ai-edge-torch-nightly 0.3.0.dev20240813__py3-none-any.whl → 0.3.0.dev20240817__py3-none-any.whl

ai_edge_torch/generative/layers/model_config.py

@@ -16,7 +16,7 @@
 from dataclasses import dataclass
 from dataclasses import field
 import enum
-from typing import Optional
+from typing import Optional, Sequence
 
 
 @enum.unique
@@ -53,6 +53,11 @@ class FeedForwardType(enum.Enum):
   GATED = enum.auto()
 
 
+class AttentionType(enum.Enum):
+  GLOBAL = enum.auto()
+  LOCAL_SLIDING = enum.auto()
+
+
 @dataclass
 class AttentionConfig:
   """Attention model's parameters."""
@@ -78,6 +83,12 @@ class AttentionConfig:
   enable_kv_cache: bool = True
   relative_attention_num_buckets: int = 0
   relative_attention_max_distance: int = 0
+  # Softcap on the output logits.
+  logit_softcap: Optional[float] = None
+  # The types of attention used in the layers of the model.
+  attn_types: Optional[Sequence[AttentionType]] = None
+  # The size of the sliding window used for local attention.
+  sliding_window_size: Optional[int] = None
 
 
 @dataclass
@@ -88,16 +99,6 @@ class ActivationConfig:
   dim_out: Optional[int] = None
 
 
-@dataclass
-class FeedForwardConfig:
-  """FeedForward module's parameters."""
-
-  type: FeedForwardType
-  activation: ActivationConfig
-  intermediate_size: int
-  use_bias: bool = False
-
-
 @dataclass
 class NormalizationConfig:
   """Normalizater parameters."""
@@ -109,6 +110,24 @@ class NormalizationConfig:
   group_num: Optional[float] = None
 
 
+@dataclass
+class FeedForwardConfig:
+  """FeedForward module's parameters."""
+
+  type: FeedForwardType
+  activation: ActivationConfig
+  intermediate_size: int
+  use_bias: bool = False
+  # The normalization applied to feed forward's input.
+  pre_ff_norm_config: NormalizationConfig = field(
+      default_factory=NormalizationConfig
+  )
+  # The normalization applied to feed forward's output.
+  post_ff_norm_config: NormalizationConfig = field(
+      default_factory=NormalizationConfig
+  )
+
+
 @dataclass
 class ModelConfig:
   """Base configurations for building a transformer architecture."""
@@ -124,8 +143,8 @@ class ModelConfig:
   pre_attention_norm_config: NormalizationConfig = field(
       default_factory=NormalizationConfig
   )
-  # The normalization applied to feed forward's input.
-  pre_ff_norm_config: NormalizationConfig = field(
+  # The normalization applied to attentions's output.
+  post_attention_norm_config: NormalizationConfig = field(
       default_factory=NormalizationConfig
   )
   # The normalization applied before LM head.
@@ -151,6 +170,9 @@ class ModelConfig:
   # Default batch size of the exported model. Default value is 1.
   batch_size: int = 1
 
+  # Softcap on the model output logits.
+  final_logit_softcap: Optional[float] = None
+
   @property
   def kv_cache_max(self) -> int:
     if self.kv_cache_max_len > 0:

ai_edge_torch/generative/layers/scaled_dot_product_attention.py

@@ -29,6 +29,7 @@ def scaled_dot_product_attention(
     head_size: int,
     mask: Optional[torch.Tensor] = None,
     scale: Optional[float] = None,
+    softcap: Optional[float] = None,
 ):
   """Scaled dot product attention.
 
@@ -53,15 +54,26 @@ def scaled_dot_product_attention(
     # Handle the GQA case, where q.shape[1] % k.shape[1] == 0.
     k = k.repeat_interleave(q.shape[1] // k.shape[1], dim=1)
     v = v.repeat_interleave(q.shape[1] // v.shape[1], dim=1)
-  y = F.scaled_dot_product_attention(
-      q,
-      k,
-      v,
-      attn_mask=mask,
-      dropout_p=0.0,
-      is_causal=mask is None,
-      scale=scale,
-  )
+  if softcap is None:
+    y = F.scaled_dot_product_attention(
+        q,
+        k,
+        v,
+        attn_mask=mask,
+        dropout_p=0.0,
+        is_causal=mask is None,
+        scale=scale,
+    )
+  else:
+    q.mul_(scale)
+    scores = q @ k.transpose(-1, -2)
+    scores = scores / softcap
+    scores = torch.tanh(scores)
+    scores = scores * softcap
+    scores = scores + mask
+    out = F.softmax(scores.float(), dim=-1).type_as(q)
+    y = torch.matmul(out, v)
+
   return y.transpose(1, 2)
 
 
@@ -72,6 +84,7 @@ def scaled_dot_product_attention_with_hlfb(
     head_size: int,
     mask: Optional[torch.Tensor] = None,
     scale: Optional[float] = None,
+    softcap: Optional[float] = None,
 ):
   """Scaled dot product attention with high-level function boundary enabled.
 
