ai-edge-torch-nightly 0.4.0.dev20250227__py3-none-any.whl → 0.4.0.dev20250301__py3-none-any.whl

ai_edge_torch/generative/examples/phi/convert_phi4_to_tflite.py

@@ -0,0 +1,80 @@
+# Copyright 2025 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Example of converting a Phi-4 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 phi4
+from ai_edge_torch.generative.utilities import converter
+from ai_edge_torch.generative.utilities.model_builder import ExportConfig
+
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/phi4'),
+    'The path to the model checkpoint, or directory holding the checkpoint.',
+)
+_OUTPUT_PATH = flags.DEFINE_string(
+    'output_path',
+    '/tmp/',
+    'The path to export the tflite model.',
+)
+_OUTPUT_NAME_PREFIX = flags.DEFINE_string(
+    'output_name_prefix',
+    'phi4',
+    'The prefix of the output tflite model name.',
+)
+_PREFILL_SEQ_LENS = flags.DEFINE_multi_integer(
+    'prefill_seq_lens',
+    (8, 64, 128, 256, 512, 1024),
+    'List of the maximum sizes of prefill input tensors.',
+)
+_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.',
+)
+_LORA_RANKS = flags.DEFINE_multi_integer(
+    'lora_ranks',
+    None,
+    'If set, the model will be converted with the provided list of LoRA ranks.',
+)
+
+
+def main(_):
+  pytorch_model = phi4.build_model(
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  converter.convert_to_tflite(
+      pytorch_model,
+      output_path=_OUTPUT_PATH.value,
+      output_name_prefix=_OUTPUT_NAME_PREFIX.value,
+      prefill_seq_len=_PREFILL_SEQ_LENS.value,
+      quantize=_QUANTIZE.value,
+      lora_ranks=_LORA_RANKS.value,
+      export_config=ExportConfig(),
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/phi/phi3.py

@@ -136,10 +136,7 @@ def _build_phi3_rope(
 
 class Phi3_5Mini(model_builder.DecoderOnlyModel):
   """A Phi-3.5 model built from the Edge Generative API layers."""
-
-  def __init__(self, config: cfg.ModelConfig):
-    super().__init__(config)
-    attn_config = self.config.block_config(0).attn_config
+  pass
 
 
 def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
@@ -150,7 +147,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       is 1024.
 
   Returns:
-    The model config for a Phi-2 model.
+    The model config for a Phi-3.5 model.
   """
   attn_config = cfg.AttentionConfig(
       num_heads=32,

ai_edge_torch/generative/examples/phi/phi4.py

@@ -0,0 +1,165 @@
+# 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-4 model up to 4K tokens, not to 128K tokens."""
+
+from functools import partial
+import math
+from typing import Tuple
+
+import ai_edge_torch.generative.layers.model_config as cfg
+from ai_edge_torch.generative.utilities import model_builder
+import ai_edge_torch.generative.utilities.loader as loading_utils
+import torch
+
+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",
+)
+
+# max_position_embeddings / original_max_position_embeddings in Phi-4 config.
+ROPE_SCALE_FACTOR = 32
+
+# ROPE short factor in Phi-4 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] * 48
+
+
+def _build_phi4_rope(
+    input_pos: int,
+    n_elem: int,
+    base: int,
+    condense_ratio: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    theta_factors: torch.Tensor,
+    scale: float,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+  """Computes Rotary Positional Embeddings for Phi-4 model.
+
+  It's a modified version of attn_utils.build_rope_cache with additional
+  arguments for Phi-4 model. It precompute Rotary Positional Embedding Sin and
+  Cos values with scaling factors for quick lookup during the inference.
+
+  Args:
+      input_pos (torch.Tensor): the given input sequence positions
+      n_elem (int): Each sequence's dimmension.
+      base (int, optional): Rope base value.
+      condense_ratio (int, optional): The ratio by which sequence indicies are
+        condensed.
+      dtype (torch.dtype, optional): Output tensor's data type.
+      device (torch.device, optional): Output tensor's data type.
+      theta_factors (torch.Tensor, optional): A tensor of shape (n_elem,) used
+        to scale the theta values.
+      scale (float, optional): A float used to scale the rope values.
+
+  Returns:
+      Tuple[torch.Tensor, torch.Tensor]: Rope's Cosine and Sine waves.
+  """
+  theta = 1.0 / (base ** (torch.arange(0, n_elem, 2).float() / n_elem))
+  theta = theta / theta_factors
+  seq_idx = input_pos / 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 Phi4Mini(model_builder.DecoderOnlyModel):
+  """A Phi-4 model built from the Edge Generative API layers."""
+  pass
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Phi-4 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-4 model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=24,
+      head_dim=128,
+      num_query_groups=8,
+      rotary_base=10000,
+      rotary_percentage=0.75,
+      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,
+  )
+
+  max_seq_len = 4096
+  # Create the RoPE callable
+  build_rope = partial(
+      _build_phi4_rope,
+      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(max_seq_len)),
+  )
+
+  config = cfg.ModelConfig(
+      vocab_size=200064,
+      num_layers=32,
+      max_seq_len=max_seq_len,
+      kv_cache_max_len=kv_cache_max_len,
+      embedding_dim=3072,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+      build_rope=build_rope,
+  )
+  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-4 has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 128
+  return config
+
+
+def build_model(checkpoint_path: str, **kwargs) -> torch.nn.Module:
+  """Instantiates the model instance and load checkpoint if provided."""
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+      model_class=Phi4Mini,
+  )

