ai-edge-quantizer-nightly 0.4.0.dev20251008__py3-none-any.whl → 0.4.0.dev20251010__py3-none-any.whl

ai_edge_quantizer/default_policy.py

@@ -178,7 +178,6 @@ DEFAULT_JSON_POLICY = """
       "INPUT",
       "OUTPUT",
       "SLICE",
-      "EMBEDDING_LOOKUP",
       "SUM",
       "SELECT",
       "SELECT_V2",
@@ -226,7 +225,6 @@ DEFAULT_JSON_POLICY = """
       "INPUT",
       "OUTPUT",
       "SLICE",
-      "EMBEDDING_LOOKUP",
       "SUM",
       "SELECT",
       "SELECT_V2",
@@ -252,8 +250,8 @@ DEFAULT_JSON_POLICY = """
       "NOT_EQUAL",
       "MIRROR_PAD"
     ],
-    "static_wi4_ai8": ["FULLY_CONNECTED", "CONV_2D", "INPUT", "OUTPUT", "EMBEDDING_LOOKUP"],
-    "static_wi4_ai16": ["FULLY_CONNECTED", "CONV_2D", "INPUT", "OUTPUT", "EMBEDDING_LOOKUP"],
+    "static_wi4_ai8": ["FULLY_CONNECTED", "CONV_2D", "INPUT", "OUTPUT"],
+    "static_wi4_ai16": ["FULLY_CONNECTED", "CONV_2D", "INPUT", "OUTPUT"],
     "dynamic_wi8_afp32": [
       "BATCH_MATMUL",
       "CONV_2D",

ai_edge_quantizer/quantizer.py

@@ -126,12 +126,16 @@ class Quantizer:
     float_model: TFLite model file path or bytearray.
     quantization_recipe: Quantization recipe .json filepath or in loaded json
       format.
+    previous_quantized_model: Optional previously quantized TFLite model file
+      path or bytearray. This is useful for validating a quantized model
+      without quantizing it again.
   """
 
   def __init__(
       self,
       float_model: Union[str, bytearray],
       quantization_recipe: Optional[Union[str, _QuantRecipe]] = None,
+      previous_quantized_model: Optional[Union[str, bytearray]] = None,
   ):
     """Initializes the quantizer.
 
@@ -139,6 +143,9 @@ class Quantizer:
       float_model: Path to the float tflite model.
       quantization_recipe: Quantization recipe in .json filepath or loaded json
         format.
+      previous_quantized_model: Path to an optional previously quantized tflite
+        model. This is useful for validating a quantized model without
+        quantizing it again.
     """
     # Use `float model` as bytes for memory efficiency.
     self.float_model: bytes = (
@@ -146,6 +153,14 @@ class Quantizer:
         if isinstance(float_model, str)
         else float_model
     )
+    if previous_quantized_model is not None:
+      self.previous_quantized_model: bytes = (
+          tfl_flatbuffer_utils.get_model_content(previous_quantized_model)
+          if isinstance(previous_quantized_model, str)
+          else previous_quantized_model
+      )
+    else:
+      self.previous_quantized_model = None
 
     self._recipe_manager: recipe_manager.RecipeManager = (
         recipe_manager.RecipeManager()
@@ -153,6 +168,7 @@ class Quantizer:
     if quantization_recipe is not None:
       self.load_quantization_recipe(quantization_recipe)
     self._result: QuantizationResult = QuantizationResult([{}], None)
+    self._quantize_called = False
 
   def load_quantization_recipe(self, recipe: Union[str, _QuantRecipe]) -> None:
     """Loads a quantization recipe.
@@ -399,7 +415,7 @@ class Quantizer:
     Raises:
       RuntimeError: If quantization recipe is empty.
     """
-
+    self._quantize_called = True
     if calibration_result is not None:
       self._ensure_model_qsv_sufficient(calibration_result)
 
@@ -445,9 +461,16 @@ class Quantizer:
       test_data = tfl_interpreter_utils.create_random_normal_input_data(
           self.float_model, num_samples=1
       )
+    if self._quantize_called:
+      quantized_model = self._result.quantized_model
+    else:
+      quantized_model = self.previous_quantized_model
+
+    if quantized_model is None:
+      raise ValueError('No quantized model available to validate.')
     return model_validator.compare_model(
         self.float_model,
-        self._result.quantized_model,
+        quantized_model,
         test_data,
         error_metrics,
         validation_utils.get_validation_func(error_metrics),

ai_edge_quantizer/quantizer_test.py

@@ -337,6 +337,21 @@ class QuantizerTest(parameterized.TestCase):
         'sequential/dense_1/MatMul', validation_result.intermediate_tensors
     )
 
+  def test_validate_with_quantized_model_arg_succeeds(self):
+    self._quantizer.quantize()
+    quantized_model = self._quantizer._result.quantized_model
+    self.assertIsNotNone(quantized_model)
+
+    new_quantizer = quantizer.Quantizer(
+        self._test_model_path, previous_quantized_model=quantized_model
+    )
+    validation_result = new_quantizer.validate()
+    validation_result = validation_result.get_signature_comparison_result()
+    self.assertIsNotNone(validation_result)
+    self.assertIn(
+        'sequential/dense_1/MatMul', validation_result.intermediate_tensors
+    )
+
   def test_load_custom_policies_succeeds(self):
 
     test_op_config = qtyping.OpQuantizationConfig(

{ai_edge_quantizer_nightly-0.4.0.dev20251008.dist-info → ai_edge_quantizer_nightly-0.4.0.dev20251010.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-quantizer-nightly
-Version: 0.4.0.dev20251008
+Version: 0.4.0.dev20251010
 Summary: A quantizer for advanced developers to quantize converted AI Edge models.
 Home-page: https://github.com/google-ai-edge/ai-edge-quantizer
 Keywords: On-Device ML,AI,Google,TFLite,Quantization,LLMs,GenAI

