ai-edge-quantizer-nightly 0.1.0.dev20250415__py3-none-any.whl → 0.5.0.dev20260103__py3-none-any.whl

ai_edge_quantizer/quantizer_test.py

@@ -51,6 +51,30 @@ def _get_calibration_data(num_samples: int = 16):
   return calibration_data
 
 
+def _is_all_signature_defs_inputs_float(model_content: bytes):
+  tfl_interpreter = tfl_interpreter_utils.create_tfl_interpreter(model_content)
+  for signature_key in tfl_interpreter.get_signature_list():
+    input_details = tfl_interpreter.get_signature_runner(
+        signature_key
+    ).get_input_details()
+    for tensor_details in input_details.values():
+      if tensor_details['dtype'] != np.float32:
+        return False
+  return True
+
+
+def _is_all_signature_defs_outputs_float(model_content: bytes):
+  tfl_interpreter = tfl_interpreter_utils.create_tfl_interpreter(model_content)
+  for signature_key in tfl_interpreter.get_signature_list():
+    output_details = tfl_interpreter.get_signature_runner(
+        signature_key
+    ).get_output_details()
+    for tensor_details in output_details.values():
+      if tensor_details['dtype'] != np.float32:
+        return False
+  return True
+
+
 class QuantizerTest(parameterized.TestCase):
 
   def setUp(self):
@@ -92,6 +116,76 @@ class QuantizerTest(parameterized.TestCase):
         new_op_config.compute_precision,
     )
 
+  def test_add_dynamic_config_succeeds(self):
+    self._quantizer.load_quantization_recipe(self._test_recipe_path)
+    scope_regex = '.*/Dense/.*'
+    self._quantizer.add_dynamic_config(
+        regex=scope_regex,
+        operation_name=qtyping.TFLOperationName.FULLY_CONNECTED,
+        num_bits=8,
+    )
+    updated_recipe = self._quantizer.get_quantization_recipe()
+    self.assertLen(updated_recipe, 2)
+
+    added_config = updated_recipe[-1]
+    self.assertEqual(added_config['regex'], scope_regex)
+    self.assertEqual(
+        added_config['op_config']['compute_precision'],
+        qtyping.ComputePrecision.INTEGER,
+    )
+    self.assertFalse(added_config['op_config']['explicit_dequantize'])
+    self.assertEqual(
+        added_config['op_config']['weight_tensor_config']['num_bits'], 8
+    )
+
+  def test_add_weight_only_config_succeeds(self):
+    self._quantizer.load_quantization_recipe(self._test_recipe_path)
+    scope_regex = '.*/Dense/.*'
+    self._quantizer.add_weight_only_config(
+        regex=scope_regex,
+        operation_name=qtyping.TFLOperationName.FULLY_CONNECTED,
+        num_bits=4,
+    )
+    updated_recipe = self._quantizer.get_quantization_recipe()
+    self.assertLen(updated_recipe, 2)
+
+    added_config = updated_recipe[-1]
+    self.assertEqual(added_config['regex'], scope_regex)
+    self.assertEqual(
+        added_config['op_config']['compute_precision'],
+        qtyping.ComputePrecision.FLOAT,
+    )
+    self.assertTrue(added_config['op_config']['explicit_dequantize'])
+    self.assertEqual(
+        added_config['op_config']['weight_tensor_config']['num_bits'], 4
+    )
+
+  def test_add_static_config_succeeds(self):
+    self._quantizer.load_quantization_recipe(self._test_recipe_path)
+    scope_regex = '.*/Dense/.*'
+    self._quantizer.add_static_config(
+        regex=scope_regex,
+        operation_name=qtyping.TFLOperationName.FULLY_CONNECTED,
+        activation_num_bits=8,
+        weight_num_bits=4,
+    )
+    updated_recipe = self._quantizer.get_quantization_recipe()
+    self.assertLen(updated_recipe, 2)
+
+    added_config = updated_recipe[-1]
+    self.assertEqual(added_config['regex'], scope_regex)
+    self.assertEqual(
+        added_config['op_config']['compute_precision'],
+        qtyping.ComputePrecision.INTEGER,
+    )
+    self.assertFalse(added_config['op_config']['explicit_dequantize'])
+    self.assertEqual(
+        added_config['op_config']['activation_tensor_config']['num_bits'], 8
+    )
+    self.assertEqual(
+        added_config['op_config']['weight_tensor_config']['num_bits'], 4
+    )
+
   def test_load_quantization_recipe_succeeds(self):
     qt = quantizer.Quantizer(self._test_model_path, None)
     qt.load_quantization_recipe(self._test_recipe_path)
@@ -118,7 +212,7 @@ class QuantizerTest(parameterized.TestCase):
     # Calibrate with empty state.
     calib_data = _get_calibration_data()
     calibration_result = self._quantizer.calibrate(calib_data)
-    self.assertLen(calibration_result, 13)
+    self.assertLen(calibration_result, 7)
 
   @parameterized.parameters(
       'recipes/default_a8w8_recipe.json',
@@ -133,7 +227,7 @@ class QuantizerTest(parameterized.TestCase):
     updated_calibration_result = self._quantizer.calibrate(
         calib_data, previous_calibration_result=calibration_result
     )
-    self.assertLen(updated_calibration_result, 13)
+    self.assertLen(updated_calibration_result, 7)
     self.assertNotEqual(
         calibration_result['StatefulPartitionedCall:0'],
         updated_calibration_result['StatefulPartitionedCall:0'],
@@ -215,6 +309,44 @@ class QuantizerTest(parameterized.TestCase):
       saved_recipe = json.load(json_file)
     self.assertEqual(saved_recipe, self._test_recipe)
 
+  def test_saved_legacy_recipe_lacks_block_size(self):
+    model_name = 'test_model'
+    legacy_recipe_path = os.path.join(
+        TEST_DATA_PREFIX_PATH,
+        'recipes/dynamic_legacy_wi8_afp32_recipe.json',
+    )
+    self._quantizer.load_quantization_recipe(legacy_recipe_path)
+    result = self._quantizer.quantize()
+    result.save(self._tmp_save_path, model_name)
+    saved_recipe_path = os.path.join(
+        self._tmp_save_path, model_name + '_recipe.json'
+    )
+    with open(saved_recipe_path) as json_file:
+      saved_recipe = json.load(json_file)
+    with open(legacy_recipe_path) as json_file:
+      legacy_recipe = json.load(json_file)
+
+    self.assertNotEqual(saved_recipe, legacy_recipe)
+
+    # Verify that the default test recipe contains 'block_size'.
+    has_block_size = False
+    for config in legacy_recipe:
+      op_config = config.get('op_config')
+      if op_config:
+        weight_config = op_config.get('weight_tensor_config')
+        if weight_config and 'block_size' in weight_config:
+          has_block_size = True
+          break
+    self.assertTrue(has_block_size)
+
+    # Verify that the saved recipe does not have 'block_size'.
+    for config in saved_recipe:
+      op_config = config.get('op_config')
+      if op_config:
+        weight_config = op_config.get('weight_tensor_config')
+        if weight_config:
+          self.assertNotIn('block_size', weight_config)
+
   def test_save_no_quantize_raise_error(self):
     error_message = 'No quantized model to save.'
     with self.assertRaisesWithPredicateMatch(
@@ -243,6 +375,34 @@ class QuantizerTest(parameterized.TestCase):
         'sequential/dense_1/MatMul', validation_result.intermediate_tensors
     )
 
