compressed-tensors 0.13.1a20260109__tar.gz → 0.13.1a20260116__tar.gz

{compressed_tensors-0.13.1a20260109/src/compressed_tensors.egg-info → compressed_tensors-0.13.1a20260116}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.13.1a20260109
+Version: 0.13.1a20260116
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/vllm-project/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.13.1a20260109 → compressed_tensors-0.13.1a20260116}/src/compressed_tensors/quantization/quant_scheme.py

@@ -303,6 +303,26 @@ INT8_W4A8 = dict(
     ),
 )
 
+# 4 bit integer weights and 8 bit FP activations quantization
+W4AFP8 = dict(
+    weights=QuantizationArgs(
+        num_bits=4,
+        type=QuantizationType.INT,
+        strategy=QuantizationStrategy.GROUP,
+        group_size=128,
+        symmetric=True,
+        dynamic=False,
+    ),
+    input_activations=QuantizationArgs(
+        num_bits=8,
+        type=QuantizationType.FLOAT,
+        strategy=QuantizationStrategy.TOKEN,
+        symmetric=True,
+        dynamic=True,
+        observer=None,
+    ),
+)
+
 # FP8 weights and FP8 activations quantization
 FP8 = dict(
     weights=QuantizationArgs(
@@ -374,6 +394,7 @@ PRESET_SCHEMES = {
     "W8A8": INT8_W8A8,
     "INT8": INT8_W8A8,  # alias for W8A8
     "W4A8": INT8_W4A8,
+    "W4AFP8": W4AFP8,
     # Float weight and activation schemes
     "FP8": FP8,
     "FP8_DYNAMIC": FP8_DYNAMIC,

{compressed_tensors-0.13.1a20260109 → compressed_tensors-0.13.1a20260116}/src/compressed_tensors/quantization/utils/mxfp4_utils.py

@@ -100,4 +100,4 @@ def generate_mxfp4_scales(x: torch.Tensor) -> torch.Tensor:
     """
     # Round to closest power of 2
     scale_power_2 = round_to_power_2(x)
-    return 127 + torch.floor(torch.log2(scale_power_2))
+    return 127 + torch.floor(torch.log2(scale_power_2)) - 2

{compressed_tensors-0.13.1a20260109 → compressed_tensors-0.13.1a20260116}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.13.1.a20260109'
+__version__ = version = '0.13.1.a20260116'
 __version_tuple__ = version_tuple = (0, 13, 1)

{compressed_tensors-0.13.1a20260109 → compressed_tensors-0.13.1a20260116/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.13.1a20260109
+Version: 0.13.1a20260116
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/vllm-project/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.13.1a20260109 → compressed_tensors-0.13.1a20260116}/tests/test_quantization/test_utils/test_mxfp4_utils.py

@@ -93,5 +93,5 @@ def test_mxfp4_scales_e2e():
     converted_ct = maybe_convert_from_mxfp4_exp(args=args, scale=scales)
 
     scales_exp = torch.log2(converted_ct)
-    block_max_exp = torch.floor(torch.log2(round_to_power_2(block_max)))
+    block_max_exp = torch.floor(torch.log2(round_to_power_2(block_max))) - 2
     assert torch.equal(scales_exp, block_max_exp)