ai_edge_torch/generative/examples/phi/verify_phi4.py

@@ -0,0 +1,69 @@
+# Copyright 2025 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Verifies the reauthored Phi-4 model."""
+
+import logging
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.phi import phi4
+from ai_edge_torch.generative.utilities import transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+import transformers
+
+
+_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-4-mini-instruct"
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
+
+  # Locate the cached dir.
+  cached_config_file = transformers.utils.cached_file(
+      checkpoint, transformers.utils.CONFIG_NAME
+  )
+  reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
+  reauthored_model = phi4.build_model(reauthored_checkpoint)
+
+  logging.info("Loading the tokenizer from: %s", checkpoint)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
+
+  verifier.verify_reauthored_model(
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
+      generate_prompts=_PROMPTS.value,
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/layers/experimental/attention.py

@@ -52,7 +52,6 @@ class TransformerBlock(nn.Module):
         config.pre_attention_norm_config,
     )
     self.atten_func = CausalSelfAttention(
-        model_config.batch_size,
         model_config.embedding_dim,
         config.attn_config,
         model_config.enable_hlfb,
@@ -119,7 +118,6 @@ class CausalSelfAttention(nn.Module):
 
   def __init__(
       self,
-      batch_size: int,
       dim: int,
       config: cfg.AttentionConfig,
       enable_hlfb: bool,
@@ -127,14 +125,12 @@ class CausalSelfAttention(nn.Module):
     """Initialize an instance of CausalSelfAttention.
 
     Args:
-      batch_size (int): batch size of the input tensor.
       dim (int): causal attention's input/output dimmension.
       config (cfg.AttentionConfig): attention specific configurations.
       enable_hlfb (bool): whether hlfb is enabled or not.
     """
     super().__init__()
     self.kv_cache = None
-    self.batch_size = batch_size
     qkv_shape = (
         config.num_heads + 2 * config.num_query_groups
     ) * config.head_dim
@@ -180,10 +176,6 @@ class CausalSelfAttention(nn.Module):
     """
     # Batch size, sequence length, embedding dimensionality.
     B, T, E = x.size()
-    assert B == self.batch_size, (
-        "batch size of input tensor must match with the batch size specified in"
-        " the model configuration."
-    )
 
     qkv = self.qkv_projection(x)
 

ai_edge_torch/generative/layers/experimental/kv_cache.py

@@ -21,23 +21,19 @@ This is an experimental implementation and is subject to change at any time.
 import dataclasses
 from typing import List, Tuple
 
-from ai_edge_torch import hlfb
 from ai_edge_torch.generative.layers import model_config
-from ai_edge_torch.generative.layers.experimental import types as types
-from ai_edge_torch.generative.utilities.dynamic_update_slice import dynamic_update_slice
+from ai_edge_torch.generative.layers.experimental import types
+from ai_edge_torch.generative.utilities import dynamic_update_slice as dus_utils
 import torch
-import torch.nn as nn
 import torch.utils._pytree as pytree
 
-BATCH_SIZE = 1
-
 
 @dataclasses.dataclass
 class KVCacheEntryBase:
   """A single cache entry that includes K and V caches.
 
