Skip to main content

​
1. Model Introduction

The DeepSeek-V3.2 series includes three model variants, each optimized for different use cases: DeepSeek-V3.2-Exp is an upgraded version of DeepSeek-V3.1-Terminus, introducing the DeepSeek Sparse Attention (DSA) mechanism through continued training. DSA is a fine-grained sparse attention mechanism powered by a lightning indexer, enabling DeepSeek-V3.2-Exp to achieve significant efficiency improvements in long-context scenarios. Recommended for general conversations, long-context processing, and efficient inference. DeepSeek-V3.2 is the standard version suitable for general tasks and conversational scenarios. For local deployment, we recommend setting the sampling parameters to temperature = 1.0, top_p = 0.95. Recommended for standard conversations and general tasks. DeepSeek-V3.2-Speciale is a special variant designed exclusively for deep reasoning tasks. This model is specifically optimized for scenarios requiring complex logical reasoning and deep thinking. However this model does not support tool calls (see below). For local deployment, we recommend setting the sampling parameters to temperature = 1.0, top_p = 0.95. Recommended for deep reasoning tasks, complex logical problems, and mathematical reasoning. DeepSeek-V3.2-NVFP4 is an NVIDIA-optimized NVFP4-quantized variant of DeepSeek-V3.2 for Blackwell devices. It uses ModelOpt FP4 quantization with a choice of MoE runner backends (flashinfer_trtllm (recommended), flashinfer_cutlass, or flashinfer_cutedsl), enabling efficient deployment with lower tensor parallelism (TP=4). It supports the same features as DeepSeek-V3.2 including tool calling, reasoning, and speculative decoding (MTP). DeepSeek-V3.2-MXFP4 is an OCP-MXFP4 optimized variant for DeepSeek-V3.2 for AMD MI300X/MI355X devices. It uses OCP MXFP4 quantization with a triton mxfp4 backend (the same backend for gptoss-120B), enabling efficient deployment with lower tensor parallelism (TP=8) in a single node. It includes the same features as DeepSeek-V3.2 including tool calling, reasoning, fp8-kv, CP, TP and speculative decoding MTP.

​
2. SGLang Installation

SGLang offers multiple installation methods. You can choose the most suitable installation method based on your hardware platform and requirements. Please refer to the official SGLang installation guide for installation instructions.

​
2.1 Docker Images

Pre-built Docker images are available for different hardware platforms:
Command
# NVIDIA H200 / B200
docker pull lmsysorg/sglang:latest

# AMD MI350 / MI355X
docker pull lmsysorg/sglang:v0.5.8-rocm700-mi35x

# AMD MI300X
# Note: 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.
docker pull lmsysorg/sglang:v0.5.8-rocm700-mi30x

# Ascend NPU (Atlas 800I A2 / A3)
docker pull lmsysorg/sglang:dsv32-a2
docker pull lmsysorg/sglang:dsv32-a3

​
3. Model Deployment

This section provides a progressive guide from quick deployment to performance optimization, suitable for users at different levels.

​
3.1 Basic Configuration

Interactive Command Generator: Use the configuration selector below to automatically generate the appropriate deployment command for your hardware platform, model variant, deployment strategy, and thinking capabilities. SGLang supports serving DeepSeek V3.2 on NVIDIA H200, B200, and AMD MI300X/MI355X GPUs.

​
3.2 Configuration Tips

  • Short-sequence MHA prefill (adaptive): For prefill sequences shorter than 2048 tokens (default threshold), the DSA backend automatically switches to standard MHA (using FlashAttention variable-length on SM90, TRT-LLM ragged MHA on SM100). To extend this to longer sequences set env var SGLANG_DSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD to a larger value (potential minor accuracy trade-off).
  • DSA prefill/decode attention kernels (--dsa-prefill-backend, --dsa-decode-backend): The dsa backend is automatically selected for DeepSeek-V3.2. Available kernels: flashmla_sparse, flashmla_kv, flashmla_auto, fa3 (Hopper only), tilelang (GPU/HPU/NPU), aiter (AMD, decode only), trtllm (Blackwell only). Defaults: Hopper BF16 KV → flashmla_sparse prefill / fa3 decode; Hopper FP8 KV → flashmla_kv both; Blackwell BF16 → flashmla_sparse / trtllm; Blackwell FP8 → trtllm both.
  • Index Cache: Reuses indexer results across layers for efficiency at negligible accuracy cost. For GLM-5 specifically, append --json-model-override-args '{"index_topk_pattern": "FFSFSSSFSSFFFSSSFFFSFSSSSSSFFSFFSFFSSFFFFFFSFFFFFSFFSSSSSSFSFFFSFSSSFSFFSFFSSS"}' for a better speed/accuracy tradeoff.
  • HiSparse (experimental): Reduces per-request GPU memory during long-context decode by offloading KV data to CPU pinned memory. Requires PD disaggregation mode (decode instance only). See HiSparse Guide.
  • NVFP4 on Blackwell: Specify --quantization modelopt_fp4 and --moe-runner-backend flashinfer_trtllm (recommended) / flashinfer_cutlass / flashinfer_cutedsl. Full example: 
    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
  • NCCL timeout: Slow model loading → add --dist-timeout 3600.

​
4. Model Invocation

​
4.1 Basic Usage

For basic API usage and request examples, please refer to:

​
4.2 Advanced Usage

​
4.2.1 Reasoning Parser

DeepSeek-V3.2 supports reasoning mode. Enable the reasoning parser during deployment to separate the thinking and content sections:
Command
sglang serve \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --reasoning-parser deepseek-v3 \
  --tp 8 \
  --host 0.0.0.0 \
  --port 30000
Streaming with Thinking Process:
Example
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:30000/v1",
    api_key="EMPTY"
)

# Enable streaming to see the thinking process in real-time
response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V3.2-Exp",
    messages=[
        {"role": "user", "content": "Solve this problem step by step: What is 15% of 240?"}
    ],
    temperature=0.7,
    max_tokens=2048,
    extra_body = {"chat_template_kwargs": {"thinking": True}},
    stream=True
)

# Process the stream
has_thinking = False
has_answer = False
thinking_started = False

for chunk in response:
    if chunk.choices and len(chunk.choices) > 0:
        delta = chunk.choices[0].delta

        # Print thinking process
        if hasattr(delta, 'reasoning_content') and delta.reasoning_content:
            if not thinking_started:
                print("=============== Thinking =================", flush=True)
                thinking_started = True
            has_thinking = True
            print(delta.reasoning_content, end="", flush=True)

        # Print answer content
        if delta.content:
            # Close thinking section and add content header
            if has_thinking and not has_answer:
                print("\n=============== Content =================", flush=True)
                has_answer = True
            print(delta.content, end="", flush=True)

print()
Output Example:
Output
=============== Thinking =================
To solve this problem, I need to calculate 15% of 240.
Step 1: Convert 15% to decimal: 15% = 0.15
Step 2: Multiply 240 by 0.15
Step 3: 240 × 0.15 = 36
=============== Content =================

The answer is 36. To find 15% of 240, we multiply 240 by 0.15, which equals 36.
Note: The reasoning parser captures the model’s step-by-step thinking process, allowing you to see how the model arrives at its conclusions.

​
4.2.2 Tool Calling

DeepSeek-V3.2 and DeepSeek-V3.2-Exp support tool calling capabilities. But they use different parameters. Enable the tool call parser: Note: DeepSeek-V3.2-Speciale does NOT support tool calling. Launch it with reasoning parser only:
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
Deployment Command: For DeepSeek-V3.2-Exp:
Command
sglang serve \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --tool-call-parser deepseekv31 \
  --reasoning-parser deepseek-v3 \
  --chat-template ./examples/chat_template/tool_chat_template_deepseekv32.jinja \
  --tp 8 \
  --host 0.0.0.0 \
  --port 30000
For DeepSeek-V3.2, use --tool-call-parser deepseekv32 and remove --chat-template. Python Example (with Thinking Process):
Example
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:30000/v1",
    api_key="EMPTY"
)

# Define available tools
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get the current weather for a location",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "The city name"
                    },
                    "unit": {
                        "type": "string",
                        "enum": ["celsius", "fahrenheit"],
                        "description": "Temperature unit"
                    }
                },
                "required": ["location"]
            }
        }
    }
]

# Make request with streaming to see thinking process
response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V3.2-Exp",
    messages=[
        {"role": "user", "content": "What's the weather in Beijing?"}
    ],
    tools=tools,
    extra_body = {"chat_template_kwargs": {"thinking": True}},
    temperature=0.7,
    stream=True
)

# Process streaming response
thinking_started = False
has_thinking = False
tool_calls_accumulator = {}

for chunk in response:
    if chunk.choices and len(chunk.choices) > 0:
        delta = chunk.choices[0].delta

        # Print thinking process
        if hasattr(delta, 'reasoning_content') and delta.reasoning_content:
            if not thinking_started:
                print("=============== Thinking =================", flush=True)
                thinking_started = True
            has_thinking = True
            print(delta.reasoning_content, end="", flush=True)

        # Accumulate tool calls
        if hasattr(delta, 'tool_calls') and delta.tool_calls:
            # Close thinking section if needed
            if has_thinking and thinking_started:
                print("\n=============== Content =================\n", flush=True)
                thinking_started = False

            for tool_call in delta.tool_calls:
                index = tool_call.index
                if index not in tool_calls_accumulator:
                    tool_calls_accumulator[index] = {
                        'name': None,
                        'arguments': ''
                    }

                if tool_call.function:
                    if tool_call.function.name:
                        tool_calls_accumulator[index]['name'] = tool_call.function.name
                    if tool_call.function.arguments:
                        tool_calls_accumulator[index]['arguments'] += tool_call.function.arguments

        # Print content
        if delta.content:
            print(delta.content, end="", flush=True)

# Print accumulated tool calls
for index, tool_call in sorted(tool_calls_accumulator.items()):
    print(f"Tool Call: {tool_call['name']}")
    print(f"   Arguments: {tool_call['arguments']}")

print()
Output Example:
Output
=============== Thinking =================
The user is asking about the weather in Beijing. I need to use the get_weather function to retrieve this information.
I should call the function with location="Beijing".
=============== Content =================

Tool Call: get_weather
   Arguments: {"location": "Beijing", "unit": "celsius"}
Note:
  • The reasoning parser shows how the model decides to use a tool
  • Tool calls are clearly marked with the function name and arguments
  • You can then execute the function and send the result back to continue the conversation
Handling Tool Call Results:
Example
# After getting the tool call, execute the function
def get_weather(location, unit="celsius"):
    # Your actual weather API call here
    return f"The weather in {location} is 22°{unit[0].upper()} and sunny."

# Send tool result back to the model
messages = [
    {"role": "user", "content": "What's the weather in Beijing?"},
    {
        "role": "assistant",
        "content": None,
        "tool_calls": [{
            "id": "call_123",
            "type": "function",
            "function": {
                "name": "get_weather",
                "arguments": '{"location": "Beijing", "unit": "celsius"}'
            }
        }]
    },
    {
        "role": "tool",
        "tool_call_id": "call_123",
        "content": get_weather("Beijing", "celsius")
    }
]

final_response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V3.2-Exp",
    messages=messages,
    temperature=0.7
)

print(final_response.choices[0].message.content)
# Output: "The weather in Beijing is currently 22°C and sunny."

​
4.2.3 Multi-Token Prediction (EAGLE Speculative Decoding)

SGLang implements Multi-Token Prediction (MTP) for DeepSeek V3.2 based on EAGLE speculative decoding. This optimization significantly improves decoding speed for small batch sizes. 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
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
Find optimal values for your workload with bench_speculative.py. The minimum viable config is --speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2.
--max-running-requests defaults to 48 for MTP. Increase it for larger batch sizes.
The spec-v2 overlap scheduler is enabled by default (SGLANG_ENABLE_SPEC_V2=True). Set SGLANG_ENABLE_SPEC_V2=0 to disable.

​
4.2.4 PD Disaggregation

Prefill-Decode (PD) disaggregation separates prefill and decode stages onto different instances, improving GPU utilization for mixed workloads. 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 30000
For production deployments (RBG / LWS-based, DeepEP EP parallelism), see multi_node_deployment docs.

