[XPU] ep+tp all2all (#4836)

This commit is contained in:
zhupengyang
2025-11-06 17:26:14 +08:00
committed by GitHub
parent 901d559aa7
commit b54eb7ad81
8 changed files with 201 additions and 17 deletions

View File

@@ -251,7 +251,7 @@ if [ ${vl_test_exit_code} -ne 0 ]; then
fi
echo "============================开始 EP4TP1 测试!============================"
echo "============================开始 EP8TP1 测试!============================"
sleep 5
rm -rf log/*
rm -f core*
@@ -290,12 +290,12 @@ stop_processes
if [ ${ep_exit_code} -ne 0 ]; then
echo "log/workerlog.0"
cat log/workerlog.0
echo "EP4TP1 相关测试失败请检查pr代码"
echo "EP8TP1 相关测试失败请检查pr代码"
exit 1
fi
echo "============================开始 EP4TP4 测试!============================"
echo "============================开始 EP8TP8 allreduce 测试!============================"
sleep 5
rm -rf log/*
rm -f core*
@@ -323,11 +323,55 @@ unset BKCL_PCIE_RING
unset XSHMEM_MODE
unset XSHMEM_QP_NUM_PER_RANK
unset BKCL_RDMA_VERBS
unset enable_expert_parallel
unset enable_tensor_parallel
stop_processes
if [ ${ep_exit_code} -ne 0 ]; then
echo "log/workerlog.0"
cat log/workerlog.0
echo "EP4TP4 相关测试失败请检查pr代码"
echo "EP8TP8 allreduce 相关测试失败请检查pr代码"
exit 1
fi
echo "============================开始 EP8TP8 all2all 测试!============================"
sleep 5
rm -rf log/*
rm -f core*
ipcrm --all=msg
xpu-smi
export XPU_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
export BKCL_ENABLE_XDR=1
export BKCL_RDMA_NICS=xgbe1,xgbe2,xgbe3,xgbe4
export BKCL_TRACE_TOPO=1
export BKCL_PCIE_RING=1
export XSHMEM_MODE=1
export XSHMEM_QP_NUM_PER_RANK=32
export BKCL_RDMA_VERBS=1
export enable_expert_parallel=1
export enable_tensor_parallel=1
export EP_TP_SPLIT_MODE=1
python -m pytest -s --timeout=600 tests/ci_use/XPU_45T/run_ep.py
ep_exit_code=$?
unset BKCL_ENABLE_XDR
unset BKCL_RDMA_NICS
unset BKCL_TRACE_TOPO
unset BKCL_PCIE_RING
unset XSHMEM_MODE
unset XSHMEM_QP_NUM_PER_RANK
unset BKCL_RDMA_VERBS
unset enable_expert_parallel
unset enable_tensor_parallel
unset EP_TP_SPLIT_MODE
stop_processes
if [ ${ep_exit_code} -ne 0 ]; then
echo "log/workerlog.0"
cat log/workerlog.0
echo "EP8TP8 all2all 相关测试失败请检查pr代码"
exit 1
fi