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

use super::*;

#[test]
fn auto_selects_validated_refinement_and_gates_experiments() -> Result<()> {
    let request = OutputHeadPlanRequest {
        input_features: 2_816,
        output_features: 262_144,
    };
    assert_eq!(
        planner(12)?.plan_output_head(request)?.execution(),
        OutputHeadExecution::AutoRefined
    );
    assert_eq!(
        planner(12)?
            .plan_output_head(OutputHeadPlanRequest {
                input_features: 4_096,
                output_features: 151_936,
            })?
            .execution(),
        OutputHeadExecution::AutoRefined
    );
    assert_execution(request, CudaOutputHeadPolicy::Bf16, OutputHeadExecution::Bf16)?;
    for (policy, execution) in [
        (CudaOutputHeadPolicy::Fp8Blockwise, OutputHeadExecution::Fp8Blockwise),
        (CudaOutputHeadPolicy::Fp8Vectorized, OutputHeadExecution::Fp8Vectorized),
        (CudaOutputHeadPolicy::Fp8Residual, OutputHeadExecution::Fp8Residual),
        (
            CudaOutputHeadPolicy::Fp8BlockVectorized,
            OutputHeadExecution::Fp8BlockVectorized,
        ),
        (CudaOutputHeadPolicy::Fp8BlockRefined, OutputHeadExecution::Fp8BlockRefined),
    ] {
        assert_execution(request, policy, execution)?;
    }
    let selected = experimental(CudaOutputHeadPolicy::Fp8Blockwise);
    assert_eq!(
        planner_with_policy(11, selected)?.plan_output_head(request)?.execution(),
        OutputHeadExecution::Bf16
    );
    Ok(())
}

fn assert_execution(
    request: OutputHeadPlanRequest,
    policy: CudaOutputHeadPolicy,
    expected: OutputHeadExecution,
) -> Result<()> {
    let planning = if policy == CudaOutputHeadPolicy::Bf16 {
        CudaPlanningPolicy {
            output_head: policy,
            ..CudaPlanningPolicy::default()
        }
    } else {
        experimental(policy)
    };
    assert_eq!(
        planner_with_policy(12, planning)?.plan_output_head(request)?.execution(),
        expected
    );
    Ok(())
}

fn experimental(output_head: CudaOutputHeadPolicy) -> CudaPlanningPolicy {
    CudaPlanningPolicy {
        numerical: CudaNumericalPolicy::Throughput,
        admission: CudaKernelAdmission::Experimental,
        output_head,
        ..CudaPlanningPolicy::default()
    }
}