compressed-tensors 0.9.4a20250408__tar.gz → 0.9.4a20250410__tar.gz

{compressed_tensors-0.9.4a20250408 → compressed_tensors-0.9.4a20250410}/.github/workflows/upload.yml

@@ -97,14 +97,21 @@ jobs:
                 run_id: ${{ inputs.run_id }}
                 asset_identifier: 'compressed*.whl'
 
+            - name: check if whl is new
+              id: check-whl
+              uses: neuralmagic/nm-actions/actions/check-whl-on-pypi@scm
+              with:
+                whl: ${{ steps.find-asset-whl.outputs.asset }}
+
             - name: upload whl to "nm-pypi"
-              if: ${{ inputs.push_to_pypi }}
+              if: ${{ inputs.push_to_pypi && steps.check-whl.outputs.status }}
               uses: neuralmagic/nm-actions/actions/gcp-upload-asset@v1.1.0
               with:
                 bucket_target: ${{ secrets.GCP_NM_PYPI_DIST }}
                 asset: ${{ steps.find-asset-whl.outputs.asset }}
 
             - name: find tar.gz
+              if: ${{ steps.check-whl.outputs.status }}
               id: find-asset-targz
               uses: neuralmagic/nm-actions/actions/find-asset@v1.1.0
               with:
@@ -112,13 +119,14 @@ jobs:
                 asset_identifier: 'compressed*.tar.gz'
 
             - name: upload tar.gz to "nm-pypi"
-              if: ${{ inputs.push_to_pypi }}
+              if: ${{ inputs.push_to_pypi && steps.check-whl.outputs.status }}
               uses: neuralmagic/nm-actions/actions/gcp-upload-asset@v1.1.0
               with:
                 bucket_target: ${{ secrets.GCP_NM_PYPI_DIST }}
                 asset: ${{ steps.find-asset-targz.outputs.asset }}
 
             - name: update "nm-pypi" index
+              if: ${{ steps.check-whl.outputs.status }}
               uses: actions/github-script@v6
               with:
                 github-token: ${{ secrets.NM_PYPI_WORKFLOW }}
@@ -133,7 +141,7 @@ jobs:
 
             # publish the wheel file to public pypi
             - name: push wheel to pypi.org
-              if: ${{ inputs.push_to_pypi }}
+              if: ${{ inputs.push_to_pypi && steps.check-whl.outputs.status }}
               uses: neuralmagic/nm-actions/actions/publish-whl@v1.0.0
               with:
                 username: ${{ secrets.PYPI_PUBLIC_USER }}
@@ -142,7 +150,7 @@ jobs:
 
             # publish the tar.gz file to public pypi
             - name: push wheel to pypi.org
-              if: ${{ inputs.push_to_pypi }}
+              if: ${{ inputs.push_to_pypi && steps.check-whl.outputs.status }}
               uses: neuralmagic/nm-actions/actions/publish-whl@v1.0.0
               with:
                 username: ${{ secrets.PYPI_PUBLIC_USER }}

{compressed_tensors-0.9.4a20250408 → compressed_tensors-0.9.4a20250410}/Makefile

@@ -1,6 +1,6 @@
 .PHONY: build docs test
 
-BUILD_ARGS := dev # set nightly to build nightly release
+BUILD_TYPE ?= dev # set nightly to build nightly release
 PYCHECKDIRS := src tests
 PYCHECKGLOBS := 'src/**/*.py' 'tests/**/*.py' 'utils/**/*.py' 'examples/**/*.py' setup.py
 # run checks on all files for the repo
@@ -28,7 +28,7 @@ test:
 # creates wheel file
 build:
 	@echo "Building the wheel for the repository";
-	BUILD_TYPE=$(BUILD_ARGS) python3 setup.py sdist bdist_wheel;
+	BUILD_TYPE=$(BUILD_TYPE) python3 setup.py sdist bdist_wheel;
 
 # clean package
 clean:

{compressed_tensors-0.9.4a20250408/src/compressed_tensors.egg-info → compressed_tensors-0.9.4a20250410}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.9.4a20250408
+Version: 0.9.4a20250410
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.9.4a20250408 → compressed_tensors-0.9.4a20250410}/src/compressed_tensors/quantization/quant_scheme.py

@@ -142,6 +142,18 @@ W4A16 = dict(
     ),
 )
 
+# 4 bit integer weights only asymmetric quantization
+W4A16_ASYM = dict(
+    weights=QuantizationArgs(
+        num_bits=4,
+        type=QuantizationType.INT,
+        strategy=QuantizationStrategy.GROUP,
+        group_size=128,
+        symmetric=False,
+        dynamic=False,
+    ),
+)
+
 # 4 bit integer weights and 8 bit activations quantization
 INT8_W4A8 = dict(
     weights=QuantizationArgs(
@@ -205,6 +217,7 @@ PRESET_SCHEMES = {
     # Integer weight only schemes
     "W8A16": W8A16,
     "W4A16": W4A16,
+    "W4A16_ASYM": W4A16_ASYM,
     # Integer weight and activation schemes
     "W8A8": INT8_W8A8,
     "INT8": INT8_W8A8,  # alias for W8A8

{compressed_tensors-0.9.4a20250408 → compressed_tensors-0.9.4a20250410}/src/compressed_tensors/quantization/utils/helpers.py

@@ -64,8 +64,11 @@ def calculate_qparams(
     :param quantization_args: settings to quantization
     :return: tuple of the calculated scale(s) and zero point(s)
     """
+    # based on the implementations for consuming quantized values,
+    # 0.0 must always be representable within the quantized range
     min_vals = torch.min(min_vals, torch.zeros_like(min_vals))
     max_vals = torch.max(max_vals, torch.zeros_like(max_vals))
+
     device = min_vals.device
 
     bit_min, bit_max = calculate_range(quantization_args, device)
@@ -84,6 +87,9 @@ def calculate_qparams(
         zero_points = torch.clamp(zero_points, bit_min, bit_max)
 
     # match zero-points to quantized type
+    # if casting to int, use round instead of truncate
+    if quantization_args.type == QuantizationType.INT:
+        zero_points = torch.round(zero_points)
     zero_points = zero_points.to(zp_dtype)
 
     if scales.ndim == 0:
@@ -96,7 +102,7 @@ def calculate_qparams(
 def compute_dynamic_scales_and_zp(value: Tensor, args: QuantizationArgs):
     """
     Returns the computed scales and zero points for dynamic activation
-    qunatization.
+    quantization.
 
     :param value: tensor to calculate quantization parameters for
     :param args: quantization args

{compressed_tensors-0.9.4a20250408 → compressed_tensors-0.9.4a20250410}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.9.4.a20250408'
+__version__ = version = '0.9.4.a20250410'
 __version_tuple__ = version_tuple = (0, 9, 4)

{compressed_tensors-0.9.4a20250408 → compressed_tensors-0.9.4a20250410/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.9.4a20250408
+Version: 0.9.4a20250410
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.