ai-edge-torch-nightly 0.3.0.dev20240926__py3-none-any.whl → 0.3.0.dev20240928__py3-none-any.whl

ai_edge_torch/generative/utilities/verifier.py

@@ -16,111 +16,129 @@
 """Common utility functions to verify the reauthored models."""
 
 import logging
-from typing import List, Optional, Union
+from typing import List
 
 from ai_edge_torch.generative.layers import kv_cache as kv_utils
 import torch
-import transformers
 
 
 class ModelWrapper(torch.nn.Module):
-  """A wrapper for the model to be verified, this could be a HuggingFace model
+  """A wrapper for the model to be verified.
 
-  or a regular PyTorch model.
+  It unifies the interface of forward() and generate() of models for the
+  verification to call.
   """
 
-  def __init__(
-      self,
-      model: torch.nn.Module,
-      model_format: str = "huggingface",
-      hf_generation_config: Optional[transformers.GenerationConfig] = None,
-  ):
+  def __init__(self, model: torch.nn.Module):
     """Initializes the wrapper.
 
     Args:
-      model (torch.nn.Module): The original model. This could be a model built
-        from HuggingFace transformers, or a regular PyTorch model.
-      model_format (str): The format of the model. It should be either
-        "huggingface" or "pytorch".
-      hf_generation_config (transformers.GenerationConfig): The HuggingFace
-        generation config. This config will only be used if the underlying model
-        is built from HuggingFace transformers.
+      model (torch.nn.Module): The model which might have different interfaces
+        of forward() and generate(). It could be a model built from HuggingFace
+        transformers, a regular PyTorch model, or a model re-authored with
+        ai_edge_torch Generative API.
     """
     super().__init__()
     self.model = model
-    self.model_format = model_format
-    self.hf_generation_config = hf_generation_config
+
+  def forward(self, tokens: torch.Tensor) -> torch.Tensor:
+    """Gets output logits by forwarding the input tokens.
+
+    Args:
+      tokens (torch.Tensor): The input tokens to forward. Its dimension is
+        expected to be (batch_size=1, kv_cache_max_len).
+
+    Returns:
+      The output logits.
+    """
+    raise NotImplementedError("forward() is not implemented.")
 
   def generate(
-      self, inputs: torch.Tensor
-  ) -> Union[transformers.utils.ModelOutput, torch.LongTensor]:
-    if self.model_format == "huggingface":
-      return self.model.generate(
-          inputs=inputs, generation_config=self.hf_generation_config
-      )
-    else:
-      raise NotImplementedError(
-          "generate() is not implemented for model format: %s"
-          % self.model_format
-      )
+      self, prompts: torch.Tensor, max_new_tokens: int
+  ) -> torch.IntTensor:
+    """Returns the response token IDs to the given prompts tensor.
 
-  def forward(
-      self,
-      inputs: torch.Tensor,
-  ):
-    return self.model.forward(inputs)
+    The maximum number of tokens to generate might be set by subclasses.
 
+    Args:
+      prompts (torch.Tensor): The input token IDs to generate with. Its shape is
+        expected to be (batch_size=1, input_ids_len).
+      max_new_tokens (int): The maximum number of response token IDs to
+        generate.
+
+    Returns:
+      The tensor of response token IDs with shape of (batch_size=1,
+      response_ids_len).
+    """
+    raise NotImplementedError("generate() is not implemented.")
 
-def forward(
-    model: torch.nn.Module,
-    tokens: torch.Tensor,
-    kv_cache: kv_utils.KVCache,
-) -> tuple[torch.Tensor, kv_utils.KVCache]:
-  """Forwards the model reauthored with ai_edge_torch Generative API.
 
-  Args:
-    model (torch.nn.Module): The model to forward. It should be a model built
-      with ai_edge_torch Generative API.
-    tokens (torch.Tensor): The input tokens to forward.
-    kv_cache (KVCache): The KV cache to forward.
+class ReauthoredModelWrapper(ModelWrapper):
+  """A wrapper for the model reauthored with ai_edge_torch Generative API."""
 
