ai-edge-torch-nightly 0.3.0.dev20240814__py3-none-any.whl → 0.3.0.dev20240822__py3-none-any.whl

ai_edge_torch/_convert/test/test_convert_multisig.py

@@ -84,6 +84,9 @@ class TestConvertMultiSignature(googletest.TestCase):
         )
     )
 
+  @googletest.skip(
+      reason="Re-enable once the tflite converter issue is fixed.",
+  )
   def test_convert_mobilenet_v2_signature_helper(self):
     """Tests the ai_edge_torch.signature helper function works."""
     torch_module = torchvision.models.mobilenet_v2().eval()

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/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.dev20240814"
+__version__ = "0.3.0.dev20240822"

{ai_edge_torch_nightly-0.3.0.dev20240814.dist-info → ai_edge_torch_nightly-0.3.0.dev20240822.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20240814
+Version: 0.3.0.dev20240822
 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.dev20240814.dist-info → ai_edge_torch_nightly-0.3.0.dev20240822.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=BlH3JqkXwVHXFYAd5rF04dUvLCthvKVqnfgO3abgh14,706
+ai_edge_torch/version.py,sha256=DjujCBc63P1BCTPwFWV93QP5GPGWMkSuuTWTRIg5YNA,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
@@ -28,7 +28,7 @@ ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/layout_partitio
 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_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=cqXJK1YALC1huw87HJVjrdc9xe0LaahRY8tVm_RsKg4,4817
 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
@@ -122,12 +122,13 @@ ai_edge_torch/hlfb/mark_pattern/__init__.py,sha256=cjTprggj_cuktSCm7-A25e7Shop3k
 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
@@ -136,8 +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.dev20240814.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.3.0.dev20240814.dist-info/METADATA,sha256=eYXq0PpFouGnXKu9vXIzyaXj8XsLDxlDn903GJFR3ak,1885
-ai_edge_torch_nightly-0.3.0.dev20240814.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-ai_edge_torch_nightly-0.3.0.dev20240814.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.3.0.dev20240814.dist-info/RECORD,,
+ai_edge_torch_nightly-0.3.0.dev20240822.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20240822.dist-info/METADATA,sha256=U8torMOW4U2TfZG7j8mwQ3egNgda43VRAtleRwZmyKw,1885
+ai_edge_torch_nightly-0.3.0.dev20240822.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20240822.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20240822.dist-info/RECORD,,