Fix CI regression

bitsandbytes/__init__.py

@@ -34,7 +34,7 @@
 if torch.cuda.is_available():
     from .backends.cuda import ops as cuda_ops
 
-if torch.xpu.is_available():
+if hasattr(torch, "xpu") and torch.xpu.is_available():
     from .backends.xpu import ops as xpu_ops
 
 

bitsandbytes/backends/utils.py

@@ -30,7 +30,9 @@
         1.0,
     ],
     dtype=torch.float32,
-    device="xpu" if torch.xpu.is_available() else "cpu",  # Only cpu/xpu use this table for now.
+    device="xpu"
+    if hasattr(torch, "xpu") and torch.xpu.is_available()
+    else "cpu",  # Only cpu/xpu use this table for now.
 )
 _FP4_QUANT_TABLE = torch.tensor(
     [
@@ -52,6 +54,8 @@
         -0.2500,
     ],
     dtype=torch.float32,
-    device="xpu" if torch.xpu.is_available() else "cpu",  # Only cpu/xpu use this table for now.
+    device="xpu"
+    if hasattr(torch, "xpu") and torch.xpu.is_available()
+    else "cpu",  # Only cpu/xpu use this table for now.
 )
 CODE = {"nf4": _NF4_QUANT_TABLE, "fp4": _FP4_QUANT_TABLE}

tests/test_functional.py

@@ -1330,6 +1330,9 @@ def test_gemv_4bit(self, device, dim, dtype, storage_type, quant_storage, double
     @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32], ids=describe_dtype)
     @pytest.mark.parametrize("double_quant", [False], ids=["DQ_True"])
     def test_gemv_eye_4bit(self, device, storage_type, dtype, double_quant):
+        if device == "cpu" and dtype == torch.bfloat16 and torch.__version__ < (2, 3):
+            pytest.skip("eye doe not support bfloat16 on CPU in torch < 2.3")
+
         dims = 10
         torch.random.manual_seed(np.random.randint(0, 412424242))
         dims = get_test_dims(0, 8192, n=dims)

tests/test_ops.py

@@ -167,9 +167,8 @@ def test_quantize_4bit(self, device, dtype, storage_dtype, quant_type, blocksize
         assert absmax.device == A.device
         assert absmax.dtype == torch.float32
 
-        # TODO: Enable it
-        if device in ("cpu", "xpu") and storage_dtype == torch.bfloat16:
-            pytest.skip("CPU bf16 storage_dtype will fail on torch op check")
+        if storage_dtype != torch.uint8:
+            pytest.xfail("opcheck fails for storage_dtype != torch.uint8")
 
         opcheck(torch.ops.bitsandbytes.quantize_4bit, (A, blocksize, quant_type, storage_dtype))