mircuda 0.3.0

Native, explicit-stream Rust gateway to NVIDIA CUDA
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

use crate::{Context, DeviceBuffer, Error, Result, Stream, bf16};

mod paged;

/// Head geometry supported by the BF16 CUTLASS fused-attention plan.
#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
pub struct FmhaBf16Spec {
    query_heads: usize,
    kv_heads: usize,
    head_dim: usize,
    value_head_dim: usize,
}

impl FmhaBf16Spec {
    /// Creates a GQA/MHA fused-attention geometry.
    pub const fn new(
        query_heads: usize,
        kv_heads: usize,
        head_dim: usize,
        value_head_dim: usize,
    ) -> Result<Self> {
        if query_heads == 0
            || kv_heads == 0
            || !query_heads.is_multiple_of(kv_heads)
            || head_dim != value_head_dim
            || !matches!(head_dim, 64 | 128)
        {
            return Err(Error::InvalidMatmulShape);
        }
        Ok(Self {
            query_heads,
            kv_heads,
            head_dim,
            value_head_dim,
        })
    }

    /// Query-head count.
    #[must_use]
    pub const fn query_heads(self) -> usize {
        self.query_heads
    }

    /// K/V-head count.
    #[must_use]
    pub const fn kv_heads(self) -> usize {
        self.kv_heads
    }
}

/// Stream-bound BF16 fused-attention plan using a bottom-right causal mask.
#[derive(Debug)]
pub struct FmhaBf16Plan {
    native: mircuda_sys::FmhaBf16Plan,
    spec: FmhaBf16Spec,
}

impl FmhaBf16Plan {
    /// Creates an asynchronous fixed-head plan.
    pub fn new(context: &Context, stream: &Stream, spec: FmhaBf16Spec) -> Result<Self> {
        let native = context.native.create_fmha_bf16_plan(
            &stream.native,
            mircuda_sys::FmhaBf16Spec {
                query_heads: spec.query_heads,
                kv_heads: spec.kv_heads,
                head_dim: spec.head_dim,
                value_head_dim: spec.value_head_dim,
            },
        )?;
        Ok(Self { native, spec })
    }

    /// Enqueues causal attention for queries aligned to the end of the K/V context.
    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &self,
        stream: &Stream,
        query: &DeviceBuffer<bf16>,
        key_pages: &DeviceBuffer<u8>,
        value_pages: &DeviceBuffer<u8>,
        output: &mut DeviceBuffer<bf16>,
        query_tokens: usize,
        context_tokens: usize,
        first_page_token: usize,
        scale: f32,
    ) -> Result<()> {
        let key_row = self.spec.kv_heads * self.spec.head_dim;
        let value_row = self.spec.kv_heads * self.spec.value_head_dim;
        let key_offset = first_page_token
            .checked_mul(key_row)
            .and_then(|value| value.checked_mul(size_of::<bf16>()))
            .ok_or(Error::InvalidMatmulShape)?;
        let value_offset = first_page_token
            .checked_mul(value_row)
            .and_then(|value| value.checked_mul(size_of::<bf16>()))
            .ok_or(Error::InvalidMatmulShape)?;
        Ok(self.native.execute(
            &stream.native,
            &query.native,
            &key_pages.native,
            &value_pages.native,
            &output.native,
            query_tokens,
            context_tokens,
            key_offset,
            value_offset,
            scale,
        )?)
    }

    /// Enqueues one bottom-right causal attention launch over packed variable-length rows.
    #[allow(clippy::too_many_arguments)]
    pub fn execute_varlen(
        &self,
        stream: &Stream,
        query: &DeviceBuffer<bf16>,
        keys: &DeviceBuffer<bf16>,
        values: &DeviceBuffer<bf16>,
        output: &mut DeviceBuffer<bf16>,
        query_starts: &DeviceBuffer<u32>,
        key_starts: &DeviceBuffer<u32>,
        batch_size: usize,
        total_query_tokens: usize,
        total_context_tokens: usize,
        max_query_tokens: usize,
        max_context_tokens: usize,
        scale: f32,
    ) -> Result<()> {
        Ok(self.native.execute_varlen(
            &stream.native,
            &query.native,
            &keys.native,
            &values.native,
            &output.native,
            &query_starts.native,
            &key_starts.native,
            batch_size,
            total_query_tokens,
            total_context_tokens,
            max_query_tokens,
            max_context_tokens,
            scale,
        )?)
    }
}