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

use std::time::Duration;

use mircuda::{Context, DeviceBuffer, Stream, bf16};

use super::super::ClampedRoutedSplitDecode;
use crate::{
    Result,
    kernels::{ClampedRoutedAttention, SplitAttentionWorkspace, SplitPagedAttention},
};

#[allow(clippy::too_many_arguments)]
pub(super) fn measure_direct(
    split: &ClampedRoutedSplitDecode,
    direct: &ClampedRoutedAttention,
    stream: &Stream,
    query: &DeviceBuffer<bf16>,
    key_pages: &DeviceBuffer<u8>,
    value_pages: &DeviceBuffer<u8>,
    table: &DeviceBuffer<u32>,
    sinks: &DeviceBuffer<bf16>,
    output: &mut DeviceBuffer<bf16>,
    blocks: usize,
    window: Option<usize>,
    scale: f32,
    contexts: &[usize],
) -> Result<(Vec<Duration>, Duration)> {
    let mut timings = Vec::with_capacity(contexts.len());
    let mut consumed = Duration::ZERO;
    for &tokens in contexts {
        let (warmup, iterations) = split.backend.auto_tuner().iterations(tokens);
        for _ in 0..warmup {
            direct.execute(
                stream, query, query, query, key_pages, value_pages, table, sinks, output, tokens,
                blocks, window, scale,
            )?;
        }
        let average = measure(split.backend.context(), stream, iterations, || {
            direct.execute(
                stream, query, query, query, key_pages, value_pages, table, sinks, output, tokens,
                blocks, window, scale,
            )
        })?;
        consumed =
            consumed.saturating_add(average.saturating_mul(iterations.saturating_add(warmup)));
        timings.push(average);
    }
    Ok((timings, consumed))
}

#[allow(clippy::too_many_arguments)]
pub(super) fn measure_split(
    split: &ClampedRoutedSplitDecode,
    partition_tokens: usize,
    stream: &Stream,
    query: &DeviceBuffer<bf16>,
    key_pages: &DeviceBuffer<u8>,
    value_pages: &DeviceBuffer<u8>,
    table: &DeviceBuffer<u32>,
    sinks: &DeviceBuffer<bf16>,
    output: &mut DeviceBuffer<bf16>,
    blocks: usize,
    window: Option<usize>,
    scale: f32,
    contexts: &[usize],
) -> Result<(Vec<Duration>, Duration)> {
    let operation = SplitPagedAttention::compile(
        split.backend.compiler(),
        split.operation.spec(),
        partition_tokens,
    )?;
    let (values, statistics) = operation.workspace_lengths();
    let mut workspace = SplitAttentionWorkspace::new(
        split.backend.pool().allocate(stream, values)?,
        split.backend.pool().allocate(stream, statistics)?,
        split.backend.pool().allocate(stream, statistics)?,
    );
    let mut timings = Vec::with_capacity(contexts.len());
    let mut consumed = Duration::ZERO;
    for &tokens in contexts {
        let (warmup, iterations) = split.backend.auto_tuner().iterations(tokens);
        for _ in 0..warmup {
            execute_split(
                split, &operation, &mut workspace, stream, query, key_pages, value_pages, table,
                sinks, output, tokens, blocks, window, scale,
            )?;
        }
        let average = measure(split.backend.context(), stream, iterations, || {
            execute_split(
                split, &operation, &mut workspace, stream, query, key_pages, value_pages, table,
                sinks, output, tokens, blocks, window, scale,
            )
        })?;
        consumed =
            consumed.saturating_add(average.saturating_mul(iterations.saturating_add(warmup)));
        timings.push(average);
    }
    Ok((timings, consumed))
}

#[allow(clippy::too_many_arguments)]
fn execute_split(
    split: &ClampedRoutedSplitDecode,
    operation: &SplitPagedAttention,
    workspace: &mut SplitAttentionWorkspace,
    stream: &Stream,
    query: &DeviceBuffer<bf16>,
    key_pages: &DeviceBuffer<u8>,
    value_pages: &DeviceBuffer<u8>,
    table: &DeviceBuffer<u32>,
    sinks: &DeviceBuffer<bf16>,
    output: &mut DeviceBuffer<bf16>,
    tokens: usize,
    blocks: usize,
    window: Option<usize>,
    scale: f32,
) -> Result<()> {
    let active = operation.execute_partitions(
        stream, query, key_pages, value_pages, table, workspace, output, tokens, blocks, window,
        scale,
    )?;
    let query_heads = sinks.len();
    let head_dim = output.len() / query_heads;
    split.merge.launch(
        stream,
        operation.configs(active)?.merge,
        (
            &workspace.values,
            &workspace.maxima,
            &workspace.denominators,
            sinks,
            output,
            u32::try_from(query_heads)?,
            u32::try_from(head_dim)?,
            u32::try_from(active)?,
            u32::try_from(operation.max_partitions())?,
        ),
    )?;
    Ok(())
}

#[allow(clippy::cast_precision_loss)]
fn measure(
    context: &Context,
    stream: &Stream,
    iterations: u32,
    mut execute: impl FnMut() -> Result<()>,
) -> Result<Duration> {
    let started = context.create_event(true)?;
    let completed = context.create_event(true)?;
    started.record(stream)?;
    for _ in 0..iterations {
        execute()?;
    }
    completed.record(stream)?;
    completed.synchronize()?;
    Ok(Duration::from_secs_f32(
        started.elapsed_ms(&completed)? / (iterations as f32 * 1_000.0),
    ))
}