ai-edge-torch-nightly 0.3.0.dev20240904__py3-none-any.whl → 0.3.0.dev20240906__py3-none-any.whl

ai_edge_torch/generative/examples/test_models/toy_model.py

@@ -71,6 +71,56 @@ class ToySingleLayerModel(torch.nn.Module):
     return self.lm_head(x)
 
 
+class ToySingleLayerModelWeightSharing(torch.nn.Module):
+
+  def __init__(self, config: cfg.ModelConfig) -> None:
+    super().__init__()
+    self.lm_head = nn.Linear(
+        config.embedding_dim, config.vocab_size, bias=config.lm_head_use_bias
+    )
+    self.tok_embedding = nn.Embedding(config.vocab_size, config.embedding_dim)
+    self.lm_head = nn.Linear(
+        config.embedding_dim,
+        config.vocab_size,
+        bias=config.lm_head_use_bias,
+    )
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+    self.transformer_block = TransformerBlock(config)
+    self.final_norm = builder.build_norm(
+        config.embedding_dim,
+        config.final_norm_config,
+    )
+    self.rope_cache = attn_utils.build_rope_cache(
+        size=config.max_seq_len,
+        dim=int(
+            config.attn_config.rotary_percentage * config.attn_config.head_dim
+        ),
+        base=10_000,
+        condense_ratio=1,
+        dtype=torch.float32,
+        device=torch.device('cpu'),
+    )
+    self.mask_cache = attn_utils.build_causal_mask_cache(
+        size=config.max_seq_len, dtype=torch.float32, device=torch.device('cpu')
+    )
+    self.config = config
+
+  @torch.inference_mode
+  def forward(self, idx: torch.Tensor, input_pos: torch.Tensor) -> torch.Tensor:
+    x = self.tok_embedding(idx)
+    cos, sin = self.rope_cache
+
+    cos = cos.index_select(0, input_pos)
+    sin = sin.index_select(0, input_pos)
+    mask = self.mask_cache.index_select(2, input_pos)
+    mask = mask[:, :, :, : self.config.max_seq_len]
+
+    x = self.transformer_block(x, (cos, sin), mask, input_pos)
+    x = self.final_norm(x)
+    res = self.lm_head(x)
+    return res
+
+
 def get_model_config() -> cfg.ModelConfig:
   attn_config = cfg.AttentionConfig(
       num_heads=32,

ai_edge_torch/generative/test/test_model_conversion.py

@@ -70,7 +70,6 @@ class TestModelConversion(googletest.TestCase):
         )
     )
 
-
   @googletest.skipIf(
       ai_edge_config.Config.use_torch_xla,
       reason="tests with custom ops are not supported on oss",
@@ -130,6 +129,7 @@ class TestModelConversion(googletest.TestCase):
     )
 
     copied_model = copy.deepcopy(pytorch_model)
+    copied_edge = copy.deepcopy(edge_model)
 
     self.assertTrue(
         model_coverage.compare_tflite_torch(
@@ -141,18 +141,15 @@ class TestModelConversion(googletest.TestCase):
         )
     )
 
-    # TODO(b/362840003): figure why this decode output has big numerical diff.
-    skip_output_check = True
-    if not skip_output_check:
-      self.assertTrue(
-          model_coverage.compare_tflite_torch(
-              edge_model,
-              copied_model,
-              (decode_token, decode_input_pos),
-              signature_name="decode",
-              num_valid_inputs=1,
-          )
-      )
+    self.assertTrue(
+        model_coverage.compare_tflite_torch(
+            copied_edge,
+            copied_model,
+            (decode_token, decode_input_pos),
+            signature_name="decode",
+            num_valid_inputs=1,
+        )
+    )
 
 
 if __name__ == "__main__":

ai_edge_torch/generative/test/test_model_conversion_large.py

@@ -82,28 +82,28 @@ class TestModelConversion(googletest.TestCase):
     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)
+    prefill_tokens = torch.full((1, 10), 0, dtype=torch.long, device="cpu")
+    prefill_tokens[0, :4] = idx
+    prefill_input_pos = torch.arange(0, 10)
 
