ai-edge-quantizer-nightly 0.0.1.dev20250221__py3-none-any.whl → 0.0.1.dev20250222__py3-none-any.whl

ai_edge_quantizer/algorithms/uniform_quantize/dequantized_weight_recovery.py

@@ -15,11 +15,13 @@
 
 """Recovers quantized weights from dequantized weights (often from QAT)."""
 
-from typing import Optional
+import dataclasses
+from typing import Any, Optional
 import numpy as np
 from ai_edge_quantizer import qtyping
+from ai_edge_quantizer.algorithms.uniform_quantize import naive_min_max_quantize
 from ai_edge_quantizer.algorithms.uniform_quantize import uniform_quantize_tensor
-
+from ai_edge_quantizer.algorithms.utils import common_utils
 
 ALGORITHM_KEY = "dequantized_weight_recovery"
 _TFLOpName = qtyping.TFLOperationName
@@ -27,24 +29,26 @@ _QuantTransformation = qtyping.QuantTransformation
 _IntType = uniform_quantize_tensor.IntType
 
 
-def _validate_recovered_scale(
-    dequant_vals: np.ndarray, scale: np.ndarray, tol: float = 1e-4
+def _validate_recovered_weights(
+    original_vals: np.ndarray,
+    quant_vals: np.ndarray,
+    scale: np.ndarray,
+    tol: float = 1e-4,
 ):
-  """Validates if the recovered quantized values match the dequantized values.
+  """Validates if recovered weights (from the quantized values) are close enough to the original ones.
 
   Args:
-      dequant_vals: The dequantized weight values.
-      scale: The scale values.
-      tol: The tolerance for the difference between the recovered and original
-        values.
+    original_vals: Original values before quantization.
+    quant_vals: Quantized values.
+    scale: Scale used for quantization.
+    tol: Tolerance for the difference between original and recovered values.
 
   Raises:
-      RuntimeError: If the maximum difference between the recovered and
-        original values exceeds the tolerance.
+    RuntimeError: If the maximum difference between original and recovered
+    values exceeds the tolerance.
   """
-  quant_vals = np.round(dequant_vals / scale)  # no need to clamp.
   recovered_vals = quant_vals * scale
-  diff = np.abs(recovered_vals - dequant_vals).flatten()
+  diff = np.abs(recovered_vals - original_vals).flatten()
   max_diff = diff.max()
   if max_diff > tol:
     raise RuntimeError(
@@ -127,5 +131,120 @@ def get_zp_scale_from_2d_dequantized_symmetric_weights(
     )
 
