Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.sglang.io/llms.txt

Use this file to discover all available pages before exploring further.

DeepSeek-V3.2 model family equips DeepSeek-V3.1-Terminus with DeepSeek Sparse Attention (DSA) through continued training. With DSA, a fine-grained sparse attention mechanism powered by a lightning indexer, DeepSeek-V3.2 achieves efficiency improvements in long-context scenarios. Note: This document is originally written for the usage of DeepSeek-V3.2-Exp model. The usage of DeepSeek-V3.2 or DeepSeek-V3.2-Speciale is the same as DeepSeek-V3.2-Exp except for the tool call parser. GLM-5 model also applies DSA (DeepSeek Sparse Attention) structure, so it can share most of the usage here, except for the reasoning parser and tool call parser.

​
Installation

​
Docker

Command
# H200/B200
docker pull lmsysorg/sglang:latest

# MI350/MI355
docker pull lmsysorg/sglang:v0.5.8-rocm700-mi35x

# MI300
# v0.5.8-rocm700-mi30x does not include PR #17504. Prefer the newest MI30x ROCm
# image tag from Docker Hub when available, or build from source (below).
docker pull lmsysorg/sglang:v0.5.8-rocm700-mi30x


# NPUs
docker pull lmsysorg/sglang:dsv32-a2
docker pull lmsysorg/sglang:dsv32-a3

​
Build From Source

Command
# Install SGLang
git clone https://github.com/sgl-project/sglang
cd sglang
pip3 install pip --upgrade
pip3 install -e "python"

​
Launch DeepSeek V3.2/GLM-5 with SGLang

To serve DeepSeek-V3.2-Exp on 8xH200/B200 GPUs:
Command
# Launch with TP + DP (Recommended)
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --enable-dp-attention

# Launch with EP + DP
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --ep 8 --dp 8 --enable-dp-attention

# Launch with Pure TP
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8

# Launch with TP on MI30x/MI35x
python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --nsa-prefill-backend tilelang --nsa-decode-backend tilelang
To serve GLM-5, just replace the --model argument with zai-org/GLM-5-FP8.

​
Configuration Tips

  • DP Attention: To enable DP Attention, please include --enable-dp-attention --dp <dp-size> in command. DP Attention is better for large concurrency scenarios.
  • TP Attention: Launching with TP attention is also supported. TP attention is better for low latency scenarios.
  • Short-sequence MHA prefill (adaptive): For short prefill sequences (default threshold: 2048 tokens), the NSA backend uses standard MHA automatically (no extra flags). On H200 (SM90) this path uses the FlashAttention variable-length kernel; on B200 (SM100) it uses TRT-LLM ragged MHA. MHA uses MHA_ONE_SHOT for best performance, which computes multi-head attention over all tokens (both cached prefix and newly extended tokens) in a single kernel invocation, avoiding the overhead of chunked KV cache processing. This achieves optimal throughput for short sequences where total sequence length fits within the chunk capacity limit.
  • MHA prefill threshold relaxation: To apply MHA attention to requests longer than 2048 tokens, please set the flag SGLANG_NSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD to a value larger than 2048. As threshold grows larger, the prefill performance can be improved, but at the cost of potential accuracy drop.
  • Choices of Attention Kernels: The attention backend is automatically set to nsa attention backend for DeepSeek V3.2 model. In this backend, different kernels for sparse prefilling/decoding are implemented, which can be specified by --nsa-prefill-backend and --nsa-decode-backend server arguments. The choices of nsa prefill/decode attention kernels include:
    • flashmla_sparse: flash_mla_sparse_fwd kernel from flash_mla library. Can run on both Hopper and Blackwell GPUs. It requires bf16 q, kv inputs.
    • flashmla_kv: flash_mla_with_kvcache kernel from flash_mla library. Can run on both Hopper and Blackwell GPUs. It requires bf16 q, fp8 k_cache inputs.
    • flashmla_auto: enables automatic selection of either flashmla_sparse or flashmla_kv kernel for prefill based on KV cache dtype, hardware, and heuristics. With BF16 KV cache, flashmla_sparse is always used on both Hopper and Blackwell. With FP8 KV cache: On Hopper (SM90), it unconditionally uses flashmla_kv; On Blackwell (SM100), it uses flashmla_sparse when total_kv_tokens < total_q_tokens * 512, otherwise falls back to flashmla_kv. The heuristics may need to be tuned if the performance of either kernel changes significantly.
    • fa3: flash_attn_with_kvcache kernel from flash_attn library. Can only run on Hopper GPUs. It requires bf16 q, kv inputs.
    • tilelang: tilelang implementation that can run on GPU, HPU and NPU.
    • aiter: Aiter kernel on AMD HPUs. Can only be used as decode kernel.
    • trtllm: trtllm-mla sparse kernel from flashinfer library. Only run on blackwell GPUs. It requires q,k,v to be uniformly bf16 or fp8_e4m3 format.
    • On the basis of performance benchmarks, the default configuration of DSA kernels on Hopper and Blackwell are set as follows :
      • Bfloat 16 kv cache: On Hopper, flashmla_sparse prefill attention, fa3 decode attention; On Blackwell, flashmla_sparse prefill attention, trtllm decode attention
      • Float8_e4m3fn KV cache: On Hopper, flashmla_kv prefill attention, flashmla_kv decode attention; On Blackwell, trtllm prefill attention and trtllm decode attention.
  • Index Cache: Introduce in this paper, IndexCache improves speed by reusing the result of indexer across different layers, only at cost of negligible accuracy loss. For GLM-5 model, we recommend appending --json-model-override-args '&#123;"index_topk_pattern": "FFSFSSSFSSFFFSSSFFFSFSSSSSSFFSFFSFFSSFFFFFFSFFFFFSFFSSSSSSFSFFFSFSSSFSFFSFFSSS"&#125;' to command for better tradeoff between speedup and performance.

