ai-edge-torch-nightly 0.3.0.dev20240923__py3-none-any.whl → 0.3.0.dev20240925__py3-none-any.whl

ai_edge_torch/generative/examples/openelm/openelm.py

@@ -161,9 +161,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
         ),
         ff_config=cfg.FeedForwardConfig(
             type=cfg.FeedForwardType.SEQUENTIAL,
-            activation=cfg.ActivationConfig(
-                cfg.ActivationType.SILU_GLU, gate_is_front=True
-            ),
+            activation=cfg.ActivationConfig(cfg.ActivationType.SILU_GLU),
             intermediate_size=get_intermediate_size(idx),
             pre_ff_norm_config=norm_config,
         ),

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/phi3.py

@@ -0,0 +1,286 @@
+# 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 = 10000,
+    condense_ratio: int = 1,
+    dtype: torch.dtype = torch.float32,
+    device: torch.device = None,
+    theta_factors: torch.Tensor = None,
+    scale: float = 1.0,
+) -> 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. Defaults to 10000.
+      condense_ratio (int, optional): The ratio by which sequence indicies are
+        condensed. Defaults to 1.
+      dtype (torch.dtype, optional): Output tensor's data type. Defaults to
+        torch.float32.
+      device (torch.device, optional): Output tensor's data type. Defaults to
+        None in which case "cpu" is used.
+      theta_factors (torch.Tensor, optional): A tensor of shape (dim,) used to
+        scale the theta values. Defaults to None.
+      scale (float, optional): A float used to scale the rope values. Defaults
+        to 1.0.
+
+  Returns:
+      Tuple[torch.Tensor, torch.Tensor]: Rope's Cosine and Sine waves.
+  """
+  if device is None:
+    device = torch.device('cpu')
+  theta = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+  if theta_factors is not None:
+    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=10_000,
+        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,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+    )
+    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_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

@@ -27,13 +27,13 @@ _PROMPTS = flags.DEFINE_multi_string(
     "Instruct: Write an email about the weather Output:",
     "The input prompts to generate answers.",
 )
-
 _MAX_NEW_TOKENS = flags.DEFINE_integer(
     "max_new_tokens",
     30,
     "The maximum size of the generated tokens.",
 )
 
+
 def main(_):
   checkpoint = kagglehub.model_download("Microsoft/phi/transformers/2")
   verifier.log_msg("Loading the original model from", checkpoint)

ai_edge_torch/generative/examples/phi/verify_phi3.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.
+# ==============================================================================
+
+"""Verifies the reauthored Phi-3.5 model."""
+
+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 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"
+  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,
+  )
+
+  # 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)
+  reauthored_model = phi3.build_model(reauthored_checkpoint)
+
+  verifier.log_msg("Loading the tokenizer from", checkpoint)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
+
+  verifier.verify_reauthored_model(
+      original_model=wrapper_model,
+      reauthored_model=reauthored_model,
+      tokenizer=tokenizer,
+      generate_prompts=_PROMPTS.value,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)

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
@@ -132,3 +133,53 @@ def get_model_config() -> cfg.ModelConfig:
   )
 
   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_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

ai_edge_torch/generative/examples/stable_diffusion/decoder.py

@@ -324,3 +324,59 @@ def get_model_config() -> unet_cfg.AutoEncoderConfig:
       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_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

ai_edge_torch/generative/examples/stable_diffusion/diffusion.py

@@ -603,7 +603,7 @@ 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 +645,71 @@ 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/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
@@ -149,31 +148,3 @@ 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]
 
@@ -87,11 +86,6 @@ 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,
@@ -133,51 +127,3 @@ 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/layers/builder.py

@@ -23,34 +23,35 @@ from torch import nn
 import torch.nn.functional as F
 
 
-def build_glu(
-    act: Callable[[torch.Tensor], torch.Tensor], gate_is_front: bool = False
-) -> Callable[[torch.Tensor], torch.Tensor]:
-  """Builds an activation function with GLU (Gated Linear Unit).
+class GeGLU(nn.Module):
+  """GeGLU is an activation function which is a variant of GELU.
 
