mct-nightly 2.2.0.20241030.505__py3-none-any.whl → 2.2.0.20241101.532__py3-none-any.whl

{mct_nightly-2.2.0.20241030.505.dist-info → mct_nightly-2.2.0.20241101.532.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.2.0.20241030.505
+Version: 2.2.0.20241101.532
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct_nightly-2.2.0.20241030.505.dist-info → mct_nightly-2.2.0.20241101.532.dist-info}/RECORD

@@ -1,4 +1,4 @@
-model_compression_toolkit/__init__.py,sha256=TIarpxVdU_SeymMap8cBKo1tIuj6XQFjDhMwO-5o8XQ,1573
+model_compression_toolkit/__init__.py,sha256=LOlTR7_5Hf3TRvNgu5Y_8a2hjaVGdzEOYKxvqq4feMo,1573
 model_compression_toolkit/constants.py,sha256=i4wYheBkIdQmsQA-axIpcT3YiSO1USNc-jaNiNE8w6E,3920
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -62,7 +62,7 @@ model_compression_toolkit/core/common/mixed_precision/__init__.py,sha256=sw7LOPN
 model_compression_toolkit/core/common/mixed_precision/bit_width_setter.py,sha256=iPuI11f3IkroC-dStdR40vKn4jpa1VL-kskCXc5z7wk,7536
 model_compression_toolkit/core/common/mixed_precision/configurable_quant_id.py,sha256=LLDguK7afsbN742ucLpmJr5TUfTyFpK1vbf2bpVr1v0,882
 model_compression_toolkit/core/common/mixed_precision/configurable_quantizer_utils.py,sha256=7dKMi5S0zQZ16m8NWn1XIuoXsKuZUg64G4-uK8-j1PQ,5177
-model_compression_toolkit/core/common/mixed_precision/distance_weighting.py,sha256=H8qYkJsk88OszUJo-Zde7vTmWiypLTg9KbbzIZ-hhvM,2812
+model_compression_toolkit/core/common/mixed_precision/distance_weighting.py,sha256=-x8edUyudu1EAEM66AuXPtgayLpzbxoLNubfEbFM5kU,2867
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_candidates_filter.py,sha256=klmaMQDeFc3IxRLf6YX4Dw1opFksbLyN10yFHdKAtLo,4875
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py,sha256=r1t025_QHshyoop-PZvL7x6UuXaeplCCU3h4VNBhJHo,4309
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_facade.py,sha256=pk8HRoShDhiUprBC4m1AFQv1SacS4hOrj0MRdbq-5gY,7556
@@ -72,7 +72,7 @@ model_compression_toolkit/core/common/mixed_precision/set_layer_to_bitwidth.py,s
 model_compression_toolkit/core/common/mixed_precision/solution_refinement_procedure.py,sha256=KifDMbm7qkSfvSl6pcZzQ82naIXzeKL6aT-VsvWZYyc,7901
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization.py,sha256=HILF7CIn-GYPvPmTFyvjWLhuLDwSGwdBcAaKFgVYrwk,4745
-model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py,sha256=3ZOI-RNp5faT-U2Og7rLW9EKwBB6ooa7-RwSsWJmquo,14022
+model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py,sha256=1y-Uy60FKbH0B3MDX0sTD4NGqoOU3jyYUQdiF3o-cFw,14893
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/ru_aggregation_methods.py,sha256=ttc8wPa_9LZansutQ2f1ss-RTzgTv739wy3qsdLzyyk,4217
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/ru_functions_mapping.py,sha256=QhuqaECEGLnYC08iD6-2XXcU7NXbPzYf1sQcjYlGak8,1682
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/ru_methods.py,sha256=WC1EHoNuo_lrzy4NRhGJ1cgmJ2IsFsbmP86mrVO3AVA,21506
