DeepSeek-V4-Flash

DeepSeek-V4-Flash#

Introduction#

DeepSeek-V4 is introducing 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 adopt the DeepSeekMoE architecture, with only minor adjustments.

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.

Environment Preparation#

Model Weight#

DeepSeek-V4-Flash-w8a8-mtp(Quantized version): require 1 Atlas 800 A3 (128G × 8) node or 1 Atlas 800 A2 (64G × 8) node. Download model weight

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

Verify Multi-node Communication(Optional)#

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

Installation#

You can using our official docker image to run DeepSeek-V4 directly.

A2 series

Start the docker image on your each node.

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

A3 series

Start the docker image on your each node.

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

Deployment#

备注

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

Single-node Deployment#

DeepSeek-V4-Flash-w8a8-mtp: can be deployed on 1 Atlas 800 A3 (128G × 8) or 1 Atlas 800 A2 (64G × 8).

Run the following scripts on each node respectively.

A2 series

Run the following script to execute online inference.

#!/usr/bin/bash
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=8
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ACL_OP_INIT_MODE=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

export USE_MULTI_GROUPS_KV_CACHE=1

export TASK_QUEUE_ENABLE=1
export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_BUFFSIZE=512

export USE_MULTI_BLOCK_POOL=1

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

vllm serve /mnt/nfs_hw/weight/DeepSeek-V4-Flash-w8a8-mtp \
  --safetensors-load-strategy 'prefetch' \
  --max-model-len 135168 \
  --max-num-batched-tokens 4096 \
  --served-model-name ds \
  --gpu-memory-utilization 0.92 \
  --max-num-seqs 16 \
  --data-parallel-size 1 \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --quantization ascend \
  --port 7000 \
  --block-size 128 \
  --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 \
  --additional-config '{"enable_cpu_binding":true,"multistream_overlap_shared_expert":false}' \
  --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[2,4,6,8,10,12,14,16,18,20,22,24,32,36,40]}' \
  --model-loader-extra-config '{"enable_multithread_load":true,"num_threads":16}' \
  --speculative-config '{"num_speculative_tokens": 1,"method": "mtp"}'

A3 series

Run the following script to execute online inference.

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 ASCEND_A3_ENABLE=1
export USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=1
export HCCL_BUFFSIZE=1024
export VLLM_ASCEND_ENABLE_FUSED_MC2=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Flash-w8a8-mtp \
    --enable-prefix-caching \
    --max_model_len 1024000 \
    --max-num-batched-tokens 8192 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.9 \
    --api-server-count 1 \
    --max-num-seqs 16 \
    --data-parallel-size 4 \
    --tensor-parallel-size 4 \
    --enable-expert-parallel \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --safetensors-load-strategy 'prefetch' \
    --quantization ascend \
    --speculative-config '{"num_speculative_tokens": 1,"method": "deepseek_mtp"}' \
    --port 8008 \
    --block-size 128 \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'\
    --async-scheduling \
    --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}'

Prefill-Decode Disaggregation#

We'd like to show the deployment guide of DeepSeek-V4 on Atlas 800 A3 (128G × 8) multi-node environment with 2P1D for better performance.

Before you start, please

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()

prepare the script run_dp_template.sh on each node.

Prefill node 1

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=2560
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 ASCEND_RT_VISIBLE_DEVICES=$1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Flash-w8a8-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 135000 \
    --max-num-batched-tokens 8192 \
    --max-num-seqs 4 \
    --no-disable-hybrid-kv-cache-manager \
    --no-enable-prefix-caching \
    --safetensors-load-strategy 'prefetch' \
    --trust-remote-code \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --gpu-memory-utilization 0.85 \
    --quantization ascend \
    --profiler-config \
        '{"profiler": "torch",
        "torch_profiler_dir": "./vllm_profile",
        "torch_profiler_with_stack": false}' \
    --enforce-eager \
    --additional_config '{"enable_cpu_binding": "True"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30100",
    "engine_id": "1",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 16,
                        "tp_size": 1
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 1
                }
        }
    }'

