DeepSeek-V4-Pro

DeepSeek-V4-Pro#

简介#

DeepSeek-V4 在 DeepSeek-V3 基础上引入了多项关键升级。

  • 流形约束超连接（mHC）以增强传统残差连接；

  • 混合注意力架构，通过 Compress-4-Attention 和 Compress-128-Attention 大幅提升长上下文效率。对于混合专家（MoE）组件，仍采用 DeepSeekMoE 架构，仅做少量调整。

DeepSeek-V4-Pro 作为 DeepSeek-V4-Pro 的最大推理努力模式，显著提升了开源模型的知识能力，牢固确立了其作为当今最佳开源模型的地位。它在编码基准测试中达到顶级性能，并在推理和智能体任务上显著缩小了与领先闭源模型的差距。

本文档将展示该模型的主要验证步骤，包括支持的特性、特性配置、环境准备、单节点和多节点部署、精度及性能评估。

环境准备#

模型权重#

  • DeepSeek-V4-Pro-w4a8-mtp（量化版本）：需要 2 个 Atlas 800 A3（128G × 8）节点或 4 个 Atlas 800 A2（64G × 8）节点。下载模型权重

建议将模型权重下载到多节点共享目录，例如 /root/.cache/

验证多节点通信（可选）#

如需部署多节点环境，需按照验证多节点通信环境验证多节点通信。

安装#

您可以直接使用我们的官方 Docker 镜像运行 DeepSeek-V4。

在每个节点上启动 Docker 镜像。

export IMAGE=quay.io/ascend/vllm-ascend:deepseekv4
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/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 \
    -v /mnt/sfs_turbo/.cache:/root/.cache \
    -it $IMAGE bash

在每个节点上启动 Docker 镜像。

export IMAGE=quay.io/ascend/vllm-ascend:deepseekv4-a3
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/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 \
    -v /mnt/sfs_turbo/.cache:/root/.cache \
    -it $IMAGE bash

部署#

备注

在本教程中，我们假设您已将模型权重下载到 /root/.cache/。您可以根据需要更改为自己的路径。

多节点部署#

  • DeepSeek-V4-Pro-w4a8-mtp：至少需要 2 个 Atlas 800 A3（128G × 8）或 4 个 Atlas 800 A2（64G × 8）。请分别在每个节点上运行以下脚本。

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 USE_MULTI_BLOCK_POOL=1
export USE_MULTI_GROUPS_KV_CACHE=1
export TASK_QUEUE_ENABLE=1
#export DYNAMIC_EPLB=true
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

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

echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
  --host 0.0.0.0 \
  --port 10010 \
  --max_model_len 133072 \
  --max-num-batched-tokens 4096 \
  --served-model-name ds-v4 \
  --gpu-memory-utilization 0.9 \
  --max-num-seqs 4 \
  --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 \
  --enable-chunked-prefill \
  --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' \
  --profiler-config '{"profiler": "torch", "torch_profiler_dir": "/path", "torch_profiler_with_stack": false}' \
  --speculative-config '{"num_speculative_tokens": 1,"method": "mtp"}' \
  --additional-config '{"ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false},"enable_cpu_binding":"True"}' \
  --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}'

Node1-Node3

local_ip="xxx"
node0_ip="xxxx"
data_parallel_start_rank=xxx

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 USE_MULTI_BLOCK_POOL=1
export USE_MULTI_GROUPS_KV_CACHE=1
export TASK_QUEUE_ENABLE=1
#export DYNAMIC_EPLB=true
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

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

echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp \
  --host 0.0.0.0 \
  --port 10010 \
  --headless \
  --max_model_len 133072 \
  --max-num-batched-tokens 4096 \
  --served-model-name ds-v4 \
  --gpu-memory-utilization 0.9 \
  --max-num-seqs 4 \
  --data-parallel-size 4 \
  --tensor-parallel-size 8 \
  --data-parallel-size-local 1 \
  --data-parallel-start-rank $data_parallel_start_rank \
  --data-parallel-address $node0_ip \
  --enable-expert-parallel \
  --quantization ascend \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --async-scheduling \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --safetensors-load-strategy 'prefetch' \
  --speculative-config '{"num_speculative_tokens": 1,"method": "mtp"}' \
  --profiler-config '{"profiler": "torch", "torch_profiler_dir": "/path", "torch_profiler_with_stack": false}' \
  --additional-config '{"ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false},"enable_cpu_binding":"True"}' \
  --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}'