​
Multi-token Prediction

SGLang implements Multi-Token Prediction (MTP) for DeepSeek V3.2 based on EAGLE speculative decoding. With this optimization, the decoding speed can be improved significantly on small batch sizes. Please look at this PR for more information. Example usage with DP Attention:
Command
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --enable-dp-attention --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
Example usage with Pure TP:
Command
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
  • 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.
  • The default value of --max-running-requests is set to 48 for MTP. For larger batch sizes, this value should be increased beyond the default value.
To enable overlap scheduler for EAGLE speculative decoding, we recommend setting the environment variable SGLANG_ENABLE_SPEC_V2=1. This can improve performance by enabling overlap scheduling between draft and verification stages.

​
Function Calling and Reasoning Parser

The usage of function calling and reasoning parser is the same as DeepSeek V3.1. Please refer to Reasoning Parser and Tool Parser documents. To launch DeepSeek-V3.2-Exp with function calling and reasoning parser:
Note: It is recommended to specify the chat-template, ensuring that you are within the sglang’s root directory.
Command
python3 -m sglang.launch_server \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --trust-remote-code \
  --tp-size 8 --dp-size 8 --enable-dp-attention \
  --tool-call-parser deepseekv31 \
  --reasoning-parser deepseek-v3 \
  --chat-template ./examples/chat_template/tool_chat_template_deepseekv32.jinja
To launch DeepSeek-V3.2 with function calling and reasoning parser:
Command
python3 -m sglang.launch_server \
  --model-path deepseek-ai/DeepSeek-V3.2 \
  --trust-remote-code \
  --tp-size 8 --dp-size 8 --enable-dp-attention \
  --tool-call-parser deepseekv32 \
  --reasoning-parser deepseek-v3
DeepSeek-V3.2-Speciale does not support tool calling, so it can only be launched with the reasoning parser:
Command
python3 -m sglang.launch_server \
  --model-path deepseek-ai/DeepSeek-V3.2-Speciale \
  --trust-remote-code \
  --tp-size 8 --dp-size 8 --enable-dp-attention \
  --reasoning-parser deepseek-v3
To launch GLM-5 with function calling and reasoning parser:
Command
python -m sglang.launch_server \
  --model zai-org/GLM-5-FP8 \
  --tp-size 8 --dp-size 8 --enable-dp-attention \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \

​
NVFP4 Checkpoint

To launch deepseek v3.2 NVFP4 checkpoint on Blackwell devices, the user needs to specify the quantization method as modelopt_fp4, and moe runner backend as one of flashinfer_trtllm(recommended), flashinfer_cutlass and flashinfer_cutedsl. Any other usage (parallelism, reasoning parser, …) is the same as FP8 checkpoint. An example launching command can be:
Command
python -m sglang.launch_server --model nvidia/DeepSeek-V3.2-NVFP4 --tp 4 --quantization modelopt_fp4 --moe-runner-backend flashinfer_trtllm --tool-call-parser deepseekv32  --reasoning-parser deepseek-v3

​
PD Disaggregation

Prefill Command:
Command
python -m sglang.launch_server \
        --model-path deepseek-ai/DeepSeek-V3.2-Exp \
        --disaggregation-mode prefill \
        --host $LOCAL_IP \
        --port $PORT \
        --tp 8 \
        --dp 8 \
        --enable-dp-attention \
        --dist-init-addr ${HOST}:${DIST_PORT} \
        --trust-remote-code \
        --disaggregation-bootstrap-port 8998 \
        --mem-fraction-static 0.9 \
Decode command:
Command
python -m sglang.launch_server \
        --model-path deepseek-ai/DeepSeek-V3.2-Exp \
        --disaggregation-mode decode \
        --host $LOCAL_IP \
        --port $PORT \
        --tp 8 \
        --dp 8 \
        --enable-dp-attention \
        --dist-init-addr ${HOST}:${DIST_PORT} \
        --trust-remote-code \
        --mem-fraction-static 0.9 \
Router command:
Command
python -m sglang_router.launch_router --pd-disaggregation \
  --prefill $PREFILL_ADDR 8998 \
  --decode $DECODE_ADDR \
  --host 127.0.0.1 \
  --port 8000 \
If you need more advanced deployment methods or production-ready deployment methods, such as RBG or LWS-based deployment, please refer to references/multi_node_deployment/rbg_pd/deepseekv32_pd.md. Additionally, you can also find startup commands for DeepEP-based EP parallelism in the aforementioned documentation.

​
Benchmarking Results

​
Accuracy Test with gsm8k

A simple accuracy benchmark can be tested with gsm8k dataset:
Command
python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1319 --parallel 1319
The result is 0.956, which matches our expectation:
Command
Accuracy: 0.956
Invalid: 0.000
Latency: 25.109 s
Output throughput: 5226.235 token/s
To test long-context accuracy, run gsm8k with --num-shots 20. The results are very close to the 8 shots results:
Output
Accuracy: 0.956
Invalid: 0.000
Latency: 29.545 s
Output throughput: 4418.617 token/s

​
Accuracy Test with gpqa-diamond

Accuracy benchmark on long context can be tested on GPQA-diamond dataset with long output tokens and thinking enabled:
Command
python3 -m sglang.test.run_eval --port 30000 --eval-name gpqa --num-examples 198 --max-tokens 128000 --repeat 8 --thinking-mode deepseek-v3
The mean accuracy over 8 runs shows 0.797, which matches the number 0.799 in official tech report.
Command
Repeat: 8, mean: 0.797
Scores: ['0.808', '0.798', '0.808', '0.798', '0.783', '0.788', '0.803', '0.793']
For DeepSeek V3.2, DeepSeek recommends setting the sampling parameters to temperature = 1.0, top_p = 0.95:
Command
python3 -m sglang.test.run_eval --port 30000 --eval-name gpqa --num-examples 198 --max-tokens 128000 --repeat 8 --top-p 0.95 --temperature 1.0 --thinking-mode deepseek-v3

Repeat: 8, mean: 0.840
Scores: ['0.848', '0.808', '0.848', '0.838', '0.879', '0.813', '0.838', '0.848']
which matches the official score, 0.824, as reported in the DeepSeek-V3.2 technical report.

​
Accuracy Test with aime 2025

Prepare the environment by installing NeMo-Skills in the docker or your own virtual environment: 
pip install git+https://github.com/NVIDIA/NeMo-Skills.git --ignore-installed blinker
Then launch the SGLang server:
Output
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --enable-dp-attention
For DeepSeek-V3.2 and DeepSeek-V3.2-Speciale:
Output
python3 -m sglang.launch_server   --model-path deepseek-ai/DeepSeek-V3.2   --trust-remote-code   --tp-size 8 --dp-size 8 --enable-dp-attention   --tool-call-parser deepseekv32   --reasoning-parser deepseek-v3
Run the following script to evaluate AIME 2025:
Output
#! /bin/bash
export NEMO_SKILLS_DISABLE_UNCOMMITTED_CHANGES_CHECK=1

