EngineX-Ascend/enginex-ascend-910-vllm
11
0
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity
Files
7f8a1b1f7add090acef165fa96461f8fad0dd3a0
enginex-ascend-910-vllm/docs/source/tutorials/models/DeepSeek-V4-Pro.md
Sun Ruoxi 7f8a1b1f7a init v0.23.0 
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00

64 KiB
Raw Blame History

DeepSeek-V4-Pro

1 Introduction

DeepSeek-V4 introduces several key upgrades over DeepSeek-V3:

  • The Manifold-Constrained Hyper-Connections (mHC) to strengthen conventional residual connections.
  • A hybrid attention architecture, which greatly improves long-context efficiency through Compress-4-Attention and Compress-128-Attention. For the Mixture-of-Experts (MoE) components, it still adopts the DeepSeekMoE architecture, with only minor adjustments.

DeepSeek-V4-Pro, the maximum reasoning effort mode of DeepSeek-V4, significantly advances the knowledge capabilities of open-source models, firmly establishing itself as the best open-source model available today. It achieves top-tier performance in coding benchmarks and significantly bridges the gap with leading closed-source models on reasoning and agentic tasks.

This document will show the main verification steps of the model, including supported features, feature configuration, environment preparation, single-node and multi-node deployment, accuracy and performance evaluation.

2 Supported Features

Refer to supported features to get the model's supported feature matrix.

Refer to feature guide to get the feature's configuration.

3 Prerequisites

3.1 Model Weight

  • DeepSeek-V4-Pro-w4a8-mtp (Quantized version): requires 2 Atlas 800 A3 (128GB × 8) nodes or 4 Atlas 800 A2 (64GB × 8) nodes. Download model weight
  • For Ascend 950DT servers, use the original DeepSeek-V4-Pro weights released by DeepSeek on Hugging Face. The Attention weights use MXFP8, while the MoE weights use MXFP4 with 4-bit weights and 8-bit activation computation (W4A8). No Ascend-specific quantization or weight conversion is required. The mixed deployment in Section 5.1 uses 2 Ascend 950DT servers (96GB × 8), while the 1P1D example in Section 5.2.3 uses 4 Prefill servers and 4 Decode servers.

It is recommended to download the model weight to the shared directory of multiple nodes, such as /root/.cache/.

3.2 Verify Multi-node Communication (Optional)

If you want to deploy a multi-node environment, you need to verify multi-node communication according to verify multi-node communication environment.

4 Installation

4.1 Docker Image Installation

Select an image based on your machine type and start the docker image on your node, refer to using docker.

:::::{tab-set} :sync-group: install

::::{tab-item} Ascend 950DT series :sync: Ascend 950DT

Start the docker image on your each node.

   :substitutions:

export IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|-a5
export NAME=vllm-ascend

docker run --rm \
    --name $NAME \
    --net=host \
    --shm-size=512g \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci2 \
    --device /dev/davinci3 \
    --device /dev/davinci4 \
    --device /dev/davinci5 \
    --device /dev/davinci6 \
    --device /dev/davinci7 \
    --device /dev/davinci_manager \
    --device /dev/hisi_hdc \
    --device /dev/ummu \
    --device /dev/uburma \
    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /etc/hccl_rootinfo.json:/etc/hccl_rootinfo.json \
    -v /etc/hixlep/:/etc/hixlep/ \
    -v /root/.cache:/root/.cache \
    -v /usr/local/sbin:/usr/local/sbin \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi \
    -itd $IMAGE bash

:::: ::::{tab-item} A3 series :sync: A3

Start the docker image on each node.

   :substitutions:

export IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|-a3
docker run --rm \
    --name vllm-ascend \
    --shm-size=512g \
    --net=host \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci2 \
    --device /dev/davinci3 \
    --device /dev/davinci4 \
    --device /dev/davinci5 \
    --device /dev/davinci6 \
    --device /dev/davinci7 \
    --device /dev/davinci8 \
    --device /dev/davinci9 \
    --device /dev/davinci10 \
    --device /dev/davinci11 \
    --device /dev/davinci12 \
    --device /dev/davinci13 \
    --device /dev/davinci14 \
    --device /dev/davinci15 \
    --device /dev/davinci_manager \
    --device /dev/devmm_svm \
    --device /dev/hisi_hdc \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /etc/hccn.conf:/etc/hccn.conf \
    -it $IMAGE bash

:::: ::::{tab-item} A2 series :sync: A2

Start the docker image on each node.

   :substitutions:

export IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|
docker run --rm \
    --name vllm-ascend \
    --shm-size=512g \
    --net=host \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci2 \
    --device /dev/davinci3 \
    --device /dev/davinci4 \
    --device /dev/davinci5 \
    --device /dev/davinci6 \
    --device /dev/davinci7 \
    --device /dev/davinci_manager \
    --device /dev/devmm_svm \
    --device /dev/hisi_hdc \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /etc/hccn.conf:/etc/hccn.conf \
    -it $IMAGE bash

:::: :::::

After a successful docker run, you can verify the running container service by executing the docker ps command.

4.2 Source Code Installation

If you don't want to use the docker image as above, you can also build all from source:

If you want to deploy a multi-node environment, you need to set up the environment on each node.

5 Online Service Deployment

:::{note} In this tutorial, we suppose you downloaded the model weight to /root/.cache/. Feel free to change it to your own path. :::

5.1 Multi-Node Online Deployment

The quantized model DeepSeek-V4-Pro-w4a8-mtp requires at least 2 Atlas 800 A3 (128GB × 8) nodes or 4 Atlas 800 A2 (64GB × 8) nodes. The mixed-deployment example for Ascend 950DT servers uses 2 servers (96GB × 8). Run the following scripts on each server respectively.

