compressed-tensors 0.11.1a20250828__tar.gz → 0.11.1a20250903__tar.gz

{compressed_tensors-0.11.1a20250828 → compressed_tensors-0.11.1a20250903}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.11.1a20250828
+Version: 0.11.1a20250903
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.11.1a20250828 → compressed_tensors-0.11.1a20250903}/examples/llama_1.1b/ex_llmcompressor_quantization.py

@@ -24,7 +24,7 @@
 from pathlib import Path
 
 import torch
-from llmcompressor.transformers import oneshot
+from llmcompressor import oneshot
 
 
 recipe = str(Path(__file__).parent / "example_quant_recipe.yaml")

{compressed_tensors-0.11.1a20250828 → compressed_tensors-0.11.1a20250903}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.11.1.a20250828'
+__version__ = version = '0.11.1.a20250903'
 __version_tuple__ = version_tuple = (0, 11, 1)

{compressed_tensors-0.11.1a20250828 → compressed_tensors-0.11.1a20250903}/src/compressed_tensors.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.11.1a20250828
+Version: 0.11.1a20250903
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.11.1a20250828 → compressed_tensors-0.11.1a20250903}/tests/test_quantization/lifecycle/test_forward.py

@@ -19,9 +19,8 @@ import pytest
 import torch
 from compressed_tensors.quantization.lifecycle.forward import (
     _process_quantization,
-    dequantize,
+    fake_quantize,
     forward_quantize,
-    quantize,
     wrap_module_forward_quantized,
 )
 from compressed_tensors.quantization.lifecycle.initialize import (
@@ -96,7 +95,7 @@ def test_forward_quantize(
 
 
 @pytest.mark.parametrize(
-    "num_bits,type,strategy,group_size,scale,zero_point,g_idx",
+    "num_bits,type,strategy,group_size,scale,zero_point,g_idx,global_scale",
     [
         (
             4,
@@ -106,6 +105,7 @@ def test_forward_quantize(
             torch.rand((1,)) * 0.01,
             torch.zeros((1,)),
             None,
+            None,
         ),
         (
             4,
@@ -115,6 +115,7 @@ def test_forward_quantize(
             torch.rand((512, 8)) * 0.01,
             torch.zeros((512, 8)),
             None,
+            None,
         ),
         (
             4,
@@ -124,6 +125,7 @@ def test_forward_quantize(
             torch.rand((512, 8)) * 0.01,
             torch.zeros((512, 8)),
             make_dummy_g_idx(1024, 128),
+            None,
         ),
         (
             8,
@@ -133,6 +135,7 @@ def test_forward_quantize(
             torch.rand((1,)) * 0.01,
             torch.zeros((1,)),
             None,
+            None,
         ),
         (
             8,
@@ -142,6 +145,7 @@ def test_forward_quantize(
             torch.rand((512, 8)) * 0.01,
             torch.zeros((512, 8)),
             None,
+            None,
         ),
         (
             8,
@@ -151,28 +155,8 @@ def test_forward_quantize(
             torch.rand((512, 8)) * 0.01,
             torch.zeros((512, 8)),
             make_dummy_g_idx(1024, 128),
+            None,
         ),
-    ],
-)
-def test_quantize(num_bits, type, strategy, group_size, scale, zero_point, g_idx):
-    args = QuantizationArgs(
-        num_bits=num_bits, type=type, strategy=strategy, group_size=group_size
-    )
-
-    x = torch.rand((512, 1024))
-    quantize(
-        x=x,
-        scale=scale,
-        zero_point=zero_point,
-        args=args,
-        dtype=args.pytorch_dtype(),
-        g_idx=g_idx,
-    )
-
-
-@pytest.mark.parametrize(
-    "num_bits,type,strategy,group_size,scale,zero_point,g_idx",
-    [
         (
             8,
             "int",
@@ -181,6 +165,7 @@ def test_quantize(num_bits, type, strategy, group_size, scale, zero_point, g_idx
             torch.rand((512, 8)) * 0.01,
             torch.zeros((512, 8)),
             None,
+            None,
         ),
         (
             8,
@@ -190,23 +175,26 @@ def test_quantize(num_bits, type, strategy, group_size, scale, zero_point, g_idx
             torch.rand((512, 8)) * 0.01,
             torch.zeros((512, 8)),
             make_dummy_g_idx(1024, 128),
+            None,
         ),
     ],
 )
-def test_dequantize(num_bits, type, strategy, group_size, scale, zero_point, g_idx):
+def test_fake_quantize_2d(
+    num_bits, type, strategy, group_size, scale, zero_point, g_idx, global_scale
+):
     args = QuantizationArgs(
         num_bits=num_bits, type=type, strategy=strategy, group_size=group_size
     )
 
-    x_q = torch.rand((512, 1024)).to(dtype=args.pytorch_dtype())
-    dequantize(
-        x_q=x_q,
+    x = torch.rand((512, 1024))
+    fake_quantize(
+        x=x,
         scale=scale,
         zero_point=zero_point,
         args=args,
-        dtype=None,
         g_idx=g_idx,
-    )
+        global_scale=global_scale,
+    )  # note that reconstruction loss is bad for uncalibrated scales
 
 
 def test_process_quantization_block_static():