-  Returns:
-    The output logits and the updated KV cache.
-  """
-  input_pos = torch.arange(0, tokens.shape[1], dtype=torch.int)
-  output = model.forward(tokens, input_pos, kv_cache)
-  return output["logits"], output["kv_cache"]
+  def _init_kv_cache(self):
+    """Returns an initialized KV cache."""
+    return kv_utils.KVCache.from_model_config(self.model.config)
 
+  def _forward_with_kv_cache(
+      self,
+      tokens: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> tuple[torch.Tensor, kv_utils.KVCache]:
+    """Forwards the model and updates an external KV cache.
 
-def generate(
-    model: torch.nn.Module, prompts: torch.Tensor, response_len: int
-) -> torch.Tensor:
-  """Generates the response to the prompts.
+    Args:
+      tokens (torch.Tensor): The input tokens to forward.
+      kv_cache (KVCache): The KV cache to forward.
 
-  It appends tokens output by the model to the prompts and feeds them back to
-  the model up to decode_len.
+    Returns:
+      The output logits and the updated KV cache.
+    """
+    input_pos = torch.arange(0, tokens.shape[1], dtype=torch.int)
+    output = self.model.forward(tokens, input_pos, kv_cache)
+    return output["logits"], output["kv_cache"]
 
-  Args:
-    model (torch.nn.Module): The model to generate. It should be a model built
-      with ai_edge_torch Generative API.
-    prompts (torch.Tensor): The prompts to generate.
-    response_len (int): The number of tokens to generate.
+  def forward(self, tokens: torch.Tensor) -> torch.Tensor:
+    logits, _ = self._forward_with_kv_cache(tokens, self._init_kv_cache())
+    return logits
 
-  Returns:
-    The generated tokens.
-  """
-  input_ids = prompts[0].int().tolist()
-  kv_cache = kv_utils.KVCache.from_model_config(model.config)
-  for _ in range(response_len - len(input_ids)):
-    logits, kv_cache = forward(model, torch.tensor([input_ids]), kv_cache)
-    generated_token = logits[0][-1].argmax().item()
-    input_ids.append(generated_token)
-  return torch.tensor([input_ids])
+  def generate(
+      self, prompts: torch.Tensor, max_new_tokens: int
+  ) -> torch.IntTensor:
+    input_ids = prompts[0].int().tolist()
+    kv_cache = self._init_kv_cache()
+    for _ in range(max_new_tokens):
+      tokens = torch.tensor([input_ids])
+      logits, kv_cache = self._forward_with_kv_cache(tokens, kv_cache)
+      generated_token = logits[0][-1].argmax().item()
+      input_ids.append(generated_token)
+    return torch.tensor([input_ids])
+
+
+class TokenizerWrapper(torch.nn.Module):
+  """A wrapper for the tokenizer used for verification."""
+
+  def __init__(self, tokenizer: torch.nn.Module):
+    """Initializes the wrapper.
+
+    Args:
+      tokenizer (torch.nn.Module): The tokenizer to wrap.
+    """
+    super().__init__()
+    self.tokenizer = tokenizer
+
+  def encode(self, prompts: str) -> torch.Tensor:
+    """Encodes the prompts to token IDs."""
+    return self.tokenizer.encode(prompts, return_tensors="pt")
+
+  def decode(self, token_ids: torch.Tensor) -> str:
+    """Decodes the token IDs to a string."""
+    return self.tokenizer.decode(token_ids)
 
 
 def verify_with_input_ids(
     original_model: ModelWrapper,
-    reauthored_model: torch.nn.Module,
+    reauthored_model: ReauthoredModelWrapper,
     input_ids: List[int],
     kv_cache_max_len: int = 1024,
     rtol: float = 1e-05,
@@ -132,8 +150,8 @@ def verify_with_input_ids(
 
