Merge pull request #3350 from AI-Hypercomputer:jimmytsai/add-template-check-for-sft

PiperOrigin-RevId: 893538846

Author: Google-ML-Automation

src/maxtext/input_pipeline/hf_data_processing.py

@@ -302,6 +302,7 @@ def preprocessing_pipeline(
   )
   operations = []
   if use_sft:
+    input_pipeline_utils.verify_chat_template_generation_prompt_logic(tokenizer)
     operations.append(
         input_pipeline_utils.SFTPromptMasking(
             text_column_name=data_column_names[0],

src/maxtext/input_pipeline/input_pipeline_utils.py

@@ -19,6 +19,8 @@
 from threading import current_thread
 from typing import Any, Iterable, TYPE_CHECKING
 
+from jinja2 import TemplateError
+
 if TYPE_CHECKING:
   import datasets
   import tensorflow as tf
@@ -178,6 +180,103 @@ def is_conversational(features, data_columns):
   return False
 
 
+def _extract_token_ids(tokens):
+  """Extracts token IDs from various tokenizer output formats.
+
+  This helper function standardizes the extraction of tokenized integer IDs
+  from common return types of Hugging Face tokenizers, including
+  `BatchEncoding` objects, dictionaries, or simple lists.
+
+  Args:
+    tokens: The object containing token IDs. Supported types include:
+      - A list of integers.
+      - A dictionary containing the `INPUT_TOKENS_KEY`.
+      - An object (e.g., `BatchEncoding`) with an attribute named `INPUT_TOKENS_KEY`.
+
+  Returns:
+    A list of integer token IDs.
+
+  Raises:
+    ValueError: If the input type is not supported or does not contain the expected key.
+  """
+  # attention masks in BatchEncoding are effectively ignored
+  if hasattr(tokens, INPUT_TOKENS_KEY):
+    return getattr(tokens, INPUT_TOKENS_KEY)
+  elif isinstance(tokens, dict) and INPUT_TOKENS_KEY in tokens:
+    return tokens[INPUT_TOKENS_KEY]
+  elif isinstance(tokens, list):
+    return tokens
+  else:
+    raise ValueError(f"Can't extract token_ids from type {type(tokens)}")
+
+
+def verify_chat_template_generation_prompt_logic(tokenizer_model):
+  """Verifies the tokenizer's chat template for correct SFT loss masking.
+
+  This function ensures that the tokens added by `add_generation_prompt=True`
+  are identical to the tokens that begin an assistant's turn in a complete
+  conversation, which is critical for masking prompt tokens during SFT loss
+  calculation.
+
+  Example of a mismatch:
+    A `ValueError` is raised if the generation prompt and the actual
+    assistant prefix do not match. For example:
+
+    - `add_generation_prompt=True` on a user message produces a prompt ending in:
+      `...<|im_start|>generation\n`
+    - A full turn with an assistant message starts the reply with:
+      `...<|im_start|>assistant\n...`
+
+    This function would fail because the tokens for "generation" do not
+    match the tokens for "assistant".
+
+  Args:
+    tokenizer_model: The Hugging Face tokenizer instance to verify.
+
+  Raises:
+    ValueError: If the `add_generation_prompt` tokens do not exactly
+      match the beginning of an assistant message in the template.
+  """
+  dummy_msgs = [{"role": "system", "content": "System message"}, {"role": "user", "content": "Test message"}]
+
+  try:
+    prompt_wo_gen_tokens = tokenizer_model.apply_chat_template(dummy_msgs, add_generation_prompt=False, tokenize=True)
+  except TemplateError:
+    max_logging.info(
+        "Tokenizer failed to apply chat template with 'system' role. "
+        "Falling back to 'user' role only for chat template verification."
+    )
+    dummy_msgs.pop(0)
+    prompt_wo_gen_tokens = tokenizer_model.apply_chat_template(dummy_msgs, add_generation_prompt=False, tokenize=True)
+  prompt_wo_gen_ids = _extract_token_ids(prompt_wo_gen_tokens)
+
+  prompt_w_gen_tokens = tokenizer_model.apply_chat_template(dummy_msgs, add_generation_prompt=True, tokenize=True)
+  prompt_w_gen_ids = _extract_token_ids(prompt_w_gen_tokens)
+
+  if prompt_w_gen_ids[: len(prompt_wo_gen_ids)] != prompt_wo_gen_ids:
+    raise ValueError("Unable to extract generation prompt tokens.")
+  # Extract the tokenized generation prompt (the expected assistant prefix)
+  assistant_prefix = prompt_w_gen_ids[len(prompt_wo_gen_ids) :]
+  full_turn_tokens = _extract_token_ids(
+      tokenizer_model.apply_chat_template(
+          dummy_msgs + [{"role": "assistant", "content": "Dummy response"}], add_generation_prompt=False, tokenize=True
+      )
+  )
+  full_turn_ids = _extract_token_ids(full_turn_tokens)
+  # Extract the actual tokens that appear right after the user message in the full turn
+  actual_prefix_in_full_turn = full_turn_ids[len(prompt_wo_gen_ids) : len(prompt_wo_gen_ids) + len(assistant_prefix)]
+
+  if actual_prefix_in_full_turn != assistant_prefix:
+    expected_str = tokenizer_model.decode(assistant_prefix)
+    actual_str = tokenizer_model.decode(actual_prefix_in_full_turn)
+    raise ValueError(
+        "Chat template generation prompt mismatch!\n"
+        f"Expected assistant prefix tokens: {assistant_prefix} ('{expected_str}')\n"
+        f"Actual prefix tokens found: {actual_prefix_in_full_turn} ('{actual_str}')\n"
+        "This means the tokenizer's chat template will break the sft masking logic."
+    )
+
+
 def _get_completion_in_chat_template(tokenizer_model, round_msgs):
   """
   Calculates the completion part of a conversation turn when formatted with a chat template.
@@ -196,18 +295,8 @@ def _get_completion_in_chat_template(tokenizer_model, round_msgs):
   # include generation_prompt as part of the prompt tokens
   prompt_tokens = tokenizer_model.apply_chat_template(round_msgs[:-1], add_generation_prompt=True, tokenize=True)
 
