more cleanup

bitsandbytes/functional.py

@@ -2388,9 +2388,9 @@ def int8_linear_matmul(A: torch.Tensor, B: torch.Tensor, out: Optional[torch.Ten
     if has_error:
         raise RuntimeError(
             f"cublasLt ran into an error!\n"
-            f"\tA: {shapeA}, B: {shapeB}, C: {shapeC}\n"
-            f"\t(lda, ldb, ldc): {(lda, ldb, ldc)}\n"
-            f"\t(m, n, k): {(m, n, k)}"
+            f"\t{shapeA=}, {shapeB=}, {shapeC=}\n"
+            f"\t{(lda, ldb, ldc)=}\n"
+            f"\t{(m, n, k)=}"
         )
 
     return out

bitsandbytes/research/autograd/_functions.py

@@ -256,8 +256,6 @@ def forward(ctx, A, B, out=None, bias=None, state: Optional[MatmulLtState] = Non
         if outlier_cols is not None and not state.has_fp16_weights:
             # extract outliers
             state.idx = outlier_cols
-
-            # outliers = F.extract_outliers(state.CxB, state.SB, state.idx.int())
             outliers = state.CB[:, state.idx.long()].clone()
             state.subB = (outliers * state.SCB.view(-1, 1) / 127.0).t().contiguous().to(A.dtype)
             CA[:, state.idx.long()] = 0

tests/test_autograd.py

@@ -255,14 +255,6 @@ def test_matmullt(dim1, dim2, dim3, dim4, funcs, dtype, req_grad, transpose, dec
                     B2 = B2.t().contiguous()
 
                 state.CB, state.SCB, _ = bnb.functional.int8_vectorwise_quant(B2.to(torch.float16))
-
-                # (
-                #     state.CB,
-                #     CBt,
-                #     state.SCB,
-                #     SCBt,
-                #     coo_tensorB,
-                # ) = bnb.functional.double_quant(B2.to(torch.float16))
                 B2 = state.CB
 
             if not transpose[0] and transpose[1]: