The SGLang and DeepSeek teams collaborated to get DeepSeek V3 FP8 running on NVIDIA and AMD GPUs **from day one**. SGLang also supports [MLA optimization](https://lmsys.org/blog/2024-09-04-sglang-v0-3/#deepseek-multi-head-latent-attention-mla-throughput-optimizations) and [DP attention](https://lmsys.org/blog/2024-12-04-sglang-v0-4/#data-parallelism-attention-for-deepseek-models), making SGLang one of the best open-source LLM engines for running DeepSeek models. SGLang is the inference engine recommended by the official [DeepSeek team](https://github.com/deepseek-ai/DeepSeek-V3/tree/main?tab=readme-ov-file#62-inference-with-sglang-recommended).
Special thanks to Meituan's Search & Recommend Platform Team and Baseten's Model Performance Team for implementing the model, and DataCrunch for providing GPU resources.
For optimizations made on the DeepSeek series models regarding SGLang, please refer to [DeepSeek Model Optimizations in SGLang](https://docs.sglang.ai/references/deepseek.html).
If you encounter errors when starting the server, ensure the weights have finished downloading. It's recommended to download them beforehand or restart multiple times until all weights are downloaded.
[MLA optimizations](https://lmsys.org/blog/2024-09-04-sglang-v0-3/#deepseek-multi-head-latent-attention-mla-throughput-optimizations) are enabled by default. Here are some optional optimizations can be enabled as needed.
- [Data Parallelism Attention](https://lmsys.org/blog/2024-12-04-sglang-v0-4/#data-parallelism-attention-for-deepseek-models): For high QPS scenarios, add the `--enable-dp-attention` argument to boost throughput.
- [Torch.compile Optimization](https://lmsys.org/blog/2024-09-04-sglang-v0-3/#torchcompile-latency-optimizations): Add `--enable-torch-compile` argument to enable it. This will take some time while server starts. The maximum batch size for torch.compile optimization can be controlled with `--torch-compile-max-bs`. It's recommended to set it between `1` and `8`. (e.g., `--torch-compile-max-bs 8`)
For example, there are two H20 nodes, each with 8 GPUs. The first node's IP is `10.0.0.1`, and the second node's IP is `10.0.0.2`. Please **use the first node's IP** for both commands.
If the command fails, try setting the `GLOO_SOCKET_IFNAME` parameter. For more information, see [Common Environment Variables](https://pytorch.org/docs/stable/distributed.html#common-environment-variables).
If the multi nodes support NVIDIA InfiniBand and encounter hanging issues during startup, consider adding the parameter `export NCCL_IB_GID_INDEX=3`. For more information, see [this](https://github.com/sgl-project/sglang/issues/3516#issuecomment-2668493307).
> **Note that the launch command here does not enable Data Parallelism Attention or `torch.compile` Optimization**. For optimal performance, please refer to the command options in [Performance Optimization Options](#option_args).
There are two H200 nodes, each with 8 GPUs. The first node's IP is `192.168.114.10`, and the second node's IP is `192.168.114.11`. Configure the endpoint to expose it to another Docker container using `--host 0.0.0.0` and `--port 40000`, and set up communications with `--dist-init-addr 192.168.114.10:20000`.
A single H200 with 8 devices can run DeepSeek V3, the dual H200 setup is just to demonstrate multi-node usage.
> **Note that the launch command here does not enable Data Parallelism Attention or `torch.compile` Optimization**. For optimal performance, please refer to the command options in [Performance Optimization Options](#option_args).
To serve DeepSeek-V3 with A100 GPUs, we need to convert the [FP8 model checkpoints](https://huggingface.co/deepseek-ai/DeepSeek-V3) to BF16 with [script](https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/fp8_cast_bf16.py) mentioned [here](https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/fp8_cast_bf16.py) first.
Since the BF16 model is over 1.3 TB, we need to prepare four A100 nodes, each with 8 80GB GPUs. Assume the first node's IP is `10.0.0.1`, and the converted model path is `/path/to/DeepSeek-V3-BF16`, we can have following commands to launch the server.
> **Note that the launch command here does not enable Data Parallelism Attention or `torch.compile` Optimization**. For optimal performance, please refer to the command options in [Performance Optimization Options](#option_args).
There are block-wise and per-channel quantization methods, and the quantization parameters have already been uploaded to Huggingface. One example is as follows:
Assuming that master node IP is `MASTER_IP`, checkpoint path is `/path/to/DeepSeek-R1-INT8` and port=5000, we can have following commands to launch the server:
> **Note that the launch command here enables `torch.compile` Optimization**. For optimal performance, please refer to the command options in [Performance Optimization Options](#option_args).
Then on the **master node**, supposing the ShareGPT data is located at `/path/to/ShareGPT_V3_unfiltered_cleaned_split.json`, you can run the following commands to benchmark the launched server:
Assuming that master node IP is `MASTER_IP`, checkpoint path is `/path/to/DeepSeek-R1-Channel-INT8` and port=5000, we can have following commands to launch the server:
The benchmarking method is the same as describted in the previous [16 x A100](https://github.com/sgl-project/sglang/tree/main/benchmark/deepseek_v3#example-serving-with-16-a100a800-with-int8-quantization) example.
### Example: Serving on any cloud or Kubernetes with SkyPilot
SkyPilot helps find cheapest available GPUs across any cloud or existing Kubernetes clusters and launch distributed serving with a single command. See details [here](https://github.com/skypilot-org/skypilot/tree/master/llm/deepseek-r1).
ValueError: Weight output_partition_size = 576 is not divisible by weight quantization block_n = 128.
```
edit your `config.json` and remove the `quantization_config` block. For example:
```json
"quantization_config": {
"activation_scheme": "dynamic",
"fmt": "e4m3",
"quant_method": "fp8",
"weight_block_size": [128, 128]
},
```
Removing this block typically resolves the error. For more details, see the discussion in [sgl-project/sglang#3491](https://github.com/sgl-project/sglang/issues/3491#issuecomment-2650779851).
