ai-hwaccel 1.1.0

Universal AI hardware accelerator detection, capability querying, and workload planning for Rust
Documentation 
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

//! Training memory estimation tests.

use crate::*;

// ---------------------------------------------------------------------------
// Training memory: cross-target comparisons
// ---------------------------------------------------------------------------

#[test]
fn training_memory_tpu_less_optimizer_than_gpu() {
    let gpu = estimate_training_memory(7000, TrainingMethod::FullFineTune, TrainingTarget::Gpu);
    let tpu = estimate_training_memory(7000, TrainingMethod::FullFineTune, TrainingTarget::Tpu);
    assert!(tpu.optimizer_gb < gpu.optimizer_gb);
}

#[test]
fn training_memory_qlora_less_than_full() {
    let full = estimate_training_memory(7000, TrainingMethod::FullFineTune, TrainingTarget::Gpu);
    let qlora =
        estimate_training_memory(7000, TrainingMethod::QLoRA { bits: 4 }, TrainingTarget::Gpu);
    assert!(qlora.total_gb < full.total_gb);
    assert!(qlora.model_gb < full.model_gb);
}

#[test]
fn training_memory_lora_less_than_full() {
    let full = estimate_training_memory(7000, TrainingMethod::FullFineTune, TrainingTarget::Gpu);
    let lora = estimate_training_memory(7000, TrainingMethod::LoRA, TrainingTarget::Gpu);
    assert!(lora.total_gb < full.total_gb);
    assert!(lora.optimizer_gb < full.optimizer_gb);
}

#[test]
fn training_memory_gaudi() {
    let est = estimate_training_memory(7000, TrainingMethod::FullFineTune, TrainingTarget::Gaudi);
    assert!(est.total_gb > 30.0);
    assert!(est.total_gb < 60.0);
    assert!(est.model_gb > 0.0);
    assert!(est.optimizer_gb > 0.0);
    assert!(est.activation_gb > 0.0);
}

// ---------------------------------------------------------------------------
// All training methods produce positive estimates
// ---------------------------------------------------------------------------

#[test]
fn training_memory_all_methods_positive() {
    let methods = [
        TrainingMethod::FullFineTune,
        TrainingMethod::LoRA,
        TrainingMethod::QLoRA { bits: 4 },
        TrainingMethod::QLoRA { bits: 8 },
        TrainingMethod::Prefix,
        TrainingMethod::DPO,
        TrainingMethod::RLHF,
        TrainingMethod::Distillation,
    ];
    let targets = [
        TrainingTarget::Gpu,
        TrainingTarget::Tpu,
        TrainingTarget::Gaudi,
        TrainingTarget::Cpu,
    ];
    for method in &methods {
        for target in &targets {
            let est = estimate_training_memory(7000, *method, *target);
            assert!(
                est.total_gb > 0.0,
                "{:?} on {:?} should have positive total",
                method,
                target
            );
            assert!(
                (est.model_gb + est.optimizer_gb + est.activation_gb - est.total_gb).abs() < 0.001,
                "{:?} on {:?}: component sum doesn't match total",
                method,
                target
            );
        }
    }
}

// ---------------------------------------------------------------------------
// DPO/RLHF need more memory than LoRA (two model copies)
// ---------------------------------------------------------------------------

#[test]
fn training_memory_dpo_more_than_lora() {
    let lora = estimate_training_memory(7000, TrainingMethod::LoRA, TrainingTarget::Gpu);
    let dpo = estimate_training_memory(7000, TrainingMethod::DPO, TrainingTarget::Gpu);
    assert!(dpo.total_gb > lora.total_gb);
}

#[test]
fn training_memory_rlhf_more_than_lora() {
    let lora = estimate_training_memory(7000, TrainingMethod::LoRA, TrainingTarget::Gpu);
    let rlhf = estimate_training_memory(7000, TrainingMethod::RLHF, TrainingTarget::Gpu);
    assert!(rlhf.total_gb > lora.total_gb);
}

#[test]
fn training_memory_distillation() {
    let full = estimate_training_memory(7000, TrainingMethod::FullFineTune, TrainingTarget::Gpu);
    let dist = estimate_training_memory(7000, TrainingMethod::Distillation, TrainingTarget::Gpu);
    // Distillation needs teacher + student, more model memory
    assert!(dist.model_gb > full.model_gb);
}

#[test]
fn training_memory_prefix_is_lightweight() {
    let full = estimate_training_memory(7000, TrainingMethod::FullFineTune, TrainingTarget::Gpu);
    let prefix = estimate_training_memory(7000, TrainingMethod::Prefix, TrainingTarget::Gpu);
    assert!(prefix.optimizer_gb < full.optimizer_gb);
    assert!(prefix.total_gb < full.total_gb);
}

// ---------------------------------------------------------------------------
// QLoRA 4-bit vs 8-bit
// ---------------------------------------------------------------------------

#[test]
fn training_memory_qlora_4bit_less_than_8bit() {
    let q4 = estimate_training_memory(7000, TrainingMethod::QLoRA { bits: 4 }, TrainingTarget::Gpu);
    let q8 = estimate_training_memory(7000, TrainingMethod::QLoRA { bits: 8 }, TrainingTarget::Gpu);
    assert!(q4.model_gb < q8.model_gb);
}

// ---------------------------------------------------------------------------
// TrainingMethod
// ---------------------------------------------------------------------------

#[test]
fn training_method_display() {
    assert_eq!(TrainingMethod::LoRA.to_string(), "lora");
    assert_eq!(TrainingMethod::QLoRA { bits: 4 }.to_string(), "qlora-4bit");
    assert_eq!(TrainingMethod::FullFineTune.to_string(), "full");
    assert_eq!(TrainingMethod::DPO.to_string(), "dpo");
    assert_eq!(TrainingMethod::RLHF.to_string(), "rlhf");
    assert_eq!(TrainingMethod::Distillation.to_string(), "distillation");
    assert_eq!(TrainingMethod::Prefix.to_string(), "prefix");
}

#[test]
fn training_method_preferred_accelerator() {
    assert_eq!(
        TrainingMethod::LoRA.preferred_accelerator(),
        AcceleratorRequirement::Gpu
    );
    assert_eq!(
        TrainingMethod::QLoRA { bits: 4 }.preferred_accelerator(),
        AcceleratorRequirement::Gpu
    );
    assert_eq!(
        TrainingMethod::FullFineTune.preferred_accelerator(),
        AcceleratorRequirement::GpuOrTpu
    );
    assert_eq!(
        TrainingMethod::DPO.preferred_accelerator(),
        AcceleratorRequirement::GpuOrTpu
    );
    assert_eq!(
        TrainingMethod::RLHF.preferred_accelerator(),
        AcceleratorRequirement::GpuOrTpu
    );
    assert_eq!(
        TrainingMethod::Distillation.preferred_accelerator(),
        AcceleratorRequirement::GpuOrTpu
    );
    assert_eq!(
        TrainingMethod::Prefix.preferred_accelerator(),
        AcceleratorRequirement::GpuOrTpu
    );
}