:::::{tab-set} :sync-group: install

::::{tab-item} A2 series :sync: A2

Node0

local_ip="xxx"
node0_ip="xxxx"

export HCCL_IF_IP=$local_ip
export IFNAME="xxx"
export GLOO_SOCKET_IFNAME="$IFNAME"
export TP_SOCKET_IFNAME="$IFNAME"
export HCCL_SOCKET_IFNAME="$IFNAME"
export HCCL_BUFFSIZE=512
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ACL_OP_INIT_MODE=1
export VLLM_ENGINE_READY_TIMEOUT_S=3600
export HCCL_OP_EXPANSION_MODE="AIV"

export TASK_QUEUE_ENABLE=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD

export HCCL_CONNECT_TIMEOUT=7200
export ASCEND_CONNECT_TIMEOUT=10000
export ASCEND_TRANSFER_TIMEOUT=10000
export VLLM_RPC_TIMEOUT=1800000

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
  --host 0.0.0.0 \
  --port 10010 \
  --max-model-len 135000 \
  --max-num-batched-tokens 4096 \
  --served-model-name dsv4 \
  --gpu-memory-utilization 0.9 \
  --max-num-seqs 16 \
  --data-parallel-size 4 \
  --tensor-parallel-size 8 \
  --data-parallel-size-local 1 \
  --data-parallel-start-rank 0 \
  --data-parallel-address $node0_ip  \
  --enable-expert-parallel \
  --quantization ascend \
  --no-enable-prefix-caching \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --async-scheduling \
  --safetensors-load-strategy 'prefetch' \
  --block-size 128 \
  --speculative-config '{
     "num_speculative_tokens": 1,
     "method": "mtp",
     "enforce_eager": true
  }' \
  --additional-config '{
     "ascend_compilation_config":{
        "enable_npugraph_ex":true,
        "enable_static_kernel":false
     },
     "enable_cpu_binding": true,
     "enable_shared_expert_dp": true,
     "multistream_overlap_shared_expert":true
  }' \
  --compilation-config '{
     "cudagraph_mode":"FULL_DECODE_ONLY"
  }' \
  --model-loader-extra-config '{
     "enable_multithread_load": "true",
     "num_threads": 128
  }'

Node1-Node3

local_ip="xxx"
node0_ip="xxxx"

export HCCL_IF_IP=$local_ip
export IFNAME="xxx"
export GLOO_SOCKET_IFNAME="$IFNAME"
export TP_SOCKET_IFNAME="$IFNAME"
export HCCL_SOCKET_IFNAME="$IFNAME"
export HCCL_BUFFSIZE=512
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ACL_OP_INIT_MODE=1
export VLLM_ENGINE_READY_TIMEOUT_S=3600
export HCCL_OP_EXPANSION_MODE="AIV"

export TASK_QUEUE_ENABLE=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD

export HCCL_CONNECT_TIMEOUT=7200
export ASCEND_CONNECT_TIMEOUT=10000
export ASCEND_TRANSFER_TIMEOUT=10000
export VLLM_RPC_TIMEOUT=1800000

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
  --host 0.0.0.0 \
  --port 10010 \
  --max-model-len 135000 \
  --max-num-batched-tokens 4096 \
  --served-model-name dsv4 \
  --gpu-memory-utilization 0.9 \
  --max-num-seqs 16 \
  --data-parallel-size 4 \
  --tensor-parallel-size 8 \
  --data-parallel-size-local 1 \
  --data-parallel-start-rank 1 \
  --data-parallel-address $node0_ip  \
  --enable-expert-parallel \
  --quantization ascend \
  --no-enable-prefix-caching \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --async-scheduling \
  --safetensors-load-strategy 'prefetch' \
  --block-size 128 \
  --headless \
  --speculative-config '{
     "num_speculative_tokens": 1,
     "method": "mtp",
     "enforce_eager": true
  }' \
  --additional-config '{
     "ascend_compilation_config":{
        "enable_npugraph_ex":true,
        "enable_static_kernel":false
     },
     "enable_cpu_binding": true,
     "enable_shared_expert_dp": true,
     "multistream_overlap_shared_expert":true
  }' \
  --compilation-config '{
     "cudagraph_mode":"FULL_DECODE_ONLY"
  }' \
  --model-loader-extra-config '{
     "enable_multithread_load": "true",
     "num_threads": 128
  }'

:::: ::::{tab-item} A3 series :sync: A3

Node0

# this obtained through ifconfig
# nic_name is the network interface name corresponding to local_ip of the current node
nic_name="xxx"
local_ip="xxx"

# The value of node0_ip must be consistent with the value of local_ip set in node0 (master node)
node0_ip="xxxx"

export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export HCCL_BUFFSIZE=2048
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export TASK_QUEUE_ENABLE=1
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
  --safetensors-load-strategy 'prefetch' \
  --max-model-len 135000  \
  --max-num-batched-tokens 4096 \
  --served-model-name dsv4 \
  --gpu-memory-utilization 0.9 \
  --max-num-seqs 32 \
  --data-parallel-size 2 \
  --data-parallel-size-local 1 \
  --data-parallel-start-rank 0 \
  --data-parallel-address $node0_ip \
  --data-parallel-rpc-port 13399 \
  --tensor-parallel-size 16 \
  --enable-expert-parallel \
  --quantization ascend \
  --port 8900 \
  --host 0.0.0.0 \
  --block-size 128 \
  --async-scheduling \
  --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
  --additional-config '
    {"ascend_compilation_config":{
        "enable_npugraph_ex":true,
        "enable_static_kernel":false
        },
    "enable_cpu_binding": true,
    "multistream_overlap_shared_expert":true}'

Node1

# this obtained through ifconfig
# nic_name is the network interface name corresponding to local_ip of the current node
nic_name="xxx"
local_ip="xxx"

