libmir-cuda 0.3.0

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

use mircuda::{DeviceBuffer, bf16};

use crate::{
    AffineQuantizedEmbedding, Bf16Embedding, CudaBackend, DenseRole, Error, MxFp4EmbeddingLookup,
    Result,
    backend::linear::{
        CheckpointProjection, CheckpointProjectionWeight, CompressedInt8Embedding,
        packed_integer_embedding,
    },
};

#[derive(Debug)]
pub(super) enum SharedRoutedEmbedding {
    Affine {
        operation: AffineQuantizedEmbedding,
        weight: crate::AffineQuantizedWeight,
        vocab: usize,
    },
    Dense {
        operation: Bf16Embedding,
        weight: crate::CudaTensor,
        vocab: usize,
    },
    MxFp4(MxFp4EmbeddingLookup),
    PackedInteger {
        operation: CompressedInt8Embedding,
        vocab: usize,
    },
}

#[derive(Debug)]
pub(super) struct SharedRoutedOutputHead {
    operation: CheckpointProjection,
    hidden: usize,
    vocab: usize,
}

impl SharedRoutedEmbedding {
    pub(super) fn new(
        backend: &CudaBackend,
        hidden: usize,
        vocab: usize,
        weight: &CheckpointProjectionWeight,
    ) -> Result<Self> {
        weight.affine_format(1, hidden, vocab)?;
        match weight {
            CheckpointProjectionWeight::Affine(weight) => {
                let config = weight.infer_config(1, hidden, vocab)?;
                Ok(Self::Affine {
                    operation: backend.prepare_affine_embedding(config, 1.0)?,
                    weight: weight.clone(),
                    vocab,
                })
            },
            CheckpointProjectionWeight::Dense(weight) => Ok(Self::Dense {
                operation: backend.prepare_bf16_embedding(vocab, hidden, 1.0)?,
                weight: weight.clone(),
                vocab,
            }),
            CheckpointProjectionWeight::DirectFp8(_) => Err(Error::InvalidExecutionPlan(
                "shared routed embedding does not support direct FP8",
            )),
            CheckpointProjectionWeight::MxFp4(weight) => {
                weight.prepare_embedding(backend, 1.0).map(Self::MxFp4)
            },
            CheckpointProjectionWeight::MxFp8(_) => {
                Err(Error::InvalidExecutionPlan("shared routed embedding does not support MXFP8"))
            },
            CheckpointProjectionWeight::NvFp4(_)
            | CheckpointProjectionWeight::NvFp4WeightOnly(_) => {
                Err(Error::InvalidExecutionPlan("shared routed embedding does not support NVFP4"))
            },
            CheckpointProjectionWeight::PackedInteger(weight) => Ok(Self::PackedInteger {
                operation: packed_integer_embedding(backend, vocab, hidden, 1.0, weight.clone())?,
                vocab,
            }),
        }
    }

    pub(super) fn execute_batch(
        &self,
        selected: &DeviceBuffer<u32>,
        tokens: usize,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        match self {
            Self::Affine { operation, weight, .. } => {
                operation.execute_batch(selected, 0, tokens, weight.tensors(), output)
            },
            Self::Dense { operation, weight, .. } => {
                operation.execute_batch(selected, 0, tokens, weight, output)
            },
            Self::PackedInteger { operation, .. } => {
                operation.execute_batch(selected, 0, tokens, output)
            },
            Self::MxFp4(operation) => operation.execute_batch(selected, 0, tokens, output),
        }
    }

    pub(super) fn validate_token(&self, token: u32) -> Result<()> {
        if let Self::MxFp4(operation) = self {
            return operation.validate_token(token);
        }
        let vocab = match self {
            Self::Affine { vocab, .. }
            | Self::Dense { vocab, .. }
            | Self::PackedInteger { vocab, .. } => *vocab,
            Self::MxFp4(_) => unreachable!(),
        };
        if usize::try_from(token)? < vocab {
            Ok(())
        } else {
            Err(Error::InvalidToken { token, vocab })
        }
    }
}

impl SharedRoutedOutputHead {
    pub(super) fn new(
        backend: &CudaBackend,
        hidden: usize,
        vocab: usize,
        weight: &CheckpointProjectionWeight,
    ) -> Result<Self> {
        Ok(Self {
            operation: CheckpointProjection::new(
                backend,
                1,
                hidden,
                vocab,
                DenseRole::OutputHead,
                weight,
            )?,
            hidden,
            vocab,
        })
    }

    pub(super) fn execute(
        &mut self,
        input: &DeviceBuffer<bf16>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        if input.len() != self.hidden || output.len() != self.vocab {
            return Err(Error::InvalidDecoderKernel("shared-routed output-head buffer mismatch"));
        }
        self.operation.execute(input, output)
    }
}