   zero_points = np.zeros_like(scales, dtype=np.int32)
-  _validate_recovered_scale(dequant_vals, scales)
   return zero_points, scales
+
+
+def get_tensor_quant_params(
+    op_info: qtyping.OpInfo,
+    tensor_quant_config: qtyping.TensorQuantizationConfig,
+    tensor_content: Optional[np.ndarray] = None,
+    tensor_qsv: Optional[dict[str, Any]] = None,
+) -> qtyping.UniformQuantParams:
+  """Get the quantization parameters for a tensor.
+
+  Args:
+    op_info: Aggregated information about the op (e.g., quantization config).
+    tensor_quant_config: The quantization config for the tensor.
+    tensor_content: The content of the tensor.
+    tensor_qsv: A dictionary containing the min/max of the tensor.
+
+  Returns:
+    The quantization parameters for the tensor.
+
+  Raises:
+    ValueError: If the quantization granularity is blockwise, or if the tensor
+    is not a 2D symmetric weight tensor.
+  """
+  # Fallback to naive_min_max_quantize.py for non-weight tensors.
+  if tensor_content is None:
+    return naive_min_max_quantize.get_tensor_quant_params(
+        op_info, tensor_quant_config, tensor_content, tensor_qsv
+    )
+
+  if tensor_quant_config.granularity == qtyping.QuantGranularity.BLOCKWISE:
+    raise ValueError(
+        "Blockwise quantization is not supported for dequantized weight"
+        " recovery."
+    )
+  if tensor_content.ndim != 2 or not tensor_quant_config.symmetric:
+    raise ValueError(
+        "Only 2D symmetric weights are supported for dequantized weight"
+        " recovery."
+    )
+
+  quantized_dim = None
+  if tensor_quant_config.granularity == qtyping.QuantGranularity.CHANNELWISE:
+    quantized_dim = common_utils.get_weight_quantized_dim(
+        op_info, tensor_content
+    )
+
+  zp, scale = get_zp_scale_from_2d_dequantized_symmetric_weights(
+      dequant_vals=tensor_content,
+      quantized_dimension=quantized_dim,
+  )
+  quant_params = qtyping.UniformQuantParams(
+      scale=scale,
+      zero_point=zp,
+      num_bits=tensor_quant_config.num_bits,
+      symmetric=tensor_quant_config.symmetric,
+      quantized_dimension=quantized_dim,
+  )
+  quantized_vars = uniform_quantize_tensor.uniform_quantize(
+      tensor_content, quant_params
+  )
+  _validate_recovered_weights(tensor_content, quantized_vars, scale)
+  return dataclasses.replace(quant_params, quantized_data=quantized_vars)
+
+
+def calibrate(
+    tfl_op: Any,
+    graph_info: qtyping.GraphInfo,
+    tensor_content_map: dict[str, np.ndarray],
+    inputs_to_ignore: Optional[list[int]] = None,
+    outputs_to_ignore: Optional[list[int]] = None,
+) -> dict[str, qtyping.QSV]:
+  """Collect quantization statistics variable (QSV, e.g., min/max) for the op.
+
+  Args:
+    tfl_op: The tfl operation.
+    graph_info: Graph information needed to perform quantization for the op.
+    tensor_content_map: A map of tensor name to tensor content.
+    inputs_to_ignore: Input tensor indices to ignore.
+    outputs_to_ignore: Output tensor indices to ignore.
+
+  Returns:
+    A dictionary with key as tensor name and value as the collected QSV.
+  """
+  # Reuse the min/max calibration algorithm from naive_min_max_quantize.py since
+  # only weights need to be handled differently.
+  return naive_min_max_quantize.min_max_calibrate(
+      tfl_op,
+      graph_info,
+      tensor_content_map,
+      inputs_to_ignore,
+      outputs_to_ignore,
+  )
+
+
+def init_qsvs(
+    op_info: qtyping.OpInfo,
+    graph_info: qtyping.GraphInfo,
+    inputs_to_ignore: Optional[list[int]] = None,
+    outputs_to_ignore: Optional[list[int]] = None,
+) -> qtyping.QSV:
+  """Initialize the QSVs.
+
+  Args:
+    op_info: Aggregated information about the op (e.g., quantization config).
+    graph_info: Graph information needed to perform quantization for the op.
+    inputs_to_ignore: Input tensor indices to ignore.
+    outputs_to_ignore: Output tensor indices to ignore.
+
+  Returns:
+    QSVs.
+  """
+  # Reuse the min/max calibration algorithm from naive_min_max_quantize.py since
+  # only weights need to be handeled differently.
+  return naive_min_max_quantize.init_qsvs(
+      op_info, graph_info, inputs_to_ignore, outputs_to_ignore
+  )

ai_edge_quantizer/algorithms/uniform_quantize/dequantized_weight_recovery_test.py

@@ -19,7 +19,6 @@ import numpy as np
 from tensorflow.python.platform import googletest
 from ai_edge_quantizer import qtyping
 from ai_edge_quantizer.algorithms.uniform_quantize import dequantized_weight_recovery
-from ai_edge_quantizer.utils import test_utils
 
 _TFLOpName = qtyping.TFLOperationName
 _TensorQuantConfig = qtyping.TensorQuantizationConfig
@@ -31,9 +30,15 @@ class DequantizedWeightRecoveryTest(parameterized.TestCase):
     super().setUp()
     self._dummy_quantized_weights = np.array([
         [1, -2, 3, 4],
-        [6, 7, -8, 5],
-        [-1, 8, -7, -4],
+        [6, 7, -6, 5],
+        [2, -6, -7, -4],
     ])
+    self._dummy_op_info = qtyping.OpInfo(
+        op=None,
+        op_name=_TFLOpName.FULLY_CONNECTED,
+        subgraph_op_index=0,
+        op_quant_config=qtyping.OpQuantizationConfig(),
+    )
 
   @parameterized.named_parameters(
       dict(
@@ -96,18 +101,104 @@ class DequantizedWeightRecoveryTest(parameterized.TestCase):
 
   @parameterized.named_parameters(
       dict(
-          testcase_name="recovery_on_wrong_dimension",
-          quantized_dimension=1,  # should be 0.
+          testcase_name="tensor-recovery-tensor-quant",
+          tensor_quant_config=qtyping.TensorQuantizationConfig(
+              num_bits=4,
+              granularity=qtyping.QuantGranularity.TENSORWISE,
+          ),
+          scale=np.array([0.1875]).reshape(1, 1),
+      ),
+      dict(
+          testcase_name="channel-recovery-channel-quant",
+          tensor_quant_config=qtyping.TensorQuantizationConfig(
+              num_bits=4,
+              granularity=qtyping.QuantGranularity.CHANNELWISE,
+          ),
           scale=np.array([0.1875, 1e-4, 12.3]).reshape(3, 1),
       ),
       dict(
-          testcase_name="tensor_recovery_for_channel_quantization",
-          quantized_dimension=None,  # should be 0.
+          testcase_name="channel-recovery-excessive-bits",
+          tensor_quant_config=qtyping.TensorQuantizationConfig(
+              num_bits=8,  # int4 is enough for the sample weights.
+              granularity=qtyping.QuantGranularity.CHANNELWISE,
+          ),
+          scale=np.array([0.1875, 1e-4, 12.3]).reshape(3, 1),
+      ),
+  )
+  def test_get_tensor_quant_params_success_with_dequantized_weights(
+      self, tensor_quant_config, scale
+  ):
+    dequant_vals = scale * self._dummy_quantized_weights
+    tensor_quant_params = dequantized_weight_recovery.get_tensor_quant_params(
+        self._dummy_op_info, tensor_quant_config, dequant_vals
+    )
+
+    if tensor_quant_config.granularity is qtyping.QuantGranularity.TENSORWISE:
+      self.assertIsNone(tensor_quant_params.quantized_dimension)
+    else:
+      self.assertEqual(tensor_quant_params.quantized_dimension, 0)
+
+    recovered_scale = tensor_quant_params.scale
+    self.assertEqual(recovered_scale.shape, scale.shape)
+    self.assertSequenceAlmostEqual(recovered_scale.flatten(), scale.flatten())
+
+    # Zero point should be zero for symmetric quantization.
+    recovered_zp = tensor_quant_params.zero_point
+    self.assertEqual(np.sum(recovered_zp), 0)
+    self.assertEqual(recovered_zp.shape, scale.shape)
+
+  def test_get_tensor_quant_params_success_with_qsv(self):
+    # Fall back to naive_min_max_quantize.py for non-weight tensors.
+    tensor_quant_params = dequantized_weight_recovery.get_tensor_quant_params(
+        self._dummy_op_info,
+        tensor_quant_config=qtyping.TensorQuantizationConfig(
+            num_bits=8,
+            granularity=qtyping.QuantGranularity.TENSORWISE,
+        ),
+        tensor_qsv={
+            "min": np.array([-1]),
+            "max": np.array([1]),
+        },
+    )
+
+    self.assertIsNone(tensor_quant_params.quantized_dimension)
+    recovered_scale = tensor_quant_params.scale
+    self.assertEqual(recovered_scale.shape, (1,))
+    self.assertSequenceAlmostEqual(recovered_scale.flatten(), [1 / 127])
+
+    # Zero point should be zero for symmetric quantization.
+    recovered_zp = tensor_quant_params.zero_point
+    self.assertEqual(np.sum(recovered_zp), 0)
+    self.assertEqual(recovered_zp.shape, (1,))
+
+  @parameterized.named_parameters(
+      dict(
+          testcase_name="recovery_on_wrong_dimension",
+          tensor_quant_config=qtyping.TensorQuantizationConfig(
+              num_bits=4,
+              granularity=qtyping.QuantGranularity.CHANNELWISE,
+          ),
           scale=np.array([0.003, 1.234, 12.65, 2.24e-4]).reshape(1, 4),
       ),
+      dict(
+          testcase_name="tensor_recovery_for_channel_quantization",
+          tensor_quant_config=qtyping.TensorQuantizationConfig(
+              num_bits=4,
+              granularity=qtyping.QuantGranularity.TENSORWISE,
+          ),
+          scale=np.array([0.1875, 1e-2, 12.3]).reshape(3, 1),
+      ),
+      dict(
+          testcase_name="insufficient_bits",
+          tensor_quant_config=qtyping.TensorQuantizationConfig(
+              num_bits=2,
+              granularity=qtyping.QuantGranularity.CHANNELWISE,
+          ),
+          scale=np.array([0.1875, 1e-2, 12.3]).reshape(3, 1),
+      ),
   )
-  def test_tensor_zp_scale_from_2d_dequantized_symmetric_weights_raises_error_big_recovery_error(
-      self, quantized_dimension, scale
+  def test_get_tensor_quant_params_raises_error_big_recovery_error(
+      self, tensor_quant_config, scale
   ):
     dequant_vals = scale * self._dummy_quantized_weights
     with self.assertRaisesRegex(
@@ -115,8 +206,8 @@ class DequantizedWeightRecoveryTest(parameterized.TestCase):
         "Failed to recover the original quantized values from dequantized"
         " values. Max diff between recovered and original values: ",
     ):
-      dequantized_weight_recovery.get_zp_scale_from_2d_dequantized_symmetric_weights(
-          dequant_vals, quantized_dimension
+      dequantized_weight_recovery.get_tensor_quant_params(
+          self._dummy_op_info, tensor_quant_config, dequant_vals
       )
 
 

{ai_edge_quantizer_nightly-0.0.1.dev20250221.dist-info → ai_edge_quantizer_nightly-0.0.1.dev20250222.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-quantizer-nightly
-Version: 0.0.1.dev20250221
+Version: 0.0.1.dev20250222
 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-0.0.1.dev20250221.dist-info → ai_edge_quantizer_nightly-0.0.1.dev20250222.dist-info}/RECORD

@@ -30,8 +30,8 @@ ai_edge_quantizer/algorithms/nonlinear_quantize/float_casting_test.py,sha256=s64
 ai_edge_quantizer/algorithms/uniform_quantize/__init__.py,sha256=lpq1g2ayg3lCPLy79t2VicYcnGKw64FfYIj1V7J-4m8,676
 ai_edge_quantizer/algorithms/uniform_quantize/common_quantize.py,sha256=wPZevOuowJczG9t4Gynzv7tIeH6zhOnaKPsfr2K_fsk,21259
 ai_edge_quantizer/algorithms/uniform_quantize/common_quantize_test.py,sha256=qMmKbWqxrCoVKbLKHn9WuCrGKPfHkEyU0Nmhokh8Qeo,2597
-ai_edge_quantizer/algorithms/uniform_quantize/dequantized_weight_recovery.py,sha256=LfwZgZwkPZVZBS6VEwaskLNw3BoeymIjxAVw3ZkjjsI,4597
-ai_edge_quantizer/algorithms/uniform_quantize/dequantized_weight_recovery_test.py,sha256=6FPYx4M2-W-SSV6iqQdggd5q5cnciqFI7Ci3Wo5Wyog,4566
+ai_edge_quantizer/algorithms/uniform_quantize/dequantized_weight_recovery.py,sha256=OTXjEZ3Ctq3ffYzisX-6HwgK_DuA7uos_aap5PiIUPE,8686
+ai_edge_quantizer/algorithms/uniform_quantize/dequantized_weight_recovery_test.py,sha256=y7BK11fkF63Ex_Jzg3fbIdy0D_Ca6HuvChVZR7Uwggc,8073
 ai_edge_quantizer/algorithms/uniform_quantize/naive_min_max_quantize.py,sha256=fBqSidFVKZmdO-xIFfwZPdIN1eLJjOik8mUZxZj2ljk,12149
 ai_edge_quantizer/algorithms/uniform_quantize/naive_min_max_quantize_test.py,sha256=Hok09dloSyBfD0oDM5VABdSZjM9JWSQhm_hDHNbFujA,7640
 ai_edge_quantizer/algorithms/uniform_quantize/uniform_quantize_tensor.py,sha256=Q_vx7YN7KMpjubsngxRdJ4bfdSIV-gmXjtVuxIkZuX4,11078
@@ -60,8 +60,8 @@ ai_edge_quantizer/utils/tfl_interpreter_utils.py,sha256=SM8H4i7Jq_nfdsJpImopHndN
 ai_edge_quantizer/utils/tfl_interpreter_utils_test.py,sha256=Op3JxtOqlrjzmYF18jnnstL1k9xiY9kKJ8S2vklKGkc,11327
 ai_edge_quantizer/utils/validation_utils.py,sha256=oYw33Sg547AqtGw-choPUJmp9SAKkV46J_ddqSsum2Q,3950
 ai_edge_quantizer/utils/validation_utils_test.py,sha256=V_qNDikPD4OPB-siOLQCWNVWTAu87h2IgNYt7teFd-o,2934
-ai_edge_quantizer_nightly-0.0.1.dev20250221.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-ai_edge_quantizer_nightly-0.0.1.dev20250221.dist-info/METADATA,sha256=3so5mv89zJCWrCAs77PXQqIq3sGmyue7jkZsmIyO_mQ,1484
-ai_edge_quantizer_nightly-0.0.1.dev20250221.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-ai_edge_quantizer_nightly-0.0.1.dev20250221.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
-ai_edge_quantizer_nightly-0.0.1.dev20250221.dist-info/RECORD,,
+ai_edge_quantizer_nightly-0.0.1.dev20250222.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+ai_edge_quantizer_nightly-0.0.1.dev20250222.dist-info/METADATA,sha256=e9r1p0vAQtBGj4RIEtBbjmiyDyUVUmdNYNU8LqfDVGk,1484
+ai_edge_quantizer_nightly-0.0.1.dev20250222.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+ai_edge_quantizer_nightly-0.0.1.dev20250222.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
+ai_edge_quantizer_nightly-0.0.1.dev20250222.dist-info/RECORD,,