Small bugfix for int8 test

Author: matthewdouglas

Diff for: bitsandbytes/autograd/_functions.py

@@ -284,7 +284,7 @@ def backward(ctx: torch.autograd.function.FunctionCtx, grad_output: torch.Tensor
                 dtype=torch.float16,
             )
 
-            if state.threshold > 0.0 and subA is not None:
+            if state.threshold > 0.0 and subA is not None and subA.numel() > 0:
                 grad_B[:, idx] += torch.matmul(grad_output.t(), subA)
 
         if req_gradA:

Diff for: tests/test_linear8bitlt.py

@@ -67,13 +67,15 @@ def test_linear_no_igemmlt(device):
 
 @pytest.mark.parametrize("device", get_available_devices())
 @pytest.mark.parametrize("has_fp16_weights", TRUE_FALSE, ids=id_formatter("has_fp16_weights"))
+@pytest.mark.parametrize("threshold", [0.0, 6.0], ids=id_formatter("threshold"))
 @pytest.mark.parametrize("serialize_before_forward", TRUE_FALSE, ids=id_formatter("serialize_before_forward"))
 @pytest.mark.parametrize("deserialize_before_cuda", TRUE_FALSE, ids=id_formatter("deserialize_before_cuda"))
 @pytest.mark.parametrize("save_before_forward", TRUE_FALSE, ids=id_formatter("save_before_forward"))
 @pytest.mark.parametrize("load_before_cuda", TRUE_FALSE, ids=id_formatter("load_before_cuda"))
 def test_linear_serialization(
     device,
     has_fp16_weights,
+    threshold,
     serialize_before_forward,
     deserialize_before_cuda,
     save_before_forward,
@@ -92,7 +94,7 @@ def test_linear_serialization(
         linear.out_features,
         linear.bias is not None,
         has_fp16_weights=has_fp16_weights,
-        threshold=6.0,
+        threshold=threshold,
     )
 
     linear_custom.weight = bnb.nn.Int8Params(
@@ -137,7 +139,7 @@ def test_linear_serialization(
         linear.out_features,
         linear.bias is not None,
         has_fp16_weights=has_fp16_weights,
-        threshold=6.0,
+        threshold=threshold,
     )
 
     if deserialize_before_cuda: