ai-edge-torch-nightly 0.3.0.dev20240817__py3-none-any.whl → 0.3.0.dev20240823__py3-none-any.whl

ai_edge_torch/_convert/signature.py

@@ -16,6 +16,7 @@
 import dataclasses
 from typing import Any, Dict, List, Optional, Tuple, Union
 
+from ai_edge_torch import lowertools
 import torch
 import torch.utils._pytree as pytree
 
@@ -53,47 +54,12 @@ class Signature:
     for i in range(args_spec.num_leaves):
       names.append(f"args_{i}")
 
-    kwargs_names = self._flat_kwarg_names(
+    kwargs_names = lowertools.flat_dict_names(
         kwargs_spec.children_specs, kwargs_spec.context
     )
     names.extend(kwargs_names)
     return names
 
-  def _flat_kwarg_names(self, specs, context) -> List[str]:
-    flat_names = []
-    if context is None:
-      for i, spec in enumerate(specs):
-        if spec.children_specs:
-          flat_names.extend([
-              f"{i}_{name}"
-              for name in self._flat_kwarg_names(
-                  spec.children_specs, spec.context
-              )
-          ])
-        else:
-          flat_names.append(f"{i}")
-    else:
-      flat_ctx = self._flatten_list(context)
-      for prefix, spec in zip(flat_ctx, specs):
-        leaf_flat_names = self._flat_kwarg_names(
-            spec.children_specs, spec.context
-        )
-        if leaf_flat_names:
-          flat_names.extend([f"{prefix}_{name}" for name in leaf_flat_names])
-        else:
-          flat_names.append(prefix)
-
-    return flat_names
-
-  def _flatten_list(self, l: List) -> List:
-    flattened = []
-    for item in l:
-      if isinstance(item, list):
-        flattened.extend(self._flatten_list(item))
-      else:
-        flattened.append(item)
-    return flattened
-
   @property
   def flat_args(self) -> tuple[Any]:
     args, kwargs = self._normalized_sample_args_kwargs

ai_edge_torch/_convert/test/test_convert.py

@@ -174,7 +174,7 @@ class TestConvert(googletest.TestCase):
     self.assertTrue(result)
 
   def test_12_outputs_model(self):
-    """Tests conversion of a model that returns multiple outputs."""
+    """Tests conversion of a model that returns more than 10 outputs."""
 
     class BasicAddModelWithMultipleOutputs(torch.nn.Module):
       """A model that returns multiple outputs."""
@@ -421,6 +421,37 @@ class TestConvert(googletest.TestCase):
         SampleModel(), args, kwargs, flat_inputs
     )
 
+  def test_convert_model_non_flat_output_dict(self):
+    """Test converting a model with non-flat output structure."""
+
+    class SampleModel(torch.nn.Module):
+
+      def forward(self, x, y, z):
+        return {"x": x, "y": TestContainer1(data_1=y, data_2=[y, z])}
+
+    args = (torch.randn(10, 10), torch.randn(10, 10), torch.randn(10, 10))
+    kwargs = dict()
+    flat_inputs = {
+        "args_0": args[0].numpy(),
+        "args_1": args[1].numpy(),
+        "args_2": args[2].numpy(),
+    }
+
+    edge_model = ai_edge_torch.convert(SampleModel().eval(), args, kwargs)
+    edge_output = edge_model(**flat_inputs)
+    np.testing.assert_almost_equal(edge_output["x"], args[0])
+    np.testing.assert_almost_equal(edge_output["y_data_1"], args[1])
+    np.testing.assert_almost_equal(edge_output["y_data_2_0"], args[1])
+    np.testing.assert_almost_equal(edge_output["y_data_2_1"], args[2])
+
+    interpreter = tf.lite.Interpreter(model_content=edge_model._tflite_model)
+    runner = interpreter.get_signature_runner("serving_default")
+    output_details = runner.get_output_details()
+    self.assertIn("x", output_details.keys())
+    self.assertIn("y_data_1", output_details.keys())
+    self.assertIn("y_data_2_0", output_details.keys())
+    self.assertIn("y_data_2_1", output_details.keys())
+
   def _compare_tflite_torch_args_kwargs(self, model, args, kwargs, flat_inputs):
     model.eval()
     edge_model = ai_edge_torch.convert(model, args, kwargs)

