Configuring Parameters for Deploying DeepSeek Prefill/Decode Instances

Prerequisites

The MindCluster cluster scheduling components (including Ascend Docker Runtime, Volcano, Ascend Device Plugin, Ascend Operator, ClusterD, and NodeD) have been installed, and the status of each component is Running.
The DeepSeek model weight file has been prepared.
You have performed the operations in Configuring the Parameter Plane Network.
Python 3.9 or later has been installed on the master node using Ascend Deployer.

Parameters

Currently, MindCluster Ascend Deployer can start only DeepSeek P/D instances in typical configurations.

Configure parameters. The following is an example. For detailed parameters description, see Table 1 and Table 2.

[master]
xx.xxx.xx.x1 ansible_ssh_user="root" ansible_ssh_pass="xxxxxxx" set_hostname="master-1"     # Replace it with the IP address of the target device.

[worker]
xx.xxx.xx.x1-xx.xxx.xx.x3 ansible_ssh_user="root" ansible_ssh_pass="xxxxxxx"
xx.xxx.xx.x5-xx.xxx.xx.x9 ansible_ssh_user="root" ansible_ssh_pass="xxxxxxx" 

[all:vars]

# Configure parameters for deploying DeepSeek P/D instances.
weight_mount_path="" 
model_weight_path="" 
mindie_image_name="" 
mindie_image_file="" 
expert_map_file=""
model_name="ds_r1"
job_id="mindie-service"
p_instances_num=2
d_instances_num=1
single_p_instance_pod_num=2
single_d_instance_pod_num=4
max_seq_len=18000
mindie_host_log_path=""

**Table 1** [master] and [worker] parameters
Parameter	Required or Not	Description
IP	Yes	Server IP address. You can enter multiple IP addresses. The details are described as follows: index: sequence number of the server within the IP address segment, for example, 1.1.1.1-1.1.1.3 set_hostname="master-{index}", where 1.1.1.1 corresponds to set_hostname=master-1. The value is of the Int type. The default value is 1. step_len: IP address step. The last IP address is retained even if it is not within the step. For example, if 1.1.1.1-1.1.1.6 step_len=2 is set, 1.1.1.1,1.1.1.3,1.1.1.5,1.1.1.6 is output. Parameters can be parsed in the batch input configuration {}, including data calculation and conversion. After the parsing, the output is a character string. For example, if 1.1.1.1-1.1.1.3 set_hostname="master-{str(index+20)+'x'}" is entered, the parsed host information is 1.1.1.1 set_hostname="master-21x" .... The IP address segment parameters support only Python 3. When IPv6 is used, the IP address cannot be passed as a dynamic parameter.
ansible_ssh_user	Yes	Account for logging in to a remote server using SSH. The account must be root.
ansible_ssh_pass	No	Password for logging in to a remote server using SSH. If SSH key-based authentication is configured and the root user is allowed for login, you do not need to set this parameter.
ansible_ssh_port	No	Port for SSH connection. You do not need to set this parameter when the default port number 22 is used. If a non-default port is used, you need to configure this parameter.

**Table 2** [all:vars] parameters
Parameter	Required or Not	Description
weight_mount_path	Yes	Host directory mounted to a container. Example: /mnt/mindie_data/
model_weight_path	Yes	Path for accessing the weight file within a container. Example: /mnt/mindie_data/deepseek_diff_level/deepseek_r1_w8a8
mindie_image_name	No	Image name (name of the loaded MindIE image). Example: mindie:dev-2.0.RC1.B091-800I-A2-py311-ubuntu22.04-aarch64
mindie_image_file	No	Path to the image file in shared storage (MindIE image not loaded yet). Example: */mnt/weight/image/mindie_image.tar.gz**
expert_map_file	No	Path to the expert deployment table. Ensure that the path can be loaded within a container. Example: /mnt/weight/16node_gsm8k.json
model_name	No	Model name, which needs to be specified during interface requests.
job_id	Yes	Namespace in Kubernetes.
p_instances_num	Yes	Number of P instances.
d_instances_num	Yes	Number of D instances.
single_p_instance_pod_num	Yes	Number of pods used by a single P instance.
single_d_instance_pod_num	Yes	Number of pods used by a single D instance.
max_seq_len	Yes	Maximum sequence length. The value must be greater than 0. Only one of the following values can be selected: 18000 68000 134000
mindie_host_log_path	No	MindIE log path. If CCAE is required, enter a non-root directory.

If a MindIE image exists in the cluster, configure mindie_image_name. If a MindIE image file exists on each node, configure mindie_image_file. Set either of the two parameters. If both are provided, mindie_image_name takes priority. If neither is provided, MindCluster Ascend Deployer will search for the image file in the ascend_deployer/resources/MindIE-image directory.
Ensure that the MindIE image corresponding to mindie_image_name and the image file corresponding to mindie_image_file exist in the cluster environment.
When using mindie_image_name, you only need to specify the master node; worker nodes may be omitted. If mindie_image_name is not used, list all worker nodes for deployment.

Parent topic: Configuring Server Installation and Deployment Parameters