更新 vllm/_custom_ops.py, vllm/model_executor/layers/quantization/utils/w8a8_utils.py

vllm/_custom_ops.py

@@ -1150,6 +1150,8 @@ def blaslt_scaled_mm(a: torch.Tensor,
     n = b.shape[0]
     k = a.shape[1]
     _, out = quant_ops.hipblaslt_w8a8_gemm(a, b, scale_a, scale_b, m, n, k, 'NT', out_dtype)
+    if bias is not None:
+        out += bias
     return out
 
 def triton_scaled_mm(a: torch.Tensor,
@@ -2486,4 +2488,4 @@ direct_register_custom_op(
     op_func=awq_gemm,
     mutates_args=[],
     fake_impl=awq_gemm_fake,
-)
\ No newline at end of file
+)

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -504,7 +504,7 @@ def apply_int8_linear(
                                     scale_a=x_scale,
                                     scale_b=weight_scale,
                                     out_dtype=input.dtype,
-                                    bias=None)
+                                    bias=bias)
     else:
         return ops.rocblas_scaled_mm(
                 x_q,