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
//! Runtime roofline model derived from contract YAML.
//!
//! Reads `roofline-model-v1.yaml` and provides performance ceiling
//! calculations. Consumer crates use this instead of hardcoded formulas.
//! Section 24: Deep Stack Integration.
use std::path::Path;
/// Performance ceilings derived from the roofline contract.
#[derive(Debug, Clone)]
pub struct RooflineCeiling {
/// Model size in bytes
pub model_bytes: f64,
/// Bandwidth-limited ceiling (tokens/sec)
pub bw_ceiling: f64,
/// Compute-limited ceiling (tokens/sec)
pub compute_ceiling: f64,
/// Effective ceiling: min(bw, compute)
pub throughput_ceiling: f64,
/// Whether bandwidth-bound or compute-bound
pub bottleneck: Bottleneck,
/// Source contract ID
pub contract_id: String,
}
/// Whether the workload is memory-bandwidth or compute bound.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Bottleneck {
Bandwidth,
Compute,
}
/// Hardware profile for roofline calculation.
#[derive(Debug, Clone)]
pub struct HardwareProfile {
/// Effective memory bandwidth in GB/s
pub bandwidth_gb_s: f64,
/// Effective compute throughput in GFLOPS
pub compute_gflops: f64,
/// Operations per token (depends on model architecture)
pub ops_per_token: f64,
}
impl HardwareProfile {
/// Create a profile for Apple M-series (conservative estimates).
pub fn apple_m_series() -> Self {
Self {
bandwidth_gb_s: 100.0, // M1 Pro ~200, M2 Ultra ~800, use conservative
compute_gflops: 1000.0, // GPU TFLOPS varies; conservative
ops_per_token: 2.0, // 2 FLOPs per parameter per token (forward pass)
}
}
/// Create a profile for NVIDIA A100.
pub fn nvidia_a100() -> Self {
Self {
bandwidth_gb_s: 2039.0, // HBM2e bandwidth
compute_gflops: 19500.0, // FP16 tensor core
ops_per_token: 2.0,
}
}
}
/// Compute roofline ceilings from contract equations + hardware profile.
///
/// Implements the 4 equations from `roofline-model-v1.yaml`:
/// 1. `model_bytes = total_params × bits_per_weight / 8`
/// 2. `bw_ceiling = effective_bandwidth_GB_s / (model_bytes / 1e9)`
/// 3. `compute_ceiling = effective_GFLOPS / ops_per_token`
/// 4. `throughput <= min(bw_ceiling, compute_ceiling)`
pub fn compute_roofline(
total_params: u64,
bits_per_weight: u32,
hw: &HardwareProfile,
) -> RooflineCeiling {
// Equation 1: model_bytes
#[allow(clippy::cast_precision_loss)]
// model params fit within f64 mantissa for all real models
let total_params_f = total_params as f64;
let model_bytes = total_params_f * f64::from(bits_per_weight) / 8.0;
// Equation 2: bw_ceiling (tokens/sec)
let model_gb = model_bytes / 1e9;
let bw_ceiling = if model_gb > 0.0 {
hw.bandwidth_gb_s / model_gb
} else {
f64::INFINITY
};
// Equation 3: compute_ceiling (tokens/sec)
let compute_ceiling = if hw.ops_per_token > 0.0 {
hw.compute_gflops * 1e9 / (total_params_f * hw.ops_per_token)
} else {
f64::INFINITY
};
// Equation 4: throughput <= min(bw_ceiling, compute_ceiling)
let throughput_ceiling = bw_ceiling.min(compute_ceiling);
let bottleneck = if bw_ceiling < compute_ceiling {
Bottleneck::Bandwidth
} else {
Bottleneck::Compute
};
RooflineCeiling {
model_bytes,
bw_ceiling,
compute_ceiling,
throughput_ceiling,
bottleneck,
contract_id: "roofline-model-v1".to_string(),
}
}
/// Try to load roofline contract from a standard path.
/// Returns the contract description if found, None if not.
pub fn load_roofline_contract(contracts_dir: &Path) -> Option<String> {
let path = contracts_dir.join("roofline-model-v1.yaml");
if path.exists() {
let content = std::fs::read_to_string(&path).ok()?;
// Extract description
for line in content.lines() {
if let Some(desc) = line.trim().strip_prefix("description:") {
return Some(desc.trim().trim_matches('"').to_string());
}
}
}
None
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn roofline_7b_q4() {
// 7B params, 4-bit quantized, on hypothetical hardware
let hw = HardwareProfile {
bandwidth_gb_s: 200.0,
compute_gflops: 5000.0,
ops_per_token: 2.0,
};
let r = compute_roofline(7_000_000_000, 4, &hw);
assert!((r.model_bytes - 3_500_000_000.0).abs() < 1.0); // 7B * 4 / 8 = 3.5GB
assert!(r.bw_ceiling > 0.0);
assert!(r.compute_ceiling > 0.0);
assert!((r.throughput_ceiling - r.bw_ceiling.min(r.compute_ceiling)).abs() < f64::EPSILON);
assert_eq!(r.bottleneck, Bottleneck::Bandwidth); // Memory-bound at 4-bit
}
#[test]
fn roofline_contract_id() {
let hw = HardwareProfile::apple_m_series();
let r = compute_roofline(1_000_000, 16, &hw);
assert_eq!(r.contract_id, "roofline-model-v1");
}
#[test]
fn bottleneck_classification() {
// Tiny model on fast hardware = compute bound
let hw = HardwareProfile {
bandwidth_gb_s: 1000.0,
compute_gflops: 1.0, // Very slow compute
ops_per_token: 2.0,
};
let r = compute_roofline(1000, 32, &hw);
assert_eq!(r.bottleneck, Bottleneck::Compute);
}
}