+  def test_validate_output_tensors_only_succeeds(self):
+    self._quantizer.quantize()
+    validation_result = self._quantizer.validate(
+        validate_output_tensors_only=True
+    )
+    validation_result = validation_result.get_signature_comparison_result()
+    self.assertIsNotNone(validation_result)
+    self.assertEmpty(validation_result.input_tensors)
+    self.assertEmpty(validation_result.constant_tensors)
+    self.assertEmpty(validation_result.intermediate_tensors)
+    self.assertNotEmpty(validation_result.output_tensors)
+    self.assertIn('StatefulPartitionedCall:0', validation_result.output_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(
@@ -284,6 +444,33 @@ class QuantizerTest(parameterized.TestCase):
         op_config=test_op_config,
     )
 
+  def test_two_pass_quantization_with_conv_and_fc_succeeds(self):
+    float_model_path = self._test_model_path
+
+    drq_recipe_path = os.path.join(
+        TEST_DATA_PREFIX_PATH, 'recipes/dynamic_wi8_afp32_hadamard_recipe.json'
+    )
+    drq_quantizer = quantizer.Quantizer(float_model_path)
+    drq_quantizer.load_quantization_recipe(drq_recipe_path)
+    drq_result = drq_quantizer.quantize()
+    drq_model_path = os.path.join(self._tmp_save_path, 'drq_model.tflite')
+    drq_result.export_model(drq_model_path)
+
+    srq_recipe_path = os.path.join(
+        TEST_DATA_PREFIX_PATH, 'recipes/default_a8w8_recipe.json'
+    )
+    srq_quantizer = quantizer.Quantizer(drq_model_path)
+    srq_quantizer.load_quantization_recipe(srq_recipe_path)
+    representative_dataset = (
+        tfl_interpreter_utils.create_random_normal_input_data(
+            drq_model_path, num_samples=1
+        )
+    )
+    calibration_result = srq_quantizer.calibrate(representative_dataset)
+    srq_result = srq_quantizer.quantize(calibration_result)
+    srq_model_path = os.path.join(self._tmp_save_path, 'srq_model.tflite')
+    srq_result.export_model(srq_model_path)
+
 
 class QuantizerBytearrayInputs(googletest.TestCase):
 
@@ -426,14 +613,12 @@ class QuantizerMultiSignatureModelTest(parameterized.TestCase):
                     'symmetric': False,
                     'granularity': 'TENSORWISE',
                     'dtype': 'INT',
-                    'block_size': 0,
                 },
                 'weight_tensor_config': {
                     'num_bits': 8,
                     'symmetric': True,
                     'granularity': 'CHANNELWISE',
                     'dtype': 'INT',
-                    'block_size': 0,
                 },
                 'compute_precision': 'INTEGER',
                 'explicit_dequantize': False,
@@ -454,8 +639,7 @@ class QuantizerMultiSignatureModelTest(parameterized.TestCase):
 
     # Quantize and expect an error about missing signature in calibration data.
     error_message = (
-        'Missing QSVs (min/max) for tensor multiply_x:0 in Signature'
-        " 'multiply'."
+        'MUL(index: 0) not found in tensor_name_to_qsv'
     )
     with self.assertRaisesWithPredicateMatch(
         ValueError, lambda err: error_message in str(err)
@@ -477,21 +661,21 @@ class QuantizerToyGemma2Test(parameterized.TestCase):
         'signature_1': [{
             'cache_0': _RNG.random(size=(1, 100, 4, 4), dtype=np.float32),
             'cache_1': _RNG.random(size=(1, 100, 4, 4), dtype=np.float32),
-            'positions': _RNG.integers(low=0, high=10, size=(1, 100)).astype(
-                np.int32
+            'positions': (
+                _RNG.integers(low=0, high=10, size=(1, 100)).astype(np.int32)
             ),
-            'tokens': _RNG.integers(low=0, high=10, size=(1, 100)).astype(
-                np.int32
+            'tokens': (
+                _RNG.integers(low=0, high=10, size=(1, 100)).astype(np.int32)
             ),
         }],
         'signature_2': [{
             'cache_0': _RNG.random(size=(1, 100, 4, 4), dtype=np.float32),
             'cache_1': _RNG.random(size=(1, 100, 4, 4), dtype=np.float32),
-            'positions': _RNG.integers(low=0, high=10, size=(1, 100)).astype(
-                np.int32
+            'positions': (
+                _RNG.integers(low=0, high=10, size=(1, 100)).astype(np.int32)
             ),
-            'tokens': _RNG.integers(low=0, high=10, size=(1, 100)).astype(
-                np.int32
+            'tokens': (
+                _RNG.integers(low=0, high=10, size=(1, 100)).astype(np.int32)
             ),
         }],
     }
@@ -508,8 +692,8 @@ class QuantizerToyGemma2Test(parameterized.TestCase):
     )
 
