From 3ededf8dd58868532dc4bb89b6ea3a25c48b9d0e Mon Sep 17 00:00:00 2001
From: lijiaming <1228575330@qq.com>
Date: Wed, 16 Apr 2025 16:54:22 +0800
Subject: [PATCH] adjust start commond for int4 deploy

---
 ...-V3&R1\351\203\250\347\275\262\346\214\207\345\215\227.md" | 4 ++--
 .../workspace/roles/prepare/files/lib/start_ds.sh             | 2 +-
 2 files changed, 3 insertions(+), 3 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 1726edc..eddb984 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"
@@ -135,8 +135,8 @@ all:
   vars:
 # 容器镜像
 # 如果本地docker上已加载镜像，改成docker image的image_name和image_tag
-image_name: hub.oepkgs.net/oedeploy/openeuler/aarch64/deepseek_hyperinfer
-    image_tag: openeuler22.03-py3.11
+    image_name: hub.oepkgs.net/oedeploy/openeuler/aarch64/mindspore
+    image_tag: 20250411  # 单机部署时，建议使用tag为20250326的容器镜像
     # 将要拉起的推理容器的名称
     container_name: openeuler_ds  # 启动之后的docker name，不能和已有镜像重名
     # 模型路径
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 49a7bd2..89209c4 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
@@ -39,7 +39,7 @@ rm -rf 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 &
+    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=32 --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
-- 
Gitee