eld_llm 0.0.1

An LLM built from scratch in Rust
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

use std::time::Instant;
use crate::tokenizers::{Tokenizer, benchmark::dev::traits::{
    TokenizerDevelopmentResult, TokenizerVocabMetrics, TokenizerInferenceMetrics,
}};

pub struct TokenizerDevelopmentBencher;

impl TokenizerDevelopmentBencher {
    pub fn new() -> Self {
        Self
    }

    pub fn run<T: Tokenizer>(&self, tokenizer: &mut T, data_path: &str, corpus_name: &str) -> TokenizerDevelopmentResult {
        // train and collect timing
        let training_metrics = tokenizer
            .train(data_path, corpus_name)
            .expect("Training failed");

        // compute vocab length stats from all token strings
        let tokens = tokenizer.vocab_tokens();
        let lengths: Vec<usize> = tokens.iter().map(|t| t.len()).collect();
        let vocab_size = lengths.len();
        let avg_token_length = if vocab_size > 0 {
            lengths.iter().sum::<usize>() as f64 / vocab_size as f64
        } else {
            0.0
        };
        let vocab_metrics = TokenizerVocabMetrics {
            vocab_size,
            avg_token_length,
            max_token_length: lengths.iter().copied().max().unwrap_or(0),
            min_token_length: lengths.iter().copied().min().unwrap_or(0),
        };

        // time encode and decode over the full corpus
        let corpus_text = std::fs::read_to_string(data_path).expect("Failed to read corpus");

        let encode_start = Instant::now();
        let encoded = tokenizer.encode(&corpus_text);
        let encode_elapsed_ns = encode_start.elapsed().as_nanos() as f64;

        let decode_start = Instant::now();
        let _ = tokenizer.decode(&encoded);
        let decode_elapsed_ns = decode_start.elapsed().as_nanos() as f64;

        let token_count = encoded.len();
        let inference_metrics = TokenizerInferenceMetrics {
            avg_encode_time_ns: encode_elapsed_ns / token_count as f64,
            avg_decode_time_ns: decode_elapsed_ns / token_count as f64,
            encode_throughput_tokens_per_sec: token_count as f64 / (encode_elapsed_ns / 1e9),
            decode_throughput_tokens_per_sec: token_count as f64 / (decode_elapsed_ns / 1e9),
            corpus_token_count: token_count,
        };

        // assemble and return the full result
        TokenizerDevelopmentResult {
            version: tokenizer.name().to_string(),
            corpus_name: corpus_name.to_string(),
            corpus_size_bytes: tokenizer.corpus().total_size_bytes(),
            training_metrics,
            vocab_metrics,
            inference_metrics,
        }
    }
}