Prefill node 2

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=2560
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 ASCEND_RT_VISIBLE_DEVICES=$1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Flash-w8a8-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 135000 \
    --max-num-batched-tokens 8192 \
    --max-num-seqs 4 \
    --no-disable-hybrid-kv-cache-manager \
    --no-enable-prefix-caching \
    --safetensors-load-strategy 'prefetch' \
    --trust-remote-code \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --gpu-memory-utilization 0.85 \
    --quantization ascend \
    --speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
    --enforce-eager \
    --additional_config '{"enable_cpu_binding": "True"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30200",
    "engine_id": "2",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 16,
                        "tp_size": 1
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 1
                }
        }
    }'

Decode node (Same as another D node)

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 DYNAMIC_EPLB="true"

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-Flash-w8a8-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 135000 \
    --max-num-batched-tokens 120 \
    --max-num-seqs 60 \
    --async-scheduling \
    --no-disable-hybrid-kv-cache-manager \
    --no-enable-prefix-caching \
    --safetensors-load-strategy 'prefetch' \
    --trust-remote-code \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --gpu-memory-utilization 0.88 \
    --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": "30300",
    "engine_id": "3",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 16,
                        "tp_size": 1
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 1
                }
        }
    }' \
    --additional-config '{
        "ascend_compilation_config":{
            "enable_npugraph_ex":true,
            "enable_static_kernel":false
        },
        "eplb_config":{
            "dynamic_eplb":true,
            "expert_heat_collection_interval":600,
            "algorithm_execution_interval":50,
            "eplb_policy_type":2,
            "num_redundant_experts":16
    },
    "enable_cpu_binding":true,
    "multistream_overlap_shared_expert":false,
    "multistream_dsa_preprocess":false,
    "recompute_scheduler_enable":true
    }'

Once the preparation is done, you can start the server with the following command on each node:

Prefill node 0

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

Prefill node 1

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

Decode node 0

# change ip to your own
python launch_online_dp.py --dp-size 32 --dp-size-local 16 --dp-rank-start 0 --dp-address xx.xx.xx.3 --dp-rpc-port 12321 --vllm-start-port 7100

Decode node 1

# change ip to your own
python launch_online_dp.py --dp-size 32 --dp-size-local 16 --dp-rank-start 16 --dp-address xx.xx.xx.3 --dp-rpc-port 12321 --vllm-start-port 7100

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

For Atlas 800 A2 series machines, we can configure the deployment(4*1P 1*4D) as follows:.

Before you start, please

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()

prepare the script run_dp_template.sh on each node.

Prefill node

unset ftp_proxy
unset https_proxy
unset http_proxy
rm -rf ~/ascend/log

nic_name="xxxxxx" #eg."enp67s0f0np0"
local_ip=`hostname -I|awk -F " " '{print$1}'`

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

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-Flash-w8a8-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 137216 \
    --max-num-batched-tokens 8192 \
    --max-num-seqs 16 \
    --enforce-eager \
    --async-scheduling \
    --no-disable-hybrid-kv-cache-manager \
    --enable-prefix-caching \
    --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 \
    --additional-config '{"enable_cpu_binding":"True"}' \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --kv-transfer-config \
    '{"kv_connector": "MooncakeHybridConnector",
    "kv_role": "kv_producer",
    "kv_port": "30000",
    "engine_id": "0",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 8,
                        "tp_size": 1
                },
                "decode": {
                        "dp_size": 32,
                        "tp_size": 1
                }
        }
    }'
            

For each P instance, only these two configuration values need to be modified: “kv_port” and “engine_id”. The “engine_id” should start from 0 and increment sequentially, while the “kv_port” (e.g., “30100”) must be unique for each P instance, such as 30000, 30100, etc.