   Args:
     original_model (ModelWrapper): The original model.
-    reauthored_model (torch.nn.Module): The model reauthored with ai_edge_torch
-      Generative API.
+    reauthored_model (ReauthoredModelWrapper): The model reauthored with
+      ai_edge_torch Generative API.
     input_ids (List[int]): The input token IDs to forward with.
     kv_cache_max_len (int): The maximum sequence length of the KV cache.
     rtol (float): The relative tolerance for the comparison.
@@ -147,13 +165,12 @@ def verify_with_input_ids(
 
   logging.info("Forwarding the original model...")
   outputs_original = original_model.forward(tokens)
-  logits_original = outputs_original.logits[0, len(input_ids) - 1, :]
+  logits_original = outputs_original[0, len(input_ids) - 1, :]
   logging.info("logits_original: %s", logits_original)
 
   logging.info("Forwarding the reauthored model...")
-  kv_cache = kv_utils.KVCache.from_model_config(reauthored_model.config)
-  outputs_reauthored = forward(reauthored_model, tokens, kv_cache)
-  logits_reauthored = outputs_reauthored[0][0, len(input_ids) - 1, :]
+  outputs_reauthored = reauthored_model.forward(tokens)
+  logits_reauthored = outputs_reauthored[0, len(input_ids) - 1, :]
   logging.info("logits_reauthored: %s", logits_reauthored)
 
   return torch.allclose(
@@ -163,9 +180,10 @@ def verify_with_input_ids(
 
 def verify_model_with_prompts(
     original_model: ModelWrapper,
-    reauthored_model: torch.nn.Module,
-    tokenizer: torch.nn.Module,
+    reauthored_model: ReauthoredModelWrapper,
+    tokenizer: TokenizerWrapper,
     prompts: str,
+    max_new_tokens: int,
 ) -> bool:
   """Verifies if the model reauthored generates the same answer of the oringal.
 
@@ -174,24 +192,24 @@ def verify_model_with_prompts(
 
   Args:
     original_model (ModelWrapper): The original model.
-    reauthored_model (torch.nn.Module): The model reauthored with ai_edge_torch
-      Generative API.
-    tokenizer (torch.nn.Module): The tokenizer.
+    reauthored_model (ReauthoredModelWrapper): The model reauthored with
+      ai_edge_torch Generative API.
+    tokenizer (TokenizerWrapper): The tokenizer.
     prompts (str): The input prompts to generate answers.
+    max_new_tokens (int): The maximum number of new tokens to generate.
 
   Returns:
     True if the model reauthored generates the same answer of the original.
   """
-  prompt_tokens = tokenizer.encode(prompts, return_tensors="pt")
+  prompt_tokens = tokenizer.encode(prompts)
 
   logging.info("Generating answer with the original model...")
-  outputs_original = original_model.generate(prompt_tokens)
+  outputs_original = original_model.generate(prompt_tokens, max_new_tokens)
   response_original = tokenizer.decode(outputs_original[0])
   logging.info("outputs_from_original_model: [[%s]]", response_original)
 
   logging.info("Generating answer with the reauthored model...")
-  generate_len = len(outputs_original[0])
-  outputs_reauthored = generate(reauthored_model, prompt_tokens, generate_len)
+  outputs_reauthored = reauthored_model.generate(prompt_tokens, max_new_tokens)
   response_reauthored = tokenizer.decode(outputs_reauthored[0])
   logging.info("outputs from reauthored model: [[%s]]", response_reauthored)
 
@@ -200,9 +218,10 @@ def verify_model_with_prompts(
 
 def verify_reauthored_model(
     original_model: ModelWrapper,
-    reauthored_model: torch.nn.Module,
-    tokenizer: torch.nn.Module,
+    reauthored_model: ReauthoredModelWrapper,
+    tokenizer: TokenizerWrapper,
     generate_prompts: List[str],
+    max_new_tokens: int = 30,
     forward_input_ids: List[List[int]] = [[1, 2, 3, 4]],
     rtol: float = 1e-05,
     atol: float = 1e-05,
@@ -219,10 +238,11 @@ def verify_reauthored_model(
 
   Args:
     original_model (ModelWrapper): The original model.
-    reauthored_model (torch.nn.Module): The model reauthored with ai_edge_torch
-      Generative API.
-    tokenizer (torch.nn.Module): The tokenizer.
+    reauthored_model (ReauthoredModelWrapper): The model reauthored with
+      ai_edge_torch Generative API.
+    tokenizer (TokenizerWrapper): The tokenizer.
     generate_prompts (List[str]): List of the input prompts to generate answers.
+    max_new_tokens (int): The maximum number of new tokens to generate.
     forward_input_ids (List[torch.Tensor]): List if ihe input token IDs to
       forward with.
     rtol (float): The relative tolerance for the comparison.
@@ -235,13 +255,13 @@ def verify_reauthored_model(
     ):
       logging.info("PASS")
     else:
-      logging.info("FAILED")
+      logging.error("FAILED")
 
   for prompts in generate_prompts:
     logging.info("Verifying the reauthored model with prompts:%s", prompts)
     if verify_model_with_prompts(
-        original_model, reauthored_model, tokenizer, prompts
+        original_model, reauthored_model, tokenizer, prompts, max_new_tokens
     ):
       logging.info("PASS")
     else:
-      logging.info("FAILED")
+      logging.error("FAILED")

