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

use mircuda::{DeviceBuffer, KernelNode, bf16};
use runtime::kv::KvWritePlan;

use super::PagedKvCache;
use crate::{Error, PagedDecodeBatch, PagedPrefillBatch, Result, kernels::KvStoreKernel};

impl PagedKvCache {
    pub(crate) fn copy_ring_slot(&mut self, source_slot: usize, target_slot: usize) -> Result<()> {
        let ring = self.ring.ok_or(Error::InvalidPagedKv("cannot copy a non-windowed KV ring"))?;
        let physical_blocks = ring.physical_blocks()?;
        let source_blocks = ring.slot_blocks(source_slot)?;
        let target_blocks = ring.slot_blocks(target_slot)?;
        copy_slot(
            &self.stream,
            &mut self.key_pages,
            physical_blocks,
            source_blocks.clone(),
            target_blocks.start,
        )?;
        copy_slot(
            &self.stream,
            &mut self.value_pages,
            physical_blocks,
            source_blocks,
            target_blocks.start,
        )
    }

    pub(crate) fn store_captured(
        &mut self,
        plan: &KvWritePlan,
        keys: &DeviceBuffer<bf16>,
        values: &DeviceBuffer<bf16>,
    ) -> Result<KernelNode<KvStoreKernel>> {
        self.validate_plan(plan)?;
        let [write] = plan.writes() else {
            return Err(Error::InvalidPagedKv("captured decode requires one KV page write"));
        };
        self.operation.execute_captured(
            &self.stream,
            keys,
            values,
            &mut self.key_pages,
            &mut self.value_pages,
            write.local_start,
            write.token_count(),
            usize::try_from(write.page.block.0)?,
            write.page_start,
        )
    }

    pub(crate) fn store_batch(
        &mut self,
        batch: &PagedDecodeBatch,
        keys: &DeviceBuffer<bf16>,
        values: &DeviceBuffer<bf16>,
    ) -> Result<()> {
        if batch.cache_config() != self.storage.cache || batch.active() == 0 {
            return Err(Error::InvalidPagedKv("batched KV metadata geometry differs"));
        }
        self.operation.execute_batch(
            &self.stream,
            keys,
            values,
            &mut self.key_pages,
            &mut self.value_pages,
            batch.tables(),
            batch.token_counts(),
            batch.active(),
            batch.max_blocks(),
        )
    }

    pub fn store_prefill_batch(
        &mut self,
        batch: &PagedPrefillBatch,
        keys: &DeviceBuffer<bf16>,
        values: &DeviceBuffer<bf16>,
    ) -> Result<()> {
        if batch.cache_config() != self.storage.cache || batch.tokens() == 0 {
            return Err(Error::InvalidPagedKv("prefill KV metadata geometry differs"));
        }
        let windowed = self.is_windowed();
        self.operation.execute_prefill_batch(
            &self.stream,
            keys,
            values,
            &mut self.key_pages,
            &mut self.value_pages,
            batch.slot_mapping(windowed),
            batch.tokens(),
        )
    }

    pub(crate) fn store_for_session(
        &mut self,
        plan: &KvWritePlan,
        session_slot: usize,
        keys: &DeviceBuffer<bf16>,
        values: &DeviceBuffer<bf16>,
    ) -> Result<usize> {
        self.validate_plan(plan)?;
        for write in plan.writes() {
            let block = match self.ring {
                Some(ring) => ring.write_block(session_slot, *write)?,
                None => usize::try_from(write.page.block.0)?,
            };
            self.operation.execute(
                &self.stream,
                keys,
                values,
                &mut self.key_pages,
                &mut self.value_pages,
                write.local_start,
                write.token_count(),
                block,
                write.page_start,
            )?;
        }
        Ok(plan.written_tokens())
    }
}

fn copy_slot(
    stream: &mircuda::Stream,
    pages: &mut DeviceBuffer<u8>,
    physical_blocks: usize,
    source_blocks: std::ops::Range<usize>,
    target_block: usize,
) -> Result<()> {
    if physical_blocks == 0 || !pages.len().is_multiple_of(physical_blocks) {
        return Err(Error::InvalidPagedKv("windowed KV page geometry is not divisible"));
    }
    let bytes_per_block = pages.len() / physical_blocks;
    let source = source_blocks.start * bytes_per_block..source_blocks.end * bytes_per_block;
    let target = target_block * bytes_per_block;
    let alias = pages.clone();
    Ok(stream.copy_device_range(&alias, source, pages, target)?)
}