ai-edge-torch-nightly 0.3.0.dev20240924__py3-none-any.whl → 0.3.0.dev20240928__py3-none-any.whl

ai_edge_torch/generative/examples/smollm/verify.py

@@ -15,43 +15,53 @@
 
 """Verifies the reauthored SmolLM-135M model."""
 
+import logging
 import pathlib
 
 from absl import app
 from absl import flags
 from ai_edge_torch.generative.examples.smollm import smollm
+from ai_edge_torch.generative.utilities import transformers_verifier
 from ai_edge_torch.generative.utilities import verifier
 import transformers
 
+
 _PROMPTS = flags.DEFINE_multi_string(
     "prompts",
     "What is the meaning of life?",
     "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 = "HuggingFaceTB/SmolLM-135M"
-  verifier.log_msg("Loading the original model from", checkpoint)
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(checkpoint),
-  )
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
+
   # Locate the cached dir.
   cached_config_file = transformers.utils.cached_file(
       checkpoint, transformers.utils.CONFIG_NAME
   )
   reauthored_checkpoint = pathlib.Path(cached_config_file).parent
-  verifier.log_msg("Building the reauthored model from", reauthored_checkpoint)
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
   reauthored_model = smollm.build_model(reauthored_checkpoint)
 
-  verifier.log_msg("Loading the tokenizer from", checkpoint)
+  logging.info("Loading the tokenizer from: %s", checkpoint)
   tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
 
   verifier.verify_reauthored_model(
-      original_model=wrapper_model,
-      reauthored_model=reauthored_model,
-      tokenizer=tokenizer,
+      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-04,
   )
 

ai_edge_torch/generative/examples/stable_diffusion/clip.py

@@ -48,7 +48,7 @@ TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
 
 
 class CLIP(nn.Module):
-  """CLIP text encoder
+  """CLIP text encoder.
 
   For details, see https://arxiv.org/abs/2103.00020
   """
@@ -86,6 +86,7 @@ class CLIP(nn.Module):
 
 
 def get_model_config() -> cfg.ModelConfig:
+  """Get configs for the CLIP of Stable Diffusion v1.5."""
   max_seq_len = 77
   vocab_size = 49408
   num_layers = 12
@@ -97,6 +98,58 @@ def get_model_config() -> cfg.ModelConfig:
       num_heads=num_heads,
       head_dim=embedding_dim // num_heads,
       num_query_groups=num_query_groups,
+      rotary_base=0,
+      rotary_percentage=0.0,
+      qkv_use_bias=True,
+      qkv_transpose_before_split=True,
+      qkv_fused_interleaved=False,
+      output_proj_use_bias=True,
+      enable_kv_cache=False,
+  )
+
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.SEQUENTIAL,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_QUICK),
+      intermediate_size=embedding_dim * 4,
+      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,
+      post_attention_norm_config=norm_config,
+  )
+
+  config = cfg.ModelConfig(
+      vocab_size=vocab_size,
+      num_layers=num_layers,
+      max_seq_len=max_seq_len,
+      embedding_dim=embedding_dim,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+  )
+
+  return config
+
+
+def get_fake_model_config() -> cfg.ModelConfig:
+  """Get fake configs for the CLIP of Stable Diffusion v1.5 for testing."""
+  max_seq_len = 6
+  vocab_size = 100
+  num_layers = 2
+  num_heads = 12
+  num_query_groups = 12
+  embedding_dim = 24
+
+  attn_config = cfg.AttentionConfig(
+      num_heads=num_heads,
+      head_dim=embedding_dim // num_heads,
+      num_query_groups=num_query_groups,
+      rotary_base=0,
       rotary_percentage=0.0,
       qkv_use_bias=True,
       qkv_transpose_before_split=True,