     self._quantizer.update_quantization_recipe(
-        regex='StatefulPartitionedCall',
-        operation_name=qtyping.TFLOperationName.FULLY_CONNECTED,
+        regex='.*',
+        operation_name=qtyping.TFLOperationName.OUTPUT,
         algorithm_key=_AlgorithmName.NO_QUANTIZE,
     )
 
@@ -521,6 +705,90 @@ class QuantizerToyGemma2Test(parameterized.TestCase):
     self._quantizer.quantize(calib_result)
     self.assertIsNotNone(self._quantizer._result.quantized_model)
 
+  def test_toy_gemma2_update_signature_defs_succeeds(self):
+
+    self.assertTrue(
+        _is_all_signature_defs_outputs_float(
+            open(self._test_model_path, 'rb').read()
+        )
+    )
+    calib_result = self._quantizer.calibrate(
+        self._toy_gemma2_calibration_dataset
+    )
+    self.assertIsNotNone(calib_result)
+    self._quantizer.quantize(calib_result)
+    self.assertIsNotNone(self._quantizer._result.quantized_model)
+    self.assertTrue(
+        _is_all_signature_defs_outputs_float(
+            self._quantizer._result.quantized_model
+        )
+    )
+
+
+class QuantizerFullyConnectedTest(parameterized.TestCase):
+
+  def setUp(self):
+    super().setUp()
+    self._tmp_save_path = self.create_tempdir().full_path
+    self._test_model_path = os.path.join(
+        TEST_DATA_PREFIX_PATH,
+        'tests/models/single_fc.tflite',
+    )
+
+    self._test_recipe_path = os.path.join(
+        TEST_DATA_PREFIX_PATH,
+        'recipes/default_a8w8_recipe.json',
+    )
+    with open(self._test_recipe_path) as json_file:
+      self._test_recipe = json.load(json_file)
+
+    self._quantizer = quantizer.Quantizer(
+        self._test_model_path, self._test_recipe_path
+    )
+
+    self._quantizer.update_quantization_recipe(
+        regex='.*',
+        operation_name=qtyping.TFLOperationName.INPUT,
+        algorithm_key=_AlgorithmName.NO_QUANTIZE,
+    )
+    self._quantizer.update_quantization_recipe(
+        regex='.*',
+        operation_name=qtyping.TFLOperationName.OUTPUT,
+        algorithm_key=_AlgorithmName.NO_QUANTIZE,
+    )
+
+  def test_fully_connected_quantization_succeeds(self):
+    calib_result = self._quantizer.calibrate(
+        tfl_interpreter_utils.create_random_normal_input_data(
+            self._test_model_path, num_samples=4
+        )
+    )
+    self.assertIsNotNone(calib_result)
+    self._quantizer.quantize(calib_result)
+    self.assertIsNotNone(self._quantizer._result.quantized_model)
+
+  def test_fully_connected_quantization_update_signature_defs_succeeds(self):
+
+    model_content = open(self._test_model_path, 'rb').read()
+    self.assertTrue(_is_all_signature_defs_inputs_float(model_content))
+    self.assertTrue(_is_all_signature_defs_outputs_float(model_content))
+
+    calib_result = self._quantizer.calibrate(
+        tfl_interpreter_utils.create_random_normal_input_data(
+            self._test_model_path, num_samples=4
+        )
+    )
+    self.assertIsNotNone(calib_result)
+    quant_result = self._quantizer.quantize(calib_result)
+    self.assertIsNotNone(quant_result.quantized_model)
+
+    self.assertTrue(
+        _is_all_signature_defs_inputs_float(quant_result.quantized_model)
+    )
+    self.assertTrue(
+        _is_all_signature_defs_outputs_float(quant_result.quantized_model)
+    )
+
 
