compressed-tensors 0.9.5a20250530__py3-none-any.whl → 0.9.5a20250602__py3-none-any.whl

compressed_tensors/quantization/lifecycle/forward.py

@@ -227,31 +227,42 @@ def _process_quantization(
             perm = torch.argsort(g_idx)
             x = safe_permute(x, perm, dim=1)
 
-        # TODO: experiment with vectorizing for loop for performance
-        end = 0
-        for index, group_count in enumerate(group_sizes):
-            sc = scale[:, index].view(-1, 1)
-            zp = zero_point[:, index].view(-1, 1) if zero_point is not None else None
-
-            start = end
-            end = start + group_count
-            if do_quantize:
-                output[:, start:end] = _quantize(
-                    x=x[:, start:end],
-                    scale=sc,
-                    zero_point=zp,
-                    q_min=q_min,
-                    q_max=q_max,
-                    args=args,
-                    dtype=dtype,
-                    global_scale=global_scale,
-                )
+        x = torch.reshape(
+            x,
+            (
+                x.shape[0],
+                ceil(x.shape[1] / group_size),
+                group_size,
+            ),
+        )
 
-            if do_dequantize:
-                input = output[:, start:end] if do_quantize else x[:, start:end]
-                output[:, start:end] = _dequantize(
-                    x_q=input, scale=sc, zero_point=zp, global_scale=global_scale
-                )
+        if do_quantize:
+            output = _quantize(
+                x=x,
+                scale=scale.unsqueeze(-1),
+                zero_point=zero_point.unsqueeze(-1) if zero_point is not None else None,
+                dtype=dtype,
+                global_scale=global_scale,
+                q_min=q_min,
+                q_max=q_max,
+                args=args,
+            )
+
+        if do_dequantize:
+            input = output if do_quantize else x
+            output = _dequantize(
+                x_q=input,
+                scale=scale.unsqueeze(-1),
+                zero_point=zero_point.unsqueeze(-1) if zero_point is not None else None,
+                global_scale=global_scale,
+            )
+
+        output = torch.reshape(
+            output,
+            (output.shape[0], output.shape[1] * output.shape[2]),
+        )
+
+        output = output.to(output_dtype)
 
         if not is_column_order:
             output = safe_permute(output, torch.argsort(perm), dim=1)

compressed_tensors/quantization/quant_args.py

@@ -53,6 +53,7 @@ class FP4_E2M1_DATA(FloatArgs):
     min = -6.0
 
     @staticmethod
+    @torch.compile
     def cast_to_fp4(x):
         sign = torch.sign(x)
         x = torch.abs(x)

compressed_tensors/quantization/utils/helpers.py

@@ -81,7 +81,7 @@ def calculate_qparams(
         currently only applied/supported for Fp4
 
     :return: tuple of the calculated scale(s) and zero point(s). For FP4, the calculated
-        scale if of dtype FP8
+        scale is of dtype FP8
     """
     # based on the implementations for consuming quantized values,
     # 0.0 must always be representable within the quantized range
@@ -490,7 +490,6 @@ def generate_global_scale(
     attempts to use the entire FP8 dtype range while mapping a per-group max
     to the FP4 max.
     """
-    scale_dtype = scale_data.dtype
     tensor_amax = torch.abs(input_tensor.data).max().to(dtype)
     global_scale = scale_data.max * quant_data.max / tensor_amax
     return global_scale.to(dtype)

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.9.5.a20250530'
+__version__ = version = '0.9.5.a20250602'
 __version_tuple__ = version_tuple = (0, 9, 5)

{compressed_tensors-0.9.5a20250530.dist-info → compressed_tensors-0.9.5a20250602.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.9.5a20250530
+Version: 0.9.5a20250602
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.9.5a20250530.dist-info → compressed_tensors-0.9.5a20250602.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=BwDcUUpFaOn_-cMqdBWktPf89WCzFmESpx94d8qAUZM,521
+compressed_tensors/version.py,sha256=HRt5ki0v2pIyEYRWy0pzJwoSKwmkkF4pTx8ox29EyvU,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
@@ -26,17 +26,17 @@ compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5y
 compressed_tensors/linear/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
 compressed_tensors/linear/compressed_linear.py,sha256=1yo9RyjA0aQ--iuIknFfcSorJn43Mn4CoV-q4JlTJ_o,4052
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
-compressed_tensors/quantization/quant_args.py,sha256=huROC8fbY899EYa2MnEmujvcBeHYLpn-e8ZEViEFASo,11804
+compressed_tensors/quantization/quant_args.py,sha256=2OpiiSdl4KidzNmjx7J8UlQoAYmt5k5GdXv_73ELw0A,11823
 compressed_tensors/quantization/quant_config.py,sha256=aFi6PKqmEX9iP9O8GVn3mEUjRDEwk_hOCbmmiq-j9oU,10198
 compressed_tensors/quantization/quant_scheme.py,sha256=IDWa1GWUbUdWCo8j78Jz6svYF5hLz89J2PVYWBBnXRc,7102
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
 compressed_tensors/quantization/lifecycle/apply.py,sha256=-OKZ-FFFfIIoeGTrho8lXx6HVWZQp3Xkn3Q-G0hU-CM,18294
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
-compressed_tensors/quantization/lifecycle/forward.py,sha256=65USJEtsp_n8X36L5y4g4ftMnhrQyRWbwKJ8RZMMiBo,14797
+compressed_tensors/quantization/lifecycle/forward.py,sha256=WFwvNebxXNUlpX5p1xG80oa8W9fz4-Xd6LCH_B_nptg,14881
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
 compressed_tensors/quantization/lifecycle/initialize.py,sha256=976sZ45ywGVzH1n4pyVhG7hnUBP1wKEWoo9cHrmKHxU,12522
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
-compressed_tensors/quantization/utils/helpers.py,sha256=I-bJcMdBFXjIUQEpnxMMN_FfQyXjojpe5w7ZIKSZ5UU,17588
+compressed_tensors/quantization/utils/helpers.py,sha256=UXs7mTItMdNyGh9NcsztO7qqxG1pmtNQqaAj_bsa7m8,17553
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
 compressed_tensors/registry/registry.py,sha256=0s15BxdGgzBv8RL4kUJCYcuDOFUh_KZYvNvLEeRqWTc,11956
 compressed_tensors/transform/__init__.py,sha256=oa5VdrE-GtDYYceXNSwj5X_ropoXLLukm6Aufcc9WhY,747
@@ -50,8 +50,8 @@ 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=DMfZBuUbA6qp_BG_zIWT3ckiEE33K9ob34s-OgzReO4,12057
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.9.5a20250530.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.9.5a20250530.dist-info/METADATA,sha256=avjHgMxk1vnX09YKjerSCov-X8mTckulmJV1xQyLk5I,7004
-compressed_tensors-0.9.5a20250530.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.9.5a20250530.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.9.5a20250530.dist-info/RECORD,,
+compressed_tensors-0.9.5a20250602.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.9.5a20250602.dist-info/METADATA,sha256=U1CBsVaqy32z2gjdWS4xVTMmlIAxk9BweHUhk51l338,7004
+compressed_tensors-0.9.5a20250602.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.9.5a20250602.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.9.5a20250602.dist-info/RECORD,,