ai-edge-torch-nightly 0.3.0.dev20250114__py3-none-any.whl

ai_edge_torch/generative/examples/paligemma/verify_image_encoder.py

@@ -0,0 +1,99 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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
+#
+#     http://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.
+# ==============================================================================
+
+"""Verifies the reauthored image encoder of PaliGemma 3B model."""
+
+import logging
+import pathlib
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.paligemma import image_encoder
+import kagglehub
+from PIL import Image
+import requests
+import torch
+import transformers
+
+_VERSION = flags.DEFINE_enum(
+    "version",
+    "2",
+    ["1", "2"],
+    "The version of PaliGemma vision model to verify.",
+)
+_IMAGE_URL = flags.DEFINE_string(
+    "image_url",
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true",
+    "The image URI to encode.",
+)
+
+_CHECKPOINT = {
+    "1": "google/paligemma-3b-mix-224",
+    "2": "google/paligemma-2/transformers/paligemma2-3b-pt-224",
+}
+
+
+def main(_):
+  if _VERSION.value == "1":
+    checkpoint = _CHECKPOINT[_VERSION.value]
+    # Locate the cached dir.
+    cached_config_file = transformers.utils.cached_file(
+        checkpoint, transformers.utils.CONFIG_NAME
+    )
+    reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  else:
+    checkpoint = kagglehub.model_download(_CHECKPOINT[_VERSION.value])
+    reauthored_checkpoint = checkpoint
+
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_full_model = (
+      transformers.PaliGemmaForConditionalGeneration.from_pretrained(checkpoint)
+  )
+  original_vision_model = original_full_model.eval().vision_tower
+
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
+  reauthored_model = image_encoder.build_image_encoder(reauthored_checkpoint)
+
+  logging.info("Loading the processor from: %s", checkpoint)
+  # It works only when GemmaTokenizerFast is available. In some environments,
+  # use_fast=False doeesn't work either if the tokenizer cannot load the
+  # sentencepiece model file properly.
+  processor = transformers.AutoProcessor.from_pretrained(checkpoint)
+
+  logging.info("Loading the image from: %s", _IMAGE_URL.value)
+  image = Image.open(requests.get(_IMAGE_URL.value, stream=True).raw)
+  pixel_values = processor(images=image, return_tensors="pt")["pixel_values"]
+
+  logging.info("Forwarding the original model...")
+  outputs_original = original_vision_model.forward(pixel_values=pixel_values)
+  outputs_original = outputs_original.last_hidden_state
+  logging.info("outputs_original: %s", outputs_original)
+
+  logging.info("Forwarding the reauthored model...")
+  outputs_reauthored = reauthored_model.forward(pixel_values=pixel_values)
+  logging.info("outputs_reauthored: %s", outputs_reauthored)
+
+  try:
+    assert torch.allclose(
+        outputs_original, outputs_reauthored, atol=1e-03, rtol=1e-04
+    )
+  except AssertionError as e:
+    logging.error("*** FAILED *** verify with an image")
+    raise e
+  else:
+    logging.info("*** PASSED *** verify with an image")
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/examples/phi/__init__.py

@@ -0,0 +1,14 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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
+#
+#     http://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.
+# ==============================================================================

