ai-edge-torch-nightly 0.3.0.dev20240926__py3-none-any.whl → 0.3.0.dev20240929__py3-none-any.whl

ai_edge_torch/generative/examples/openelm/verify.py

@@ -20,6 +20,7 @@ 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
 
@@ -29,15 +30,18 @@ _PROMPTS = flags.DEFINE_multi_string(
     "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"
   logging.info("Loading the original model from: %s", checkpoint)
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(
-          checkpoint, trust_remote_code=True
-      ),
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(
+      checkpoint, trust_remote_code=True
   )
 
   # Locate the cached dir.
@@ -53,10 +57,13 @@ def main(_):
   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/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

@@ -97,15 +97,15 @@ ROPE_SHORT_FACTOR = [
 ]
 
 
-def build_rope_cache(
+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,
+    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.
 
@@ -116,26 +116,20 @@ def build_rope_cache(
   Args:
       size (int): The size of the built cache.
       dim (int): Each sequence's dimmension.
-      base (int, optional): Rope base value. Defaults to 10000.
+      base (int, optional): Rope base value.
       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.
+        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. Defaults to None.
-      scale (float, optional): A float used to scale the rope values. Defaults
-        to 1.0.
+        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.
   """
-  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
+  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
@@ -167,10 +161,10 @@ class Phi3_5Mini(nn.Module):
         config.final_norm_config,
     )
     attn_config = block_config.attn_config
-    self.rope_cache = build_rope_cache(
+    self.rope_cache = _build_rope_cache(
         size=config.kv_cache_max,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
+        base=attn_config.rotary_base,
         condense_ratio=1,
         dtype=torch.float32,
         device=torch.device("cpu"),
@@ -181,8 +175,6 @@ class Phi3_5Mini(nn.Module):
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
         size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
     )
     self.config = config
 
@@ -238,6 +230,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       num_heads=32,
       head_dim=96,
       num_query_groups=32,
+      rotary_base=10000,
       rotary_percentage=1.0,
       qkv_transpose_before_split=True,
   )

ai_edge_torch/generative/examples/phi/verify.py

@@ -19,6 +19,7 @@ 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
@@ -39,12 +40,7 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
 def main(_):
   checkpoint = kagglehub.model_download("Microsoft/phi/transformers/2")
   logging.info("Loading the original model from: %s", 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,
-  )
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
 
   logging.info("Building the reauthored model from: %s", checkpoint)
   reauthored_model = phi2.build_model(checkpoint)
@@ -53,10 +49,13 @@ def main(_):
   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

@@ -21,6 +21,7 @@ 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
 
@@ -40,12 +41,7 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
 def main(_):
   checkpoint = "microsoft/Phi-3.5-mini-instruct"
   logging.info("Loading the original model from: %s", 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,
-  )
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
 
   # Locate the cached dir.
   cached_config_file = transformers.utils.cached_file(
@@ -59,10 +55,13 @@ def main(_):
   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,
   )
 
 

ai_edge_torch/generative/examples/qwen/__init__.py

@@ -0,0 +1,14 @@
+# 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.
+# ==============================================================================

ai_edge_torch/generative/examples/qwen/convert_to_tflite.py

@@ -0,0 +1,81 @@
+# 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 Qwen 2.5 models to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.qwen import qwen
+from ai_edge_torch.generative.utilities import converter
+
+_MODEL_SIZE = flags.DEFINE_enum(
+    'model_size',
+    '3b',
+    ['0.5b', '1.5b', '3b'],
+    'The size of the model to convert.',
+)
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/qwen'),
+    '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.',
+)
+
+_BUILDER = {
+    '0.5b': qwen.build_0_5b_model,
+    '1.5b': qwen.build_1_5b_model,
+    '3b': qwen.build_3b_model,
+}
+
+
+def main(_):
+  pytorch_model = _BUILDER[_MODEL_SIZE.value](
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  quant_suffix = 'q8' if _QUANTIZE.value else 'f32'
+  model_size = _MODEL_SIZE.value.replace('.', '_')
+  output_filename = f'qwen_{model_size}_{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/qwen/qwen.py

