hf2q 0.1.13

Pure Rust CLI for converting HuggingFace models to hardware-optimized formats and serving them over an OpenAI-compatible API on Apple Silicon
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

use std::collections::BTreeSet;

use super::{
    reject, valid_nonnegative, valid_sha256, CandidateReceipt, CandidateRejectionCode,
    CandidateRejectionReason, PerformanceMeasurement, SelectionContract,
};

pub(super) fn check_performance(
    contract: &SelectionContract,
    candidate: &CandidateReceipt,
    reasons: &mut Vec<CandidateRejectionReason>,
) {
    let mut seen = BTreeSet::new();
    for measurement in &candidate.performance {
        if !seen.insert(measurement.regime) {
            reject(
                reasons,
                CandidateRejectionCode::DuplicatePerformanceMeasurement,
                format!(
                    "duplicate performance measurement: {:?}",
                    measurement.regime
                ),
            );
        }
        if !valid_sha256(&measurement.workload_sha256)
            || !valid_nonnegative(measurement.median_tokens_per_second)
            || !valid_nonnegative(measurement.median_semantic_ttft_ms)
            || measurement.peak_mlx_bytes == 0
            || measurement.measured_runs == 0
            || measurement.tokens_per_run == 0
            || !measurement.output_quality_cases.is_complete()
        {
            reject(
                reasons,
                CandidateRejectionCode::InvalidPerformanceMeasurement,
                format!(
                    "invalid performance measurement for {:?}",
                    measurement.regime
                ),
            );
        }
    }

    for requirement in &contract.required_regimes {
        let Some(measurement) = candidate
            .performance
            .iter()
            .find(|measurement| measurement.regime == requirement.regime)
        else {
            reject(
                reasons,
                CandidateRejectionCode::MissingPerformanceRegime,
                format!(
                    "required performance regime is missing: {:?}",
                    requirement.regime
                ),
            );
            continue;
        };

        if !valid_nonnegative(measurement.median_tokens_per_second)
            || measurement.median_tokens_per_second < requirement.min_tokens_per_second
        {
            reject(
                reasons,
                CandidateRejectionCode::ThroughputGateFailed,
                format!("throughput gate failed for {:?}", requirement.regime),
            );
        }
        if measurement.workload_sha256 != requirement.workload_sha256 {
            reject(
                reasons,
                CandidateRejectionCode::WorkloadIdentityMismatch,
                format!("workload identity mismatch for {:?}", requirement.regime),
            );
        }
        if let Some(max_ttft_ms) = requirement.max_semantic_ttft_ms {
            if !valid_nonnegative(measurement.median_semantic_ttft_ms)
                || measurement.median_semantic_ttft_ms > max_ttft_ms
            {
                reject(
                    reasons,
                    CandidateRejectionCode::SemanticTtftGateFailed,
                    format!("semantic TTFT gate failed for {:?}", requirement.regime),
                );
            }
        }
        if measurement.peak_mlx_bytes > contract.max_peak_mlx_bytes {
            reject(
                reasons,
                CandidateRejectionCode::MemoryGateFailed,
                format!("memory gate failed for {:?}", requirement.regime),
            );
        }
        if measurement.warmup_runs < requirement.min_warmup_runs
            || measurement.measured_runs < requirement.min_measured_runs
            || measurement.tokens_per_run < requirement.min_tokens_per_run
        {
            reject(
                reasons,
                CandidateRejectionCode::InsufficientBenchmarkEvidence,
                format!(
                    "insufficient benchmark evidence for {:?}",
                    requirement.regime
                ),
            );
        }
    }
}

pub(super) fn compare_candidates(
    contract: &SelectionContract,
    left: &CandidateReceipt,
    right: &CandidateReceipt,
) -> std::cmp::Ordering {
    let left_primary = primary_measurement(contract, left);
    let right_primary = primary_measurement(contract, right);

    right_primary
        .median_tokens_per_second
        .total_cmp(&left_primary.median_tokens_per_second)
        .then_with(|| {
            left_primary
                .median_semantic_ttft_ms
                .total_cmp(&right_primary.median_semantic_ttft_ms)
        })
        .then_with(|| left.artifact_bytes.cmp(&right.artifact_bytes))
        .then_with(|| left.recipe.candidate_id.cmp(&right.recipe.candidate_id))
}

fn primary_measurement<'a>(
    contract: &SelectionContract,
    candidate: &'a CandidateReceipt,
) -> &'a PerformanceMeasurement {
    // Eligibility already proved this measurement exists exactly once.
    candidate
        .performance
        .iter()
        .find(|measurement| measurement.regime == contract.primary_regime)
        .expect("eligible candidate must contain primary performance evidence")
}