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
140
141
142
143
144
145
146
147
148
149
150

use std::collections::HashMap;

use uuid::Uuid;

use super::{GenerationExecution, SpatialVisionPrefill};
use crate::{
    CudaBackend, CudaClampedRoutedModelSession, CudaClampedRoutedModelTemplate,
    CudaSharedRoutedModelSession, CudaSharedRoutedModelTemplate, Error, Result,
    engine::execution::{Output, generation_output},
};

pub(in crate::engine) struct MixedMixerExecution {
    template: CudaSharedRoutedModelTemplate,
    sessions: HashMap<Uuid, CudaSharedRoutedModelSession>,
}

pub(in crate::engine) struct SinkAttentionExecution {
    template: CudaClampedRoutedModelTemplate,
    sessions: HashMap<Uuid, CudaClampedRoutedModelSession>,
}

impl MixedMixerExecution {
    pub(in crate::engine) fn new(template: CudaSharedRoutedModelTemplate) -> Self {
        Self { template, sessions: HashMap::new() }
    }
}

impl SinkAttentionExecution {
    pub(in crate::engine) fn new(template: CudaClampedRoutedModelTemplate) -> Self {
        Self { template, sessions: HashMap::new() }
    }
}

impl GenerationExecution for MixedMixerExecution {
    fn prefill_chunk_len(&self, remaining: usize) -> usize {
        remaining
    }

    fn prefill_chunk(
        &mut self,
        backend: &CudaBackend,
        request: &runtime::backend::PrefillRequest,
        tokens: &[u32],
        offset: usize,
        table: &runtime::kv::BlockTable,
        final_chunk: bool,
    ) -> Result<Option<Output>> {
        if offset == 0 {
            self.sessions.insert(request.session_id, self.template.instantiate()?);
        }
        let session = required(&mut self.sessions, request.session_id)?;
        session.prefill(request.session_id, tokens, table)?;
        if final_chunk {
            generation_output(backend, session, request.sampling_logits).map(Some)
        } else {
            Ok(None)
        }
    }

    fn decode(
        &mut self,
        backend: &CudaBackend,
        request: &runtime::backend::DecodeRequest,
        _use_device_token: bool,
    ) -> Result<Output> {
        let session = required(&mut self.sessions, request.session_id)?;
        session.decode(request.session_id, request.token_id, &request.block_table)?;
        generation_output(backend, session, request.sampling_logits)
    }

    fn clear_sessions(&mut self) {
        self.sessions.clear();
    }

    fn release_session(&mut self, session: Uuid) {
        self.sessions.remove(&session);
    }

    fn prefill_spatial_vision(
        &mut self,
        backend: &CudaBackend,
        input: SpatialVisionPrefill<'_>,
    ) -> Result<Output> {
        let mut session = self.template.instantiate()?;
        session.prefill_vision(
            input.session,
            input.tokens,
            input.positions,
            input.table,
            input.image_span,
            input.image,
            input.position_delta,
        )?;
        let output = generation_output(backend, &mut session, input.sampling)?;
        self.sessions.insert(input.session, session);
        Ok(output)
    }
}

impl GenerationExecution for SinkAttentionExecution {
    fn prefill_chunk_len(&self, remaining: usize) -> usize {
        remaining.min(256)
    }

    fn prefill_chunk(
        &mut self,
        backend: &CudaBackend,
        request: &runtime::backend::PrefillRequest,
        tokens: &[u32],
        offset: usize,
        table: &runtime::kv::BlockTable,
        final_chunk: bool,
    ) -> Result<Option<Output>> {
        if offset == 0 {
            self.sessions.insert(request.session_id, self.template.instantiate()?);
        }
        let session = required(&mut self.sessions, request.session_id)?;
        session.prefill(request.session_id, tokens, table)?;
        if final_chunk {
            generation_output(backend, session, request.sampling_logits).map(Some)
        } else {
            Ok(None)
        }
    }

    fn decode(
        &mut self,
        backend: &CudaBackend,
        request: &runtime::backend::DecodeRequest,
        _use_device_token: bool,
    ) -> Result<Output> {
        let session = required(&mut self.sessions, request.session_id)?;
        session.decode(request.session_id, request.token_id, &request.block_table)?;
        generation_output(backend, session, request.sampling_logits)
    }

    fn clear_sessions(&mut self) {
        self.sessions.clear();
    }

    fn release_session(&mut self, session: Uuid) {
        self.sessions.remove(&session);
    }
}

fn required<T>(sessions: &mut HashMap<Uuid, T>, id: Uuid) -> Result<&mut T> {
    sessions
        .get_mut(&id)
        .ok_or_else(|| Error::State("CUDA decoder session is not initialized".into()))
}