@@ -86,6 +99,9 @@ def scaled_dot_product_attention_with_hlfb(
     The output tensor of scaled_dot_product_attention.
   """
 
+  if softcap is not None:
+    raise NotImplementedError("SDPA with HLFB not available with softcap.")
+
   if scale is None:
     scale = 1.0 / math.sqrt(head_size)
 

ai_edge_torch/generative/test/test_model_conversion.py

@@ -16,7 +16,7 @@
 import copy
 
 import ai_edge_torch
-from ai_edge_torch.generative.examples.gemma import gemma
+from ai_edge_torch.generative.examples.gemma import gemma, gemma2
 from ai_edge_torch.generative.examples.phi2 import phi2
 from ai_edge_torch.generative.examples.test_models import toy_model_with_kv_cache  # NOQA
 from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
@@ -202,6 +202,34 @@ class TestModelConversion(googletest.TestCase):
           )
       )
 
+  def test_gemma2(self):
+    self.skipTest("b/338288901")
+    config = gemma2.get_fake_model_config_2b_for_test()
+    model = gemma2.Gemma2(config)
+    model.eval()
+
+    idx = torch.from_numpy(np.array([[1, 2, 3, 4]]))
+    tokens = torch.full((1, 10), 0, dtype=torch.long, device="cpu")
+    tokens[0, :4] = idx
+    input_pos = torch.arange(0, 10)
+
+    edge_model = ai_edge_torch.convert(model, (tokens, input_pos))
+
+    # TODO: b/338288901 - re-enable test to check output tensors.
+    skip_output_check = True
+    if not skip_output_check:
+      # TODO(talumbau, haoliang): debug numerical diff.
+      self.assertTrue(
+          model_coverage.compare_tflite_torch(
+              edge_model,
+              model,
+              (tokens, input_pos),
+              num_valid_inputs=1,
+              atol=1e-2,
+              rtol=1e-5,
+          )
+      )
+
   def test_phi2(self):
     self.skipTest("b/338288901")
     config = phi2.get_fake_model_config_for_test()

ai_edge_torch/generative/utilities/loader.py

@@ -107,7 +107,9 @@ class ModelLoader:
     ff_gate_proj: str = None
 
     pre_attn_norm: str = None
+    post_attn_norm: str = None
     pre_ff_norm: str = None
+    post_ff_norm: str = None
     embedding: str = None
     embedding_position: str = None
     final_norm: str = None
@@ -258,6 +260,26 @@ class ModelLoader:
             f"{ff_gate_proj_name}.bias"
         )
 
+    if self._names.pre_ff_norm is not None:
+      pre_ff_norm_name = self._names.pre_ff_norm.format(idx)
+      converted_state[f"{prefix}.ff.pre_ff_norm.weight"] = state.pop(
+          f"{pre_ff_norm_name}.weight"
+      )
+      if f"{pre_ff_norm_name}.bias" in state:
+        converted_state[f"{prefix}.ff.pre_ff_norm.bias"] = state.pop(
+            f"{pre_ff_norm_name}.bias"
+        )
+
+    if self._names.post_ff_norm is not None:
+      post_ff_norm_name = self._names.post_ff_norm.format(idx)
+      converted_state[f"{prefix}.ff.post_ff_norm.weight"] = state.pop(
+          f"{post_ff_norm_name}.weight"
+      )
+      if f"{post_ff_norm_name}.bias" in state:
+        converted_state[f"{prefix}.ff.post_ff_norm.bias"] = state.pop(
+            f"{post_ff_norm_name}.bias"
+        )
+
   def _map_attention(
       self,
       idx: int,
@@ -325,14 +347,14 @@ class ModelLoader:
             f"{pre_attn_norm_name}.bias"
         )
 
-    if self._names.pre_ff_norm is not None:
-      pre_ff_norm_name = self._names.pre_ff_norm.format(idx)
-      converted_state[f"{prefix}.pre_ff_norm.weight"] = state.pop(
-          f"{pre_ff_norm_name}.weight"
+    if self._names.post_attn_norm is not None:
+      post_attn_norm_name = self._names.post_attn_norm.format(idx)
+      converted_state[f"{prefix}.post_atten_norm.weight"] = state.pop(
+          f"{post_attn_norm_name}.weight"
       )
-      if f"{pre_ff_norm_name}.bias" in state:
-        converted_state[f"{prefix}.pre_ff_norm.bias"] = state.pop(
-            f"{pre_ff_norm_name}.bias"
+      if f"{post_attn_norm_name}.bias" in state:
+        converted_state[f"{prefix}.post_atten_norm.bias"] = state.pop(
+            f"{post_attn_norm_name}.bias"
         )
 
   def _fuse_qkv(

ai_edge_torch/generative/utilities/t5_loader.py

@@ -113,7 +113,7 @@ class ModelLoader:
 
     pre_attn_norm: str = None
     pre_cross_attn_norm: str = None
-    pre_ff_norm: str = None
+    post_attn_norm: str = None
     embedding: str = None
     final_norm: str = None
     lm_head: str = None
@@ -484,14 +484,14 @@ class ModelLoader:
             state.pop(f"{pre_cross_attn_norm_name}.bias")
         )
 
-    if names.pre_ff_norm is not None:
-      pre_ff_norm_name = names.pre_ff_norm.format(idx)
-      converted_state[f"{prefix}.pre_ff_norm.weight"] = state.pop(
-          f"{pre_ff_norm_name}.weight"
+    if names.post_attn_norm is not None:
+      post_attn_norm_name = names.post_attn_norm.format(idx)
+      converted_state[f"{prefix}.post_atten_norm.weight"] = state.pop(
+          f"{post_attn_norm_name}.weight"
       )
-      if f"{pre_ff_norm_name}.bias" in state:
-        converted_state[f"{prefix}.pre_ff_norm.bias"] = state.pop(
-            f"{pre_ff_norm_name}.bias"
+      if f"{post_attn_norm_name}.bias" in state:
+        converted_state[f"{prefix}.post_atten_norm.bias"] = state.pop(
+            f"{post_attn_norm_name}.bias"
         )
 
   def _fuse_qkv(

ai_edge_torch/hlfb/test/test_mark_pattern.py

@@ -51,7 +51,12 @@ class TestMarkPattern(googletest.TestCase):
     mark_pattern.mark_pattern(exported_program.graph_module, pattern)
     mlir = _export_stablehlo_mlir(exported_program)
 
-    self.assertEqual(mlir.count('stablehlo.composite "test.add"'), 2)
+    lowertools.assert_string_count(
+        self,
+        mlir,
+        {'stablehlo.composite "test.add"': 2},
+        {"stablehlo.custom_call @mark_tensor": 6},
+    )
 
   def test_mark_pattern_with_attr_builder(self):
     class TestModel(torch.nn.Module):
@@ -72,9 +77,15 @@ class TestMarkPattern(googletest.TestCase):
     mark_pattern.mark_pattern(exported_program.graph_module, pattern)
     mlir = _export_stablehlo_mlir(exported_program)
 
-    self.assertEqual(mlir.count('stablehlo.composite "test.add"'), 2)
-    self.assertEqual(
-        mlir.count('composite_attributes = {alias = "test.test_add"}'), 2
+    lowertools.assert_string_count(
+        self,
+        mlir,
+        {
+            'stablehlo.composite "test.add"': 2,
+            'composite_attributes = {alias = "test.test_add"}': 2,
+        },
+        {"stablehlo.custom_call @mark_tensor": 6},
+        {'{"alias": "test.test_add"}': 2},
     )
 
   def test_mark_pattern_with_scalar_attr_tracker(self):
@@ -104,9 +115,17 @@ class TestMarkPattern(googletest.TestCase):
     mark_pattern.mark_pattern(exported_program.graph_module, pattern)
     mlir = _export_stablehlo_mlir(exported_program)
 
-    self.assertEqual(mlir.count('stablehlo.composite "test.log_softmax"'), 5)
-    self.assertEqual(mlir.count("composite_attributes = {dim = 0 : i64}"), 3)
-    self.assertEqual(mlir.count("composite_attributes = {dim = 1 : i64}"), 2)
+    lowertools.assert_string_count(
+        self,
+        mlir,
+        {
+            'stablehlo.composite "test.log_softmax"': 5,
+            "composite_attributes = {dim = 0 : i64}": 3,
+            "composite_attributes = {dim = 1 : i64}": 2,
+        },
+        {"stablehlo.custom_call @mark_tensor": 10},
+        {'{"dim": 0}': 3, '{"dim": 1}': 2},
+    )
 
   def test_mark_tangent_model_and_pattern_input(self):
     class TestModel(torch.nn.Module):
@@ -128,7 +147,12 @@ class TestMarkPattern(googletest.TestCase):
     mark_pattern.mark_pattern(exported_program.graph_module, pattern)
     mlir = _export_stablehlo_mlir(exported_program)
 
