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

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

use super::AutoClampedExperts;
use crate::{
    CudaBackend, CudaConfig, CudaTensor, ExecutionPhase, PlanSource, Result,
    backend::{
        clamped_routed::{
            ClampedRoutedConfig,
            weights::{ClampedRoutedExpertWeights, NativeExpertWeights},
        },
        tuning::{ClampedMoeStorage, MoeProfileExecution, MoeProfileRequest},
    },
    kernels::{ClampedRoutedKernels, ClampedRoutedSpec},
};

const EXPERTS: usize = 4;
const TOP_K: usize = 2;
const WIDTH: usize = 32;

#[test]
fn synthetic_clamped_experts_autotune_decode_and_prefill() -> Result<()> {
    let backend = CudaBackend::new(CudaConfig::default())?;
    let weights = native_weights(&backend)?;
    run(&backend, &weights, ExecutionPhase::Decode, 1)?;
    run(&backend, &weights, ExecutionPhase::Prefill, 2)
}

fn run(
    backend: &CudaBackend,
    weights: &ClampedRoutedExpertWeights,
    phase: ExecutionPhase,
    tokens: usize,
) -> Result<()> {
    let config = config();
    let kernels = ClampedRoutedKernels::compile(backend.compiler(), kernel_spec(config, tokens))?;
    let mut experts = AutoClampedExperts::new(backend, config, tokens, phase, weights, kernels)
        .ok_or(crate::Error::InvalidExecutionPlan("missing synthetic clamped experts"))?;
    let input = upload(backend, &vec![bf16::from_f32(1.0); tokens * WIDTH])?;
    let selections = u32::try_from(tokens * TOP_K)?;
    let expert_count = u32::try_from(EXPERTS)?;
    let selected =
        upload(backend, &(0..selections).map(|value| value % expert_count).collect::<Vec<_>>())?;
    let routing = upload(backend, &vec![bf16::from_f32(0.5); tokens * TOP_K])?;
    let mut activated = backend.pool().allocate(backend.stream(), tokens * TOP_K * WIDTH)?;
    let mut partial = backend.pool().allocate(backend.stream(), tokens * TOP_K * WIDTH)?;
    let mut output = backend.pool().allocate(backend.stream(), tokens * WIDTH)?;

    experts
        .execute(weights, &input, &selected, &routing, &mut activated, &mut partial, &mut output)?;
    backend.synchronize()?;
    let profile = MoeProfileRequest::clamped(
        phase,
        tokens,
        EXPERTS,
        TOP_K,
        WIDTH,
        WIDTH,
        ClampedMoeStorage::Native,
    );
    assert!(matches!(
        backend.auto_tuner().lookup_moe(profile),
        Some((MoeProfileExecution::Clamped(_), PlanSource::MeasuredStartup))
    ));
    Ok(())
}

fn native_weights(backend: &CudaBackend) -> Result<ClampedRoutedExpertWeights> {
    let gate_rows = 2 * WIDTH;
    Ok(ClampedRoutedExpertWeights::Native(Box::new(NativeExpertWeights {
        gate_up_blocks: u8_tensor(
            backend,
            "gate-up-blocks",
            vec![EXPERTS, gate_rows, 1, 16],
            0x11,
        )?,
        gate_up_scales: u8_tensor(backend, "gate-up-scales", vec![EXPERTS, gate_rows, 1], 127)?,
        gate_up_bias: bf16_tensor(backend, "gate-up-bias", EXPERTS * gate_rows)?,
        down_blocks: u8_tensor(backend, "down-blocks", vec![EXPERTS, WIDTH, 1, 16], 0x11)?,
        down_scales: u8_tensor(backend, "down-scales", vec![EXPERTS, WIDTH, 1], 127)?,
        down_bias: bf16_tensor(backend, "down-bias", EXPERTS * WIDTH)?,
    })))
}

fn u8_tensor(
    backend: &CudaBackend,
    name: &str,
    shape: Vec<usize>,
    value: u8,
) -> Result<CudaTensor> {
    let elements = shape.iter().product();
    Ok(CudaTensor::from_u8(
        name.into(),
        shape,
        upload(backend, &vec![value; elements])?,
    ))
}

fn bf16_tensor(backend: &CudaBackend, name: &str, elements: usize) -> Result<CudaTensor> {
    Ok(CudaTensor::from_bf16(
        name.into(),
        vec![elements],
        upload(backend, &vec![bf16::from_f32(0.0); elements])?,
    ))
}

fn upload<T: DeviceElement>(backend: &CudaBackend, values: &[T]) -> Result<DeviceBuffer<T>> {
    let mut host = backend.context().allocate_pinned(values.len())?;
    host.copy_from_slice(values)?;
    let mut device = backend.pool().allocate(backend.stream(), values.len())?;
    backend.stream().copy_to_device(&mut host, &mut device)?;
    Ok(device)
}

fn config() -> ClampedRoutedConfig {
    ClampedRoutedConfig {
        vocab: 32,
        hidden: WIDTH,
        intermediate: WIDTH,
        query_heads: 1,
        kv_heads: 1,
        head_dim: WIDTH,
        experts: EXPERTS,
        top_k: TOP_K,
        epsilon: 1.0e-5,
        scale: 1.0,
        theta: 150_000.0,
        factor: 32.0,
        initial_context: 4_096.0,
        beta_fast: 32.0,
        beta_slow: 1.0,
        swiglu_limit: 7.0,
    }
}

fn kernel_spec(config: ClampedRoutedConfig, tokens: usize) -> ClampedRoutedSpec {
    ClampedRoutedSpec {
        tokens,
        hidden: config.hidden,
        intermediate: config.intermediate,
        query_heads: config.query_heads,
        kv_heads: config.kv_heads,
        head_dim: config.head_dim,
        top_k: config.top_k,
        theta: config.theta,
        factor: config.factor,
        initial_context: config.initial_context,
        beta_fast: config.beta_fast,
        beta_slow: config.beta_slow,
        swiglu_limit: config.swiglu_limit,
    }
}