@@ -111,7 +111,7 @@ model_compression_toolkit/core/common/quantization/quantization_fn_selection.py,
 model_compression_toolkit/core/common/quantization/quantization_params_fn_selection.py,sha256=MwIOBZ4BlZSTIOG75PDvlI3JmZ6t8YjPc1VP9Adei60,3847
 model_compression_toolkit/core/common/quantization/quantize_graph_weights.py,sha256=N005MSvx8UypVpa7XrxNrB2G732n2wHj3RmLyjTgd3I,2728
 model_compression_toolkit/core/common/quantization/quantize_node.py,sha256=cdzGNWfT4MRogIU8ehs0tr3lVjnzAI-jeoS9b4TwVBo,2854
-model_compression_toolkit/core/common/quantization/set_node_quantization_config.py,sha256=IjqFX0EGk4YCTaQsJp4-UycCVc2Ec6GTbu890dkGVns,21318
+model_compression_toolkit/core/common/quantization/set_node_quantization_config.py,sha256=cclpyw9uDe0rXmM6AReH0Q-7SAvft9ksumRXoB5kQUc,21149
 model_compression_toolkit/core/common/quantization/quantization_params_generation/__init__.py,sha256=eCDGwsWYLU6z7qbEVb4TozMW_nd5VEP_iCJ6PcvyEPw,1486
 model_compression_toolkit/core/common/quantization/quantization_params_generation/error_functions.py,sha256=fzUvqmXVgzp_IV5ER-20kKzl4m8U_shZsAKs-ehhjFo,23887
 model_compression_toolkit/core/common/quantization/quantization_params_generation/lut_kmeans_params.py,sha256=RL-PklAjGyC-26anSt8fU07a6pB_LBQFQy9o4e9giN0,8739
@@ -570,8 +570,8 @@ tests_pytest/pytorch/gptq/test_annealing_cfg.py,sha256=hGC7L6mp3N1ygcJ3OctgS_Fz2
 tests_pytest/pytorch/gptq/test_gradual_act_quantization.py,sha256=Dg2cg1X8u9Jxm7Y6tlZIGH81EPoW_vYorcdDExdj02w,4630
 tests_pytest/pytorch/trainable_infrastructure/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
 tests_pytest/pytorch/trainable_infrastructure/test_linear_annealing.py,sha256=zErt9tOu7oupjpv08cvd1Cxvdk9qvP7GMUP6EhefK0c,1814
-mct_nightly-2.2.0.20241030.505.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.2.0.20241030.505.dist-info/METADATA,sha256=qsOde1itXo-3-_7ZxVb1WRYi2y3mzNyq-ei-JidqgqM,20830
-mct_nightly-2.2.0.20241030.505.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-mct_nightly-2.2.0.20241030.505.dist-info/top_level.txt,sha256=csdfSXhtRnpWYRzjZ-dRLIhOmM2TEdVXUxG05A5fgb8,39
-mct_nightly-2.2.0.20241030.505.dist-info/RECORD,,
+mct_nightly-2.2.0.20241101.532.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+mct_nightly-2.2.0.20241101.532.dist-info/METADATA,sha256=-KcA4QsljqwU33XyaBeGMdZz_S4GxsfkJ2eHRkEqpWw,20830
+mct_nightly-2.2.0.20241101.532.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+mct_nightly-2.2.0.20241101.532.dist-info/top_level.txt,sha256=csdfSXhtRnpWYRzjZ-dRLIhOmM2TEdVXUxG05A5fgb8,39
+mct_nightly-2.2.0.20241101.532.dist-info/RECORD,,

model_compression_toolkit/__init__.py

@@ -27,4 +27,4 @@ from model_compression_toolkit import data_generation
 from model_compression_toolkit import pruning
 from model_compression_toolkit.trainable_infrastructure.keras.load_model import keras_load_quantized_model
 