ai_edge_torch/_convert/test/test_convert_multisig.py

@@ -94,8 +94,8 @@ class TestConvertMultiSignature(googletest.TestCase):
     signature_name = "large_input"
 
     edge_model = ai_edge_torch.signature(
-        signature_name, torch_module, args
-    ).convert(torch_module, large_args)
+        signature_name, torch_module, large_args
+    ).convert(torch_module, args)
 
     self.assertTrue(
         model_coverage.compare_tflite_torch(edge_model, torch_module, args)

ai_edge_torch/generative/examples/gemma/gemma2.py

@@ -203,7 +203,7 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       final_norm_config=norm_config,
       parallel_residual=False,
       lm_head_use_bias=False,
-      enable_hlfb=False,
+      enable_hlfb=True,
       final_logit_softcap=30.0,
   )
   return config
@@ -242,7 +242,6 @@ def define_and_run_2b() -> None:
   out = model.forward(tokens, input_pos)
   out_final = out[0, 8, :]
   assert torch.allclose(gemma2_goldens, out_final, atol=1e-04)
-  print(out)
 
 
 if __name__ == "__main__":

ai_edge_torch/generative/layers/scaled_dot_product_attention.py

@@ -99,14 +99,16 @@ 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)
 
+  attrs = {"scale": scale}
+
+  if softcap is not None:
+    attrs["logit_cap"] = softcap
+
   builder = StableHLOCompositeBuilder(
-      name="odml.scaled_dot_product_attention", attr={"scale": scale}
+      name="odml.scaled_dot_product_attention", attr=attrs
   )
   q, k, v, mask = builder.mark_inputs(q, k, v, mask)
 

ai_edge_torch/generative/utilities/loader.py

@@ -72,7 +72,7 @@ def load_pytorch_statedict(full_path: str):
   patterns = []
   if os.path.isdir(full_path):
     patterns.append(os.path.join(full_path, "*.bin"))
-    patterns.append(os.path.join(full_path, "*.pt"))
+    patterns.append(os.path.join(full_path, "*pt"))
   else:
     patterns.append(full_path)
   for pattern in patterns:
@@ -149,6 +149,7 @@ class ModelLoader:
           enabled.
     """
     state = self._loader(self._file_name)
+    state = state["model_state_dict"] if "model_state_dict" in state else state
     converted_state = dict()
     if self._names.embedding is not None:
       converted_state["tok_embedding.weight"] = state.pop(
@@ -200,7 +201,7 @@ class ModelLoader:
       if glob.glob(os.path.join(self._file_name, "*.safetensors")):
         return load_safetensors
       if glob.glob(os.path.join(self._file_name, "*.bin")) or glob.glob(
-          os.path.join(self._file_name, "*.pt")
+          os.path.join(self._file_name, "*pt")
       ):
         return load_pytorch_statedict
 

ai_edge_torch/lowertools/__init__.py

@@ -14,4 +14,5 @@
 # ==============================================================================
 
 from ._shim import *
+from .common_utils import flat_dict_names
 from .test_utils import *

ai_edge_torch/lowertools/common_utils.py

@@ -14,10 +14,63 @@
 # ==============================================================================
 
 import logging
+from typing import List
 
 from ai_edge_torch._convert import signature as signature_module
 import tensorflow as tf
 import torch
+import torch.utils._pytree as pytree
+
+
+def _flatten_list(l: List) -> List:
+  flattened = []
+  for item in l:
+    if isinstance(item, list):
+      flattened.extend(_flatten_list(item))
+    else:
+      flattened.append(item)
+  return flattened
+
+
+def flat_dict_names(
+    tree_spec: pytree.TreeSpec, context: pytree.Context
+) -> List[str]:
+  """Given a TreeSpec, this produces a list of names for the leaves.
+
+  The list of names embeddeds the structure of the tree_spec. A nesting level is
+  indicated by an `_` and elements in a list are indicated by `_<index>`.
+
+  TODO b/361601485: The flattening of names is not collision-free and needs to
+  be revised.
+
+  Args:
+    tree_spec: The TreeSpec to extract the names from.
+    context: The context used to check if the provided spec belongs to a
+      dictionary or a list.
+
+  Returns:
+    A list of flattened names.
+  """
+  flat_names = []
+  if context is None:
+    for i, spec in enumerate(tree_spec):
+      if spec.children_specs:
+        flat_names.extend([
+            f"{i}_{name}"
+            for name in flat_dict_names(spec.children_specs, spec.context)
+        ])
+      else:
+        flat_names.append(f"{i}")
+  else:
+    flat_ctx = _flatten_list(context)
+    for prefix, spec in zip(flat_ctx, tree_spec):
+      leaf_flat_names = flat_dict_names(spec.children_specs, spec.context)
+      if leaf_flat_names:
+        flat_names.extend([f"{prefix}_{name}" for name in leaf_flat_names])
+      else:
+        flat_names.append(prefix)
+
+  return flat_names
 
 
 def _torch_to_tf_variable(torch_tensor: torch.Tensor):

ai_edge_torch/lowertools/odml_torch_utils.py

@@ -38,6 +38,7 @@ class MergedBundle:
   """A bundle of MlirLowered that has been merged."""
 
   bundles: list[odml_torch.export.MlirLowered]
+  exported_programs: list[torch.export.ExportedProgram]
   deduped_tf_vars: list[tf.Variable]
 
 
@@ -74,12 +75,16 @@ def _extract_call_args(
   return call_args
 
 
-def _wrap_as_tf_func(bundle, tf_state_dict):
+def _wrap_as_tf_func(
+    bundle: export.MlirLowered,
+    tf_state_dict: Dict[str, tf.Variable],
+    exported_program: torch.export.ExportedProgram,
+):
   def inner(*args):
     t_outs = [torch_dtype_to_tf(sig.dtype) for sig in bundle.output_signature]
     s_outs = [_get_shape_with_dynamic(sig) for sig in bundle.output_signature]
     call_args = _extract_call_args(bundle, args, tf_state_dict)
-    return tfxla.call_module(
+    call_module_return = tfxla.call_module(
         tuple(call_args),
         version=5,
         Tout=t_outs,  # dtype information
@@ -87,6 +92,14 @@ def _wrap_as_tf_func(bundle, tf_state_dict):
         function_list=[],
         module=bundle.module_bytecode,
     )
+    spec = exported_program.call_spec.out_spec
+
+    # The module returning a flat array.
+    if not spec.context:
+      return call_module_return
+
+    flat_names = common_utils.flat_dict_names(spec.children_specs, spec.context)
+    return {name: value for name, value in zip(flat_names, call_module_return)}
 
   return inner
 
@@ -128,8 +141,10 @@ def merged_bundle_to_tfl_model(
       for bundle, sig in zip(merged_bundle.bundles, signatures)
   ]
   tf_functions = [
-      _wrap_as_tf_func(bundle, tf_state_dict)
-      for bundle in merged_bundle.bundles
+      _wrap_as_tf_func(bundle, tf_state_dict, ep)
+      for bundle, ep in zip(
+          merged_bundle.bundles, merged_bundle.exported_programs
+      )
   ]
 
   tf_module = tf.Module()
@@ -202,7 +217,9 @@ def merge_mlir_bundles(
   )
 
   merged_bundle = MergedBundle(
-      bundles=bundles.copy(), deduped_tf_vars=deduped_vars
+      bundles=bundles.copy(),
+      exported_programs=exported_programs,
+      deduped_tf_vars=deduped_vars,
   )
   for bundle, signature in zip(merged_bundle.bundles, signatures):
     bundle.state_dict = state_dict

ai_edge_torch/lowertools/torch_xla_utils.py

@@ -51,6 +51,7 @@ MlirBundle = stablehlo.StableHLOModelBundle
 class MergedBundle:
 
   bundle: stablehlo.StableHLOModelBundle
+  exported_programs: list[torch.export.ExportedProgram]
   deduped_tf_vars: list[tf.Variable]
 