ai_edge_torch/generative/examples/phi/convert_phi3_to_tflite.py

@@ -0,0 +1,80 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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
+#
+#     http://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.
+# ==============================================================================
+
+"""Example of converting a Phi-3.5 model to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.phi import phi3
+from ai_edge_torch.generative.utilities import converter
+from ai_edge_torch.generative.utilities.model_builder import ExportConfig
+
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/phi3'),
+    'The path to the model checkpoint, or directory holding the checkpoint.',
+)
+_OUTPUT_PATH = flags.DEFINE_string(
+    'output_path',
+    '/tmp/',
+    'The path to export the tflite model.',
+)
+_OUTPUT_NAME_PREFIX = flags.DEFINE_string(
+    'output_name_prefix',
+    'phi3',
+    'The prefix of the output tflite model name.',
+)
+_PREFILL_SEQ_LENS = flags.DEFINE_multi_integer(
+    'prefill_seq_lens',
+    (8, 64, 128, 256, 512, 1024),
+    'List of the maximum sizes of prefill input tensors.',
+)
+_KV_CACHE_MAX_LEN = flags.DEFINE_integer(
+    'kv_cache_max_len',
+    1280,
+    'The maximum size of KV cache buffer, including both prefill and decode.',
+)
+_QUANTIZE = flags.DEFINE_bool(
+    'quantize',
+    True,
+    'Whether the model should be quantized.',
+)
+_LORA_RANKS = flags.DEFINE_multi_integer(
+    'lora_ranks',
+    None,
+    'If set, the model will be converted with the provided list of LoRA ranks.',
+)
+
+
+def main(_):
+  pytorch_model = phi3.build_model(
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  converter.convert_to_tflite(
+      pytorch_model,
+      output_path=_OUTPUT_PATH.value,
+      output_name_prefix=_OUTPUT_NAME_PREFIX.value,
+      prefill_seq_len=_PREFILL_SEQ_LENS.value,
+      quantize=_QUANTIZE.value,
+      lora_ranks=_LORA_RANKS.value,
+      export_config=ExportConfig(),
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/phi/convert_to_tflite.py

@@ -0,0 +1,80 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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
+#
+#     http://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.
+# ==============================================================================
+
+"""Example of converting a Phi-2 model to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.phi import phi2
+from ai_edge_torch.generative.utilities import converter
+from ai_edge_torch.generative.utilities.model_builder import ExportConfig
+
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/phi2'),
+    'The path to the model checkpoint, or directory holding the checkpoint.',
+)
+_OUTPUT_PATH = flags.DEFINE_string(
+    'output_path',
+    '/tmp/',
+    'The path to export the tflite model.',
+)
+_OUTPUT_NAME_PREFIX = flags.DEFINE_string(
+    'output_name_prefix',
+    'phi2',
+    'The prefix of the output tflite model name.',
+)
+_PREFILL_SEQ_LENS = flags.DEFINE_multi_integer(
+    'prefill_seq_lens',
+    (8, 64, 128, 256, 512, 1024),
+    'List of the maximum sizes of prefill input tensors.',
+)
+_KV_CACHE_MAX_LEN = flags.DEFINE_integer(
+    'kv_cache_max_len',
+    1280,
+    'The maximum size of KV cache buffer, including both prefill and decode.',
+)
+_QUANTIZE = flags.DEFINE_bool(
+    'quantize',
+    True,
+    'Whether the model should be quantized.',
+)
+_LORA_RANKS = flags.DEFINE_multi_integer(
+    'lora_ranks',
+    None,
+    'If set, the model will be converted with the provided list of LoRA ranks.',
+)
+
+
+def main(_):
+  pytorch_model = phi2.build_model(
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  converter.convert_to_tflite(
+      pytorch_model,
+      output_path=_OUTPUT_PATH.value,
+      output_name_prefix=_OUTPUT_NAME_PREFIX.value,
+      prefill_seq_len=_PREFILL_SEQ_LENS.value,
+      quantize=_QUANTIZE.value,
+      lora_ranks=_LORA_RANKS.value,
+      export_config=ExportConfig(),
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/phi/phi2.py

