diff --git a/.github/workflows/_accuracy_test.yml b/.github/workflows/_accuracy_test.yml
index 2e112f055c..e2c8d40dbf 100644
--- a/.github/workflows/_accuracy_test.yml
+++ b/.github/workflows/_accuracy_test.yml
@@ -180,9 +180,7 @@ jobs:
           -e TZ="Asia/Shanghai" \
           -e "no_proxy=localhost,127.0.0.1,0.0.0.0,bcebos.com,.bcebos.com,bj.bcebos.com,su.bcebos.com,paddle-ci.gz.bcebos.com,apiin.im.baidu.com,baidu-int.com,.baidu.com,aliyun.com,gitee.com,pypi.tuna.tsinghua.edu.cn,.tuna.tsinghua.edu.cn" \
           --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          # Avoid using pip cache to ensure the wheel is updated to the latest version
-          wget -q --no-proxy https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/latest/paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
-          python -m pip install paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/2b9f8b689bc8988f97a5ede056c8c81bfa0332c2/paddlepaddle_gpu-3.3.1.post20260420+2b9f8b689bc-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
 
           pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
 
diff --git a/.github/workflows/_base_test.yml b/.github/workflows/_base_test.yml
index 99813aa427..dd3c55b7c7 100644
--- a/.github/workflows/_base_test.yml
+++ b/.github/workflows/_base_test.yml
@@ -213,9 +213,7 @@ jobs:
           -e TZ="Asia/Shanghai" \
           -e "no_proxy=localhost,127.0.0.1,0.0.0.0,bcebos.com,.bcebos.com,bj.bcebos.com,su.bcebos.com,paddle-ci.gz.bcebos.com,apiin.im.baidu.com,baidu-int.com,.baidu.com,aliyun.com,gitee.com,pypi.tuna.tsinghua.edu.cn,.tuna.tsinghua.edu.cn" \
           --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          # Avoid using pip cache to ensure the wheel is updated to the latest version
-          wget -q --no-proxy https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/latest/paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
-          python -m pip install paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/2b9f8b689bc8988f97a5ede056c8c81bfa0332c2/paddlepaddle_gpu-3.3.1.post20260420+2b9f8b689bc-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
 
           pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
 
diff --git a/.github/workflows/_build_linux.yml b/.github/workflows/_build_linux.yml
index 8ce0b3fef6..8d19261218 100644
--- a/.github/workflows/_build_linux.yml
+++ b/.github/workflows/_build_linux.yml
@@ -189,9 +189,7 @@ jobs:
             elif [[ "${PADDLEVERSION}" != "" ]];then
               python -m pip install paddlepaddle-gpu==${PADDLEVERSION} -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
             else
-              # Avoid using pip cache to ensure the wheel is updated to the latest version
-              wget -q --no-proxy https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/latest/paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
-              python -m pip install paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
+              python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/2b9f8b689bc8988f97a5ede056c8c81bfa0332c2/paddlepaddle_gpu-3.3.1.post20260420+2b9f8b689bc-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
             fi
 
             pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
diff --git a/.github/workflows/_gpu_4cards_case_test.yml b/.github/workflows/_gpu_4cards_case_test.yml
index a393552ddb..c4b771c15f 100644
--- a/.github/workflows/_gpu_4cards_case_test.yml
+++ b/.github/workflows/_gpu_4cards_case_test.yml
@@ -208,9 +208,7 @@ jobs:
           cd FastDeploy
           git diff origin/${BASE_REF}..HEAD --unified=0 > diff.txt
 
-          # Avoid using pip cache to ensure the wheel is updated to the latest version
-          wget -q --no-proxy https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/latest/paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
-          python -m pip install paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/2b9f8b689bc8988f97a5ede056c8c81bfa0332c2/paddlepaddle_gpu-3.3.1.post20260420+2b9f8b689bc-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
           pip config set global.extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
 
           python -m pip install -r scripts/unittest_requirement.txt
diff --git a/.github/workflows/_logprob_test_linux.yml b/.github/workflows/_logprob_test_linux.yml
index 1de1f93ada..7219435606 100644
--- a/.github/workflows/_logprob_test_linux.yml
+++ b/.github/workflows/_logprob_test_linux.yml
@@ -189,9 +189,7 @@ jobs:
           -e TZ="Asia/Shanghai" \
           -e "no_proxy=localhost,127.0.0.1,0.0.0.0,bcebos.com,.bcebos.com,bj.bcebos.com,su.bcebos.com,paddle-ci.gz.bcebos.com,apiin.im.baidu.com,baidu-int.com,.baidu.com,aliyun.com,gitee.com,pypi.tuna.tsinghua.edu.cn,.tuna.tsinghua.edu.cn" \
           --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          # Avoid using pip cache to ensure the wheel is updated to the latest version
-          wget -q --no-proxy https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/latest/paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
-          python -m pip install paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/2b9f8b689bc8988f97a5ede056c8c81bfa0332c2/paddlepaddle_gpu-3.3.1.post20260420+2b9f8b689bc-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
 
           pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
 
diff --git a/.github/workflows/_pre_ce_test.yml b/.github/workflows/_pre_ce_test.yml
index 91ba176bcd..0669d503d8 100644
--- a/.github/workflows/_pre_ce_test.yml
+++ b/.github/workflows/_pre_ce_test.yml
@@ -201,9 +201,7 @@ jobs:
           --gpus "\"device=${DEVICES}\"" ${docker_image} /bin/bash -c '
           git config --global --add safe.directory /workspace/FastDeploy
           cd FastDeploy
-          # Avoid using pip cache to ensure the wheel is updated to the latest version
-          wget -q --no-proxy https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/latest/paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
-          python -m pip install paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/2b9f8b689bc8988f97a5ede056c8c81bfa0332c2/paddlepaddle_gpu-3.3.1.post20260420+2b9f8b689bc-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
           python -m pip install ${fd_wheel_url}
           bash scripts/run_pre_ce.sh
           '
diff --git a/.github/workflows/_stable_test.yml b/.github/workflows/_stable_test.yml
index b8b544468d..8678490f9d 100644
--- a/.github/workflows/_stable_test.yml
+++ b/.github/workflows/_stable_test.yml
@@ -193,9 +193,7 @@ jobs:
           -e TZ="Asia/Shanghai" \
           -e "no_proxy=localhost,127.0.0.1,0.0.0.0,bcebos.com,.bcebos.com,bj.bcebos.com,su.bcebos.com,paddle-ci.gz.bcebos.com,apiin.im.baidu.com,baidu-int.com,.baidu.com,aliyun.com,gitee.com,pypi.tuna.tsinghua.edu.cn,.tuna.tsinghua.edu.cn" \
           --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          # Avoid using pip cache to ensure the wheel is updated to the latest version
-          wget -q --no-proxy https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/latest/paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
-          python -m pip install paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/2b9f8b689bc8988f97a5ede056c8c81bfa0332c2/paddlepaddle_gpu-3.3.1.post20260420+2b9f8b689bc-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
 
           pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
 
diff --git a/.github/workflows/_unit_test_coverage.yml b/.github/workflows/_unit_test_coverage.yml
index 086223b006..e69bc31108 100644
--- a/.github/workflows/_unit_test_coverage.yml
+++ b/.github/workflows/_unit_test_coverage.yml
@@ -223,9 +223,7 @@ jobs:
           git config --global --add safe.directory /workspace/FastDeploy
           cd FastDeploy
           git diff origin/${BASE_REF}..HEAD --unified=0 > diff.txt
-          # Avoid using pip cache to ensure the wheel is updated to the latest version
-          wget -q --no-proxy https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/latest/paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl
-          python -m pip install paddlepaddle_gpu-0.0.0-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
+          python -m pip install https://paddle-qa.bj.bcebos.com/paddle-pipeline/Release-TagBuild-Training-Linux-Gpu-Cuda12.6-Cudnn9.5-Trt10.5-Mkl-Avx-Gcc11-SelfBuiltPypiUse/2b9f8b689bc8988f97a5ede056c8c81bfa0332c2/paddlepaddle_gpu-3.3.1.post20260420+2b9f8b689bc-cp310-cp310-linux_x86_64.whl --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
           pip config set global.extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
 
           python -m pip install -r scripts/unittest_requirement.txt
diff --git a/tests/ce/server/test_compare_top_logprobs.py b/tests/ce/server/test_compare_top_logprobs.py
index 7bd1d3d5d2..c0b3dc1ba3 100644
--- a/tests/ce/server/test_compare_top_logprobs.py
+++ b/tests/ce/server/test_compare_top_logprobs.py
@@ -102,8 +102,11 @@ def test_compare_top_logprobs():
             {"role": "user", "content": "牛顿的三大运动定律是什么？"},
         ],
     }
-
-    compare_top_logprobs(data, top_logprobs_values=[5, 10])
+    # Using [5, 7] instead of larger gaps (e.g., [5, 10]) due to behavior changes
+    # introduced in https://github.com/PaddlePaddle/Paddle/pull/78659.
+    # After this PR, top_logprobs across different k are no longer guaranteed
+    # to be strict prefixes of each other. When k increases, slight variations
+    compare_top_logprobs(data, top_logprobs_values=[5, 7])
 
 
 if __name__ == "__main__":