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.3.0.dev20240926"
+__version__ = "0.3.0.dev20240928"

{ai_edge_torch_nightly-0.3.0.dev20240926.dist-info → ai_edge_torch_nightly-0.3.0.dev20240928.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20240926
+Version: 0.3.0.dev20240928
 Summary: Supporting PyTorch models with the Google AI Edge TFLite runtime.
 Home-page: https://github.com/google-ai-edge/ai-edge-torch
 Keywords: On-Device ML,AI,Google,TFLite,PyTorch,LLMs,GenAI

{ai_edge_torch_nightly-0.3.0.dev20240926.dist-info → ai_edge_torch_nightly-0.3.0.dev20240928.dist-info}/RECORD

@@ -3,7 +3,7 @@ ai_edge_torch/config.py,sha256=FMWeCH2b7HYILBvaI1iZNnYCO4WAhDOwBZBmIE-xrF0,909
 ai_edge_torch/conftest.py,sha256=r0GTrhMRhlmOGrrkvumHN8hkmyug6WvF60vWq8wRIBI,758
 ai_edge_torch/fx_pass_base.py,sha256=D86Gw3pIRcpnTebUPKlnPbPGJae1S6Fw4DZZ3ZkD0zw,3730
 ai_edge_torch/model.py,sha256=N-pNpTxzhaFGhWhnSGd70lBzb9VlEhTOq5mddU7bvvI,5542
-ai_edge_torch/version.py,sha256=e_e6TIee2wiwaLShw_LBUsVRwNFQHZYOv7WtAcrMix4,706
+ai_edge_torch/version.py,sha256=YiCjdglLzSPYyRq64U8zJSgWDFqJs-t2JSzuA0bYYzA,706
 ai_edge_torch/_convert/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/_convert/conversion.py,sha256=5uPwHhmc6kwiIz-CqaiHDejf2SOWMHrb-rYEHm69wKc,3801
 ai_edge_torch/_convert/conversion_utils.py,sha256=Sr8qXVcTwc-ZnZmK7yxVrIOOp1S_vNrwzC0zUvLTI2o,2160
@@ -41,32 +41,38 @@ ai_edge_torch/generative/examples/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQe
 ai_edge_torch/generative/examples/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/gemma/convert_gemma1_to_tflite.py,sha256=evmUj_4yygQthSRU-ke-Xn1qFNDCZKbegqINWfruKwU,2184
 ai_edge_torch/generative/examples/gemma/convert_gemma2_to_tflite.py,sha256=RZDs6oY-NLYrPNtfuJDweIHzGUL2kzpIc3AW_1p8gGg,2186
-ai_edge_torch/generative/examples/gemma/gemma1.py,sha256=cahMzvJNJfShIw4uqoBRX5iBZrI3rvsha6wpNHzeYJ0,6369
-ai_edge_torch/generative/examples/gemma/gemma2.py,sha256=UziyJVrR_QXE_vFAagjnn1KluMM74coI89-UcdGTpkQ,9243
+ai_edge_torch/generative/examples/gemma/gemma1.py,sha256=kxWmmoVvtLP5auB3UXA2vsvZmSnpBs4SBixzYeAXzVA,6255
+ai_edge_torch/generative/examples/gemma/gemma2.py,sha256=7VF5RYJ8QhROQNIlx-QovO-y6-jFp_EHgAkBNChZaqE,9066
 ai_edge_torch/generative/examples/gemma/verify_gemma1.py,sha256=ip-Gmk4CI5f0GWSdAIdrectxQWJ0t328KCsA4nfHuGg,1736
-ai_edge_torch/generative/examples/gemma/verify_gemma2.py,sha256=p0wmhUDkNj6m9QZVssHjSc2HzIxci_1XMvZqS1X3xK4,1818
-ai_edge_torch/generative/examples/gemma/verify_util.py,sha256=vQp_QN8nRXv81-wrI8h8722jV68VEUMGSrpXa5kRk68,5175
+ai_edge_torch/generative/examples/gemma/verify_gemma2.py,sha256=K77k-JpdhIwm3tbBnzpw8HQsFRwAVyszxRo82fR6-q4,1762