@@ -0,0 +1,107 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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
+#
+#     http://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.
+# ==============================================================================
+
+"""Example of building a Phi-2 model."""
+
+import ai_edge_torch.generative.layers.model_config as cfg
+from ai_edge_torch.generative.utilities import model_builder
+import ai_edge_torch.generative.utilities.loader as loading_utils
+from torch import nn
+
+TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
+    ff_up_proj="model.layers.{}.mlp.fc1",
+    ff_down_proj="model.layers.{}.mlp.fc2",
+    attn_query_proj="model.layers.{}.self_attn.q_proj",
+    attn_key_proj="model.layers.{}.self_attn.k_proj",
+    attn_value_proj="model.layers.{}.self_attn.v_proj",
+    attn_output_proj="model.layers.{}.self_attn.dense",
+    pre_attn_norm="model.layers.{}.input_layernorm",
+    embedding="model.embed_tokens",
+    final_norm="model.final_layernorm",
+    lm_head="lm_head",
+)
+
+
+class Phi2(model_builder.DecoderOnlyModel):
+  """A Phi-2 model built from the Edge Generative API layers."""
+  pass
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Phi-2 model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a Phi-2 model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=32,
+      head_dim=80,
+      num_query_groups=32,
+      rotary_base=10000,
+      rotary_percentage=0.4,
+      qkv_use_bias=True,
+      output_proj_use_bias=True,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.SEQUENTIAL,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
+      intermediate_size=10240,
+      use_bias=True,
+  )
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.LAYER_NORM,
+  )
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      parallel_residual=True,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=51200,
+      num_layers=32,
+      max_seq_len=2048,
+      kv_cache_max_len=kv_cache_max_len,
+      embedding_dim=2560,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      lm_head_use_bias=True,
+      lm_head_share_weight_with_embedding=False,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
+  config = get_model_config(kv_cache_max_len)
+  config.vocab_size = 128
+  config.num_layers = 2
+  config.max_seq_len = 2 * kv_cache_max_len
+  # Phi-2 has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 128
+  return config
+
+
+def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+      model_class=Phi2,
+  )

