compressed-tensors 0.12.1a20251001__tar.gz → 0.12.2__tar.gz

compressed_tensors-0.12.2/.github/workflows/post-release-nightly-build.yml

@@ -0,0 +1,15 @@
+name: Post-release nightly build & release
+
+on:
+  release:
+    types: [released]
+
+jobs:
+  BUILD-TEST-NIGHTLY:
+    name: Post-release nightly build & release
+    uses: ./.github/workflows/trigger-all.yml
+    with:
+      wf_category: NIGHTLY
+      push_to_pypi: true
+      gitref: main
+    secrets: inherit

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/.github/workflows/trigger-all.yml

@@ -5,6 +5,21 @@ on:
       # * is a special character in YAML so you have to quote this string
       - cron: '30 0 * * *'  # nightly run
 
+    workflow_call:
+        inputs:
+            wf_category:
+                description: "workflow category, must be 'NIGHTLY' or 'RELEASE' (default: NIGHTLY)"
+                type: string
+                default: NIGHTLY
+            push_to_pypi:
+                description: "when set and tests pass, then '.whl' & '.tar.gz' will be pushed to public pypi"
+                type: boolean
+                default: false
+            gitref:
+                description: "git commit hash or tag name"
+                type: string
+                default: 'main'
+
     workflow_dispatch:
         inputs:
             wf_category:
@@ -36,5 +51,4 @@ jobs:
                             {"python":"3.10.12","label":"k8s-util","timeout":"40"},
                             {"python":"3.9.17","label":"k8s-h100-solo","timeout":"40"},
                             {"python":"3.12.6","label":"k8s-a100-duo","timeout":"40"}]'
-
         secrets: inherit

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.1a20251001
+Version: 0.12.2
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/src/compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -344,6 +344,55 @@ class ModelCompressor:
                     format, config=quantization_config
                 )
 
+    def get_missing_module_keys(self, model: Module) -> List[str]:
+        """
+        Identifies the expected missing weight keys in the compressed state_dict.
+
+        When a model undergoes sparsity or quantization compression, certain
+        weight tensors may be absent from the checkpoint by virtue of compression.
+        This function determines which weight keys are missing based on the
+        applied compression techniques.
+
+        :param model: The PyTorch model to check for missing keys.
+        :return: A list of missing keys expected in the compressed state_dict.
+        """
+        missing_keys = set()
+
+        # Determine missing keys due to sparsity compression
+        if (
+            self.sparsity_compressor
+            and self.sparsity_config.format != CompressionFormat.dense.value
+        ):
+            sparse_targets = match_named_modules(
+                model=model,
+                targets=self.sparsity_config.targets,
+                ignore=self.sparsity_config.ignore,
+            )
+
+            missing_keys.update(
+                merge_names(target_name, "weight")
+                for target_name, _module in sparse_targets
+            )
+
+        # Determine missing keys due to pack quantization
+        if (
+            self.quantization_compressor
+            and self.quantization_config.format
+            == CompressionFormat.pack_quantized.value
+        ):
+            for scheme in self.quantization_config.config_groups.values():
+                quant_targets = match_named_modules(
+                    model=model,
+                    targets=scheme.targets,
+                    ignore=self.quantization_config.ignore,
+                )
+                missing_keys.update(
+                    merge_names(target_name, "weight")
+                    for target_name, _module in quant_targets
+                )
+
+        return list(missing_keys)
+
     def get_unexpected_file_keys(self, model: Module) -> List[str]:
         """
         Identifies extra keys introduced by the compression process in the

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/src/compressed_tensors/quantization/utils/helpers.py

@@ -165,7 +165,7 @@ def compute_dynamic_scales_and_zp(
 
     keep_dims = True
     if args.strategy == QuantizationStrategy.TOKEN:
-        dim = {1, 2}
+        dim = {0, 1}
         reduce_dims = tuple(idx for idx in range(value.ndim) if idx not in dim)
     elif args.strategy == QuantizationStrategy.TENSOR:
         reduce_dims = None

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/src/compressed_tensors/transform/apply.py

@@ -15,9 +15,9 @@
 from typing import Dict
 
 import torch
-from accelerate.utils import has_offloaded_params
 from compressed_tensors import TRANSFORM_CONFIG_NAME
 from compressed_tensors.transform import TransformConfig, TransformFactory
+from compressed_tensors.utils.offload import has_offloaded_params
 
 
 __all__ = ["apply_transform_config"]

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/src/compressed_tensors/utils/__init__.py

@@ -18,7 +18,6 @@ from .internal import *
 from .match import *
 from .offload import *
 from .permutations_24 import *
-from .permute import *
 from .safetensors_load import *
 from .semi_structured_conversions import *
 from .type import *

