ai-edge-torch-nightly 0.5.0.dev20250516__py3-none-any.whl → 0.5.0.dev20250518__py3-none-any.whl

ai_edge_torch/generative/examples/phi/verify_phi4.py

@@ -15,15 +15,10 @@
 
 """Verifies the reauthored Phi-4 model."""
 
-import logging
-import pathlib
 
 from absl import app
 from absl import flags
-from ai_edge_torch.generative.examples.phi import phi4
-from ai_edge_torch.generative.utilities import transformers_verifier
-from ai_edge_torch.generative.utilities import verifier
-import transformers
+from ai_edge_torch.generative.examples.phi import verify_util
 
 
 _PROMPTS = flags.DEFINE_multi_string(
@@ -39,29 +34,11 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
 
 
 def main(_):
-  checkpoint = "microsoft/Phi-4-mini-instruct"
-  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
-  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
-  reauthored_model = phi4.build_model(reauthored_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,
+  verify_util.verify_phi(
+      version="v4",
+      checkpoint_dir="microsoft/Phi-4-mini-instruct",
       max_new_tokens=_MAX_NEW_TOKENS.value,
+      prompts=_PROMPTS.value,
   )
 
 

ai_edge_torch/generative/examples/phi/verify_util.py

@@ -0,0 +1,84 @@
+# Copyright 2025 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.
+# ==============================================================================
+"""Utils for verifying the Phi model."""
+import logging
+import os
+import pathlib
+
+from ai_edge_torch.generative.examples.phi import phi2, phi3, phi4
+from ai_edge_torch.generative.utilities import loader
+from ai_edge_torch.generative.utilities import transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+import transformers
+
+
+DEFAULT_PROMPTS = ["Instruct: Write an email about the weather Output:"]
+
+_BUILDER = {
+    "v2": phi2.build_model,
+    "v3": phi3.build_model,
+    "v4": phi4.build_model,
+}
+
+
+def verify_phi(
+    version: str,
+    checkpoint_dir: str,
+    weight_filename: str = "model.safetensors",
+    max_new_tokens: int = 30,
+    initialize_from_local: bool = True,
+    prompts: list[str] | None = None,
+    atol: float = 1e-04,
+) -> bool:
+  """Verifies the reauthored Phi model with a custom loader."""
+  logging.info("Loading the original model from: %s", checkpoint_dir)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(
+      checkpoint_dir
+  )
+
+  logging.info("Building the reauthored model from: %s", checkpoint_dir)
+  custom_loader = (
+      None
+      if initialize_from_local
+      else loader.get_custom_loader("", "safetensors")
+  )
+
+  if initialize_from_local:
+    # Locate the cached dir.
+    cached_config_file = transformers.utils.cached_file(
+        checkpoint_dir, transformers.utils.CONFIG_NAME
+    )
+    reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  else:
+    reauthored_checkpoint = os.path.join(checkpoint_dir, weight_filename)
+
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
+  reauthored_model = _BUILDER[version](
+      checkpoint_path=reauthored_checkpoint,
+      custom_loader=custom_loader,
+  )
+
+  logging.info("Loading the tokenizer from: %s", checkpoint_dir)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint_dir)
+  return verifier.verify_reauthored_model(
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
+      generate_prompts=DEFAULT_PROMPTS if prompts is None else prompts,
+      max_new_tokens=max_new_tokens,
+      atol=atol,
+  )

ai_edge_torch/generative/examples/qwen/qwen.py

@@ -53,9 +53,7 @@ def get_3b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       intermediate_size=11008,
   )
   norm_config = cfg.NormalizationConfig(
-      type=cfg.NormalizationType.RMS_NORM,
-      epsilon=1e-06,
-      enable_hlfb=True,
+      type=cfg.NormalizationType.RMS_NORM, epsilon=1e-06
   )
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
@@ -71,7 +69,6 @@ def get_3b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       kv_cache_max_len=kv_cache_max_len,
       block_configs=block_config,
       final_norm_config=norm_config,
-      enable_hlfb=True,
   )
   return config
 

ai_edge_torch/generative/examples/qwen/verify.py

@@ -15,15 +15,10 @@
 
 """Verifies the reauthored Qwen 2.5 0.5B, 1.5B, and 3B models."""
 