+ai_edge_torch/generative/examples/gemma/verify_util.py,sha256=sqltZbnyKemNvKqqi9d09i74gP-PPQFodRYfDfnhycQ,4933
+ai_edge_torch/generative/examples/llama/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/llama/convert_3b_to_tflite.py,sha256=_OrerrTA6tvP9Tnwj601QO95Cm8PlOiYP-mxvtmBmb4,2186
+ai_edge_torch/generative/examples/llama/convert_to_tflite.py,sha256=GGo6Kxiwqva4JfurGx3InU3nROW70XtYvxUwEf_6mBQ,2180
+ai_edge_torch/generative/examples/llama/llama.py,sha256=5vlh2Z8vEPH8Z4LoHoFYCcuOQynx4mbVE37v3yMl1hE,7162
+ai_edge_torch/generative/examples/llama/verify.py,sha256=7xwKM_yzLCrmFsYj1UbsjW58ZG8Yic0xw1GFkdydrCU,2525
+ai_edge_torch/generative/examples/llama/verify_3b.py,sha256=IijBWqLXINOfwayM-8EIpc7OcC6Nj5CnberStx-vDSk,2528
 ai_edge_torch/generative/examples/openelm/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/openelm/convert_to_tflite.py,sha256=85FVEt6cKFP2UzCLC78tAkbwGlGhAArtG7Wa75NxJik,2185
-ai_edge_torch/generative/examples/openelm/openelm.py,sha256=VcU8A0B9nQR-FTPHXqNHSHZzeIZZ_As4yvKZMnoU2P4,7482
-ai_edge_torch/generative/examples/openelm/verify.py,sha256=_tglrzub_qAVDHGacriyzPFlKeIFhZ4KnDGkZ22g7So,2127
+ai_edge_torch/generative/examples/openelm/openelm.py,sha256=hxbpvk0fNswzbqZfGteflqKMmkH7yzeMuW6r29s_xnQ,7374
+ai_edge_torch/generative/examples/openelm/verify.py,sha256=VkigoqhAr8ew95neb3TifYv-SLOSheaWKv2AH0iKDrc,2441
 ai_edge_torch/generative/examples/phi/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/phi/convert_phi3_to_tflite.py,sha256=rkbTtMaqSVG48cm-NTxR_LDgZmXAEBqayTm9O49oMXc,2171
 ai_edge_torch/generative/examples/phi/convert_to_tflite.py,sha256=3go690yX6PFeXMdpY7y4JZorAwxX0HT_b_pKZieauvk,2169
-ai_edge_torch/generative/examples/phi/phi2.py,sha256=YwAszA53aOjvaMJ5wua2-5rP79N21Un_Y5yBCfFSYNU,6189
-ai_edge_torch/generative/examples/phi/phi3.py,sha256=DIDzpG8DZkWDcWsAVkcxzxIC3U3352uVI3zMoYZD16U,9554
-ai_edge_torch/generative/examples/phi/verify.py,sha256=HJ7RkE0CTtpTdu_pgW4VX37_4Q20Ow6nG1DtA_eXsZE,2161
-ai_edge_torch/generative/examples/phi/verify_phi3.py,sha256=7ygNNafJBu803JhxQm0aXMccCy6YX6-D-DT9k4QCyL0,2344
+ai_edge_torch/generative/examples/phi/phi2.py,sha256=82SEKRwtKfT9VcNQaykGmemiov_XaXWLi4Zyw9Vtmj0,6075
+ai_edge_torch/generative/examples/phi/phi3.py,sha256=Xh-l7TQdXYZJ9PViRVk2_y91Ec7Yntn0UpkuzRIG3T8,9231
+ai_edge_torch/generative/examples/phi/verify.py,sha256=YPFCdbnfmvq38fbpBNr0kHPfSZo4p3_6WkLJAW3pLPo,2177
+ai_edge_torch/generative/examples/phi/verify_phi3.py,sha256=kVYaBVvddfQng0IyZGxyTJEzhiPO0G4VFJm2WOc2Q94,2360
 ai_edge_torch/generative/examples/smollm/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/smollm/convert_to_tflite.py,sha256=zPrDTDeRVWFi9DS32uNi-RLpzOStFOk5MhNla4ixeew,2179
