self.input_norm.execute(
            stream,
            input,
            bf16_tensor(&self.input_norm_weight)?,
            &mut self.scratch.normalized,
        )?;
        self.attention.execute_with_image_span(
            &self.scratch.normalized,
            positions,
            state,
            write_plan,
            table,
            start_position,
            window,
            image_span,
            &mut self.scratch.attention,
        )?;
        self.residual
            .add(stream, input, &self.scratch.attention, &mut self.scratch.residual)?;
        self.post_attention_norm.execute(
            stream,
            &self.scratch.residual,
            bf16_tensor(&self.post_attention_norm_weight)?,
            &mut self.scratch.normalized,
        )?;
        self.moe.execute(&self.scratch.normalized, &mut self.scratch.moe)?;
        self.residual.add(stream, &self.scratch.residual, &self.scratch.moe, output)
    }

    fn validate(&self, input: &DeviceBuffer<bf16>, output: &DeviceBuffer<bf16>) -> Result<()> {
        let expected = self.scratch.residual.len();
        if input.len() != expected || output.len() != expected {
            return Err(Error::InvalidDecoderKernel(
                "gated full-attention MoE layer buffer mismatch",
            ));
        }
        Ok(())
    }
}

fn bf16_tensor(tensor: &CudaTensor) -> Result<&DeviceBuffer<bf16>> {
    tensor.as_bf16().ok_or_else(|| Error::DTypeMismatch {
        name: tensor.name().into(),
        expected: "BF16",
    })
}