w4a8 默认使用deepgemm的masked接口

vllm/model_executor/layers/fused_moe/batched_deep_gemm_moe.py

@@ -41,9 +41,9 @@ from vllm.model_executor.layers.activation import SiluAndMul
 from lightop import fuse_silu_mul_quant_ep, m_grouped_w4a8_gemm_nt_masked
 from lmslim.layers.gemm.int8_utils import per_token_quant_int8
 if has_deep_gemm():
-    from deepgemm import m_grouped_w8a8_gemm_nt_masked
+    from deepgemm import m_grouped_w8a8_gemm_nt_masked, m_grouped_w4a8_gemm_nt_masked
 else:
-    from lightop import m_grouped_w8a8_gemm_nt_masked
+    from lightop import m_grouped_w8a8_gemm_nt_masked, m_grouped_w4a8_gemm_nt_masked