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.12.1.a20251001'
-__version_tuple__ = version_tuple = (0, 12, 1)
+__version__ = version = '0.12.2'
+__version_tuple__ = version_tuple = (0, 12, 2)

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/src/compressed_tensors.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.1a20251001
+Version: 0.12.2
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/src/compressed_tensors.egg-info/SOURCES.txt

@@ -10,6 +10,7 @@ setup.py
 .github/scripts/step-status
 .github/workflows/build-test.yml
 .github/workflows/build.yml
+.github/workflows/post-release-nightly-build.yml
 .github/workflows/quality-check.yaml
 .github/workflows/report.yml
 .github/workflows/test-check.yaml
@@ -95,7 +96,6 @@ src/compressed_tensors/utils/internal.py
 src/compressed_tensors/utils/match.py
 src/compressed_tensors/utils/offload.py
 src/compressed_tensors/utils/permutations_24.py
-src/compressed_tensors/utils/permute.py
 src/compressed_tensors/utils/safetensors_load.py
 src/compressed_tensors/utils/semi_structured_conversions.py
 src/compressed_tensors/utils/type.py
@@ -132,7 +132,6 @@ tests/test_quantization/lifecycle/test_apply.py
 tests/test_quantization/lifecycle/test_dynamic_lifecycle.py
 tests/test_quantization/lifecycle/test_enabled.py
 tests/test_quantization/lifecycle/test_forward.py
-tests/test_quantization/lifecycle/test_helpers.py
 tests/test_quantization/lifecycle/test_initialize.py
 tests/test_quantization/lifecycle/test_lifecycle.py
 tests/test_quantization/test_configs/__init__.py

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2}/tests/test_compressors/model_compressors/test_model_compressor.py

@@ -450,6 +450,26 @@ def test_multiple_quant_compressors():
     assert all(format in compressor.quantization_compressor for format in formats)
 
 
+@pytest.mark.parametrize(
+    "model, sparsity_config, quantization_config, expected",
+    [
+        (
+            TwoLayerModel(),
+            get_bitmask_sparsity_config(targets=["re:.*layer1$"]),
+            create_quantization_config(bits=8, type="int", strategy="channel"),
+            {"layer1.weight"},
+        )
+    ],
+)
+def test_get_missing_keys(model, sparsity_config, quantization_config, expected):
+    model_compressor = ModelCompressor(
+        sparsity_config=sparsity_config, quantization_config=quantization_config
+    )
+
+    actual = model_compressor.get_missing_module_keys(model)
+    assert len(actual) == len(expected) and all(key in actual for key in expected)
+
+
 @pytest.mark.parametrize(
     "model, sparsity_config, quantization_config, expected",
     [

compressed_tensors-0.12.1a20251001/src/compressed_tensors/utils/permute.py

@@ -1,33 +0,0 @@
-# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#    http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import torch
-from compressed_tensors.utils.helpers import deprecated
-
-
-__all__ = ["safe_permute"]
-
-
-@deprecated("Tensor.index_select")
-def safe_permute(value: torch.Tensor, perm: torch.Tensor, dim: int = 0) -> torch.Tensor:
-    """
-    Perform out-of-place permutation without using torch.Tensor.index_put_,
-    whose implementation is missing for datatypes such as `torch.float8_e4m3fn`
-
-    :param value: tensor to permute
-    :param perm: permutation map
-    :param dim: dimension along which to apply permutation
-    :return: permuted value
-    """
-    return value.index_select(dim, perm)

compressed_tensors-0.12.1a20251001/tests/test_quantization/lifecycle/test_helpers.py

@@ -1,49 +0,0 @@
-# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#    http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-
-import pytest
-import torch
-from compressed_tensors.utils.permute import safe_permute
-from tests.testing_utils import requires_gpu
-
-
-@requires_gpu
-@pytest.mark.unit
-@pytest.mark.filterwarnings("ignore::DeprecationWarning")
-@pytest.mark.parametrize(
-    "dtype",
-    [
-        torch.int8,
-        torch.int16,
-        torch.int32,
-        torch.int64,
-        torch.bfloat16,
-        torch.float16,
-        torch.float32,
-        torch.float64,
-        torch.float8_e4m3fn,
-    ],
-)
-@pytest.mark.parametrize(
-    "device", [torch.device("cpu"), torch.device("cuda"), torch.device("meta")]
-)
-def test_safe_permute(dtype: torch.dtype, device: torch.device):
-    value = torch.tensor([[0, 1, 2, 3]], dtype=dtype, device=device)
-    perm = torch.tensor([3, 1, 0, 2], device=device)
-
-    result = safe_permute(value, perm, dim=-1)
-
-    if device.type != "meta":
-        assert torch.equal(result.squeeze(0), perm.to(result.dtype))