-    edge_model = ai_edge_torch.convert(model, (tokens, input_pos))
+    edge_model = ai_edge_torch.signature(
+        "prefill", model, (prefill_tokens, prefill_input_pos)
+    ).convert()
     edge_model.set_interpreter_builder(
         self._interpreter_builder(edge_model.tflite_model())
     )
 
-    # TODO(b/362840003): debug numerical diff.
-    skip_output_check = True
-    if not skip_output_check:
-      self.assertTrue(
-          model_coverage.compare_tflite_torch(
-              edge_model,
-              model,
-              (tokens, input_pos),
-              num_valid_inputs=1,
-              atol=1e-2,
-              rtol=1e-5,
-          )
-      )
+    self.assertTrue(
+        model_coverage.compare_tflite_torch(
+            edge_model,
+            model,
+            (prefill_tokens, prefill_input_pos),
+            signature_name="prefill",
+            num_valid_inputs=1,
+            atol=1e-2,
+            rtol=1e-5,
+        )
+    )
 
   @googletest.skipIf(
       ai_edge_config.Config.use_torch_xla,

ai_edge_torch/generative/test/test_quantize.py

@@ -25,16 +25,16 @@ from ai_edge_torch.generative.quantize.quant_attrs import Granularity
 from ai_edge_torch.generative.quantize.quant_attrs import Mode
 from ai_edge_torch.quantize import quant_config
 from ai_edge_torch.testing import model_coverage
-from parameterized import parameterized
 import torch
 
 from absl.testing import absltest as googletest
+from absl.testing import parameterized
 
 
-class TestVerifyRecipes(googletest.TestCase):
+class TestVerifyRecipes(parameterized.TestCase):
   """Unit tests that check for model quantization recipes."""
 
-  @parameterized.expand([
+  @parameterized.parameters([
       (Dtype.FP32, Dtype.FP32),
       (Dtype.INT8, Dtype.INT8),
       (Dtype.INT8, Dtype.FP16),
@@ -52,7 +52,7 @@ class TestVerifyRecipes(googletest.TestCase):
           with self.assertRaises(ValueError):
             quant_recipe.LayerQuantRecipe(activation, weight, m, a, g).verify()
 
-  @parameterized.expand([
+  @parameterized.parameters([
       (
           Dtype.FP32,
           Dtype.INT8,
@@ -88,7 +88,7 @@ class TestVerifyRecipes(googletest.TestCase):
     ).verify()
 
 
-class TestQuantizeConvert(googletest.TestCase):
+class TestQuantizeConvert(parameterized.TestCase):
   """Test conversion with quantization."""
 
   def _attention_int8_dynamic_recipe() -> quant_config.QuantConfig:
@@ -105,17 +105,13 @@ class TestQuantizeConvert(googletest.TestCase):
         )
     )
 
-  @parameterized.expand([
+  @parameterized.parameters([
       (quant_recipes.full_fp16_recipe()),
       (quant_recipes.full_int8_dynamic_recipe()),
       (quant_recipes.full_int8_weight_only_recipe()),
       (_attention_int8_dynamic_recipe()),
       (_feedforward_int8_dynamic_recipe()),
   ])
-  @googletest.skipIf(
-      not config.Config.use_torch_xla,
-      reason="Not working with odml_torch at the moment.",
-  )
   def test_quantize_convert_toy_sizes(self, quant_config):
     config = toy_model.get_model_config()
     pytorch_model = toy_model.ToySingleLayerModel(config)
@@ -132,6 +128,23 @@ class TestQuantizeConvert(googletest.TestCase):
         "Quantized model isn't smaller than F32 model.",
     )
 
+  def test_quantize_convert_toy_weight_sharing(self):
+    config = toy_model.get_model_config()
+    pytorch_model = toy_model.ToySingleLayerModelWeightSharing(config)
+    idx = torch.unsqueeze(torch.arange(0, 100), 0)
+    input_pos = torch.arange(0, 100)
+
+    quant_config = quant_recipes.full_int8_dynamic_recipe()
+    quantized_model = ai_edge_torch.convert(
+        pytorch_model, (idx, input_pos), quant_config=quant_config
+    )
+    float_model = ai_edge_torch.convert(pytorch_model, (idx, input_pos))
+    self.assertLess(
+        len(quantized_model._tflite_model),
+        len(float_model._tflite_model),
+        "Quantized model isn't smaller than F32 model.",
+    )
+
   def test_quantize_convert_compare_toy(self):
     self.skipTest("b/338288901")
     config = toy_model_with_kv_cache.get_model_config()

ai_edge_torch/generative/utilities/loader.py

@@ -208,7 +208,7 @@ class ModelLoader:
     if self._file_name.endswith(".safetensors"):
       return load_safetensors
 
