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
159
160
161
162
163
164
165
166
167
168
169
170
171

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

use super::{super::experts::ExpertWeights, PrefillMoeBlockBf16};
use crate::{
    DecodeMoeBlockBf16, DecodeMoeBlockWeights, PagedPrefillBatch, Result,
    backend::{attention::ImageAttentionSpan, block::scalar},
};

impl PrefillMoeBlockBf16 {
    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &mut self,
        state: &mut DecodeMoeBlockBf16,
        input: &DeviceBuffer<bf16>,
        weights: DecodeMoeBlockWeights<'_>,
        write_plan: &KvWritePlan,
        table: &BlockTable,
        start_position: usize,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.execute_masked(state, input, weights, write_plan, table, start_position, output, None)
    }

    #[allow(clippy::too_many_arguments)]
    pub(in crate::backend) fn execute_masked(
        &mut self,
        state: &mut DecodeMoeBlockBf16,
        input: &DeviceBuffer<bf16>,
        weights: DecodeMoeBlockWeights<'_>,
        write_plan: &KvWritePlan,
        table: &BlockTable,
        start_position: usize,
        output: &mut DeviceBuffer<bf16>,
        image: Option<ImageAttentionSpan>,
    ) -> Result<()> {
        self.attention.execute_masked(
            &mut state.attention,
            input,
            weights.attention,
            write_plan,
            table,
            start_position,
            &mut self.scratch.attention,
            image,
        )?;
        self.execute_tail(input, weights, &state.expert_weights, output)
    }

    pub fn execute_batch(
        &mut self,
        state: &mut DecodeMoeBlockBf16,
        input: &DeviceBuffer<bf16>,
        weights: DecodeMoeBlockWeights<'_>,
        batch: &PagedPrefillBatch,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.attention.execute_batch(
            &mut state.attention,
            input,
            weights.attention,
            batch,
            &mut self.scratch.attention,
        )?;
        self.execute_tail(input, weights, &state.expert_weights, output)
    }

    fn execute_tail(
        &mut self,
        input: &DeviceBuffer<bf16>,
        weights: DecodeMoeBlockWeights<'_>,
        expert_weights: &ExpertWeights,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.post_attention_norm.execute(
            &self.scratch.attention,
            weights.post_attention_norm,
            &mut self.scratch.attention_norm,
        )?;
        self.hidden_ops.add(
            &self.stream,
            input,
            &self.scratch.attention_norm,
            &mut self.scratch.hidden,
        )?;
        self.execute_dense(weights)?;
        self.execute_experts(weights, expert_weights)?;
        self.hidden_ops.add(
            &self.stream,
            &self.scratch.dense,
            &self.scratch.expert_norm,
            &mut self.scratch.feed_forward,
        )?;
        self.post_feed_forward_norm.execute(
            &self.scratch.feed_forward,
            weights.post_feed_forward_norm,
            &mut self.scratch.feed_forward_norm,
        )?;
        self.hidden_ops.add(
            &self.stream,
            &self.scratch.hidden,
            &self.scratch.feed_forward_norm,
            &mut self.scratch.residual,
        )?;
        self.hidden_ops.multiply_scalar(
            &self.stream,
            &self.scratch.residual,
            scalar(weights.layer_scalar)?,
            output,
        )
    }

    fn execute_dense(&mut self, weights: DecodeMoeBlockWeights<'_>) -> Result<()> {
        self.pre_dense_norm.execute(
            &self.scratch.hidden,
            weights.pre_dense_norm,
            &mut self.scratch.normalized,
        )?;
        self.dense_gate_up.execute(
            &self.scratch.normalized,
            weights.dense_gate_up,
            &mut self.scratch.dense_gate_up,
        )?;
        self.dense_activation.execute(
            &self.stream,
            &self.scratch.dense_gate_up,
            &mut self.scratch.dense_activated,
            self.config.activation.into(),
        )?;
        self.dense_down.execute(
            &self.scratch.dense_activated,
            weights.dense_down,
            &mut self.scratch.normalized,
        )?;
        self.post_dense_norm.execute(
            &self.scratch.normalized,
            weights.post_dense_norm,
            &mut self.scratch.dense,
        )
    }

    fn execute_experts(
        &mut self,
        weights: DecodeMoeBlockWeights<'_>,
        expert_weights: &ExpertWeights,
    ) -> Result<()> {
        let selection = self.router.execute(&self.scratch.hidden, weights.router)?;
        self.pre_expert_norm.execute(
            &self.scratch.hidden,
            weights.pre_expert_norm,
            &mut self.scratch.normalized,
        )?;
        self.experts.execute(
            &self.scratch.normalized,
            selection.indices,
            selection.weights,
            expert_weights,
            &mut self.scratch.expert,
        )?;
        self.post_expert_norm.execute(
            &self.scratch.expert,
            weights.post_expert_norm,
            &mut self.scratch.expert_norm,
        )
    }

    #[must_use]
    pub const fn tokens(&self) -> usize {
        self.tokens
    }
}