-ai_edge_torch/generative/examples/smollm/smollm.py,sha256=hyhMk-b5762Q2xmjdD47g85dcbBSNJXNPIsifm1DRto,3239
-ai_edge_torch/generative/examples/smollm/verify.py,sha256=FhnNiISi8JZWQaA-AaL9giS7IzkBAvcHkuDFBISj4Zg,2038
+ai_edge_torch/generative/examples/smollm/smollm.py,sha256=dal8vnZjQd6vR7sc76-FYGDKUlVjOlfUALV-pwbXJGc,3264
+ai_edge_torch/generative/examples/smollm/verify.py,sha256=HXYcCjDJMylVL3Pc9HU-UXqtpjtIU25o1YhPiX30aPU,2361
 ai_edge_torch/generative/examples/stable_diffusion/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/stable_diffusion/attention.py,sha256=kDWG6MlIGa89zC5KSRcJlw2c4ITuw8KcchtfmF55f4g,3545
-ai_edge_torch/generative/examples/stable_diffusion/clip.py,sha256=2RMi5UmfMT4Ep68ZLJsqF-fMvEumNVkIwqtsRli9HhA,6068
+ai_edge_torch/generative/examples/stable_diffusion/clip.py,sha256=lwWrKY1NpnbvHQRenpltVN65QlzjWmSScl5CLSipBkc,6110
 ai_edge_torch/generative/examples/stable_diffusion/convert_to_tflite.py,sha256=vfMGI03UL_gfB561t2kzIHuScwnsUmqaPWxgvq_1T5A,5043
-ai_edge_torch/generative/examples/stable_diffusion/decoder.py,sha256=ZTRD56e8MsdGPJr7vpLa4Ju_BFw_b-FUgXgd-SO5MBw,15665
-ai_edge_torch/generative/examples/stable_diffusion/diffusion.py,sha256=6FAnevL8ZfCK2YCSPivarUH0Z8wGKSmnPpJNC0OI5A8,33680
+ai_edge_torch/generative/examples/stable_diffusion/decoder.py,sha256=ClXNntmh0PF3s6U3C7SW3tyVrsSSrV2kyz-_RF4BcqA,15715
+ai_edge_torch/generative/examples/stable_diffusion/diffusion.py,sha256=mBEAUYjV1qDJy9ZAsHtm9RGce0Mbzv0VoPZpdcQl1mk,33730
 ai_edge_torch/generative/examples/stable_diffusion/encoder.py,sha256=CAPsW84A8f00nS6fLFeh_XUjCPsDCA5UxHOUsMrLfSU,3450
 ai_edge_torch/generative/examples/stable_diffusion/pipeline.py,sha256=x9lEEENGNbpx6VTf_LTVudd9d6bs9tLvFUKTl252zEY,8623
 ai_edge_torch/generative/examples/stable_diffusion/tokenizer.py,sha256=xychak9hdLd6ieXBYEwrK2BkF8NRZWZSSCijIsESpBA,3420
@@ -78,16 +84,16 @@ ai_edge_torch/generative/examples/stable_diffusion/samplers/k_lms.py,sha256=ZE6H
 ai_edge_torch/generative/examples/stable_diffusion/samplers/sampler.py,sha256=RxR5rw0wFFm_5CfAY-3-EIz83vhM9EKye8Bb5zBb0Ok,1341
 ai_edge_torch/generative/examples/t5/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/t5/convert_to_tflite.py,sha256=HHtZTtUh3QgE4F74-ru_8n1pt6cqfbObw12xoaMJ7NQ,4596
-ai_edge_torch/generative/examples/t5/t5.py,sha256=OZ67knK-UB1dBjxydG-Jwkp0Z3FzOCqGPTdg5aBFu4w,21328
+ai_edge_torch/generative/examples/t5/t5.py,sha256=gFTmPi-xB8pcPRgoF3DJxvH_fT-KWTb8ii77P5UbKR0,21263
 ai_edge_torch/generative/examples/t5/t5_attention.py,sha256=l01oYyJo77INzRwN4xqXquaFQPvCFBFF5zOnmGVb3Hg,8731
 ai_edge_torch/generative/examples/test_models/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/test_models/convert_toy_model.py,sha256=6-WaNHckq_LlXMVTh8x90MGWeWq2bu_T_XQd3w9FnGg,3261