-import logging
-import pathlib
 
 from absl import app
 from absl import flags
-from ai_edge_torch.generative.examples.qwen import qwen
-from ai_edge_torch.generative.utilities import transformers_verifier
-from ai_edge_torch.generative.utilities import verifier
-import transformers
+from ai_edge_torch.generative.examples.qwen import verify_util
 
 
 _MODEL_SIZE = flags.DEFINE_enum(
@@ -49,38 +44,13 @@ _CHECKPOINT = {
     "3b": "Qwen/Qwen2.5-3B-Instruct",
 }
 
-_BUILDER = {
-    "0.5b": qwen.build_0_5b_model,
-    "1.5b": qwen.build_1_5b_model,
-    "3b": qwen.build_3b_model,
-}
-
 
 def main(_):
-  checkpoint = _CHECKPOINT[_MODEL_SIZE.value]
-  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
-  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
-  reauthored_model = _BUILDER[_MODEL_SIZE.value](reauthored_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,
+  verify_util.verify_qwen(
+      model_size=_MODEL_SIZE.value,
+      checkpoint_dir=_CHECKPOINT[_MODEL_SIZE.value],
       max_new_tokens=_MAX_NEW_TOKENS.value,
-      atol=1e-04,
+      prompts=_PROMPTS.value,
   )
 
 

ai_edge_torch/generative/examples/qwen/verify_util.py

@@ -0,0 +1,83 @@
+# Copyright 2025 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.
+# ==============================================================================
+"""Utils for verifying the Qwen model."""
+import logging
+import os
+import pathlib
+
+from ai_edge_torch.generative.examples.qwen import qwen
+from ai_edge_torch.generative.utilities import loader
+from ai_edge_torch.generative.utilities import transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+import transformers
+
+
+_BUILDER = {
+    "0.5b": qwen.build_0_5b_model,
+    "1.5b": qwen.build_1_5b_model,
+    "3b": qwen.build_3b_model,
+}
+
+DEFAULT_PROMPTS = ["What is the meaning of life?"]
+
+
+def verify_qwen(
+    model_size: str,
+    checkpoint_dir: str,
+    weight_filename: str = "model.safetensors",
+    max_new_tokens: int = 30,
+    initialize_from_local: bool = True,
+    prompts: list[str] | None = None,
+) -> bool:
+  """Verifies the reauthored Llama 3.2 model with a custom loader."""
+  logging.info("Loading the original model from: %s", checkpoint_dir)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(
+      checkpoint_dir
+  )
+
+  logging.info("Building the reauthored model from: %s", checkpoint_dir)
+  custom_loader = (
+      None
+      if initialize_from_local
+      else loader.get_custom_loader("", "safetensors")
+  )
+
+  if initialize_from_local:
+    # Locate the cached dir.
+    cached_config_file = transformers.utils.cached_file(
+        checkpoint_dir, transformers.utils.CONFIG_NAME
+    )
+    reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  else:
+    reauthored_checkpoint = os.path.join(checkpoint_dir, weight_filename)
+
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
+  reauthored_model = _BUILDER[model_size](
+      checkpoint_path=reauthored_checkpoint,
+      custom_loader=custom_loader,
+  )
+
+  logging.info("Loading the tokenizer from: %s", checkpoint_dir)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint_dir)
+  return verifier.verify_reauthored_model(
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
+      generate_prompts=DEFAULT_PROMPTS if prompts is None else prompts,
+      max_new_tokens=max_new_tokens,
+      atol=1e-04,
+  )

ai_edge_torch/generative/examples/qwen_vl/decoder.py

@@ -97,7 +97,7 @@ def get_decoder_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       intermediate_size=11008,
   )
   norm_config = cfg.NormalizationConfig(
-      type=cfg.NormalizationType.RMS_NORM, epsilon=1e-06, enable_hlfb=True
+      type=cfg.NormalizationType.RMS_NORM, epsilon=1e-06
   )
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
@@ -113,7 +113,6 @@ def get_decoder_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       kv_cache_max_len=kv_cache_max_len,
       block_configs=block_config,
       final_norm_config=norm_config,
-      enable_hlfb=True,
   )
   return config
 

ai_edge_torch/generative/examples/qwen_vl/image_encoder.py

@@ -332,8 +332,7 @@ def get_image_encoder_config(image_size: Tuple[int, int]) -> QwenVLImageConfig:
       use_bias=True,
   )
   norm_config = cfg.NormalizationConfig(
-      type=cfg.NormalizationType.RMS_NORM,
-      epsilon=1e-6,
+      type=cfg.NormalizationType.RMS_NORM, epsilon=1e-6
   )
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
@@ -359,7 +358,6 @@ def get_image_encoder_config(image_size: Tuple[int, int]) -> QwenVLImageConfig:
       window_size=112,
       spatial_merge_size=2,
       full_atten_block_indexes=[7, 15, 23, 31],
-      enable_hlfb=True,
   )
   return config
 

ai_edge_torch/generative/examples/smollm/smollm.py

@@ -51,9 +51,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
       intermediate_size=1536,
   )
-  norm_config = cfg.NormalizationConfig(
-      type=cfg.NormalizationType.RMS_NORM, enable_hlfb=True
-  )
+  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
       ff_config=ff_config,
@@ -68,7 +66,6 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       kv_cache_max_len=kv_cache_max_len,
       block_configs=block_config,
       final_norm_config=norm_config,
-      enable_hlfb=True,
   )
   return config
 

ai_edge_torch/generative/examples/smollm/verify.py

@@ -15,15 +15,9 @@
 
 """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
+from ai_edge_torch.generative.examples.smollm import verify_util
 
 
 _PROMPTS = flags.DEFINE_multi_string(
@@ -47,38 +41,13 @@ _CHECKPOINT = {
     "v2": "HuggingFaceTB/SmolLM2-135M-Instruct",
 }
 
-_BUILDER = {
-    "v1": smollm.build_model,
-    "v2": smollm.build_model_v2,
-}
-
 
 def main(_):
-  checkpoint = _CHECKPOINT[_MODEL_VERSION.value]
-  builder = _BUILDER[_MODEL_VERSION.value]
-  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
-  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
-  reauthored_model = builder(reauthored_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,
+  verify_util.verify_smollm_135m(
+      model_version=_MODEL_VERSION.value,
+      checkpoint_dir=_CHECKPOINT[_MODEL_VERSION.value],
       max_new_tokens=_MAX_NEW_TOKENS.value,
-      atol=1e-04,
+      prompts=_PROMPTS.value,
   )
 
 

ai_edge_torch/generative/examples/smollm/verify_util.py

@@ -0,0 +1,81 @@
+# Copyright 2025 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.
+# ==============================================================================
+"""Utils for verifying the SmoLLM model."""
+import logging
+import os
+import pathlib
+
+from ai_edge_torch.generative.examples.smollm import smollm
+from ai_edge_torch.generative.utilities import loader
+from ai_edge_torch.generative.utilities import transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+import transformers
+
+_BUILDER = {
+    "v1": smollm.build_model,
+    "v2": smollm.build_model_v2,
+}
+
+DEFAULT_PROMPTS = ["What is the meaning of life?"]
+
+
+def verify_smollm_135m(
+    model_version: str,
+    checkpoint_dir: str,
+    weight_filename: str = "model.safetensors",
+    max_new_tokens: int = 30,
+    initialize_from_local: bool = True,
+    prompts: list[str] | None = None,
+) -> bool:
+  """Verifies the reauthored SmoLLM model with a custom loader."""
+  logging.info("Loading the original model from: %s", checkpoint_dir)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(
+      checkpoint_dir
+  )
+
+  logging.info("Building the reauthored model from: %s", checkpoint_dir)
+  custom_loader = (
+      None
+      if initialize_from_local
+      else loader.get_custom_loader("", "safetensors")
+  )
+
+  if initialize_from_local:
+    # Locate the cached dir.
+    cached_config_file = transformers.utils.cached_file(
+        checkpoint_dir, transformers.utils.CONFIG_NAME
+    )
+    reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  else:
+    reauthored_checkpoint = os.path.join(checkpoint_dir, weight_filename)
+
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
+  reauthored_model = _BUILDER[model_version](
+      checkpoint_path=reauthored_checkpoint,
+      custom_loader=custom_loader,
+  )
+
+  logging.info("Loading the tokenizer from: %s", checkpoint_dir)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint_dir)
+  return verifier.verify_reauthored_model(
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
+      generate_prompts=DEFAULT_PROMPTS if prompts is None else prompts,
+      max_new_tokens=max_new_tokens,
+      atol=1e-04,
+  )

ai_edge_torch/generative/examples/stable_diffusion/clip.py

@@ -113,7 +113,9 @@ def get_model_config() -> cfg.ModelConfig:
       use_bias=True,
   )
 