-    if self._file_name.endswith(".bin") or self._file_name.endswith(".pt"):
+    if self._file_name.endswith(".bin") or self._file_name.endswith("pt"):
       return load_pytorch_statedict
 
     raise ValueError("File format not supported.")

ai_edge_torch/lowertools/odml_torch_utils.py

@@ -21,6 +21,7 @@ from ai_edge_torch import odml_torch
 from ai_edge_torch._convert import conversion_utils
 from ai_edge_torch._convert import signature as signature_module
 from ai_edge_torch.lowertools import common_utils
+from ai_edge_torch.lowertools import translate_recipe
 from ai_edge_torch.odml_torch import export
 from ai_edge_torch.odml_torch import export_utils
 from ai_edge_torch.quantize import quant_config as qcfg
@@ -186,10 +187,29 @@ def merged_bundle_to_tfl_model(
     converter._experimental_enable_composite_direct_lowering = True
     converter.model_origin_framework = "PYTORCH"
 
+    conversion_utils.set_tfl_converter_quant_flags(converter, quant_config)
+    if (
+        quant_config is not None
+        and quant_config._quantizer_mode
+        == quant_config._QuantizerMode.AI_EDGE_QUANTIZER
+    ):
+      translated_recipe = translate_recipe.translate_to_ai_edge_recipe(
+          quant_config.generative_recipe
+      )
+
     conversion_utils.apply_tfl_converter_flags(converter, _tfl_converter_flags)
 
     tflite_model = converter.convert()
 
+    if (
+        quant_config is not None
+        and quant_config._quantizer_mode
+        == quant_config._QuantizerMode.AI_EDGE_QUANTIZER
+    ):
+      tflite_model = translate_recipe.quantize_model(
+          tflite_model, translated_recipe
+      )
+
   return tflite_model
 
 

ai_edge_torch/lowertools/torch_xla_utils.py

@@ -25,8 +25,8 @@ from typing import Any, Dict, Optional, Tuple, Union
 from ai_edge_torch import model
 from ai_edge_torch._convert import conversion_utils
 from ai_edge_torch._convert import signature as signature_module
-from ai_edge_torch.generative.quantize.ai_edge_quantizer_glue import translate_recipe  # NOQA
 from ai_edge_torch.lowertools import common_utils
+from ai_edge_torch.lowertools import translate_recipe
 from ai_edge_torch.quantize import quant_config as qcfg
 import torch
 from torch_xla import stablehlo

ai_edge_torch/{generative/quantize/ai_edge_quantizer_glue → lowertools}/translate_recipe.py

@@ -17,7 +17,8 @@ from ai_edge_quantizer import quantizer
 from ai_edge_torch.generative.quantize import quant_attrs
 from ai_edge_torch.generative.quantize import quant_recipe
 
-_OpExecutionMode = quantizer.qtyping.OpExecutionMode
+_ComputePrecision = quantizer.qtyping.ComputePrecision
+_QuantGranularity = quantizer.qtyping.QuantGranularity
 _OpName = quantizer.qtyping.TFLOperationName
 _TensorQuantConfig = quantizer.qtyping.TensorQuantizationConfig
 _OpQuantConfig = quantizer.qtyping.OpQuantizationConfig
@@ -50,21 +51,31 @@ def _get_dtype_from_dtype(
     return quantizer.qtyping.TensorDataType.INT
 
 
-def _get_execution_mode_from_mode(mode: quant_attrs.Mode) -> _OpExecutionMode:
+def _get_compute_precision_from_mode(
+    mode: quant_attrs.Mode,
+) -> _ComputePrecision:
   if mode == quant_attrs.Mode.DYNAMIC_RANGE:
-    return _OpExecutionMode.DRQ
+    return _ComputePrecision.INTEGER
   elif mode == quant_attrs.Mode.WEIGHT_ONLY:
-    return _OpExecutionMode.WEIGHT_ONLY
+    return _ComputePrecision.FLOAT
   raise ValueError('Unimplemented execution mode')
 
