cubek-attention 0.2.0

CubeK: Attention Kernels
Documentation 
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

use cubecl::{Runtime, client::ComputeClient, prelude::TensorBinding};

use crate::routines::{
    Routine, blackbox_accelerated::BlackboxAcceleratedRoutine, unit::UnitRoutine,
};
use crate::{
    definition::AttentionSetupError,
    definition::{AttentionDims, AttentionGlobalTypes, AttentionOptions, AttentionProblem},
    launch::args::{TensorArgs, TensorInputsLaunch},
    routines::DeviceSettings,
};

use crate::components::batch::BatchAttentionFamily;

#[derive(Debug, Clone)]
pub enum BlueprintStrategy<R: Routine> {
    /// Use a predefined blueprint
    Forced(R::Blueprint),
    /// Allows to give limited settings information, and the rest is inferred from it
    Inferred(R::Strategy),
}

#[derive(Debug, Clone)]
pub enum Strategy {
    BlackboxAccelerated(BlueprintStrategy<BlackboxAcceleratedRoutine>),
    Unit(BlueprintStrategy<UnitRoutine>),
}

#[allow(clippy::result_large_err, clippy::too_many_arguments)]
pub fn launch_ref<R: Runtime>(
    strategy: Strategy,
    client: &ComputeClient<R>,
    query: TensorBinding<R>,
    key: TensorBinding<R>,
    value: TensorBinding<R>,
    mask: Option<TensorBinding<R>>,
    out: TensorBinding<R>,
    attention_global_types: &AttentionGlobalTypes,
    attention_options: AttentionOptions,
) -> Result<(), AttentionSetupError> {
    match strategy {
        Strategy::BlackboxAccelerated(strategy) => {
            launch_attention::<R, BlackboxAcceleratedRoutine>(
                client,
                query,
                key,
                value,
                mask,
                out,
                attention_global_types,
                strategy,
                attention_options,
            )
        }
        Strategy::Unit(strategy) => launch_attention::<R, UnitRoutine>(
            client,
            query,
            key,
            value,
            mask,
            out,
            attention_global_types,
            strategy,
            attention_options,
        ),
    }
}

#[allow(clippy::too_many_arguments)]
pub fn launch_attention<R: Runtime, A: Routine>(
    client: &ComputeClient<R>,
    query: TensorBinding<R>,
    key: TensorBinding<R>,
    value: TensorBinding<R>,
    mask: Option<TensorBinding<R>>,
    out: TensorBinding<R>,
    global_dtypes: &AttentionGlobalTypes,
    strategy: BlueprintStrategy<A>,
    attention_options: AttentionOptions,
) -> Result<(), AttentionSetupError> {
    let definition = AttentionProblem {
        dims: AttentionDims {
            batch: query.shape[0],
            num_heads: query.shape[1],
            seq_q: query.shape[2],
            head_dim: query.shape[3],
            seq_kv: key.shape[2],
            val_dim: value.shape[3],
        },
        masked: mask.is_some(),
        global_dtypes: global_dtypes.clone(),
        options: attention_options,
        address_type: [
            query.required_address_type(global_dtypes.query.size()),
            key.required_address_type(global_dtypes.key.size()),
            value.required_address_type(global_dtypes.value.size()),
            mask.as_ref()
                .map(|mask| mask.required_address_type(global_dtypes.mask.size()))
                .unwrap_or_default(),
            out.required_address_type(global_dtypes.out.size()),
        ]
        .into_iter()
        .max()
        .unwrap_or_default(),
    };

    let device_settings = DeviceSettings::new(client, &definition);

    let launch_info = A::prepare(&definition, &device_settings, strategy)?;

    // This allows an expand_config error to be caught by the client rather than the server.
    // Then the server can re-run expand config assuming a valid blueprint
    <A as Routine>::BatchAttention::expand_config(
        client.properties(),
        launch_info.blueprint.clone(),
        &launch_info.dtypes,
    )?;

    let result = unsafe {
        <A as Routine>::BatchAttention::launch_unchecked::<TensorArgs, R>(
            client,
            launch_info.cube_dim,
            launch_info.cube_count_plan.resolve(),
            launch_info.address_type,
            TensorInputsLaunch::new(
                query.into_tensor_arg(),
                key.into_tensor_arg(),
                value.into_tensor_arg(),
                mask.map(|it| it.into_tensor_arg()).into(),
            ),
            out.into_tensor_arg(),
            cubek_std::cube_count::cube_mapping_launch(&launch_info.cube_count_plan),
            &launch_info.dtypes,
            &device_settings.vector_sizes,
            launch_info.blueprint,
        )
    };

    match result {
        Ok(_) => Ok(()),
        Err(err) => Err(AttentionSetupError::Execution(err)),
    }
}