​
4.2.5 DSA Long-Sequence Context Parallel and PP/CP

SGLang provides two context parallel (CP) modes for long-sequence workloads, controlled with --dsa-prefill-cp-mode. In-sequence splitting (--dsa-prefill-cp-mode in-seq-split): Each CP rank handles a uniform shard of the sequence; KV cache is gathered via all-gather. Batch size is restricted to 1 during prefill. See PR #12065.
Command
# In-seq splitting mode — EP + DP, batch size 1
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp \
  --tp 8 --ep 8 --dp 2 --enable-dp-attention \
  --enable-dsa-prefill-context-parallel --attn-cp-size 4 \
  --dsa-prefill-cp-mode in-seq-split --max-running-requests 32
Round-robin splitting (--dsa-prefill-cp-mode round-robin-split, default): Distributes tokens by token_idx % cp_size. Supports fused MoE, FP8 KV cache, and multi-batch prefill. Cannot be combined with DP attention. See PR #13959.
Command
# Round-robin splitting — FusedMoE + CP8
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp \
  --tp 8 --enable-dsa-prefill-context-parallel --attn-cp-size 8 \
  --dsa-prefill-cp-mode round-robin-split --max-running-requests 32
PP + CP (multi-node): Combines Pipeline Parallelism and Context Parallelism for cross-node scaling. The production-optimized configurations below have been verified on Hopper: We suggested DP2 + MTP for local deployment of agentic workflow with DeepSeek V3.2 on Hopper platform:
Command
export SGLANG_DEEPEP_LL_COMBINE_SEND_NUM_SMS=32
export SGLANG_SET_CPU_AFFINITY=1

# Test workload ISL/OSL=1k/1k, raw tap : 4948.16 toks/sec, MAX ITL 5970
#   dp 2 : 5019.54  toks/sec, MAX ITL 7233
#   dp 4 : 4942.82  toks/sec, MAX ITL 35654
#   dp 2 + mtp : 6842.51 toks/sec, MAX ITL 3081
sglang_args=$(echo serve \
  --model-path $MAPPED_MODEL_PATH \
  --nccl-init $MASTER_ADDR:$MASTER_PORT --nnodes 2 --node-rank $RANK --tp 16 \
  --dp 2 --enable-dp-attention --page-size 64 \
  --trust-remote-code --host "0.0.0.0" --port 30000 \
  --log-requests \
  --context-length 65536 --max-running-requests 128 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 2 --speculative-eagle-topk 1 --speculative-num-draft-tokens 3 \
  --allow-auto-truncate --enable-metrics \
  --tool-call-parser deepseekv32 --reasoning-parser deepseek-v3 \
  --served-model-name DeepSeek-V3.2-Opt-dp2-mtp
)

sglang_args=($sglang_args)

sglang "${sglang_args[@]}" 2>&1 | tee $LOG_DIR/$RANK.log
CP + PP + EP + DP CP is currently enabled with PP=2 on Hopper platform and we can reduce TP=16 to TP=8 from standalone deployment:
Command
# verified on Hopper platform
sglang_args=$(echo serve \
  --model-path $MAPPED_MODEL_PATH \
  --nccl-init $MASTER_ADDR:$MASTER_PORT --nnodes 2 --node-rank $RANK --tp 8 --pp-size 2 --dp 1 --enable-dp-attention \
  --moe-a2a-backend deepep --ep-size 16  \
  --page-size 128 \
  --chunked-prefill-size 16384 \
  --attention-backend dsa \
  --dsa-prefill-backend flashmla_sparse \
  --dsa-decode-backend flashmla_sparse \
  --enable-dsa-prefill-context-parallel \
  --dsa-prefill-cp-mode round-robin-split \
  --cuda-graph-max-bs 128 \
  --max-running-requests 128 \
  --trust-remote-code --host "0.0.0.0" --port 30000 \
  --log-requests \
  --context-length 65536 \
  --allow-auto-truncate --enable-metrics \
  --tool-call-parser deepseekv32 --reasoning-parser deepseek-v3 \
  --served-model-name DeepSeek-V3.2-dsa-pp-cp-ep-dp
)

sglang_args=($sglang_args)

sglang "${sglang_args[@]}" 2>&1 | tee $LOG_DIR/$RANK.log
fp8 KV + CP + PP With FP8 KV, we can have less memory footprint. This can be combined with various parallel schemes:
Command
# verified in Hopper platform
dp=1

dp_config=" \
  --dp 1 --enable-dp-attention \
"

cp_config=" \
  --enable-dsa-prefill-context-parallel \
"

if [ "$dp" -eq 1 ]; then

cp_config=" \
  $cp_config \
  --dsa-prefill-cp-mode round-robin-split \
"

else
cp_config=" \
  $cp_config \
  --dsa-prefill-cp-mode in-seq-split \
"
fi

# see discussion : https://github.com/sgl-project/sglang/pull/12065
sglang_args=$(echo serve \
  --model-path $MAPPED_MODEL_PATH \
  --nccl-init $MASTER_ADDR:$MASTER_PORT --nnodes 2 --node-rank $RANK --tp 8 --pp-size 2 --pp-async-batch-depth 1 \
  $dp_config \
  --trust-remote-code --host "0.0.0.0" --port 30000 \
  --log-requests \
  --context-length 65536 --max-running-requests 128 \
  $cp_config \
  --kv-cache-dtype fp8_e4m3 \
  --allow-auto-truncate --enable-metrics \
  --tool-call-parser deepseekv32 --reasoning-parser deepseek-v3 \
  --served-model-name DeepSeek-V3.2-Opt-fp8kv-pp2-cp4
)

sglang_args=($sglang_args)

sglang "${sglang_args[@]}" 2>&1 | tee $LOG_DIR/$RANK.log

​
5. Benchmark

​
5.1 Speed Benchmark on Blackwell

Test Environment:
  • Hardware: NVIDIA B200 GPU (8x)
  • Model: DeepSeek-V3.2-Exp
  • Tensor Parallelism: 8
  • sglang version: 0.5.6
We use SGLang’s built-in benchmarking tool to conduct performance evaluation on the ShareGPT_Vicuna_unfiltered dataset. This dataset contains real conversation data and can better reflect performance in actual use scenarios. To simulate real-world usage patterns, we configure each request with 1024 input tokens and 1024 output tokens, representing typical medium-length conversations with detailed responses.

​
5.1.1 Latency-Sensitive Benchmark

  • Model Deployment Command:
Command
sglang serve \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --tp 8 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --host 0.0.0.0 \
  --port 30000
  • Benchmark Command:
Command
python3 -m sglang.bench_serving \
  --backend sglang \
  --host 127.0.0.1 \
  --port 30000 \
  --model deepseek-ai/DeepSeek-V3.2-Exp \
  --random-input-len 1024 \
  --random-output-len 1024 \
  --num-prompts 10 \
  --max-concurrency 1
  • Test Results:
Output
============ Serving Benchmark Result ============
Backend:                                 sglang
Traffic request rate:                    inf
Max request concurrency:                 1
Successful requests:                     10
Benchmark duration (s):                  29.11
Total input tokens:                      1972
Total input text tokens:                 1972
Total input vision tokens:               0
Total generated tokens:                  2784
Total generated tokens (retokenized):    2777
Request throughput (req/s):              0.34
Input token throughput (tok/s):          67.73
Output token throughput (tok/s):         95.62
Peak output token throughput (tok/s):    157.00
Peak concurrent requests:                3
Total token throughput (tok/s):          163.36
Concurrency:                             1.00
Accept length:                           2.46
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   2909.74
Median E2E Latency (ms):                 3088.27
P90 E2E Latency (ms):                    4200.62
P99 E2E Latency (ms):                    5588.52
---------------Time to First Token----------------
Mean TTFT (ms):                          317.58
Median TTFT (ms):                        191.31
P99 TTFT (ms):                           740.79
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          9.09
Median TPOT (ms):                        9.25
P99 TPOT (ms):                           11.73
---------------Inter-Token Latency----------------
Mean ITL (ms):                           9.35
Median ITL (ms):                         7.64
P95 ITL (ms):                            22.81
P99 ITL (ms):                            23.33
Max ITL (ms):                            31.45
==================================================

​
5.1.2 Throughput-Sensitive Benchmark

  • Model Deployment Command:
Command
sglang serve \
  --model-path deepseek-ai/DeepSeek-V3.2-Exp \
  --tp 8 \
  --ep 8 \
  --dp 8 \
  --enable-dp-attention \
  --host 0.0.0.0 \
  --port 30000
  • Benchmark Command:
Command
python3 -m sglang.bench_serving \
  --backend sglang \
  --host 127.0.0.1 \
  --port 30000 \
  --model deepseek-ai/DeepSeek-V3.2-Exp \
  --random-input-len 1024 \
  --random-output-len 1024 \
  --num-prompts 1000 \
  --max-concurrency 100
  • Test Results:
Output
============ Serving Benchmark Result ============
Backend:                                 sglang
Traffic request rate:                    inf
Max request concurrency:                 100
Successful requests:                     1000
Benchmark duration (s):                  219.09
Total input tokens:                      301701
Total input text tokens:                 301701
Total input vision tokens:               0
Total generated tokens:                  188375
Total generated tokens (retokenized):    187443
Request throughput (req/s):              4.56
Input token throughput (tok/s):          1377.06
Output token throughput (tok/s):         859.80
Peak output token throughput (tok/s):    2465.00
Peak concurrent requests:                109
Total token throughput (tok/s):          2236.86
Concurrency:                             88.05
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   19291.23
Median E2E Latency (ms):                 11927.39
---------------Time to First Token----------------
Mean TTFT (ms):                          530.36
Median TTFT (ms):                        444.00
P99 TTFT (ms):                           1504.78
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          106.16
Median TPOT (ms):                        106.69
P99 TPOT (ms):                           221.12
---------------Inter-Token Latency----------------
Mean ITL (ms):                           100.46
Median ITL (ms):                         41.73
P95 ITL (ms):                            225.67
P99 ITL (ms):                            392.37
Max ITL (ms):                            975.03
==================================================

​
5.2 Accuracy Benchmark

​
5.2.1 GSM8K Benchmark

  • Benchmark Command:
Command
python3 -m sglang.test.few_shot_gsm8k --num-questions 200 --port 30000
  • Test Results:
    • DeepSeek-V3.2-Exp 
      Accuracy: 0.980
Invalid: 0.000
Latency: 19.128 s
Output throughput: 965.919 token/s
  • Full GSM8K (1319 questions) — for a stricter accuracy check, run the full set 8-shot:
Command
python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1319 --parallel 1319
  • 8-shot: 
    Accuracy: 0.956
Invalid: 0.000
Latency: 25.109 s
Output throughput: 5226.235 token/s
  • 20-shot (long-context; stays close to the 8-shot result): 
    Accuracy: 0.956
Invalid: 0.000
Latency: 29.545 s
Output throughput: 4418.617 token/s

​
5.2.2 MMLU Benchmark

  • Benchmark Command:
Command
cd sglang
bash benchmark/mmlu/download_data.sh
python3 benchmark/mmlu/bench_sglang.py --nsub 10 --port 30000
  • Test Results:
    • DeepSeek-V3.2-Exp 
      subject: abstract_algebra, #q:100, acc: 0.780
subject: anatomy, #q:135, acc: 0.874
subject: astronomy, #q:152, acc: 0.961
subject: business_ethics, #q:100, acc: 0.860
subject: clinical_knowledge, #q:265, acc: 0.925
subject: college_biology, #q:144, acc: 0.972
subject: college_chemistry, #q:100, acc: 0.660
subject: college_computer_science, #q:100, acc: 0.880
subject: college_mathematics, #q:100, acc: 0.840
subject: college_medicine, #q:173, acc: 0.879
Total latency: 7.961
Average accuracy: 0.879

​
5.2.3 GPQA-Diamond Benchmark

  • Benchmark Command:
Command
python3 -m sglang.test.run_eval --port 30000 --eval-name gpqa --num-examples 198 --max-tokens 128000 --repeat 8 --thinking-mode deepseek-v3
  • Test Results (model: deepseek-ai/DeepSeek-V3.2-Exp, 8×B200):
    • Default (temperature=0): mean 0.797 over 8 runs — closely matches the official GPQA-Diamond score of 79.9 for DeepSeek-V3.2-Exp reported in its model card
    • With temperature=1.0, top_p=0.95 (as recommended by DeepSeek):
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']

​
5.2.4 AIME 2025 Benchmark

Results on AIME 2025 (8×B200), evaluated with NeMo-Skills:
Modelpass@1 avg-of-4majority@4pass@4
DeepSeek-V3.2-Exp87.50% ± 1.67%90.00%90.00%
DeepSeek-V3.292.50% ± 1.67%94.71%96.67%
DeepSeek-V3.2-Speciale95.00% ± 1.92%95.83%100.00%
Reproduction. Install NeMo-Skills, launch the server with the tool-call and reasoning parsers, then run ns eval:
Command
pip install git+https://github.com/NVIDIA/NeMo-Skills.git --ignore-installed blinker

export NEMO_SKILLS_DISABLE_UNCOMMITTED_CHANGES_CHECK=1
ns prepare_data aime25

ns eval \
  --benchmarks=aime25:4 \
  --server_type=sglang \
  --model=deepseek-ai/DeepSeek-V3.2-Exp \
  --server_address=http://localhost:30000/v1 \
  --output_dir=nemo_skills_aime25_output \
  ++chat_template_kwargs.thinking=true \
  ++inference.temperature=1.0 \
  ++inference.top_p=0.95 \
  ++inference.tokens_to_generate=64000
  # Use ++inference.tokens_to_generate=120000 for the DeepSeek-V3.2-Speciale model

​
5.3 Speed Benchmark on Hopper

Test Environment:
  • Hardware: NVIDIA H800 GPU (16x)
  • Model: DeepSeek-V3.2
  • Tensor Parallelism: 16
  • sglang version: 0.5.9

​
5.3.1 Latency-Sensitive Benchmark

  • Model Deployment Command:
Command
export SGLANG_DEEPEP_LL_COMBINE_SEND_NUM_SMS=32
export SGLANG_SET_CPU_AFFINITY=1

# Test workload ISL/OSL=1k/1k, raw tap : 4948.16 toks/sec, MAX ITL 5970
#   dp 2 : 5019.54  toks/sec, MAX ITL 7233
#   dp 4 : 4942.82  toks/sec, MAX ITL 35654
#   dp 2 + mtp : 6842.51 toks/sec, MAX ITL 3081
sglang_args=$(echo serve \
  --model-path $MAPPED_MODEL_PATH \
  --nccl-init $MASTER_ADDR:$MASTER_PORT --nnodes 2 --node-rank $RANK --tp 16 \
  --dp 2 --enable-dp-attention --page-size 64 \
  --trust-remote-code --host "0.0.0.0" --port 30000 \
  --log-requests \
  --context-length 65536 --max-running-requests 128 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 2 --speculative-eagle-topk 1 --speculative-num-draft-tokens 3 \
  --allow-auto-truncate --enable-metrics \
  --tool-call-parser deepseekv32 --reasoning-parser deepseek-v3 \
  --served-model-name DeepSeek-V3.2-Opt-dp2-mtp
)

sglang_args=($sglang_args)

sglang "${sglang_args[@]}" 2>&1 | tee $LOG_DIR/$RANK.log
  • Benchmark Command:
Command
python3 -m sglang.bench_serving \
  --backend sglang \
  --host $MASTER_ADDR \
  --port 30000 \
  --model deepseek-ai/DeepSeek-V3.2 \
  --random-input-len 1024 \
  --random-output-len 1024 \
  --num-prompts 10 \
  --max-concurrency 1
  • Test Results:
Output
============ Serving Benchmark Result ============
Backend:                                 sglang
Traffic request rate:                    64.0
Max request concurrency:                 1
Successful requests:                     10
Benchmark duration (s):                  48.96
Total input tokens:                      6101
Total input text tokens:                 6101
Total generated tokens:                  4220
Total generated tokens (retokenized):    4217
Request throughput (req/s):              0.20
Input token throughput (tok/s):          124.62
Output token throughput (tok/s):         86.20
Peak output token throughput (tok/s):    113.00
Peak concurrent requests:                2
Total token throughput (tok/s):          210.81
Concurrency:                             1.00
Accept length:                           3.27
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   4893.12
Median E2E Latency (ms):                 3742.47
P90 E2E Latency (ms):                    8877.37
P99 E2E Latency (ms):                    10769.85
---------------Time to First Token----------------
Mean TTFT (ms):                          199.88
Median TTFT (ms):                        176.15
P99 TTFT (ms):                           272.49
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          10.99
Median TPOT (ms):                        10.88
P99 TPOT (ms):                           13.93
---------------Inter-Token Latency----------------
Mean ITL (ms):                           11.15
Median ITL (ms):                         8.86
P95 ITL (ms):                            17.29
P99 ITL (ms):                            33.71
Max ITL (ms):                            36.84
==================================================

