Add safety comments (#1651)

* Unsafe comment for from_u32_unchecked

* Add safety comments and type assertion for HashSet parallel iteration

* Add safety comment for String splice

* fixes

* fmt

* pos

tokenizers/src/models/bpe/trainer.rs

@@ -454,7 +454,7 @@ impl BpeTrainer {
         // 3. Tokenize words
         //
         self.update_progress(&progress, word_counts.len(), "Tokenize words");
-        let (words, counts) =
+        let (mut words, counts) =
             self.tokenize_words(word_counts, &mut word_to_id, &mut id_to_word, &progress);
         self.finalize_progress(&progress, words.len());
 
@@ -530,14 +530,29 @@ impl BpeTrainer {
             merges.push((top.pair, new_token_id));
 
             // Merge the new pair in every words
-            let changes = top
+            // Safety: This is just a type assertion, the code below may no longer be safe
-                .pos
+            // if the type of `pos` changes
+            let pos: &HashSet<usize> = &top.pos;
+
+            let words_len = words.len();
+            struct WordPtr(*mut Word);
+            // Safety: We do not actually use this for concurrent access to the same memory,
+            // only to different chunks within the same allocation.
+            unsafe impl Sync for WordPtr {}
+            let word_start = WordPtr(words.as_mut_ptr());
+
+            let changes = pos
                 .maybe_par_iter()
                 .flat_map(|&i| {
-                    let word = &words[i] as *const _ as *mut Word;
+                    // Safety:
-                    // We can merge each of these words in parallel here because each position
+                    // We are producing a valid pointer since we are indexing in bounds
-                    // can be there only once (HashSet). So this is safe.
+                    //
+                    // We can access each `word` here in parallel because each position
+                    // can be there only once (pos is a HashSet).
                     unsafe {
+                        assert!(i < words_len);
+                        // This is words[i], but avoids needing to go through &T (which triggers UB)
+                        let word = word_start.0.add(i);
                         // let word: &mut Word = &mut (*word);
                         (*word)
                             .merge(top.pair.0, top.pair.1, new_token_id, max_token_length)

tokenizers/src/pre_tokenizers/byte_level.rs

@@ -28,6 +28,9 @@ pub(crate) fn bytes_char() -> HashMap<u8, char> {
         }
     }
 
+    // Safety: cs contains all values from bs (between 0 and 255),
+    // and some values of value 2⁸ + n, where n is between 0 and 255. This is between 255 and 512.
+    // Both ranges are valid UTF-32 values (which is fully saturated until 0xD000)
     bs.into_iter()
         .zip(cs)
         .map(|(f, t)| (f, unsafe { std::char::from_u32_unchecked(t) }))

tokenizers/src/tokenizer/normalizer.rs

@@ -411,8 +411,16 @@ impl NormalizedString {
             .collect::<String>();
 
         self.alignments.splice(n_range.clone(), alignments);
+
+        // This bounds check already happens above (`self.normalized[n_range.clone()]`), but future
+        // code could change to mutate `self` or `self.normalized` in the interim.
+        // Perform it again and hope the optimizer collapses it.
+        assert!(self.normalized.get(n_range.clone()).is_some());
         unsafe {
             self.normalized
+                // Safety: This is safe as long as we do not splice across a
+                // UTF-8 character, and we only add UTF-8 text. `normalized` is a String
+                // so the latter is trivially true, and we assert for the former above.
                 .as_mut_vec()
                 .splice(n_range, normalized.bytes());
         }