Merge remote-tracking branch 'origin' into kylesayrs/reduce-quantized-compression-memory

Author: kylesayrs

.github/workflows/build-test.yml

@@ -21,7 +21,7 @@ on:
       build_label:
         description: "requested runner label for build (specifies instance)"
         type: string
-        default: ubuntu-24.04
+        default: ubuntu-22.04
 
       # test related parameters
       test_configs:
@@ -67,6 +67,7 @@ jobs:
 
     REPORT:
         needs: [BUILD, TEST]
+        if: success() || failure()
         uses: ./.github/workflows/report.yml
         with:
             label: rh-reporter

.github/workflows/build.yml

@@ -70,11 +70,13 @@ jobs:
               id: checkout
               uses: actions/checkout@v4
               with:
+                  fetch-depth: 0
+                  fetch-tags: true
                   ref: ${{ inputs.gitref }}
 
             - name: build
               id: build
-              uses: neuralmagic/nm-actions/actions/build-ml-whl@v1.12.0
+              uses: neuralmagic/nm-actions/actions/build-ml-whl@v1.18.0
               with:
                   dev: false
                   release: ${{ inputs.wf_category == 'RELEASE' }}

.github/workflows/upload.yml

@@ -97,28 +97,36 @@ jobs:
                 run_id: ${{ inputs.run_id }}
                 asset_identifier: 'compressed*.whl'
 
+            - name: check if whl is new
+              id: check-whl
+              uses: neuralmagic/nm-actions/actions/check-whl-on-pypi@v1.19.0
+              with:
+                whl: ${{ steps.find-asset-whl.outputs.asset }}
+
             - name: upload whl to "nm-pypi"
-              if: ${{ inputs.push_to_pypi }}
+              if: ${{ steps.check-whl.outputs.status == '0' && inputs.push_to_pypi }}
               uses: neuralmagic/nm-actions/actions/gcp-upload-asset@v1.1.0
               with:
                 bucket_target: ${{ secrets.GCP_NM_PYPI_DIST }}
                 asset: ${{ steps.find-asset-whl.outputs.asset }}
 
             - name: find tar.gz
+              if: ${{ steps.check-whl.outputs.status == '0' && inputs.push_to_pypi }}
               id: find-asset-targz
               uses: neuralmagic/nm-actions/actions/find-asset@v1.1.0
               with:
                 run_id: ${{ inputs.run_id }}
                 asset_identifier: 'compressed*.tar.gz'
 
             - name: upload tar.gz to "nm-pypi"
-              if: ${{ inputs.push_to_pypi }}
+              if: ${{ steps.check-whl.outputs.status =='0' && inputs.push_to_pypi }}
               uses: neuralmagic/nm-actions/actions/gcp-upload-asset@v1.1.0
               with:
                 bucket_target: ${{ secrets.GCP_NM_PYPI_DIST }}
                 asset: ${{ steps.find-asset-targz.outputs.asset }}
 
             - name: update "nm-pypi" index
+              if: ${{ steps.check-whl.outputs.status == '0' && inputs.push_to_pypi }}
               uses: actions/github-script@v6
               with:
                 github-token: ${{ secrets.NM_PYPI_WORKFLOW }}
@@ -133,7 +141,7 @@ jobs:
 
             # publish the wheel file to public pypi
             - name: push wheel to pypi.org
-              if: ${{ inputs.push_to_pypi }}
+              if: ${{ steps.check-whl.outputs.status == '0' && inputs.push_to_pypi }}
               uses: neuralmagic/nm-actions/actions/publish-whl@v1.0.0
               with:
                 username: ${{ secrets.PYPI_PUBLIC_USER }}
@@ -142,7 +150,7 @@ jobs:
 
             # publish the tar.gz file to public pypi
             - name: push wheel to pypi.org
-              if: ${{ inputs.push_to_pypi }}
+              if: ${{ steps.check-whl.outputs.status == '0' && inputs.push_to_pypi }}
               uses: neuralmagic/nm-actions/actions/publish-whl@v1.0.0
               with:
                 username: ${{ secrets.PYPI_PUBLIC_USER }}

.gitignore

@@ -25,6 +25,7 @@ share/python-wheels/
 .installed.cfg
 *.egg
 MANIFEST
+src/compressed_tensors/version.py
 
 # PyInstaller
 #  Usually these files are written by a python script from a template

Makefile

@@ -1,6 +1,6 @@
 .PHONY: build docs test
 
-BUILD_ARGS :=  # set nightly to build nightly release
+BUILD_TYPE ?= dev # set nightly to build nightly release
 PYCHECKDIRS := src tests
 PYCHECKGLOBS := 'src/**/*.py' 'tests/**/*.py' 'utils/**/*.py' 'examples/**/*.py' setup.py
 # run checks on all files for the repo
@@ -28,10 +28,10 @@ test:
 # creates wheel file
 build:
 	@echo "Building the wheel for the repository";
-	python3 setup.py sdist bdist_wheel $(BUILD_ARGS);
+	BUILD_TYPE=$(BUILD_TYPE) python3 setup.py sdist bdist_wheel;
 
 # clean package
 clean:
 	@echo "Cleaning up";
 	rm -rf .pytest_cache;
-	find $(PYCHECKDIRS) | grep -E "(__pycache__|\.pyc|\.pyo)" | xargs rm -rf;
+	find $(PYCHECKDIRS) | grep -E "(__pycache__|\.pyc|\.pyo)" | xargs rm -rf;