   The chaches are built based on the provided config with the shape of
-  (batch_size=1, kv_cache_max, num_query_groups, head_dim).
+  (batch_size, kv_cache_max, num_query_groups, head_dim).
   """
 
   k_cache: torch.Tensor
@@ -46,10 +42,8 @@ class KVCacheEntryBase:
   @classmethod
   def _from_model_config(
       cls,
-      kv_cache_max: int,
-      config: model_config.AttentionConfig,
-      k_shape: Tuple,
-      v_shape: Tuple,
+      k_shape: Tuple[int, ...],
+      v_shape: Tuple[int, ...],
       dtype: torch.dtype = torch.float32,
       device: torch.device = None,
   ) -> "KVCacheEntryBase":
@@ -66,12 +60,11 @@ class KVCacheEntryBase:
       config: model_config.AttentionConfig,
       dtype: torch.dtype = torch.float32,
       device: torch.device = None,
+      batch_size: int = 1,
   ) -> "KVCacheEntryBase":
     """Build an instance of the class based on model config."""
-    shape = (BATCH_SIZE, kv_cache_max, config.num_query_groups, config.head_dim)
-    return cls._from_model_config(
-        kv_cache_max, config, shape, shape, dtype, device
-    )
+    shape = (batch_size, kv_cache_max, config.num_query_groups, config.head_dim)
+    return cls._from_model_config(shape, shape, dtype, device)
 
 
 @dataclasses.dataclass
@@ -93,24 +86,22 @@ class KVCacheEntryTransposed(KVCacheEntryBase):
       config: model_config.AttentionConfig,
       dtype: torch.dtype = torch.float32,
       device: torch.device = None,
+      batch_size: int = 1,
   ) -> "KVCacheEntryBase":
     """Build an instance of the class based on model config."""
-    num_kv_heads = config.num_query_groups
     k_shape = (
-        1,
-        BATCH_SIZE * num_kv_heads,
+        batch_size,
+        config.num_query_groups,
         kv_cache_max,
         config.head_dim,
-    )  # 1, bk, s, h
+    )  # b, k, s, h
     v_shape = (
-        1,
-        BATCH_SIZE * num_kv_heads,
+        batch_size,
+        config.num_query_groups,
         config.head_dim,
         kv_cache_max,
-    )  # 1, bk, h, s
-    return cls._from_model_config(
-        kv_cache_max, config, k_shape, v_shape, dtype, device
-    )
+    )  # b, k, h, s
+    return cls._from_model_config(k_shape, v_shape, dtype, device)
 
 
 @dataclasses.dataclass
@@ -126,6 +117,7 @@ class KVCacheBase:
       config: model_config.ModelConfig,
       dtype: torch.dtype = torch.float32,
       device: torch.device = None,
+      batch_size: int = 1,
   ) -> "KVCacheBase":
     caches = [
         kv_entry_cls.from_model_config(
@@ -133,6 +125,7 @@ class KVCacheBase:
             config.block_config(idx).attn_config,
             dtype,
             device,
+            batch_size,
         )
         for idx in range(config.num_layers)
     ]
@@ -145,6 +138,7 @@ class KVCacheBase:
       config: model_config.ModelConfig,
       dtype: torch.dtype = torch.float32,
       device: torch.device = None,
+      batch_size: int = 1,
   ) -> "KVCacheBase":
     """Build an instance of the class based on model config.
 
@@ -154,12 +148,19 @@ class KVCacheBase:
           Defaults to torch.float32.
         device (torch.device, optional): The device placement of the cache
           tensors. Defaults to None.
+        batch_size (int, optional): The batch size of the cache tensors.
+          Defaults to 1.
 
