mct-nightly 2.2.0.20250107.15510__py3-none-any.whl → 2.2.0.20250107.164940__py3-none-any.whl

{mct_nightly-2.2.0.20250107.15510.dist-info → mct_nightly-2.2.0.20250107.164940.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.2.0.20250107.15510
+Version: 2.2.0.20250107.164940
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License

{mct_nightly-2.2.0.20250107.15510.dist-info → mct_nightly-2.2.0.20250107.164940.dist-info}/RECORD

@@ -1,4 +1,4 @@
-model_compression_toolkit/__init__.py,sha256=2YnbOvrfArKHSGDUnbQ59awLfgOIYR-Owy8e4iFZ8PE,1573
+model_compression_toolkit/__init__.py,sha256=A5js8ho45WYpU_ZDFPmwshjYKWICZlIgdFrShfdEJGo,1573
 model_compression_toolkit/constants.py,sha256=i_R6uXBfO1ph_X6DNJych2x59SUojfJbn7dNjs_mZnc,3846
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -361,10 +361,10 @@ model_compression_toolkit/gptq/common/gradual_activation_quantization.py,sha256=
 model_compression_toolkit/gptq/common/regularization_factory.py,sha256=hyunpXepVeHyoAFJw6zNLK-3ZHBmiut3lmNisJN_L3E,2514
 model_compression_toolkit/gptq/keras/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
 model_compression_toolkit/gptq/keras/gptq_keras_implementation.py,sha256=axBwnCSjq5xk-xGymOwSOqjp39It-CVtGcCTRTf0E_4,1248
-model_compression_toolkit/gptq/keras/gptq_loss.py,sha256=2hzWzsbuVd5XcL85NM57YeOyHxRY0qMArKn8NvQ1UWw,7643
+model_compression_toolkit/gptq/keras/gptq_loss.py,sha256=k5s7D4CTqbYCHgydyevw1c2p3S2TZCECHNvK79QGE2U,7797
 model_compression_toolkit/gptq/keras/gptq_training.py,sha256=0WGiP7Gs4xX3FBs1PNaZ7w3hWRigwQXqYjBrs_-x32o,23241
 model_compression_toolkit/gptq/keras/graph_info.py,sha256=zwoeHX67nJJ5-zYLjzvMXS9TLsy9BsizARbZiDVjVSA,4473
-model_compression_toolkit/gptq/keras/quantization_facade.py,sha256=meRKqpzZe2Irf21L_rN_mkr5dqPTJHzfSFBeqv4Csp4,18536
+model_compression_toolkit/gptq/keras/quantization_facade.py,sha256=jUAjkIszziedftaQBSmjEL6tYEYpHhlFpSgw2X9OTf4,18672
 model_compression_toolkit/gptq/keras/quantizer/__init__.py,sha256=-DK1CDXvlsnEbki4lukZLpl6Xrbo91_jcqxXlG5Eg6Q,963
 model_compression_toolkit/gptq/keras/quantizer/base_keras_gptq_quantizer.py,sha256=Rbl9urzkmACvVxICSEyJ02qFOBxWK0UQWtysFJzBVZw,4899
 model_compression_toolkit/gptq/keras/quantizer/quant_utils.py,sha256=Vt7Qb8i4JsE4sFtcjpfM4FTXTtfV1t6SwfoNH8a_Iaw,5055
@@ -380,7 +380,7 @@ model_compression_toolkit/gptq/pytorch/gptq_loss.py,sha256=_07Zx_43bnNokwR5S8phI
 model_compression_toolkit/gptq/pytorch/gptq_pytorch_implementation.py,sha256=tECPTavxn8EEwgLaP2zvxdJH6Vg9jC0YOIMJ7857Sdc,1268
 model_compression_toolkit/gptq/pytorch/gptq_training.py,sha256=WtehnyiYXdUXf8-uNpV0mdsalF7YF7eKnL7tcFrzZoE,19549
 model_compression_toolkit/gptq/pytorch/graph_info.py,sha256=4mVM-VvnBaA64ACVdOe6wTGHdMSa2UTLIUe7nACLcdo,4008
-model_compression_toolkit/gptq/pytorch/quantization_facade.py,sha256=kMSq9mrpcgMBRgrEKfMBHaJG6HhGRYnuiDzF4ofckwo,16581
+model_compression_toolkit/gptq/pytorch/quantization_facade.py,sha256=HSFpx6JgjxGhU-0jA0z85sOOgSjCq6gzDOSkmuksZVE,16713
 model_compression_toolkit/gptq/pytorch/quantizer/__init__.py,sha256=ZHNHo1yzye44m9_ht4UUZfTpK01RiVR3Tr74-vtnOGI,968
 model_compression_toolkit/gptq/pytorch/quantizer/base_pytorch_gptq_quantizer.py,sha256=fKg-PNOhGBiL-4eySS9Fyw0GkA76Pq8jT_HbJuJ8iZU,4143
 model_compression_toolkit/gptq/pytorch/quantizer/quant_utils.py,sha256=OocYYRqvl7rZ37QT0hTzfJnWGiNCPskg7cziTlR7TRk,3893
@@ -525,8 +525,8 @@ model_compression_toolkit/xquant/pytorch/model_analyzer.py,sha256=b93o800yVB3Z-i
 model_compression_toolkit/xquant/pytorch/pytorch_report_utils.py,sha256=3jNiV5Z4BVw9cEWuLKNOlLuLdr0EMuKg6eYnSiAq3LU,3952
 model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
 model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=mkoEktLFFHtEKzzFRn_jCnxjhJolK12TZ5AQeDHzUO8,9767
-mct_nightly-2.2.0.20250107.15510.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.2.0.20250107.15510.dist-info/METADATA,sha256=hzg6vLVTb7Mr6NUgWR73ODaN4AIjmActNPhRlH7pGfM,26463
-mct_nightly-2.2.0.20250107.15510.dist-info/WHEEL,sha256=A3WOREP4zgxI0fKrHUG8DC8013e3dK3n7a6HDbcEIwE,91
-mct_nightly-2.2.0.20250107.15510.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.2.0.20250107.15510.dist-info/RECORD,,
+mct_nightly-2.2.0.20250107.164940.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+mct_nightly-2.2.0.20250107.164940.dist-info/METADATA,sha256=9IxOhXh5nRY6Ck-0wINZ4huD8os58zC-X_15pdsxHCo,26464
+mct_nightly-2.2.0.20250107.164940.dist-info/WHEEL,sha256=A3WOREP4zgxI0fKrHUG8DC8013e3dK3n7a6HDbcEIwE,91
+mct_nightly-2.2.0.20250107.164940.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.2.0.20250107.164940.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__ = "2.2.0.20250107.015510"
+__version__ = "2.2.0.20250107.164940"