Decode node(Same as another D node)

unset ftp_proxy
unset https_proxy
unset http_proxy
rm -rf ~/ascend/log

nic_name="xxxxxxx" #eg. "enp67s0f0np0"
local_ip=`hostname -I|awk -F " " '{print$1}'`

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

export USE_MULTI_GROUPS_KV_CACHE=1
export USE_MULTI_BLOCK_POOL=1

export VLLM_TORCH_PROFILER_WITH_STACK=0
export VLLM_TORCH_PROFILER_DIR="./profiling"


export ASCEND_RT_VISIBLE_DEVICES=$1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Flash-w8a8-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 137216 \
    --max-num-batched-tokens 60 \
    --max-num-seqs 30 \
    --async-scheduling \
    --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 \
    --quantization ascend \
    --profiler-config \
        '{"profiler": "torch",
        "torch_profiler_dir": "./profiling",
        "torch_profiler_with_stack": false}' \
    --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": "30400",
    "engine_id": "4",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 8,
                        "tp_size": 1
                },
                "decode": {
                        "dp_size": 32,
                        "tp_size": 1
                }
        }
    }' \
    --additional-config '{
        "ascend_compilation_config":{
            "enable_npugraph_ex":true,
            "enable_static_kernel":false
        },
        "eplb_config":{
            "dynamic_eplb":false,
            "expert_heat_collection_interval":600,
            "algorithm_execution_interval":50,
            "eplb_policy_type":2,
            "num_redundant_experts":32
    },
    "enable_cpu_binding":true,
    "multistream_overlap_shared_expert":false,
    "multistream_dsa_preprocess":false,
    "recompute_scheduler_enable":true
    }' 

Once the preparation is done, you can start the server with the following command on each node:

Prefill node

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

For each P instance, only the --dp-address parameter differs and must be configured as the IP address of the service within the same subnet as the other instances.

Decode node

# change ip to your own
python launch_online_dp.py --dp-size 32 --tp-size 1 --dp-size-local 8 --dp-rank-start x --dp-address x.x.x.x --dp-rpc-port 12321 --vllm-start-port 7100

For each D instance, only the --dp-rank-start parameter differs, which should be configured as 0, 8, 16, and 24 respectively.Each instance’s --dp-address must be set to the IP address of the main D node, which is the IP of the Decode instance with --dp-rank-start set to 0.

The proxy is also implemented by referring Prefill-Decode Disaggregation (Deepseek).

For ultra-long sequence scenarios, support can be achieved by adjusting the PD (Prefill/Decode) ratio and the model parallelism strategy. For example, in a 1M sequence scenario, a 1*4P-1*4D ratio can be used, with the model parallelism set to DP4TP8 mode.

Functional Verification#

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

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
    }'

Accuracy Evaluation#

Here are two accuracy evaluation methods.

Using AISBench#

Refer to Using AISBench for details.
After execution, you can get the result.

Using Language Model Evaluation Harness#

As an example, take the gsm8k dataset as a test dataset, and run accuracy evaluation of DeepSeek-V4 in online mode.

Refer to Using lm_eval for lm_eval installation.
Run lm_eval to execute the accuracy evaluation.

lm_eval \
  --model local-completions \
  --model_args model=/root/.cache/Eco-Tech/DeepSeek-V4-Flash-w8a8-mtp,base_url=http://127.0.0.1:8006/v1/completions,tokenized_requests=False,trust_remote_code=True \
  --tasks gsm8k \
  --output_path ./

After execution, you can get the result.

Performance#

Using AISBench#

Refer to Using AISBench for performance evaluation for details.

Using vLLM Benchmark#

Run performance evaluation of DeepSeek-V4-Flash-w8a8-mtp as an example.

Refer to vllm benchmark for more details.

There are three vllm bench subcommand:

latency: Benchmark the latency of a single batch of requests.
serve: Benchmark the online serving throughput.
throughput: Benchmark offline inference throughput.

Take the serve as an example. Run the code as follows.

export VLLM_USE_MODELSCOPE=true
vllm bench serve --model /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Flash-w8a8-mtp  --dataset-name random --random-input 200 --num-prompt 200 --request-rate 1 --save-result --result-dir ./

DeepSeek-V4-Flash

目录

DeepSeek-V4-Flash#

Introduction#

Environment Preparation#

Model Weight#

Verify Multi-node Communication(Optional)#

Installation#

Deployment#

Single-node Deployment#

Prefill-Decode Disaggregation#

Functional Verification#

Accuracy Evaluation#

Using AISBench#

Using Language Model Evaluation Harness#

Performance#

Using AISBench#

Using vLLM Benchmark#