-    self.assertEqual(mlir.count('stablehlo.composite "test.relu'), 1)
+    lowertools.assert_string_count(
+        self,
+        mlir,
+        {'stablehlo.composite "test.relu"': 1},
+        {"stablehlo.custom_call @mark_tensor": 2},
+    )
 
 
 if __name__ == "__main__":

ai_edge_torch/hlfb/test/test_stablehlo_composite_builder.py

@@ -16,6 +16,7 @@
 
 import math
 
+from ai_edge_torch import config
 from ai_edge_torch import lowertools
 from ai_edge_torch.hlfb import StableHLOCompositeBuilder
 import torch
@@ -29,6 +30,10 @@ def _export_stablehlo_mlir(model, args):
   return lowertools.exported_program_to_mlir_text(ep)
 
 
+@googletest.skipIf(
+    not config.Config.use_torch_xla,
+    reason="The odml_torch counter part is in odml_torch.",
+)
 class TestStableHLOCompositeBuilder(googletest.TestCase):
 
   def test_build_composite(self):

ai_edge_torch/lowertools/__init__.py

@@ -14,3 +14,4 @@
 # ==============================================================================
 
 from ._shim import *
+from .test_utils import *

ai_edge_torch/lowertools/odml_torch_utils.py

@@ -28,6 +28,7 @@ import tensorflow as tf
 import torch
 
 from tensorflow.compiler.tf2xla.python import xla as tfxla
+from tensorflow.lite.python import conversion_metadata_schema_py_generated as conversion_metadata_fb
 
 MlirBundle = odml_torch.export.MlirLowered
 
@@ -162,7 +163,9 @@ def merged_bundle_to_tfl_model(
     )
 
     converter = tf.lite.TFLiteConverter.from_saved_model(temp_dir_path)
+    converter._set_original_model_type(conversion_metadata_fb.ModelType.PYTORCH)
     converter._experimental_enable_composite_direct_lowering = True
+    converter.model_origin_framework = "PYTORCH"
 
     conversion_utils.apply_tfl_converter_flags(converter, _tfl_converter_flags)
 

ai_edge_torch/lowertools/test_utils.py

@@ -0,0 +1,60 @@
+# 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.
+# ==============================================================================
+
+import re
+from typing import Optional
+from ai_edge_torch import config
+from tensorflow.python.platform import googletest
+
+
+def _extract_backend_configs(mlir):
+  mlir = mlir.replace("\\22", '"')
+  configs = []
+  for match in re.finditer(r"backend_config\s*=\s*\"(\{.*\})\"", mlir):
+    configs.append(match.group(1))
+  return "\n".join(configs)
+
+
+def assert_string_count(
+    test_case: googletest.TestCase,
+    mlir: str,
+    torch_xla_pattern_counter: dict[str, int],
+    odml_torch_pattern_counter: dict[str, int],
+    odml_torch_attr_counter: Optional[dict[str, int]] = None,
+):
+
+  if odml_torch_attr_counter is None:
+    odml_torch_attr_counter = {}
+
+  if config.Config.use_torch_xla:
+    for key in torch_xla_pattern_counter:
+      test_case.assertEqual(
+          mlir.count(key),
+          torch_xla_pattern_counter[key],
+      )
+  else:
+    for key in odml_torch_pattern_counter:
+      test_case.assertEqual(
+          mlir.count(key),
+          odml_torch_pattern_counter[key],
+      )
+    backend_configs = _extract_backend_configs(mlir)
+    print("backend_configs:")
+    print(backend_configs)
+    for key in odml_torch_attr_counter:
+      test_case.assertEqual(
+          backend_configs.count(key),
+          odml_torch_attr_counter[key],
+      )

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.3.0.dev20240813"
+__version__ = "0.3.0.dev20240817"

{ai_edge_torch_nightly-0.3.0.dev20240813.dist-info → ai_edge_torch_nightly-0.3.0.dev20240817.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20240813
+Version: 0.3.0.dev20240817
 Summary: Supporting PyTorch models with the Google AI Edge TFLite runtime.
 Home-page: https://github.com/google-ai-edge/ai-edge-torch
 Keywords: On-Device ML,AI,Google,TFLite,PyTorch,LLMs,GenAI

