libmir-metal 0.1.0

Metal 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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187

use models::layout::AttentionLayerType;

use super::*;

#[test]
fn keeps_recurrent_gated_delta_state_on_the_explicit_gpu_stream() -> Result<()> {
    let stream = Stream::new_gpu()?;
    let mut state = GatedDeltaState::new()?;
    let inputs = inputs(&[2.0, 4.0, 3.0, 5.0], 2)?;
    let output = state.update(inputs.as_borrowed(), &stream)?;
    output.async_eval()?;
    stream.synchronize()?;

    assert_close(&output.to_vec_f32()?, &[1.0, 2.0, 1.75, 3.0]);
    assert_eq!(state.offset()?, 2);
    assert_close(&state.values()?.to_vec_f32_on_stream(&stream)?, &[1.75, 3.0]);
    Ok(())
}

#[test]
fn uses_the_metal_recurrence_kernel_for_supported_key_dimensions() -> Result<()> {
    let stream = Stream::new_gpu()?;
    let mut state = GatedDeltaState::new()?;
    let mut query_and_key = vec![0.0; 64];
    query_and_key[0] = 1.0;
    query_and_key[32] = 1.0;
    let inputs = GatedDeltaInputs {
        query: &Array::from_f32(&query_and_key, &[1, 2, 1, 32])?,
        key: &Array::from_f32(&query_and_key, &[1, 2, 1, 32])?,
        value: &Array::from_f32(&[2.0, 4.0], &[1, 2, 1, 1])?,
        alpha: &Array::from_f32(&[0.0, 0.0], &[1, 2, 1])?,
        beta: &Array::from_f32(&[0.0, 0.0], &[1, 2, 1])?,
        a_log: &Array::from_f32(&[0.0], &[1])?,
        dt_bias: &Array::from_f32(&[0.0], &[1])?,
    };
    let output = state.update(inputs, &stream)?;

    output.async_eval()?;
    stream.synchronize()?;
    assert_close(&output.to_vec_f32()?, &[1.0, 2.25]);
    assert_close(&state.values()?.to_vec_f32_on_stream(&stream)?[..1], &[2.25]);
    Ok(())
}

#[test]
fn fused_decode_matches_the_general_recurrence_step() -> Result<()> {
    let stream = Stream::new_gpu()?;
    let mut general = GatedDeltaState::new()?;
    let mut fused = GatedDeltaState::new()?;
    for step in 0..2 {
        let general_inputs = decode_inputs(step)?;
        let fused_inputs = decode_inputs(step)?;
        let general_output = general.update(general_inputs.as_borrowed(), &stream)?;
        let fused_output = fused.update_fused(fused_inputs.as_borrowed(), &stream)?;
        general_output.async_eval()?;
        fused_output.async_eval()?;
        stream.synchronize()?;
        assert_eq!(general_output.to_vec_f32()?, fused_output.to_vec_f32()?);
        assert_eq!(
            general.values()?.to_vec_f32_on_stream(&stream)?,
            fused.values()?.to_vec_f32_on_stream(&stream)?,
        );
    }
    Ok(())
}

#[test]
fn snapshots_recurrent_state_without_a_host_round_trip() -> Result<()> {
    let stream = Stream::new_gpu()?;
    let mut original = GatedDeltaState::new()?;
    let first = inputs(&[2.0, 4.0], 1)?;
    let _first_output = original.update(first.as_borrowed(), &stream)?;
    let mut snapshot = original.snapshot()?;
    let second = inputs(&[3.0, 5.0], 1)?;
    let original_output = original.update(second.as_borrowed(), &stream)?;
    let snapshot_inputs = inputs(&[3.0, 5.0], 1)?;
    let snapshot_output = snapshot.update(snapshot_inputs.as_borrowed(), &stream)?;

    original_output.async_eval()?;
    snapshot_output.async_eval()?;
    stream.synchronize()?;
    assert_close(&original_output.to_vec_f32()?, &snapshot_output.to_vec_f32()?);
    assert_eq!(original.offset()?, 2);
    assert_eq!(snapshot.offset()?, 2);
    Ok(())
}