model_compression_toolkit/gptq/keras/gptq_loss.py

@@ -144,8 +144,9 @@ def activation_mse(flp_act_list,
         loss_values_list.append(point_loss)
         bias_loss_list.append(bias_loss)
     if weights_for_average_loss is not None:
-        return tf.reduce_sum(weights_for_average_loss * tf.stack(loss_values_list)), \
-               tf.reduce_mean(tf.stack(bias_loss_list))
+        print(f"weights_for_average_loss.shape: {weights_for_average_loss.shape}")
+        print(f"tf.stack(loss_values_list).shape: {tf.stack(loss_values_list).shape}")
+        return tf.reduce_sum(weights_for_average_loss * tf.stack(loss_values_list)), tf.reduce_mean(tf.stack(bias_loss_list))
     else:
         return tf.reduce_mean(tf.stack(loss_values_list)), tf.reduce_mean(tf.stack(bias_loss_list))
 
@@ -187,4 +188,4 @@ class GPTQMultipleTensorsLoss:
                                                         weights_for_average_loss=weights_for_average_loss,
                                                         norm_loss=self.norm_loss)
 
-        return loss_act
+        return loss_act

model_compression_toolkit/gptq/keras/quantization_facade.py

@@ -115,7 +115,6 @@ if FOUND_TF:
         if regularization_factor is None:
             regularization_factor = REG_DEFAULT_SLA if use_hessian_sample_attention else REG_DEFAULT
 
-        loss = loss or GPTQMultipleTensorsLoss()
         hessian_weights_config = None
         if use_hessian_sample_attention:
             if not use_hessian_based_weights:    # pragma: no cover
@@ -129,7 +128,10 @@ if FOUND_TF:
             hessian_weights_config = GPTQHessianScoresConfig(per_sample=False,
                                                              hessians_num_samples=GPTQ_HESSIAN_NUM_SAMPLES,
                                                              hessian_batch_size=hessian_batch_size)
-
+        
+        # If a loss was not passed (and was not initialized due to use_hessian_sample_attention), use the default loss
+        loss = loss or GPTQMultipleTensorsLoss()
+        
         if isinstance(gradual_activation_quantization, bool):
             gradual_quant_config = GradualActivationQuantizationConfig() if gradual_activation_quantization else None
         elif isinstance(gradual_activation_quantization, GradualActivationQuantizationConfig):

model_compression_toolkit/gptq/pytorch/quantization_facade.py

@@ -104,7 +104,6 @@ if FOUND_TORCH:
         if regularization_factor is None:
             regularization_factor = REG_DEFAULT_SLA if use_hessian_sample_attention else REG_DEFAULT
 
-        loss = loss or multiple_tensors_mse_loss
         hessian_weights_config = None
         if use_hessian_sample_attention:
             if not use_hessian_based_weights:    # pragma: no cover
@@ -118,6 +117,9 @@ if FOUND_TORCH:
             hessian_weights_config = GPTQHessianScoresConfig(per_sample=False,
                                                              hessians_num_samples=GPTQ_HESSIAN_NUM_SAMPLES,
                                                              hessian_batch_size=hessian_batch_size)
+            
+        # If a loss was not passed (and was not initialized due to use_hessian_sample_attention), use the default loss
+        loss = loss or multiple_tensors_mse_loss
 
         if isinstance(gradual_activation_quantization, bool):
             gradual_quant_config = GradualActivationQuantizationConfig() if gradual_activation_quantization else None