{ai_edge_torch_nightly-0.3.0.dev20240813.dist-info → ai_edge_torch_nightly-0.3.0.dev20240817.dist-info}/RECORD

@@ -2,7 +2,7 @@ ai_edge_torch/__init__.py,sha256=48qP37uHT90YPs4eIUQxCiWVwqGEX3idCUs6mQKvX1U,116
 ai_edge_torch/config.py,sha256=PCd9PVrbUNeVIUDFUCnW4goDWU4bjouK28yMYU6VOi0,877
 ai_edge_torch/conftest.py,sha256=r0GTrhMRhlmOGrrkvumHN8hkmyug6WvF60vWq8wRIBI,758
 ai_edge_torch/model.py,sha256=5DYNpFVwvI1w0JbAC1hn83NJVGS1WPX7n742419PMqs,4558
-ai_edge_torch/version.py,sha256=C9Lsgh_kXnELi8xLPUgnmDTLFOKW5S5z6lXAVwLMypU,706
+ai_edge_torch/version.py,sha256=C4yfJq9TbtZBH5gwhPSUtBgiIe04GkxvCq5TImNopww,706
 ai_edge_torch/_convert/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/_convert/conversion.py,sha256=kcv_QgNgeyDmrqwdzHicGNP68w6zF7GJg7YkMEIXp4Q,3759
 ai_edge_torch/_convert/conversion_utils.py,sha256=Sr8qXVcTwc-ZnZmK7yxVrIOOp1S_vNrwzC0zUvLTI2o,2160
@@ -42,22 +42,24 @@ ai_edge_torch/generative/examples/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQe
 ai_edge_torch/generative/examples/experimental/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/experimental/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/experimental/gemma/convert_to_tflite.py,sha256=lpiPFSh3SJd6WwuZ0QegSva3__iSz2tUD7L7QfkAe4I,3085
-ai_edge_torch/generative/examples/experimental/gemma/gemma.py,sha256=QoFbUUCTJrW1IYZg0vfb2-K-X0q1-NJFbWNGPQGwBgk,6688
+ai_edge_torch/generative/examples/experimental/gemma/gemma.py,sha256=8313wSsddvuxZ5ZYVdaITBV2FF1k22dcCujnq0UZvKs,6699
 ai_edge_torch/generative/examples/experimental/phi/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/experimental/phi/convert_to_tflite.py,sha256=DavrdGmqUgoThsGNRv3LXMW5tvJdYEvj66Hf1XRqkXU,3055
 ai_edge_torch/generative/examples/experimental/phi/phi2.py,sha256=u-VJX5mjzQKspXtAhNi53LCITtag-3nCaRTKdk5Z1sc,6231
 ai_edge_torch/generative/examples/experimental/tiny_llama/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/experimental/tiny_llama/convert_to_tflite.py,sha256=xPVvHQjLJHFiRv_-Fy2sDm0Aft7SG8SXiV6o3rF03cQ,3108
-ai_edge_torch/generative/examples/experimental/tiny_llama/tiny_llama.py,sha256=GOLLd9yCBnlNXeW7xrVy1wjOltcTbRdSpiJycbMj8TA,6372
+ai_edge_torch/generative/examples/experimental/tiny_llama/tiny_llama.py,sha256=zQYtyk3xYdiRAnzMKN58Q_wgTQFnDujxp6L4RFQjiD4,6383
 ai_edge_torch/generative/examples/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/gemma/convert_gemma2_to_tflite.py,sha256=pseJExH35lSAK0ZtzSHB1sFtRtF_EuT2xcSpGU0gKVI,2524
 ai_edge_torch/generative/examples/gemma/convert_to_tflite.py,sha256=w589IJETATd6Z9_1XCIWbrlCV3E92X_5ac3VVCVFXG0,2522
-ai_edge_torch/generative/examples/gemma/gemma.py,sha256=5Dn9JgJiXN-hWGQj9YqCr8Iik8mh5s0dX0VfyY8KDDo,6236
+ai_edge_torch/generative/examples/gemma/gemma.py,sha256=cCki-0cKvmGxK4Md6dRNdPDWZUyhkJUI854OCTFf3h0,6262
+ai_edge_torch/generative/examples/gemma/gemma2.py,sha256=j-zxJ-JNRnQ_kDzUESmsyy_a_4IxWZ510HmIImc0LDc,8240