-__version__ = "2.2.0.20241030.000505"
+__version__ = "2.2.0.20241101.000532"

model_compression_toolkit/core/common/mixed_precision/distance_weighting.py

@@ -71,3 +71,6 @@ class MpDistanceWeighting(Enum):
 
     def __call__(self, distance_matrix: np.ndarray) -> np.ndarray:
         return self.value(distance_matrix)
+
+    def __deepcopy__(self, memo):
+        return self

model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py

@@ -12,11 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+import copy
+
 import numpy as np
 from typing import Callable, Any, Dict, Tuple
 
 from model_compression_toolkit.constants import FLOAT_BITWIDTH, BITS_TO_BYTES
-from model_compression_toolkit.core import FrameworkInfo, ResourceUtilization, CoreConfig
+from model_compression_toolkit.core import FrameworkInfo, ResourceUtilization, CoreConfig, QuantizationErrorMethod
 from model_compression_toolkit.core.common import Graph
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
 from model_compression_toolkit.core.common.graph.edge import EDGE_SINK_INDEX
@@ -57,7 +59,7 @@ def compute_resource_utilization_data(in_model: Any,
 
 
     """
-
+    core_config = _create_core_config_for_ru(core_config)
     # We assume that the resource_utilization_data API is used to compute the model resource utilization for
     # mixed precision scenario, so we run graph preparation under the assumption of enabled mixed precision.
     if transformed_graph is None:
@@ -222,6 +224,8 @@ def requires_mixed_precision(in_model: Any,
     Returns: A boolean indicating if mixed precision is needed.
     """
     is_mixed_precision = False
+    core_config = _create_core_config_for_ru(core_config)
+
     transformed_graph = graph_preparation_runner(in_model,
                                                  representative_data_gen,
                                                  core_config.quantization_config,
@@ -247,3 +251,21 @@ def requires_mixed_precision(in_model: Any,
     is_mixed_precision |= target_resource_utilization.total_memory < total_weights_memory_bytes + max_activation_tensor_size_bytes
     is_mixed_precision |= target_resource_utilization.bops < bops_count
     return is_mixed_precision
+
+
+def _create_core_config_for_ru(core_config: CoreConfig) -> CoreConfig:
+    """
+    Create a core config to use for resource utilization computation.
+
+    Args:
+        core_config: input core config
+
+    Returns:
+        Core config for resource utilization.
+    """
+    core_config = copy.deepcopy(core_config)
+    # For resource utilization graph_preparation_runner runs with gptq=False (the default value). HMSE is not supported
+    # without GPTQ and will raise an error later so we replace it with MSE.
+    if core_config.quantization_config.weights_error_method == QuantizationErrorMethod.HMSE:
+        core_config.quantization_config.weights_error_method = QuantizationErrorMethod.MSE
+    return core_config

model_compression_toolkit/core/common/quantization/set_node_quantization_config.py

@@ -58,13 +58,10 @@ def set_quantization_configuration_to_graph(graph: Graph,
 
     if quant_config.weights_error_method == QuantizationErrorMethod.HMSE:
         if not running_gptq:
-            Logger.warning(f"The HMSE error method for parameters selection is only supported when running GPTQ "
-                           f"optimization due to long execution time that is not suitable for basic PTQ. "
-                           f"Using the default MSE error method instead.")
-            quant_config.weights_error_method = QuantizationErrorMethod.MSE
-        else:
-            Logger.warning("Using the HMSE error method for weights quantization parameters search. "
-                           "Note: This method may significantly increase runtime during the parameter search process.")
+            raise ValueError(f"The HMSE error method for parameters selection is only supported when running GPTQ "
+                             f"optimization due to long execution time that is not suitable for basic PTQ.")
+        Logger.warning("Using the HMSE error method for weights quantization parameters search. "
+                       "Note: This method may significantly increase runtime during the parameter search process.")
 
     nodes_to_manipulate_bit_widths = {} if bit_width_config is None else bit_width_config.get_nodes_to_manipulate_bit_widths(graph)