diff --git a/fastdeploy/model_executor/layers/quantization/fp8_utils.py b/fastdeploy/model_executor/layers/quantization/fp8_utils.py
index 65d30d4004d..52cc1665386 100644
--- a/fastdeploy/model_executor/layers/quantization/fp8_utils.py
+++ b/fastdeploy/model_executor/layers/quantization/fp8_utils.py
@@ -80,8 +80,10 @@ def load_deep_gemm():
 
                 logger.info("Detected sm100, use PFCC DeepGEMM")
         else:
-            logger.info("use FastDeploy DeepGEMM")
-            import fastdeploy.model_executor.ops.gpu.deep_gemm as deep_gemm
+            paddle.enable_compat(scope={"deep_gemm"})
+            import deep_gemm as deep_gemm
+
+            logger.info("Use PFCC DeepGEMM")
     else:
         deep_gemm = None
     return deep_gemm
diff --git a/requirements.txt b/requirements.txt
index 66ed714045b..ff1ae91aceb 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -49,4 +49,5 @@ py-cpuinfo
 flashinfer-python-paddle
 flash_mask @ https://paddle-qa.bj.bcebos.com/ernie/flash_mask-4.0.post20260128-py3-none-any.whl
 arctic_inference @ https://paddle-qa.bj.bcebos.com/ernie/arctic_inference-0.1.3-cp310-cp310-linux_x86_64.whl
+deep-gemm-cpp @ https://paddle-qa.bj.bcebos.com/ernie/deep_gemm_cpp-2.2.0%2Blocal-cp310-cp310-linux_x86_64.whl
 transformers>=4.55.1,<5.0.0