-  If gate_is_front is True,
-    f(x) = act(x) * y
-  otherwise,
-    f(x) = x * act(y),
-  where x is the first half of the input and y is the second half of the input.
+  GeGLU(x) = (xW+b) * GELU(xV+c)
+  See: https://arxiv.org/abs/2002.05202v1
+  """
 
-  Args:
-    act (Callable[[torch.Tensor], torch.Tensor]): activation function to apply
-      to the gate.
-    gate_is_front: whether the gate is in front half of the input. Other part is
-      the output in GLU.
+  def __init__(self, d_in: int, d_out: int):
+    super().__init__()
+    self.proj = nn.Linear(d_in, d_out * 2)
 
-  Returns:
-    A callable activation function with GLU.
+  def forward(self, x: torch.Tensor):
+    x, gate = self.proj(x).chunk(2, dim=-1)
+    return x * F.gelu(gate)
+
+
+class SwiGLU(nn.Module):
+  """SwiGLU is an activation function which is a variant of GLU.
+
+  SwiGLU is same as SiLU_GLU, because The SiLU function is also known as the
+  swish function.
+
+  SwiGLU(x) = Swish(xW+b) * (xV+c)
+  See: https://paperswithcode.com/method/swiglu
   """
 
-  def _glu(x):
+  def forward(self, x: torch.Tensor):
     x, y = x.chunk(2, dim=-1)
-    if gate_is_front:
-      return act(x) * y
-    return x * act(y)
-
-  return _glu
+    return F.silu(x) * y
 
 
 def build_norm(dim: int, config: cfg.NormalizationConfig):
@@ -151,10 +152,10 @@ def get_activation(config: cfg.ActivationConfig):
     # See: https://github.com/hendrycks/GELUs
     return lambda x: x * F.sigmoid(1.702 * x)
   elif config.type == cfg.ActivationType.GE_GLU:
-    return build_glu(F.gelu, config.gate_is_front)
+    return GeGLU(config.dim_in, config.dim_out)
   elif config.type == cfg.ActivationType.RELU:
     return F.relu
   elif config.type == cfg.ActivationType.SILU_GLU:
-    return build_glu(F.silu, config.gate_is_front)
+    return SwiGLU()
   else:
     raise ValueError("Unsupported activation type.")

ai_edge_torch/generative/layers/model_config.py

@@ -118,9 +118,9 @@ class AttentionConfig:
 @dataclass
 class ActivationConfig:
   type: ActivationType = ActivationType.LINEAR
-  # Whether to GLU gate is the front part instead of the back part of input
-  # when ActivationType is `GE_GLU` or `SILU_GLU`.
-  gate_is_front: bool = False
+  # Dimension of input and output, used in GeGLU.
+  dim_in: Optional[int] = None
+  dim_out: Optional[int] = None
 
 
 @dataclass

ai_edge_torch/generative/layers/normalization.py

@@ -183,8 +183,16 @@ def group_norm_with_hlfb(
   """
   x = torch.permute(x, (0, 2, 3, 1))
 
+  # TODO: b/366544750 - Change "reduction_axes" field as an array, rather than
+  # int32 when the bug is fixed.
   builder = StableHLOCompositeBuilder(
-      name="odml.group_norm", attr={"num_groups": num_groups, "eps": eps}
+      name="odml.group_norm",
+      attr={
+          "num_groups": num_groups,
+          "epsilon": eps,
+          "reduction_axes": 3,
+          "channel_axis": 3,
+      },
   )
   x, w, b = builder.mark_inputs(x, w, b)
   x = torch.permute(x, (0, 3, 1, 2))
@@ -206,7 +214,7 @@ def layer_norm_with_hlfb(
   """Layer Normalization with high-level function boundary enabled.
 
   Args:
-    x (torch.Tensor): Input tensor for Layer Normalization.
+    x (torch.Tensor): Input tensor for Layer Normalization, with BCHW shape.
     w (torch.Tensor): The weight tensor for the normalization.
     b (torch.Tensor): The bias tensor for the normalization.
     eps (float): A small float value to ensure numerical stability.
@@ -216,7 +224,10 @@ def layer_norm_with_hlfb(
   Returns:
     The output tensor of Layer Normalization.
   """
-  builder = StableHLOCompositeBuilder(name="odml.layer_norm", attr={"eps": eps})
+  builder = StableHLOCompositeBuilder(
+      name="odml.group_norm",
+      attr={"num_groups": 1, "epsilon": eps, "channel_axis": 1},
+  )
   x, w, b = builder.mark_inputs(x, w, b)
   if use_input_shape:
     normalized_shape = x.shape

ai_edge_torch/generative/layers/unet/blocks_2d.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 # ==============================================================================
 
-from typing import List, Optional, Tuple
+from typing import List, Optional, Tuple, Union
 
 from ai_edge_torch.generative.layers.attention import CrossAttention
 from ai_edge_torch.generative.layers.attention import SelfAttention
@@ -416,7 +416,7 @@ class DownEncoderBlock2D(nn.Module):
       time_emb: Optional[torch.Tensor] = None,
       context_tensor: Optional[torch.Tensor] = None,
       output_hidden_states: bool = False,
-  ) -> torch.Tensor | Tuple[torch.Tensor, List[torch.Tensor]]:
+  ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]]]:
     """Forward function of the DownEncoderBlock2D.
 
     Args:

ai_edge_torch/generative/test/test_model_conversion_large.py

@@ -21,7 +21,11 @@ from ai_edge_torch.generative.examples.gemma import gemma1
 from ai_edge_torch.generative.examples.gemma import gemma2
 from ai_edge_torch.generative.examples.openelm import openelm
 from ai_edge_torch.generative.examples.phi import phi2
+from ai_edge_torch.generative.examples.phi import phi3
 from ai_edge_torch.generative.examples.smollm import smollm
+from ai_edge_torch.generative.examples.stable_diffusion import clip as sd_clip
+from ai_edge_torch.generative.examples.stable_diffusion import decoder as sd_decoder
+from ai_edge_torch.generative.examples.stable_diffusion import diffusion as sd_diffusion
 from ai_edge_torch.generative.layers import kv_cache
 from ai_edge_torch.generative.test import utils as test_utils
 import numpy as np
@@ -109,6 +113,17 @@ class TestModelConversion(googletest.TestCase):
         config, pytorch_model, "serving_default", atol=1e-3, rtol=1e-3
     )
 
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_phi3(self):
+    config = phi3.get_fake_model_config()
+    pytorch_model = phi3.Phi3_5Mini(config).eval()
+    self._test_model(
+        config, pytorch_model, "prefill", atol=1e-5, rtol=1e-5
+    )
+
   @googletest.skipIf(
       ai_edge_config.Config.use_torch_xla,
       reason="tests with custom ops are not supported on oss",
@@ -127,6 +142,110 @@ class TestModelConversion(googletest.TestCase):
     pytorch_model = openelm.OpenELM(config).eval()
     self._test_model(config, pytorch_model, "prefill", atol=1e-4, rtol=1e-5)
 
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_stable_diffusion_clip(self):
+    config = sd_clip.get_fake_model_config()
+    prompt_tokens = torch.from_numpy(
+        np.array([[1, 2, 3, 4, 5, 6]], dtype=np.int32)
+    )
+
+    pytorch_model = sd_clip.CLIP(config).eval()
+    torch_output = pytorch_model(prompt_tokens)
+
+    edge_model = ai_edge_torch.signature(
+        "encode", pytorch_model, (prompt_tokens,)
+    ).convert()
+    edge_model.set_interpreter_builder(
+        self._interpreter_builder(edge_model.tflite_model())
+    )
+    edge_output = edge_model(
+        prompt_tokens.numpy(),
+        signature_name="encode",
+    )
+    self.assertTrue(
+        np.allclose(
+            edge_output,
+            torch_output.detach().numpy(),
+            atol=1e-4,
+            rtol=1e-5,
+        )
+    )
+
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_stable_diffusion_diffusion(self):
+    config = sd_diffusion.get_fake_model_config(2)
+    latents = torch.from_numpy(
+        np.random.normal(size=(2, 4, 8, 8)).astype(np.float32)
+    )
+    context = torch.from_numpy(
+        np.random.normal(size=(2, 4, 4)).astype(np.float32)
+    )
+    time_embedding = torch.from_numpy(
+        np.random.normal(size=(2, 2)).astype(np.float32)
+    )
+
+    pytorch_model = sd_diffusion.Diffusion(config).eval()
+    torch_output = pytorch_model(latents, context, time_embedding)
+
+    edge_model = ai_edge_torch.signature(
+        "diffusion", pytorch_model, (latents, context, time_embedding)
+    ).convert()
+    edge_model.set_interpreter_builder(
+        self._interpreter_builder(edge_model.tflite_model())
+    )
+    edge_output = edge_model(
+        latents.numpy(),
+        context.numpy(),
+        time_embedding.numpy(),
+        signature_name="diffusion",
+    )
+    self.assertTrue(
+        np.allclose(
+            edge_output,
+            torch_output.detach().numpy(),
+            atol=1e-4,
+            rtol=1e-5,
+        )
+    )
+
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_stable_diffusion_decoder(self):
+    config = sd_decoder.get_fake_model_config()
+    latents = torch.from_numpy(
+        np.random.normal(size=(1, 4, 64, 64)).astype(np.float32)
+    )
+
+    pytorch_model = sd_decoder.Decoder(config).eval()
+    torch_output = pytorch_model(latents)
+
+    edge_model = ai_edge_torch.signature(
+        "decode", pytorch_model, (latents,)
+    ).convert()
+    edge_model.set_interpreter_builder(
+        self._interpreter_builder(edge_model.tflite_model())
+    )
+    edge_output = edge_model(
+        latents.numpy(),
+        signature_name="decode",
+    )
+    self.assertTrue(
+        np.allclose(
+            edge_output,
+            torch_output.detach().numpy(),
+            atol=1e-4,
+            rtol=1e-5,
+        )
+    )
+
 
