Skip to content

fix: store key-value pairs in the hashmap preserving the order of insertions #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: store key-value pairs in the hashmap preserving the order of insertions #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 33 additions & 10 deletions 002-Rust-bindings-to-Python/rust-llm/Cargo.lock
View file
Open in desktop

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

1 change: 1 addition & 0 deletions 002-Rust-bindings-to-Python/rust-llm/Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,3 +12,4 @@ pyo3 = { version = "0.20.0", features = ["extension-module"] }
lazy_static = "1.4.0"
base64 = "0.22.1"
unicode_categories = "0.1"
indexmap = "2.5.0"
145 changes: 85 additions & 60 deletions 002-Rust-bindings-to-Python/rust-llm/src/bpe.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@

use pyo3::prelude::*;
use pyo3::prepare_freethreaded_python;
use indexmap::IndexMap;
use std::collections::BTreeMap;
use std::fs::File;
use std::io::{BufRead, BufReader, Write};
Expand All @@ -14,9 +15,9 @@ use crate::helpers::{get_stats, merge, render_token, build_vocab, b_as_literal};
// The #[pyclass] macro from PyO3 is used to map this as a Python class
#[pyclass]
pub struct BPETokenizer {
merges: BTreeMap<(i32, i32), i32>, // (pair, idx)
merges: IndexMap<(i32, i32), i32>, // (pair, idx)
pattern: String, // pattern to split the text into tokens
special_tokens: BTreeMap<String, i32>, // special tokens (string, idx)
special_tokens: IndexMap<String, i32>, // special tokens (string, idx)
vocab: BTreeMap<i32, Vec<u8>>, // (idx, token)
}

Expand All @@ -27,10 +28,10 @@ impl BPETokenizer {
#[new]
pub fn new() -> Self {
prepare_freethreaded_python();
let merges = BTreeMap::<(i32, i32), i32>::new();
let pattern = String::new();
let special_tokens = BTreeMap::<String, i32>::new();
let vocab = build_vocab(&merges, &special_tokens);
let merges: IndexMap<(i32, i32), i32> = IndexMap::<(i32, i32), i32>::new();
let pattern: String = String::new();
let special_tokens: IndexMap<String, i32> = IndexMap::<String, i32>::new();
let vocab: BTreeMap<i32, Vec<u8>> = build_vocab(&merges, &special_tokens);

BPETokenizer {
merges,
Expand All @@ -40,7 +41,7 @@ impl BPETokenizer {
}
}

fn train(&mut self, text: &str, vocab_size: usize, verbose: bool) {
pub fn train(&mut self, text: &str, vocab_size: usize, verbose: bool) {
assert!(vocab_size >= 256);
let num_merges = vocab_size - 256;

Expand All @@ -58,29 +59,36 @@ impl BPETokenizer {

for i in 0..num_merges {
let stats = get_stats(&ids);
if stats.is_empty() {
break;
}
let pair = *stats.iter().max_by_key(|&(_, count)| count).unwrap().0;
let idx = 256 + i as i32;
ids = merge(&ids, pair, idx);

self.merges.insert(pair, idx);

// Separate immutable borrow before mutable borrow
let pair_1_len = self.vocab[&pair.1].len();
if let Some(vocab_entry) = self.vocab.get_mut(&pair.0) {
let mut new_token = Vec::with_capacity(vocab_entry.len() + pair_1_len);
new_token.extend_from_slice(vocab_entry);
new_token.extend_from_slice(&self.vocab[&pair.1]);
self.vocab.insert(idx, new_token);
if let (Some(vocab_entry_p0), Some(vocab_entry_p1)) = (self.vocab.get(&pair.0), self.vocab.get(&pair.1)) {
let mut merged_vec = Vec::with_capacity(vocab_entry_p0.len() + vocab_entry_p1.len());
merged_vec.extend_from_slice(vocab_entry_p0);
merged_vec.extend_from_slice(vocab_entry_p1);
self.vocab.insert(idx, merged_vec);
}
else {
// Handle the case where p0 or p1 does not exist in vocab
eprintln!("Warning: Missing key in vocab for pair ({}, {})", p0, p1);
}

if verbose {
println!(
"merge {}/{}: {:?} -> {} ({:?}) had {} occurrences",
"merge {}/{}: {:?} -> {} ({:?}, binary: {:?}) had {} occurrences",
i + 1,
num_merges,
pair,
idx,
self.vocab[&idx],
b_as_literal(&self.vocab[&idx]),
stats[&pair]
);
}
Expand All @@ -89,17 +97,66 @@ impl BPETokenizer {
#[cfg(debug_assertions)]
{
println!("[DEBUG] Merges: {:?}", self.merges);
println!("[DEBUG] Vocabulary: ");
for (key, value) in &self.vocab {
let formatted_value = b_as_literal(value);
println!("{}: {}", key, formatted_value);
}
}
}

pub fn encode(&self, text: &str, verbose: bool) -> Vec<i32> {
if verbose {
println!("[DEBUG] Encoding text: {}", text);
}
let mut ids: Vec<i32> = text.bytes().map(|b| b as i32).collect();
let merges = &self.merges;
if verbose {
println!("[DEBUG] Initial IDs: {:?}", ids);
println!("[DEBUG] Merges loaded: {:?}", merges);
}