ai_edge_torch/generative/examples/stable_diffusion/decoder.py

@@ -295,6 +295,64 @@ def get_model_config() -> unet_cfg.AutoEncoderConfig:
           enable_kv_cache=False,
           qkv_transpose_before_split=True,
           qkv_fused_interleaved=False,
+          rotary_base=0,
+          rotary_percentage=0.0,
+      ),
+      enable_hlfb=False,
+  )
+
+  mid_block_config = unet_cfg.MidBlock2DConfig(
+      in_channels=block_out_channels[-1],
+      normalization_config=norm_config,
+      activation_config=layers_cfg.ActivationConfig(
+          layers_cfg.ActivationType.SILU
+      ),
+      num_layers=1,
+      attention_block_config=att_config,
+  )
+
+  config = unet_cfg.AutoEncoderConfig(
+      in_channels=in_channels,
+      latent_channels=latent_channels,
+      out_channels=out_channels,
+      activation_config=layers_cfg.ActivationConfig(
+          layers_cfg.ActivationType.SILU
+      ),
+      block_out_channels=block_out_channels,
+      scaling_factor=scaling_factor,
+      layers_per_block=layers_per_block,
+      normalization_config=norm_config,
+      mid_block_config=mid_block_config,
+  )
+  return config
+
+
+def get_fake_model_config() -> unet_cfg.AutoEncoderConfig:
+  """Get fake configs for the Decoder of Stable Diffusion v1.5 for testing."""
+  in_channels = 3
+  latent_channels = 4
+  out_channels = 3
+  block_out_channels = [2, 4]
+  scaling_factor = 0.18215
+  layers_per_block = 2
+
+  norm_config = layers_cfg.NormalizationConfig(
+      layers_cfg.NormalizationType.GROUP_NORM, group_num=2
+  )
+
+  att_config = unet_cfg.AttentionBlock2DConfig(
+      dim=block_out_channels[-1],
+      normalization_config=norm_config,
+      attention_config=layers_cfg.AttentionConfig(
+          num_heads=1,
+          head_dim=block_out_channels[-1],
+          num_query_groups=1,
+          qkv_use_bias=True,
+          output_proj_use_bias=True,
+          enable_kv_cache=False,
+          qkv_transpose_before_split=True,
+          qkv_fused_interleaved=False,
+          rotary_base=0,
           rotary_percentage=0.0,
       ),
       enable_hlfb=False,

ai_edge_torch/generative/examples/stable_diffusion/diffusion.py

@@ -199,6 +199,7 @@ def build_attention_config(
     num_heads,
     dim,
     num_query_groups,
+    rotary_base=0,
     rotary_percentage=0.0,
     qkv_transpose_before_split=True,
     qkv_use_bias=False,
@@ -211,6 +212,7 @@ def build_attention_config(
       num_heads=num_heads,
       head_dim=dim // num_heads,
       num_query_groups=num_query_groups,
+      rotary_base=rotary_base,
       rotary_percentage=rotary_percentage,
       qkv_transpose_before_split=qkv_transpose_before_split,
       qkv_use_bias=qkv_use_bias,
@@ -603,7 +605,7 @@ def get_model_config(batch_size: int) -> unet_cfg.DiffusionModelConfig:
   # Transformer configs.
   transformer_num_attention_heads = 8
   transformer_batch_size = batch_size
-  transformer_cross_attention_dim = 768  # Embedding fomr CLIP model
+  transformer_cross_attention_dim = 768  # Embedding from CLIP model
   transformer_pre_conv_norm_config = layers_cfg.NormalizationConfig(
       layers_cfg.NormalizationType.GROUP_NORM, epsilon=1e-6, group_num=32
   )
@@ -645,3 +647,71 @@ def get_model_config(batch_size: int) -> unet_cfg.DiffusionModelConfig:
       final_norm_config=final_norm_config,
       final_activation_type=final_activation_type,
   )
+
+
+def get_fake_model_config(batch_size: int) -> unet_cfg.DiffusionModelConfig:
+  """Get fake configs for the Diffusion model of Stable Diffusion v1.5 for testing.
+
+  Args:
+    batch_size (int): the batch size of input.
+
+  Retruns:
+    The configuration of diffusion model of Stable Diffusion v1.5.
+  """
+  in_channels = 4
+  out_channels = 4
+  block_out_channels = [2, 4, 8, 8]
+  layers_per_block = 1
+  downsample_padding = 1
+
+  # Residual configs.
+  residual_norm_config = layers_cfg.NormalizationConfig(
+      layers_cfg.NormalizationType.GROUP_NORM, group_num=2
+  )
+  residual_activation_type = layers_cfg.ActivationType.SILU
+
+  # Transformer configs.
+  transformer_num_attention_heads = 1
+  transformer_batch_size = batch_size
+  transformer_cross_attention_dim = 4  # Embedding from CLIP model
+  transformer_pre_conv_norm_config = layers_cfg.NormalizationConfig(
+      layers_cfg.NormalizationType.GROUP_NORM, epsilon=1e-6, group_num=2
+  )
+  transformer_norm_config = layers_cfg.NormalizationConfig(
+      layers_cfg.NormalizationType.LAYER_NORM
+  )
+  transformer_ff_activation_type = layers_cfg.ActivationType.GE_GLU
+
+  # Time embedding configs.
+  time_embedding_dim = 2
+  time_embedding_blocks_dim = 4
+
+  # Mid block configs.
+  mid_block_layers = 1
+
+  # Finaly layer configs.
+  final_norm_config = layers_cfg.NormalizationConfig(
+      layers_cfg.NormalizationType.GROUP_NORM, group_num=2
+  )
+  final_activation_type = layers_cfg.ActivationType.SILU
+
+  return unet_cfg.DiffusionModelConfig(
+      in_channels=in_channels,
+      out_channels=out_channels,
+      block_out_channels=block_out_channels,
+      layers_per_block=layers_per_block,
+      downsample_padding=downsample_padding,
+      residual_norm_config=residual_norm_config,
+      residual_activation_type=residual_activation_type,
+      transformer_batch_size=transformer_batch_size,
+      transformer_num_attention_heads=transformer_num_attention_heads,
+      transformer_cross_attention_dim=transformer_cross_attention_dim,
+      transformer_pre_conv_norm_config=transformer_pre_conv_norm_config,
+      transformer_norm_config=transformer_norm_config,
+      transformer_ff_activation_type=transformer_ff_activation_type,
+      mid_block_layers=mid_block_layers,
+      time_embedding_dim=time_embedding_dim,
+      time_embedding_blocks_dim=time_embedding_blocks_dim,
+      final_norm_config=final_norm_config,
+      final_activation_type=final_activation_type,
+  )