-  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.LAYER_NORM)
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.LAYER_NORM, enable_hlfb=False
+  )
 
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
@@ -129,7 +131,6 @@ def get_model_config() -> cfg.ModelConfig:
       embedding_dim=embedding_dim,
       block_configs=block_config,
       final_norm_config=norm_config,
-      enable_hlfb=True,
   )
 
   return config
@@ -164,7 +165,9 @@ def get_fake_model_config() -> cfg.ModelConfig:
       use_bias=True,
   )
 
-  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.LAYER_NORM)
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.LAYER_NORM, enable_hlfb=False
+  )
 
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
@@ -180,7 +183,6 @@ def get_fake_model_config() -> cfg.ModelConfig:
       embedding_dim=embedding_dim,
       block_configs=block_config,
       final_norm_config=norm_config,
-      enable_hlfb=True,
   )
 
   return config

ai_edge_torch/generative/examples/t5/t5.py

@@ -393,8 +393,7 @@ def get_model_config_t5() -> cfg.ModelConfig:
   )
   # T5 Confirmed as RMS Norm and eps = 1e-6 TJA.
   norm_config = cfg.NormalizationConfig(
-      type=cfg.NormalizationType.RMS_NORM,
-      epsilon=1e-6,
+      type=cfg.NormalizationType.RMS_NORM, epsilon=1e-6, enable_hlfb=False
   )
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
@@ -411,7 +410,6 @@ def get_model_config_t5() -> cfg.ModelConfig:
       block_configs=block_config,
       final_norm_config=norm_config,
       lm_head_use_bias=False,
-      enable_hlfb=True,
   )
   return config
 

ai_edge_torch/generative/examples/test_models/toy_model.py

@@ -138,7 +138,9 @@ def get_model_config() -> cfg.ModelConfig:
       activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
       intermediate_size=256,
   )
-  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM, enable_hlfb=False
+  )
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
       ff_config=ff_config,
@@ -152,5 +154,6 @@ def get_model_config() -> cfg.ModelConfig:
       embedding_dim=128,
       block_configs=block_config,
       final_norm_config=norm_config,
+      enable_hlfb=False,
   )
   return config

ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py

@@ -108,7 +108,9 @@ def get_model_config() -> cfg.ModelConfig:
       activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
       intermediate_size=256,
   )
-  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM, enable_hlfb=False
+  )
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
       ff_config=ff_config,
@@ -122,7 +124,6 @@ def get_model_config() -> cfg.ModelConfig:
       embedding_dim=128,
       block_configs=block_config,
       final_norm_config=norm_config,
-      enable_hlfb=True,
   )
   return config
 

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -51,9 +51,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
       intermediate_size=5632,
   )
-  norm_config = cfg.NormalizationConfig(
-      type=cfg.NormalizationType.RMS_NORM, enable_hlfb=True
-  )
+  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
   block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
       ff_config=ff_config,
@@ -69,7 +67,6 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       block_configs=block_config,
       final_norm_config=norm_config,
       lm_head_share_weight_with_embedding=False,
-      enable_hlfb=True,
   )
   return config
 

ai_edge_torch/generative/examples/tiny_llama/verify.py

@@ -15,15 +15,10 @@
 
 """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
+from ai_edge_torch.generative.examples.tiny_llama import verify_util
 
 
 _PROMPTS = flags.DEFINE_multi_string(
@@ -39,32 +34,10 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
 
 
 def main(_):
-  checkpoint = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
-  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
-  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
-  reauthored_model = tiny_llama.build_model(str(reauthored_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,
+  verify_util.verify_tiny_llama(
+      checkpoint_dir="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
       max_new_tokens=_MAX_NEW_TOKENS.value,
-      atol=1e-04,
+      prompts=_PROMPTS.value,
   )