ai_edge_torch/generative/examples/phi/phi3.py

@@ -0,0 +1,219 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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
+#
+#     http://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.
+# ==============================================================================
+
+"""Example of building a Phi-3.5 model up to 4K tokens, not to 128K tokens."""
+
+from functools import partial
+import math
+from typing import Tuple
+
+import ai_edge_torch.generative.layers.model_config as cfg
+from ai_edge_torch.generative.utilities import model_builder
+import ai_edge_torch.generative.utilities.loader as loading_utils
+import torch
+
+TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
+    ff_up_proj="model.layers.{}.mlp.gate_up_proj",
+    ff_down_proj="model.layers.{}.mlp.down_proj",
+    attn_fused_qkv_proj="model.layers.{}.self_attn.qkv_proj",
+    attn_output_proj="model.layers.{}.self_attn.o_proj",
+    pre_attn_norm="model.layers.{}.input_layernorm",
+    post_attn_norm="model.layers.{}.post_attention_layernorm",
+    embedding="model.embed_tokens",
+    final_norm="model.norm",
+    lm_head="lm_head",
+)
+
+# max_position_embeddings / original_max_position_embeddings in Phi-3.5 config.
+ROPE_SCALE_FACTOR = 32
+
+# ROPE short factor in Phi-3.5 config. According to LOPE paper and its code in
+# https://github.com/microsoft/LongRoPE, these values had been searched with
+# min=1.0, step-0.01 to optimize the errors of sample dataset.
+ROPE_SHORT_FACTOR = [
+    1.0,
+    1.0199999809265137,
+    1.0299999713897705,
+    1.0299999713897705,
+    1.0499999523162842,
+    1.0499999523162842,
+    1.0499999523162842,
+    1.0499999523162842,
+    1.0499999523162842,
+    1.0699999332427979,
+    1.0999999046325684,
+    1.1099998950958252,
+    1.1599998474121094,
+    1.1599998474121094,
+    1.1699998378753662,
+    1.2899998426437378,
+    1.339999794960022,
+    1.679999828338623,
+    1.7899998426437378,
+    1.8199998140335083,
+    1.8499997854232788,
+    1.8799997568130493,
+    1.9099997282028198,
+    1.9399996995925903,
+    1.9899996519088745,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0799996852874756,
+    2.0899996757507324,
+    2.189999580383301,
+    2.2199995517730713,
+    2.5899994373321533,
+    2.729999542236328,
+    2.749999523162842,
+    2.8399994373321533,
+]
+
+
+def _build_phi3_rope(
+    input_pos: int,
+    n_elem: int,
+    base: int,
+    condense_ratio: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    theta_factors: torch.Tensor,
+    scale: float,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+  """Computes Rotary Positional Embeddings for Phi-3.5 model.
+
+  It's a modified version of attn_utils.build_rope_cache with additional
+  arguments for Phi-3.5 model. It precompute Rotary Positional Embedding Sin and
+  Cos values with scaling factors for quick lookup during the inference.
+
+  Args:
+      input_pos (torch.Tensor): the given input sequence positions
+      n_elem (int): Each sequence's dimmension.
+      base (int, optional): Rope base value.
+      condense_ratio (int, optional): The ratio by which sequence indicies are
+        condensed.
+      dtype (torch.dtype, optional): Output tensor's data type.
+      device (torch.device, optional): Output tensor's data type.
+      theta_factors (torch.Tensor, optional): A tensor of shape (n_elem,) used
+        to scale the theta values.
+      scale (float, optional): A float used to scale the rope values.
+
+  Returns:
+      Tuple[torch.Tensor, torch.Tensor]: Rope's Cosine and Sine waves.
+  """
+  theta = 1.0 / (base ** (torch.arange(0, n_elem, 2).float() / n_elem))
+  theta = theta / theta_factors
+  seq_idx = input_pos / condense_ratio
+  idx_theta = torch.outer(seq_idx, theta)
+  cos = torch.cos(idx_theta).to(dtype=dtype, device=device) * scale
+  sin = torch.sin(idx_theta).to(dtype=dtype, device=device) * scale
+  return cos, sin
+
+
+class Phi3_5Mini(model_builder.DecoderOnlyModel):
+  """A Phi-3.5 model built from the Edge Generative API layers."""
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__(config)
+    attn_config = self.config.block_config(0).attn_config
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Phi-3.5 model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a Phi-2 model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=32,
+      head_dim=96,
+      num_query_groups=32,
+      rotary_base=10000,
+      rotary_percentage=1.0,
+      qkv_transpose_before_split=True,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.SEQUENTIAL,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU_GLU),
+      intermediate_size=8192,
+  )
+  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
+
+  max_seq_len = 4096
+  # Create the RoPE callable
+  build_rope = partial(
+      _build_phi3_rope,
+      condense_ratio=1,
+      dtype=torch.float32,
+      device=torch.device("cpu"),
+      theta_factors=torch.tensor(ROPE_SHORT_FACTOR),
+      scale=math.sqrt(1 + math.log(ROPE_SCALE_FACTOR) / math.log(max_seq_len)),
+  )
+
+  config = cfg.ModelConfig(
+      vocab_size=32064,
+      num_layers=32,
+      max_seq_len=max_seq_len,
+      kv_cache_max_len=kv_cache_max_len,
+      embedding_dim=3072,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      lm_head_share_weight_with_embedding=False,
+      enable_hlfb=True,
+      build_rope=build_rope,
+  )
+  return config
+
+
+def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
+  config = get_model_config(kv_cache_max_len)
+  config.vocab_size = 128
+  config.num_layers = 2
+  config.max_seq_len = 2 * kv_cache_max_len
+  # Phi-3.5 has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 128
+  return config
+
+
+def build_model(checkpoint_path: str, **kwargs) -> torch.nn.Module:
+  """Instantiates the model instance and load checkpoint if provided."""
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+      model_class=Phi3_5Mini,
+  )

ai_edge_torch/generative/examples/phi/verify.py

@@ -0,0 +1,64 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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
+#
+#     http://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.
+# ==============================================================================
+
+"""Verifies the reauthored Phi-2 model."""
+import logging
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.phi import phi2
+from ai_edge_torch.generative.utilities import transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+import kagglehub
+import transformers
+
+
+_PROMPTS = flags.DEFINE_multi_string(
+    "prompts",
+    "Instruct: Write an email about the weather Output:",
+    "The input prompts to generate answers.",
+)
+_MAX_NEW_TOKENS = flags.DEFINE_integer(
+    "max_new_tokens",
+    30,
+    "The maximum size of the generated tokens.",
+)
+
+
+def main(_):
+  checkpoint = kagglehub.model_download("Microsoft/phi/transformers/2")
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
+
+  logging.info("Building the reauthored model from: %s", checkpoint)
+  reauthored_model = phi2.build_model(checkpoint)
+
+  logging.info("Loading the tokenizer from: %s", checkpoint)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
+
+  verifier.verify_reauthored_model(
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
+      generate_prompts=_PROMPTS.value,
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+      atol=1e-03,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)