Add roundtrip ckpt conversion tests in CI

src/maxtext/checkpoint_conversion/to_maxtext.py

@@ -51,6 +51,7 @@
 
 import argparse
 from functools import partial
+import importlib.util
 import json
 import os
 import sys
@@ -59,7 +60,6 @@
 from typing import Any, Callable, List, Sequence
 import absl
 import ml_dtypes
-import torch
 import flax.linen as nn
 from huggingface_hub import hf_hub_download, list_repo_files
 import jax
@@ -79,6 +79,21 @@
 from safetensors import safe_open
 
 
+def _lazy_import(name: str):
+  spec = importlib.util.find_spec(name)
+  if spec is None:
+    return None
+  loader = importlib.util.LazyLoader(spec.loader)
+  spec.loader = loader
+  module = importlib.util.module_from_spec(spec)
+  sys.modules[name] = module
+  loader.exec_module(module)
+  return module
+
+
+torch = _lazy_import("torch")
+
+
 absl.logging.set_verbosity(absl.logging.INFO)  # for max_logging.log
 
 
@@ -807,7 +822,7 @@ def _merged_getter(key):
 def main(
     args: Sequence[str],
     lazy_load_tensors: bool = False,
-    eager_load_method: str = "transformers",
+    eager_load_method: str = "safetensors",
     hf_model_path: str | None = None,
     revision: str | None = None,
     save_dtype: str = "bfloat16",
@@ -894,9 +909,9 @@ def main(
       #      (e.g., Multi-Token Prediction weights (`layers.61`) in DeepSeek-V3).
       #
       # Recommendation:
-      # - Use 'transformers' as the default for backward compatibility of mapping.
-      # - 'safetensors' is an interchangeable and valid alternative for most models,
-      #   and is strictly required if the model or specific weights lack Transformers support.
+      # - Use 'safetensors' as the default. Since transformers 5.8.0, model initialization
+      #   changed and the 'transformers' method may produce different key structures.
+      # - Use 'transformers' only if explicitly needed for backward-compatible key mapping (e.g. Gemma3).
       if eager_load_method == "transformers":
         max_logging.log("Eager load with Transformers backend, from_pretrained with auto dtype")
         # For auto mode, loaded dtype is the same as `dtype` specified in config.json (or `torch_dtype` for older version)

src/maxtext/configs/pyconfig.py

@@ -295,7 +295,10 @@ def __getattr__(self, attr: str) -> Any:
     # This is necessary for proper pickling/unpickling support
     flat_config = object.__getattribute__(self, "_flat_config")
     if attr in flat_config:
-      return flat_config[attr]
+      val = flat_config[attr]
+      if isinstance(val, dict) and attr in ("debug", "rl", "lora"):
+        return getattr(object.__getattribute__(self, "_pydantic_config"), attr)
+      return val
     raise AttributeError(f"'{self.__class__.__name__}' object has no attribute '{attr}'")
 
   def __setattr__(self, attr: str, value: Any) -> None:

tests/integration/checkpoint_conversion_test.py

@@ -0,0 +1,93 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Integration test for checkpoint conversion."""
+
+import os
+import subprocess
+import sys
+import tempfile
+import unittest
+import pytest
+
+pytestmark = [pytest.mark.integration_test]
+
+
+class Qwen3CheckpointConversionTest(unittest.TestCase):
+  """Tests HuggingFace to Orbax checkpoint conversion."""
+
+  @pytest.mark.cpu_only
+  def test_qwen3_30b_a3b_roundtrip_conversion(self):
+    model_name = "qwen3-30b-a3b"
+    base_num_decoder_layers = 2
+    with tempfile.TemporaryDirectory() as tmpdir:
+      to_mt_cmd = [
+          sys.executable,
+          "-m",
+          "maxtext.checkpoint_conversion.to_maxtext",
+          f"model_name={model_name}",
+          f"base_output_directory={tmpdir}",
+          f"base_num_decoder_layers={base_num_decoder_layers}",
+          "override_model_config=True",
+          "scan_layers=False",
+          "hardware=cpu",
+          "skip_jax_distributed_system=True",
+          "checkpoint_storage_use_ocdbt=False",
+          "checkpoint_storage_use_zarr3=False",
+          "--save_dtype=bfloat16",
+          "--lazy_load_tensors=True",
+      ]
+      env = os.environ.copy()
+      env["JAX_PLATFORMS"] = "cpu"
+      env["HF_HOME"] = tmpdir
+
+      print("Running checkpoint conversion command:", " ".join(to_mt_cmd))
+      # Inherit stdout and stderr from the parent process to stream logs in real time
+      subprocess.run(to_mt_cmd, env=env, check=True)
+
+      # Verify output directory exists and contains items
+      # Output structure: tmpdir/0/items
+      expected_dir = os.path.join(tmpdir, "0", "items")
+      self.assertTrue(os.path.exists(expected_dir), f"Expected checkpoint directory {expected_dir} does not exist.")
+      self.assertTrue(len(os.listdir(expected_dir)) > 0, f"Checkpoint directory {expected_dir} is empty.")
+
+      # Roundtrip conversion back to HuggingFace format
+      expected_hf_dir = os.path.join(tmpdir, "hf_safetensor", "qwen3-30b-a3b")
+      to_hf_cmd = [
+          sys.executable,
+          "-m",
+          "maxtext.checkpoint_conversion.to_huggingface",
+          f"model_name={model_name}",
+          f"load_parameters_path={expected_dir}",
+          f"base_output_directory={expected_hf_dir}",
+          f"base_num_decoder_layers={base_num_decoder_layers}",
+          "override_model_config=True",
+          "scan_layers=false",
+          "weight_dtype=bfloat16",
+          "hardware=cpu",
+          "skip_jax_distributed_system=True",
+          "--override_model_architecture=True",
+      ]
+      print("Running roundtrip checkpoint conversion command (to HF):", " ".join(to_hf_cmd))
+      subprocess.run(to_hf_cmd, env=env, check=True)
+
+      # Verify HF output directory exists and contains safetensors/config files
+      self.assertTrue(
+          os.path.exists(expected_hf_dir), f"Expected HF checkpoint directory {expected_hf_dir} does not exist."
+      )
+      self.assertTrue(len(os.listdir(expected_hf_dir)) > 0, f"HF checkpoint directory {expected_hf_dir} is empty.")
+
+
+if __name__ == "__main__":
+  unittest.main()