DeepSeek-V3 is a strong Mixture-of-Experts (MoE) language model similar to OpenAI ChatGPT.
It supports 671B total parameters with 37B activated for each token.
To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2.
Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance.
It was pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities.
Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models.
Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training.
In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks.
DeepSeek-V3 achieves a significant breakthrough in inference speed over previous models.
It tops the leaderboard among open-source models and rivals the most advanced closed-source models globally.
Architecture: Innovative Load Balancing Strategy and Training Objective
On top of the efficient architecture of DeepSeek-V2, we pioneer an auxiliary-loss-free strategy for load balancing, which minimizes the performance degradation that arises from encouraging load balancing.
We investigate a Multi-Token Prediction (MTP) objective and prove it beneficial to model performance. It can also be used for speculative decoding for inference acceleration.
Pre-Training: Towards Ultimate Training Efficiency
We design an FP8 mixed precision training framework and, for the first time, validate the feasibility and effectiveness of FP8 training on an extremely large-scale model.
Through co-design of algorithms, frameworks, and hardware, we overcome the communication bottleneck in cross-node MoE training, nearly achieving full computation-communication overlap.
This significantly enhances our training efficiency and reduces the training costs, enabling us to further scale up the model size without additional overhead.
At an economical cost of only 2.664M H800 GPU hours, we complete the pre-training of DeepSeek-V3 on 14.8T tokens, producing the currently strongest open-source base model. The subsequent training stages after pre-training require only 0.1M GPU hours.
Post-Training: Knowledge Distillation from DeepSeek-R1
They introduced an innovative methodology to distill reasoning capabilities from the long-Chain-of-Thought (CoT) model, specifically from one of the DeepSeek R1 series models, into standard LLMs, particularly DeepSeek-V3.
Their pipeline elegantly incorporates the verification and reflection patterns of R1 into DeepSeek-V3 and notably improves its reasoning performance. Meanwhile, we also maintain a control over the output style and length of DeepSeek-V3.
How to Run Locally
DeepSeek-V3 can be deployed locally using the following hardware and open-source community software:
- DeepSeek-Infer Demo: We provide a simple and lightweight demo for FP8 and BF16 inference.
- SGLang: Fully support the DeepSeek-V3 model in both BF16 and FP8 inference modes, with Multi-Token Prediction coming soon.
- LMDeploy: Enables efficient FP8 and BF16 inference for local and cloud deployment.
- TensorRT-LLM: Currently supports BF16 inference and INT4/8 quantization, with FP8 support coming soon.
- vLLM: Support DeepSeek-V3 model with FP8 and BF16 modes for tensor parallelism and pipeline parallelism.
- AMD GPU: Enables running the DeepSeek-V3 model on AMD GPUs via SGLang in both BF16 and FP8 modes.
- Huawei Ascend NPU: Supports running DeepSeek-V3 on Huawei Ascend devices.
Since FP8 training is natively adopted in our framework, we only provide FP8 weights. If you require BF16 weights for experimentation, you can use the provided conversion script to perform the transformation.
Here is an example of converting FP8 weights to BF16:
cd inference
python fp8_cast_bf16.py --input-fp8-hf-path /path/to/fp8_weights --output-bf16-hf-path /path/to/bf16_weights
Inference with DeepSeek-Infer Demo (example only)
Model Weights & Demo Code Preparation
First, clone our DeepSeek-V3 GitHub repository:
git clone https://github.com/deepseek-ai/DeepSeek-V3.git
Navigate to the inference folder and install dependencies listed in requirements.txt.
cd DeepSeek-V3/inference pip install -r requirements.txt
Download the model weights from Hugging Face, and put them into /path/to/DeepSeek-V3 folder.
Model Weights Conversion
Convert Hugging Face model weights to a specific format:
python convert.py --hf-ckpt-path /path/to/DeepSeek-V3 --save-path /path/to/DeepSeek-V3-Demo --n-experts 256 --model-parallel 16
Run
Then you can chat with DeepSeek-V3:
torchrun --nnodes 2 --nproc-per-node 8 generate.py --node-rank $RANK --master-addr $ADDR --ckpt-path /path/to/DeepSeek-V3-Demo --config configs/config_671B.json --interactive --temperature 0.7 --max-new-tokens 200
Or batch inference on a given file:
torchrun --nnodes 2 --nproc-per-node 8 generate.py --node-rank $RANK --master-addr $ADDR --ckpt-path /path/to/DeepSeek-V3-Demo --config configs/config_671B.json --input-file $FILE
6.2 Inference with SGLang (recommended)
SGLang currently supports MLA optimizations, DP Attention, FP8 (W8A8), FP8 KV Cache, and Torch Compile, delivering state-of-the-art latency and throughput performance among open-source frameworks.
Notably, SGLang v0.4.1 fully supports running DeepSeek-V3 on both NVIDIA and AMD GPUs, making it a highly versatile and robust solution.
SGLang also supports multi-node tensor parallelism, enabling you to run this model on multiple network-connected machines.
Multi-Token Prediction (MTP) is in development, and progress can be tracked in the optimization plan.
Here are the launch instructions from the SGLang team: https://github.com/sgl-project/sglang/tree/main/benchmark/deepseek_v3
6.3 Inference with LMDeploy (recommended)
LMDeploy, a flexible and high-performance inference and serving framework tailored for large language models, now supports DeepSeek-V3. It offers both offline pipeline processing and online deployment capabilities, seamlessly integrating with PyTorch-based workflows.
For comprehensive step-by-step instructions on running DeepSeek-V3 with LMDeploy, please refer to here: InternLM/lmdeploy#2960
6.4 Inference with TRT-LLM (recommended)
TensorRT-LLM now supports the DeepSeek-V3 model, offering precision options such as BF16 and INT4/INT8 weight-only. Support for FP8 is currently in progress and will be released soon. You can access the custom branch of TRTLLM specifically for DeepSeek-V3 support through the following link to experience the new features directly: https://github.com/NVIDIA/TensorRT-LLM/tree/deepseek/examples/deepseek_v3.
6.5 Inference with vLLM (recommended)
vLLM v0.6.6 supports DeepSeek-V3 inference for FP8 and BF16 modes on both NVIDIA and AMD GPUs. Aside from standard techniques, vLLM offers pipeline parallelism allowing you to run this model on multiple machines connected by networks. For detailed guidance, please refer to the vLLM instructions. Please feel free to follow the enhancement plan as well.
6.6 Recommended Inference Functionality with AMD GPUs
In collaboration with the AMD team, we have achieved Day-One support for AMD GPUs using SGLang, with full compatibility for both FP8 and BF16 precision. For detailed guidance, please refer to the SGLang instructions.
6.7 Recommended Inference Functionality with Huawei Ascend NPUs
The MindIE framework from the Huawei Ascend community has successfully adapted the BF16 version of DeepSeek-V3. For step-by-step guidance on Ascend NPUs, please follow the instructions here.
7. License
This code repository is licensed under the MIT License. The use of DeepSeek-V3 Base/Chat models is subject to the Model License. DeepSeek-V3 series (including Base and Chat) supports commercial use.
Contact – [email protected]