compressed-tensors 0.9.4a20250408__py3-none-any.whl → 0.9.4a20250410__py3-none-any.whl

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/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/version.py

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

{compressed_tensors-0.9.4a20250408.dist-info → compressed_tensors-0.9.4a20250410.dist-info}/METADATA

@@ -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.dist-info → compressed_tensors-0.9.4a20250410.dist-info}/RECORD

@@ -1,6 +1,6 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=dU0WxLg_un23vp7nx7GfLU01yj3Z9Aru2yP4cp8c0-c,520
+compressed_tensors/version.py,sha256=ttCtltINyBJn0un6r4KDJzCF49mLYcu-E_khCO2nku8,520
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=x8dQrWVEurynXw03yHJZTaAmrRTOsdZJoHjmvs0IKwk,7002
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
@@ -27,7 +27,7 @@ compressed_tensors/linear/compressed_linear.py,sha256=_m6XpNcI53eeSHO8VdiuAM6UBT
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
 compressed_tensors/quantization/quant_args.py,sha256=sKpb8DcNObidjXjNol1Tn_Iih3ZXBycSp-fyz68TGhY,9117
 compressed_tensors/quantization/quant_config.py,sha256=MxSUcb5dOqMN6LFyD5K2h8X0TvEtcWIAoiUJqD2dHGE,10159
-compressed_tensors/quantization/quant_scheme.py,sha256=eQ0JrRZ80GX69fpwW87VzPzzhajhk4mUaJScjk82OY4,6010
+compressed_tensors/quantization/quant_scheme.py,sha256=yz0oMbbwp7QZXXd2k5KIJu-Q6aTqg2929VdUzZ7vysM,6324
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
 compressed_tensors/quantization/lifecycle/apply.py,sha256=lZmCCSm1_o79iUAy460w6Bv9FaOvntVisMdS-dN9fnk,16594
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
@@ -35,7 +35,7 @@ compressed_tensors/quantization/lifecycle/forward.py,sha256=DOWouUqfaLA4Qhg-ojVV
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
 compressed_tensors/quantization/lifecycle/initialize.py,sha256=sK3PLm69N91QepBuq-83Qd2Br6XcOmRDpD5qo_WWNJo,7469
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
-compressed_tensors/quantization/utils/helpers.py,sha256=DBP-sGRpGAY01K0LFE7qqonNj4hkTYL_mXrMs2LtAD8,14100
+compressed_tensors/quantization/utils/helpers.py,sha256=-wX0H7zVysJ67jRRCGbx6BfxbMU_1sqffTf5YUIpPiU,14391
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
 compressed_tensors/registry/registry.py,sha256=vRcjVB1ITfSbfYUaGndBBmqhip_5vsS62weorVg0iXo,11896
 compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
@@ -45,8 +45,8 @@ compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVy
 compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
 compressed_tensors/utils/safetensors_load.py,sha256=5SeM2hzLh77Ne8Vk7qR6-km7cf8bhov41ExpWITqX3A,11470
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.9.4a20250408.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.9.4a20250408.dist-info/METADATA,sha256=CXl80o7QymLek4-pwpCHF9L3-OgIowJ2KDmfi8r-YBs,7004
-compressed_tensors-0.9.4a20250408.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
-compressed_tensors-0.9.4a20250408.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.9.4a20250408.dist-info/RECORD,,
+compressed_tensors-0.9.4a20250410.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.9.4a20250410.dist-info/METADATA,sha256=EKD2EHV9dh7GHjeO25QvapLLR3JabT4WC04IU6K-cLg,7004
+compressed_tensors-0.9.4a20250410.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+compressed_tensors-0.9.4a20250410.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.9.4a20250410.dist-info/RECORD,,