ai-edge-torch-nightly 0.3.0.dev20241003__py3-none-any.whl → 0.3.0.dev20241005__py3-none-any.whl

ai_edge_torch/generative/examples/stable_diffusion/clip.py

@@ -75,9 +75,7 @@ class CLIP(nn.Module):
     )
 
   @torch.inference_mode
-  def forward(self, tokens: torch.LongTensor) -> torch.FloatTensor:
-    tokens = tokens.type(torch.int)
-
+  def forward(self, tokens: torch.IntTensor) -> torch.FloatTensor:
     state = self.tok_embedding(tokens) + self.tok_embedding_position
     for layer in self.transformer_blocks:
       state = layer(state, mask=self.mask_cache)

ai_edge_torch/generative/examples/stable_diffusion/convert_to_tflite.py

@@ -13,47 +13,54 @@
 # limitations under the License.
 # ==============================================================================
 
-import argparse
 import os
-from pathlib import Path
-from typing import Optional
+import pathlib
 
+from absl import app
+from absl import flags
 import ai_edge_torch
-import ai_edge_torch.generative.examples.stable_diffusion.clip as clip
-import ai_edge_torch.generative.examples.stable_diffusion.decoder as decoder
-import ai_edge_torch.generative.examples.stable_diffusion.diffusion as diffusion
-from ai_edge_torch.generative.examples.stable_diffusion.encoder import Encoder
-import ai_edge_torch.generative.examples.stable_diffusion.util as util
+from ai_edge_torch.generative.examples.stable_diffusion import clip
+from ai_edge_torch.generative.examples.stable_diffusion import decoder
+from ai_edge_torch.generative.examples.stable_diffusion import diffusion
+from ai_edge_torch.generative.examples.stable_diffusion import util
 from ai_edge_torch.generative.quantize import quant_recipes
-import ai_edge_torch.generative.utilities.stable_diffusion_loader as stable_diffusion_loader
+from ai_edge_torch.generative.utilities import stable_diffusion_loader
 import torch
 
-arg_parser = argparse.ArgumentParser()
-arg_parser.add_argument(
-    '--clip_ckpt',
-    type=str,
+_CLIP_CKPT = flags.DEFINE_string(
+    'clip_ckpt',
+    None,
     help='Path to source CLIP model checkpoint',
     required=True,
 )
-arg_parser.add_argument(
-    '--diffusion_ckpt',
-    type=str,
+
+_DIFFUSION_CKPT = flags.DEFINE_string(
+    'diffusion_ckpt',
+    None,
     help='Path to source diffusion model checkpoint',
     required=True,
 )
-arg_parser.add_argument(
-    '--decoder_ckpt',
-    type=str,
+
+_DECODER_CKPT = flags.DEFINE_string(
+    'decoder_ckpt',
+    None,
     help='Path to source image decoder model checkpoint',
     required=True,
 )
-arg_parser.add_argument(
-    '--output_dir',
-    type=str,
+
+_OUTPUT_DIR = flags.DEFINE_string(
+    'output_dir',
+    None,
     help='Path to the converted TF Lite directory.',
     required=True,
 )
 
+_QUANTIZE = flags.DEFINE_bool(
+    'quantize',
+    help='Whether to quantize the model during conversion.',
+    default=True,
+)
+
 
 @torch.inference_mode
 def convert_stable_diffusion_to_tflite(
@@ -111,7 +118,7 @@ def convert_stable_diffusion_to_tflite(
   time_embedding = util.get_time_embedding(timestamp)
 
   if not os.path.exists(output_dir):
-    Path(output_dir).mkdir(parents=True, exist_ok=True)
+    pathlib.Path(output_dir).mkdir(parents=True, exist_ok=True)
 
   quant_config = (
       quant_recipes.full_int8_weight_only_recipe() if quantize else None
@@ -142,14 +149,15 @@ def convert_stable_diffusion_to_tflite(
   ).export(f'{output_dir}/decoder.tflite')
 
 
-if __name__ == '__main__':
-  args = arg_parser.parse_args()
+def main(_):
   convert_stable_diffusion_to_tflite(
-      output_dir=args.output_dir,
-      clip_ckpt_path=args.clip_ckpt,
-      diffusion_ckpt_path=args.diffusion_ckpt,
-      decoder_ckpt_path=args.decoder_ckpt,
-      image_height=512,
-      image_width=512,
-      quantize=True,
+      output_dir=_OUTPUT_DIR.value,
+      clip_ckpt_path=_CLIP_CKPT.value,
+      diffusion_ckpt_path=_DIFFUSION_CKPT.value,
+      decoder_ckpt_path=_DECODER_CKPT.value,
+      quantize=_QUANTIZE.value,
   )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/test/test_model_conversion.py

@@ -43,31 +43,40 @@ class TestModelConversion(googletest.TestCase):
         )
     )
 