-  # attention masks in BatchEncoding are effectively ignored
-  if hasattr(prompt_completion_tokens, INPUT_TOKENS_KEY):
-    prompt_completion_ids = getattr(prompt_completion_tokens, INPUT_TOKENS_KEY)
-    prompt_ids = getattr(prompt_tokens, INPUT_TOKENS_KEY)
-  elif isinstance(prompt_completion_tokens, dict) and INPUT_TOKENS_KEY in prompt_completion_tokens:
-    prompt_completion_ids = prompt_completion_tokens[INPUT_TOKENS_KEY]
-    prompt_ids = prompt_tokens[INPUT_TOKENS_KEY]
-  elif isinstance(prompt_completion_tokens, list):
-    prompt_completion_ids = prompt_completion_tokens
-    prompt_ids = prompt_tokens
-  else:
-    raise ValueError(f"Can't handle the chat template output of type {type(prompt_completion_tokens)}")
+  prompt_completion_ids = _extract_token_ids(prompt_completion_tokens)
+  prompt_ids = _extract_token_ids(prompt_tokens)
 
   completion_tokens = prompt_completion_ids[len(prompt_ids) :]
   completion_in_chat_template = tokenizer_model.decode(completion_tokens, skip_special_tokens=False)

tests/post_training/unit/sft_data_processing_test.py

@@ -22,17 +22,19 @@
 import os.path
 import numpy as np
 import jax
+import re
 from jax.sharding import Mesh
 from jax.experimental import mesh_utils
 from datasets import Dataset
 import transformers
 from parameterized import parameterized_class
-
+from unittest.mock import patch
 from maxtext.configs import pyconfig
 from maxtext.utils.globals import MAXTEXT_PKG_DIR, MAXTEXT_CONFIGS_DIR, MAXTEXT_ASSETS_ROOT
 from maxtext.input_pipeline import hf_data_processing
 from maxtext.input_pipeline import input_pipeline_interface
 from maxtext.input_pipeline.hf_data_processing import _get_pad_id
+from maxtext.input_pipeline.input_pipeline_utils import verify_chat_template_generation_prompt_logic
 
 PROMPT_DATA = [
     [
@@ -484,5 +486,51 @@ def test_system_message_not_at_beginning(self):
       self.get_data_iterator(dataset, ["messages"])
 
 
+@pytest.mark.external_training
+class SFTChatTemplateLogicTest(unittest.TestCase):
+  LLAMA_TOKENIZER_PATH = os.path.join(MAXTEXT_ASSETS_ROOT, "llama2-chat-tokenizer")
+
+  @classmethod
+  def setUpClass(cls):
+    super().setUpClass()
+    if not os.path.exists(cls.LLAMA_TOKENIZER_PATH):
+      exit_code = subprocess.call(
+          [
+              "gsutil",
+              "cp",
+              "-r",
+              "gs://maxtext-dataset/hf/llama2-chat-tokenizer",
+              os.path.join(MAXTEXT_ASSETS_ROOT, ""),
+          ]
+      )
+      if exit_code != 0:
+        raise ValueError("Failed to download llama tokenizer")
+
+  def setUp(self):
+    super().setUp()
+    self.qwen3_tokenizer = transformers.AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
+    self.llama2_tokenizer = transformers.AutoTokenizer.from_pretrained(self.LLAMA_TOKENIZER_PATH)
+
+  def test_tokenizer_w_generation_prompt(self):
+    verify_chat_template_generation_prompt_logic(self.qwen3_tokenizer)
+
+  def test_tokenizer_wo_generation_prompt(self):
+    verify_chat_template_generation_prompt_logic(self.llama2_tokenizer)
+
+  def test_failure_path_with_modified_template(self):
+    """Verifies the function correctly raises a ValueError on a bad template."""
+    # Replace the role within the existing add_generation_prompt block with a deliberately faulty one.
+    fault_chat_template = re.sub(
+        r"(\{%-?\s*if add_generation_prompt\s*%\}.*?<\|im_start\|>)assistant(.*?\{%-?\s*endif\s*%\})",
+        r"\1wrong_role\2",
+        self.qwen3_tokenizer.chat_template,
+        flags=re.DOTALL,
+    )
+    with patch.object(self.qwen3_tokenizer, "chat_template", fault_chat_template):
+      # Verify that our function catches the mismatch and raises the expected error
+      with self.assertRaisesRegex(ValueError, "Chat template generation prompt mismatch!"):
+        verify_chat_template_generation_prompt_logic(self.qwen3_tokenizer)
+
+
 if __name__ == "__main__":
   unittest.main()