@@ -0,0 +1,141 @@
+# 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 Qwen 2.5 models."""
+
+import copy
+
+from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+import ai_edge_torch.generative.layers.model_config as cfg
+import ai_edge_torch.generative.utilities.loader as loading_utils
+from torch import nn
+
+TENSOR_NAMES = copy.copy(tiny_llama.TENSOR_NAMES)
+# Qwen re-uses the embedding as the head projection layer.
+TENSOR_NAMES.lm_head = None
+
+
+class Qwen(tiny_llama.TinyLlama):
+  """A Qwen model built from the Edge Generative API layers.
+
+  Qwen 2.5 shares the same architecture as TinyLlama.
+  """
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__(config)
+    # Qwen re-uses the embedding as the head projection layer.
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+
+
+def get_3b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 2.5 3B model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a SmolLM model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=16,
+      head_dim=128,
+      num_query_groups=2,
+      rotary_base=1000000,
+      rotary_percentage=1.0,
+      qkv_use_bias=True,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
+      intermediate_size=11008,
+  )
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM,
+      epsilon=1e-06,
+  )
+  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=151936,
+      num_layers=36,
+      max_seq_len=32768,
+      embedding_dim=2048,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_1_5b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 2.5 1B model."""
+  config = get_3b_model_config(kv_cache_max_len)
+  # Qwen has only one block config.
+  block_config = config.block_config(0)
+  block_config.attn_config.num_heads = 12
+  block_config.ff_config.intermediate_size = 8960
+  config.num_layers = 28
+  config.embedding_dim = 1536
+  return config
+
+
+def get_0_5b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 2.5 0.5B model."""
+  config = get_3b_model_config(kv_cache_max_len)
+  # Qwen has only one block config.
+  block_config = config.block_config(0)
+  block_config.attn_config.num_heads = 14
+  block_config.attn_config.head_dim = 64
+  block_config.ff_config.intermediate_size = 4864
+  config.num_layers = 24
+  config.embedding_dim = 896
+  return config
+
+
+def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
+  config = get_3b_model_config(**kwargs)
+  config.vocab_size = 128
+  config.num_layers = 2
+  # Qwen has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 64
+  return config
+
+
+def _build_model(checkpoint_path: str, config: cfg.ModelConfig) -> nn.Module:
+  model = Qwen(config)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  # Since embedding and lm-head use the same weight, we need to set strict
+  # to False.
+  loader.load(model, strict=False)
+  model.eval()
+  return model
+
+
+def build_3b_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  return _build_model(checkpoint_path, get_3b_model_config(**kwargs))
+
+
+def build_1_5b_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  return _build_model(checkpoint_path, get_1_5b_model_config(**kwargs))
+
+
+def build_0_5b_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  return _build_model(checkpoint_path, get_0_5b_model_config(**kwargs))

ai_edge_torch/generative/examples/qwen/verify.py

@@ -0,0 +1,88 @@
+# 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 Qwen 2.5 0.5B, 1.5B, and 3B models."""
+
+import logging
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.qwen import qwen
+from ai_edge_torch.generative.utilities import transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+import transformers
+
+
+_MODEL_SIZE = flags.DEFINE_enum(
+    "model_size",
+    "3b",
+    ["0.5b", "1.5b", "3b"],
+    "The size of the model to verify.",
+)
+_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.",
+)
+
+_CHECKPOINT = {
+    "0.5b": "Qwen/Qwen2.5-0.5B-Instruct",
+    "1.5b": "Qwen/Qwen2.5-1.5B-Instruct",
+    "3b": "Qwen/Qwen2.5-3B-Instruct",
+}
+
+_BUILDER = {
+    "0.5b": qwen.build_0_5b_model,
+    "1.5b": qwen.build_1_5b_model,
+    "3b": qwen.build_3b_model,
+}
+
+
+def main(_):
+  checkpoint = _CHECKPOINT[_MODEL_SIZE.value]
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
+
+  # Locate the cached dir.
+  cached_config_file = transformers.utils.cached_file(
+      checkpoint, transformers.utils.CONFIG_NAME
+  )
+  reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
+  reauthored_model = _BUILDER[_MODEL_SIZE.value](reauthored_checkpoint)
+
+  logging.info("Loading the tokenizer from: %s", checkpoint)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
+
+  verifier.verify_reauthored_model(
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
+      generate_prompts=_PROMPTS.value,
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+      atol=1e-04,
+  )
+
+
+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(