{ai_edge_quantizer_nightly-0.4.0.dev20251008.dist-info → ai_edge_quantizer_nightly-0.4.0.dev20251010.dist-info}/RECORD

@@ -5,7 +5,7 @@ ai_edge_quantizer/algorithm_manager_api_test.py,sha256=w6bSONvXkX6bzXAGc0-7b6gND
 ai_edge_quantizer/calibrator.py,sha256=Sms7_AIHPH9G5xFaz5Ef3a5gPhxuIWQI8d2LUM8C96I,12071
 ai_edge_quantizer/calibrator_test.py,sha256=ZLzIMWB2FSFU4TOatDioYuwp_kLh8iSCefZ5_Q9FU7s,11900
 ai_edge_quantizer/conftest.py,sha256=SxCz-5LlRD_lQm4hQc4c6IGG7DS8d7IyEWY9gnscPN0,794
-ai_edge_quantizer/default_policy.py,sha256=6eJA0eX5Npv8lw_0EDS5iPldInoURQKEDhDZ272VG1Q,11770
+ai_edge_quantizer/default_policy.py,sha256=DpQUe0ETpRtx2qWwvrj9pFAhfyEwES1mVTTak8kyPzM,11678
 ai_edge_quantizer/model_modifier.py,sha256=U70JByv6CItP8tg4bdyMfX-R3UlwylAGSviZkF_FSAM,10468
 ai_edge_quantizer/model_modifier_test.py,sha256=CV4pgMEQkBJr_qbYR720TO8HBCutbEYLHptDHgdQMUE,7274
 ai_edge_quantizer/model_validator.py,sha256=Hj0_5o-Oa3dSlJ3ryVjRhvsyelHNyek1GrtG9buMczg,13153
@@ -13,8 +13,8 @@ ai_edge_quantizer/model_validator_test.py,sha256=EeqOP_mrZsnZ3rug756s0ryDDqd2KgI
 ai_edge_quantizer/params_generator.py,sha256=0w-sDGk84sVNkXoduon1wDqq30sGOHVgBVbdg44QVF4,20153
 ai_edge_quantizer/params_generator_test.py,sha256=RDYoRZDJfEZRtjlTAU2kZ_4t3JHOqEHxfJX9V4ETAhg,40597
 ai_edge_quantizer/qtyping.py,sha256=rp2jdmCuSsP6Ay8rD7NxDCpbFkRNkbYP29Uwe0xBfnA,17196
-ai_edge_quantizer/quantizer.py,sha256=ckAEOnnBxuCKZuvlzdChevCKPuE-IeDPHCNtFTWr250,17857
-ai_edge_quantizer/quantizer_test.py,sha256=bh4IowxRF249p_XKIKQ0f17PmeDddfcOUzvQ2ht1L0E,26530
+ai_edge_quantizer/quantizer.py,sha256=teYeONdIS31IAY6ubLujCRi1t6lYAd0LkC8dRPxQdbw,18919
+ai_edge_quantizer/quantizer_test.py,sha256=9BVwt7oyM8IsSC7jN73nI0O-4MikBkymm_FigJnSeCM,27117
 ai_edge_quantizer/recipe.py,sha256=MEkfQ2Sg3KAE9LAORHWcbjYNPg06EUbwc1d-VspQA2U,6461
 ai_edge_quantizer/recipe_manager.py,sha256=6l2uq8KL23KLu9OQDmPGkxrFiwHrdDB9xnn-ni8WdEM,15036
 ai_edge_quantizer/recipe_manager_test.py,sha256=qjgGUF-wggXnSXqZ5khmqrDMIQI5CShk52IVWTahq6s,36817
@@ -74,8 +74,8 @@ ai_edge_quantizer/utils/tfl_interpreter_utils.py,sha256=EoVjI_hplX_Rml3hfRsGmQOi
 ai_edge_quantizer/utils/tfl_interpreter_utils_test.py,sha256=6fjkM-rycZ95L4yfvlr0TN6RlrhfPzxNUYrZaYO_F0A,12013
 ai_edge_quantizer/utils/validation_utils.py,sha256=0sOdH4pzk_Pwh1r8O47iaECRng1Xn0ABn9GVc8UPNcY,4994
 ai_edge_quantizer/utils/validation_utils_test.py,sha256=1sblJWHLTYTbn1Qi9rwnrREOSXRy5KwHAWSwgI1e_aU,3697
-ai_edge_quantizer_nightly-0.4.0.dev20251008.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-ai_edge_quantizer_nightly-0.4.0.dev20251008.dist-info/METADATA,sha256=OhwG-gg_4N_0x8rm_4M_-aCX0mBjdPxj__btANE2Sok,1508
-ai_edge_quantizer_nightly-0.4.0.dev20251008.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-ai_edge_quantizer_nightly-0.4.0.dev20251008.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
-ai_edge_quantizer_nightly-0.4.0.dev20251008.dist-info/RECORD,,
+ai_edge_quantizer_nightly-0.4.0.dev20251010.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+ai_edge_quantizer_nightly-0.4.0.dev20251010.dist-info/METADATA,sha256=n_gZ1KzZVjtGwWh9EEnE7lCZ8Qt1AzkUjDbGcmFvs-I,1508
+ai_edge_quantizer_nightly-0.4.0.dev20251010.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+ai_edge_quantizer_nightly-0.4.0.dev20251010.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
+ai_edge_quantizer_nightly-0.4.0.dev20251010.dist-info/RECORD,,