Update tokenization_code_llama_fast.py (huggingface#26576)

* Update tokenization_code_llama_fast.py * Update test_tokenization_code_llama.py * Update test_tokenization_code_llama.py
blbadger · Nov 8, 2023 · 6333195 · 6333195
1 parent 0f8dbc0
commit 6333195
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 1 deletion.
diff --git a/src/transformers/models/code_llama/tokenization_code_llama_fast.py b/src/transformers/models/code_llama/tokenization_code_llama_fast.py
@@ -278,7 +278,7 @@ def set_infilling_processor(self, reset, suffix_first=False, add_special_tokens=
         special_tokens = [(self.bos_token, self.bos_token_id)] if self.add_bos_token and add_special_tokens else []
         if suffix_first:
             # format as " <PRE> <SUF>{suf} <MID> {pre}"
-            pair += [self.prefix_token, self.suffix_token, "$A", self.middle_token, "$B"]
+            pair += [self.prefix_token, self.suffix_token, "$B", self.middle_token, "$A"]
             special_tokens += [
                 (self.prefix_token, self.prefix_id),
                 (self.suffix_token, self.suffix_id),

diff --git a/tests/models/code_llama/test_tokenization_code_llama.py b/tests/models/code_llama/test_tokenization_code_llama.py
@@ -643,3 +643,15 @@ def main():
         input_ids = tokenizer.encode(PROMPTS[0])
         self.assertEqual(input_ids, tokenizer.encode(prefix, suffix=suffix))
         self.assertEqual(tokenizer.encode(prefix, suffix=suffix), tokenizer_fast.encode(prefix, suffix=suffix))
+
+        # Adding suffix_first check for infilling tasks
+        suffix_first_formatted_prompt = tokenizer.tokenize(PROMPTS[0], suffix_first=True)
+        self.assertEqual(suffix_first_formatted_prompt, tokenizer_fast.tokenize(PROMPTS[0], suffix_first=True))
+        prefix, suffix = PROMPTS[0].split("<FILL_ME>")
+        self.assertEqual(suffix_first_formatted_prompt, tokenizer.tokenize(prefix, suffix, suffix_first=True))
+        self.assertEqual(suffix_first_formatted_prompt, tokenizer_fast.tokenize(prefix, suffix, suffix_first=True))
+
+        prefix, suffix = PROMPTS[0].split("<FILL_ME>")
+        suffix_first_input_ids = tokenizer.encode(PROMPTS[0], suffix_first=True)
+        self.assertEqual(suffix_first_input_ids, tokenizer.encode(prefix, suffix=suffix, suffix_first=True))
+        self.assertEqual(suffix_first_input_ids, tokenizer_fast.encode(prefix, suffix=suffix, suffix_first=True))