 if __name__ == "__main__":
   googletest.main()

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.3.0.dev20240923"
+__version__ = "0.3.0.dev20240925"

{ai_edge_torch_nightly-0.3.0.dev20240923.dist-info → ai_edge_torch_nightly-0.3.0.dev20240925.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20240923
+Version: 0.3.0.dev20240925
 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
@@ -30,6 +30,7 @@ Requires-Dist: tabulate
 Requires-Dist: torch>=2.4.0
 Requires-Dist: torch-xla>=2.4.0
 Requires-Dist: tf-nightly>=2.18.0.dev20240722
+Requires-Dist: ai-edge-litert-nightly
 Requires-Dist: ai-edge-quantizer-nightly
 
 Library that supports converting PyTorch models into a .tflite format, which can

{ai_edge_torch_nightly-0.3.0.dev20240923.dist-info → ai_edge_torch_nightly-0.3.0.dev20240925.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=oxtOOEY9LJkV5vRrgr1EoSjAjuetYVNq7WQqMuauRkc,706
+ai_edge_torch/version.py,sha256=UXj1-90S3RDoHwYSmy9VdMC0Sm3EHt9ESLZbi3hnWus,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
@@ -48,22 +48,25 @@ ai_edge_torch/generative/examples/gemma/verify_gemma2.py,sha256=kSzn1ITJXqrtNQax
 ai_edge_torch/generative/examples/gemma/verify_util.py,sha256=HBK2d8FcWFoxVDF5zk9sLSbKZEtwZQhX-K_zm4AvQtQ,5160
 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=gGkHELNrt4xqnu11fCh3sJbZ7OsPyvoiF1J1aKCs5r8,7532
+ai_edge_torch/generative/examples/openelm/openelm.py,sha256=VcU8A0B9nQR-FTPHXqNHSHZzeIZZ_As4yvKZMnoU2P4,7482
 ai_edge_torch/generative/examples/openelm/verify.py,sha256=QdFKymQSCYFJcYVvA63u5uIsn1YxJ0JZD5UqN6gxraI,2112
 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/verify.py,sha256=QPYX6weEZGMEXt_Vb2hNARPAECQBKzx-KCivd4dzOrw,2145
+ai_edge_torch/generative/examples/phi/phi3.py,sha256=DIDzpG8DZkWDcWsAVkcxzxIC3U3352uVI3zMoYZD16U,9554
+ai_edge_torch/generative/examples/phi/verify.py,sha256=5pQ0Bt8vGl8uTpkgXvOx8G7_rju0Gi8mIEr5NtRSAbs,2145
+ai_edge_torch/generative/examples/phi/verify_phi3.py,sha256=o1UTqpimkeX3MDjgdG1QTQkoZHvCEnGClA0J0WB3wJ4,2328
 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=G2dAcl-VhAbx1E1PEqM6hpzPF24HqFZaz7UBEpJSQ3w,2022
 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=tL6w2dr6VP66IXjSKo9StDNP-wl0RO3fh6dIliiYlFA,4656
+ai_edge_torch/generative/examples/stable_diffusion/clip.py,sha256=2RMi5UmfMT4Ep68ZLJsqF-fMvEumNVkIwqtsRli9HhA,6068
 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=slieF2-QcDCwd4DRZ7snsZIphT97IXpp4plRRsRSwL8,13983
-ai_edge_torch/generative/examples/stable_diffusion/diffusion.py,sha256=7o-5oJARCm4fhRwmNv84ofmajP5MMIS102vj4d8eeRQ,31248
+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/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
@@ -78,8 +81,9 @@ ai_edge_torch/generative/examples/t5/convert_to_tflite.py,sha256=HHtZTtUh3QgE4F7
 ai_edge_torch/generative/examples/t5/t5.py,sha256=OZ67knK-UB1dBjxydG-Jwkp0Z3FzOCqGPTdg5aBFu4w,21328