 
-def _get_channelwise_from_granularity(
+def _get_explicit_dequant_from_mode(mode: quant_attrs.Mode) -> bool:
+  if mode == quant_attrs.Mode.DYNAMIC_RANGE:
+    return False
+  elif mode == quant_attrs.Mode.WEIGHT_ONLY:
+    return True
+  raise ValueError('Unimplemented execution mode')
+
+
+def _get_granularity(
     granularity: quant_attrs.Granularity,
 ) -> bool:
   if granularity == quant_attrs.Granularity.CHANNELWISE:
-    return True
-  elif granularity == quant_attrs.Granularity.NONE:
-    return False
+    return _QuantGranularity.CHANNELWISE
+  if granularity == quant_attrs.Granularity.NONE:
+    return _QuantGranularity.TENSORWISE
   raise ValueError('Unimplemented granularity')
 
 
@@ -88,12 +99,13 @@ def _set_quant_config(
           weight_tensor_config=_TensorQuantConfig(
               num_bits=_get_nbits_from_dtype(layer_recipe.weight_dtype),
               symmetric=True,
-              channel_wise=_get_channelwise_from_granularity(
-                  layer_recipe.granularity
-              ),
+              granularity=_get_granularity(layer_recipe.granularity),
               dtype=_get_dtype_from_dtype(layer_recipe.weight_dtype),
           ),
-          execution_mode=_get_execution_mode_from_mode(layer_recipe.mode),
+          compute_precision=_get_compute_precision_from_mode(layer_recipe.mode),
+          explicit_dequantize=_get_explicit_dequant_from_mode(
+              layer_recipe.mode
+          ),
       ),
       algorithm_key=_get_algorithm_key_from_algorithm(layer_recipe.algorithm),
   )

ai_edge_torch/odml_torch/export.py

@@ -227,12 +227,9 @@ def exported_program_to_mlir(
     exported_program: torch.export.ExportedProgram,
 ) -> MlirLowered:
   """Lower the exported program to MLIR."""
-  if torch.__version__ >= "2.2":
-    # torch version 2.1 didn't expose this yet
-    exported_program = exported_program.run_decompositions()
-    exported_program = exported_program.run_decompositions(
-        lowerings.decompositions()
-    )
+  exported_program = exported_program.run_decompositions(
+      lowerings.decompositions()
+  )
 
   with export_utils.create_ir_context() as context, ir.Location.unknown():
 

ai_edge_torch/odml_torch/jax_bridge/_wrap.py

@@ -35,7 +35,7 @@ jax.config.update("jax_enable_x64", True)
 
 def _lower_to_ir_text(
     jaxfn, args, kwargs, ir_input_names: list[str] = None
-) -> str:
+) -> tuple[str, list[ir.Value]]:
   args = utils.tree_map_list_to_tuple(args)
   kwargs = utils.tree_map_list_to_tuple(kwargs)
 
@@ -74,7 +74,9 @@ def _lower_to_ir_text(
             x for x in pytree.tree_flatten(arg)[0] if isinstance(x, ir.Value)
         ]
 
-  def new_lowering(*args, **jax_lower_static_kwargs):
+  def lower_wrapper(*args):
+    nonlocal jax_lower_static_kwargs
+
     jaxfn_args = []
     jaxfn_kwargs = jax_lower_static_kwargs.copy()
     for name, arg in zip(jax_lower_argnames, args):
@@ -85,11 +87,7 @@ def _lower_to_ir_text(
 
     return jaxfn(*jaxfn_args, **jaxfn_kwargs)
 
-  return (
-      jax.jit(new_lowering, static_argnames=static_argnames)
-      .lower(*jax_lower_args, **jax_lower_static_kwargs)
-      .as_text()
-  ), ir_inputs
+  return jax.jit(lower_wrapper).lower(*jax_lower_args).as_text(), ir_inputs
 
 
 def wrap(jaxfn: Callable[Any, Any], ir_input_names: list[str] = None):

ai_edge_torch/odml_torch/lowerings/registry.py

@@ -52,6 +52,7 @@ class LoweringRegistry:
 
 
 global_registry = LoweringRegistry()
+global_registry.decompositions.update(torch._decomp.core_aten_decompositions())
 global_registry.decompositions.update(
     torch._decomp.get_decompositions([
         torch.ops.aten.upsample_nearest2d,
@@ -70,6 +71,13 @@ global_registry.decompositions.update(
     ])
 )
 