 ai_edge_torch/generative/examples/phi2/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/phi2/convert_to_tflite.py,sha256=ON6zLO-nFS8eJ2yhyWzT5x2Somr-Ca-VjpjT7OGFU10,2506
 ai_edge_torch/generative/examples/phi2/phi2.py,sha256=C_kFYsPrEQ9GJCnc6h-jh8B5qQryvEpI6O6t4FBxg1I,5858
 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=VR09iAnj1e-sr-oam2rh24Wnb_JdZZQvpJIjylfgnS8,4468
+ai_edge_torch/generative/examples/stable_diffusion/clip.py,sha256=0WniBWQ6_NcQc5WycX3YRRX7Os9AGQSxfc1m2HKBqg8,4479
 ai_edge_torch/generative/examples/stable_diffusion/convert_to_tflite.py,sha256=7ra36nM5tQwSw-vi6QCFLx5IssZhT-6yVK4H3XsAc4w,5044
 ai_edge_torch/generative/examples/stable_diffusion/decoder.py,sha256=slieF2-QcDCwd4DRZ7snsZIphT97IXpp4plRRsRSwL8,13983
 ai_edge_torch/generative/examples/stable_diffusion/diffusion.py,sha256=7oUIJ6HO0vmlhFdkXpqGm9KTB-eM4Ob9VrHSDlIGFOg,30926
@@ -72,27 +74,27 @@ ai_edge_torch/generative/examples/stable_diffusion/samplers/k_lms.py,sha256=ZE6H
 ai_edge_torch/generative/examples/stable_diffusion/samplers/sampler.py,sha256=RxR5rw0wFFm_5CfAY-3-EIz83vhM9EKye8Bb5zBb0Ok,1341
 ai_edge_torch/generative/examples/t5/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/t5/convert_to_tflite.py,sha256=CZVuNEL8OHPkdsz70WOvNpTJ9LFkiDnlwgJiXfUZCVk,4548
-ai_edge_torch/generative/examples/t5/t5.py,sha256=6Rkisv7UI2w5KV8ogPPzeIiPWYwDLfFfSIncqD7Eenc,20854
-ai_edge_torch/generative/examples/t5/t5_attention.py,sha256=gp7DV8pv4FwICQhYlUYfYZ7BE5jzDIsD_V3a_4-T4Ds,8492
+ai_edge_torch/generative/examples/t5/t5.py,sha256=Zobw5BV-PC0nlU9Z6fzb2O07rMeU8vGIk-KtKp9D_H0,20871
+ai_edge_torch/generative/examples/t5/t5_attention.py,sha256=1lvbSlzyBwmd5Bs7-Up_v4iJQkCPIJx2RmMkLgy7l2Q,8508
 ai_edge_torch/generative/examples/test_models/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/test_models/toy_model.py,sha256=DhxOrIKe-tilBjbh1q4MsmCmmKMc4c1BPUzhnaJDD6M,3955
-ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py,sha256=bW0QB-_h9cfwAQf11AxFxOBq3HrEep_UlpBjXz3JSew,5801
-ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=CRja_CT0_eaH16rSDxwHKJS_CGUJMW0Fxd4r45Ii8Uo,4833
+ai_edge_torch/generative/examples/test_models/toy_model.py,sha256=LfWO_gSr1f66V1pxAc6yh21mtaJs7TVeuO9748zXBnE,3963
+ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py,sha256=l9swUKTcDtnTibNSNExaMgLvDeJ4Er2tVh5ZW1EtRgk,5809
+ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=mQkcpSe6HlRLMkIRCEHc9ZXL7jxEp9RWSGUQjjd-r2w,4841
 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=CLRqO7ycMbpy7J3_Czp1sLx6hcdwGD9zVq04yRba0e8,2550
-ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=nu3Il8Vxe7JwM8-AnGNXoGoZ9eVXKHMYEAqVEP-gwe8,5929
+ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=mXXFYJfo8yegSOFOndCR0oYxFPchYb9vTJ4ThXGIFLU,5940
 ai_edge_torch/generative/fx_passes/__init__.py,sha256=fmNNXawJ722M4cTUuTx289rT0NHxBEsOy_k8baqCOms,1173
 ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py,sha256=sXis0U4u-RoIp_NyrmWJNnqFqpqRuZOrhfsJIO6rMps,2028
