fix bug

GPTQ-for-LLaMa/autograd_4bit.py

@@ -131,6 +131,7 @@ class AutogradMatmul4bit(torch.autograd.Function):
     def backward(ctx, grad_output):
         qweight, scales, zeros = ctx.saved_tensors
         buffer = get_buffer(qweight.shape, dtype=scales.dtype, device=qweight.device)
+        quant.quant_cuda.vecquant4recons(qweight, buffer, scales, zeros)
         grad = torch.matmul(grad_output, buffer.T)
         return grad, None, None, None
 
@@ -229,4 +230,4 @@ def load_llama_model_4bit_low_ram(config_path, model_path, half=False):
     print(f"Loaded the model in {(time.time()-t0):.2f} seconds.")
     
     return model, tokenizer
-    
+