 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/toy_model.py,sha256=QyLeCqDnk71WvvFH68g9UeF-HytonSk1ItGF9dc7Zj8,5854
-ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=e_Kqm5dStSrNE9_aIYC-vYJRsqLn-hJVkmR4QjYqZI0,5913
+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/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
@@ -89,15 +93,15 @@ ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py,sha256=hhxSQvkD
 ai_edge_torch/generative/layers/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/layers/attention.py,sha256=Z0Y_G8IG0LmvLX2u9D8__Fkr22szB-az6wMNnZpzhkA,13233
 ai_edge_torch/generative/layers/attention_utils.py,sha256=68GXGR2HSWBFViTxX7cHifzVG-kcLS2IL2tQJPIpupg,7344
-ai_edge_torch/generative/layers/builder.py,sha256=toT9Tl1x9o5KbG-eGOEViUr4fd_4f-XLZdMQT0Ae5_8,5130
+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=d0Y-EFb4Rr7iLZ4Bsdf1i92KuhY1BXRqyeUN2kuu510,6923
-ai_edge_torch/generative/layers/normalization.py,sha256=l_36uFdruJwqqyubnBTM0M-iGiJfeFafyXKPPK8KHVo,6713
+ai_edge_torch/generative/layers/model_config.py,sha256=l5Rb3h3GK2pux-Lg3BONTD6b7klxXqUbDDtYs_bGKLk,6879
+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
 ai_edge_torch/generative/layers/unet/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/layers/unet/blocks_2d.py,sha256=c8rtlfDaeKmUfiiTKPmQhNW-U5vW9jFB2pPPcvT6qsc,27527
+ai_edge_torch/generative/layers/unet/blocks_2d.py,sha256=JwndhL3Z31TvkdGlAoTL5PQzmKfHdRWaaE1EbaMI4Gs,27540
 ai_edge_torch/generative/layers/unet/builder.py,sha256=zAqWXdimmMrQRhmE_t9XkS68mh6PSrzwb-2NZZXrR5I,1901
 ai_edge_torch/generative/layers/unet/model_config.py,sha256=8ze9kVWMuyZVQcgK7hWYw9TM1W9lXD-2j0iMHlxoGX4,9267
 ai_edge_torch/generative/quantize/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
@@ -111,7 +115,7 @@ 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=mAK8Pm4mgGyilDSBtFazCRDetoqYKKB0sGC83MPKE0M,4494
+ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=IzW2HjXS2-zePZM-qEuXL4zclnGvYsNw-6tuDSeNna4,8163
 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
@@ -166,8 +170,8 @@ 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.dev20240923.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.3.0.dev20240923.dist-info/METADATA,sha256=BgwLxDJ3AOPVn0fkngAQpf3YdmShufhMt3bANFevtiQ,1859
-ai_edge_torch_nightly-0.3.0.dev20240923.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-ai_edge_torch_nightly-0.3.0.dev20240923.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.3.0.dev20240923.dist-info/RECORD,,
+ai_edge_torch_nightly-0.3.0.dev20240925.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20240925.dist-info/METADATA,sha256=5KsshdZ4-3X193HkoO2ukceyDEdWGvb8ZEMcw88qt7k,1897
+ai_edge_torch_nightly-0.3.0.dev20240925.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20240925.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20240925.dist-info/RECORD,,