Files

Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>

2026-08-27 15:11:51 +08:00

49 KiB

Raw Blame History

DeepSeek-V4-Flash

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-Flash is the lightweight variant of the DeepSeek-V4 family, suitable for high-throughput and low-latency serving scenarios.

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-Flash-w8a8-mtp (Quantized version): requires 1 Atlas 800 A3 (128GB × 8) node or 1 Atlas 800 A2 (64GB × 8) node. Download model weight
For Ascend 950DT servers, use the original DeepSeek-V4-Flash 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. Both the mixed-deployment example in Section 5.1 and the intra-server 1P1D example in Section 5.2.3 use one Ascend 950DT server (96GB × 8). The mixed deployment uses four NPUs, while the 1P1D deployment assigns four NPUs to Prefill and four NPUs to Decode.

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:

Install vllm-ascend from source, refer to installation.

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 Single-Node Online Deployment

Single-node deployment completes both Prefill and Decode within the same node. The quantized model DeepSeek-V4-Flash-w8a8-mtp can be deployed on 1 Atlas 800 A3 (128GB × 8) or 1 Atlas 800 A2 (64GB × 8). The original DeepSeek-V4-Flash weights can be deployed on 1 Ascend 950DT server (96GB × 8).

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

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

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 LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export HCCL_BUFFSIZE=1024
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export TASK_QUEUE_ENABLE=1
export HCCL_OP_EXPANSION_MODE="AIV"

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Flash-w8a8-mtp \
    --max-model-len 133120 \
    --max-num-batched-tokens 8192 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.9 \
    --max-num-seqs 32 \
    --data-parallel-size 1 \
    --tensor-parallel-size 8 \
    --enable-expert-parallel \
    --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 \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --quantization ascend \
    --port 8900 \
    --block-size 128 \
    --speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
    --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,
    "enable_dsa_cp": true,
    "multistream_overlap_shared_expert":true}'

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

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 LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export HCCL_BUFFSIZE=1024
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export TASK_QUEUE_ENABLE=1
export HCCL_OP_EXPANSION_MODE="AIV"

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V4-Flash-w8a8-mtp \
    --max-model-len 1048576 \
    --max-num-batched-tokens 10240 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.9 \
    --api-server-count 1 \
    --max-num-seqs 64 \
    --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' \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --quantization ascend \
    --port 8900 \
    --block-size 128 \
    --speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
    --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":true}'

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

Run the following script to execute mixed online inference on one Ascend 950DT server.

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=0,1,2,3

vllm serve /root/.cache/DeepSeek-V4-Flash \
    --max-model-len 1048576 \
    --safetensors-load-strategy prefetch \
    --max-num-batched-tokens 4096 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.9 \
    --enable-expert-parallel \
    --async-scheduling \
    --max-num-seqs 64 \
    --port 8900 \
    --block-size 32 \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --data-parallel-size 4 \
    --api-server-count 1 \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --speculative-config '{"num_speculative_tokens": 1, "method": "deepseek_mtp"}' \
    --additional-config '{"enable_cpu_binding": true, "multistream_overlap_shared_expert": true, "enable_shared_expert_dp": true}'

:::: :::::

Key Parameter Descriptions:

--max-model-len specifies the maximum context length - that is, the sum of input and output tokens for a single request. 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.
--data-parallel-size sets the global number of data parallel ranks, while --tensor-parallel-size sets the tensor parallel size within each DP rank. Configure them together 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.
--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 W8A8 model used in the A2 and A3 configurations. The original DeepSeek-V4-Flash weights used on Ascend 950DT do not require this option.
--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_dsa_cp enables DSA context parallelism, enable_shared_expert_dp enables data parallelism for shared experts, and multistream_overlap_shared_expert overlaps shared expert computation for better MoE throughput. DSA-CP depends on FlashComm1, and both options must be enabled explicitly.
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 single-node deployment mode, Prefill (prompt processing) and Decode (token generation) tasks run on the same set of NPUs. This can lead to two issues:

Prefill preemption interrupts Decode: Prefill is a compute-intensive task that processes the entire input context at once, while Decode generates tokens one by one. When a new user request arrives, its Prefill phase can preempt and interrupt ongoing Decode tasks, causing jitter and higher time-per-output-token (TPOT) latency.
Inflexible resource allocation: Prefill and Decode have fundamentally different computational characteristics — Prefill is compute-bound and memory-bandwidth-intensive, while Decode is memory-bandwidth-bound. Running them on the same hardware forces a compromise that satisfies neither optimally.

PD (Prefill-Decode) separation addresses these issues 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 on Atlas 800 A3 nodes (128GB × 8), Atlas 800 A2 nodes (64GB × 8), and Ascend 950DT servers (96GB × 8).

5.2.1 A3 Series PD Separation Deployment

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

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.

Prepare the script run_dp_template.sh on each node.

Prefill node

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

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 VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export HCCL_OP_EXPANSION_MODE="AIV"
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
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 dsv4 \
    --max-model-len 1048576 \
    --max-num-batched-tokens 8192 \
    --max-num-seqs 16 \
    --no-disable-hybrid-kv-cache-manager \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --no-enable-prefix-caching \
    --safetensors-load-strategy 'prefetch' \
    --speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \
    --trust-remote-code \
    --block-size 128 \
    --tokenizer-mode deepseek_v4 \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --gpu-memory-utilization 0.9 \
    --quantization ascend \
    --enforce-eager \
    --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": 4
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 1
                }
        }
    }'

Decode node

nic_name="xxxx" # change to your own nic name
local_ip=xx.xx.xx.2 # 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
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 dsv4 \
    --max-model-len 1048576 \
    --max-num-batched-tokens 120 \
    --max-num-seqs 60 \
    --async-scheduling \
    --block-size 128 \
    --no-disable-hybrid-kv-cache-manager \
    --no-enable-prefix-caching \
    --safetensors-load-strategy 'prefetch' \
    --trust-remote-code \
    --tokenizer-mode deepseek_v4 \
    --model-loader-extra-config='{"enable_multithread_load": "true", "num_threads": 128}' \
    --tool-call-parser deepseek_v4 \
    --enable-auto-tool-choice \
    --reasoning-parser deepseek_v4 \
    --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": "30100",
    "engine_id": "1",
    "kv_connector_extra_config": {
                "prefill": {
                        "dp_size": 4,
                        "tp_size": 4
                },
                "decode": {
                        "dp_size": 16,
                        "tp_size": 1
                }
        }
    }' \
    --additional-config '{
        "ascend_compilation_config":{
            "enable_npugraph_ex":true,
            "enable_static_kernel":false
        },
        "enable_cpu_binding":true,
        "multistream_overlap_shared_expert":true,
        "recompute_scheduler_enable":true
    }'

Start the server with the following command on each node.

Prefill node

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

Decode node

# 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

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-Flash on Atlas 800 A2 (64GB × 8) multi-node environment with 4*1P 1*4D 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 (4 P nodes share the same script)

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.

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}'`

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

export ASCEND_RT_VISIBLE_DEVICES=$1
export TASK_QUEUE_ENABLE=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 dsv4 \
    --max-model-len 135000 \
    --max-num-batched-tokens 4096 \
    --max-num-seqs 16 \
    --block-size 128 \
    --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' \
    --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 \
    --additional-config '{"enable_cpu_binding": true, "enable_shared_expert_dp": true}' \
    --speculative-config '{"num_speculative_tokens": 1, "method": "mtp","enforce_eager": true}' \
    --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
                }
        }
    }'

Decode node (4 D nodes share the same script)

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}'`

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

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 dsv4 \
    --max-model-len 135000 \
    --max-num-batched-tokens 60 \
    --max-num-seqs 30 \
    --async-scheduling \
    --block-size 128 \
    --no-disable-hybrid-kv-cache-manager \
    --no-enable-prefix-caching \
    --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 \
    --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": "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
        },
       "enable_cpu_binding":true,
       "multistream_overlap_shared_expert":true,
       "recompute_scheduler_enable":true
    }'

Start the server with the following command on each node.
1. 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.
2. 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.
Deploy the P-D disaggregation proxy.

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

5.2.3 Ascend 950DT Series PD Separation Deployment

This section shows an intra-server 1P1D deployment on one Ascend 950DT server (96GB × 8). The Prefill instance uses NPUs 0-3 with DP4/TP1, and the Decode instance uses NPUs 4-7 with DP4/TP1. You can add Prefill or Decode instances as needed based on the workload's input/output characteristics and service requirements. An expert-parallel size (ep_size) of 4 is recommended for each instance. The prefill and decode parallel settings in --kv-transfer-config must match the actual engine settings after scaling.

Before starting the service, mount /etc/hixlep/ into the container and replace nic_name, local_ip, and the model path with values from your environment.

Prepare run_prefill.sh.

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

nic_name="xxx"
local_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=204
export HCCL_CONNECT_TIMEOUT=120
export HCCL_BUFFSIZE=512
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
export ASCEND_LOCAL_COMM_RES_PATH=/etc/hixlep/
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

vllm serve /root/.cache/DeepSeek-V4-Flash \
    --host $local_ip \
    --port 8000 \
    --data-parallel-size 4 \
    --data-parallel-address $local_ip \
    --data-parallel-rpc-port 12321 \
    --tensor-parallel-size 1 \
    --max-model-len 1048576 \
    --max-num-batched-tokens 3072 \
    --served-model-name dsv4 \
    --gpu-memory-utilization 0.85 \
    --enable-expert-parallel \
    --async-scheduling \
    --max-num-seqs 8 \
    --block-size 32 \
    --api-server-count 1 \
    --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": 4,
          "tp_size": 1
        },
        "decode": {
          "dp_size": 4,
          "tp_size": 1
        },
        "ascend_local_comm_res_path": "/etc/hixlep"
      }
    }' \
    --additional-config '{"enable_cpu_binding": true}'

Prepare run_decode.sh.

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

nic_name="xxx"
local_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=1024
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
export ASCEND_LOCAL_COMM_RES_PATH=/etc/hixlep/
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ASCEND_RT_VISIBLE_DEVICES=4,5,6,7

vllm serve /root/.cache/DeepSeek-V4-Flash \
    --host $local_ip \
    --port 8001 \
    --data-parallel-size 4 \
    --data-parallel-address $local_ip \
    --data-parallel-rpc-port 12325 \
    --tensor-parallel-size 1 \
    --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 56 \
    --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": 1, "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": 4,
          "tp_size": 1
        },
        "decode": {
          "dp_size": 4,
          "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}'

Start the Prefill and Decode services in separate terminals.

# Prefill terminal
bash run_prefill.sh

# Decode terminal
bash run_decode.sh

Deploy the P-D disaggregation proxy.

Refer to Prefill-Decode Disaggregation (Deepseek) and configure the Prefill endpoint as <local_ip>:8000 and the Decode endpoint as <local_ip>:8001.

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.
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.
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.

5.2.4 Ultra-Long Sequence Deployment

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.

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 is the accuracy evaluation method using AISBench.

Using AISBench

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

dataset	version	metric	mode	vllm-api-general-chat	note
GPQA	-	accuracy	gen	88.17	1 Atlas 800 A3 (128GB × 8)
GSM8K	-	accuracy	gen	96.30	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	16 (A3)	DeepSeek-V4-Flash-w8a8-mtp	Use dp4 tp4 to balance memory capacity and compute efficiency
High Throughput / Long Context (1M)	Single-Node Mixed	4 (Ascend 950DT)	DeepSeek-V4-Flash	Use dp4 tp1 with expert parallelism
High Throughput	1P1D deployment	32 (A3)	DeepSeek-V4-Flash-w8a8-mtp	dp16 tp1 on both P and D nodes; balanced latency and throughput
Long Context (1M)	Single-Node (A3)	8 (A3)	DeepSeek-V4-Flash-w8a8-mtp	Use dp4 tp4 to balance memory capacity and compute efficiency
Long Context (1M)	1P1D deployment	32 (A3)	DeepSeek-V4-Flash-w8a8-mtp	dp16 tp1 on both P and D nodes; balanced latency and throughput
Long Context (1M)	1P1D deployment	8 (Ascend 950DT)	DeepSeek-V4-Flash	Split one Ascend 950DT server into P and D groups; both groups use dp4 tp1

Table 2: Detailed Node Configuration

Scenario	Configuration	NPUs	TP	DP	Max Num Seqs	Max Num Batched Tokens	Max Model Len	MTP Speculation Num
High Throughput (A3)	Server / Single Machine	8	4	4	64	10240	1048576	1
Single-Node Mixed (Ascend 950DT)	Server / Single Machine	4	1	4	64	4096	1048576	1
Long Context (1M, A3)	Server / Single Machine	8	4	4	64	10240	1048576	1
PD Separation (A3)	Server-P Node	8	4	4	16	8192	1048576	1
PD Separation (A3)	Server-D Node	8	1	16	60	120	1048576	1
PD Separation (Ascend 950DT)	Prefill Group	4	1	4	8	3072	1048576	1
PD Separation (Ascend 950DT)	Decode Group	4	1	4	56	256	1048576	1

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.

49 KiB Raw Blame History Unescape Escape

DeepSeek-V4-Flash

1 Introduction

2 Supported Features

3 Prerequisites

3.1 Model Weight

3.2 Verify Multi-node Communication (Optional)

4 Installation

4.1 Docker Image Installation

4.2 Source Code Installation

5 Online Service Deployment

5.1 Single-Node Online Deployment

5.2 Multi-Node PD Separation Deployment

5.2.1 A3 Series PD Separation Deployment

5.2.2 A2 Series PD Separation Deployment

5.2.3 Ascend 950DT Series PD Separation Deployment

5.2.4 Ultra-Long Sequence Deployment

6 Functional Verification

7 Accuracy Evaluation

Using AISBench

8 Performance Evaluation

Using AISBench

Using vLLM Benchmark

9 Performance Tuning

9.1 Recommended Configurations

Table 1: Scenario Overview

Table 2: Detailed Node Configuration

9.2 Tuning Guidelines

9.2.1 General Tuning Reference

10 FAQ

49 KiB

Raw Blame History