#[test]
fn snapshots_mixed_linear_and_key_value_cache_at_a_common_offset() -> Result<()> {
    let mut cache = DecoderCache::new_hybrid_linear(
        &[AttentionLayerType::Linear, AttentionLayerType::Full],
        16,
    )?;

    assert_eq!(cache.cached_tokens()?, 0);
    assert_eq!(cache.snapshot_at(0)?.cached_tokens()?, 0);
    assert!(cache.snapshot_at(1).is_err());
    cache.reset()?;
    Ok(())
}

#[test]
fn keeps_depthwise_convolution_history_in_a_state_snapshot() -> Result<()> {
    let stream = Stream::new_gpu()?;
    let mut original = GatedDeltaState::new()?;
    let weight = Array::from_f32(&[1.0, 1.0], &[1, 2, 1])?;
    let first = Array::from_f32(&[1.0, 2.0], &[1, 2, 1])?;
    let first_output = original.convolve_silu(&first, &weight, &stream)?;
    let mut snapshot = original.snapshot()?;
    let next = Array::from_f32(&[3.0], &[1, 1, 1])?;
    let original_output = original.convolve_silu(&next, &weight, &stream)?;
    let snapshot_output = snapshot.convolve_silu(&next, &weight, &stream)?;

    first_output.async_eval()?;
    original_output.async_eval()?;
    snapshot_output.async_eval()?;
    stream.synchronize()?;
    assert_eq!(first_output.shape()?, vec![1, 2, 1]);
    assert_close(&original_output.to_vec_f32()?, &snapshot_output.to_vec_f32()?);
    Ok(())
}

struct Inputs {
    query: Array,
    key: Array,
    value: Array,
    alpha: Array,
    beta: Array,
    a_log: Array,
    dt_bias: Array,
}

impl Inputs {
    fn as_borrowed(&self) -> GatedDeltaInputs<'_> {
        GatedDeltaInputs {
            query: &self.query,
            key: &self.key,
            value: &self.value,
            alpha: &self.alpha,
            beta: &self.beta,
            a_log: &self.a_log,
            dt_bias: &self.dt_bias,
        }
    }
}

fn inputs(values: &[f32], tokens: i32) -> Result<Inputs> {
    let heads = 2;
    Ok(Inputs {
        query: Array::from_f32(
            &vec![1.0; usize::try_from(tokens * heads)? / 2],
            &[1, tokens, 1, 1],
        )?,
        key: Array::from_f32(&vec![1.0; usize::try_from(tokens * heads)? / 2], &[1, tokens, 1, 1])?,
        value: Array::from_f32(values, &[1, tokens, heads, 1])?,
        alpha: Array::from_f32(&vec![0.0; usize::try_from(tokens * heads)?], &[1, tokens, heads])?,
        beta: Array::from_f32(&vec![0.0; usize::try_from(tokens * heads)?], &[1, tokens, heads])?,
        a_log: Array::from_f32(&[0.0, 0.0], &[heads])?,
        dt_bias: Array::from_f32(&[0.0, 0.0], &[heads])?,
    })
}

fn decode_inputs(step: usize) -> Result<Inputs> {
    let scale = f32::from(u16::try_from(step)?) + 1.0;
    let mut query = vec![0.0; 32];
    let mut key = vec![0.0; 32];
    for index in 0..32 {
        query[index] = scale * f32::from(u16::try_from(index + 1)?) / 64.0;
        key[index] = f32::from(u16::try_from(33 - index)?) / 64.0;
    }
    Ok(Inputs {
        query: Array::from_f32(&query, &[1, 1, 1, 32])?,
        key: Array::from_f32(&key, &[1, 1, 1, 32])?,
        value: Array::from_f32(&[0.25, -0.5, 0.75, 1.25], &[1, 1, 2, 2])?,
        alpha: Array::from_f32(&[-0.3, 0.2], &[1, 1, 2])?,
        beta: Array::from_f32(&[0.4, -0.7], &[1, 1, 2])?,
        a_log: Array::from_f32(&[0.1, -0.2], &[2])?,
        dt_bias: Array::from_f32(&[0.05, -0.1], &[2])?,
    })
}

fn assert_close(actual: &[f32], expected: &[f32]) {
    assert_eq!(actual.len(), expected.len());
    for (actual, expected) in actual.iter().zip(expected) {
        assert!((actual - expected).abs() < 1.0e-5, "{actual} != {expected}");
    }
}