xc-llm-ascend/docs/source/tutorials/DeepSeek-V3.2-Exp.md

DeepSeek-V3.2-Exp

Introduction

DeepSeek-V3.2-Exp is a sparse attention model. The main architecture is similar to DeepSeek-V3.1, but with a sparse attention mechanism, which is designed to explore and validate optimizations for training and inference efficiency in long-context 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.

Supported Features

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

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

Environment Preparation

Model Weight

• DeepSeek-V3.2-Exp(BF16 version): require 2 Atlas 800 A3 (64G × 16) nodes or 4 Atlas 800 A2 (64G × 8) nodes. Download model weight
• DeepSeek-V3.2-Exp-w8a8(Quantized version): require 1 Atlas 800 A3 (64G × 16) node or 2 Atlas 800 A2 (64G × 8) nodes. 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 and install extra operator for supporting DeepSeek-V3.2-Exp.

:::{note} Only AArch64 architecture are supported currently due to extra operator's installation limitations. :::

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

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

1. Start the docker image on your node, refer to using docker.

2. Install the package custom-ops to make the kernels available.

wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/CANN-custom_ops-sfa-linux.aarch64.run
chmod +x ./CANN-custom_ops-sfa-linux.aarch64.run
./CANN-custom_ops-sfa-linux.aarch64.run --quiet
export ASCEND_CUSTOM_OPP_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize:${ASCEND_CUSTOM_OPP_PATH}
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/op_api/lib/:${LD_LIBRARY_PATH}
wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a3/custom_ops-1.0-cp311-cp311-linux_aarch64.whl
pip install custom_ops-1.0-cp311-cp311-linux_aarch64.whl

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

1. Start the docker image on your node, refer to using docker.

2. Install the package custom-ops to make the kernels available.

wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a2/CANN-custom_ops-sfa-linux.aarch64.run
chmod +x ./CANN-custom_ops-sfa-linux.aarch64.run
./CANN-custom_ops-sfa-linux.aarch64.run --quiet
export ASCEND_CUSTOM_OPP_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize:${ASCEND_CUSTOM_OPP_PATH}
export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/op_api/lib/:${LD_LIBRARY_PATH}
wget https://vllm-ascend.obs.cn-north-4.myhuaweicloud.com/vllm-ascend/a2/custom_ops-1.0-cp311-cp311-linux_aarch64.whl
pip install custom_ops-1.0-cp311-cp311-linux_aarch64.whl

:::: :::::

In addition, 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.

• Install extra operator for supporting DeepSeek-V3.2-Exp, refer to the above tab.

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

Deployment

Single-node Deployment

Only the quantized model DeepSeek-V3.2-Exp-w8a8 can be deployed on 1 Atlas 800 A3.

Run the following script to execute online inference.

#!/bin/sh
export VLLM_USE_MODELSCOPE=true

vllm serve vllm-ascend/DeepSeek-V3.2-Exp-W8A8 \
--host 0.0.0.0 \
--port 8000 \
--tensor-parallel-size 16 \
--seed 1024 \
--quantization ascend \
--served-model-name deepseek_v3.2 \
--max-num-seqs 16 \
--max-model-len 17450 \
--max-num-batched-tokens 17450 \
--enable-expert-parallel \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--additional-config '{"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'

Multi-node Deployment

• DeepSeek-V3.2-Exp: require 2 Atlas 800 A3 (64G × 16) nodes or 4 Atlas 800 A2 (64G × 8).
• DeepSeek-V3.2-Exp-w8a8: require 2 Atlas 800 A2 (64G × 8).

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

::::{tab-item} DeepSeek-V3.2-Exp A3 series :sync: A3

Run the following scripts on two nodes respectively.

Node 0

#!/bin/sh

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

export VLLM_USE_MODELSCOPE=True
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=100
export HCCL_BUFFSIZE=1024

vllm serve /root/.cache/Modelers_Park/DeepSeek-V3.2-Exp \
--host 0.0.0.0 \
--port 8000 \
--data-parallel-size 2 \
--data-parallel-size-local 1 \
--data-parallel-address $local_ip \
--data-parallel-rpc-port 13389 \
--tensor-parallel-size 16 \
--seed 1024 \
--served-model-name deepseek_v3.2 \
--enable-expert-parallel \
--max-num-seqs 16 \
--max-model-len 17450 \
--max-num-batched-tokens 17450 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.9 \
--additional-config '{"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'

Node 1

#!/bin/sh

# 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 VLLM_USE_MODELSCOPE=True
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=100
export HCCL_BUFFSIZE=1024

vllm serve /root/.cache/Modelers_Park/DeepSeek-V3.2-Exp \
--host 0.0.0.0 \
--port 8000 \
--headless \
--data-parallel-size 2 \
--data-parallel-size-local 1 \
--data-parallel-start-rank 1 \
--data-parallel-address $node0_ip \
--data-parallel-rpc-port 13389 \
--tensor-parallel-size 16 \
--seed 1024 \
--served-model-name deepseek_v3.2 \
--max-num-seqs 16 \
--max-model-len 17450 \
--max-num-batched-tokens 17450 \
--enable-expert-parallel \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--additional-config '{"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'

:::: ::::{tab-item} DeepSeek-V3.2-Exp-W8A8 A2 series :sync: A2

Run the following scripts on two nodes respectively.

Node 0

#!/bin/sh

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

export VLLM_USE_MODELSCOPE=True
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=100
export HCCL_BUFFSIZE=1024
export HCCL_OP_EXPANSION_MODE="AIV"
export PYTORCH_NPU_ALLOC_CONF="expandable_segments:True"

vllm serve vllm-ascend/DeepSeek-V3.2-Exp-W8A8 \
--host 0.0.0.0 \
--port 8000 \
--data-parallel-size 2 \
--data-parallel-size-local 1 \
--data-parallel-address $local_ip \
--data-parallel-rpc-port 13389 \
--tensor-parallel-size 8 \
--seed 1024 \
--served-model-name deepseek_v3.2 \
--enable-expert-parallel \
--max-num-seqs 16 \
--max-model-len 17450 \
--max-num-batched-tokens 17450 \
--trust-remote-code \
--quantization ascend \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.9 \
--additional-config '{"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'

Node 1

#!/bin/sh

# 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 VLLM_USE_MODELSCOPE=True
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=100
export HCCL_BUFFSIZE=1024
export HCCL_OP_EXPANSION_MODE="AIV"
export PYTORCH_NPU_ALLOC_CONF="expandable_segments:True"

vllm serve vllm-ascend/DeepSeek-V3.2-Exp-W8A8 \
--host 0.0.0.0 \
--port 8000 \
--headless \
--data-parallel-size 2 \
--data-parallel-size-local 1 \
--data-parallel-start-rank 1 \
--data-parallel-address $node0_ip \
--data-parallel-rpc-port 13389 \
--tensor-parallel-size 8 \
--seed 1024 \
--served-model-name deepseek_v3.2 \
--max-num-seqs 16 \
--max-model-len 17450 \
--max-num-batched-tokens 17450 \
--enable-expert-parallel \
--trust-remote-code \
--quantization ascend \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--additional-config '{"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]}}'

:::: :::::

Prefill-Decode Disaggregation

Not supported yet.

Functional Verification

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

curl http://<node0_ip>:<port>/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "deepseek_v3.2",
        "prompt": "The future of AI is",
        "max_tokens": 50,
        "temperature": 0
    }'

Accuracy Evaluation

Here are two accuracy evaluation methods.

Using AISBench

1. Refer to Using AISBench for details.

2. After execution, you can get the result, here is the result of DeepSeek-V3.2-Exp-W8A8 in vllm-ascend:0.11.0rc0 for reference only.

dataset	version	metric	mode	vllm-api-general-chat
cevaldataset	-	accuracy	gen	92.20

Using Language Model Evaluation Harness

As an example, take the gsm8k dataset as a test dataset, and run accuracy evaluation of DeepSeek-V3.2-Exp-W8A8 in online mode.

1. Refer to Using lm_eval for lm_eval installation.

2. Run lm_eval to execute the accuracy evaluation.

lm_eval \
  --model local-completions \
  --model_args model=/root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-Exp-W8A8,base_url=http://127.0.0.1:8000/v1/completions,tokenized_requests=False,trust_remote_code=True \
  --tasks gsm8k \
  --output_path ./

3. After execution, you can get the result, here is the result of DeepSeek-V3.2-Exp-W8A8 in vllm-ascend:0.11.0rc0 for reference only.

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.9591	±	0.0055
gsm8k	3	strict-match	5	exact_match	↑	0.9583	±	0.0055

Performance

Using AISBench

Refer to Using AISBench for performance evaluation for details.

Using vLLM Benchmark

Run performance evaluation of DeepSeek-V3.2-Exp-W8A8 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 vllm-ascend/DeepSeek-V3.2-Exp-W8A8  --dataset-name random --random-input 200 --num-prompt 200 --request-rate 1 --save-result --result-dir ./

After about several minutes, you can get the performance evaluation result.