compressed-tensors 0.12.1a20251001__tar.gz → 0.12.2a20251002__tar.gz

compressed_tensors-0.12.2a20251002/.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-0.12.2a20251002}/.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-0.12.2a20251002}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.1a20251001
+Version: 0.12.2a20251002
 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-0.12.2a20251002}/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-0.12.2a20251002}/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.a20251002'
+__version_tuple__ = version_tuple = (0, 12, 2)

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2a20251002}/src/compressed_tensors.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.1a20251001
+Version: 0.12.2a20251002
 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-0.12.2a20251002}/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

{compressed_tensors-0.12.1a20251001 → compressed_tensors-0.12.2a20251002}/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",
     [