-  def _test_model_with_kv_cache(self, config, pytorch_model):
+  def _get_params(self, enable_hlfb: bool):
+    """Returns a model, edge model and the kwargs to use for testing."""
+    config = toy_model_with_kv_cache.get_model_config()
+    config.enable_hlfb = enable_hlfb
+    pytorch_model = toy_model_with_kv_cache.ToyModelWithKVCache(config).eval()
     tokens, input_pos = torch.tensor([[1]], dtype=torch.int), torch.tensor(
         [10], dtype=torch.int
     )
     kv = kv_cache.KVCache.from_model_config(config)
+    kwargs = {
+        "tokens": tokens,
+        "input_pos": input_pos,
+        "kv_cache": kv,
+    }
 
     edge_model = ai_edge_torch.convert(
         pytorch_model,
-        sample_kwargs={
-            "tokens": tokens,
-            "input_pos": input_pos,
-            "kv_cache": kv,
-        },
+        sample_kwargs=kwargs,
     )
     edge_model.set_interpreter_builder(
         self._interpreter_builder(edge_model.tflite_model())
     )
+    return pytorch_model, edge_model, kwargs
+
+  def _test_model_with_kv_cache(self, enable_hlfb: bool):
+    pytorch_model, edge_model, kwargs = self._get_params(enable_hlfb)
 
     self.assertTrue(
         test_utils.compare_tflite_torch(
             edge_model,
             pytorch_model,
-            tokens,
-            input_pos,
-            kv,
+            kwargs["tokens"],
+            kwargs["input_pos"],
+            kwargs["kv_cache"],
             signature_name="serving_default",
             atol=1e-5,
             rtol=1e-5,
@@ -79,19 +88,31 @@ class TestModelConversion(googletest.TestCase):
       reason="tests with custom ops are not supported on oss",
   )
   def test_toy_model_with_kv_cache(self):
-    config = toy_model_with_kv_cache.get_model_config()
-    pytorch_model = toy_model_with_kv_cache.ToyModelWithKVCache(config).eval()
-    self._test_model_with_kv_cache(config, pytorch_model)
+    self._test_model_with_kv_cache(enable_hlfb=False)
 
   @googletest.skipIf(
       ai_edge_config.Config.use_torch_xla,
       reason="tests with custom ops are not supported on oss",
   )
   def test_toy_model_with_kv_cache_with_hlfb(self):
-    config = toy_model_with_kv_cache.get_model_config()
-    config.enable_hlfb = True
-    pytorch_model = toy_model_with_kv_cache.ToyModelWithKVCache(config).eval()
-    self._test_model_with_kv_cache(config, pytorch_model)
+    self._test_model_with_kv_cache(enable_hlfb=True)
+
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_toy_model_has_ekv_op(self):
+    """Tests that the model has the external kv cache op."""
+    _, edge_model, _ = self._get_params(enable_hlfb=True)
+    interpreter_ = interpreter.InterpreterWithCustomOps(
+        custom_op_registerers=["GenAIOpsRegisterer"],
+        model_content=edge_model.tflite_model(),
+        experimental_default_delegate_latest_features=True,
+    )
+
+    # pylint: disable=protected-access
+    op_names = [op["op_name"] for op in interpreter_._get_ops_details()]
+    self.assertIn("odml.update_external_kv_cache", op_names)
 
   def _test_multisig_model(self, config, pytorch_model, atol, rtol):
     # prefill

ai_edge_torch/lowertools/translate_recipe.py

@@ -156,8 +156,8 @@ def translate_to_ai_edge_recipe(
 
 
 def quantize_model(
-    model: bytearray, recipe: quantizer.recipe_manager.ModelQuantizationRecipe
+    model: bytes, recipe: quantizer.recipe_manager.ModelQuantizationRecipe
 ) -> bytearray:
-  qt = quantizer.Quantizer(bytearray(model), recipe)
+  qt = quantizer.Quantizer(model, recipe)
   result = qt.quantize()
   return result.quantized_model

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.3.0.dev20241003"
+__version__ = "0.3.0.dev20241005"

{ai_edge_torch_nightly-0.3.0.dev20241003.dist-info → ai_edge_torch_nightly-0.3.0.dev20241005.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20241003
+Version: 0.3.0.dev20241005
 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.dev20241003.dist-info → ai_edge_torch_nightly-0.3.0.dev20241005.dist-info}/RECORD