# The value of node0_ip must be consistent with the value of local_ip set in node0 (master node)
node0_ip="xxxx"

export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export HCCL_BUFFSIZE=2048
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export TASK_QUEUE_ENABLE=1
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
  --safetensors-load-strategy 'prefetch' \
  --max-model-len 135000  \
  --max-num-batched-tokens 4096 \
  --served-model-name dsv4 \
  --gpu-memory-utilization 0.9 \
  --max-num-seqs 32 \
  --data-parallel-size 2 \
  --data-parallel-size-local 1 \
  --data-parallel-start-rank 1 \
  --data-parallel-address $node0_ip \
  --data-parallel-rpc-port 13399 \
  --tensor-parallel-size 16 \
  --enable-expert-parallel \
  --quantization ascend \
  --port 8900 \
  --host 0.0.0.0 \
  --block-size 128 \
  --async-scheduling \
  --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
  --additional-config '
    {"ascend_compilation_config":{
        "enable_npugraph_ex":true,
        "enable_static_kernel":false
        },
    "enable_cpu_binding": true,
    "multistream_overlap_shared_expert":true}'

:::: ::::{tab-item} Ascend 950DT series :sync: Ascend 950DT

Set node0_ip to the service IP of Node0 on both nodes.

Node0

source /root/.bashrc

nic_name="xxx"
local_ip="xx.xx.xx.1"
node0_ip="xx.xx.xx.1"

export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export HCCL_ALGO=level0:fullmesh
export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=2040
export HCCL_CONNECT_TIMEOUT=1200
export HCCL_BUFFSIZE=512
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

vllm serve /root/.cache/DeepSeek-V4-Pro \
    --host $local_ip \
    --port 8900 \
    --data-parallel-size 2 \
    --data-parallel-size-local 1 \
    --data-parallel-start-rank 0 \
    --data-parallel-address $node0_ip \
    --data-parallel-rpc-port 6987 \
    --tensor-parallel-size 8 \
    --max-model-len 200000 \
    --max-num-batched-tokens 8192 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.94 \
    --enable-expert-parallel \
    --async-scheduling \
    --max-num-seqs 40 \
    --block-size 128 \
    --api-server-count 1 \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --safetensors-load-strategy prefetch \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --speculative-config '{"num_speculative_tokens": 3, "method": "mtp", "enforce_eager": true}' \
    --additional-config '{"enable_cpu_binding": true, "multistream_overlap_shared_expert": true, "multistream_dsa_preprocess": false, "dp_allreduce_on_npu": false, "enable_flashcomm1": true, "enable_dsa_cp": true}'

Node1

source /root/.bashrc

nic_name="xxx"
local_ip="xx.xx.xx.2"
node0_ip="xx.xx.xx.1"

export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export HCCL_ALGO=level0:fullmesh
export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=2040
export HCCL_CONNECT_TIMEOUT=1200
export HCCL_BUFFSIZE=512
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

vllm serve /root/.cache/DeepSeek-V4-Pro \
    --data-parallel-size 2 \
    --data-parallel-size-local 1 \
    --data-parallel-start-rank 1 \
    --data-parallel-address $node0_ip \
    --data-parallel-rpc-port 6987 \
    --tensor-parallel-size 8 \
    --max-model-len 200000 \
    --max-num-batched-tokens 8192 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.94 \
    --enable-expert-parallel \
    --async-scheduling \
    --max-num-seqs 40 \
    --block-size 128 \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --safetensors-load-strategy prefetch \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --speculative-config '{"num_speculative_tokens": 3, "method": "mtp", "enforce_eager": true}' \
    --additional-config '{"enable_cpu_binding": true, "multistream_overlap_shared_expert": true, "multistream_dsa_preprocess": false, "dp_allreduce_on_npu": false, "enable_flashcomm1": true, "enable_dsa_cp": true}' \
    --headless

:::: :::::

Key Parameter Descriptions:

  • --data-parallel-size sets the global number of data parallel ranks, and --data-parallel-size-local sets the number of DP ranks on the current node.
  • --data-parallel-start-rank specifies the starting data parallel rank of the current node. Each node must be set to a unique value (e.g., Node0 = 0, Node1 = 1).
  • --data-parallel-address specifies the IP address of the data parallel master node (Node0). It must be consistent across all nodes.
  • --data-parallel-rpc-port is the DP RPC port. Use the same value on all nodes and ensure the port is available.
  • --tensor-parallel-size sets the tensor parallel size within each DP rank. Configure it together with the DP sizes according to the deployment topology and available NPUs.
  • On Ascend 950DT, DeepSeek-V4 does not currently support standalone tensor parallelism (TP-only). TP partitions only the wq_b, wo_a, and wo_b linear layers, so data parallelism is recommended. When TP is required, use it together with DSA-CP and enable FlashComm (enable_flashcomm1) and DSA-CP (enable_dsa_cp), as in the two-server DeepSeek-V4-Pro mixed deployment.
  • --enable-expert-parallel enables expert parallelism for MoE layers. Do not mix MoE tensor parallelism and expert parallelism in the same MoE layer.
  • --headless (used on non-master nodes) disables the API server on the node, since only the master node serves requests.
  • --max-model-len specifies the maximum context length. Adjust it according to your actual scenario.
  • --max-num-seqs indicates the maximum number of requests that each DP group is allowed to process. If the number of requests sent to the service exceeds this limit, the excess requests will remain in a waiting state and will not be scheduled. Note that the time spent in the waiting state is also counted in metrics such as TTFT and TPOT. Therefore, when testing performance, it is generally recommended that --max-num-seqs * --data-parallel-size >= the actual total concurrency.
  • --max-num-batched-tokens is the maximum number of tokens processed in one scheduler step. A larger value can improve prefill efficiency but consumes more activation memory.
  • --no-enable-prefix-caching indicates that prefix caching is disabled. To enable it, remove this option.
  • --block-size sets the KV cache block size. To enable the experimental 4K prefix cache hit support, change it from 128 to 32.
  • --quantization ascend enables Ascend quantization for the W4A8 model.
  • --speculative-config configures the MTP (Multi-Token Prediction) speculative decoding to accelerate inference.
  • --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' enables full ACL graph execution in the decode phase to reduce scheduling latency.
  • --async-scheduling enables asynchronous scheduling to overlap CPU scheduling with NPU computation.
  • --additional-config enables Ascend-specific optimizations. enable_npugraph_ex enables enhanced ACL graph execution, enable_static_kernel: false keeps static-kernel compilation disabled, enable_cpu_binding enables Ascend-native CPU binding, enable_shared_expert_dp enables data parallelism for shared experts, and multistream_overlap_shared_expert overlaps shared expert computation for better MoE throughput.
  • VLLM_ASCEND_ENABLE_FLASHCOMM1=1 enables the FlashComm communication optimization.

