libmir-cuda 0.1.0

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

use mircuda::{DeviceBuffer, PinnedBuffer, bf16};

use super::CudaMoeModelTemplate;
use crate::{
    CudaDecodeBatch, CudaModelSessionConfig, CudaMoeModelSession, Error, Result,
    backend::model::{
        batch::DecodeBatchSource,
        layer::{DecoderLayerTemplate, SessionLayer},
        prefill::PrefillTokenBuffer,
    },
};

impl CudaMoeModelTemplate {
    /// Creates private activation, plan, K/V, and graph state for one session.
    pub fn instantiate(&self) -> Result<CudaMoeModelSession> {
        self.instantiate_with_config(CudaModelSessionConfig::default())
    }

    /// Creates one preallocated decode bucket without copying model weights.
    pub fn instantiate_decode_batch(&self, batch_size: usize) -> Result<CudaDecodeBatch> {
        let caches = self.allocate_shared_kv()?;
        self.instantiate_decode_batch_with_caches(batch_size, &caches)
    }

    pub(crate) fn instantiate_decode_batch_with_caches(
        &self,
        batch_size: usize,
        caches: &[crate::PagedKvCache],
    ) -> Result<CudaDecodeBatch> {
        let attention = self
            .layers
            .first()
            .ok_or(Error::InvalidDecoderKernel("CUDA model template requires layers"))?
            .attention();
        CudaDecodeBatch::new(
            &self.backend,
            DecodeBatchSource {
                embedding: self.embedding.clone(),
                final_norm: self.final_norm.clone(),
                output_head: &self.output_head,
                layers: &self.layers,
                caches,
                attention,
                vocab: self.decoder.vocab_size,
                hidden: self.decoder.hidden_size,
                embedding_scale: self.embedding_scale,
            },
            batch_size,
        )
    }

    /// Creates one session with an explicit prefill allocation policy.
    pub fn instantiate_with_config(
        &self,
        config: CudaModelSessionConfig,
    ) -> Result<CudaMoeModelSession> {
        let caches = self.allocate_shared_kv()?;
        self.instantiate_with_config_and_caches(config, &caches)
    }

    pub(crate) fn instantiate_with_config_and_caches(
        &self,
        config: CudaModelSessionConfig,
        caches: &[crate::PagedKvCache],
    ) -> Result<CudaMoeModelSession> {
        let config = config.validate()?;
        let hidden = self.decoder.hidden_size;
        let vocab = self.decoder.vocab_size;
        let first = self.backend.inner.pool.allocate::<bf16>(&self.backend.inner.stream, hidden)?;
        let second =
            self.backend.inner.pool.allocate::<bf16>(&self.backend.inner.stream, hidden)?;
        let layers = instantiate_layers(&self.layers, caches, &first, &second)?;
        let prefill_elements = config
            .prefill_chunk_tokens
            .checked_mul(hidden)
            .ok_or(Error::InvalidDecoderKernel("CUDA prefill activation size overflow"))?;
        let prefill_first = self
            .backend
            .inner
            .pool
            .allocate::<bf16>(&self.backend.inner.stream, prefill_elements)?;
        let prefill_second = self
            .backend
            .inner
            .pool
            .allocate::<bf16>(&self.backend.inner.stream, prefill_elements)?;
        let logits = self.backend.inner.pool.allocate::<bf16>(&self.backend.inner.stream, vocab)?;
        let input_token = self.backend.inner.pool.allocate::<u32>(&self.backend.inner.stream, 1)?;
        let token_staging: PinnedBuffer<u32> = self.backend.inner.context.allocate_pinned(1)?;
        let epsilon = self
            .layers
            .first()
            .ok_or(Error::InvalidDecoderKernel("CUDA model template requires layers"))?
            .attention()
            .rms_norm_epsilon;
        CudaMoeModelSession::new(
            self.backend.prepare_bf16_embedding(vocab, hidden, self.embedding_scale)?,
            self.backend.prepare_rms_norm_bf16(1, hidden, epsilon)?,
            self.output_head.instantiate(&self.backend)?,
            self.embedding.clone(),
            self.final_norm.clone(),
            layers,
            self.backend.prepare_device_sampler_bf16(vocab)?,
            self.backend.inner.stream.clone(),
            input_token,
            token_staging,
            PrefillTokenBuffer::new(&self.backend, config)?,
            crate::kernels::SelectRowBf16::compile(&self.backend.inner.compiler, hidden)?,
            first,
            second,
            prefill_first,
            prefill_second,
            logits,
        )
    }
}

fn instantiate_layers(
    templates: &[DecoderLayerTemplate],
    caches: &[crate::PagedKvCache],
    first: &DeviceBuffer<bf16>,
    second: &DeviceBuffer<bf16>,
) -> Result<Vec<SessionLayer>> {
    if templates.len() != caches.len() {
        return Err(Error::InvalidDecoderKernel("CUDA session cache count differs from layers"));
    }
    templates
        .iter()
        .zip(caches)
        .enumerate()
        .map(|(index, (template, cache))| {
            let (input, output) = if index.is_multiple_of(2) {
                (first, second)
            } else {
                (second, first)
            };
            template.instantiate(input, output, cache.clone())
        })
        .collect()
}