-ai_edge_torch/generative/examples/test_models/toy_model.py,sha256=LTuzres5DHmrMT6U9rCrGf6vmR9SmopmB8sO6Cd2NxQ,5255
-ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=xDYTh4m3vBEb6r3_ERhmj5qILW7YdVDAnZ-fitgYONg,4450
+ai_edge_torch/generative/examples/test_models/toy_model.py,sha256=4113jZK-Hu3kYop__WTc8Bq-bG6YzQtADbxHtYPEB4w,5036
+ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=ZpjSIiayjTEVwg5Q1vI9Iy5tq1YSF5zaVDF4HTp_Z2s,4353
 ai_edge_torch/generative/examples/tiny_llama/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/tiny_llama/convert_to_tflite.py,sha256=ekxd8efjMgEvauUu3PidWOC-DszPHn5sqU753F7sJIM,2201
-ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=tlWpa7Aun3u3w5b-9EBtW7olhmSf8W-tn5bKUIwC-ys,6044
-ai_edge_torch/generative/examples/tiny_llama/verify.py,sha256=Yt49ndXTL_fUfjrQjx-AyhD4x796IOTjuBegiQe7_Yc,2111
+ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=aSNHOAar5yPnGAeKsv8zrqYhOq9RR_7hwqHUMBb2mkM,5930
+ai_edge_torch/generative/examples/tiny_llama/verify.py,sha256=7Bk8z033M-BCXJ299fpQNXYAudBbZoDQp9934xcvg50,2426
 ai_edge_torch/generative/fx_passes/__init__.py,sha256=jrzCB3ZyY_t5jJM1e2Czdt3DjAIL43R0_a-T-I7wOzw,1155
 ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py,sha256=hhxSQvkDMv0isZJhmuLiod66ZODaJ8uSPSVTJVHBabQ,1931
 ai_edge_torch/generative/layers/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
@@ -96,7 +102,7 @@ ai_edge_torch/generative/layers/attention_utils.py,sha256=68GXGR2HSWBFViTxX7cHif
 ai_edge_torch/generative/layers/builder.py,sha256=oE8DdqLA-oWkBC2zySSCh8JNAJg_hk8-W_UoMSrgDVk,5088
 ai_edge_torch/generative/layers/feed_forward.py,sha256=hdICat-8gW7-vxDAevJQ8NQ-mynllPiqLdXQMF6JMnc,4189
 ai_edge_torch/generative/layers/kv_cache.py,sha256=2El7kZYnQRCRcVc63xgiAdBh9oVOksDu35p9XggvaGE,6148
-ai_edge_torch/generative/layers/model_config.py,sha256=l5Rb3h3GK2pux-Lg3BONTD6b7klxXqUbDDtYs_bGKLk,6879
+ai_edge_torch/generative/layers/model_config.py,sha256=Fa0eFCMlyfdwd3cM1drhP9vlXRhIguDrglsHn4ax2_w,6948
 ai_edge_torch/generative/layers/normalization.py,sha256=cpo88JUXbF9j3sJTU4JuwOap9ryGV05C1QkPij-YQwU,6999
 ai_edge_torch/generative/layers/rotary_position_embedding.py,sha256=CZqOoibLcHvUgrgaIIWAlmk3XgE2inzx340MN-npLoU,1347
 ai_edge_torch/generative/layers/scaled_dot_product_attention.py,sha256=gXxh3papKy4FBpGEX7VyZ7rZ1Js6aHK70Q6DKrVSckY,4154
@@ -115,7 +121,7 @@ ai_edge_torch/generative/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudj
 ai_edge_torch/generative/test/test_kv_cache.py,sha256=W6Bh0gYDzmwb0j9HdD5_D7Z7FPToP2HSyFrmwIXuFqo,3793
