compressed-tensors 0.9.5a20250509__py3-none-any.whl → 0.9.5a20250512__py3-none-any.whl

compressed_tensors/quantization/quant_args.py

@@ -24,6 +24,8 @@ from pydantic import BaseModel, Field, field_validator, model_validator
 
 __all__ = [
     "FP8_DTYPE",
+    "FP8_E4M3_DATA",
+    "FP4_E2M1_DATA",
     "QuantizationType",
     "QuantizationStrategy",
     "QuantizationArgs",
@@ -31,6 +33,48 @@ __all__ = [
     "ActivationOrdering",
 ]
 
+
+class FloatArgs:
+    exponent: int
+    mantissa: int
+    bits: int
+    max: float
+    min: float
+    dtype: Optional[torch.dtype] = None
+
+
+class FP4_E2M1_DATA(FloatArgs):
+    exponent = 2
+    mantissa = 1
+    bits = 4
+    max = 6.0
+    min = -6.0
+
+    @staticmethod
+    def cast_to_fp4(x):
+        sign = torch.sign(x)
+        x = torch.abs(x)
+        x[(x >= 0.0) & (x <= 0.25)] = 0.0
+        x[(x > 0.25) & (x < 0.75)] = 0.5
+        x[(x >= 0.75) & (x <= 1.25)] = 1.0
+        x[(x > 1.25) & (x < 1.75)] = 1.5
+        x[(x >= 1.75) & (x <= 2.5)] = 2.0
+        x[(x > 2.5) & (x < 3.5)] = 3.0
+        x[(x >= 3.5) & (x <= 5.0)] = 4.0
+        x[x > 5.0] = 6.0
+        return x * sign
+
+
+class FP8_E4M3_DATA(FloatArgs):
+    exponent = 4
+    mantissa = 3
+    bits = 8
+    max = torch.finfo(torch.float8_e4m3fn).max
+    min = torch.finfo(torch.float8_e4m3fn).min
+    dtype = torch.float8_e4m3fn
+
+
+# TODO: Remove soon in favour of a more descriptive FloatArgs
 FP8_DTYPE = torch.float8_e4m3fn
 
 
@@ -234,7 +278,10 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
 
     def pytorch_dtype(self) -> torch.dtype:
         if self.type == QuantizationType.FLOAT:
-            return FP8_DTYPE
+            if self.num_bits == 8:
+                return FP8_E4M3_DATA.dtype
+            else:
+                raise NotImplementedError("Only num_bits in (8) are supported")
         elif self.type == QuantizationType.INT:
             if self.num_bits <= 8:
                 return torch.int8
@@ -263,7 +310,12 @@ def round_to_quantized_type(
     """
     original_dtype = tensor.dtype
     if args.type == QuantizationType.FLOAT:
-        rounded = tensor.to(FP8_DTYPE)
+        if args.num_bits == 8:
+            rounded = tensor.to(FP8_E4M3_DATA.dtype)
+        elif args.num_bits == 4:
+            rounded = FP4_E2M1_DATA.cast_to_fp4(tensor)
+        else:
+            raise NotImplementedError("Only num_bits in (4, 8) are supported")
     elif args.type == QuantizationType.INT:
         rounded = torch.round(tensor)
     else:

compressed_tensors/quantization/quant_scheme.py

@@ -100,6 +100,17 @@ def is_preset_scheme(name: str) -> bool:
 
 UNQUANTIZED = dict()
 
+NVFP4A16 = dict(
+    weights=QuantizationArgs(
+        num_bits=4,
+        type=QuantizationType.FLOAT,
+        strategy=QuantizationStrategy.GROUP,
+        symmetric=True,
+        dynamic=False,
+        group_size=16,
+    )
+)
+
 # 8 bit integer weights and 8 bit activations quantization
 INT8_W8A8 = dict(
     weights=QuantizationArgs(
@@ -225,4 +236,5 @@ PRESET_SCHEMES = {
     # Float weight and activation schemes
     "FP8": FP8,
     "FP8_DYNAMIC": FP8_DYNAMIC,
+    "NVFP4A16": NVFP4A16,
 }

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.9.5.a20250509'
+__version__ = version = '0.9.5.a20250512'
 __version_tuple__ = version_tuple = (0, 9, 5)

{compressed_tensors-0.9.5a20250509.dist-info → compressed_tensors-0.9.5a20250512.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.9.5a20250509
+Version: 0.9.5a20250512
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.9.5a20250509.dist-info → compressed_tensors-0.9.5a20250512.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=MJNvDWAetuDB0OYncH2iIZtFhxnO8XXV5IHnTcj9e6k,521
+compressed_tensors/version.py,sha256=Yd5MPtXGvm9XWCPM_O99KCK_9DwKmcl2OenuJ4MxlUI,521
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=nvWsv4xEw1Tkxkxth6TmHplDYXfBeP22xWxOsZERyDY,7204
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
@@ -26,9 +26,9 @@ compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5y
 compressed_tensors/linear/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
 compressed_tensors/linear/compressed_linear.py,sha256=_m6XpNcI53eeSHO8VdiuAM6UBTdpDhn5Ivd8iRMwEKc,3980
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
-compressed_tensors/quantization/quant_args.py,sha256=sKpb8DcNObidjXjNol1Tn_Iih3ZXBycSp-fyz68TGhY,9117
+compressed_tensors/quantization/quant_args.py,sha256=2m4WJWBnNjkU-3rVR_f2a6p_BZMvGfMrPyOui8JUwWk,10487
 compressed_tensors/quantization/quant_config.py,sha256=MxSUcb5dOqMN6LFyD5K2h8X0TvEtcWIAoiUJqD2dHGE,10159
-compressed_tensors/quantization/quant_scheme.py,sha256=yz0oMbbwp7QZXXd2k5KIJu-Q6aTqg2929VdUzZ7vysM,6324
+compressed_tensors/quantization/quant_scheme.py,sha256=0FpN3R7bVn8rQ18Vp0NuDVpoilTZ7X8vk9zp_8AndwY,6578
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
 compressed_tensors/quantization/lifecycle/apply.py,sha256=DOoxH4jM8r0270GGGUFOpRrgwaisiJi7TV-Q6E8qM8E,18067
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
@@ -46,8 +46,8 @@ 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=kkkUDmS1H40MFy6FDP-DFGiAYbtqke6bKE7YrAtORtA,11499
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.9.5a20250509.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.9.5a20250509.dist-info/METADATA,sha256=2BVor0VJtcZHud4QBjYS5OhLXob0nF9dmeP08RUSA5k,7004
-compressed_tensors-0.9.5a20250509.dist-info/WHEEL,sha256=DnLRTWE75wApRYVsjgc6wsVswC54sMSJhAEd4xhDpBk,91
-compressed_tensors-0.9.5a20250509.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.9.5a20250509.dist-info/RECORD,,
+compressed_tensors-0.9.5a20250512.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.9.5a20250512.dist-info/METADATA,sha256=gArKa7gy0jdBGF5PbuNLVh_ZmnXq4CdEp1-7grxpjnw,7004
+compressed_tensors-0.9.5a20250512.dist-info/WHEEL,sha256=DnLRTWE75wApRYVsjgc6wsVswC54sMSJhAEd4xhDpBk,91
+compressed_tensors-0.9.5a20250512.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.9.5a20250512.dist-info/RECORD,,