     Returns:
         KVCacheBase: The created cache object.
     """
+    assert batch_size == 1, "Batch size must be 1 for KV Cache."
     return cls._from_model_config(
-        KVCacheEntryBase, config=config, dtype=dtype, device=device
+        KVCacheEntryBase,
+        config=config,
+        dtype=dtype,
+        device=device,
+        batch_size=batch_size,
     )
 
   def flatten(self) -> List[torch.Tensor]:
@@ -177,9 +178,14 @@ class KVCacheBTNH(KVCacheBase):
       config: model_config.ModelConfig,
       dtype: torch.dtype = torch.float32,
       device: torch.device = None,
+      batch_size: int = 1,
   ) -> "KVCacheBTNH":
     return cls._from_model_config(
-        KVCacheEntryBTNH, config=config, dtype=dtype, device=device
+        KVCacheEntryBTNH,
+        config=config,
+        dtype=dtype,
+        device=device,
+        batch_size=batch_size,
     )
 
 
@@ -192,9 +198,14 @@ class KVCacheTransposed(KVCacheBase):
       config: model_config.ModelConfig,
       dtype: torch.dtype = torch.float32,
       device: torch.device = None,
+      batch_size: int = 1,
   ) -> "KVCacheBTNH":
     return cls._from_model_config(
-        KVCacheEntryTransposed, config=config, dtype=dtype, device=device
+        KVCacheEntryTransposed,
+        config=config,
+        dtype=dtype,
+        device=device,
+        batch_size=batch_size,
     )
 
 
@@ -258,7 +269,6 @@ def update(
     input_pos: torch.Tensor,
     k_slice: torch.Tensor,
     v_slice: torch.Tensor,
-    use_dus: bool = True,
 ) -> KVCacheEntryBase:
   """Out of place update of Cache buffer.
 
@@ -309,6 +319,10 @@ def _update_kv_impl(
   positions = input_pos.clone()
   k_slice_indices = _get_slice_indices(positions, cache_dim, k_ts_idx)
   v_slice_indices = _get_slice_indices(positions, cache_dim, v_ts_idx)
-  k = dynamic_update_slice(cache.k_cache, k_slice, [x for x in k_slice_indices])
-  v = dynamic_update_slice(cache.v_cache, v_slice, [x for x in v_slice_indices])
+  k = dus_utils.dynamic_update_slice(
+      cache.k_cache, k_slice, [x for x in k_slice_indices]
+  )
+  v = dus_utils.dynamic_update_slice(
+      cache.v_cache, v_slice, [x for x in v_slice_indices]
+  )
   return KVCacheEntryTransposed(k, v)

ai_edge_torch/generative/test/test_model_conversion_large.py

@@ -27,6 +27,7 @@ from ai_edge_torch.generative.examples.paligemma import decoder2
 from ai_edge_torch.generative.examples.paligemma import paligemma
 from ai_edge_torch.generative.examples.phi import phi2
 from ai_edge_torch.generative.examples.phi import phi3
+from ai_edge_torch.generative.examples.phi import phi4
 from ai_edge_torch.generative.examples.qwen import qwen
 from ai_edge_torch.generative.examples.qwen_vl import qwen_vl
 from ai_edge_torch.generative.examples.smollm import smollm
@@ -139,6 +140,15 @@ class TestModelConversion(googletest.TestCase):
     pytorch_model = phi3.Phi3_5Mini(config).eval()
     self._test_model(config, pytorch_model, "prefill", atol=1e-5, rtol=1e-5)
 
+  @googletest.skipIf(
+      ai_edge_torch.config.in_oss,
+      reason="tests with custom ops are not supported in oss",
+  )
+  def test_phi4(self):
+    config = phi4.get_fake_model_config()
+    pytorch_model = phi4.Phi4Mini(config).eval()
+    self._test_model(config, pytorch_model, "prefill", atol=1e-3, rtol=1e-5)
+
   @googletest.skipIf(
       ai_edge_torch.config.in_oss,
       reason="tests with custom ops are not supported in oss",