Common Issues Tip: If you encounter issues, please refer to the Public FAQ for troubleshooting.

Service Verification:

curl http://<node0_ip>:8900/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "dsv4",
        "messages": [
            {
                "role": "user",
                "content": "Who are you?"
            }
        ],
        "max_tokens": 256,
        "temperature": 0
    }'

Expected Result:

The service returns HTTP 200 OK with a JSON response containing the choices field.

5.2 Multi-Node PD Separation Deployment

We recommend using Mooncake for deployment: Mooncake.

In the standard deployment mode, Prefill (prompt processing) and Decode (token generation) tasks run on the same set of NPUs. PD (Prefill-Decode) separation addresses this by running Prefill and Decode on dedicated node groups, each configured independently. This architecture is recommended for production deployments with concurrent multi-user workloads, where stable latency and high throughput are both required.

The following sections describe PD separation deployment in multi-node environments using Atlas 800 A3 nodes (128GB × 8), Atlas 800 A2 nodes (64GB × 8), or Ascend 950DT servers (96GB × 8).

5.2.1 A3 Series PD Separation Deployment

This section shows the deployment guide of DeepSeek-V4-Pro on Atlas 800 A3 (128GB × 8) multi-node environment with 1P1D for better performance.

Before you start, please:

  1. Prepare the script launch_online_dp.py on each node.

    import argparse
import multiprocessing
import os
import subprocess
import sys

def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--dp-size",
        type=int,
        required=True,
        help="Data parallel size."
    )
    parser.add_argument(
        "--tp-size",
        type=int,
        default=1,
        help="Tensor parallel size."
    )
    parser.add_argument(
        "--dp-size-local",
        type=int,
        default=-1,
        help="Local data parallel size."
    )
    parser.add_argument(
        "--dp-rank-start",
        type=int,
        default=0,
        help="Starting rank for data parallel."
    )
    parser.add_argument(
        "--dp-address",
        type=str,
        required=True,
        help="IP address for data parallel master node."
    )
    parser.add_argument(
        "--dp-rpc-port",
        type=str,
        default=12345,
        help="Port for data parallel master node."
    )
    parser.add_argument(
        "--vllm-start-port",
        type=int,
        default=9000,
        help="Starting port for the engine."
    )
    return parser.parse_args()

args = parse_args()
dp_size = args.dp_size
tp_size = args.tp_size
dp_size_local = args.dp_size_local
if dp_size_local == -1:
    dp_size_local = dp_size
dp_rank_start = args.dp_rank_start
dp_address = args.dp_address
dp_rpc_port = args.dp_rpc_port
vllm_start_port = args.vllm_start_port

def run_command(visible_devices, dp_rank, vllm_engine_port):
    command = [
        "bash",
        "./run_dp_template.sh",
        visible_devices,
        str(vllm_engine_port),
        str(dp_size),
        str(dp_rank),
        dp_address,
        dp_rpc_port,
        str(tp_size),
    ]
    subprocess.run(command, check=True)

if __name__ == "__main__":
    template_path = "./run_dp_template.sh"
    if not os.path.exists(template_path):
        print(f"Template file {template_path} does not exist.")
        sys.exit(1)

    processes = []
    num_cards = dp_size_local * tp_size
    for i in range(dp_size_local):
        dp_rank = dp_rank_start + i
        vllm_engine_port = vllm_start_port + i
        visible_devices = ",".join(str(x) for x in range(i * tp_size, (i + 1) * tp_size))
        process = multiprocessing.Process(target=run_command,
                                        args=(visible_devices, dp_rank,
                                                vllm_engine_port))
        processes.append(process)
        process.start()

    for process in processes:
        process.join()

    Parameter descriptions:

    Parameter Type Required Default Description
    --dp-size int Yes - Data parallel size (total number of DP ranks across all nodes).
    --tp-size int No 1 Tensor parallel size within each DP rank.
    --dp-size-local int No (same as --dp-size) Number of DP ranks on the current node. If not set, defaults to --dp-size.
    --dp-rank-start int No 0 Starting rank offset for data parallel ranks on this node.
    --dp-address str Yes - IP address of the data parallel master node.
    --dp-rpc-port str No 12345 RPC port for data parallel master communication.
    --vllm-start-port int No 9000 Starting port for each vLLM engine instance on this node.

  2. Prepare the script run_dp_template.sh on each node.

    1. Prefill node 0

      nic_name="xxxx" # change to your own nic name
local_ip=xx.xx.xx.1 # change to your own ip

export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=204
export HCCL_CONNECT_TIMEOUT=120
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_BUFFSIZE=1024
export TASK_QUEUE_ENABLE=1
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export ASCEND_RT_VISIBLE_DEVICES=$1
export VLLM_ASCEND_ENABLE_FUSED_MC2=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
    --host 0.0.0.0 \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --enable-expert-parallel \
    --seed 1024 \
    --served-model-name auto \
    --max-model-len 131072 \
    --max-num-batched-tokens 4096 \
    --max-num-seqs 16 \
    --no-disable-hybrid-kv-cache-manager \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --safetensors-load-strategy 'prefetch' \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --trust-remote-code \
    --gpu-memory-utilization 0.92 \
    --quantization ascend \
    --block-size 128 \
    --enforce-eager \
    --speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
    --additional-config '{"enable_cpu_binding": true, "enable_dsa_cp": true}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30200",
    "engine_id": "1",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 2,
                        "tp_size": 16
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 2
                }
        }
    }'