ai_edge_torch/generative/examples/t5/t5.py

@@ -335,8 +335,6 @@ class T5Decoder(nn.Module):
 
     self.dec_attn_mask_cache = attn_utils.build_causal_mask_cache(
         size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
     )
 
   @torch.inference_mode

ai_edge_torch/generative/examples/test_models/convert_toy_model.py

@@ -0,0 +1,105 @@
+# 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.
+# ==============================================================================
+# A toy example which has a single-layer transformer block.
+from absl import app
+import ai_edge_torch
+from ai_edge_torch import lowertools
+from ai_edge_torch.generative.examples.test_models import toy_model
+from ai_edge_torch.generative.examples.test_models import toy_model_with_kv_cache
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
+import torch
+
+KV_CACHE_MAX_LEN = 100
+
+
+def convert_toy_model(_) -> None:
+  """Converts a toy model to tflite."""
+  model = toy_model.ToySingleLayerModel(toy_model.get_model_config())
+  idx = torch.unsqueeze(torch.arange(0, KV_CACHE_MAX_LEN), 0)
+  input_pos = torch.arange(0, KV_CACHE_MAX_LEN)
+  print('running an inference')
+  print(
+      model.forward(
+          idx,
+          input_pos,
+      )
+  )
+
+  # Convert model to tflite.
+  print('converting model to tflite')
+  edge_model = ai_edge_torch.convert(
+      model,
+      (
+          idx,
+          input_pos,
+      ),
+  )
+  edge_model.export('/tmp/toy_model.tflite')
+
+
+def _export_stablehlo_mlir(model, args):
+  ep = torch.export.export(model, args)
+  return lowertools.exported_program_to_mlir_text(ep)
+
+
+def convert_toy_model_with_kv_cache(_) -> None:
+  """Converts a toy model with kv cache to tflite."""
+  dump_mlir = False
+
+  config = toy_model_with_kv_cache.get_model_config()
+  model = toy_model_with_kv_cache.ToyModelWithKVCache(config)
+  model.eval()
+  print('running an inference')
+  kv = kv_utils.KVCache.from_model_config(config)
+
+  tokens, input_pos = toy_model_with_kv_cache.get_sample_prefill_inputs()
+  decode_token, decode_input_pos = (
+      toy_model_with_kv_cache.get_sample_decode_inputs()
+  )
+  print(model.forward(tokens, input_pos, kv))
+
+  if dump_mlir:
+    mlir_text = _export_stablehlo_mlir(model, (tokens, input_pos, kv))
+    with open('/tmp/toy_model_with_external_kv.stablehlo.mlir', 'w') as f:
+      f.write(mlir_text)
+
+  # Convert model to tflite with 2 signatures (prefill + decode).
+  print('converting toy model to tflite with 2 signatures (prefill + decode)')
+  edge_model = (
+      ai_edge_torch.signature(
+          'prefill',
+          model,
+          sample_kwargs={
+              'tokens': tokens,
+              'input_pos': input_pos,
+              'kv_cache': kv,
+          },
+      )
+      .signature(
+          'decode',
+          model,
+          sample_kwargs={
+              'tokens': decode_token,
+              'input_pos': decode_input_pos,
+              'kv_cache': kv,
+          },
+      )
+      .convert()
+  )
+  edge_model.export('/tmp/toy_external_kv_cache.tflite')
+
+
+if __name__ == '__main__':
+  app.run(convert_toy_model)

