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

use mircuda::{
    CompileOptions, Compiler, DeviceBuffer, LaunchConfig, Stream, TypedKernel, cuda_export,
    cuda_kernel_file,
};

use super::{PagedAttentionSpec, validate_attention};
use crate::{
    Result,
    kernels::geometry::{narrow, product, require},
};

cuda_export!(PagedKvGatherKernel = "libmir_cuda_gather_paged_kv_bf16"(
    key_pages: &DeviceBuffer<u8>, value_pages: &DeviceBuffer<u8>,
    block_table: &DeviceBuffer<u32>, keys: &mut DeviceBuffer<u8>,
    values: &mut DeviceBuffer<u8>, context_tokens: u32, block_count: u32,
    block_size: u32, key_width: u32, value_width: u32,
));

cuda_export!(BatchedPagedKvGatherKernel = "libmir_cuda_gather_paged_kv_batch_bf16"(
    key_pages: &DeviceBuffer<u8>, value_pages: &DeviceBuffer<u8>,
    block_tables: &DeviceBuffer<u32>, context_starts: &DeviceBuffer<u32>,
    keys: &mut DeviceBuffer<mircuda::bf16>, values: &mut DeviceBuffer<mircuda::bf16>,
    batch_size: u32, max_blocks: u32, block_size: u32,
    key_width: u32, value_width: u32,
));

/// Gathers logical BF16 K/V rows from physical pages into contiguous buffers.
#[derive(Clone, Debug)]
pub struct PagedKvGather {
    kernel: TypedKernel<PagedKvGatherKernel>,
    spec: PagedAttentionSpec,
}

#[derive(Clone, Debug)]
pub struct BatchedPagedKvGather {
    kernel: TypedKernel<BatchedPagedKvGatherKernel>,
    spec: PagedAttentionSpec,
}

impl PagedKvGather {
    pub fn compile(compiler: &Compiler, spec: PagedAttentionSpec) -> Result<Self> {
        validate_attention(spec)?;
        let source = cuda_kernel_file!("../../../kernels/paged_gather_bf16.cu");
        let module = compiler.compile(source, &CompileOptions::default())?;
        Ok(Self { kernel: module.kernel()?, spec })
    }

    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &self,
        stream: &Stream,
        key_pages: &DeviceBuffer<u8>,
        value_pages: &DeviceBuffer<u8>,
        block_table: &DeviceBuffer<u32>,
        keys: &mut DeviceBuffer<u8>,
        values: &mut DeviceBuffer<u8>,
        context_tokens: usize,
        block_count: usize,
    ) -> Result<()> {
        let key_width = product(self.spec.kv_heads, self.spec.head_dim)?;
        let value_width = product(self.spec.kv_heads, self.spec.value_head_dim)?;
        require(
            "gathered prefill keys",
            product(product(context_tokens, key_width)?, size_of::<u16>())?,
            keys.len(),
        )?;
        require(
            "gathered prefill values",
            product(product(context_tokens, value_width)?, size_of::<u16>())?,
            values.len(),
        )?;
        require("gathered prefill table", self.spec.max_blocks, block_table.len())?;
        Ok(self.kernel.launch(
            stream,
            LaunchConfig {
                grid: (narrow(context_tokens)?, 1, 1),
                block: (256, 1, 1),
                shared_memory_bytes: 0,
            },
            (
                key_pages,
                value_pages,
                block_table,
                keys,
                values,
                narrow(context_tokens)?,
                narrow(block_count)?,
                narrow(self.spec.block_size)?,
                narrow(key_width)?,
                narrow(value_width)?,
            ),
        )?)
    }
}

impl BatchedPagedKvGather {
    pub fn compile(compiler: &Compiler, spec: PagedAttentionSpec) -> Result<Self> {
        validate_attention(spec)?;
        let source = cuda_kernel_file!("../../../kernels/paged_gather_bf16.cu");
        let module = compiler.compile(source, &CompileOptions::default())?;
        Ok(Self { kernel: module.kernel()?, spec })
    }

    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &self,
        stream: &Stream,
        key_pages: &DeviceBuffer<u8>,
        value_pages: &DeviceBuffer<u8>,
        block_tables: &DeviceBuffer<u32>,
        context_starts: &DeviceBuffer<u32>,
        keys: &mut DeviceBuffer<mircuda::bf16>,
        values: &mut DeviceBuffer<mircuda::bf16>,
        batch_size: usize,
        total_context_tokens: usize,
        max_context_tokens: usize,
    ) -> Result<()> {
        let key_width = product(self.spec.kv_heads, self.spec.head_dim)?;
        let value_width = product(self.spec.kv_heads, self.spec.value_head_dim)?;
        require(
            "batched gathered prefill keys",
            product(total_context_tokens, key_width)?,
            keys.len(),
        )?;
        require(
            "batched gathered prefill values",
            product(total_context_tokens, value_width)?,
            values.len(),
        )?;
        require(
            "batched gathered prefill tables",
            product(batch_size, self.spec.max_blocks)?,
            block_tables.len(),
        )?;
        require("batched gathered prefill starts", batch_size + 1, context_starts.len())?;
        Ok(self.kernel.launch(
            stream,
            LaunchConfig {
                grid: (narrow(max_context_tokens)?, narrow(batch_size)?, 1),
                block: (256, 1, 1),
                shared_memory_bytes: 0,
            },
            (
                key_pages,
                value_pages,
                block_tables,
                context_starts,
                keys,
                values,
                narrow(batch_size)?,
                narrow(self.spec.max_blocks)?,
                narrow(self.spec.block_size)?,
                narrow(key_width)?,
                narrow(value_width)?,
            ),
        )?)
    }
}