    2. Prefill node 1

      nic_name="xxxx" # change to your own nic name
local_ip=xx.xx.xx.2 # change to your own ip

export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=204
export HCCL_CONNECT_TIMEOUT=120
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_BUFFSIZE=1024
export TASK_QUEUE_ENABLE=1
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export ASCEND_RT_VISIBLE_DEVICES=$1
export VLLM_ASCEND_ENABLE_FUSED_MC2=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
    --host 0.0.0.0 \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --enable-expert-parallel \
    --seed 1024 \
    --served-model-name auto \
    --max-model-len 131072 \
    --max-num-batched-tokens 4096 \
    --max-num-seqs 16 \
    --no-disable-hybrid-kv-cache-manager \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --safetensors-load-strategy 'prefetch' \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --trust-remote-code \
    --gpu-memory-utilization 0.92 \
    --quantization ascend \
    --block-size 128 \
    --enforce-eager \
    --speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
    --additional-config '{"enable_cpu_binding": true, "enable_dsa_cp": true}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30200",
    "engine_id": "1",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 2,
                        "tp_size": 16
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 2
                }
        }
    }'

    3. Decode node (Same as another D node)

      nic_name="xxxx" # change to your own nic name
local_ip=xx.xx.xx.3/4 # change to your own ip

export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export HCCL_OP_EXPANSION_MODE="AIV"
export TASK_QUEUE_ENABLE=1
export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=2000
export HCCL_CONNECT_TIMEOUT=1200
export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_BUFFSIZE=1024
export ASCEND_RT_VISIBLE_DEVICES=$1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
    --host 0.0.0.0 \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --enable-expert-parallel \
    --seed 1024 \
    --served-model-name auto \
    --max-model-len 131072 \
    --max-num-batched-tokens 120 \
    --max-num-seqs 60 \
    --async-scheduling \
    --block-size 128 \
    --no-enable-prefix-caching \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --no-disable-hybrid-kv-cache-manager \
    --safetensors-load-strategy 'prefetch' \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --trust-remote-code \
    --gpu-memory-utilization 0.9 \
    --quantization ascend \
    --speculative-config '{"num_speculative_tokens": 1, "method":"mtp", "enforce_eager": true}' \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_consumer",
    "kv_port": "30800",
    "engine_id": "8",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 2,
                        "tp_size": 16
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 2
                }
        }
    }' \
    --additional-config '{
        "ascend_compilation_config":{
            "enable_npugraph_ex":true,
            "enable_static_kernel":false
        },
    "enable_cpu_binding":true,
    "recompute_scheduler_enable":true
    }'

  3. Start the server with the following command on each node.

    1. Prefill node 0

      # change ip to your own
python launch_online_dp.py --dp-size 2 --tp-size 16 --dp-size-local 1 --dp-rank-start 0 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100

    2. Prefill node 1

      # change ip to your own
python launch_online_dp.py --dp-size 2 --tp-size 16 --dp-size-local 1 --dp-rank-start 1 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100

    3. Decode node 0

      # change ip to your own
python launch_online_dp.py --dp-size 16 --tp-size 2 --dp-size-local 8 --dp-rank-start 0 --dp-address xx.xx.xx.3 --dp-rpc-port 12321 --vllm-start-port 7100

    4. Decode node 1

      # change ip to your own
python launch_online_dp.py --dp-size 16 --tp-size 2 --dp-size-local 8 --dp-rank-start 8 --dp-address xx.xx.xx.3 --dp-rpc-port 12321 --vllm-start-port 7100

  4. Deploy the P-D disaggregation proxy.

    Refer to Prefill-Decode Disaggregation (Deepseek) to deploy the P-D disaggregation proxy.

5.2.2 A2 Series PD Separation Deployment

This section shows the deployment guide of DeepSeek-V4-Pro on Atlas 800 A2 (64GB × 8) multi-node environment with 1P1D for better performance.

Before you start, please:

  1. Prepare the script launch_online_dp.py on each node.

    import argparse
import multiprocessing
import os
import subprocess
import sys

def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--dp-size",
        type=int,
        required=True,
        help="Data parallel size."
    )
    parser.add_argument(
        "--tp-size",
        type=int,
        default=1,
        help="Tensor parallel size."
    )
    parser.add_argument(
        "--dp-size-local",
        type=int,
        default=-1,
        help="Local data parallel size."
    )
    parser.add_argument(
        "--dp-rank-start",
        type=int,
        default=0,
        help="Starting rank for data parallel."
    )
    parser.add_argument(
        "--dp-address",
        type=str,
        required=True,
        help="IP address for data parallel master node."
    )
    parser.add_argument(
        "--dp-rpc-port",
        type=str,
        default=12345,
        help="Port for data parallel master node."
    )
    parser.add_argument(
        "--vllm-start-port",
        type=int,
        default=9000,
        help="Starting port for the engine."
    )
    return parser.parse_args()

args = parse_args()
dp_size = args.dp_size
tp_size = args.tp_size
dp_size_local = args.dp_size_local
if dp_size_local == -1:
    dp_size_local = dp_size
dp_rank_start = args.dp_rank_start
dp_address = args.dp_address
dp_rpc_port = args.dp_rpc_port
vllm_start_port = args.vllm_start_port

def run_command(visible_devices, dp_rank, vllm_engine_port):
    command = [
        "bash",
        "./run_dp_template.sh",
        visible_devices,
        str(vllm_engine_port),
        str(dp_size),
        str(dp_rank),
        dp_address,
        dp_rpc_port,
        str(tp_size),
    ]
    subprocess.run(command, check=True)

if __name__ == "__main__":
    template_path = "./run_dp_template.sh"
    if not os.path.exists(template_path):
        print(f"Template file {template_path} does not exist.")
        sys.exit(1)

    processes = []
    num_cards = dp_size_local * tp_size
    for i in range(dp_size_local):
        dp_rank = dp_rank_start + i
        vllm_engine_port = vllm_start_port + i
        visible_devices = ",".join(str(x) for x in range(i * tp_size, (i + 1) * tp_size))
        process = multiprocessing.Process(target=run_command,
                                        args=(visible_devices, dp_rank,
                                                vllm_engine_port))
        processes.append(process)
        process.start()

    for process in processes:
        process.join()

