DeepSeek Usage

SGLang provides several optimizations specifically designed for the DeepSeek model to boost its inference speed. This document outlines current optimizations for DeepSeek. Additionally, the SGLang team is actively developing enhancements for DeepSeek V3.

Launch DeepSeek V3 with SGLang

SGLang is recognized as one of the top engines for DeepSeek model inference. To run DeepSeek V3/R1 models, the requirements are as follows:

Weight Type	Configuration
Full precision FP8 (recommended)	8 x H200
	8 x MI300X
	2 x 8 x H100/800/20
Full precision BF16	2 x 8 x H200
	2 x 8 x MI300X
	4 x 8 x H100/800/20
	4 x 8 x A100/A800
Quantized weights (AWQ)	8 x H100/800/20
	8 x A100/A800
Quantized weights (int8)	16 x A100/800
	32 x L40S

Detailed commands for reference:

• 8 x H200

• 8 x MI300X

• 2 x 8 x H200

• 4 x 8 x A100

• 8 x A100 (AWQ)

• 16 x A100 (int8)

• 32 x L40S (int8)

Download Weights

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. Please refer to DeepSeek V3 official guide to download the weights.

Caching torch.compile

The DeepSeek series have huge model weights, it takes some time to compile the model with torch.compile for the first time if you have added the flag --enable-torch-compile. You can refer here to optimize the caching of compilation results, so that the cache can be used to speed up the next startup.

Launch with One node of 8 H200

Please refer to the example. Note that Deepseek V3 is already in FP8. So we should not run it with any quantization arguments like --quantization fp8 --kv-cache-dtype fp8_e5m2. Also, --enable-dp-attention can be useful to improve for Deepseek V3/R1’s throughput. Please refer to Data Parallelism Attention for detail.

Running examples on Multi-node

• Serving with two H20*8 nodes.

• Serving with two H200*8 nodes and docker.

• Serving with four A100*8 nodes.

Optimizations

Multi-head Latent Attention (MLA) Throughput Optimizations

Description: MLA is an innovative attention mechanism introduced by the DeepSeek team, aimed at improving inference efficiency. SGLang has implemented specific optimizations for this, including:

• Weight Absorption: By applying the associative law of matrix multiplication to reorder computation steps, this method balances computation and memory access and improves efficiency in the decoding phase.

• MLA Attention Backends: Currently SGLang supports different optimized MLA attention backends, including FlashAttention3, Flashinfer, and Triton backends. It can be set with --attention-backend argument.

• FP8 Quantization: W8A8 FP8 and KV Cache FP8 quantization enables efficient FP8 inference. Additionally, we have implemented Batched Matrix Multiplication (BMM) operator to facilitate FP8 inference in MLA with weight absorption.

• CUDA Graph & Torch.compile: Both MLA and Mixture of Experts (MoE) are compatible with CUDA Graph and Torch.compile, which reduces latency and accelerates decoding speed for small batch sizes.

• Chunked Prefix Cache: Chunked prefix cache optimization can increase throughput by cutting prefix cache into chunks, processing them with multi-head attention and merging their states. Its improvement can be significant when doing chunked prefill on long sequences. Currently this optimization is only available for FlashAttention3 backend.

Overall, with these optimizations, we have achieved up to 7x acceleration in output throughput compared to the previous version.

Usage: MLA optimization is enabled by default. To disable chunked prefix cache feature for mla, use disable-chunked-prefix-cache.

Reference: Check Blog and Slides for more details.

Data Parallelism Attention

Description: This optimization involves data parallelism (DP) for the MLA attention mechanism of DeepSeek Series Models, which allows for a significant reduction in the KV cache size, enabling larger batch sizes. Each DP worker independently handles different types of batches (prefill, decode, idle), which are then synchronized before and after processing through the Mixture-of-Experts (MoE) layer.

With data parallelism attention enabled, we have achieved up to 1.9x decoding throughput improvement compared to the previous version.

Usage:

• This optimization is aimed at improving throughput and should be used for scenarios with high QPS (Queries Per Second). It can be enabled by --enable-dp-attention for DeepSeek models.

• Since v0.4.4, DP and TP attention can be flexibly combined. For example, to deploy DeepSeek-V3/R1 on 2 node with 8*H100, you can specify --tp 16 and --dp 2, which means for attention part there are 2 DP groups, and in each DP group there are 8 TP groups.

Reference: Check Blog.

Multi Node Tensor Parallelism

Description: For users with limited memory on a single node, SGLang supports serving DeepSeek Series Models, including DeepSeek V3, across multiple nodes using tensor parallelism. This approach partitions the model parameters across multiple GPUs or nodes to handle models that are too large for one node’s memory.

Usage: Check here for usage examples.

