Merge remote-tracking branch 'upstream/multi-backend-refactor' into fix_docs

Author: root

.github/workflows/python-package.yml

@@ -58,6 +58,7 @@ jobs:
   # This job matrix builds the CUDA versions of the libraries for platforms that support CUDA (Linux x64/aarch64 + Windows x64)
   ##
   build-shared-libs-cuda:
+    if: github.ref_name != 'multi-backend-refactor'
     strategy:
       matrix:
         os: [ubuntu-latest, windows-latest]
@@ -148,7 +149,7 @@ jobs:
   build-wheels:
     needs:
       - build-shared-libs
-      - build-shared-libs-cuda
+      # - build-shared-libs-cuda  reduce the pkg size + build times for the preview release
       - build-shared-libs-rocm
     strategy:
       matrix:
@@ -166,6 +167,13 @@ jobs:
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          fetch-depth: 1  # shallow clone
+      - name: Fetch tags for dynamic versioning in setup.py
+        run: |
+          git fetch --depth=1 origin --tags
+          echo "Available Git tags:"
+          git tag -n
       - name: Download build artifact
         uses: actions/download-artifact@v4
         with:
@@ -183,7 +191,8 @@ jobs:
           python-version: ${{ matrix.python-version }}
           cache: pip
       - run: pip install build wheel
-      - run: python -m build .
+      # for now need to do the below instead of prior `python -m build .`, which didn't allow us to access git tags
+      - run: python -m build --sdist && python -m build --wheel
       - name: Determine and Set Platform Tag, then Tag Wheel
         shell: bash
         run: |
@@ -197,6 +206,45 @@ jobs:
           path: dist/bitsandbytes-*.whl
           retention-days: 7
 
+  upload-pre-release-wheels:
+    name: Create release and upload artifacts
+    runs-on: ubuntu-latest
+    if: github.ref_name == 'multi-backend-refactor'
+    permissions:
+      contents: write
+    needs:
+      - build-wheels
+    steps:
+      - name: Download and rename artifacts
+        uses: actions/download-artifact@v4
+        with:
+          path: tmp/
+          pattern: "bdist_wheel_*"
+          merge-multiple: true
+      - name: Inspect tmp directory after downloading artifacts
+        run: ls -alFR tmp/
+      - name: Move and rename wheel files with pattern replacement
+        run: |
+          mkdir -p wheels/
+          find tmp/ -type f -name '*.whl' -print0 | while IFS= read -r -d '' wheel; do
+            wheel_filename=$(basename "$wheel")
+            # Remove the gith hash, e.g. `+1234567`, for a stable download link on the multi-backend pre-release
+            cleaned_filename=$(echo "$wheel_filename" | sed -E 's/\+[0-9a-f]{7}-/-/g')
+            mv "$wheel" "wheels/$cleaned_filename"
+          done
+      - name: Inspect wheels directory after renaming files
+        run: ls -alFR wheels/
+      - name: Create release and upload artifacts
+        uses: softprops/[email protected]
+        with:
+          files: wheels/*.whl
+          prerelease: true
+          name: Multi-Backend Preview
+          tag_name: continuous-release_multi-backend-refactor
+          make_latest: false
+          draft: false
+          target_commitish: ${{ github.sha }}
+
   audit-wheels:
     needs: build-wheels
     runs-on: ubuntu-latest

.gitignore

@@ -151,6 +151,8 @@ dmypy.json
 # vim
 *.swp
 
+# BNB-specific stuff
 dependencies
 cuda_build
 output/
+bitsandbytes/_version.py

CMakeLists.txt

@@ -82,6 +82,11 @@ if(BUILD_CUDA)
     # This needs to be added *before* we try to enable the CUDA language so CMake's compiler check passes.
     if(MSVC AND MSVC_VERSION VERSION_GREATER_EQUAL 1940)
         string(APPEND CMAKE_CUDA_FLAGS " --allow-unsupported-compiler")
+
+        # This is needed to build with VS2022 17.11+ and CUDA < 12.4.
+        if (MSVC_VERSION VERSION_GREATER_EQUAL 1941)
+            string(APPEND CMAKE_CUDA_FLAGS " -D_ALLOW_COMPILER_AND_STL_VERSION_MISMATCH")
+        endif()
     endif()
 
     enable_language(CUDA) # This will fail if CUDA is not found

bitsandbytes/__init__.py

@@ -3,6 +3,9 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
+# Import the dynamically generated version from _version.py  (see setup.py)
+from ._version import __version__  # isort: skip # type: ignore
+
 import torch
 
 from . import research, utils
@@ -73,5 +76,3 @@
     "optim.optimizer.Optimizer8bit": False,
     "optim.optimizer.MockArgs": False,
 }
-
-__version__ = "0.43.3.dev"

bitsandbytes/nn/modules.py

@@ -471,6 +471,7 @@ def forward(self, x: torch.Tensor):
             and not hasattr(self.weight.quant_state, "op_context")
             and self.weight.quant_state.shape[1] % self.weight.quant_state.blocksize == 0
             and self.weight.quant_state.quant_type == "nf4"
+            and x.requires_grad == False
         ):
             enable_ipex_fusion(self.weight, self.weight.quant_state)
 

docs/source/contributing.mdx

@@ -5,8 +5,9 @@
 
 ### Setup pre-commit hooks
 - Install pre-commit hooks with `pip install pre-commit`.
-- Run `pre-commit autoupdate` once to configure the hooks.
-- Re-run `pre-commit autoupdate` every time a new hook got added.
+- Run `pre-commit install` once to install the hooks, so they will be run on every commit.
+- If the hooks introduce changes, they'll be visible with `git diff`. Review them and `git add` them if everything is fine, then re-execute the before commit, it should pass now.
+- If you want to manually trigger the hooks, you may do `pre-commit run --all-files`
 
 Now all the pre-commit hooks will be automatically run when you try to commit and if they introduce some changes, you need to re-add the changed files before being able to commit and push.