# Fix for matmul_4bit out Parameter Issue

Author: ved1beta

bitsandbytes/autograd/_functions.py

@@ -299,9 +299,6 @@ def backward(ctx: torch.autograd.function.FunctionCtx, grad_output: torch.Tensor
 
 
 class MatMul4Bit(torch.autograd.Function):
-    # forward is the same, but we added the fallback for pre-turing GPUs
-    # backward is mostly the same, but adds one extra clause (see "elif state.CxB is not None")
-
     @staticmethod
     def forward(ctx, A, B, out=None, bias=None, quant_state: Optional[F.QuantState] = None):
         # default of pytorch behavior if inputs are empty
@@ -319,7 +316,15 @@ def forward(ctx, A, B, out=None, bias=None, quant_state: Optional[F.QuantState]
 
         # 1. Dequantize
         # 2. MatmulnN
-        output = torch.nn.functional.linear(A, F.dequantize_4bit(B, quant_state).to(A.dtype).t(), bias)
+        # Use linear function which correctly handles 1D and 2D inputs
+        result = torch.nn.functional.linear(A, F.dequantize_4bit(B, quant_state).to(A.dtype).t(), bias)
+
+        # If out is provided, resize it if necessary and copy the result
+        if out is not None:
+            if out.shape != result.shape:
+                out.resize_(result.shape)
+            out.copy_(result)
+            result = out
 
         # 3. Save state
         ctx.state = quant_state
@@ -330,7 +335,7 @@ def forward(ctx, A, B, out=None, bias=None, quant_state: Optional[F.QuantState]
         else:
             ctx.tensors = (None, None)
 
-        return output
+        return result
 
     @staticmethod
     def backward(ctx, grad_output):
@@ -385,9 +390,14 @@ def matmul_4bit(
             )
             return MatMul4Bit.apply(A, B, out, bias, quant_state)
         else:
-            out = F.gemv_4bit(A, B.t(), out, state=quant_state)
+            # For 1D case, we'll use the MatMul4Bit implementation which correctly handles out parameter
+            if out is not None and A.dim() == 1:
+                return MatMul4Bit.apply(A, B, out, bias, quant_state)
+
+            # For other cases, use gemv_4bit
+            result = F.gemv_4bit(A, B.t(), out, state=quant_state)
             if bias is not None:
-                out += bias
-            return out
+                result += bias
+            return result
     else:
         return MatMul4Bit.apply(A, B, out, bias, quant_state)

bitsandbytes/backends/cuda/ops.py

@@ -427,11 +427,18 @@ def _(
     blocksize: int,
     out: torch.Tensor,
 ) -> None:
+    expected_shape = (*A.shape[:-1], shapeB[0])
+
+    if len(A.shape) == 1 and len(out.shape) == 2 and out.shape[0] == 1:
+        out = out.view(shapeB[0])
+        expected_shape = (shapeB[0],)
+
     torch._check(
-        out.shape == (*A.shape[:-1], shapeB[0]),
-        lambda: f"Expected out.shape == {(*A.shape[:-1], shapeB[0])}, got {out.shape}",
+        out.shape == expected_shape,
+        lambda: f"Expected out.shape == {expected_shape}, got {out.shape}",
     )
     torch._check(out.dtype == A.dtype, lambda: f"Expected out.dtype == {A.dtype}, got {out.dtype}")
+
     _gemv_4bit_impl(A, B, shapeB, absmax, code, blocksize, out=out)
 
 
@@ -445,10 +452,13 @@ def _gemv_4bit_impl(
     out: torch.Tensor,
 ) -> None:
     torch._check_is_size(blocksize)
-    torch._check(
-        A.numel() == A.size(-1),
-        lambda: f"A must be a vector with leading dimensions of 1, got {A.shape}",
-    )
+
+    is_1d = A.dim() == 1
+    if is_1d:
+        A_reshaped = A.view(1, -1)
+    else:
+        A_reshaped = A
+
     torch._check(
         A.dtype in [torch.float16, torch.bfloat16, torch.float32],
         lambda: f"A must be float16, bfloat16, or float32, got {A.dtype}",
@@ -465,11 +475,16 @@ def _gemv_4bit_impl(
     k = ct.c_int32(shapeB[1])
 
     lda = m
-    ldb = ct.c_int32((A.shape[-1] + 1) // 2)
+    ldb = ct.c_int32((A_reshaped.shape[-1] + 1) // 2)
     ldc = m
 
     stream = _get_tensor_stream(A)
 
+    if is_1d and out.dim() > 1:
+        out_view = out.view(-1)
+    else:
+        out_view = out
+
     with _cuda_device_of(A):
         if A.dtype == torch.float16:
             lib.cgemm_4bit_inference_naive_fp16(