 ai_edge_torch/generative/layers/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/layers/attention.py,sha256=xq10Gw4GudK4M2eY8-H4fi3qmpmZCfE-CziAXDZvqiQ,12177
-ai_edge_torch/generative/layers/attention_utils.py,sha256=2hzBVZvWCqqLfI-f3RJA1hi6T8cuaIJBPt8cdjQCA5s,6420
-ai_edge_torch/generative/layers/builder.py,sha256=JvPmwrG8_M4-kO2MM6sDZhpS32Wx3wVVhlVO4yPJKJ0,4161
-ai_edge_torch/generative/layers/feed_forward.py,sha256=RukSYr9h_DehcYVZWLS_rfCTY73Uj__pTRUatjxJtv8,2788
+ai_edge_torch/generative/layers/attention.py,sha256=2UujQePRJ1LK02PN-hGcuMu0ooCJC6ETfPvzEYVFyho,12284
+ai_edge_torch/generative/layers/attention_utils.py,sha256=68GXGR2HSWBFViTxX7cHifzVG-kcLS2IL2tQJPIpupg,7344
+ai_edge_torch/generative/layers/builder.py,sha256=xb7rjADv3Jm4qfmlYtg6oLLe7ReDE9UjsEqiejPpDD8,4346
+ai_edge_torch/generative/layers/feed_forward.py,sha256=uto7xtwx6jPkk1GZ2x7pSTentQzRrPSKw4_PSE12ahA,3525
 ai_edge_torch/generative/layers/kv_cache.py,sha256=Ob8QeXWW5xt-6hcGA0uoC48eRQ8lfvKca8JbWtFx2CE,3082
-ai_edge_torch/generative/layers/model_config.py,sha256=CTvKFwsBR3Rc-Kf73NA7k0799m1WnEvaEBKCnnfNkyo,4961
+ai_edge_torch/generative/layers/model_config.py,sha256=WpZ9djUBAZddyeSODHDaVMG37EQqfzGGrlMPi8AA-Hc,5752
 ai_edge_torch/generative/layers/normalization.py,sha256=u8lv0p-ktKcRqCDlOqZQa9WQcfDK9JM2IaUQFQdn7xs,1860
 ai_edge_torch/generative/layers/rotary_position_embedding.py,sha256=CZqOoibLcHvUgrgaIIWAlmk3XgE2inzx340MN-npLoU,1347
-ai_edge_torch/generative/layers/scaled_dot_product_attention.py,sha256=6WMe-A5KSSujQcZ34hIeSnnor3AXrw10cQ5FKy-30IU,3390
+ai_edge_torch/generative/layers/scaled_dot_product_attention.py,sha256=x2bOmrTgOISXcb06IDP7X3xgftpPpxOjBXw_OxTMVns,3874
 ai_edge_torch/generative/layers/unet/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/layers/unet/blocks_2d.py,sha256=4a0wh0co8Avz1wvxS3XqsgrgL5G-X1GSARI5Rj3L-xg,26995
 ai_edge_torch/generative/layers/unet/builder.py,sha256=zAqWXdimmMrQRhmE_t9XkS68mh6PSrzwb-2NZZXrR5I,1901
@@ -109,23 +111,24 @@ ai_edge_torch/generative/quantize/ai_edge_quantizer_glue/translate_recipe.py,sha
 ai_edge_torch/generative/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/test/test_experimental_ekv.py,sha256=T5-O2RVLJTH7v9w1_uBfp-Y7o3sdGzYq2Tj2wLRNHyI,4357
 ai_edge_torch/generative/test/test_loader.py,sha256=1ZqAq0HY5uIioumsReOVIsbGBx0WkYcl18PvttdJKrk,3381
-ai_edge_torch/generative/test/test_model_conversion.py,sha256=4RTB1oPA2eWPyuof2-ZB1BxVKzKy5Q9vCux7psmV6zc,7615
+ai_edge_torch/generative/test/test_model_conversion.py,sha256=52ciFy_Qol2Xuym6P6EqdL29oai35LSWGvsUwyEdFTo,8477
 ai_edge_torch/generative/test/test_quantize.py,sha256=3SmJm7Kq98gAneU6IGwwJrJYCVH1qwWR6oUxPfb6qiI,5346
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
-ai_edge_torch/generative/utilities/loader.py,sha256=XfVRvwvZyQuofctxIedLNDKQrsy9UlRr4wpScZJLWcw,11779
+ai_edge_torch/generative/utilities/loader.py,sha256=bAWZ7FM4v_pPnX_AmEdGxHkDH65QdL-MjIP3PxscZmI,12649
 ai_edge_torch/generative/utilities/stable_diffusion_loader.py,sha256=pKp3AMSbS3otCvgwJRF5M1l4JRNKk-aCKimXzIMSrds,35679
