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
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156

use super::projection::OutputHeadWeight;
use crate::{
    CudaBackend, CudaTensor, Error, OutputHeadExecution, OutputHeadPlanRequest, Result,
    kernels::{Fp8OutputKernels, Fp8OutputSpec, Fp8ResidualWeightBuffers},
};

mod instantiate;

/// Model-owned output weight prepared once and shared by all sessions.
#[derive(Clone, Debug)]
pub struct CudaOutputHeadTemplate {
    pub(super) exact: CudaTensor,
    pub(super) weight: OutputHeadWeight,
    pub(super) kernels: Option<Fp8OutputKernels>,
    pub(super) execution: OutputHeadExecution,
    pub(super) input_features: usize,
    pub(super) output_features: usize,
}

impl CudaOutputHeadTemplate {
    pub(crate) fn prepare(
        backend: &CudaBackend,
        source: CudaTensor,
        input_features: usize,
        output_features: usize,
    ) -> Result<Self> {
        let plan = backend
            .execution_planner()
            .plan_output_head(OutputHeadPlanRequest { input_features, output_features })?;
        tracing::debug!(
            input_features,
            output_features,
            execution = ?plan.execution(),
            source = ?plan.source(),
            "planned CUDA output head"
        );
        match plan.execution() {
            OutputHeadExecution::Bf16 => Ok(Self {
                exact: source.clone(),
                weight: OutputHeadWeight::Bf16(source),
                kernels: None,
                execution: OutputHeadExecution::Bf16,
                input_features,
                output_features,
            }),
            execution
            @ (OutputHeadExecution::Fp8Blockwise | OutputHeadExecution::Fp8Vectorized) => {
                Self::prepare_fp8(backend, &source, input_features, output_features, execution)
            },
            OutputHeadExecution::Fp8Residual => {
                Self::prepare_residual(backend, &source, input_features, output_features)
            },
            execution @ (OutputHeadExecution::AutoRefined
            | OutputHeadExecution::Fp8BlockVectorized
            | OutputHeadExecution::Fp8BlockRefined) => {
                super::block::prepare(backend, &source, input_features, output_features, execution)
            },
        }
    }

    fn prepare_fp8(
        backend: &CudaBackend,
        source: &CudaTensor,
        input_features: usize,
        output_features: usize,
        execution: OutputHeadExecution,
    ) -> Result<Self> {
        let source_weight = bf16(source)?;
        let spec = Fp8OutputSpec::new(input_features, output_features)?;
        let kernels = Fp8OutputKernels::compile(&backend.inner.compiler, spec)?;
        let mut weight = backend
            .inner
            .pool
            .allocate::<u8>(&backend.inner.stream, spec.weight_elements()?)?;
        let mut scales = backend
            .inner
            .pool
            .allocate::<f32>(&backend.inner.stream, spec.weight_scale_elements()?)?;
        let mut row_scales =
            backend.inner.pool.allocate::<f32>(&backend.inner.stream, output_features)?;
        kernels.quantize_weight(
            &backend.inner.stream,
            source_weight,
            &mut weight,
            &mut scales,
            &mut row_scales,
        )?;
        tracing::debug!(input_features, output_features, "prepared blockwise E4M3 output head");
        Ok(Self {
            exact: source.clone(),
            weight: OutputHeadWeight::Fp8 { weight, scales, row_scales },
            kernels: Some(kernels),
            execution,
            input_features,
            output_features,
        })
    }

    fn prepare_residual(
        backend: &CudaBackend,
        source: &CudaTensor,
        input_features: usize,
        output_features: usize,
    ) -> Result<Self> {
        let source_weight = bf16(source)?;
        let spec = Fp8OutputSpec::new(input_features, output_features)?;
        let kernels = Fp8OutputKernels::compile(&backend.inner.compiler, spec)?;
        let mut weight = backend
            .inner
            .pool
            .allocate::<u8>(&backend.inner.stream, spec.weight_elements()?)?;
        let mut block_scales = backend
            .inner
            .pool
            .allocate::<f32>(&backend.inner.stream, spec.weight_scale_elements()?)?;
        let mut row_scales =
            backend.inner.pool.allocate::<f32>(&backend.inner.stream, output_features)?;
        let mut residual = backend
            .inner
            .pool
            .allocate::<u8>(&backend.inner.stream, spec.residual_elements()?)?;
        let mut residual_scales = backend
            .inner
            .pool
            .allocate::<f32>(&backend.inner.stream, spec.residual_scale_elements()?)?;
        let mut buffers = Fp8ResidualWeightBuffers::new(
            &mut weight,
            &mut block_scales,
            &mut row_scales,
            &mut residual,
            &mut residual_scales,
        );
        kernels.quantize_weight_residual(&backend.inner.stream, source_weight, &mut buffers)?;
        tracing::debug!(input_features, output_features, "prepared FP8 plus INT4 output head");
        Ok(Self {
            exact: source.clone(),
            weight: OutputHeadWeight::Fp8Residual {
                weight,
                row_scales,
                residual,
                residual_scales,
            },
            kernels: Some(kernels),
            execution: OutputHeadExecution::Fp8Residual,
            input_features,
            output_features,
        })
    }
}

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