+torch._decomp.remove_decompositions(
+    global_registry.decompositions,
+    [
+        torch.ops.aten.roll,
+    ],
+)
+
 
 def lookup(op):
   return global_registry.lookup(op)

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.3.0.dev20240904"
+__version__ = "0.3.0.dev20240906"

{ai_edge_torch_nightly-0.3.0.dev20240904.dist-info → ai_edge_torch_nightly-0.3.0.dev20240906.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20240904
+Version: 0.3.0.dev20240906
 Summary: Supporting PyTorch models with the Google AI Edge TFLite runtime.
 Home-page: https://github.com/google-ai-edge/ai-edge-torch
 Keywords: On-Device ML,AI,Google,TFLite,PyTorch,LLMs,GenAI
@@ -30,7 +30,7 @@ Requires-Dist: tabulate
 Requires-Dist: torch>=2.4.0
 Requires-Dist: torch-xla>=2.4.0
 Requires-Dist: tf-nightly>=2.18.0.dev20240722
-Requires-Dist: ai-edge-quantizer-nightly==0.0.1.dev20240718
+Requires-Dist: ai-edge-quantizer-nightly
 
 Library that supports converting PyTorch models into a .tflite format, which can
 then be run with TensorFlow Lite and MediaPipe.  This enables applications for

{ai_edge_torch_nightly-0.3.0.dev20240904.dist-info → ai_edge_torch_nightly-0.3.0.dev20240906.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=NYV6Mkaje_ditIEI_s_7nLP_-8i4kbGM8nRzieVkbUI,5397
-ai_edge_torch/version.py,sha256=YC_qiN7DiHKGt_u71WY_mNK2BaoZr6mvPw7s7rWHA84,706
+ai_edge_torch/version.py,sha256=vEc_GracKJpLkIs6M45gCFWkBMuXTjmvfvJnfXBSyrs,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
@@ -77,7 +77,7 @@ ai_edge_torch/generative/examples/t5/convert_to_tflite.py,sha256=CZVuNEL8OHPkdsz
 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=LfWO_gSr1f66V1pxAc6yh21mtaJs7TVeuO9748zXBnE,3963
+ai_edge_torch/generative/examples/test_models/toy_model.py,sha256=5wj2RmQRIwD6O_R_pp-A_7gKGSdHWDSXyis97r1ELVI,5622
 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
@@ -106,16 +106,14 @@ ai_edge_torch/generative/quantize/quant_recipe.py,sha256=tKnuJq6hPD23JPCB9nPAlE1
 ai_edge_torch/generative/quantize/quant_recipe_utils.py,sha256=4fgmP_GgeiFUOkIaC9ZZXC12eO3DQZdrWDXRz5YXiwU,2270
 ai_edge_torch/generative/quantize/quant_recipes.py,sha256=0Kvr_o7pbMnE8VMe6Ml0FBxkHM6RJ3C14B2I1mjItjc,2030
 ai_edge_torch/generative/quantize/supported_schemes.py,sha256=FjdycEOvxRgBmQdZVufetPvkDoD7rUowIOSKV9oV5Kk,1418
-ai_edge_torch/generative/quantize/ai_edge_quantizer_glue/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/quantize/ai_edge_quantizer_glue/translate_recipe.py,sha256=sSHc_4hUEvi-3KmqbpqWbrRKBjCI1AOctM3dr2EH3vk,5263
 ai_edge_torch/generative/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/test/test_experimental_ekv.py,sha256=8qv_eVtJW9GPvBEf2hPQe3tpdJ33XShya6MCX1FqrZM,4355
 ai_edge_torch/generative/test/test_loader.py,sha256=_y5EHGgoNOmCuYonsB81UJScHVsTAQXUVd44czMAw6k,3379
-ai_edge_torch/generative/test/test_model_conversion.py,sha256=zgQN0I9z8Xm0HcAGJIrnGzZWXEgNN041n4C5MXMNMqA,5286
-ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=o3l7HFHP-sg8aHeLNTSpMF91YovPODjp4QzYUnSJiIE,4479
-ai_edge_torch/generative/test/test_quantize.py,sha256=JEsk9SAkHK0SFm44K_quISc5yBBS6yvtBP1MDyFHdFw,5344
+ai_edge_torch/generative/test/test_model_conversion.py,sha256=b3InJ8Rx03YtHpE9h-j0pSXAY1cCf-dLlx4Y5LSJnRQ,5174
+ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=9JXcd-rX8MpsYeEWUFEXf783GOwYOLY64KzDfFdmRJ8,4484
+ai_edge_torch/generative/test/test_quantize.py,sha256=kY_NRpF-v1i4clqI1CFFWEagJv-5PzBDkeJ2fInl9_w,5913
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
-ai_edge_torch/generative/utilities/loader.py,sha256=QFZ2lkeoYQ9MZ1CAFVxBHG4OT192SH74UtJCvbDsdeI,12727
+ai_edge_torch/generative/utilities/loader.py,sha256=6J0aAP6-6LySeqeYIHKcchr5T9cVtSO34aoDr3V9gxY,12726
 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
@@ -128,13 +126,14 @@ ai_edge_torch/hlfb/test/test_stablehlo_composite_builder.py,sha256=j8WpeS-mz3Zr4
 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=Z7p-ivOHtddktpnHrlDm_dSoTxJOdEjFXIGQbzjgwQo,4504
-ai_edge_torch/lowertools/odml_torch_utils.py,sha256=GKfW1X-QSFffQdVlBuD-bNpP265xcdUlfBY3-9I4f_o,7447
+ai_edge_torch/lowertools/odml_torch_utils.py,sha256=K5dZ_fFDL3GWKo0IoY4OC_GX5MY-guY-MqteolyV9hg,8098
 ai_edge_torch/lowertools/test_utils.py,sha256=bPgc2iXX16KYtMNvmsRdKfrCY6UJmcfitfCOvHoD7Oc,1930
-ai_edge_torch/lowertools/torch_xla_utils.py,sha256=-SRm9YNsIGsaVd5Cyp2PP-tdLBJH8EDoMFAa2y89a1w,9043
+ai_edge_torch/lowertools/torch_xla_utils.py,sha256=n6G3pFGmHar7kgKDsdTB74kv1PUuTTu1XjV7R-QizzE,9003
+ai_edge_torch/lowertools/translate_recipe.py,sha256=DNzD0VD35YZDqiZjAF1IyIPSzUGPDpE0jvFCCYIzpnc,5667
 ai_edge_torch/odml_torch/__init__.py,sha256=S8jOzE9nLof-6es3XDiGJRN-9H_XTxsVm9dE7lD3RWo,812
