compressed-tensors-nightly 0.5.0.20240808__py3-none-any.whl → 0.5.0.20240810__py3-none-any.whl

compressed_tensors/quantization/lifecycle/apply.py

@@ -14,7 +14,7 @@
 
 import logging
 import re
-from collections import OrderedDict
+from collections import OrderedDict, defaultdict
 from copy import deepcopy
 from typing import Dict, Iterable, List, Optional
 from typing import OrderedDict as OrderedDictType
@@ -125,13 +125,14 @@ def apply_quantization_config(model: Module, config: QuantizationConfig) -> Dict
             target_to_scheme[target] = scheme
 
     # list of submodules to ignore
-    ignored_submodules = []
+    ignored_submodules = defaultdict(list)
     # mark appropriate layers for quantization by setting their quantization schemes
     for name, submodule in iter_named_leaf_modules(model):
         # potentially fix module name to remove FSDP wrapper prefix
         name = fix_fsdp_module_name(name)
-        if find_name_or_class_matches(name, submodule, config.ignore):
-            ignored_submodules.append(name)
+        if matches := find_name_or_class_matches(name, submodule, config.ignore):
+            for match in matches:
+                ignored_submodules[match].append(name)
             continue  # layer matches ignore list, continue
         targets = find_name_or_class_matches(name, submodule, target_to_scheme)
         if targets:
@@ -200,7 +201,14 @@ def apply_quantization_status(model: Module, status: QuantizationStatus):
         model.apply(initialize_module_for_quantization)
 
     if current_status < status >= QuantizationStatus.CALIBRATION > current_status:
-        model.apply(set_module_for_calibration)
+        # only quantize weights up front when our end goal state is calibration,
+        # weight quantization parameters are already loaded for frozen/compressed
+        quantize_weights_upfront = status == QuantizationStatus.CALIBRATION
+        model.apply(
+            lambda module: set_module_for_calibration(
+                module, quantize_weights_upfront=quantize_weights_upfront
+            )
+        )
     if current_status < status >= QuantizationStatus.FROZEN > current_status:
         model.apply(freeze_module_quantization)
 

compressed_tensors/quantization/lifecycle/calibration.py

@@ -28,7 +28,7 @@ __all__ = [
 _LOGGER = logging.getLogger(__name__)
 
 
-def set_module_for_calibration(module: Module):
+def set_module_for_calibration(module: Module, quantize_weights_upfront: bool = True):
     """
     marks a layer as ready for calibration which activates observers
     to update scales and zero points on each forward pass
@@ -36,6 +36,8 @@ def set_module_for_calibration(module: Module):
     apply to full model with `model.apply(set_module_for_calibration)`
 
     :param module: module to set for calibration
+    :param quantize_weights_upfront: whether to automatically
+       run weight quantization at the start of calibration
     """
     if not getattr(module, "quantization_scheme", None):
         # no quantization scheme nothing to do
@@ -49,7 +51,7 @@ def set_module_for_calibration(module: Module):
             "to re-calibrate a frozen module"
         )
 
-    if module.quantization_scheme.weights is not None:
+    if quantize_weights_upfront and module.quantization_scheme.weights is not None:
         # set weight scale and zero_point up front, calibration data doesn't affect it
         observer = module.weight_observer
 

{compressed_tensors_nightly-0.5.0.20240808.dist-info → compressed_tensors_nightly-0.5.0.20240810.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.5.0.20240808
+Version: 0.5.0.20240810
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors_nightly-0.5.0.20240808.dist-info → compressed_tensors_nightly-0.5.0.20240810.dist-info}/RECORD

@@ -19,8 +19,8 @@ compressed_tensors/quantization/quant_args.py,sha256=Vc_tWSTcbZZsMJlACpLq4JEPvGx
 compressed_tensors/quantization/quant_config.py,sha256=NpVu8YJ4Xw2pIQW_PGaNaml8kx1bUnxkvb0jBYWbKdE,9971
 compressed_tensors/quantization/quant_scheme.py,sha256=_RKOFJI0T5xJVBLX63UeYkSY4EFAecsBnqzUIVBjeU0,6014
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=MXE2E7GfIfRRfhrdGy2Og3AZOz5N59B0ZGFcsD89y6c,821
-compressed_tensors/quantization/lifecycle/apply.py,sha256=FDKw6AaQjOwNtcX6oFXNx5b_bwm6BpuFtly1Ll6-WFE,13451
-compressed_tensors/quantization/lifecycle/calibration.py,sha256=n-m4xwa9Ds2xrltp6r0rhuVLJhE8bQ1LnifrcrbA-ig,2448
+compressed_tensors/quantization/lifecycle/apply.py,sha256=UGnccJ4QPXU14lRqJkre95A_Nn5jYMAE0mWnsMi26_s,13900
+compressed_tensors/quantization/lifecycle/calibration.py,sha256=zE5jtW-e5j8vrLO7FqhX3oUSNhjrg4FsRFiG6vDu7ME,2637
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=VreB10xPwgSLQQlTu20UCrFpRS--cA7-lx5s7nrPPrg,2247
 compressed_tensors/quantization/lifecycle/forward.py,sha256=6PSXYcf-R1dOY8zsuIWnBaoyARNymYc3-qvV6-L7SlI,12397
 compressed_tensors/quantization/lifecycle/frozen.py,sha256=h1XYt89MouBTf3jTYLG_6OdFxIu5q2N8tPjsy6J4E6Y,1726
@@ -41,8 +41,8 @@ compressed_tensors/utils/offload.py,sha256=qAMwoFT3WEQ9nB_SegE12ob8ghDugddQseE6z
 compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVyah6BUUir_StT28,2530
 compressed_tensors/utils/safetensors_load.py,sha256=0MheXwx1jeY12PeISppiSIZHs6rmN2YddwPpFb9V67I,8527
 compressed_tensors/utils/semi_structured_conversions.py,sha256=g1EZHzdv-ko7ufPX430dp7wE33o6FWJXuSP4zZydCu0,13488
-compressed_tensors_nightly-0.5.0.20240808.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.5.0.20240808.dist-info/METADATA,sha256=-jkh63msCndnafIYuIJOYXfc-rlfREqnYk5ILDk3pRg,5680
-compressed_tensors_nightly-0.5.0.20240808.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-compressed_tensors_nightly-0.5.0.20240808.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.5.0.20240808.dist-info/RECORD,,
+compressed_tensors_nightly-0.5.0.20240810.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.5.0.20240810.dist-info/METADATA,sha256=yHlRu1oQOe7H-xo9pOH-qA3Nc8yuPNKtOz5GoEg1spQ,5680
+compressed_tensors_nightly-0.5.0.20240810.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+compressed_tensors_nightly-0.5.0.20240810.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.5.0.20240810.dist-info/RECORD,,