 ai_edge_torch/generative/test/test_loader.py,sha256=8y74ChO3CZCfEi1eCf3-w47kRgAI4qPYCXpi8rTQXMA,3378
 ai_edge_torch/generative/test/test_model_conversion.py,sha256=s-EVLOQGjIeVtgNI8Ggs37pkRdErAliT6NhrrFigPOE,5459
-ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=IzW2HjXS2-zePZM-qEuXL4zclnGvYsNw-6tuDSeNna4,8163
+ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=kCm-L3rWbPj25E_QEbkSLiaCk3y23SjKJs-MG-EwKug,8545
 ai_edge_torch/generative/test/test_quantize.py,sha256=8geJhKwYBU20m0mdGPD1BUFwQ0lZKNtCB04SOLO18y4,5980
 ai_edge_torch/generative/test/utils.py,sha256=YvEhO2HIj1LkBs5du1UxY-cGRW9HMyAYsOUhgsTrTpA,1796
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
@@ -123,7 +129,8 @@ ai_edge_torch/generative/utilities/converter.py,sha256=MQUg2ZLmfk_2csWmQWKD_II0b
 ai_edge_torch/generative/utilities/loader.py,sha256=b9iotIhVDX-Zc9XjIDUaLxnV395AyBnkQe3dV5YA7Co,13297
 ai_edge_torch/generative/utilities/stable_diffusion_loader.py,sha256=dqPD9qRXEWtU3ombslOC-BE2l_dMwHoCNu7NsIJhsso,36158
 ai_edge_torch/generative/utilities/t5_loader.py,sha256=tEsfy8-ymzbbjOIc-oesXF3yGyyWtJgFXn2s7VOavt8,16961
-ai_edge_torch/generative/utilities/verifier.py,sha256=vo5haJqtae-M_s4qLMh1_ItU0zluwqekV8jQ-cEjXCQ,8807
+ai_edge_torch/generative/utilities/transformers_verifier.py,sha256=8sp9m_FMcXn7nqOrochtu2jIANkJKhnhIBUmH0ZTDR4,1549
+ai_edge_torch/generative/utilities/verifier.py,sha256=wQ4EtIED_a6FRsaOXeoQVZiHNx07esOYCQYbDVLgZ2o,9520
 ai_edge_torch/hlfb/__init__.py,sha256=sH4um75na-O8tzxN6chFyp6Y4xnexsE7kUQpZySv6dE,735
 ai_edge_torch/hlfb/mark_pattern/__init__.py,sha256=cjTprggj_cuktSCm7-A25e7Shop3k63ylp7sdZmtZ8o,4790
 ai_edge_torch/hlfb/mark_pattern/passes.py,sha256=pjkKcI1nHECPluAt87cFBrt1DP0f3ge7rHq1NhCkBIE,1936
@@ -170,8 +177,8 @@ ai_edge_torch/quantize/quant_config.py,sha256=U0KisSW-uZkoMJcy-ZP9W57p3tsa594fr9
 ai_edge_torch/testing/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/testing/model_coverage/__init__.py,sha256=5P8J6Zk5YYtDvTBucFvB9NGSRI7Gw_24WnrbhXgycEE,765
 ai_edge_torch/testing/model_coverage/model_coverage.py,sha256=UPB448aMDUyC0HNYVqio2rcJPnDN0tBQMP08J6vPYew,4718
-ai_edge_torch_nightly-0.3.0.dev20240926.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.3.0.dev20240926.dist-info/METADATA,sha256=r2brEhCKGcZsIaGpoJLT3DXvY9Yi-55_A-ZwjVtPTnk,1897
-ai_edge_torch_nightly-0.3.0.dev20240926.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-ai_edge_torch_nightly-0.3.0.dev20240926.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.3.0.dev20240926.dist-info/RECORD,,
+ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/METADATA,sha256=3HuAFZTfvmU787dVypwpmUvo4DdZSekGsqGimO-oPfM,1897
+ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/RECORD,,