ns prepare_data aime25

PORT=30000
BACKEND=sglang
MODEL="deepseek-ai/DeepSeek-V3.2-Exp" # Should be changed to the model name
MODEL_NAME="dsv32-fp8"

echo "Starting AIME25 evaluation with model $MODEL on port $PORT using backend $BACKEND..."
ns eval \
  --benchmarks=aime25:4 \
  --server_type=$BACKEND \
  --model=$MODEL \
  --server_address=http://localhost:${PORT}/v1 \
  --output_dir=nemo_skills_aime25_${MODEL_NAME}_output_${BACKEND}_$(date +%Y%m%d_%H%M%S) \
  ++chat_template_kwargs.thinking=true \
  ++inference.temperature=1.0 \
  ++inference.top_p=0.95 \
  ++inference.tokens_to_generate=64000
  # ++inference.tokens_to_generate=120000 for Speciale model
Test results (8*B200): DeepSeek-V3.2-Exp:
evaluation_modenum_entriesavg_tokensgen_secondssymbolic_correctno_answer
pass@1[avg-of-4]3015040167387.50% ± 1.67%0.00%
majority@43015040167390.00%0.00%
pass@43015040167390.00%0.00%
DeepSeek-V3.2:
evaluation_modenum_entriesavg_tokensgen_secondssymbolic_correctno_answer
pass@1[avg-of-4]3013550163292.50% ± 1.67%0.00%
majority@43013550163294.71%0.00%
pass@43013550163296.67%0.00%
DeepSeek-V3.2-Speciale:
evaluation_modenum_entriesavg_tokensgen_secondssymbolic_correctno_answer
pass@1[avg-of-4]3024155358395.00% ± 1.92%0.00%
majority@43024155358395.83%0.00%
pass@430241553583100.00%0.00%

​
DSA long sequence context parallel optimization(experimental)

Note: This feature is only verified on Hopper machines For context parallel in DeepSeek V3.2 model, we provide two different modes of splitting tokens, which can be controlled with argument --nsa-prefill-cp-mode.

​
In sequence splitting

The first mode can be enabled by --nsa-prefill-cp-mode in-seq-split. This mode implements context parallel for DSA by splitting the sequence uniformly between context parallel ranks. At attention stage, each cp rank computes the indexer results of sharded sequence, and collects the whole kv cache through all gather operator. Add attn_cp_size for communication group for context parallel. Note that the in-sequence splitting mode has the following restrictions:
  • The batch size is restricted to 1 for prefill batches
  • moe_dense_tp_size=1, moe_a2a_backend = "deepep"
  • To ensure cp_size > 1, the passed in tp_size must be larger than dp_size
For more details, please refer to PR https://github.com/sgl-project/sglang/pull/12065. Example:
Command
# In-seq splitting mode launched with EP + DP
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp  --tp 8 --ep 8 --dp 2 --enable-dp-attention --enable-nsa-prefill-context-parallel --attn-cp-size 4 --nsa-prefill-cp-mode in-seq-split --max-running-requests 32

​
Round robin splitting (default setting)

This mode can be enabled by specifying the parameter --nsa-prefill-cp-mode round-robin-split, which distributes tokens across ranks based on token_idx % cp_size. In this scenario, compared to the in-sequence splitting method, it additionally supports the fused MoE backend (the fused MoE backend may deliver better performance than DeepEP in single-machine scenarios), FP8 KV-cache, and multi-batch prefill inference. However, it cannot be enabled with DP attention together. For more details, please refer to PR https://github.com/sgl-project/sglang/pull/13959. Example usage:
Command
# Launch with FusedMoe + CP8
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp  --tp 8 --enable-nsa-prefill-context-parallel  --attn-cp-size 8 --nsa-prefill-cp-mode round-robin-split --max-running-requests 32

​
Pipeline Parallel + Context Parallel (PP + CP)

This mode combines Pipeline Parallelism (PP) and Context Parallelism (CP) to scale across multiple nodes, which can achieve better throughput and Time To First Token (TTFT). Note that this method has only been tested on H20 96G.