Block-wise FP8

Description: SGLang implements block-wise FP8 quantization with two key optimizations:

• Activation: E4M3 format using per-token-per-128-channel sub-vector scales with online casting.

• Weight: Per-128x128-block quantization for better numerical stability.

• DeepGEMM: The DeepGEMM kernel library optimized for FP8 matrix multiplications.

Usage: The activation and weight optimization above are turned on by default for DeepSeek V3 models. DeepGEMM is enabled by default on NVIDIA Hopper GPUs and disabled by default on other devices. DeepGEMM can also be manually turned off by setting the environment variable SGL_ENABLE_JIT_DEEPGEMM=0.

Before serving the DeepSeek model, precompile the DeepGEMM kernels using:

python3 -m sglang.compile_deep_gemm --model deepseek-ai/DeepSeek-V3 --tp 8 --trust-remote-code

The precompilation process typically takes around 10 minutes to complete.

Multi-token Prediction

Description: SGLang implements DeepSeek V3 Multi-Token Prediction (MTP) based on EAGLE speculative decoding. With this optimization, the decoding speed can be improved by 1.8x for batch size 1 and 1.5x for batch size 32 respectively on H200 TP8 setting.

Usage: Add arguments --speculative-algorithm, --speculative-draft-model-path, --speculative-num-steps, --speculative-eagle-topk and --speculative-num-draft-tokens to enable this feature. For example:

python3 -m sglang.launch_server --model-path deepseek-ai/DeepSeek-V3-0324 --speculative-algorithm EAGLE --speculative-draft-model-path lmsys/DeepSeek-V3-0324-NextN --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2 --trust-remote-code --tp 8

• The draft model are available at huggingface: lmsys/DeepSeek-V3-0324-NextN, lmsys/DeepSeek-R1-NextN. It can also be exported from original DeepSeek-V3/R1 model with export_deepseek_nextn.py script.

• The best configuration for --speculative-num-steps, --speculative-eagle-topk and --speculative-num-draft-tokens can be searched with bench_speculative.py script for given batch size. The minimum configuration is --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2, which can achieve speedup for larger batch sizes. When using FlashInfer MLA wrapper (--attention-backend flashinfer) with speculative decoding, set the --speculative-eagle-topk parameter to 1. The FlashAttention 3 backend also only supports --speculative-eagle-topk 1.

• To enable DeepSeek MTP for large batch sizes (>32), there are some parameters should be changed (Reference this discussion):

  • Adjust --max-running-requests to a larger number. The default value is 32 for MTP. For larger batch sizes, you should increase this value beyond the default value.

  • Set --cuda-graph-bs. It’s a list of batch sizes for cuda graph capture. The default captured batch sizes for speculative decoding is set here. You can include more batch sizes into it.

Reasoning Content for DeepSeek R1

See Separate Reasoning.

Function calling for DeepSeek Models

Add arguments --tool-call-parser deepseekv3 to enable this feature. For example (running on 1 * H20 node):

python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-V3-0324 --tp 8 --port 30000 --host 0.0.0.0 --mem-fraction-static 0.9 --disable-cuda-graph --tool-call-parser deepseekv3

Sample Request:

curl "http://127.0.0.1:30000/v1/chat/completions" \
-H "Content-Type: application/json" \
-d '{"temperature": 0, "max_tokens": 100, "model": "deepseek-ai/DeepSeek-V3-0324", "tools": [{"type": "function", "function": {"name": "query_weather", "description": "Get weather of an city, the user should supply a city first", "parameters": {"type": "object", "properties": {"city": {"type": "string", "description": "The city, e.g. Beijing"}}, "required": ["city"]}}}], "messages": [{"role": "user", "content": "Hows the weather like in Qingdao today"}]}'

Expected Response

{"id": "62af80528930423a82c806651ec66e7c", "object": "chat.completion", "created": 1744431333, "model": "deepseek-ai/DeepSeek-V3-0324", "choices": [{"index": 0, "message": {"role": "assistant", "content": null, "reasoning_content": null, "tool_calls": [{"id": "0", "type": "function", "function": {"name": "query_weather", "arguments": "{\\"city\\": \\"Guangzhou\\"}"}}]}, "logprobs": null, "finish_reason": "tool_calls", "matched_stop": null}], "usage": {"prompt_tokens": 118, "total_tokens": 140, "completion_tokens": 22, "prompt_tokens_details": null}}

Important Notes:

1. Use a lower "temperature" value for better results.

2. Currently, the function calling implementation for deepseek is incompatible with streaming requests.

FAQ

1. Question: What should I do if model loading takes too long and NCCL timeout occurs?

   Answer: You can try to add --dist-timeout 3600 when launching the model, this allows for 1-hour timeout.