ai_edge_torch/generative/examples/smollm/verify.py

@@ -21,6 +21,7 @@ import pathlib
 from absl import app
 from absl import flags
 from ai_edge_torch.generative.examples.smollm import smollm
+from ai_edge_torch.generative.utilities import transformers_verifier
 from ai_edge_torch.generative.utilities import verifier
 import transformers
 
@@ -30,14 +31,18 @@ _PROMPTS = flags.DEFINE_multi_string(
     "What is the meaning of life?",
     "The input prompts to generate answers.",
 )
+_MAX_NEW_TOKENS = flags.DEFINE_integer(
+    "max_new_tokens",
+    30,
+    "The maximum size of the generated tokens.",
+)
 
 
 def main(_):
   checkpoint = "HuggingFaceTB/SmolLM-135M"
   logging.info("Loading the original model from: %s", checkpoint)
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(checkpoint),
-  )
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
+
   # Locate the cached dir.
   cached_config_file = transformers.utils.cached_file(
       checkpoint, transformers.utils.CONFIG_NAME
@@ -50,10 +55,13 @@ def main(_):
   tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
 
   verifier.verify_reauthored_model(
-      original_model=wrapper_model,
-      reauthored_model=reauthored_model,
-      tokenizer=tokenizer,
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
       generate_prompts=_PROMPTS.value,
+      max_new_tokens=_MAX_NEW_TOKENS.value,
       atol=1e-04,
   )
 

ai_edge_torch/generative/examples/stable_diffusion/clip.py

@@ -98,6 +98,7 @@ def get_model_config() -> cfg.ModelConfig:
       num_heads=num_heads,
       head_dim=embedding_dim // num_heads,
       num_query_groups=num_query_groups,
+      rotary_base=0,
       rotary_percentage=0.0,
       qkv_use_bias=True,
       qkv_transpose_before_split=True,
@@ -148,6 +149,7 @@ def get_fake_model_config() -> cfg.ModelConfig:
       num_heads=num_heads,
       head_dim=embedding_dim // num_heads,
       num_query_groups=num_query_groups,
+      rotary_base=0,
       rotary_percentage=0.0,
       qkv_use_bias=True,
       qkv_transpose_before_split=True,

ai_edge_torch/generative/examples/stable_diffusion/decoder.py

@@ -295,6 +295,7 @@ def get_model_config() -> unet_cfg.AutoEncoderConfig:
           enable_kv_cache=False,
           qkv_transpose_before_split=True,
           qkv_fused_interleaved=False,
+          rotary_base=0,
           rotary_percentage=0.0,
       ),
       enable_hlfb=False,
@@ -351,6 +352,7 @@ def get_fake_model_config() -> unet_cfg.AutoEncoderConfig:
           enable_kv_cache=False,
           qkv_transpose_before_split=True,
           qkv_fused_interleaved=False,
+          rotary_base=0,
           rotary_percentage=0.0,
       ),
       enable_hlfb=False,

ai_edge_torch/generative/examples/stable_diffusion/diffusion.py

@@ -199,6 +199,7 @@ def build_attention_config(
     num_heads,
     dim,
     num_query_groups,
+    rotary_base=0,
     rotary_percentage=0.0,
     qkv_transpose_before_split=True,
     qkv_use_bias=False,
@@ -211,6 +212,7 @@ def build_attention_config(
       num_heads=num_heads,
       head_dim=dim // num_heads,
       num_query_groups=num_query_groups,
+      rotary_base=rotary_base,
       rotary_percentage=rotary_percentage,
       qkv_transpose_before_split=qkv_transpose_before_split,
       qkv_use_bias=qkv_use_bias,

ai_edge_torch/generative/examples/t5/t5.py

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