[Bugfix] Fix precision loss in LoRA-wrapped RowParallelLinear by fusing bias into GEMM

vllm/lora/layers/row_parallel_linear.py

@@ -63,23 +63,18 @@ def forward(
             input_parallel = splitted_input[self.tp_rank].contiguous()
 
         # Matrix multiply.
-        output_parallel = self.apply(input_parallel)
+        bias_ = (
+            None
+            if (self.tp_rank > 0 or self.base_layer.skip_bias_add)
+            else self.base_layer.bias
+        )
+        output_parallel = self.apply(input_parallel, bias_)
         if self.base_layer.reduce_results and self.tp_size > 1:
-            output_ = tensor_model_parallel_all_reduce(output_parallel)
-        else:
-            output_ = output_parallel
-
-        if not self.base_layer.skip_bias_add:
-            output = (
-                output_ + self.base_layer.bias
-                if self.base_layer.bias is not None
-                else output_
-            )
-            output_bias = None
+            output = tensor_model_parallel_all_reduce(output_parallel)
         else:
-            output = output_
-            output_bias = self.base_layer.bias
+            output = output_parallel
 
+        output_bias = self.base_layer.bias if self.base_layer.skip_bias_add else None
         if not self.base_layer.return_bias:
             return output
 
@@ -120,7 +115,7 @@ def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
         return lora_b
 
     def apply(self, x: torch.Tensor, bias: torch.Tensor | None = None) -> torch.Tensor:
-        output = self.base_layer.quant_method.apply(self.base_layer, x)
+        output = self.base_layer.quant_method.apply(self.base_layer, x, bias)
 
         x = x.view(-1, x.shape[-1])
         output, out_orig_shape = output.view(-1, output.shape[-1]), output.shape