  2. Prepare the script run_dp_template.sh on each node.

    1. Prefill node (4 P nodes share the same script)

      nic_name="xxxx" # change to your own nic name
local_ip=xx.xx.xx.1/2/3/4 # change to your own ip

export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export HCCL_OP_EXPANSION_MODE="AIV"
export TASK_QUEUE_ENABLE=1

export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=204
export HCCL_CONNECT_TIMEOUT=1200

export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_BUFFSIZE=1024

sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export ASCEND_RT_VISIBLE_DEVICES=$1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
    --host 0.0.0.0 \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --enable-expert-parallel \
    --seed 1024 \
    --served-model-name dsv4 \
    --max-model-len 133072 \
    --max-num-batched-tokens 4096 \
    --max-num-seqs 16 \
    --no-disable-hybrid-kv-cache-manager \
    --trust-remote-code \
    --gpu-memory-utilization 0.9 \
    --quantization ascend \
    --safetensors-load-strategy 'prefetch' \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --enforce-eager \
    --no-enable-prefix-caching \
    --speculative-config '{"num_speculative_tokens": 1, "method":"mtp", "enforce_eager": true}' \
    --additional-config '{"enable_cpu_binding": true, "enable_shared_expert_dp": true, "enable_dsa_cp": true}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30000",
    "engine_id": "0",
    "kv_connector_extra_config": {
            "prefill": {
                "dp_size": 4,
                "tp_size": 8
            },
            "decode": {
                "dp_size": 8,
                "tp_size": 4
            }
      }
  }'

    2. Decode node (4 D nodes share the same script)

      nic_name="xxxx" # change to your own nic name
local_ip=xx.xx.xx.5/6/7/8 # change to your own ip

export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export HCCL_OP_EXPANSION_MODE="AIV"
export TASK_QUEUE_ENABLE=1

export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=204
export HCCL_CONNECT_TIMEOUT=1200

export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_BUFFSIZE=1024

sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

export ASCEND_RT_VISIBLE_DEVICES=$1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
    --host 0.0.0.0 \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --enable-expert-parallel \
    --seed 1024 \
    --served-model-name dsv4 \
    --max-model-len 133072 \
    --max-num-batched-tokens 120 \
    --max-num-seqs 60 \
    --async-scheduling \
    --block-size 128 \
    --no-disable-hybrid-kv-cache-manager \
    --trust-remote-code \
    --gpu-memory-utilization 0.9 \
    --quantization ascend \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --safetensors-load-strategy 'prefetch' \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --no-enable-prefix-caching \
    --speculative-config '{"num_speculative_tokens": 1, "method":"mtp", "enforce_eager": true}' \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_consumer",
    "kv_port": "30100",
    "engine_id": "1",
    "kv_connector_extra_config": {
            "prefill": {
                "dp_size": 4,
                "tp_size": 8
            },
            "decode": {
                "dp_size": 8,
                "tp_size": 4
            }
      }
  }' \
    --additional-config '{"ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false}, "enable_cpu_binding":true, "recompute_scheduler_enable":true}'

  3. Start the server with the following command on each node.

    1. Prefill node 0

      # change ip to your own
python launch_online_dp.py --dp-size 4 --tp-size 8 --dp-size-local 1 --dp-rank-start 0 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100

    2. Prefill node 1

      # change ip to your own
python launch_online_dp.py --dp-size 4 --tp-size 8 --dp-size-local 1 --dp-rank-start 1 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100

    3. Prefill node 2

      # change ip to your own
python launch_online_dp.py --dp-size 4 --tp-size 8 --dp-size-local 1 --dp-rank-start 2 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100

    4. Prefill node 3

      # change ip to your own
python launch_online_dp.py --dp-size 4 --tp-size 8 --dp-size-local 1 --dp-rank-start 3 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100

    5. Decode node 0

      # change ip to your own
python launch_online_dp.py --dp-size 8 --tp-size 4 --dp-size-local 2 --dp-rank-start 0 --dp-address xx.xx.xx.2 --dp-rpc-port 12321 --vllm-start-port 7100

    6. Decode node 1

      # change ip to your own
python launch_online_dp.py --dp-size 8 --tp-size 4 --dp-size-local 2 --dp-rank-start 2 --dp-address xx.xx.xx.2 --dp-rpc-port 12321 --vllm-start-port 7100

    7. Decode node 2

      # change ip to your own
python launch_online_dp.py --dp-size 8 --tp-size 4 --dp-size-local 2 --dp-rank-start 4 --dp-address xx.xx.xx.2 --dp-rpc-port 12321 --vllm-start-port 7100

    8. Decode node 3

      # change ip to your own
python launch_online_dp.py --dp-size 8 --tp-size 4 --dp-size-local 2 --dp-rank-start 6 --dp-address xx.xx.xx.2 --dp-rpc-port 12321 --vllm-start-port 7100

  4. Deploy the P-D disaggregation proxy.

    Refer to Prefill-Decode Disaggregation (Deepseek) to deploy the P-D disaggregation proxy.

5.2.3 Ascend 950DT Series PD Separation Deployment

This section shows a logical 1P1D deployment on Ascend 950DT servers. The Prefill group uses 4 servers and the Decode group uses 4 servers. Each group contains 32 DP ranks with DP32/TP1; every server starts 8 local DP ranks. The prefill and decode settings in --kv-transfer-config must therefore both use dp_size: 32 and tp_size: 1.

