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/llama/verify_3b.py

@@ -0,0 +1,73 @@
+# 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 Llama 3.2-3B model."""
+
+import logging
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.llama import 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",
+    "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 = "meta-llama/Llama-3.2-3B-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 = llama.build_3b_model(reauthored_checkpoint)
+
+  logging.info("Loading the tokenizer from: %s", checkpoint)
+  # Llama tokenizer_config.json sets a fast tokenizer class explicitly,
+  # "PreTrainedTokenizerFast". It works only when the fast tokenizer is
+  # available.
+  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-04,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/examples/openelm/openelm.py

@@ -68,15 +68,10 @@ class OpenELM(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
 
@@ -154,6 +149,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
             num_heads=num_heads[idx],
             head_dim=128,
             num_query_groups=num_query_groups[idx],
+            rotary_base=10000,
             rotary_percentage=1.0,
             qkv_transpose_before_split=True,
             query_norm_config=norm_config,

ai_edge_torch/generative/examples/openelm/verify.py

@@ -15,28 +15,33 @@
 
 """Verifies the reauthored OpenELM-3B model."""
 
+import logging
 import pathlib
-
 from absl import app
 from absl import flags
 from ai_edge_torch.generative.examples.openelm import openelm
+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 = "apple/OpenELM-3B"
-  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.
@@ -44,18 +49,21 @@ def main(_):
       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 = openelm.build_model(reauthored_checkpoint)
 
   tokenizer_checkpoint = "meta-llama/Llama-2-7b-hf"
-  verifier.log_msg("Loading the tokenizer from", tokenizer_checkpoint)
+  logging.info("Loading the tokenizer from: %s", tokenizer_checkpoint)
   tokenizer = transformers.AutoTokenizer.from_pretrained(tokenizer_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,
   )
 
 

ai_edge_torch/generative/examples/phi/convert_phi3_to_tflite.py

@@ -0,0 +1,68 @@
+# 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
+
+_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.',
+)
+_TFLITE_PATH = flags.DEFINE_string(
+    'tflite_path',
+    '/tmp/',
+    'The tflite file path to export.',
+)
+_PREFILL_SEQ_LEN = flags.DEFINE_integer(
+    'prefill_seq_len',
+    1024,
+    'The maximum size of prefill input tensor.',
+)
+_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.',
+)
+
+
+def main(_):
+  pytorch_model = phi3.build_model(
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  quant_suffix = 'q8' if _QUANTIZE.value else 'f32'
+  output_filename = f'phi3_{quant_suffix}_seq{_PREFILL_SEQ_LEN.value}_ekv{_KV_CACHE_MAX_LEN.value}.tflite'
+  converter.convert_to_tflite(
+      pytorch_model,
+      tflite_path=os.path.join(_TFLITE_PATH.value, output_filename),
+      prefill_seq_len=_PREFILL_SEQ_LEN.value,
+      quantize=_QUANTIZE.value,
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/phi/phi2.py

@@ -65,15 +65,10 @@ class Phi2(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
 
@@ -129,6 +124,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       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,