-ai_edge_torch/generative/utilities/t5_loader.py,sha256=jz2qnDtH6oyxcqaBwEVfiiKmq_93LTDeUKNJ2cWpLwg,16856
+ai_edge_torch/generative/utilities/t5_loader.py,sha256=_UXcc1QKT-S92hikfo-fTBFhnYLzROqcyRqKonVsqj4,16885
 ai_edge_torch/hlfb/__init__.py,sha256=sH4um75na-O8tzxN6chFyp6Y4xnexsE7kUQpZySv6dE,735
 ai_edge_torch/hlfb/mark_pattern/__init__.py,sha256=cjTprggj_cuktSCm7-A25e7Shop3k63ylp7sdZmtZ8o,4790
 ai_edge_torch/hlfb/mark_pattern/passes.py,sha256=pjkKcI1nHECPluAt87cFBrt1DP0f3ge7rHq1NhCkBIE,1936
 ai_edge_torch/hlfb/mark_pattern/pattern.py,sha256=uiYRfzD1T8deCEAGfdAFusRbI41m14zeTt0Lz5lNT3M,9808
 ai_edge_torch/hlfb/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/hlfb/test/test_mark_pattern.py,sha256=oYB0RPW-tHOwW9gQFH9GtHKO_Mmh1lkoiemXmTfySqc,4383
-ai_edge_torch/hlfb/test/test_stablehlo_composite_builder.py,sha256=3vSX5E9ZFFhTPZZX6TMiAsGa_kaXABbN851bRbTFsC0,8297
-ai_edge_torch/lowertools/__init__.py,sha256=0M9TOR80sS5y6dikOsIFYx0P9IomqAdNIuYpgkP4PcI,693
+ai_edge_torch/hlfb/test/test_mark_pattern.py,sha256=e53YNSO2w7Sd9Y717jAr6WKjnXq34Tx_52hXRGtGs3A,4833
+ai_edge_torch/hlfb/test/test_stablehlo_composite_builder.py,sha256=7Qbba7GJCBc-J1TUwWIvrpBK0Hwza9nift7sKpW2YVE,8449
+ai_edge_torch/lowertools/__init__.py,sha256=uKGibEN7n4Tqbe0HiXOEEXWmPL9AUmh34xaYA9yx2sg,719
 ai_edge_torch/lowertools/_shim.py,sha256=ilL7x1ebUBj1clg7bagrX4y_nVSHiGrvDrOVfuTeenE,3039
 ai_edge_torch/lowertools/common_utils.py,sha256=emClsZ_MBlbLG_0BBtyLpkdz4dMWp6SyrNioygRBylk,2973
-ai_edge_torch/lowertools/odml_torch_utils.py,sha256=5XVI2ovptp1wsDZcyyaZDgT4oUa1McOiE-PrKhXNhFo,6316
+ai_edge_torch/lowertools/odml_torch_utils.py,sha256=32cak8uiXFIVdkaYFhIW1fWG4NzLrYq-w8xK0pNkhYc,6547
+ai_edge_torch/lowertools/test_utils.py,sha256=vsjaX3Ix2U1163jVUNSJgK9io2WNUtJjRvNFE9DrqF4,1932
 ai_edge_torch/lowertools/torch_xla_utils.py,sha256=-g0NldtVOTCQtX3V2XEjuCQO_I52nSNQlu0r_rIS2IE,8635
 ai_edge_torch/quantize/__init__.py,sha256=aB5dXot04bqyUhpsDFvxt9CIi15QAC4euvqOndJ0XLU,714
 ai_edge_torch/quantize/pt2e_quantizer.py,sha256=CKIEhs9jCcna64qj1jFH9zEbMbRdyeGV_TmSqEBPjes,15741
@@ -134,8 +137,8 @@ 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.dev20240813.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.3.0.dev20240813.dist-info/METADATA,sha256=TMkI635DYqK0Fg6W6tZbg8ZTT54_9QkkCcd3XOxjyho,1885
-ai_edge_torch_nightly-0.3.0.dev20240813.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-ai_edge_torch_nightly-0.3.0.dev20240813.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.3.0.dev20240813.dist-info/RECORD,,
+ai_edge_torch_nightly-0.3.0.dev20240817.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20240817.dist-info/METADATA,sha256=GZgUf21m2RQYBvHxmeTujeniBxbbUTVpQQB9vjNSTaM,1885
+ai_edge_torch_nightly-0.3.0.dev20240817.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20240817.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20240817.dist-info/RECORD,,