 if __name__ == '__main__':
   googletest.main()

ai_edge_quantizer/recipe.py

@@ -15,51 +15,163 @@
 
 """Quantization recipe module."""
 
+from ai_edge_quantizer import algorithm_manager
+from ai_edge_quantizer import qtyping
+from ai_edge_quantizer import recipe_manager
 
-def dynamic_wi8_afp32():
-  """Returns a dynamic quantization recipe with int8 weights and float32 activation."""
-  return [
-      dict({
-          'regex': '.*',
-          'operation': '*',
-          'algorithm_key': 'min_max_uniform_quantize',
-          'op_config': {
-              'weight_tensor_config': {
-                  'num_bits': 8,
-                  'symmetric': True,
-                  'granularity': 'CHANNELWISE',
-                  'dtype': 'INT',
-                  'block_size': 0,
-              },
-              'compute_precision': 'INTEGER',
-              'explicit_dequantize': False,
-              'skip_checks': False,
-          },
-      })
-  ]
+AlgorithmName = algorithm_manager.AlgorithmName
 
 
-def dynamic_wi4_afp32():
-  """Returns a dynamic quantization recipe with int4 weights and float32 activation."""
-  return [
-      dict({
-          'regex': '.*',
-          'operation': '*',
-          'algorithm_key': 'min_max_uniform_quantize',
-          'op_config': {
-              'weight_tensor_config': {
-                  'num_bits': 4,
-                  'symmetric': True,
-                  'granularity': 'CHANNELWISE',
-                  'dtype': 'INT',
-                  'block_size': 0,
-              },
-              'compute_precision': 'INTEGER',
-              'explicit_dequantize': False,
-              'skip_checks': False,
-          },
-      })
-  ]
+def dynamic_wi8_afp32(
+    algorithm_key: AlgorithmName = AlgorithmName.MIN_MAX_UNIFORM_QUANT,
+):
+  """Returns a dynamic quantization recipe with int8 weights and float32 activation.
+
+  All supported ops will be quantized with int8 weights and float32 activations,
+  which will be dynamically quantized to int8 during inference to enable int8
+  compute. The model quality may suffer due to the on-the-fly quantization. If
+  quality is a concern, consider using weight-only quantization.
+
+  Args:
+    algorithm_key: The algorithm to use for quantization.
+
+  Returns:
+    A dynamic quantization recipe.
+  """
+  rp_manager = recipe_manager.RecipeManager()
+  rp_manager.add_dynamic_config(
+      regex='.*',
+      operation_name=qtyping.TFLOperationName.ALL_SUPPORTED,
+      num_bits=8,
+      algorithm_key=algorithm_key,
+  )
+  return rp_manager.get_quantization_recipe()
+
+
+def dynamic_wi4_afp32(
+    algorithm_key: AlgorithmName = AlgorithmName.MIN_MAX_UNIFORM_QUANT,
+):
+  """Returns a dynamic quantization recipe with int4 weights and float32 activation.
+
+  All supported ops will be quantized with int4 weights and float32 activations,
+  which will be dynamically quantized to int4 during inference to enable int4
+  compute.
+
+  Args:
+    algorithm_key: The algorithm to use for quantization.
+
+  Returns:
+    A dynamic quantization recipe.
+  """
+  rp_manager = recipe_manager.RecipeManager()
+  rp_manager.add_dynamic_config(
+      regex='.*',
+      operation_name=qtyping.TFLOperationName.ALL_SUPPORTED,
+      num_bits=4,
+      algorithm_key=algorithm_key,
+  )
+  return rp_manager.get_quantization_recipe()
+
+
+def weight_only_wi8_afp32(
+    algorithm_key: AlgorithmName = AlgorithmName.MIN_MAX_UNIFORM_QUANT,
+):
+  """Returns a weight-only quantization recipe with int8 weights and float32 activation.
+
+  All supported ops will be quantized with int8 weights and float32 activations.
+  The weights will be explicitly dequantized before being fed into the op to
+  enable float compute thus retain model quality. If latency is a concern,
+  consider using dynamic range quantization.
+
+  Args:
+    algorithm_key: The algorithm to use for quantization.
+
+  Returns:
+    A weight-only quantization recipe.
+  """
+  rp_manager = recipe_manager.RecipeManager()
+  rp_manager.add_weight_only_config(
+      regex='.*',
+      operation_name=qtyping.TFLOperationName.ALL_SUPPORTED,
+      num_bits=8,
+      algorithm_key=algorithm_key,
+  )
+  return rp_manager.get_quantization_recipe()
+
+
+def weight_only_wi4_afp32(
+    algorithm_key: AlgorithmName = AlgorithmName.MIN_MAX_UNIFORM_QUANT,
+):
+  """Returns a weight-only quantization recipe with int4 weights and float32 activation.
+
+  All supported ops will be quantized with int4 weights and float32 activations.
+  The weights will be explicitly dequantized before being fed into the op to
+  enable float compute thus retain model quality.
+
+  Args:
+    algorithm_key: The algorithm to use for quantization.
+
+  Returns:
+    A weight-only quantization recipe.
+  """
+  rp_manager = recipe_manager.RecipeManager()
+  rp_manager.add_weight_only_config(
+      regex='.*',
+      operation_name=qtyping.TFLOperationName.ALL_SUPPORTED,
+      num_bits=4,
+      algorithm_key=algorithm_key,
+  )
+  return rp_manager.get_quantization_recipe()
+
+
+def static_wi8_ai8(
+    algorithm_key: AlgorithmName = AlgorithmName.MIN_MAX_UNIFORM_QUANT,
+):
+  """Returns a static quantization recipe with int8 weights and int8 activations.
+
+  All supported ops will be quantized with int8 weights and int8 activations.
+  Calibration is needed to use this recipe.
+
+  Args:
+    algorithm_key: The algorithm to use for quantization.
+
+  Returns:
+    A static quantization recipe.
+  """
+  rp_manager = recipe_manager.RecipeManager()
+  rp_manager.add_static_config(
+      regex='.*',
+      operation_name=qtyping.TFLOperationName.ALL_SUPPORTED,
+      activation_num_bits=8,
+      weight_num_bits=8,
+      algorithm_key=algorithm_key,
+  )
+  return rp_manager.get_quantization_recipe()
+
+
+def static_wi8_ai16(
+    algorithm_key: AlgorithmName = AlgorithmName.MIN_MAX_UNIFORM_QUANT,
+):
+  """Returns a static quantization recipe with int8 weights and int16 activations.
+
+  All supported ops will be quantized with int8 weights and int16 activations.
+  Calibration is needed to use this recipe.
+
+  Args:
+    algorithm_key: The algorithm to use for quantization.
+
+  Returns:
+    A static quantization recipe.
+  """
+  rp_manager = recipe_manager.RecipeManager()
+  rp_manager.add_static_config(
+      regex='.*',
+      operation_name=qtyping.TFLOperationName.ALL_SUPPORTED,
+      activation_num_bits=16,
+      weight_num_bits=8,
+      algorithm_key=algorithm_key,
+  )
+  return rp_manager.get_quantization_recipe()
 
 
 def dynamic_legacy_wi8_afp32():