cram-codec 0.1.0

Error-bounded scientific-data compression: bit-accounted quantized tensor train (QTT) codec
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

//! KV-cache codec CLI (see `cram_codec::kvcodec`).
//!
//!     kvc encode IN.f32 OUT --shape HxTxD --kind k|v (--rmse r | --maxabs e | --bpv b)
//!         [--post] [--sink n] [--sinkdiv n] [--block n] [--noar] [--theta t] [--prefill n]
//!     kvc decode IN OUT.f32 [--range t0:t1] [--reps n]
//!
//! `--prefill n` fixes the header from the first n tokens and appends the rest
//! one token at a time (the streaming path); default is one shot.

use std::time::Instant;

use cram_codec::kvcodec::{step_for_bpv, Kind, KvEncoder, KvParams, KvStream};

fn usage() -> ! {
    eprintln!("usage: kvc encode IN OUT --shape HxTxD --kind k|v (--rmse r|--maxabs e|--bpv b) [--post] [--sink n] [--sinkdiv n] [--block n] [--noar] [--theta t] [--prefill n]\n       kvc decode IN OUT [--range t0:t1] [--reps n]");
    std::process::exit(2)
}

fn flag(args: &[String], name: &str) -> Option<String> {
    args.iter()
        .position(|a| a == name)
        .map(|i| args.get(i + 1).cloned().unwrap_or_else(|| usage()))
}

fn read_f32(path: &str) -> Vec<f32> {
    std::fs::read(path)
        .unwrap()
        .as_chunks::<4>()
        .0
        .iter()
        .map(|&c| f32::from_le_bytes(c))
        .collect()
}

fn main() {
    let args: Vec<String> = std::env::args().collect();
    if args.len() < 4 {
        usage()
    }
    match args[1].as_str() {
        "encode" => {
            let x = read_f32(&args[2]);
            let shape: Vec<usize> = flag(&args, "--shape")
                .unwrap_or_else(|| usage())
                .split('x')
                .map(|s| s.parse().unwrap())
                .collect();
            let (h, n, d) = (shape[0], shape[1], shape[2]);
            let kind = match flag(&args, "--kind").as_deref() {
                Some("k") => Kind::K,
                Some("v") => Kind::V,
                _ => usage(),
            };
            let mut p = KvParams::new(kind, h, d, 1.0);
            p.pre_rope = !args.iter().any(|a| a == "--post");
            p.ar = !args.iter().any(|a| a == "--noar");
            p.gpu = args.iter().any(|a| a == "--gpu");
            if let Some(v) = flag(&args, "--sink") {
                p.sink = v.parse().unwrap();
            }
            if let Some(v) = flag(&args, "--sinkdiv") {
                p.sink_div = v.parse().unwrap();
            }
            if let Some(v) = flag(&args, "--block") {
                p.block = v.parse().unwrap();
            }
            if let Some(v) = flag(&args, "--theta") {
                p.theta = v.parse().unwrap();
            }
            let t = Instant::now();
            p.step = if let Some(r) = flag(&args, "--rmse") {
                r.parse::<f64>().unwrap() * 12f64.sqrt()
            } else if let Some(e) = flag(&args, "--maxabs") {
                // pre-RoPE pair rotation can move a per-element error by sqrt(2)
                let e: f64 = e.parse().unwrap();
                if kind == Kind::K && p.pre_rope {
                    e * std::f64::consts::SQRT_2
                } else {
                    2.0 * e
                }
            } else if let Some(b) = flag(&args, "--bpv") {
                step_for_bpv(p.clone(), &x, n, b.parse().unwrap())
            } else {
                usage()
            };
            let pre: usize = flag(&args, "--prefill")
                .map(|v| v.parse().unwrap())
                .unwrap_or(n);
            let slice = |t0: usize, t1: usize| -> Vec<f32> {
                let mut v = Vec::with_capacity(h * (t1 - t0) * d);
                for hh in 0..h {
                    v.extend_from_slice(&x[(hh * n + t0) * d..(hh * n + t1) * d]);
                }
                v
            };
            let mut e = KvEncoder::new(p.clone(), &slice(0, pre), pre);
            let ta = Instant::now();
            for t in pre..n {
                e.append(&slice(t, t + 1), 1);
            }
            let append_s = ta.elapsed().as_secs_f64();
            let bytes = e.to_bytes();
            let enc = t.elapsed().as_secs_f64();
            std::fs::write(&args[3], &bytes).unwrap();
            println!(
                "{{\"bytes\":{},\"bpv\":{},\"step\":{},\"encode_s\":{enc},\"append_s\":{append_s},\"appended\":{}}}",
                bytes.len(),
                8.0 * bytes.len() as f64 / x.len() as f64,
                p.step,
                n - pre
            );
        }
        "decode" => {
            let bytes = std::fs::read(&args[2]).unwrap();
            let s = KvStream::parse(&bytes).unwrap();
            let (t0, t1) = match flag(&args, "--range") {
                Some(r) => {
                    let (a, b) = r.split_once(':').unwrap();
                    (a.parse().unwrap(), b.parse().unwrap())
                }
                None => (0, s.tokens),
            };
            let reps: usize = flag(&args, "--reps")
                .map(|v| v.parse().unwrap())
                .unwrap_or(1);
            let t = Instant::now();
            let mut y = Vec::new();
            for _ in 0..reps {
                y = s.decode_range(t0, t1);
            }
            let dec = t.elapsed().as_secs_f64() / reps as f64;
            let out: Vec<u8> = y.iter().flat_map(|v| v.to_le_bytes()).collect();
            std::fs::write(&args[3], out).unwrap();
            println!(
                "{{\"decode_s\":{dec},\"ns_per_value\":{}}}",
                1e9 * dec / y.len().max(1) as f64
            );
        }
        _ => usage(),
    }
}