From 945d5140a7b274b2f155d6b2eb203590bee48f1d Mon Sep 17 00:00:00 2001
From: lijiaming <1228575330@qq.com>
Date: Fri, 11 Apr 2025 10:33:03 +0800
Subject: [PATCH] support int4 & adjust start_ds command

---
 ...50\347\275\262\346\214\207\345\215\227.md" | 12 ++---
 .../roles/prepare/files/lib/ray_start.sh      |  4 ++
 .../roles/prepare/files/lib/set_env.sh        | 26 ++++++----
 .../roles/prepare/files/lib/start_docker.sh   |  5 +-
 .../roles/prepare/files/lib/start_ds.sh       | 51 +++++++++++--------
 .../workspace/roles/prepare/files/prepare.sh  | 45 +++++-----------
 6 files changed, 71 insertions(+), 72 deletions(-)

diff --git "a/doc/deepseek/DeepSeek-V3&R1\351\203\250\347\275\262\346\214\207\345\215\227.md" "b/doc/deepseek/DeepSeek-V3&R1\351\203\250\347\275\262\346\214\207\345\215\227.md"
index e70be46..1726edc 100644
--- "a/doc/deepseek/DeepSeek-V3&R1\351\203\250\347\275\262\346\214\207\345\215\227.md"
+++ "b/doc/deepseek/DeepSeek-V3&R1\351\203\250\347\275\262\346\214\207\345\215\227.md"
@@ -99,14 +99,14 @@ sh mindspore-deepseek/workspace/roles/prepare/files/lib/ascend_prepare.sh
 **Step1：下载oedeploy工具（下载到控制节点）**
 
 ```shell
-# 下载插件包并解压
-wget https://repo.oepkgs.net/openEuler/rpm/openEuler-24.03-LTS/contrib/oedp/plugins/mindspore-deepseek.tar.gz
-
-tar zxvf mindspore-deepseek.tar.gz
 # 下载安装oedp工具，例如:
 wget https://repo.oepkgs.net/openEuler/rpm/openEuler-24.03-LTS/contrib/oedp/aarch64/Packages/oedp-1.0.0-2.oe2503.aarch64.rpm
 
 yum localinstall oedp-1.0.0-2.oe2503.aarch64.rpm
+# 下载插件包
+git clone https://gitee.com/openeuler/llm_solution.git
+
+cd llm_solution/script/mindspore-deepseek
 ```
 
 **Step2：调整oedeploy配置文件**
@@ -353,11 +353,11 @@ npu-smi set -t reset -i $id -c $chip_id
 
 该步骤在宿主机执行，需在所有节点执行
 
-**Step1:** 可使用even-iso.py绑核脚本，进行细粒度绑核提升性能
+**Step1:** 可用绑核脚本，进行细粒度绑核提升性能
 
 ```shell
 # 所有节点执行
-python ./lib/even-iso.py
+python ./lib/fine-grained-bind-cann.py
 ```
 
 
diff --git a/script/mindspore-deepseek/workspace/roles/prepare/files/lib/ray_start.sh b/script/mindspore-deepseek/workspace/roles/prepare/files/lib/ray_start.sh
index 826eb12..c4d243a 100644
--- a/script/mindspore-deepseek/workspace/roles/prepare/files/lib/ray_start.sh
+++ b/script/mindspore-deepseek/workspace/roles/prepare/files/lib/ray_start.sh
@@ -8,6 +8,10 @@ source $current_path/config.cfg
 source $ENV_FILE
 ray_start() {
     ps -ef | grep "python" | grep -v grep | awk '{print $2}' | xargs kill
+    if [ $NODE_NUM -eq 1 ]; then
+        echo "单机部署无需启动ray"
+        return
+    fi
     ray stop
 
     if [ "$1" ]; then
diff --git a/script/mindspore-deepseek/workspace/roles/prepare/files/lib/set_env.sh b/script/mindspore-deepseek/workspace/roles/prepare/files/lib/set_env.sh
index e43b877..076198a 100644
--- a/script/mindspore-deepseek/workspace/roles/prepare/files/lib/set_env.sh
+++ b/script/mindspore-deepseek/workspace/roles/prepare/files/lib/set_env.sh
@@ -20,6 +20,7 @@ export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 export ASCEND_TOTAL_MEMORY_GB=64
 export HCCL_CONNECT_TIMEOUT=7200
 export MS_COMPILER_CACHE_ENABLE=1
+export CPU_AFFINITY=0
 '
 