@@ -3,7 +3,7 @@ ai_edge_torch/config.py,sha256=FMWeCH2b7HYILBvaI1iZNnYCO4WAhDOwBZBmIE-xrF0,909
 ai_edge_torch/conftest.py,sha256=r0GTrhMRhlmOGrrkvumHN8hkmyug6WvF60vWq8wRIBI,758
 ai_edge_torch/fx_pass_base.py,sha256=D86Gw3pIRcpnTebUPKlnPbPGJae1S6Fw4DZZ3ZkD0zw,3730
 ai_edge_torch/model.py,sha256=N-pNpTxzhaFGhWhnSGd70lBzb9VlEhTOq5mddU7bvvI,5542
-ai_edge_torch/version.py,sha256=WKaZCocAyLb42oFdC07BQ6qpSfohXBwt-HKGV7S2fXw,706
+ai_edge_torch/version.py,sha256=y5TOP0Z8qFsjIuJuJtSmzOUpHyTa9UH46RdJjtRWYQA,706
 ai_edge_torch/_convert/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/_convert/conversion.py,sha256=5uPwHhmc6kwiIz-CqaiHDejf2SOWMHrb-rYEHm69wKc,3801
 ai_edge_torch/_convert/conversion_utils.py,sha256=Sr8qXVcTwc-ZnZmK7yxVrIOOp1S_vNrwzC0zUvLTI2o,2160
@@ -71,8 +71,8 @@ ai_edge_torch/generative/examples/smollm/smollm.py,sha256=M5qAcSUE5gxOSfq24a8lZk
 ai_edge_torch/generative/examples/smollm/verify.py,sha256=HXYcCjDJMylVL3Pc9HU-UXqtpjtIU25o1YhPiX30aPU,2361
 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=lwWrKY1NpnbvHQRenpltVN65QlzjWmSScl5CLSipBkc,6110
-ai_edge_torch/generative/examples/stable_diffusion/convert_to_tflite.py,sha256=vfMGI03UL_gfB561t2kzIHuScwnsUmqaPWxgvq_1T5A,5043
+ai_edge_torch/generative/examples/stable_diffusion/clip.py,sha256=5M4auM33SgCTODt0VT8TO-EVILruqGDRiNILBPeB83Y,6072
+ai_edge_torch/generative/examples/stable_diffusion/convert_to_tflite.py,sha256=i9mcBITt4jJqKLA4Qdt3uFotCrglv14tPg8VnqsVnaI,5004
 ai_edge_torch/generative/examples/stable_diffusion/decoder.py,sha256=ClXNntmh0PF3s6U3C7SW3tyVrsSSrV2kyz-_RF4BcqA,15715
 ai_edge_torch/generative/examples/stable_diffusion/diffusion.py,sha256=mBEAUYjV1qDJy9ZAsHtm9RGce0Mbzv0VoPZpdcQl1mk,33730
 ai_edge_torch/generative/examples/stable_diffusion/encoder.py,sha256=CAPsW84A8f00nS6fLFeh_XUjCPsDCA5UxHOUsMrLfSU,3450
@@ -122,7 +122,7 @@ ai_edge_torch/generative/quantize/supported_schemes.py,sha256=FjdycEOvxRgBmQdZVu
 ai_edge_torch/generative/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/test/test_kv_cache.py,sha256=W6Bh0gYDzmwb0j9HdD5_D7Z7FPToP2HSyFrmwIXuFqo,3793
 ai_edge_torch/generative/test/test_loader.py,sha256=9mQUeeZKOVApOWSWl2cN9c10axZjMKM1-0Zd823CCS4,3449
-ai_edge_torch/generative/test/test_model_conversion.py,sha256=-qB-JEIfPFNlpGyJA1TYo_5fawTdyf1C6ee8cP4kYOY,5530
+ai_edge_torch/generative/test/test_model_conversion.py,sha256=a4TzSw8KMxEafirxqkykZi-WgTs5Z7wHp-J1AfjRDzA,6353
 ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=bVCm_mubuGszCBON6oRjQXcBgPZqlVmmOaLWwhZJLio,9060
 ai_edge_torch/generative/test/test_quantize.py,sha256=8geJhKwYBU20m0mdGPD1BUFwQ0lZKNtCB04SOLO18y4,5980
 ai_edge_torch/generative/test/utils.py,sha256=YvEhO2HIj1LkBs5du1UxY-cGRW9HMyAYsOUhgsTrTpA,1796
@@ -147,7 +147,7 @@ ai_edge_torch/lowertools/common_utils.py,sha256=Z7p-ivOHtddktpnHrlDm_dSoTxJOdEjF
 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=S7RWzauts-15xP6VYuM3aAd9cyAGHstYD2A4dlv3d30,9059
-ai_edge_torch/lowertools/translate_recipe.py,sha256=DNzD0VD35YZDqiZjAF1IyIPSzUGPDpE0jvFCCYIzpnc,5667
+ai_edge_torch/lowertools/translate_recipe.py,sha256=ymkBpFqAUiupRWqrPOWiVphKcXR1K5vHK0RjgBFtxlE,5652
 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
@@ -180,8 +180,8 @@ 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.dev20241003.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.3.0.dev20241003.dist-info/METADATA,sha256=a6Q1LozCx-4NWkm1EKZJFeCJTYiTNUSigoVwRevV0oc,1897
-ai_edge_torch_nightly-0.3.0.dev20241003.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-ai_edge_torch_nightly-0.3.0.dev20241003.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.3.0.dev20241003.dist-info/RECORD,,
+ai_edge_torch_nightly-0.3.0.dev20241005.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20241005.dist-info/METADATA,sha256=O3P5ofz2aERMO1xbvIC7Z4RWsUNLJOZgn4pxEH3ftRc,1897
+ai_edge_torch_nightly-0.3.0.dev20241005.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20241005.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20241005.dist-info/RECORD,,