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

echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

export USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export VLLM_ASCEND_ENABLE_FUSED_MC2=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 16 \
  --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":"deepseek_mtp"}' \
  --additional-config '{"enable_cpu_binding": "true", "ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false}}' \

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

echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

export USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export VLLM_ASCEND_ENABLE_FUSED_MC2=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 16 \
  --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":"deepseek_mtp"}' \
  --additional-config '{"enable_cpu_binding": "true", "ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false}}' \

Prefill-Decode 分离#

我们将展示 DeepSeek-V4 在 Atlas 800 A3（128G × 8）多节点环境下采用 1P1D 配置以获得更好性能的部署指南。

开始前，请

  1. 在每个节点上准备脚本 launch_online_dp.py。

    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. 在每个节点上准备脚本 run_dp_template.sh。

    1. Prefill 节点 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 USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export VLLM_ASCEND_ENABLE_FUSED_MC2=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 auto \
    --max-model-len 131072 \
    --max-num-batched-tokens 8192 \
    --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' \
    --trust-remote-code \
    --gpu-memory-utilization 0.9 \
    --quantization ascend \
    --block-size 128 \
    --enforce-eager \
    --speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
    --additional_config '{"enable_cpu_binding": "True"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30200",
    "engine_id": "1",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 4,
                        "tp_size": 8
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 2
                }
        }
    }'

    2. Prefill 节点 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 USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export VLLM_ASCEND_ENABLE_FUSED_MC2=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 auto \
    --max-model-len 131072 \
    --max-num-batched-tokens 8192 \
    --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' \
    --trust-remote-code \
    --gpu-memory-utilization 0.9 \
    --quantization ascend \
    --block-size 128 \
    --enforce-eager \
    --speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
    --additional_config '{"enable_cpu_binding": "True"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30200",
    "engine_id": "1",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 4,
                        "tp_size": 8
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 2
                }
        }
    }'

    3. Decode 节点（与另一个 D 节点相同）

      nic_name="xxxx" # change to your own nic name
local_ip=xx.xx.xx.xx # 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 USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=1
export VLLM_ASCEND_ENABLE_FUSED_MC2=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 auto \
    --max-model-len 131072 \
    --max-num-batched-tokens 120 \
    --max-num-seqs 60 \
    --async-scheduling \
    --block-size 128 \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --no-disable-hybrid-kv-cache-manager \
    --no-enable-prefix-caching \
    --safetensors-load-strategy 'prefetch' \
    --trust-remote-code \
    --gpu-memory-utilization 0.9 \
    --quantization ascend \
    --speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_consumer",
    "kv_port": "30200",
    "engine_id": "2",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 4,
                        "tp_size": 8
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 2
                }
        }
    }' \
    --additional-config '{
        "ascend_compilation_config":{
            "enable_npugraph_ex":true,
            "enable_static_kernel":false
        },
    "enable_cpu_binding":true,
    "multistream_overlap_shared_expert":false,
    "multistream_dsa_preprocess":false,
    "recompute_scheduler_enable":true
    }'

准备完成后，您可以在每个节点上使用以下命令启动服务器：

  1. Prefill 节点 0

# change ip to your own
python launch_online_dp.py --dp-size 4 --tp-size 8 --dp-size-local 2 --dp-rank-start 0 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100

  1. Prefill 节点 1

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

  1. Decode 节点 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

  1. Decode 节点 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

最后，请参考 Prefill-Decode 分离 (Deepseek) 部署 P-D 分离代理。

A2 Prefill-Decode 分离#

我们将展示 DeepSeek-V4 在 Atlas 800 A2（64G × 8）多节点环境下采用 1P1D 配置以获得更好性能的部署指南。

开始前，请

  1. 在每个节点上准备脚本 launch_online_dp.py。

    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. 在每个节点上准备脚本 run_dp_template.sh。

    1. Prefill 节点（与另一个 P 节点相同）

      nic_name="xxxx" # change to your own nic name
local_ip=`hostname -I|awk -F " " '{print$1}'`

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 USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=1
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 deepseek_v4 \
    --max_model_len 133072 \
    --max-num-batched-tokens 8192 \
    --max-num-seqs 16 \
    --no-disable-hybrid-kv-cache-manager \
    --trust-remote-code \
    --gpu-memory-utilization 0.9 \
    --quantization ascend \
    --safetensors-load-strategy 'prefetch' \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --enforce-eager \
    --enable-prefix-caching \
    --speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
    --additional_config '{"enable_cpu_binding": "True"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30000",
    "engine_id": "0",
    "kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
    "kv_connector_extra_config": {
              "prefill": {
                      "dp_size": 4,
                      "tp_size": 8
               },
               "decode": {
                      "dp_size": 8,
                      "tp_size": 4
               }
        }
    }'

    2. Decode 节点（与另一个 D 节点相同）

      nic_name="xxxx" # change to your own nic name
local_ip=`hostname -I|awk -F " " '{print$1}'`

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 USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=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 deepseek_v4 \
    --max-model-len 133072 \
    --max-num-batched-tokens 120 \
    --max-num-seqs 60 \
    --async-scheduling \
    --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' \
    --no-enable-prefix-caching \
    --speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
    --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_module_path": "vllm_ascend.distributed.mooncake_connector",
    "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,"multistream_overlap_shared_expert":false,"multistream_dsa_preprocess":false,"recompute_scheduler_enable":true}'

准备完成后，您可以在每个节点上使用以下命令启动服务器：

  1. Prefill 节点 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

  1. Prefill 节点 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

  1. Prefill 节点 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

  1. Prefill 节点 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

  1. Decode 节点 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

  1. Decode 节点 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

  1. Decode 节点 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

  1. Decode 节点 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

最后，请参考 Prefill-Decode 分离 (Deepseek) 部署 P-D 分离代理。

功能验证#

服务器启动后，您可以使用输入提示查询模型：

curl http://<node0_ip>:<port>/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "deepseek_v4",
        "messages": [
            {
                "role": "user",
                "content": "Who are you?"
            }
        ],
        "max_tokens": 256,
        "temperature": 0
    }'

精度评估#

以下是两种精度评估方法。

使用 AISBench#

  1. 详情请参考使用 AISBench。

  2. 执行后即可获取结果。

使用 Language Model Evaluation Harness#

例如，以 gsm8k 数据集作为测试数据集，在线模式下运行 DeepSeek-V4 的精度评估。

  1. lm_eval 的安装请参考使用 lm_eval。

  2. 运行 lm_eval 执行精度评估。

lm_eval \
  --model local-completions \
  --model_args model=/root/.cache/Eco-Tech/DeepSeek-V4-Pro-w4a8-mtp,base_url=http://127.0.0.1:8006/v1/completions,tokenized_requests=False,trust_remote_code=True \
  --tasks gsm8k \
  --output_path ./

  1. 执行后即可获取结果。

性能#

使用 AISBench#

详情请参考使用 AISBench 进行性能评估。

使用 vLLM Benchmark#

以 DeepSeek-V4-Pro-w4a8-mtp 为例运行性能评估。

更多详情请参考 vllm benchmark。

vllm bench 包含三个子命令：

  • latency：基准测试单批请求的延迟。

  • serve：基准测试在线服务吞吐量。

  • throughput：基准测试离线推理吞吐量。

以 serve 为例，运行以下代码。

export VLLM_USE_MODELSCOPE=true
vllm bench serve --model /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Pro-w4a8-mtp  --dataset-name random --random-input 200 --num-prompt 200 --request-rate 1 --save-result --result-dir ./