 NET_ENV="
@@ -28,18 +29,24 @@ export TP_SOCKET_IFNAME=$RAY_DEVICE
 export HCCL_SOCKET_IFNAME=$RAY_DEVICE
 "
 
-if [ $NODE_NUM -eq 2 ]; then
-	YAML_FILE='/root/Python-3.11/lib/python3.11/site-packages/research/deepseek3/deepseek_r1_671b/predict_deepseek_r1_671b_w8a8.yaml'
+if [ $NODE_NUM -eq 1 ]; then
+    YAML_FILE='/usr/local/Python-3.11/lib/python3.11/site-packages/research/deepseek3/deepseek_r1_671b/predict_deepseek_r1_671b_w8a8.yaml'
+    cat $YAML_FILE | grep gptq-pergroup
+    if [ $? -ne 0 ]; then
+        sed -e 's/model_parallel:.*/model_parallel: 8/' -i $YAML_FILE
+        sed -e "s/quant_method:.*/quant_method: 'gptq-pergroup'/" -i $YAML_FILE
+        sed -e 's/weight_dtype/#weight_dtype/' -i $YAML_FILE
+        sed -e 's/activation_dtype/#activation_dtype/' -i $YAML_FILE
+    fi
+elif [ $NODE_NUM -eq 2 ]; then
+    YAML_FILE='/usr/local/Python-3.11/lib/python3.11/site-packages/research/deepseek3/deepseek_r1_671b/predict_deepseek_r1_671b_w8a8.yaml'
 elif [ $NODE_NUM -eq 4 ]; then
-	YAML_FILE='/root/Python-3.11/lib/python3.11/site-packages/research/deepseek3/deepseek_r1_671b/predict_deepseek_r1_671b.yaml'
+    YAML_FILE='/usr/local/Python-3.11/lib/python3.11/site-packages/research/deepseek3/deepseek_r1_671b/predict_deepseek_r1_671b.yaml'
 fi
 
-
 # 修改权重类型
 sed -e 's/^load_ckpt_format.*/load_ckpt_format: "'$MODEL_TYPE'"/' -i $YAML_FILE
-if [ "$MODEL_TYPE" = "ckpt" ]; then
-    sed -e 's/^auto_trans_ckpt.*/auto_trans_ckpt: False/' -i $YAML_FILE
-fi
+sed -e 's/^auto_trans_ckpt.*/auto_trans_ckpt: False/' -i $YAML_FILE
 
 YAML_ENV="export MINDFORMERS_MODEL_CONFIG=$YAML_FILE"
 
@@ -51,7 +58,8 @@ if grep -q "openeuler_deepseek_env_config" /root/.bashrc; then
 fi
 
 echo "$ENV_ARG" >> $ENV_FILE
-echo "$NET_ENV" >> $ENV_FILE
 echo "$YAML_ENV" >> $ENV_FILE
+if [ $NODE_NUM -ne 1 ]; then
+    echo "$NET_ENV" >> $ENV_FILE
+fi
 source $ENV_FILE
-
diff --git a/script/mindspore-deepseek/workspace/roles/prepare/files/lib/start_docker.sh b/script/mindspore-deepseek/workspace/roles/prepare/files/lib/start_docker.sh
index 5e97f7f..cd1421c 100644
--- a/script/mindspore-deepseek/workspace/roles/prepare/files/lib/start_docker.sh
+++ b/script/mindspore-deepseek/workspace/roles/prepare/files/lib/start_docker.sh
@@ -8,6 +8,7 @@ current_path=$(
 source $current_path/config.cfg
 # 安装docker
 yum install docker -y
+systemctl restart docker
 
 # 检测镜像是否已被拉取
 docker images | grep $IMAGE_NAME | grep $IMAGE_TAG
@@ -25,7 +26,7 @@ if [ $IS_STOP_OTHER_CONTAINER -ne 0 ]; then
 fi
 
 # 如果存在名称相同的容器，则直接使用
-docker ps -a | grep $IMAGE_NAME:$IMAGE_TAG | grep $CONTAINER_NAME
+docker ps -a | grep $IMAGE_NAME:$IMAGE_TAG | grep -w $CONTAINER_NAME
 if [ $? -eq 0 ]; then
     echo "发现容器 $CONTAINER_NAME 已存在，直接使用"
     docker start $CONTAINER_NAME
@@ -33,7 +34,7 @@ if [ $? -eq 0 ]; then
 fi
 
 # 如果存在名称相同，但镜像不同容器，则报错
-docker ps -a | grep $CONTAINER_NAME
+docker ps -a | grep -w $CONTAINER_NAME
 if [ $? -eq 0 ]; then
     echo "发现容器名称 $CONTAINER_NAME 已被使用，请排查"
     exit 1
diff --git a/script/mindspore-deepseek/workspace/roles/prepare/files/lib/start_ds.sh b/script/mindspore-deepseek/workspace/roles/prepare/files/lib/start_ds.sh
index 0a637ae..49a7bd2 100644
--- a/script/mindspore-deepseek/workspace/roles/prepare/files/lib/start_ds.sh
+++ b/script/mindspore-deepseek/workspace/roles/prepare/files/lib/start_ds.sh
@@ -8,33 +8,40 @@ source $current_path/config.cfg
 source $ENV_FILE
 # 仅主节点运行
 
-if [ $NODE_NUM -eq 2 ]; then
-	NPU_NUM=16.0
-    PARALLEL=16
-elif [ $NODE_NUM -eq 4 ]; then
-	NPU_NUM=32.0
-    PARALLEL=32
-fi
-
-ray_status=0
-for i in {1..10}; do
-    ray status | grep "$NPU_NUM NPU"
-    if [ $? -eq 0 ]; then
-        echo "ray集群已全部拉起"
-        ray_status=1
-        break
+if [ $NODE_NUM -ne 1 ]; then
+    if [ $NODE_NUM -eq 2 ]; then
+        NPU_NUM=16.0
+        PARALLEL=16
+    elif [ $NODE_NUM -eq 4 ]; then
+        NPU_NUM=32.0
+        PARALLEL=32
     fi
-    sleep 3
-done
 
-if [ $ray_status -eq 0 ]; then
-    echo "ray集群超时"
-    exit 1
+    ray_status=0
+    for i in {1..10}; do
+        ray status | grep "$NPU_NUM NPU"
+        if [ $? -eq 0 ]; then
+            echo "ray集群已全部拉起"
+            ray_status=1
+            break
+        fi
+        sleep 3
+    done
+
+    if [ $ray_status -eq 0 ]; then
+        echo "ray集群超时"
+        exit 1
+    fi
 fi
 
 #拉起服务
 rm -rf ds.log
-nohup python3 -m vllm_mindspore.entrypoints vllm.entrypoints.openai.api_server --model "$MODEL_PATH" --port=$LLM_PORT --trust_remote_code --tensor_parallel_size=$PARALLEL --max-num-seqs=192 --max_model_len=32768 --max-num-batched-tokens=16384 --block-size=32 --gpu-memory-utilization=0.93 --num-scheduler-steps=8 --distributed-executor-backend=ray &> ds.log &
+if [ $NODE_NUM -ne 1 ]; then
+    nohup python3 -m vllm_mindspore.entrypoints vllm.entrypoints.openai.api_server --model "$MODEL_PATH" --port=$LLM_PORT --trust_remote_code --tensor_parallel_size=$PARALLEL --max-num-seqs=192 --max_model_len=32768 --max-num-batched-tokens=16384 --block-size=128 --gpu-memory-utilization=0.93 --num-scheduler-steps=8 --disable-async-output-proc --distributed-executor-backend=ray &> ds.log &
+else
+    nohup python3 -m vllm_mindspore.entrypoints vllm.entrypoints.openai.api_server --model "$MODEL_PATH" --port=$LLM_PORT --trust_remote_code --tensor_parallel_size=8 --max-num-seqs=192 --max_model_len=32768 --max-num-batched-tokens=16384 --block-size=128 --gpu-memory-utilization=0.93 --num-scheduler-steps=8 --disable-async-output-proc &> ds.log &
+fi
+
 #检测推理服务是否拉起
 llm_status=0
 for i in {1..7200}; do
@@ -50,4 +57,4 @@ done
 if [ $llm_status -eq 0 ]; then
     echo "推理服务拉起超时，请手动确认"
     exit 1
-fi
\ No newline at end of file
+fi
diff --git a/script/mindspore-deepseek/workspace/roles/prepare/files/prepare.sh b/script/mindspore-deepseek/workspace/roles/prepare/files/prepare.sh
index c000112..0ba0d2f 100644
--- a/script/mindspore-deepseek/workspace/roles/prepare/files/prepare.sh
+++ b/script/mindspore-deepseek/workspace/roles/prepare/files/prepare.sh
@@ -16,49 +16,28 @@ main() {
     systemctl stop firewalld
     systemctl stop iptables
 
-    # 检查防火墙是否启动，如果启动则检查端口是否在防火墙白名单中，如果不存在则添加到白名单中
-    status=$(systemctl status firewalld | grep -E "Active" | awk -F":" '{print $2}' | awk -F" " '{print $1}')
-    if [[ "${status}" == "active" ]]; then
-        # ray 端口防火墙检查
-        port_ray=$(firewall-cmd --query-port=$RAY_PORT/tcp)
-        if [[ "${port_ray}" == "no" ]]; then
-            port_ray=$(firewall-cmd --zone=public --add-port=$RAY_PORT/tcp --permanent)
-            firewall-cmd --reload
-        fi
-        port_ray=$(firewall-cmd --query-port=$RAY_PORT/tcp)
-        if [[ "${port_ray}" != "yes" ]]; then
-            echo -e "防火墙开启 $RAY_PORT端口失败"
-            exit 1
-        fi
-        port_llm=$(firewall-cmd --query-port=$LLM_PORT/tcp)
-        if [[ "${port_llm}" == "no" ]]; then
-            port_llm=$(firewall-cmd --zone=public --add-port=$LLM_PORT/tcp --permanent)
-            firewall-cmd --reload
-        fi
-        port_llm=$(firewall-cmd --query-port=$LLM_PORT/tcp)
-        if [[ "${port_llm}" != "yes" ]]; then
-            echo -e "防火墙开启 $LLM_PORT端口失败"
-            exit 1
-        fi
-    fi
-
-    # 检测需要部署的节点ip数量
-    if [ [ $NODE_NUM -ne 2 ] && [ $NODE_NUM -ne 4 ] ]; then
-        echo "当前仅支持两/四节点部署,当前数量是$NODE_NUM"
-        exit 1
-    fi
-
     # 1. 启动Docker容器并复制文件
     $current_path/lib/start_docker.sh
     cp_into_container
 
     # 2. 执行组网检查
-    $current_path/lib/net_check.sh
+    if [ $NODE_NUM -ne 1 ]; then
+        $current_path/lib/net_check.sh
+    fi
 
     #进入容器执行
     # 3. 设置容器内环境变量
     docker exec -it $CONTAINER_NAME /workspace/lib/set_env.sh
 
+    # 4. 进行绑核
+    echo 3 > /proc/sys/vm/drop_caches
+    pip install psutil
+    python3 $current_path/lib/fine-grained-bind-cann.py
+    if [ $? -ne 0 ]; then
+        echo "细粒度线程绑核失败，请确保驱动版本>=24.1.0"
+        exit 1
+    fi
+
 }
 
 # 执行主函数
-- 
Gitee