Merge branch 'main' into int8

Author: matthewdouglas

.github/workflows/python-package.yml

@@ -163,45 +163,36 @@ jobs:
     needs:
       - build-wheels
     steps:
-      - name: Download artifacts to tmp directory
+      - 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
+      - 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
+          # exclude macos wheels for now
+          find tmp/ -type f -name '*.whl' ! -name '*macos*' -print0 | while IFS= read -r -d '' wheel; do
             wheel_filename=$(basename "$wheel")
-            if [[ $wheel_filename == *linux*x86_64* ]]; then
-              mv "$wheel" wheels/bnb-linux-x86_64.whl
-            elif [[ $wheel_filename == *linux*aarch64* ]]; then
-              mv "$wheel" wheels/bnb-linux-aarch64.whl
-            elif [[ $wheel_filename == *macosx*x86_64* ]]; then
-              mv "$wheel" wheels/bnb-macos-x86_64.whl
-            elif [[ $wheel_filename == *macosx*arm64* ]]; then
-              mv "$wheel" wheels/bnb-macos-arm64.whl
-            elif [[ $wheel_filename == *win*amd64* ]]; then
-              mv "$wheel" wheels/bnb-windows-x86_64.whl
-            else
-              echo "Unknown wheel format: $wheel_filename"
-              exit 1
-            fi
+            # Remove the gith hash, e.g. `+1234567`, for a stable download link on the multi-backend pre-release
+            cleaned_filename=$(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
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          GITHUB_CONTINUOUS_RELEASE_TYPE: prerelease
-          GITHUB_CONTINUOUS_RELEASE_TAG: continuous-release_main
-        run: |
-          wget -q https://github.com/TheAssassin/pyuploadtool/releases/download/continuous/pyuploadtool-x86_64.AppImage
-          chmod +x pyuploadtool-x86_64.AppImage
-          ./pyuploadtool-x86_64.AppImage --appimage-extract-and-run wheels/*.whl
+        uses: softprops/[email protected]
+        with:
+          files: wheels/*.whl
+          prerelease: true
+          name: Latest `main` wheel
+          tag_name: continuous-release_main
+          make_latest: false
+          draft: false
+          target_commitish: ${{ github.sha }}
 
   audit-wheels:
     needs: build-wheels

.pre-commit-config.yaml

@@ -1,13 +1,13 @@
 repos:
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.3.2
+    rev: v0.6.9
     hooks:
       - id: ruff
         args:
           - --fix
       - id: ruff-format
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.5.0
+    rev: v5.0.0
     hooks:
       - id: check-merge-conflict
       - id: check-yaml
@@ -18,6 +18,6 @@ repos:
         args:
           - --fix=lf
   - repo: https://github.com/crate-ci/typos
-    rev: v1.18.2
+    rev: v1.26.0
     hooks:
       - id: typos

README.md

@@ -12,15 +12,19 @@ There are ongoing efforts to support further hardware backends, i.e. Intel CPU +
 
 **[https://huggingface.co/docs/bitsandbytes/main](https://huggingface.co/docs/bitsandbytes/main)**
 
-## ALPHA TESTERS WANTED: `multi-backend-refactor` AMD GPU + Intel CPU/GPU specific BNB backend implementations
+## `bitsandbytes` multi-backend _alpha_ release is out!
 
-We're in the process of a complex refactor in order to allow the support of additional hardware backends, other than CUDA, in BNB. The efforts around this are already quite far along and there's plenty of functionality already in place that is in need for users to take a hands-on approach! Mac support will likely soon also see progress. However, I recommend waiting 2 weeks until the device abstraction has further consolidated (**breaking changes upcoming**).
+🚀 Big news! After months of hard work and incredible community contributions, we're thrilled to announce the **bitsandbytes multi-backend _alpha_ release**! 💥
 
