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

use mircuda::{DeviceBuffer, bf16};

use crate::{
    Bf16LinearPackWeights, BlockFp8LinearWeight, CudaBackend, CudaTensor,
    DecodeAttentionOutputWeight, DecodeAttentionWeights, DecodeQkvWeights, DenseExecution,
    DensePlanRequest, DenseRole, Error, ExecutionPhase, Fp8ResidualLinearWeight, NvFp4LinearWeight,
    Result,
};

#[derive(Clone)]
enum CapturedOutputWeight {
    Bf16(CudaTensor),
    NvFp4(NvFp4LinearWeight),
    BlockFp8 {
        exact: CudaTensor,
        quantized: BlockFp8LinearWeight,
    },
    Fp8Int4 {
        exact: CudaTensor,
        quantized: Fp8ResidualLinearWeight,
    },
}

#[derive(Clone)]
enum CapturedQkvWeights {
    Bf16(Bf16LinearPackWeights<3>),
    NvFp4([NvFp4LinearWeight; 3]),
}

#[derive(Clone)]
pub(in crate::backend) struct CapturedAttentionWeights {
    input_norm: CudaTensor,
    qkv: CapturedQkvWeights,
    query_norm: DeviceBuffer<bf16>,
    key_norm: DeviceBuffer<bf16>,
    output: CapturedOutputWeight,
}

impl CapturedAttentionWeights {
    pub(in crate::backend) fn prepare(
        backend: &CudaBackend,
        weights: DecodeAttentionWeights<'_>,
    ) -> Result<Self> {
        let output = weights.output.bf16().ok_or(Error::InvalidExecutionPlan(
            "optimized BF16 graph preparation received NVFP4 output weight",
        ))?;
        let [output_features, input_features] = output.shape() else {
            return Err(Error::InvalidLinearWeight {
                name: output.name().into(),
                expected: [0, 0],
                actual: output.shape().to_vec(),
            });
        };
        let request = DensePlanRequest {
            phase: ExecutionPhase::Decode,
            role: DenseRole::AttentionOutput,
            tokens: 1,
            input_features: *input_features,
            output_features: *output_features,
        };
        let output = match backend.execution_planner().plan_dense(request)?.execution() {
            DenseExecution::BlockFp8Vector => CapturedOutputWeight::BlockFp8 {
                exact: output.clone(),
                quantized: backend.prepare_block_fp8_linear_weight(output)?,
            },
            DenseExecution::Fp8Int4Vector => CapturedOutputWeight::Fp8Int4 {
                exact: output.clone(),
                quantized: backend.prepare_fp8_residual_linear_weight(output)?,
            },
            DenseExecution::Matrix | DenseExecution::Vector => {
                CapturedOutputWeight::Bf16(output.clone())
            },
        };
        Ok(Self {
            input_norm: weights.input_norm.clone(),
            qkv: CapturedQkvWeights::from(weights.qkv),
            query_norm: weights.query_norm.clone(),
            key_norm: weights.key_norm.clone(),
            output,
        })
    }

    pub(in crate::backend) fn borrow(&self) -> DecodeAttentionWeights<'_> {
        DecodeAttentionWeights {
            input_norm: &self.input_norm,
            qkv: self.qkv.borrow(),
            query_norm: &self.query_norm,
            key_norm: &self.key_norm,
            output: match &self.output {
                CapturedOutputWeight::Bf16(weight) => DecodeAttentionOutputWeight::Bf16(weight),
                CapturedOutputWeight::NvFp4(weight) => DecodeAttentionOutputWeight::NvFp4(weight),
                CapturedOutputWeight::BlockFp8 { exact, quantized } => {
                    DecodeAttentionOutputWeight::BlockFp8 { exact, quantized }
                },
                CapturedOutputWeight::Fp8Int4 { exact, quantized } => {
                    DecodeAttentionOutputWeight::Fp8Int4 { exact, quantized }
                },
            },
        }
    }
}

impl From<DecodeAttentionWeights<'_>> for CapturedAttentionWeights {
    fn from(weights: DecodeAttentionWeights<'_>) -> Self {
        Self {
            input_norm: weights.input_norm.clone(),
            qkv: CapturedQkvWeights::from(weights.qkv),
            query_norm: weights.query_norm.clone(),
            key_norm: weights.key_norm.clone(),
            output: CapturedOutputWeight::from(weights.output),
        }
    }
}

impl From<DecodeQkvWeights<'_>> for CapturedQkvWeights {
    fn from(weights: DecodeQkvWeights<'_>) -> Self {
        match weights {
            DecodeQkvWeights::Bf16(weights) => Self::Bf16(weights.clone()),
            DecodeQkvWeights::NvFp4(weights) => Self::NvFp4(weights.map(Clone::clone)),
        }
    }
}

impl CapturedQkvWeights {
    fn borrow(&self) -> DecodeQkvWeights<'_> {
        match self {
            Self::Bf16(weights) => DecodeQkvWeights::Bf16(weights),
            Self::NvFp4(weights) => {
                DecodeQkvWeights::NvFp4([&weights[0], &weights[1], &weights[2]])
            },
        }
    }
}

impl From<DecodeAttentionOutputWeight<'_>> for CapturedOutputWeight {
    fn from(weight: DecodeAttentionOutputWeight<'_>) -> Self {
        match weight {
            DecodeAttentionOutputWeight::Bf16(weight) => Self::Bf16(weight.clone()),
            DecodeAttentionOutputWeight::NvFp4(weight) => Self::NvFp4(weight.clone()),
            DecodeAttentionOutputWeight::BlockFp8 { exact, quantized } => Self::BlockFp8 {
                exact: exact.clone(),
                quantized: quantized.clone(),
            },
            DecodeAttentionOutputWeight::Fp8Int4 { exact, quantized } => Self::Fp8Int4 {
                exact: exact.clone(),
                quantized: quantized.clone(),
            },
        }
    }
}