Before you start, mount /etc/hixlep/ into the container and complete the following steps:

  1. Prepare the script launch_online_dp.py on each node.

    import argparse
import multiprocessing
import os
import subprocess
import sys

def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--dp-size",
        type=int,
        required=True,
        help="Data parallel size."
    )
    parser.add_argument(
        "--tp-size",
        type=int,
        default=1,
        help="Tensor parallel size."
    )
    parser.add_argument(
        "--dp-size-local",
        type=int,
        default=-1,
        help="Local data parallel size."
    )
    parser.add_argument(
        "--dp-rank-start",
        type=int,
        default=0,
        help="Starting rank for data parallel."
    )
    parser.add_argument(
        "--dp-address",
        type=str,
        required=True,
        help="IP address for data parallel master node."
    )
    parser.add_argument(
        "--dp-rpc-port",
        type=str,
        default=12345,
        help="Port for data parallel master node."
    )
    parser.add_argument(
        "--vllm-start-port",
        type=int,
        default=9000,
        help="Starting port for the engine."
    )
    return parser.parse_args()

args = parse_args()
dp_size = args.dp_size
tp_size = args.tp_size
dp_size_local = args.dp_size_local
if dp_size_local == -1:
    dp_size_local = dp_size
dp_rank_start = args.dp_rank_start
dp_address = args.dp_address
dp_rpc_port = args.dp_rpc_port
vllm_start_port = args.vllm_start_port

def run_command(visible_devices, dp_rank, vllm_engine_port):
    command = [
        "bash",
        "./run_dp_template.sh",
        visible_devices,
        str(vllm_engine_port),
        str(dp_size),
        str(dp_rank),
        dp_address,
        dp_rpc_port,
        str(tp_size),
    ]
    subprocess.run(command, check=True)

if __name__ == "__main__":
    template_path = "./run_dp_template.sh"
    if not os.path.exists(template_path):
        print(f"Template file {template_path} does not exist.")
        sys.exit(1)

    processes = []
    num_cards = dp_size_local * tp_size
    for i in range(dp_size_local):
        dp_rank = dp_rank_start + i
        vllm_engine_port = vllm_start_port + i
        visible_devices = ",".join(str(x) for x in range(i * tp_size, (i + 1) * tp_size))
        process = multiprocessing.Process(target=run_command,
                                        args=(visible_devices, dp_rank,
                                                vllm_engine_port))
        processes.append(process)
        process.start()

    for process in processes:
        process.join()

  2. Prepare run_dp_template.sh on each Prefill node.

    #!/usr/bin/env bash
source /root/.bashrc

nic_name="xxx"
local_ip="xx.xx.xx.x"

export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export HCCL_ALGO=level0:fullmesh
export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=2040
export HCCL_CONNECT_TIMEOUT=1200
export HCCL_BUFFSIZE=512
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ASCEND_RT_VISIBLE_DEVICES=$1

vllm serve /root/.cache/DeepSeek-V4-Pro \
    --host $local_ip \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --max-model-len 1048576 \
    --max-num-batched-tokens 4096 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.85 \
    --enable-expert-parallel \
    --async-scheduling \
    --max-num-seqs 8 \
    --block-size 32 \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --trust-remote-code \
    --enforce-eager \
    --no-disable-hybrid-kv-cache-manager \
    --speculative-config '{"num_speculative_tokens": 1, "method": "mtp"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
      "kv_role": "kv_producer",
      "kv_port": "36010",
      "engine_id": "1",
      "kv_connector_extra_config": {
        "prefill": {
          "dp_size": 32,
          "tp_size": 1
        },
        "decode": {
          "dp_size": 32,
          "tp_size": 1
        },
        "ascend_local_comm_res_path": "/etc/hixlep"
      }
    }' \
    --additional-config '{"enable_cpu_binding": true}'

  3. Prepare run_dp_template.sh on each Decode node.

    #!/usr/bin/env bash
source /root/.bashrc

nic_name="xxx"
local_ip="xx.xx.xx.x"

export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export HCCL_ALGO=level0:fullmesh
export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000
export HCCL_EXEC_TIMEOUT=2040
export HCCL_CONNECT_TIMEOUT=1200
export HCCL_BUFFSIZE=1024
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ASCEND_RT_VISIBLE_DEVICES=$1

vllm serve /root/.cache/DeepSeek-V4-Pro \
    --host $local_ip \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --max-model-len 1048576 \
    --max-num-batched-tokens 256 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.92 \
    --enable-expert-parallel \
    --async-scheduling \
    --max-num-seqs 32 \
    --block-size 32 \
    --no-enable-prefix-caching \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --trust-remote-code \
    --no-disable-hybrid-kv-cache-manager \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --speculative-config '{"num_speculative_tokens": 3, "method": "mtp", "enforce_eager": false}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
      "kv_role": "kv_consumer",
      "kv_port": "36010",
      "engine_id": "1",
      "kv_connector_extra_config": {
        "prefill": {
          "dp_size": 32,
          "tp_size": 1
        },
        "decode": {
          "dp_size": 32,
          "tp_size": 1
        },
        "ascend_local_comm_res_path": "/etc/hixlep"
      }
    }' \
    --additional-config '{"enable_cpu_binding": true, "recompute_scheduler_enable": true, "multistream_overlap_shared_expert": true}'

  4. Start the server with the following command on each node. Set local_ip in run_dp_template.sh to the current node IP before starting the service.

    1. Prefill node 0

      # change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start 0 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 8000

    2. Prefill node 1

      # change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start 8 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 8000

    3. Prefill node 2

      # change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start 16 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 8000

    4. Prefill node 3

      # change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start 24 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 8000

    5. Decode node 0

      # change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start 0 --dp-address xx.xx.xx.5 --dp-rpc-port 12325 --vllm-start-port 8000

    6. Decode node 1

      # change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start 8 --dp-address xx.xx.xx.5 --dp-rpc-port 12325 --vllm-start-port 8000

    7. Decode node 2

      # change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start 16 --dp-address xx.xx.xx.5 --dp-rpc-port 12325 --vllm-start-port 8000

    8. Decode node 3

      # change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start 24 --dp-address xx.xx.xx.5 --dp-rpc-port 12325 --vllm-start-port 8000

  5. Deploy the P-D disaggregation proxy.

    Refer to Prefill-Decode Disaggregation (Deepseek) and register all Prefill and Decode engine endpoints started above. Each node uses ports 8000 through 8007.

Key Parameter Descriptions:

  • --no-disable-hybrid-kv-cache-manager keeps the hybrid KV cache manager enabled. DeepSeek-V4 KV Pool deployments require this flag; otherwise, the service may OOM during startup.
  • --enforce-eager forces eager execution on prefill nodes instead of graph compilation.
  • enable_dsa_cp: true enables DSA context parallelism on prefill nodes. DSA-CP and FlashComm1 must be enabled separately when both are required.
  • kv_connector_extra_config.prefill.dp_size/tp_size and decode.dp_size/tp_size must match the actual global DP and TP layout on the prefill and decode sides.
  • VLLM_ASCEND_ENABLE_FLASHCOMM1=1: enables the communication optimization function on the prefill nodes.
  • VLLM_ASCEND_ENABLE_FUSED_MC2=1: enables the Fused MC2 fusion operator to accelerate communication on prefill nodes (A3 series).
  • recompute_scheduler_enable: true: enables the recomputation scheduler. When the KV Cache of the decode node is insufficient, requests will be sent to the prefill node to recompute the KV Cache. In the PD separation scenario, enable this configuration only on decode nodes.
  • MooncakeHybridConnector: the KV transfer connector used for PD separation, transferring KV Cache between prefill and decode nodes.
  • enable_shared_expert_dp: true: enables data parallelism for shared experts, applicable to MoE models.
  • On Ascend 950DT, ascend_local_comm_res_path specifies the local communication resource directory used by the KV connector. The directory must be available at the same path in the container.

Deployment Verification:

After the PD separation service is fully started, send a request through the proxy port on the prefill master node to verify that Prefill and Decode nodes are working correctly together. Refer to Prefill-Decode Disaggregation (Deepseek) for the proxy verification method.

Common Issues Tip: If you encounter issues with PD separation deployment, please refer to the Public FAQ for troubleshooting.

6 Functional Verification

Once your server is started, you can query the model with input prompts:

In <node0_ip>:, use the IP address and port number of the primary node. If the primary and standby nodes are separated, use the IP address and port number of the proxy node.

curl http://<node0_ip>:<port>/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "dsv4",
        "messages": [
            {
                "role": "user",
                "content": "Who are you?"
            }
        ],
        "max_tokens": 256,
        "temperature": 0
    }'

Expected Result:

The service returns HTTP 200 OK with a JSON response containing the choices field.

7 Accuracy Evaluation

Here are two accuracy evaluation methods.

Using AISBench

  1. Refer to Using AISBench for details.

  2. After execution, you can get the result.

dataset version metric mode vllm-api-general-chat note
GPQA - accuracy gen 89.90 1 Atlas 800 A3 (128GB × 8)
GSM8K - accuracy gen 96.21 1 Atlas 800 A3 (128GB × 8)

8 Performance Evaluation

Using AISBench

Refer to Using AISBench for performance evaluation for details.

Using vLLM Benchmark

Refer to vllm benchmark for more details.

9 Performance Tuning

9.1 Recommended Configurations

Note

: The following configurations are validated in specific test environments and are for reference only. The optimal configuration depends on factors such as maximum input/output length, prefix cache hit rate, precision requirements, and deployment machine ratios. It is recommended to refer to Section 9.2 for tuning based on actual conditions.

Table 1: Scenario Overview

*Total NPUs indicates the total number of NPUs used across all nodes.

Scenario Deployment Mode *Total NPUs Weight Version Key Considerations
High Throughput Single-Node Mixed 32 (A3) DeepSeek-V4-Pro-w4a8-mtp Use dp2 tp16 to balance memory capacity and compute efficiency
High Throughput 1P1D deployment 64 (A3) DeepSeek-V4-Pro-w4a8-mtp dp16 tp2 or dp2 tp16, depending on memory and concurrency
Long Context (1M) Single-Node Mixed 32 (A3) DeepSeek-V4-Pro-w4a8-mtp Use dp2 tp16 to balance memory capacity and compute efficiency
Long Context (1M) 1P1D deployment 64 (A3) DeepSeek-V4-Pro-w4a8-mtp dp2 tp16 on both P and D nodes; balanced latency and throughput
High Throughput Multi-Node Mixed 16 (Ascend 950DT) DeepSeek-V4-Pro Use dp2 tp8 across two Ascend 950DT servers
Long Context (1M) 1P1D deployment 64 (Ascend 950DT) DeepSeek-V4-Pro Use dp32 tp1 on both P and D groups

Table 2: Detailed Node Configuration

Scenario Configuration NPUs TP DP Max Num Seqs Max Num Batched Tokens Max Model Len MTP Speculation Num
Multi-Node (A3) Node0 / Node1 8 16 2 32 4096 135000 1
PD Separation (A3) Prefill Node 8 16 2 16 4096 131072 1
PD Separation (A3) Decode Node 8 2 16 60 120 131072 1
Multi-Node (Ascend 950DT) Node0 / Node1 16 8 2 40 8192 200000 3
PD Separation (Ascend 950DT) Prefill Group 32 1 32 8 4096 1048576 1
PD Separation (Ascend 950DT) Decode Group 32 1 32 32 256 1048576 3

For complete startup commands and parameter descriptions, please refer to the deployment examples in Chapter 5.

9.2 Tuning Guidelines

9.2.1 General Tuning Reference

Please refer to the Public Performance Tuning Documentation for tuning methods.

Please refer to the Feature Guide for detailed feature descriptions.

10 FAQ

For common environment, installation, and general parameter issues, please refer to the Public FAQ; this chapter only covers model-specific issues.