ai_edge_torch/generative/examples/test_models/toy_model.py

@@ -15,13 +15,12 @@
 # A toy example which has a single-layer transformer block.
 from typing import Tuple
 
-import ai_edge_torch
+from ai_edge_torch.generative.layers import builder
 from ai_edge_torch.generative.layers.attention import TransformerBlock
 import ai_edge_torch.generative.layers.attention_utils as attn_utils
-import ai_edge_torch.generative.layers.builder as builder
 import ai_edge_torch.generative.layers.model_config as cfg
 import torch
-import torch.nn as nn
+from torch import nn
 
 RoPECache = Tuple[torch.Tensor, torch.Tensor]
 KV_CACHE_MAX_LEN = 100
@@ -45,13 +44,10 @@ class ToySingleLayerModel(torch.nn.Module):
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.max_seq_len,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device('cpu'),
+        base=attn_config.rotary_base,
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.max_seq_len, dtype=torch.float32, device=torch.device('cpu')
+        size=config.max_seq_len,
     )
     self.config = config
 
@@ -94,13 +90,10 @@ class ToySingleLayerModelWeightSharing(torch.nn.Module):
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.max_seq_len,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device('cpu'),
+        base=attn_config.rotary_base,
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.max_seq_len, dtype=torch.float32, device=torch.device('cpu')
+        size=config.max_seq_len,
     )
     self.config = config
 
@@ -125,6 +118,7 @@ def get_model_config() -> cfg.ModelConfig:
       num_heads=32,
       head_dim=4,
       num_query_groups=4,
+      rotary_base=10000,
       rotary_percentage=1.0,
       enable_kv_cache=False,
   )
@@ -149,31 +143,3 @@ def get_model_config() -> cfg.ModelConfig:
       final_norm_config=norm_config,
   )
   return config
-
-
-def define_and_run() -> None:
-  model = ToySingleLayerModel(get_model_config())
-  idx = torch.unsqueeze(torch.arange(0, KV_CACHE_MAX_LEN), 0)
-  input_pos = torch.arange(0, KV_CACHE_MAX_LEN)
-  print('running an inference')
-  print(
-      model.forward(
-          idx,
-          input_pos,
-      )
-  )
-
-  # Convert model to tflite.
-  print('converting model to tflite')
-  edge_model = ai_edge_torch.convert(
-      model,
-      (
-          idx,
-          input_pos,
-      ),
-  )
-  edge_model.export('/tmp/toy_model.tflite')
-
-
-if __name__ == '__main__':
-  define_and_run()

ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py

@@ -17,15 +17,14 @@
 
 from typing import Tuple
 
-import ai_edge_torch
-from ai_edge_torch import lowertools
+from absl import app
 from ai_edge_torch.generative.layers import attention
 from ai_edge_torch.generative.layers import builder
 from ai_edge_torch.generative.layers import kv_cache as kv_utils
 import ai_edge_torch.generative.layers.attention_utils as attn_utils
 import ai_edge_torch.generative.layers.model_config as cfg
 import torch
-import torch.nn as nn
+from torch import nn
 
 RoPECache = Tuple[torch.Tensor, torch.Tensor]
 
@@ -52,13 +51,10 @@ class ToyModelWithKVCache(torch.nn.Module):
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.max_seq_len,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device('cpu'),
+        base=attn_config.rotary_base,
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.max_seq_len, dtype=torch.float32, device=torch.device('cpu')
+        size=config.max_seq_len,
     )
     self.config = config
 
@@ -87,16 +83,12 @@ class ToyModelWithKVCache(torch.nn.Module):
     return {'logits': self.lm_head(x), 'kv_cache': updated_kv_cache}
 
 
-def _export_stablehlo_mlir(model, args):
-  ep = torch.export.export(model, args)
-  return lowertools.exported_program_to_mlir_text(ep)
-
-
 def get_model_config() -> cfg.ModelConfig:
   attn_config = cfg.AttentionConfig(
       num_heads=32,
       head_dim=4,
       num_query_groups=4,
+      rotary_base=10000,
       rotary_percentage=1.0,
   )
   ff_config = cfg.FeedForwardConfig(
@@ -133,51 +125,3 @@ def get_sample_decode_inputs() -> Tuple[torch.Tensor, torch.Tensor]:
   tokens = torch.tensor([[1]], dtype=torch.int)
   input_pos = torch.tensor([10])
   return tokens, input_pos
-
-
-def define_and_run() -> None:
-  dump_mlir = False
-
-  config = get_model_config()
-  model = ToyModelWithExternalKV(config)
-  model.eval()
-  print('running an inference')
-  kv = kv_utils.KVCache.from_model_config(config)
-
-  tokens, input_pos = get_sample_prefill_inputs()
-  decode_token, decode_input_pos = get_sample_decode_inputs()
-  print(model.forward(tokens, input_pos, kv))
-
-  if dump_mlir:
-    mlir_text = _export_stablehlo_mlir(model, (tokens, input_pos, kv))
-    with open('/tmp/toy_model_with_external_kv.stablehlo.mlir', 'w') as f:
-      f.write(mlir_text)
-
-  # Convert model to tflite with 2 signatures (prefill + decode).
-  print('converting toy model to tflite with 2 signatures (prefill + decode)')
-  edge_model = (
-      ai_edge_torch.signature(
-          'prefill',
-          model,
-          sample_kwargs={
-              'tokens': tokens,
-              'input_pos': input_pos,
-              'kv_cache': kv,
-          },
-      )
-      .signature(
-          'decode',
-          model,
-          sample_kwargs={
-              'tokens': decode_token,
-              'input_pos': decode_input_pos,
-              'kv_cache': kv,
-          },
-      )
-      .convert()
-  )
-  edge_model.export('/tmp/toy_external_kv_cache.tflite')
-
-
-if __name__ == '__main__':
-  define_and_run()

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -67,15 +67,10 @@ class TinyLlama(nn.Module):
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.kv_cache_max,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
+        base=attn_config.rotary_base,
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
         size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
     )
     self.config = config
 
@@ -132,6 +127,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       num_heads=32,
       head_dim=64,
       num_query_groups=4,
+      rotary_base=10000,
       rotary_percentage=1.0,
   )
   ff_config = cfg.FeedForwardConfig(

ai_edge_torch/generative/examples/tiny_llama/verify.py

@@ -15,45 +15,55 @@
 
 """Verifies the reauthored TinyLlama-1.1B model."""
 
+import logging
 import pathlib
 
 from absl import app
 from absl import flags
 from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+from ai_edge_torch.generative.utilities import transformers_verifier
 from ai_edge_torch.generative.utilities import verifier
 import transformers
 
+
 _PROMPTS = flags.DEFINE_multi_string(
     "prompts",
     "Show me the program to add 2 and 3.",
     "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 = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
-  verifier.log_msg("Loading the original model from", checkpoint)
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(
-          checkpoint, trust_remote_code=True
-      ),
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(
+      checkpoint, trust_remote_code=True
   )
+
   # Locate the cached dir.
   cached_config_file = transformers.utils.cached_file(
       checkpoint, transformers.utils.CONFIG_NAME
   )
   reauthored_checkpoint = pathlib.Path(cached_config_file).parent
-  verifier.log_msg("Building the reauthored model from", reauthored_checkpoint)
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
   reauthored_model = tiny_llama.build_model(reauthored_checkpoint)
 
-  verifier.log_msg("Loading the tokenizer from", checkpoint)
+  logging.info("Loading the tokenizer from: %s", checkpoint)
   tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
 
   verifier.verify_reauthored_model(
-      original_model=wrapper_model,
-      reauthored_model=reauthored_model,
-      tokenizer=tokenizer,
+      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-04,
   )