README.md

@@ -31,7 +31,7 @@ pip install compressed-tensors
 
 Nightly release:
 ```bash
-pip install compressed-tensors-nightly
+pip install --pre compressed-tensors
 ```
 
 ### From Source

pyproject.toml

@@ -1,3 +1,10 @@
+[build-system]
+requires = ["setuptools", "wheel", "setuptools_scm>8"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools_scm]
+version_file = "src/compressed_tensors/version.py"
+
 [tool.black]
 line-length = 88
 target-version = ['py36']

setup.py

@@ -17,50 +17,66 @@
 from typing import List, Dict, Tuple
 
 
-def get_release_and_version(package_path: str) -> Tuple[bool, bool, str, str, str, str]:
-    """
-    Load version and release info from compressed-tensors package
-    """
-    # compressed-tensors/src/compressed_tensors/version.py always exists, default source of truth
-    version_path = os.path.join(package_path, "version.py")
-
-    # exec() cannot set local variables so need to manually
-    locals_dict = {}
-    exec(open(version_path).read(), globals(), locals_dict)
-    is_release = locals_dict.get("is_release", False)
-    version = locals_dict.get("version", "unknown")
-    version_major = locals_dict.get("version_major", "unknown")
-    version_minor = locals_dict.get("version_minor", "unknown")
-    version_bug = locals_dict.get("version_bug", "unknown")
-
-    print(f"Loaded version {version} from {version_path}")
-
-    return (
-        is_release,
-        version,
-        version_major,
-        version_minor,
-        version_bug,
+# Set the build type using an environment variable to give us
+# different package names based on the reason for the build.
+VALID_BUILD_TYPES = {"release", "nightly", "dev"}
+BUILD_TYPE = os.environ.get("BUILD_TYPE", "dev")
+if BUILD_TYPE not in VALID_BUILD_TYPES:
+    raise ValueError(
+        f"Unsupported build type {BUILD_TYPE!r}, must be one of {VALID_BUILD_TYPES}"
     )
 
+from setuptools_scm import ScmVersion
+
+def version_func(version: ScmVersion) -> str:
+    from setuptools_scm.version import guess_next_version
+
+    if BUILD_TYPE == "nightly":
+        # Nightly builds use alpha versions to ensure they are marked
+        # as pre-releases on pypi.org.
+        return version.format_next_version(
+            guess_next=guess_next_version,
+            fmt="{guessed}.a{node_date:%Y%m%d}",
+        )
+
+    if (
+        BUILD_TYPE == "release"
+        and not version.dirty
+        and (version.exact or version.node is None)
+    ):
+        # When we have a tagged version, use that without modification.
+        return version.format_with("{tag}")
+
+    # In development mode or when the local repository is dirty, treat
+    # it is as local development version.
+    return version.format_next_version(
+        guess_next=guess_next_version,
+        fmt="{guessed}.dev{distance}",
+    )
 
-package_path = os.path.join(
-    os.path.dirname(os.path.realpath(__file__)), "src", "compressed_tensors"
-)
-(
-    is_release,
-    version,
-    version_major,
-    version_minor,
-    version_bug,
-) = get_release_and_version(package_path)
 
-version_nm_deps = f"{version_major}.{version_minor}.0"
+def localversion_func(version: ScmVersion) -> str:
+    from setuptools_scm.version import get_local_node_and_date
+
+    # When we are building nightly versions, we guess the next release
+    # and add the date as an alpha version. We cannot publish packages
+    # with local versions, so we do not add one.
+    if BUILD_TYPE == "nightly":
+        return ""
+
+    # When we have an exact tag, with no local changes, do not append
+    # anything to the local version field.
+    if (
+        BUILD_TYPE == "release"
+        and not version.dirty
+        and (version.exact or version.node is None)
+    ):
+        return ""
 
-if is_release:
-    _PACKAGE_NAME = "compressed-tensors"
-else:
-    _PACKAGE_NAME = "compressed-tensors-nightly"
+    # In development mode or when the local repository is dirty,
+    # return a string that includes the git SHA (node) and a date,
+    # formatted as a local version tag.
+    return get_local_node_and_date(version)
 
 
 def _setup_long_description() -> Tuple[str, str]:
@@ -81,8 +97,11 @@ def _setup_extras() -> Dict:
     }
 
 setup(
-    name=_PACKAGE_NAME,
-    version=version,
+    name="compressed-tensors",
+    use_scm_version={
+        "version_scheme": version_func,
+        "local_scheme": localversion_func,
+    },
     author="Neuralmagic, Inc.",
     author_email="support@neuralmagic.com",
     license="Apache 2.0",

src/compressed_tensors/compressors/quantized_compressors/base.py

@@ -18,7 +18,7 @@
 
 import torch
 from compressed_tensors.compressors.base import BaseCompressor
-from compressed_tensors.quantization import QuantizationScheme
+from compressed_tensors.quantization import QuantizationStrategy
 from compressed_tensors.utils import (
     get_nested_mappings_from_state_dict,
     get_nested_weight_mappings,
@@ -134,6 +134,17 @@ def compress(
 
         return compressed_dict
 
+    def _check_if_zp_pack_quantized(self, quant_args):
+        from compressed_tensors.compressors import PackedQuantizationCompressor
+
+        if isinstance(self, PackedQuantizationCompressor):
+            if not quant_args.symmetric and quant_args.strategy in [
+                QuantizationStrategy.GROUP.value,
+                QuantizationStrategy.CHANNEL.value,
+            ]:
+                return True
+        return False
+
     def decompress(
         self,
         path_to_model_or_tensors: Union[str, Path, Dict[str, Any]],