Merge branch 'main' into matmul4bit

Author: ved1beta

.github/workflows/tests.yml

@@ -93,24 +93,32 @@ jobs:
           path: output/${{ matrix.os }}/${{ matrix.arch }}/*
           retention-days: 7
 
-  cpu-tests:
+  test-cpu:
     if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
     needs: build-cpu
     strategy:
       fail-fast: false
       matrix:
         os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025, macos-15]
-        torch_version: ["2.6.0", "2.7.0"]
+        # Test with the oldest supported torch version and the two newest.
+        torch_version: ["2.2.2", "2.6.0", "2.7.0"]
         include:
           - os: ubuntu-22.04
             arch: x86_64
             runner: banb-aws-general-8-plus-use1-public-80
           - os: ubuntu-22.04-arm
             arch: aarch64
+          - os: ubuntu-22.04-arm
+            arch: aarch64
+            torch_version: "2.5.1"
           - os: windows-2025
             arch: x86_64
           - os: macos-15
             arch: arm64
+        exclude:
+          - os: ubuntu-22.04-arm
+            torch_version: "2.2.2"
+
     runs-on: ${{ matrix.runner || matrix.os }}
     env:
       BNB_TEST_DEVICE: cpu
@@ -129,22 +137,71 @@ jobs:
         with:
           python-version: 3.9
 
+      - name: Setup MSVC
+        if: startsWith(matrix.os, 'windows')
+        uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl for torch.compile
+
       - name: Install dependencies
         run: |
           pip install torch==${{ matrix.torch_version }} --index-url https://download.pytorch.org/whl/cpu
           pip install -e ".[test]"
           pip install pytest-cov
 
+      # We need to downgrade to numpy<2 for torch<2.3 compatibility.
+      - name: Downgrade NumPy
+        if: startsWith(matrix.torch_version, '2.2.')
+        run: pip install "numpy<2"
+
+      - name: Show installed packages
+        run: pip list
+
+      - name: Show environment information
+        run: python -m torch.utils.collect_env
+
+      - name: Run tests
+        run: pytest --durations=100
+
+  test-cpu-ipex:
+    if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
+    needs: build-cpu
+    runs-on: banb-aws-general-8-plus-use1-public-80
+    env:
+      BNB_TEST_DEVICE: cpu
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Download build artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: lib_cpu_ubuntu-22.04_x86_64
+          path: bitsandbytes/
+          merge-multiple: true
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: 3.9
+
+      - name: Install dependencies
+        run: |
+          pip install torch==2.7.0 --index-url https://download.pytorch.org/whl/cpu
+          pip install intel_extension_for_pytorch==2.7.0 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/cpu/us/
+          pip install -e ".[test]"
+          pip install pytest-cov
+
       - name: Show installed packages
         run: pip list
 
       - name: Show environment information
         run: python -m torch.utils.collect_env
 
+      - name: IPEX smoke test
+        run: python -c "import torch; import intel_extension_for_pytorch as ipex; print(torch.__version__); print(ipex.__version__);"
+
       - name: Run tests
         run: pytest --durations=100
 
-  # cuda-aarch64-tests:
+  # test-cuda-aarch64:
   #   if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
   #   needs: build-cuda
   #   strategy:
@@ -165,9 +222,7 @@ jobs:
   #     - name: Show pip packages
   #       run: pip list
 
-
-
-  cuda-tests:
+  test-cuda:
     if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
     needs: build-cuda
     strategy:
@@ -179,7 +234,7 @@ jobs:
         cuda_version: ["11.8.0", "12.6.3", "12.8.1"]
         include:
           - cuda_version: "11.8.0"
-            torch_version: "2.4.1"
+            torch_version: "2.2.2"
             pypi_index: "https://download.pytorch.org/whl/cu118"
           - cuda_version: "12.6.3"
             torch_version: "2.6.0"
@@ -188,18 +243,40 @@ jobs:
             torch_version: "2.7.0"
             pypi_index: "https://download.pytorch.org/whl/cu128"
 
-          # L40S runners
+
+          # Linux L40S runners
           - os: ubuntu-22.04
             gpu: L40S
             runner: bandb-aws-g6e-4xlarge-plus-use1-public-80
 
-          # T4 runners
+          # Linux T4 runners
           - os: ubuntu-22.04
             gpu: T4
             runner: bandb-aws-g4dn-4xlarge-plus-use1-public-80
+
+          # Specific Windows runners using cu118
           - os: windows-2025
+            arch: x86_64
             gpu: T4
             runner: CUDA-Windows-x64
+            cuda_version: "11.8.0"
+            torch_version: "2.2.0"
+            pypi_index: "https://download.pytorch.org/whl/cu118"
+          - os: windows-2025
+            arch: x86_64
+            gpu: T4
+            runner: CUDA-Windows-x64
+            cuda_version: "11.8.0"
+            torch_version: "2.6.0"
+            pypi_index: "https://download.pytorch.org/whl/cu118"
+          - os: windows-2025
+            arch: x86_64
+            gpu: T4
+            runner: CUDA-Windows-x64
+            cuda_version: "11.8.0"
+            torch_version: "2.7.0"
+            pypi_index: "https://download.pytorch.org/whl/cu118"
+
         exclude:
           # Our current T4 Windows runner has a driver too old (471.11)
           # and cannot support CUDA 12+. Skip for now.
@@ -238,6 +315,11 @@ jobs:
           pip install -e ".[test]"
           pip install pytest-cov
 
+        # We need to downgrade to numpy<2 for torch<2.3 compatibility.
+      - name: Downgrade NumPy
+        if: startsWith(matrix.torch_version, '2.2.')
+        run: pip install "numpy<2"
+
       - name: Show installed packages
         run: pip list
 

benchmarking/int8/row_scale_benchmark.py

@@ -34,6 +34,9 @@
 if torch.cuda.is_available():
     from .backends.cuda import ops as cuda_ops
 
+if hasattr(torch, "xpu") and torch.xpu.is_available():
+    from .backends.xpu import ops as xpu_ops
+
 
 def _import_backends():
     """
@@ -64,4 +67,4 @@ def _import_backends():
     "optim.optimizer.MockArgs": False,
 }
 
-__version__ = "0.46.0.dev0"
+__version__ = "0.47.0.dev0"

bitsandbytes/__init__.py

@@ -4,6 +4,8 @@
 
 import torch
 
+from .cextension import ipex_cpu, ipex_xpu
+
 _IS_TORCH_GTE_24 = False
 
 if hasattr(torch.library, "register_fake"):
@@ -327,3 +329,22 @@ def _(
     )
     torch._check(out.device == A.device, lambda: f"Expected out.device == {A.device}, got {out.device}")
     torch._check(out.dtype == A.dtype, lambda: f"Expected out.dtype == {A.dtype}, got {out.dtype}")
+
+
+if ipex_cpu or ipex_xpu:
+    # Register the dequantize_nf4_ipex implementation
+    torch.library.define(
+        "bitsandbytes::dequantize_nf4_ipex",
+        "(Tensor A, Tensor absmax, int blocksize, int[] shape, ScalarType dtype) -> Tensor",
+    )
+
+    @register_fake("bitsandbytes::dequantize_nf4_ipex")
+    def _(
+        A: torch.Tensor,
+        absmax: torch.Tensor,
+        blocksize: int,
+        shape: Sequence[int],
+        dtype: torch.dtype,
+    ) -> torch.Tensor:
+        torch._check_is_size(blocksize)
+        return torch.empty(shape, dtype=dtype, device=A.device)