 ai_edge_torch/odml_torch/_torch_future.py,sha256=jSYHf1CMTJzMizPMbu2b39hAt0ZTR6gQLq67GMe9KTo,2336
 ai_edge_torch/odml_torch/_torch_library.py,sha256=Lw1gqL2HWNRspdTwNhIkYAHDyafHedHtkXyKKxn-Wss,805
-ai_edge_torch/odml_torch/export.py,sha256=OXN6jipwFtBvQ9XdyeDGQTQ_-UnCxPYnLc_WW7xF0aI,10469
+ai_edge_torch/odml_torch/export.py,sha256=_n43AlaTLvAK6r1szs47gSBqp-x19ZNCNtyFIWzuE4Q,10322
 ai_edge_torch/odml_torch/export_utils.py,sha256=q84U69ZQ82hLXw-xncJ8IW-K71Xux-NWlzZTs7hdZWA,5127
 ai_edge_torch/odml_torch/tf_integration.py,sha256=lTFJPPEijLPFmn6qq2jbpVTQOo0YaOTK36kK6rCiyIE,5956
 ai_edge_torch/odml_torch/composite/__init__.py,sha256=71GM_gDZxJyo38ZSoYSwhZX3xKA9rknO93JS9kw9w_c,778
@@ -144,7 +143,7 @@ ai_edge_torch/odml_torch/debuginfo/__init__.py,sha256=9ag6-WWRG50rPCtIV7OpIokEKu
 ai_edge_torch/odml_torch/debuginfo/_build.py,sha256=1xCXOs3-9UcsOyLFH0uyQwLu7c06iYFTo0NQ7Ckbl2I,1465
 ai_edge_torch/odml_torch/debuginfo/_op_polyfill.py,sha256=IvOBQyROI9WHS3umHRxsDW-1YElU9BPWzKtJA2eKWOI,1739
 ai_edge_torch/odml_torch/jax_bridge/__init__.py,sha256=Jco5zvejxuyl9xHQxZICAKbkgH7x38qPlwUUpD7S15Q,730
-ai_edge_torch/odml_torch/jax_bridge/_wrap.py,sha256=hXvhKtbH7lGytm6QZOKpTmaLJN3kfENBcSIKQ39ReXA,5478
+ai_edge_torch/odml_torch/jax_bridge/_wrap.py,sha256=drN3L0uTsSjkluKgt6Ngq7b5HLReE_7iAitHpZ9PKqE,5428
 ai_edge_torch/odml_torch/jax_bridge/utils.py,sha256=T8isGc896VrHZ6c_L5pYmLpolQ7ibcOlgWfPuVFPzIg,2264
 ai_edge_torch/odml_torch/lowerings/__init__.py,sha256=GqYk6oBJw7KWeG4_6gxSu_OvYhjJcC2FpGzWPPEdH6w,933
 ai_edge_torch/odml_torch/lowerings/_basic.py,sha256=wV8AUK8dvjLUy3qjqw_IxpiYVDWUMPNZRfi3XYE_hDs,6972
@@ -152,7 +151,7 @@ ai_edge_torch/odml_torch/lowerings/_batch_norm.py,sha256=PaLI0BB6pdBW1VyfW8VTOT_
 ai_edge_torch/odml_torch/lowerings/_convolution.py,sha256=B6BILeu-UlwGB1O6g7111X1TaIFznsfxXrB72ygBsBA,3885
 ai_edge_torch/odml_torch/lowerings/_jax_lowerings.py,sha256=I0Y4IK7Zap8m6xfxMw7DfQ9Mg4htKOoypdHVAMHqx9c,10669
 ai_edge_torch/odml_torch/lowerings/context.py,sha256=jslcCv7r_HtImSRTxJwHAUV_QCu9Jub51lovmoBkmFA,1295
-ai_edge_torch/odml_torch/lowerings/registry.py,sha256=dcnxq8vV9rxSQqXkjSg9it7l6oP_sdfH8kIZdQNkQ_4,2653
+ai_edge_torch/odml_torch/lowerings/registry.py,sha256=ES3x_RJ22T5rlmMrlomex2DdcZbhlyVJ7_HS3rjz3Uk,2851
 ai_edge_torch/odml_torch/lowerings/utils.py,sha256=NczqpsSd3Fn7yVcPC3qllemiZxxDAZgcW1T5l8-W9fE,5593
 ai_edge_torch/odml_torch/passes/__init__.py,sha256=AVwIwUTMx7rXacKjGy4kwrtMd3XB2v_ncdc40KOjUqQ,1245
 ai_edge_torch/quantize/__init__.py,sha256=aB5dXot04bqyUhpsDFvxt9CIi15QAC4euvqOndJ0XLU,714
@@ -162,8 +161,8 @@ 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.dev20240904.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.3.0.dev20240904.dist-info/METADATA,sha256=hyD9hQtmGBnd2MHwoSkhRnlupk00oncKRsN5PQe8iQk,1878
-ai_edge_torch_nightly-0.3.0.dev20240904.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-ai_edge_torch_nightly-0.3.0.dev20240904.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.3.0.dev20240904.dist-info/RECORD,,
+ai_edge_torch_nightly-0.3.0.dev20240906.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20240906.dist-info/METADATA,sha256=u4yKvulxsV9xZmKSKnNO6L_FE8P_Iy96IZ0UL_voxAE,1859
+ai_edge_torch_nightly-0.3.0.dev20240906.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20240906.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20240906.dist-info/RECORD,,

ai_edge_torch/generative/quantize/ai_edge_quantizer_glue/__init__.py

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