while ids.len() >= 2 {
let stats = get_stats(&ids);
if verbose {
println!("[DEBUG] Current IDs: {:?}", ids);
println!("[DEBUG] Stats: {:?}", stats);
}
let pair = stats
.keys()
.min_by_key(|&&p| merges.get(&p).unwrap_or(&i32::MAX))
.copied()
.unwrap();

if !merges.contains_key(&pair) {
break;
}

let new_id = merges[&pair];
ids = merge(&ids, pair, new_id);
if verbose {
println!("[DEBUG] Pair: {:?}", pair);
println!("[DEBUG] New ID for merge: {}", new_id);
}
}

if verbose {
println!("[DEBUG] Encoding complete. Final IDs: {:?}", ids);
}
ids
}

pub fn decode(&self, ids: Vec<i32>) -> String {
let mut text_bytes = Vec::with_capacity(ids.len() * 2); // Pre-allocate assuming 2 bytes per id on average

for &id in &ids {
text_bytes.extend_from_slice(&self.vocab[&id]);
if let Some(vocab_entry) = self.vocab.get(&id) {
text_bytes.extend_from_slice(vocab_entry);
}
else {
// Handle the case where p0 or p1 does not exist in vocab
eprintln!("Warning: Missing key in vocab for id {}", id);
}
}
String::from_utf8_lossy(&text_bytes).into_owned()
}
Expand All @@ -121,7 +178,7 @@ impl BPETokenizer {

let vocab_file = format!("{}.vocab", file_prefix);
let mut vocab_file = File::create(&vocab_file)?;
let inverted_merges: BTreeMap<i32, (i32, i32)> = self.merges
let inverted_merges: IndexMap<i32, (i32, i32)> = self.merges
.iter()
.map(|(&(p0, p1), &idx)| (idx, (p0, p1)))
.collect();
Expand All @@ -131,9 +188,9 @@ impl BPETokenizer {
if let Some(&(idx0, idx1)) = inverted_merges.get(&idx) {
let s0 = render_token(&self.vocab[&idx0]);
let s1 = render_token(&self.vocab[&idx1]);
writeln!(vocab_file, "[{}][{}] -> [{}] {} \n", s0, s1, s, idx)?;
writeln!(vocab_file, "[{}][{}] -> [{}] {}", s0, s1, s, idx)?;
} else {
writeln!(vocab_file, "[{}] {} \n", s, idx)?;
writeln!(vocab_file, "[{}] {}", s, idx)?;
}
}
Ok(())
Expand All @@ -154,15 +211,15 @@ impl BPETokenizer {
self.pattern = lines.next().unwrap().unwrap().trim().to_string();

let num_special_tokens = lines.next().unwrap().unwrap().parse::<i32>().unwrap();
let mut special_tokens = BTreeMap::<String, i32>::new();
let mut special_tokens = IndexMap::<String, i32>::new();

for _ in 0..num_special_tokens {
let line = lines.next().unwrap().unwrap();
let parts: Vec<&str> = line.split_whitespace().collect(); // getting special, special_idx
special_tokens.insert(parts[0].to_string(), parts[1].parse::<i32>().unwrap());
}

let mut merges = BTreeMap::<(i32, i32), i32>::new();
let mut merges = IndexMap::<(i32, i32), i32>::new();
let mut idx = 256;
for line in lines {
let line = line.unwrap();
Expand All @@ -173,50 +230,18 @@ impl BPETokenizer {
idx += 1;
}

if verbose {
for ((p0, p1), idx) in &merges {
println!("({},{}): {}", p0, p1, idx);
}
}

self.merges = merges;

self.special_tokens = special_tokens;
self.vocab = build_vocab(&self.merges, &self.special_tokens);
Ok(())
}

pub fn encode(&self, text: &str, verbose: bool) -> Vec<i32> {
if verbose {
println!("[DEBUG] Encoding text: {}", text);
}
let mut ids: Vec<i32> = text.bytes().map(|b| b as i32).collect();
let merges = &self.merges;
if verbose {
println!("[DEBUG] Initial IDs: {:?}", ids);
println!("[DEBUG] Merges loaded: {:?}", merges);
}

while ids.len() >= 2 {
let stats = get_stats(&ids);
if verbose {
println!("[DEBUG] Current IDs: {:?}", ids);
println!("[DEBUG] Stats: {:?}", stats);
}
let pair = stats
.keys()
.min_by_key(|&&p| merges.get(&p).unwrap_or(&i32::MAX))
.copied()
.unwrap();

if !merges.contains_key(&pair) {
break;
}

let new_id = merges[&pair];
ids = merge(&ids, pair, new_id);
if verbose {
println!("[DEBUG] Pair: {:?}", pair);
println!("[DEBUG] New ID for merge: {}", new_id);
}
}

if verbose {
println!("[DEBUG] Encoding complete. Final IDs: {:?}", ids);
}
ids
}
}
}
Loading