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
150
151
152
153
154
155
156
157
158
159
160

use mircuda::{DeviceBuffer, bf16};
use runtime::kv::BlockTable;

use super::{PagedAttentionBf16, pages::contiguous_page_token};
use crate::{PagedKvCache, Result};

impl PagedAttentionBf16 {
    /// Executes causal attention for every query token in a contiguous chunk.
    #[allow(clippy::too_many_arguments)]
    pub fn execute_prefill(
        &mut self,
        query: &DeviceBuffer<bf16>,
        cache: &PagedKvCache,
        table: &BlockTable,
        output: &mut DeviceBuffer<bf16>,
        query_tokens: usize,
        start_position: usize,
        window: Option<usize>,
        scale: f32,
    ) -> Result<()> {
        self.execute_prefill_masked(
            query, cache, table, output, query_tokens, start_position, window, scale, None,
        )
    }

    /// Executes causal prefill while opening one bidirectional image block.
    #[allow(clippy::too_many_arguments)]
    pub(in crate::backend) fn execute_prefill_masked(
        &mut self,
        query: &DeviceBuffer<bf16>,
        cache: &PagedKvCache,
        table: &BlockTable,
        output: &mut DeviceBuffer<bf16>,
        query_tokens: usize,
        start_position: usize,
        window: Option<usize>,
        scale: f32,
        image: Option<(usize, usize)>,
    ) -> Result<()> {
        self.validate(cache, table)?;
        self.update_table(table)?;
        let context = start_position
            .checked_add(query_tokens)
            .ok_or(crate::Error::InvalidPagedKv("prefill attention context overflow"))?;
        if window.is_none() && image.is_none() {
            if let (Some(first_page_token), Some(fmha)) =
                (contiguous_page_token(table, context), &self.fmha)
            {
                return Ok(fmha.execute(
                    &self.stream,
                    query,
                    cache.key_pages(),
                    cache.value_pages(),
                    output,
                    query_tokens,
                    context,
                    first_page_token,
                    scale,
                )?);
            }
            if self.fmha.is_some() {
                return self.execute_gathered_fmha(
                    query,
                    cache,
                    output,
                    query_tokens,
                    context,
                    table.blocks().len(),
                    scale,
                );
            }
        }
        self.prefill.execute(
            &self.stream,
            query,
            cache.key_pages(),
            cache.value_pages(),
            &self.table_device,
            output,
            query_tokens,
            start_position,
            table.blocks().len(),
            window,
            scale,
            image,
        )
    }

    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &mut self,
        query: &DeviceBuffer<bf16>,
        cache: &PagedKvCache,
        table: &BlockTable,
        output: &mut DeviceBuffer<bf16>,
        window: Option<usize>,
        scale: f32,
    ) -> Result<()> {
        self.ensure_tuned(query, cache, table, output, window, scale);
        let token_count = table.token_len();
        let visible = window.map_or(token_count, |limit| token_count.min(limit));
        if visible >= usize::try_from(self.split_threshold)? {
            return self.execute_split(query, cache, table, output, window, scale);
        }
        self.execute_direct(query, cache, table, output, window, scale)
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) fn execute_direct(
        &mut self,
        query: &DeviceBuffer<bf16>,
        cache: &PagedKvCache,
        table: &BlockTable,
        output: &mut DeviceBuffer<bf16>,
        window: Option<usize>,
        scale: f32,
    ) -> Result<()> {
        self.validate(cache, table)?;
        self.update_table(table)?;
        self.operation.execute(
            &self.stream,
            query,
            cache.key_pages(),
            cache.value_pages(),
            &self.table_device,
            output,
            table.token_len(),
            table.blocks().len(),
            window,
            scale,
        )
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) fn execute_split(
        &mut self,
        query: &DeviceBuffer<bf16>,
        cache: &PagedKvCache,
        table: &BlockTable,
        output: &mut DeviceBuffer<bf16>,
        window: Option<usize>,
        scale: f32,
    ) -> Result<()> {
        self.validate(cache, table)?;
        self.update_table(table)?;
        self.split.execute(
            &self.stream,
            query,
            cache.key_pages(),
            cache.value_pages(),
            &self.table_device,
            &mut self.split_workspace,
            output,
            table.token_len(),
            table.blocks().len(),
            window,
            scale,
        )
    }
}