Merge pull request #1893 from ailuntz/fix/matmul-4bit-out

Honor out in matmul_4bit

Author: Abdennacer-Badaoui

bitsandbytes/autograd/_functions.py

@@ -318,6 +318,9 @@ 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)
+        if out is not None:
+            out.copy_(output)
+            output = out
 
         # 3. Save state
         ctx.state = quant_state