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.4.0.dev20250227"
+__version__ = "0.4.0.dev20250301"

{ai_edge_torch_nightly-0.4.0.dev20250227.dist-info → ai_edge_torch_nightly-0.4.0.dev20250301.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.4.0.dev20250227
+Version: 0.4.0.dev20250301
 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-0.4.0.dev20250227.dist-info → ai_edge_torch_nightly-0.4.0.dev20250301.dist-info}/RECORD

@@ -2,7 +2,7 @@ ai_edge_torch/__init__.py,sha256=8sPR_5uXJA4NEE0nIwNdSl-ADOJEoR8hAgYvBQDY70Y,120
 ai_edge_torch/_config.py,sha256=AiqhbcheF7j_ozIGDLC89k1we95aVgFDa-tR6h7UI0s,2529
 ai_edge_torch/conftest.py,sha256=r0GTrhMRhlmOGrrkvumHN8hkmyug6WvF60vWq8wRIBI,758
 ai_edge_torch/model.py,sha256=N-pNpTxzhaFGhWhnSGd70lBzb9VlEhTOq5mddU7bvvI,5542
-ai_edge_torch/version.py,sha256=K2jtDrBNGi74j_uQYVUT6MJ2-aQFKkKy5ZYur9iWdVU,706
+ai_edge_torch/version.py,sha256=MENyVQGKk5h6YnKhfVQlzGJnWaGJrL8J86HAtU_LAQM,706
 ai_edge_torch/_convert/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/_convert/conversion.py,sha256=gpXQnifODU-mWxkUZw_3ov1lEYBw1SPVIcqj5k7pTGo,5550
 ai_edge_torch/_convert/conversion_utils.py,sha256=Sr8qXVcTwc-ZnZmK7yxVrIOOp1S_vNrwzC0zUvLTI2o,2160
@@ -84,11 +84,14 @@ ai_edge_torch/generative/examples/paligemma/verify_decoder2.py,sha256=tm-UfLr0Ye
 ai_edge_torch/generative/examples/paligemma/verify_image_encoder.py,sha256=vNm-wTT8BD6zbX6GocfP1QrVoHl0zSvuVxoXN36eeiU,3540
 ai_edge_torch/generative/examples/phi/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/phi/convert_phi3_to_tflite.py,sha256=CaI_-Vtd0j9FoWIDd8q5z4CFsGYUhTwEWGvMGaXICuU,2514
+ai_edge_torch/generative/examples/phi/convert_phi4_to_tflite.py,sha256=hu_fMYqHU_bxE3DzE-sNj8YSrsFLmErnNRZOODVXZjE,2512
 ai_edge_torch/generative/examples/phi/convert_to_tflite.py,sha256=g-MvEibJT_iIhkec2VGtFFA_iP54VCq9mY4KxwAYF08,2512
 ai_edge_torch/generative/examples/phi/phi2.py,sha256=c6PYCky7yJn6MVIYOCTx8S_CH27kOPmJbRZcI95nbZs,3477
-ai_edge_torch/generative/examples/phi/phi3.py,sha256=7pwHStGEPOuO1DnWiiavioRQXskMqbJMv3ctFNFuBU0,7075
+ai_edge_torch/generative/examples/phi/phi3.py,sha256=ddo52Inl5ub81q460cEyKhnsC3txellRErut-_qtBbM,6949
+ai_edge_torch/generative/examples/phi/phi4.py,sha256=OkMwLGe8l2JEAgOFi19AdbNBl1xp1djZBZo8MJP58ho,5732
 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/phi/verify_phi4.py,sha256=BoCa5kUBRHtMQ-5ql6yD4pG4xHJMyUiQlpMOWVx-JgY,2356
 ai_edge_torch/generative/examples/qwen/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/qwen/convert_to_tflite.py,sha256=tqvXVGNdDehdak9-5DDisACs9VlTwr8eFwcjQ_kZxgc,2776
 ai_edge_torch/generative/examples/qwen/qwen.py,sha256=Zi_qiQ1JPokXZ95jgSEnQp3F-LKzFCvWvFLKhJjnASo,4199
@@ -147,8 +150,8 @@ ai_edge_torch/generative/layers/normalization.py,sha256=MbwH-n80Fob5YvjBzdqDjBiz
 ai_edge_torch/generative/layers/rotary_position_embedding.py,sha256=975zR202MdIrILJ7blceAcxrNqX1ZCN0ECKG1gz-bV8,2655
 ai_edge_torch/generative/layers/scaled_dot_product_attention.py,sha256=vp8dVx6tOe99neJhpbrtIt5fvN5NFw19JVH1v0yi5Mg,4154
 ai_edge_torch/generative/layers/experimental/__init__.py,sha256=nz-K0h8DfiATHzR6s1_bCw2akUmHWffU1bDRSkIzSqI,592
-ai_edge_torch/generative/layers/experimental/attention.py,sha256=KC1UkIhaPx2DNRfkxCXO7eZZMeNm2UxkjFi-fB8HVhw,9212
-ai_edge_torch/generative/layers/experimental/kv_cache.py,sha256=gE_q8YoSzOhGgbSm0K91jXkbFKnFJpuYf-hxMzLNw78,8976
+ai_edge_torch/generative/layers/experimental/attention.py,sha256=95djjlJItDVuSNE3BL0b6u3lQoIhmmdvaik7qBBvQA0,8909
+ai_edge_torch/generative/layers/experimental/kv_cache.py,sha256=VN4gn4ylaVOwaTR5EXKv0YTVgpQ850bmjGLCgCCI1ps,9267
 ai_edge_torch/generative/layers/experimental/scaled_dot_product_attention.py,sha256=1vMh1L3uYX4ptKQMWcAjxkL1v2-g0jmOiuai8ydp0dc,2879
 ai_edge_torch/generative/layers/experimental/types.py,sha256=bPPxw6TOCZVWdeDP3vCbOnjNP5-bdUMmfsfO-EtdazQ,2847
 ai_edge_torch/generative/layers/unet/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
@@ -168,7 +171,7 @@ ai_edge_torch/generative/test/test_kv_cache.py,sha256=2AulHBS3hC4b_68PNNBkRVOryp
 ai_edge_torch/generative/test/test_loader.py,sha256=9mQUeeZKOVApOWSWl2cN9c10axZjMKM1-0Zd823CCS4,3449
 ai_edge_torch/generative/test/test_lora.py,sha256=6QIM6RLTc2HrodGpp_aS3OxM9Rco2KAzEnYgotkg41M,5310
 ai_edge_torch/generative/test/test_model_conversion.py,sha256=jfqkECCX7XKHeBAuDXrkwQJf0vM72eG3LMc5rluha84,6191
-ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=bXJwDxSPgxVKp-_6BsEmMA3TuMUaUNiZoYomNounxco,14416
+ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=-v2Vj7Qdd3GyBn4k7BWVgyGzrbcL30Su3nxZYLtwkCs,14787
 ai_edge_torch/generative/test/test_quantize.py,sha256=bEJMhpQ9bIDUZVBXTW888728FcH-i3SyE4JSZZUgU0A,6071
 ai_edge_torch/generative/test/utils.py,sha256=tF6aCfAGJnc9dmzCnZCEOuKNVimfWOqscv9og0DDLHU,2656
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
@@ -230,8 +233,8 @@ 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-0.4.0.dev20250227.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.4.0.dev20250227.dist-info/METADATA,sha256=cHcz3adq1WwVddazAJ06h7SKITJm70eMpFVjoNa2Jw4,1966
-ai_edge_torch_nightly-0.4.0.dev20250227.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-ai_edge_torch_nightly-0.4.0.dev20250227.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.4.0.dev20250227.dist-info/RECORD,,
+ai_edge_torch_nightly-0.4.0.dev20250301.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.4.0.dev20250301.dist-info/METADATA,sha256=VbeGOSHuc6HIM269rYt6xGOlKC_Pr6_EDGFlCVXa7qg,1966
+ai_edge_torch_nightly-0.4.0.dev20250301.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+ai_edge_torch_nightly-0.4.0.dev20250301.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.4.0.dev20250301.dist-info/RECORD,,