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

use mircuda::{DeviceBuffer, bf16};
use runtime::{
    backend::{DecodeRequest, DecodeSequence, PrefillRequest},
    kv::BlockTable,
};
use uuid::Uuid;

use crate::{CudaBackend, Error, Result, engine::execution::Output};

mod graph;
mod session;

pub(in crate::engine) use graph::GraphExecution;
pub(in crate::engine) use session::{MixedMixerExecution, SinkAttentionExecution};

pub(in crate::engine) trait GenerationExecution: Send {
    fn prefix_replay_tokens(&self) -> Option<usize> {
        None
    }

    // Admission may assume a retained checkpoint; replay remains the
    // correctness fallback when that checkpoint is absent.
    fn prefix_admission(&self) -> Option<(usize, usize, usize)> {
        self.prefix_replay_tokens().map(|replay| (replay, replay, 0))
    }

    fn restore_prefix(
        &mut self,
        _request: &PrefillRequest,
        _minimum: usize,
        _maximum: usize,
    ) -> Result<Option<usize>> {
        Ok(None)
    }

    fn prefill_chunk_len(&self, remaining: usize) -> usize;

    fn prefill_chunk(
        &mut self,
        backend: &CudaBackend,
        request: &PrefillRequest,
        tokens: &[u32],
        offset: usize,
        table: &BlockTable,
        final_chunk: bool,
    ) -> Result<Option<Output>>;

    fn prefill_batch_chunk(
        &mut self,
        backend: &CudaBackend,
        chunks: &[PrefillChunk<'_>],
    ) -> Result<Vec<Option<Output>>> {
        chunks
            .iter()
            .map(|chunk| {
                self.prefill_chunk(
                    backend,
                    chunk.request,
                    chunk.tokens,
                    chunk.offset,
                    chunk.table,
                    chunk.final_chunk,
                )
            })
            .collect()
    }

    fn decode(
        &mut self,
        backend: &CudaBackend,
        request: &DecodeRequest,
        use_device_token: bool,
    ) -> Result<Output>;

    fn decode_batch(
        &mut self,
        _backend: &CudaBackend,
        _sequences: &[DecodeSequence],
    ) -> Result<Option<Vec<Output>>> {
        Ok(None)
    }

    fn clear_sessions(&mut self);

    fn release_session(&mut self, session: Uuid);

    fn prefill_pooled_vision(
        &mut self,
        _backend: &CudaBackend,
        _input: PooledVisionPrefill<'_>,
    ) -> Result<Output> {
        Err(unsupported("pooled vision prefill"))
    }

    fn prefill_spatial_vision(
        &mut self,
        _backend: &CudaBackend,
        _input: SpatialVisionPrefill<'_>,
    ) -> Result<Output> {
        Err(unsupported("spatial-merge vision prefill"))
    }
}

pub(in crate::engine) struct PrefillChunk<'a> {
    pub request: &'a PrefillRequest,
    pub tokens: &'a [u32],
    pub offset: usize,
    pub table: &'a BlockTable,
    pub final_chunk: bool,
}

pub(in crate::engine) struct PooledVisionPrefill<'a> {
    pub session: Uuid,
    pub tokens: &'a [u32],
    pub image: &'a DeviceBuffer<bf16>,
    pub image_span: (usize, usize),
    pub bidirectional: bool,
    pub table: &'a BlockTable,
    pub sampling: runtime::backend::SamplingLogits,
}

pub(in crate::engine) struct SpatialVisionPrefill<'a> {
    pub session: Uuid,
    pub tokens: &'a [u32],
    pub positions: &'a [u32],
    pub image: &'a DeviceBuffer<bf16>,
    pub image_span: (usize, usize),
    pub position_delta: i32,
    pub table: &'a BlockTable,
    pub sampling: runtime::backend::SamplingLogits,
}

fn unsupported(operation: &'static str) -> Error {
    Error::MissingCapability {
        operation,
        storage: "lowered CUDA generation execution".into(),
        geometry: "model execution boundary".into(),
        requirement: "the lowered mixer plan must implement this multimodal operation",
    }
}