 
@@ -58,9 +59,9 @@ def exported_program_to_mlir(
     exported_program: torch.export.ExportedProgram,
     sample_args: tuple[torch.Tensor],
 ) -> stablehlo.StableHLOModelBundle:
-  # Setting export_weights to False here so that pytorch/xla avoids copying the weights
-  # to a numpy array which would lead to memory bloat. This means that the state_dict
-  # in the returned bundle is going to be empty.
+  # Setting export_weights to False here so that pytorch/xla avoids copying the
+  # weights to a numpy array which would lead to memory bloat. This means that
+  # the state_dict in the returned bundle is going to be empty.
   return stablehlo.exported_program_to_stablehlo(
       exported_program,
       stablehlo.StableHLOExportOptions(
@@ -96,7 +97,9 @@ def merge_mlir_bundles(
         bundle.additional_constants
     )
   return MergedBundle(
-      bundle=new_shlo_model_bundle, deduped_tf_vars=deduped_tf_vars
+      bundle=new_shlo_model_bundle,
+      exported_programs=exported_programs,
+      deduped_tf_vars=deduped_tf_vars,
   )
 
 
@@ -108,7 +111,9 @@ def _get_shape_with_dynamic(signature: stablehlo.VariableSignature):
 
 
 def _wrap_as_tf_func(
-    func: stablehlo.StableHLOFunc, bundle: stablehlo.StableHLOModelBundle
+    func: stablehlo.StableHLOFunc,
+    bundle: stablehlo.StableHLOModelBundle,
+    exported_program: torch.export.ExportedProgram,
 ):
   def inner(*args):
     type_info = [sig.dtype for sig in func.meta.output_signature]
@@ -116,7 +121,7 @@ def _wrap_as_tf_func(
         _get_shape_with_dynamic(sig) for sig in func.meta.output_signature
     ]
     call_args = stablehlo._extract_call_parameters(args, func.meta, bundle)
-    return tfxla.call_module(
+    call_module_return = tfxla.call_module(
         tuple(call_args),
         version=5,
         Tout=type_info,
@@ -124,15 +129,16 @@ def _wrap_as_tf_func(
         function_list=[],
         module=func.bytecode,
     )
+    spec = exported_program.call_spec.out_spec
 
-  return inner
+    # The module returning a flat array.
+    if not spec.context:
+      return call_module_return
 
+    flat_names = common_utils.flat_dict_names(spec.children_specs, spec.context)
+    return {name: value for name, value in zip(flat_names, call_module_return)}
 
-def _make_tf_function(
-    bundle: stablehlo.StableHLOModelBundle = None,
-):
-  bundle = bundle if bundle is None else bundle
-  return [_wrap_as_tf_func(func, bundle) for func in bundle.stablehlo_funcs]
+  return inner
 
 
 def _make_tf_signature(
@@ -205,7 +211,12 @@ def merged_bundle_to_tfl_model(
       for func, sig in zip(shlo_bundle.stablehlo_funcs, signatures)
   )
 
-  tf_functions = _make_tf_function(shlo_bundle)
+  tf_functions = [
+      _wrap_as_tf_func(func, shlo_bundle, ep)
+      for func, ep in zip(
+          shlo_bundle.stablehlo_funcs, merged_bundle.exported_programs
+      )
+  ]
 
   tf_module.f = []
   for tf_sig, func in zip(tf_signatures, tf_functions):

ai_edge_torch/model.py

@@ -21,6 +21,7 @@ PyTorch models can be converted to this representation through
 from __future__ import annotations
 
 import abc
+import re
 
 import numpy.typing as npt
 import tensorflow as tf
@@ -115,11 +116,18 @@ class TfLiteModel(Model):
     inputs = {**inputs, **kwargs}
     outputs = runner(**inputs)
 
-    return (
-        outputs['output_0']
-        if len(outputs) == 1
-        else [outputs[f'output_{idx}'] for idx in range(len(outputs))]
-    )
+    # When attempting to run a model, check if all the output tensors are named
+    # output_<number>. If so, assume the pytorch model returned a tuple and not
+    # a dictionary.
+    output_heuristic = lambda key: bool(re.search(r'output_\d+', key))
+    if all(output_heuristic(key) for key in outputs.keys()):
+      return (
+          outputs['output_0']
+          if len(outputs) == 1
+          else [outputs[f'output_{idx}'] for idx in range(len(outputs))]
+      )
+
+    return outputs
 
   def export(self, path: str) -> None:
     """Serializes the edge model to disk.

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.3.0.dev20240817"
+__version__ = "0.3.0.dev20240823"

{ai_edge_torch_nightly-0.3.0.dev20240817.dist-info → ai_edge_torch_nightly-0.3.0.dev20240823.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20240817
+Version: 0.3.0.dev20240823
 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.dev20240817.dist-info → ai_edge_torch_nightly-0.3.0.dev20240823.dist-info}/RECORD

@@ -1,13 +1,13 @@
 ai_edge_torch/__init__.py,sha256=48qP37uHT90YPs4eIUQxCiWVwqGEX3idCUs6mQKvX1U,1168
 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=C4yfJq9TbtZBH5gwhPSUtBgiIe04GkxvCq5TImNopww,706
+ai_edge_torch/model.py,sha256=7tox6sdFIlCYPLDYpjFcD8cPTSivURCL_VV6-Dt5Sfc,4910
+ai_edge_torch/version.py,sha256=a63GrjqX4sRjk0WbC_0gGhT-ax_TLEi4iCEw0Iys7bw,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
 ai_edge_torch/_convert/converter.py,sha256=ezmaATnQi7NWDo37LUb-hEXtZSmT7_AT6vqXC6Fcq1o,8615
-ai_edge_torch/_convert/signature.py,sha256=zSSG4_u38pOw08pHgXssZDK_EanACDx7lcI-PzBgh5A,3313
+ai_edge_torch/_convert/signature.py,sha256=rGpBNss3Y9FCCCcdBwDo16KqavJi8N5P0M_6Wnmoz6Y,2320
 ai_edge_torch/_convert/to_channel_last_io.py,sha256=_31phf7TYgZY2ftpNbrdlB1RhDium1lz_BXEQ6IsMFc,2893
 ai_edge_torch/_convert/fx_passes/__init__.py,sha256=hVuNftOcZIpwkUcPab44mhFmi9Z1f1REV5o3j39Sf-w,2818
 ai_edge_torch/_convert/fx_passes/_pass_base.py,sha256=WVYZuocpygHAzk9u1GNoGowAIOHTlJXyA_NklmYkRms,1672
@@ -26,9 +26,9 @@ ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/layout_partitio
 ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/layout_partitioners/greedy.py,sha256=L_x8BrF7UDah-SYl-pG11I6CIckdU9kBTUHcmwW4cts,2420
 ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/layout_partitioners/min_cut.py,sha256=mzfL9cf0qBnpmxM_OlMQFvQsEZV2B_Mia9yEJV4J7rI,7135
 ai_edge_torch/_convert/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/_convert/test/test_convert.py,sha256=y0ZRivdglGx217rnacze8N6nd7aafk28NkbBFUSa9DQ,13121
+ai_edge_torch/_convert/test/test_convert.py,sha256=tvj7fWHHmA9ddtcu-Fp3lJ6emaAQMrtK9wCG0cjgRAo,14413
 ai_edge_torch/_convert/test/test_convert_composites.py,sha256=CBiOqq-m7QT2ggBI1jBl9MkTIT5d0nK1tA0BUga0LGs,7994
-ai_edge_torch/_convert/test/test_convert_multisig.py,sha256=4jm5blAfzLMjvrJt0ntuG_Fgy4Ie3SoUOGBOy9tf6bg,4725
+ai_edge_torch/_convert/test/test_convert_multisig.py,sha256=0xIkahEU26Qx9GGn6Dm05ObIqJvsCdh692dREcaHEdE,4725
 ai_edge_torch/_convert/test/test_to_channel_last_io.py,sha256=jLAmyHw5llT2ff8qA8mem3eVN57e_o5EpBnW72ZtP2I,3026
 ai_edge_torch/debug/__init__.py,sha256=N05Mmvi41KgSuK0JhuMejERESgP8QekiGdp9_PEyuKU,742
 ai_edge_torch/debug/culprit.py,sha256=7UYVpVWpiCXbMAyThVtHt_kc_poT7sCTh5UUPvcycgk,14832
@@ -53,7 +53,7 @@ ai_edge_torch/generative/examples/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIX
 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=cCki-0cKvmGxK4Md6dRNdPDWZUyhkJUI854OCTFf3h0,6262
-ai_edge_torch/generative/examples/gemma/gemma2.py,sha256=j-zxJ-JNRnQ_kDzUESmsyy_a_4IxWZ510HmIImc0LDc,8240
+ai_edge_torch/generative/examples/gemma/gemma2.py,sha256=q9Zil66EvRKrSpLVQHxKHu_8NL0HAgY2FbtThoTZVUY,8226
 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
@@ -94,7 +94,7 @@ ai_edge_torch/generative/layers/kv_cache.py,sha256=Ob8QeXWW5xt-6hcGA0uoC48eRQ8lf
 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=x2bOmrTgOISXcb06IDP7X3xgftpPpxOjBXw_OxTMVns,3874
+ai_edge_torch/generative/layers/scaled_dot_product_attention.py,sha256=VW-VP8e7FTSPCdu-6DVxpwNrIdgX0R_kq6F6MSEiyXE,3848
 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
@@ -114,7 +114,7 @@ ai_edge_torch/generative/test/test_loader.py,sha256=1ZqAq0HY5uIioumsReOVIsbGBx0W
 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=bAWZ7FM4v_pPnX_AmEdGxHkDH65QdL-MjIP3PxscZmI,12649
+ai_edge_torch/generative/utilities/loader.py,sha256=QFZ2lkeoYQ9MZ1CAFVxBHG4OT192SH74UtJCvbDsdeI,12727
 ai_edge_torch/generative/utilities/stable_diffusion_loader.py,sha256=pKp3AMSbS3otCvgwJRF5M1l4JRNKk-aCKimXzIMSrds,35679
 ai_edge_torch/generative/utilities/t5_loader.py,sha256=_UXcc1QKT-S92hikfo-fTBFhnYLzROqcyRqKonVsqj4,16885
 ai_edge_torch/hlfb/__init__.py,sha256=sH4um75na-O8tzxN6chFyp6Y4xnexsE7kUQpZySv6dE,735
@@ -124,12 +124,12 @@ ai_edge_torch/hlfb/mark_pattern/pattern.py,sha256=uiYRfzD1T8deCEAGfdAFusRbI41m14
 ai_edge_torch/hlfb/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 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/__init__.py,sha256=A8WBXvWtuFYYWtNTqPD7waVntLaSVAnSMwx5ugjZBIw,761
 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=32cak8uiXFIVdkaYFhIW1fWG4NzLrYq-w8xK0pNkhYc,6547
+ai_edge_torch/lowertools/common_utils.py,sha256=Z7p-ivOHtddktpnHrlDm_dSoTxJOdEjFXIGQbzjgwQo,4504
+ai_edge_torch/lowertools/odml_torch_utils.py,sha256=EA2ylE4abCyC1G3lYuNouPGx0lmwtIZe7c42dtX0-3g,7146
 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/lowertools/torch_xla_utils.py,sha256=-SRm9YNsIGsaVd5Cyp2PP-tdLBJH8EDoMFAa2y89a1w,9043
 ai_edge_torch/quantize/__init__.py,sha256=aB5dXot04bqyUhpsDFvxt9CIi15QAC4euvqOndJ0XLU,714
 ai_edge_torch/quantize/pt2e_quantizer.py,sha256=CKIEhs9jCcna64qj1jFH9zEbMbRdyeGV_TmSqEBPjes,15741
 ai_edge_torch/quantize/pt2e_quantizer_utils.py,sha256=eARD1LxLi5m7Z0n_psAkeX_AtUp4fNkE--oECBfivv4,36208
@@ -137,8 +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.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,,
+ai_edge_torch_nightly-0.3.0.dev20240823.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20240823.dist-info/METADATA,sha256=OPYTq5RQCL2lvIFeBKOwbFusi4rq_Qo2ytgn_JQTVb0,1885
+ai_edge_torch_nightly-0.3.0.dev20240823.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20240823.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20240823.dist-info/RECORD,,