ai_edge_torch/generative/examples/phi/phi3.py

@@ -0,0 +1,279 @@
+# 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."""
+
+import math
+from typing import Tuple
+
+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 ai_edge_torch.generative.utilities.loader as loading_utils
+import torch
+from torch import nn
+
+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_rope_cache(
+    size: int,
+    dim: int,
+    base: int,
+    condense_ratio: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    theta_factors: torch.Tensor,
+    scale: float,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+  """Precomputes 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:
+      size (int): The size of the built cache.
+      dim (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 (dim,) 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, dim, 2).float() / dim))
+  theta = theta / theta_factors
+  seq_idx = torch.arange(size) / 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(nn.Module):
+  """A Phi-3.5 model built from the Edge Generative API layers."""
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__()
+
+    # Construct model layers.
+    self.lm_head = nn.Linear(
+        config.embedding_dim, config.vocab_size, bias=config.lm_head_use_bias
+    )
+    self.tok_embedding = nn.Embedding(
+        config.vocab_size, config.embedding_dim, padding_idx=0
+    )
+    # Phi-3.5 has only one block config.
+    block_config = config.block_config(0)
+    self.transformer_blocks = nn.ModuleList(
+        attention.TransformerBlock(block_config, config)
+        for _ in range(config.num_layers)
+    )
+    self.final_norm = builder.build_norm(
+        config.embedding_dim,
+        config.final_norm_config,
+    )
+    attn_config = block_config.attn_config
+    self.rope_cache = _build_rope_cache(
+        size=config.kv_cache_max,
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
+        base=attn_config.rotary_base,
+        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(config.max_seq_len)
+        ),
+    )
+    self.mask_cache = attn_utils.build_causal_mask_cache(
+        size=config.kv_cache_max,
+    )
+    self.config = config
+
+  @torch.inference_mode
+  def forward(
+      self,
+      tokens: torch.Tensor,
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> dict[torch.Tensor, kv_utils.KVCache]:
+    _, seq_len = tokens.size()
+    assert self.config.max_seq_len >= seq_len, (
+        f"Cannot forward sequence of length {seq_len}, max seq length is only"
+        f" {self.config.max_seq_len}"
+    )
+    assert len(self.transformer_blocks) == len(kv_cache.caches), (
+        "The number of transformer blocks and the number of KV cache entries"
+        " must be the same."
+    )
+
+    cos, sin = self.rope_cache
+    cos = cos.index_select(0, input_pos)
+    sin = sin.index_select(0, input_pos)
+    mask = self.mask_cache.index_select(2, input_pos)
+    mask = mask[:, :, :, : self.config.kv_cache_max]
+
+    x = self.tok_embedding(tokens)
+
+    updated_kv_entires = []
+    for i, block in enumerate(self.transformer_blocks):
+      kv_entry = kv_cache.caches[i] if kv_cache else None
+      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
+      if kv_entry:
+        updated_kv_entires.append(kv_entry)
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
+
+    x = self.final_norm(x)
+    logits = self.lm_head(x)  # (b, t, vocab_size)
+    return {"logits": logits, "kv_cache": updated_kv_cache}
+
+
+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,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=32064,
+      num_layers=32,
+      max_seq_len=4096,
+      kv_cache_max_len=kv_cache_max_len,
+      embedding_dim=3072,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      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-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) -> nn.Module:
+  """Instantiates the model instance and load checkpoint if provided."""
+  config = get_model_config(**kwargs)
+  model = Phi3_5Mini(config)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  loader.load(model)
+  model.eval()
+  return model

ai_edge_torch/generative/examples/phi/verify.py

@@ -14,20 +14,22 @@
 # ==============================================================================
 
 """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,
@@ -37,25 +39,23 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
 
 def main(_):
   checkpoint = kagglehub.model_download("Microsoft/phi/transformers/2")
-  verifier.log_msg("Loading the original model from", checkpoint)
-  generation_config = transformers.GenerationConfig.from_pretrained(checkpoint)
-  generation_config.max_new_tokens = _MAX_NEW_TOKENS.value
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(checkpoint),
-      hf_generation_config=generation_config,
-  )
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
 
-  verifier.log_msg("Building the reauthored model from", checkpoint)
+  logging.info("Building the reauthored model from: %s", checkpoint)
   reauthored_model = phi2.build_model(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-03,
   )
 

ai_edge_torch/generative/examples/phi/verify_phi3.py

@@ -0,0 +1,69 @@
+# 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-3.5 model."""
+
+import logging
+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 transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+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 = "microsoft/Phi-3.5-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 = phi3.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,
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/examples/smollm/smollm.py

@@ -54,6 +54,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       num_heads=9,
       head_dim=64,
       num_query_groups=3,
+      rotary_base=10000,
       rotary_percentage=1.0,
   )
   ff_config = cfg.FeedForwardConfig(