-Currently, you still need to compile from source, after checking out the `multi-backend-refactor` branch (instructions WIP, but [the current docs on the compilation from source](https://huggingface.co/docs/bitsandbytes/main/en/installation#compile-from-source) are a good starting point; [feel free to share tips / input in this Github discussion](https://github.com/TimDettmers/bitsandbytes/discussions/1219). We'll soon enable nightly releases to make this much easier for you!
+Now supporting:
+- 🔥 **AMD GPUs** (ROCm)
+- ⚡ **Intel CPUs** & **GPUs**
 
-Please give feedback to us in [this dedicated Github Discussion space](https://github.com/TimDettmers/bitsandbytes/discussions/categories/catch-all-alpha-testing-the-multi-backend-refactor)!
+We’d love your early feedback! 🙏
 
-We're super excited about these recent developments and grateful for any constructive input or support that you can give to help us make this a reality. BNB is a community project and we're excited for your collaboration 🤗
+👉 [Instructions for your `pip install` here](https://huggingface.co/docs/bitsandbytes/main/en/installation#multi-backend)
+
+We're super excited about these recent developments and grateful for any constructive input or support that you can give to help us make this a reality (e.g. helping us with the upcoming Apple Silicon backend or reporting bugs). BNB is a community project and we're excited for your collaboration 🤗
 
 ## License
 

_typos.toml

@@ -4,8 +4,10 @@
 extend-ignore-re = [
     "@Ther-nul",  # valid Github user
 ]
-
-[default.extend-identifiers]
+extend-ignore-identifiers-re = [
+    ".*arange.*",
+    ".*ARANGE.*",
+]
 
 [type.py.extend-words]
 "BA" = "BA"  # used as a commented-out variable in tests

bitsandbytes/functional.py

@@ -1860,7 +1860,7 @@ def percentile_clipping(grad: Tensor, gnorm_vec: Tensor, step: int, percentile:
     gnorm_vec: torch.Tensor
         Vector of gradient norms. 100 elements expected.
     step: int
-        The current optimiation steps (number of past gradient norms).
+        The current optimization steps (number of past gradient norms).
 
     """
     prev_device = pre_call(grad.device)

csrc/kernels.cu

@@ -2703,7 +2703,7 @@ template <int THREADS, int ITEMS_PER_THREAD, int TILE_ROWS, int TILE_COLS, int T
                   //const int global_col = base_row; // block offset for col
                   if((base_col + subrow_loop_row + jrow + warp_id < outRows) && (base_row+warp_lane < rows))
                   {
-                    // each row hae 32 columns and is offset by 1 to prevent bank conflict during storage into smem
+                    // each row has 32 columns and is offset by 1 to prevent bank conflict during storage into smem
                     char data = smem_data[(subrow_loop_row + jrow + warp_id)*33 + warp_lane];
 
                     // each 32 columns we have new tile
@@ -2742,7 +2742,7 @@ template <int THREADS, int ITEMS_PER_THREAD, int TILE_ROWS, int TILE_COLS, int T
                   //const int global_col = base_row; // block offset for col
                   if((base_col + subrow_loop_row + jrow + warp_id < outRows) && (base_row+warp_lane < rows))
                   {
-                    // each row hae 32 columns and is offset by 1 to prevent bank conflict during storage into smem
+                    // each row has 32 columns and is offset by 1 to prevent bank conflict during storage into smem
                     char data = smem_data[(subrow_loop_row + jrow + warp_id)*33 + warp_lane];
 
                     // each 32 columns we have new tile
@@ -2819,7 +2819,7 @@ template <int THREADS, int ITEMS_PER_THREAD, int TILE_ROWS, int TILE_COLS, int T
 										//const int global_col = base_row; // block offset for col
 										if((base_col + subrow_loop_row + jrow + warp_id < outRows) && (base_row+warp_lane < rows))
 										{
-											// each row hae 32 columns and is offset by 1 to prevent bank conflict during storage into smem
+											// each row has 32 columns and is offset by 1 to prevent bank conflict during storage into smem
 											char data = smem_data[(subrow_loop_row + jrow + warp_id)*33 + warp_lane];
 
 											// each 32 columns we have new tile