​
Standard Usage

To launch with PP=2 and CP (via round-robin-split mode) on 2 nodes. This configuration uses the fused MoE kernel by default, which generally provides better performance. For related development details, please refer to: Node 0:
Command
export SGLANG_PP_LAYER_PARTITION=30,31
python3 -m sglang.launch_server \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --nnodes 2 --node-rank 0 \
  --dist-init-addr <HEAD_NODE_IP>:62001 \
  --tp 8 --pp-size 2 \
  --dp-size 1 --moe-dense-tp-size 1 \
  --enable-nsa-prefill-context-parallel \
  --attn-cp-size 8 \
  --nsa-prefill-cp-mode round-robin-split \
  --trust-remote-code \
  --disable-radix-cache \
  --mem-fraction-static 0.8 \
  --max-running-requests 128 \
  --chunked-prefill-size 16384 \
  --cuda-graph-max-bs 8 \
  --page-size 64 \
  --watchdog-timeout 3600 \
  --host 0.0.0.0 --port 8000 \
  --tool-call-parser deepseekv32
Node 1:
Command
export SGLANG_PP_LAYER_PARTITION=30,31
python3 -m sglang.launch_server \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --nnodes 2 --node-rank 1 \
  --dist-init-addr <HEAD_NODE_IP>:62001 \
  --tp 8 --pp-size 2 \
  --dp-size 1 --moe-dense-tp-size 1 \
  --enable-nsa-prefill-context-parallel \
  --attn-cp-size 8 \
  --nsa-prefill-cp-mode round-robin-split \
  --trust-remote-code \
  --disable-radix-cache \
  --mem-fraction-static 0.8 \
  --max-running-requests 128 \
  --chunked-prefill-size 16384 \
  --cuda-graph-max-bs 8 \
  --page-size 64 \
  --watchdog-timeout 3600 \
  --host 0.0.0.0 --port 8000 \
  --tool-call-parser deepseekv32

​
PD Disaggregation with PP + CP

If using PD (Prefill-Decode) Disaggregation, the Prefill nodes can be configured with PP + CP as follows. Prefill Node 0:
Command
python -m sglang.launch_server \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --served-model-name deepseek-v32 \
  --nnodes 2 --node-rank 0 \
  --dist-init-addr <PREFILL_HEAD_IP>:20102 \
  --tp 8 --pp-size 2 \
  --dp-size 1 --moe-dense-tp-size 1 \
  --enable-nsa-prefill-context-parallel \
  --attn-cp-size 8 \
  --nsa-prefill-cp-mode round-robin-split  \
  --disaggregation-ib-device mlx5_bond_0,mlx5_bond_1,mlx5_bond_2,mlx5_bond_3 \
  --trust-remote-code \
  --disable-radix-cache \
  --max-running-requests 512 \
  --chunked-prefill-size 4096 \
  --context-length 131072 \
  --mem-fraction-static 0.9 \
  --page-size 64 \
  --enable-metrics \
  --collect-tokens-histogram \
  --tokenizer-worker-num 8 \
  --host 0.0.0.0 --port 30000
Prefill Node 1:
Command
python -m sglang.launch_server \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --served-model-name deepseek-v32-prefill \
  --nnodes 2 --node-rank 1 \
  --dist-init-addr <PREFILL_HEAD_IP>:20102 \
  --tp 8 --pp-size 2 \
  --dp-size 1 --moe-dense-tp-size 1 \
  --enable-nsa-prefill-context-parallel \
  --attn-cp-size 8 \
  --nsa-prefill-cp-mode round-robin-split  \
  --disaggregation-ib-device mlx5_bond_0,mlx5_bond_1,mlx5_bond_2,mlx5_bond_3 \
  --trust-remote-code \
  --disable-radix-cache \
  --max-running-requests 512 \
  --chunked-prefill-size 4096 \
  --context-length 131072 \
  --mem-fraction-static 0.9 \
  --page-size 64 \
  --enable-metrics \
  --collect-tokens-histogram \
  --tokenizer-worker-num 8 \
  --host 0.0.0.0 --port 30000
For the Decode nodes, it is recommended to use the EP mode.

​
HiSparse: Hierarchical Sparse Attention for DSA (experimental)

HiSparse reduces per-request GPU memory during decode by keeping only a small “hot” KV buffer on GPU while storing complete KV data in CPU pinned memory. A CUDA kernel dynamically swaps in the top-k most relevant KV entries from host memory on each decode step. This enables significantly higher decode concurrency for long-context DSA models. HiSparse currently requires PD disaggregation mode and is enabled on the decode instance only. For detailed design, configuration, and deployment instructions, see the HiSparse Guide.