add ci

Author: iosmers

scripts/run_ci_xpu.sh

@@ -20,13 +20,12 @@ echo "uninstall org"
 python -m pip uninstall paddlepaddle-xpu -y
 python -m pip uninstall fastdeploy-xpu -y
 
-python -m pip install paddlepaddle-xpu -i https://www.paddlepaddle.org.cn/packages/nightly/xpu-p800/
+python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Xpu-P800-SelfBuiltPypiUse/latest/paddlepaddle_xpu-0.0.0-cp310-cp310-linux_x86_64.whl
 
 echo "build whl"
-bash custom_ops/xpu_ops/download_dependencies.sh develop
+bash custom_ops/xpu_ops/download_dependencies.sh stable
 export CLANG_PATH=$(pwd)/custom_ops/xpu_ops/third_party/xtdk
 export XVLLM_PATH=$(pwd)/custom_ops/xpu_ops/third_party/xvllm
-
 bash build.sh || exit 1
 
 echo "pip others"
@@ -54,7 +53,8 @@ python -m fastdeploy.entrypoints.openai.api_server \
     --num-gpu-blocks-override 16384 \
     --max-model-len 32768 \
     --max-num-seqs 128 \
-    --quantization wint4   > server.log 2>&1 &
+    --quantization wint4  \
+    --load-choices default  > server.log 2>&1 &
 
 sleep 60
 # 探活
@@ -121,7 +121,8 @@ python -m fastdeploy.entrypoints.openai.api_server \
     --num-gpu-blocks-override 16384 \
     --max-model-len 32768 \
     --max-num-seqs 64 \
-    --quantization "W4A8"   > server.log 2>&1 &
+    --quantization "W4A8" \
+    --load-choices default  > server.log 2>&1 &
 
 sleep 60
 # 探活
@@ -191,7 +192,8 @@ python -m fastdeploy.entrypoints.openai.api_server \
     --enable-mm \
     --mm-processor-kwargs '{"video_max_frames": 30}' \
     --limit-mm-per-prompt '{"image": 10, "video": 3}' \
-    --reasoning-parser ernie-45-vl > server.log 2>&1 &
+    --reasoning-parser ernie-45-vl \
+    --load-choices default > server.log 2>&1 &
 
 sleep 60
 # 探活
@@ -283,4 +285,4 @@ if [ ${ep_exit_code} -ne 0 ]; then
     cat log/workerlog.0
     echo "EP并行 相关测试失败，请检查pr代码"
     exit 1
-fi
+fi

tests/ci_use/XPU_45T/run_ep.py

@@ -44,6 +44,7 @@ def test_fd_ep():
         quantization="wint4",
         engine_worker_queue_port=engine_worker_queue_port,
         max_num_seqs=8,
+        load_choices="default",
     )
 
     try: