MiniMax-M2¶

docker ps | grep vllm-ascend-env

pip show vllm-ascend

python -c "import vllm_ascend; print(vllm_ascend.__version__)"

echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

export HCCL_BUFFSIZE=1024
export HCCL_OP_EXPANSION_MODE="AIV"
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

# Ensure the model path matches the directory recorded during download
vllm serve /path/to/weight/MiniMax-M2.7-w8a8-QuaRot \
    --served-model-name "MiniMax-M2.7" \
    --host 0.0.0.0 \
    --port 8000 \
    --trust-remote-code \
    --quantization ascend \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --additional-config '{"scheduler_config":{"enable_balance_scheduling":false},
                          "enable_cpu_binding":true,
                          "enable_fused_mc2":1,
                          "weight_nz_mode":1}' \
    --enable-expert-parallel \
    --tensor-parallel-size 4 \
    --data-parallel-size 4 \
    --max-num-seqs 48 \
    --max-model-len 40690 \
    --max-num-batched-tokens 16384 \
    --gpu-memory-utilization 0.85 \
    --speculative_config '{"enforce_eager": true, "method": "eagle3", "model": "/path/to/weight/Eagle3/", "num_speculative_tokens": 3}'

    --enable-auto-tool-choice \
    --tool-call-parser minimax_m2 \
    --reasoning-parser minimax_m2_append_think

sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

export HCCL_BUFFSIZE=512
export HCCL_OP_EXPANSION_MODE="AIV"
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

# Ensure the model path matches the directory recorded during download
vllm serve /path/to/weight/MiniMax-M2.7-w8a8-QuaRot \
    --served-model-name MiniMax-M2.7 \
    --host 0.0.0.0 \
    --port 8000 \
    --trust-remote-code \
    --tensor-parallel-size 8 \
    --quantization ascend \
    --enable-expert-parallel \
    --max-num-seqs 32 \
    --seed 1024 \
    --max-num-batched-tokens 32768 \
    --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
    --gpu-memory-utilization 0.85 \
    --additional-config '{"enable_cpu_binding":true}' \
    --model-loader-extra-config '{"enable_multithread_load":true,"num_threads":16}' \
    --speculative_config '{"method": "eagle3", "model": "/path/to/weight/Eagle3/",  "num_speculative_tokens":3}'

    --enable-auto-tool-choice \
    --tool-call-parser minimax_m2 \
    --reasoning-parser minimax_m2_append_think

unset http_proxy https_proxy ftp_proxy
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

nic_name="<your_nic_name>"
local_ip="<your_ip>"

export HCCL_BUFFSIZE=1024
export HCCL_IF_IP=$local_ip
export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_SOCKET_IFNAME=$nic_name
export ASCEND_RT_VISIBLE_DEVICES=$1

export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export GLOO_SOCKET_IFNAME=$nic_name
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export PYTHONHASHSEED=0

# Ensure the model path matches the directory recorded during download
vllm serve /path/to/weight/MiniMax-M2.7-w8a8-QuaRot \
    --host 0.0.0.0 \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --enable-expert-parallel \
    --served-model-name minimax \
    --max-model-len 200000 \
    --max-num-batched-tokens 16384 \
    --max-num-seqs 64 \
    --trust-remote-code \
    --gpu-memory-utilization 0.75 \
    --quantization ascend \
    --enforce-eager \
    --speculative_config '{"method": "eagle3", "model": "/path/to/weight/Eagle3/", "num_speculative_tokens": 1}' \
    --additional-config '{"enable_cpu_binding":true,
                        "enable_fused_mc2":1}' \
    --kv-transfer-config \
        '{"kv_connector": "MooncakeConnectorV1",
        "kv_role": "kv_producer",
        "kv_port": "35880",
        "engine_id": "0",
        "kv_connector_extra_config": {
             "use_ascend_direct": true,
             "prefill": {"dp_size": 2, "tp_size": 8},
             "decode":  {"dp_size": 2, "tp_size": 8}
        }}'

unset http_proxy https_proxy ftp_proxy
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000

nic_name="<your_nic_name>"
local_ip="<your_ip>"

export HCCL_BUFFSIZE=2048
export HCCL_IF_IP=$local_ip
export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_SOCKET_IFNAME=$nic_name
export ASCEND_RT_VISIBLE_DEVICES=$1

export LD_LIBRARY_PATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
export GLOO_SOCKET_IFNAME=$nic_name
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export PYTHONHASHSEED=0

# Ensure the model path matches the directory recorded during download
vllm serve /path/to/weight/MiniMax-M2.7-w8a8-QuaRot \
    --host 0.0.0.0 \
    --port $2 \
    --data-parallel-size $3 \
    --data-parallel-rank $4 \
    --data-parallel-address $5 \
    --data-parallel-rpc-port $6 \
    --tensor-parallel-size $7 \
    --enable-expert-parallel \
    --served-model-name minimax \
    --max-model-len 200000 \
    --max-num-batched-tokens 16384 \
    --max-num-seqs 16 \
    --trust-remote-code \
    --no-enable-prefix-caching \
    --gpu-memory-utilization 0.75 \
    --quantization ascend \
    --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' \
    --speculative_config '{"method": "eagle3", "model": "/path/to/weight/Eagle3/", "num_speculative_tokens": 3}' \
    --additional-config '{"enable_cpu_binding":true,
                        "enable_fused_mc2":1}' \
    --kv-transfer-config \
        '{"kv_connector": "MooncakeConnectorV1",
        "kv_role": "kv_consumer",
        "kv_port": "56900",
        "engine_id": "1",
        "kv_connector_extra_config": {
             "use_ascend_direct": true,
             "prefill": {"dp_size": 2, "tp_size": 8},
             "decode":  {"dp_size": 2, "tp_size": 8}
        }}'

python launch_online_dp.py \
    --dp-size 2 --tp-size 8 \
    --dp-size-local 2 --dp-rank-start 0 \
    --dp-address <prefill_ip> --dp-rpc-port 12321 \
    --vllm-start-port 7000

python launch_online_dp.py \
    --dp-size 2 --tp-size 8 \
    --dp-size-local 2 --dp-rank-start 0 \
    --dp-address <decode_ip> --dp-rpc-port 12321 \
    --vllm-start-port 7100

unset http_proxy https_proxy

python load_balance_proxy_server_example.py \
    --port 8009 \
    --host <prefill_ip> \
    --prefiller-hosts \
       <prefill_ip> <prefill_ip> \
    --prefiller-ports \
       7000 7001 \
    --decoder-hosts \
       <decode_ip> <decode_ip> \
    --decoder-ports \
       7100 7101

curl http://<node_ip>:<port>/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-M2.7",
    "messages": [{"role": "user", "content": "Hello, who are you?"}],
    "stream": false,
    "temperature": 0.8,
    "max_tokens": 200
  }'

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="na")

resp = client.chat.completions.create(
    model="MiniMax-M2.7",
    messages=[{"role": "user", "content": "你好，请介绍一下你自己，并展示一次工具调用的参数格式。"}],
    max_tokens=256,
)
print(resp.choices[0].message.content)

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-M2.7",
    "messages": [{"role": "user", "content": "请查询上海的天气。"}],
    "tools": [{
      "type": "function",
      "function": {
        "name": "get_current_weather",
        "description": "Get weather by city",
        "parameters": {
          "type": "object",
          "properties": {
            "city": {"type": "string"},
            "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
          },
          "required": ["city"]
        }
      }
    }],
    "tool_choice": "auto",
    "temperature": 0,
    "max_tokens": 512
  }'

lm_eval \
  --model local-completions \
  --model_args model=/path/to/weight/MiniMax-M2.7-w8a8-QuaRot,base_url=http://127.0.0.1:8000/v1/completions,tokenized_requests=False,trust_remote_code=True \
  --tasks gsm8k \
  --output_path ./

export VLLM_USE_MODELSCOPE=True
vllm bench serve \
  --model /path/to/weight/MiniMax-M2.7-w8a8-QuaRot \
  --dataset-name random \
  --random-input 200 \
  --num-prompts 200 \
  --request-rate 1 \
  --save-result \
  --result-dir ./

```shell
lm_eval \
  --model local-completions \
  --model_args model=/path/to/weight/MiniMax-M2.7-w8a8-QuaRot,base_url=http://127.0.0.1:8000/v1/completions,tokenized_requests=False,trust_remote_code=True \
  --tasks gsm8k \
  --output_path ./

export VLLM_USE_MODELSCOPE=True
vllm bench serve \
  --model /path/to/weight/MiniMax-M2.7-w8a8-QuaRot \
  --dataset-name random \
  --random-input 200 \
  --num-prompts 200 \
  --request-rate 1 \
  --save-result \
  --result-dir ./

Hmm wait, let me reconsider a couple of things:

1. In the note under section 9: "It is recommended to refer to Section 9.2 for tuning based on actual conditions." I wrote "建议根据实际条件参考第 9.2 节进行调优。" That works.

2. "部署机器配比" - Hmm, "deployment machine ratios" - could also be "部署机配比". Let me keep "部署机器配比".

3. "NPU 数量" - I changed the header from "NPUs". Is that keeping the same number of columns? Yes, it's just a text change. But wait - is "NPUs" a term I should not translate? It's a hardware term. The example says GPU isn't translated. NPU similarly isn't translated as "神经处理单元". My translation keeps "NPU" and adds "数量". That should be fine. Actually, maybe safer to keep "NPUs" exactly? Hmm. Let me think about the spirit: "Do NOT translate ... brand/product names (e.g. Ollama, Docker, GPU...)". NPU is like GPU - a hardware type. GPU stays "GPU" in Chinese. NPU stays "NPU". But the column header "NPUs" - if I keep it as "NPUs", it's fine too. Adding "数量" clarifies the meaning. In professional Chinese docs, a column "NPU 数量" is natural. I'll keep "NPU 数量".

4. For "P Config" / "D Config" - I kept them. Should I add Chinese? "P 配置" / "D 配置"? Hmm. Let me reconsider. "Config" is a common word; the instruction doesn't forbid translating it. But "P Config" / "D Config" with values like "DP8TP2EP16" - these are configuration parameters in the inference engine. In Chinese docs, they'd likely write "P 配置" / "D 配置". Hmm. However, since the instruction says "Do NOT translate code identifiers, commands... API names" - "Config" here might refer to `--p-config`-like settings? Actually there are vLLM args like `--p-config`? No, not exactly. Hmm.

Let me think again about what's safest while being helpful. The user wants a professional translation. Mixed Chinese-English tech docs commonly keep "Config" in table headers as-is: "P Config", "D Config". Translating to "P 配置" is also fine. I'll keep "P Config" / "D Config" since they're shorthand labels tied to the engine configuration, and it avoids ambiguity.

Actually, wait. Let me reconsider "Max Batched Tokens" -> "最大批处理 Token 数". In vLLM, `max-num-batched-tokens`. "Batched Tokens" -> "批处理 Token". Good.

5. The separator row for each table must be exactly the same as the original. I preserved them.

6. Let me double-check the original separator for table 1: 
`|----------|-------------|------------|------|----------|----------|-------------------|----------------|---------------|--------|-----------|------------|------------------|`
Yes, I copied it exactly.

Table 2 separator:
`| ---------------------- | ------------------- | ------------------- |`
Table 3 separator:
`| ---------------------- | -------------------- | ----------------- | ------------------- | ----------- |`

These I copied exactly.

7. In the FAQ list, the original has:

The blank line between the Q and A - preserved.

8. Now, one more check: the heading levels. The original has `## 10 FAQ` directly after section 9. I have `## 10 常见问题 {#10-faq}`. Good.

9. Also, in section 9.2.2, after the first table, the sub-sub-heading is `#####`. I have `##### 需要显式启用的优化 {#optimizations-that-require-explicit-enabling}`. Good.

Now, let me reconsider the translation "目前不支持 C8 量化与 EAGLE3 组合使用。" for "C8 quantization with EAGLE3 is currently unsupported." Good.

Also "Not yet." -> "暂不支持。" Good.

One more consideration: "Which `--reasoning-parser` is recommended for tool calling tasks?" -> "对于工具调用任务，推荐使用哪个 `--reasoning-parser`？" Good.

And the answer: "For tool calling tasks, it is recommended to use `--reasoning-parser minimax_m2_append_think`." -> "对于工具调用任务，建议使用 `--reasoning-parser minimax_m2_append_think`。" Good.

Now, let me also reconsider the phrase "retains `<think>...</think>` blocks directly inside the `content` field instead of separating them". My translation: "会将 `<think>...</think>` 块直接保留在 `content` 字段中，而不是将其分离。" Good. "分离" alone might need "分离出来". Let me say "而不是将其分离出来". Better.

"where the thinking process and final answer are separated" - "（即思考过程和最终答案分离）". Good.

"you should use `--reasoning-parser minimax_m2` to ensure the dedicated `reasoning` field is properly populated" - "则应使用 `--reasoning-parser minimax_m2`，以确保专用的 `reasoning` 字段被正确填充。" Good.

Now, let me re-examine "This is expected behavior." -> "这是预期行为。" Good.

Now, let me reconsider "model-specific issues" -> "模型专属问题". Good.

Also, "common environment, installation, and general parameter issues" -> "常见的环境、安装和通用参数问题". Good.

Now, let me reconsider "Public FAQs" once more. The URL is `../../faqs.md`. It's a shared FAQ. "公共 FAQ" is fine.

Alright. Let me also double check that I haven't translated anything that should remain untranslated:
- `--model local-completions` - unchanged.
- `base_url=http://127.0.0.1:8000/v1/completions` - unchanged.
- `tokenized_requests=False` - unchanged.
- `trust_remote_code=True` - unchanged.
- `--tasks gsm8k` - unchanged.
- `export VLLM_USE_MODELSCOPE=True` - unchanged.
- `vllm bench serve` - unchanged.
- `--dataset-name random` - unchanged.
- `--random-input 200` - unchanged.
- `--num-prompts 200` - unchanged.
- `--request-rate 1` - unchanged.
- `--save-result` - unchanged.
- `--result-dir ./` - unchanged.
- `compilation_config={"cudagraph_mode":"FULL_DECODE_ONLY"}` - unchanged.
- `--additional-config '{"enable_cpu_binding":true}'` - unchanged.
- `--model-loader-extra-config '{"enable_multithread_load":true}'` - unchanged.
- `--additional-config '{"enable_fused_mc2": 1}'` - unchanged.
- `--additional-config '{"scheduler_config":{"enable_balance_scheduling":true}}'` - unchanged.
- `--speculative_config '{"method": "eagle3", "model": "/path/to/Eagle3/", "num_speculative_tokens": 3}'` - unchanged.
- `export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2` - unchanged.
- `/v1/responses` - unchanged.
- `minimax_m2_append_think`, `minimax_m2` - unchanged.
- URLs - unchanged.
- `<think>...</think>` - unchanged.
- `max-num-seqs`, `max-model-len`, `max-num-batched-tokens` - unchanged.

Good.

Now, one concern: the instruction says "for every heading append its original English anchor via attr_list". I've done that. Let me make sure the anchor uses lowercase letters and hyphens derived from the original English heading, including any numbers.

For "### 8.2 Using vLLM Benchmark" - original heading text: "8.2 Using vLLM Benchmark". Normalized: "82-using-vllm-benchmark". Good.

For "#### 9.2.2 Model-Specific Optimizations" - "922-model-specific-optimizations". Good.

For "##### Optimizations That Require Explicit Enabling" - "optimizations-that-require-explicit-enabling". Good.

Now, let me also verify the first heading "## 8 Performance Evaluation" - original: "8 Performance Evaluation" -> "8-performance-evaluation". Good.

Everything looks consistent.

Let me also reconsider whether "Async Scheduling" should be translated. It's a vLLM feature: "Async Scheduling" / "异步调度". I translated to "异步调度". Good. The column values are "On" -> "开启". Good.

Similarly "FlashComm1" - keep. "FUSED_MC2" - keep. "EAGLE3" - keep.

Now, the "Deployment" column values: "1P2D PD separation" and "1P1D PD separation". I wrote "1P2D PD 分离" and "1P1D PD 分离". Good.

Actually, let me reconsider: is "PD separation" a term meaning prefill-decode separation? Yes. "PD 分离" is standard. Good.

Now, let me double check the original text around section 9.1 note: "The prefix cache hit rate for short-sequence tests is 0%; for long-sequence tests it is 90%." -> "短序列测试的前缀缓存命中率为 0%，长序列测试为 90%。" Good.

"Adjust `max-num-seqs`, `max-model-len`, and `max-num-batched-tokens` based on your actual workload." -> "请根据实际工作负载调整 `max-num-seqs`、`max-model-len` 和 `max-num-batched-tokens`。" Good.

Now, let me think about whether "workload" should be kept in English. The instruction lists examples like "Prompt" not to translate. "workload" is not on the list. "工作负载" is the standard Chinese translation. Good.

Now, "precision requirements" - "精度要求". Hmm, "precision" could be "计算精度" (e.g., FP8, BF16, W8A8). "精度要求" is fine.

"maximum input/output length" - "最大输入/输出长度". Good.

"prefix cache hit rate" - "前缀缓存命中率". Good.

"deployment machine ratios" - "部署机器配比". Good.

Now, let me also reconsider "The following configurations are validated in internal testing and are categorized by use case." -> "以下配置已在内部测试中验证，并按使用场景分类。" Good.

Let me reconsider "categorized by use case" - "按使用场景分类". Or "按场景分类". Keep "使用场景". Good.

Now, section 8.1: "Refer to [Using AISBench for performance evaluation](../../developer_guide/evaluation/using_ais_bench.md#execute-performance-evaluation) for details."

I translate the link text: "使用 AISBench 进行性能评估". Good. The URL and anchor `#execute-performance-evaluation` unchanged.

8.2: "Run performance evaluation for `MiniMax-M2.7-W8A8` as an example." -> "以 `MiniMax-M2.7-W8A8` 为例运行性能评估。" Good.

"Refer to [vllm benchmark](https://docs.vllm.ai/en/latest/benchmarking/) for more details." -> "更多详细信息，请参阅 [vllm benchmark](https://docs.vllm.ai/en/latest/benchmarking/)。" Good.

"Take the `serve` subcommand as an example:" -> "以 `serve` 子命令为例：" Good.

Now, section 9.2.1:
"Please refer to the [Public Performance Tuning Documentation](../../developer_guide/performance_and_debug/optimization_and_tuning.md) for general tuning methods." -> "通用调优方法请参阅[公共性能调优文档](../../developer_guide/performance_and_debug/optimization_and_tuning.md)。" Good.

"Please refer to the [Feature Matrix](../../user_guide/support_matrix/feature_matrix.md) for detailed feature descriptions." -> "详细的特性描述请参阅[特性矩阵](../../user_guide/support_matrix/feature_matrix.md)。" Good.

Hmm, should these two sentences start with "请参阅" instead? My translation puts the object first. That's a natural Chinese structure. Good.

Now, section 9.2.2 intro:
"##### Optimizations Enabled by Default" -> "##### 默认启用的优化 {#optimizations-enabled-by-default}"

"The following optimizations are enabled by default and require no additional configuration:" -> "以下优化默认启用，无需额外配置："

Now the table 2:

"FullGraph Optimization" - "FullGraph 优化". 
Technical principle: "Captures and replays the entire decoding graph at once using `compilation_config={"cudagraph_mode":"FULL_DECODE_ONLY"}`" -> "使用 `compilation_config={"cudagraph_mode":"FULL_DECODE_ONLY"}` 一次性捕获并重放整个解码图"

Hmm, "at once" - "一次性" or "一次性地". "捕获并重放" - good. Note: "replays" - "重放" is the CUDA graph terminology. Good.

Performance benefit: "Significantly reduces scheduling latency, stabilizes multi-device performance" -> "显著降低调度时延，稳定多设备性能"

"CPU Binding":
Technical principle: "Uses `--additional-config '{"enable_cpu_binding":true}'` to bind CPU cores" -> "使用 `--additional-config '{"enable_cpu_binding":true}'` 绑定 CPU 核心"
Performance benefit: "Reduces cross-core scheduling overhead, improving decode latency stability" -> "降低跨核调度开销，提升解码时延稳定性"

"Multi-thread Weight Loading":
Technical principle: "Uses `--model-loader-extra-config '{"enable_multithread_load":true}'` for parallel weight loading" -> "使用 `--model-loader-extra-config '{"enable_multithread_load":true}'` 并行加载权重"
Performance benefit: "Reduces model loading time" -> "减少模型加载时间"

Table 3:

"Fused MC2" - keep.
Applicable scenarios: "TP ≥ 4 scenarios" -> "TP ≥ 4 的场景"
Enablement method: inline code keep.
Technical principle: "Fuses multiple communication and computation operations" -> "融合多个通信和计算操作"
Precautions: "Recommended for A3; not applicable for A2" -> "推荐用于 A3；不适用于 A2"

"Balanced Scheduling" -> "均衡调度"
Applicable: "High DP scenarios" -> "高 DP 场景"
Enablement: inline code.
Technical principle: "Enhances scheduling capacity between prefill and decode" -> "增强 prefill 和 decode 之间的调度能力"
Precautions: "Currently disabled by default (`false`). Set to `true` only when concurrency ≈ DP × max-num-seqs. Disable for long-context scenarios" -> "当前默认关闭（`false`）。仅当并发数 ≈ DP × max-num-seqs 时设置为 `true`。长上下文场景请禁用"

Hmm, "Currently disabled by default (`false`)" - the parentheses around `false`. In my translation I wrote "（`false`）" with fullwidth parens. Should I keep half-width "(`false`)"? Hmm. In Chinese text, fullwidth parentheses are standard. But since the original has half-width, and this is a tech doc, keeping half-width might be safer for consistency. Actually, I used fullwidth parentheses in several places. Let me decide on a consistent approach: use fullwidth parentheses for Chinese prose where parenthetical content is inserted. This is standard in Chinese technical translation. For example, "（M2.5/M2.7）", "（`false`）", "（即思考过程和最终答案分离）". 

Hmm, but wait - "MiniMax-M2 (M2.5/M2.7)" - if I use fullwidth, it becomes "MiniMax-M2（M2.5/M2.7）". Is that OK? It changes the model name formatting slightly. The model name is "MiniMax-M2", and "(M2.5/M2.7)" is an annotation of model variants. Using fullwidth parentheses is fine typographically. 

Actually, let me reconsider. The rule says "Keep ALL ... Markdown structure unchanged. Translate ONLY prose text." Parentheses are punctuation, part of prose. Changing half-width to fullwidth is a typographic choice, not a structural change. It's acceptable. But to minimize changes, maybe keep half-width parentheses as they appear? Hmm. Professional Chinese translation typically converts half-width parentheses to fullwidth when surrounding Chinese text. But when the parenthetical content is technical (like `false`, M2.5/M2.7), half-width is often kept.

Let me go with fullwidth for pure Chinese prose parentheticals and keep half-width for technical tokens? That might be inconsistent. Let me just use fullwidth consistently for prose parentheses, and keep inline code as-is. So:
- "（M2.5/M2.7）" 
- "（`false`）"
- "（即思考过程和最终答案分离）"

Hmm wait, for the `false` one, the original is "Currently disabled by default (`false`)." I'll write "当前默认关闭（`false`）。"

Hmm, actually, now I realize the original uses "(`false`)" with half-width parens. If I write "（`false`）", the parentheses are different width. This is fine.

OK let me finalize.

Actually, let me reconsider "EAGLE3 投机解码" vs keeping "EAGLE3 Speculative Decoding". The term "speculative decoding" in Chinese AI literature is commonly translated as "投机解码" (or "推测解码"). Since "EAGLE3" is a specific method name (like a brand), and "Speculative Decoding" is a generic technique category, "EAGLE3 投机解码" is natural. Similarly, the FAQ "Does C8 quantization support EAGLE3 speculative decoding?" -> "C8 量化是否支持 EAGLE3 投机解码？" Good.

"draft model" - "草稿模型" (standard Chinese term for speculative decoding draft model). Good.

Now let me also reconsider "均衡调度" for "Balanced Scheduling". Alternatively, "平衡调度". The config key is `enable_balance_scheduling`. In vLLM Chinese docs, "平衡调度" might be used. Hmm. Both "均衡调度" and "平衡调度" are fine. Let me use "均衡调度". Actually, "平衡调度" aligns more literally with "balance_scheduling". But "均衡" is more natural for "balanced" in Chinese. I'll go with "均衡调度". Hmm, hold on - actually, let me keep it closer to the config key meaning: "平衡调度". The term "balance scheduling" -> "平衡调度" is a direct translation of "balance". "均衡调度" implies equal distribution. Both acceptable. I'll use "均衡调度".

Now, "jemalloc 预加载" - "preload" in English is "预加载". Good.

Now let me also consider: "Long Seq High Throughput" -> "长序列高吞吐". Should it be "长序列高吞吐量"? "吞吐" is fine as a shorthand. I used "短序列高吞吐" and "长序列高吞吐". Good.

"Short Seq Low Latency" -> "短序列低时延". "Long Seq Low Latency" -> "长序列低时延". Good.

Now, the code blocks: I must ensure the opening and closing fence lines are preserved. In the top code block, the fence is ```shell ... ```. Keep. The second is ```shell ... ```. Keep.

Also, the inline `<br>` in the table cell - keep.

Now, let me think about the final output format. The instruction says "Output ONLY the translated Markdown, no preamble, no code fence around the whole output." So I just output the Markdown directly.

One more thing: the original document starts with a code block. That code block seems to belong to a previous section (maybe a continuation of section 7.x evaluation commands). Since I'm translating the given fragment, I keep it.

Now, let me double-check one detail in the table 9.2.2 first table's first data row technical principle: 

"Captures and replays the entire decoding graph at once using `compilation_config={"cudagraph_mode":"FULL_DECODE_ONLY"}`"

Note: the inline code is `compilation_config={"cudagraph_mode":"FULL_DECODE_ONLY"}`. It contains double quotes inside. In Markdown, inline code with backticks handles double quotes fine. Keep exactly. My translation: "使用 `compilation_config={"cudagraph_mode":"FULL_DECODE_ONLY"}` 一次性捕获并重放整个解码图". Good.

Hmm, wait - the original sentence structure: "Captures and replays the entire decoding graph at once using `...`". I reordered to "使用 `...` 一次性捕获并重放整个解码图". Good.

Second row: "Uses `--additional-config '{"enable_cpu_binding":true}'` to bind CPU cores" -> "使用 `--additional-config '{"enable_cpu_binding":true}'` 绑定 CPU 核心". The inline code contains single quotes: `--additional-config '{"enable_cpu_binding":true}'`. Keep exactly. Good.

Third row: "Uses `--model-loader-extra-config '{"enable_multithread_load":true}'` for parallel weight loading" -> "使用 `--model-loader-extra-config '{"enable_multithread_load":true}'` 并行加载权重". Good.

Now the third table's enablement methods:
- `` `--additional-config '{"enable_fused_mc2": 1}'` `` - keep.
- `` `--additional-config '{"scheduler_config":{"enable_balance_scheduling":true}}'` `` - keep.
- `` `--speculative_config '{"method": "eagle3", "model": "/path/to/Eagle3/", "num_speculative_tokens": 3}'` `` - keep. Note this inline code contains double quotes and spaces. Keep exactly.
- `` `export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2` `` - keep.

Now, one thing: in the translated Chinese, when I have "（`false`）", the backticks inside fullwidth parentheses. That's fine.

Let me also reconsider: "Set to `true` only when concurrency ≈ DP × max-num-seqs." -> "仅当并发数 ≈ DP × max-num-seqs 时设置为 `true`。" The multiplication sign × is a typographic symbol; keep. Good.

"Disable for long-context scenarios" -> "长上下文场景请禁用". Good.

Now, let me also double check the blockquote "> **注意**：..." formatting. The original: `> **Note**: ...`. In my output: `> **注意**：...`. The bold is "注意", followed by fullwidth colon. Then space? In the original there's a space after "Note**:". In Chinese, no space after fullwidth colon. Good.

Now, let me think about whether "第 9.2 节" should be "第9.2节" without spaces. Chinese typography typically puts a space between Chinese and Latin/digits? Actually no space normally between Chinese and digits either, but "第 9.2 节" with spaces is common in tech docs to improve readability. Hmm, Chinese typesetting convention: no space between Chinese characters and adjacent digits/Latin, but many Chinese tech docs do insert spaces. Let me keep "第 9.2 节" with spaces, consistent with my other mixed-script text like "TP ≥ 4 的场景" (which has spaces).

Actually, let me reconsider. In my translations, I have "TP ≥ 4 的场景", "高 DP 场景", "P/D", "EAGLE3 投机解码". These mixed-script phrases with spaces around Latin terms are typical in Chinese tech writing. So "第 9.2 节" with spaces is fine.

Similarly, "C8 量化" - space between C8 and 量化. "EAGLE3 投机解码" - space. "M2.5/M2.7" - no Chinese adjacent. "MiniMax-M2 (M2.5/M2.7) 模型" - hmm, if I use fullwidth parens: "MiniMax-M2（M2.5/M2.7）模型" - there's no space between the closing paren and 模型. With fullwidth parens, no space is needed. With half-width, a space would be typical: "MiniMax-M2 (M2.5/M2.7) 模型". Let me decide: I'll use half-width parentheses to match the original exactly: "MiniMax-M2 (M2.5/M2.7) 模型". Hmm, but then I have a space before "(" which matches the original "(M2.5/M2.7)". Actually the original is "MiniMax-M2 (M2.5/M2.7)" with a space before "(". So I'll keep it exactly: "MiniMax-M2 (M2.5/M2.7) 模型". Wait, the original text: "MiniMax-M2 (M2.5/M2.7) model-specific issues." So the phrase in the original: "MiniMax-M2 (M2.5/M2.7)" with space before paren. Keep that. Then "模型" follows. So: "本章仅涵盖 MiniMax-M2 (M2.5/M2.7) 模型的专属问题。" That has a space between ")" and "模型". That's fine and consistent with mixed-script spacing.

Similarly for the "currently disabled by default (`false`)" - I'll keep half-width parens to match original: "当前默认关闭 (`false`)。" Hmm, but Chinese text with half-width parens: "当前默认关闭 (`false`)。" That's also acceptable. Actually, wait - the original: "Currently disabled by default (`false`)." I could translate as "当前默认关闭 (`false`)。" Hmm, mixing: after "关闭" a space then half-width paren. In Chinese tech docs, both styles appear. 

Let me make a decision: use half-width parentheses consistently when the parenthetical content is technical (contains code/numbers), and fullwidth when it's pure Chinese annotation. But honestly, for simplicity and to preserve original formatting, let me keep half-width parentheses everywhere in the translated text where the original had half-width parentheses, EXCEPT where I'm adding a Chinese parenthetical that didn't exist. For example, "（即思考过程和最终答案分离）" is my addition to explain "where the thinking process and final answer are separated". Since the original has "where ...", not a parenthetical. Hmm, actually the original has "(where the thinking process and final answer are separated)" - with half-width parens. So I should keep half-width: "(即思考过程和最终答案分离)". Let me keep half-width to match.

OK so decision: keep half-width parentheses as in the original, since these are within prose and match the original punctuation style.

So:
- "MiniMax-M2 (M2.5/M2.7) 模型"
- "当前默认关闭 (`false`)。"
- "(即思考过程和最终答案分离)"

Hmm, but for "当前默认关闭 (`false`)。", the half-width paren after Chinese char - it's fine: "当前默认关闭 (`false`)。"

Also "(90% 缓存命中)" in the table - original "(90% cache hit)" with half-width parens. Keep half-width: "(90% 缓存命中)". Good, I already planned that.

Now, let me reconsider "EAGLE3 投机解码" vs "EAGLE3 Speculative Decoding" in table 3 row name. The feature is named "EAGLE3 Speculative Decoding" in the original. If I translate to "EAGLE3 投机解码", it's a translated phrase with brand name kept. Fine.

Now, one more check: "Optimizations Enabled by Default" - I translated to "默认启用的优化". The anchor from original: "optimizations-enabled-by-default". Good.

"Optimizations That Require Explicit Enabling" -> "需要显式启用的优化". Anchor: "optimizations-that-require-explicit-enabling". Good.

Now, let me also verify I haven't missed any prose. Let me go through the original line by line:

1. ```shell ... ``` - code, keep. ✓
2. "## 8 Performance Evaluation" - heading, translated. ✓
3. "### 8.1 Using AISBench" - heading, translated. ✓
4. "Refer to [Using AISBench for performance evaluation](...) for details." - prose, translated. ✓
5. "### 8.2 Using vLLM Benchmark" - heading, translated. ✓
6. "Run performance evaluation for `MiniMax-M2.7-W8A8` as an example." - prose, translated. ✓
7. "Refer to [vllm benchmark](...) for more details." - prose, translated. ✓
8. "Take the `serve` subcommand as an example:" - prose, translated. ✓
9. ```shell ... ``` - code, keep. ✓
10. "## 9 Performance Tuning" - heading. ✓
11. "> **Note**: ..." - blockquote prose, translated. ✓
12. "### 9.1 Recommended Configurations" - heading. ✓
13. "The following configurations are validated in internal testing and are categorized by use case." - prose, translated. ✓
14. Table 1 - translated cells. ✓
15. "> **Note**: ..." - blockquote, translated. ✓
16. "### 9.2 Tuning Guidelines" - heading. ✓
17. "#### 9.2.1 General Tuning Reference" - heading. ✓
18. "Please refer to ... for general tuning methods." - prose, translated. ✓
19. "Please refer to ... for detailed feature descriptions." - prose, translated. ✓
20. "#### 9.2.2 Model-Specific Optimizations" - heading. ✓
21. "##### Optimizations Enabled by Default" - heading. ✓
22. "The following optimizations are enabled by default and require no additional configuration:" - prose, translated. ✓
23. Table 2 - translated cells. ✓
24. "##### Optimizations That Require Explicit Enabling" - heading. ✓
25. Table 3 - translated cells. ✓
26. "## 10 FAQ" - heading. ✓
27. "For common environment, installation, and general parameter issues, please refer to the [Public FAQs](../../faqs.md). This chapter only covers MiniMax-M2 (M2.5/M2.7) model-specific issues." - prose, translated. ✓
28. Three FAQ list items - translated. ✓

Everything covered.

Now, one more thought on "Performance Evaluation" heading translation. "性能评估" is standard. "性能评测" is also common. I'll use "性能评估".

"Performance Tuning" -> "性能调优". Standard. Good.

"Recommended Configurations" -> "推荐配置". Good.

"Tuning Guidelines" -> "调优指南". Good.

"General Tuning Reference" -> "通用调优参考". Good.

"Model-Specific Optimizations" -> "模型特定优化". Good.

"FAQ" -> "常见问题". Good.

Now, let me reconsider the first sentence of section 10: "For common environment, installation, and general parameter issues, please refer to the [Public FAQs](../../faqs.md)."

My translation: "有关常见的环境、安装和通用参数问题，请参阅[公共 FAQ](../../faqs.md)。"

Good. Hmm, "公共 FAQ" - maybe "通用 FAQ"? The word "Public" is used likely to mean a shared documentation (公共文档), as opposed to model-specific. "公共 FAQ" is fine.

Now, let me finalize. Let me also make sure I don't have any trailing issues like missing newlines between code blocks and headings. I'll ensure proper blank lines.

Also note: the output should not be wrapped in a code fence. So no ``` at the very beginning or end beyond those belonging to the original code blocks.

Let me write the final output carefully.

Wait, one more consideration: the original has "## 8 Performance Evaluation" immediately after the first code block. In my output, I need a blank line between the closing ``` of the code block and the heading. Actually, in the original there's a blank line. I'll include one.

Also, between "以 `serve` 子命令为例：" and the code block - there's a blank line. Keep.

Also, between the code block and "## 9 性能调优" - blank line. Keep.

Also, in section 9.1, between the intro sentence and the table - blank line. Keep.

Between the table and the note blockquote - blank line. Keep.

Between sections - blank lines. Keep.

Now, let me also make sure the list item indentation in the FAQ is preserved. The original:

The "A:" line is indented with two spaces. I'll keep two-space indent.

Let me write the final answer now.

Actually, let me reconsider two translation details once more:

1. "C8 quantization" - "C8 量化" - quantization is "量化". Good.

2. "the standard reasoning semantics of `/v1/responses`" - "`/v1/responses` 的标准 reasoning 语义". The word "reasoning" is a field/concept name here, not translated. Good.

3. "properly populated" - "被正确填充". Good.

4. "draft model" - "草稿模型". Good.

5. "cross-core scheduling overhead" - "跨核调度开销". Good.

6. "decoding graph" - "解码图". Good.

7. "schedule" - I used "调度" in various places. Good.

8. "Throughput" - "吞吐". Good.

9. "Low Latency" - "低时延". Good.

All right, I'm confident in the translation. Let me produce the final output.</think>```shell
lm_eval \
  --model local-completions \
  --model_args model=/path/to/weight/MiniMax-M2.7-w8a8-QuaRot,base_url=http://127.0.0.1:8000/v1/completions,tokenized_requests=False,trust_remote_code=True \
  --tasks gsm8k \
  --output_path ./

export VLLM_USE_MODELSCOPE=True
vllm bench serve \
  --model /path/to/weight/MiniMax-M2.7-w8a8-QuaRot \
  --dataset-name random \
  --random-input 200 \
  --num-prompts 200 \
  --request-rate 1 \
  --save-result \
  --result-dir ./