​
5.3.2 Throughput-Sensitive Benchmark

We simply use the same deployment method and vary the throughput by maximizing concurrencies:
Command
python3 -m sglang.bench_serving \
  --backend sglang \
  --host $MASTER_ADDR \
  --port 30000 \
  --model deepseek-ai/DeepSeek-V3.2 \
  --random-input-len 1024 \
  --random-output-len 1024 \
  --num-prompts 2048 \
  --max-concurrency 1024 # see picture below why we use 1024 for concurrency, hence num prompts 2048
DeepSeek 3.2 can steadily support concurrency up to 1024 and when concurrency is greater than 128, the TTFT increase sharply: DeepSeek V3.2 Concurrency ISL/OSL=1024/128 Performance record:
Output
============ Serving Benchmark Result ============
Backend:                                 sglang
Traffic request rate:                    64.0
Max request concurrency:                 1024
Successful requests:                     2048
Benchmark duration (s):                  408.09
Total input tokens:                      1048992
Total input text tokens:                 1048992
Total generated tokens:                  1032734
Total generated tokens (retokenized):    1031817
Request throughput (req/s):              5.02
Input token throughput (tok/s):          2570.50
Output token throughput (tok/s):         2530.66
Peak output token throughput (tok/s):    5092.00
Peak concurrent requests:                1035
Total token throughput (tok/s):          5101.16
Concurrency:                             763.41
Accept length:                           3.26
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   152117.70
Median E2E Latency (ms):                 181704.84
P90 E2E Latency (ms):                    215924.77
P99 E2E Latency (ms):                    231679.59
---------------Time to First Token----------------
Mean TTFT (ms):                          127729.28
Median TTFT (ms):                        170098.94
P99 TTFT (ms):                           185705.73
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          49.18
Median TPOT (ms):                        48.48
P99 TPOT (ms):                           77.24
---------------Inter-Token Latency----------------
Mean ITL (ms):                           48.46
Median ITL (ms):                         52.11
P95 ITL (ms):                            110.26
P99 ITL (ms):                            200.63
Max ITL (ms):                            2666.37
==================================================
By adding --random-range-ratio 1, we could get even higher statistical numbers:
Output
============ Serving Benchmark Result ============
Backend:                                 sglang
Traffic request rate:                    64.0
Max request concurrency:                 1024
Successful requests:                     2048
Benchmark duration (s):                  612.87
Total input tokens:                      2097152
Total input text tokens:                 2097152
Total generated tokens:                  2097152
Total generated tokens (retokenized):    2096201
Request throughput (req/s):              3.34
Input token throughput (tok/s):          3421.84
Output token throughput (tok/s):         3421.84
Peak output token throughput (tok/s):    9077.00
Peak concurrent requests:                1039
Total token throughput (tok/s):          6843.68
Concurrency:                             772.66
Accept length:                           3.26
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   231222.27
Median E2E Latency (ms):                 289846.24
P90 E2E Latency (ms):                    314480.41
P99 E2E Latency (ms):                    320392.27
---------------Time to First Token----------------
Mean TTFT (ms):                          194081.02
Median TTFT (ms):                        252945.22
P99 TTFT (ms):                           279637.50
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          36.31
Median TPOT (ms):                        36.73
P99 TPOT (ms):                           46.33
---------------Inter-Token Latency----------------
Mean ITL (ms):                           36.31
Median ITL (ms):                         23.18
P95 ITL (ms):                            96.79
P99 ITL (ms):                            135.81
Max ITL (ms):                            3121.00
==================================================