Instructions to use iFlytekOpenSource/Domux with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use iFlytekOpenSource/Domux with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="iFlytekOpenSource/Domux")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("iFlytekOpenSource/Domux")
model = AutoModelForMultimodalLM.from_pretrained("iFlytekOpenSource/Domux", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use iFlytekOpenSource/Domux with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "iFlytekOpenSource/Domux"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iFlytekOpenSource/Domux",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/iFlytekOpenSource/Domux

SGLang

How to use iFlytekOpenSource/Domux with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "iFlytekOpenSource/Domux" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iFlytekOpenSource/Domux",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "iFlytekOpenSource/Domux" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iFlytekOpenSource/Domux",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use iFlytekOpenSource/Domux with Docker Model Runner:
```
docker model run hf.co/iFlytekOpenSource/Domux
```

[HER Hack-Astron #4] Domux 全量 4057 条官方评测：复现 97.58% 准确率并分析失败模式

by gebibd - opened 3 days ago

Discussion

gebibd

3 days ago

发布到：https://huggingface.co/iFlytekOpenSource/Domux/discussions
发布后把讨论编号记下来（案例 frontmatter 的 channels 要用）。

一句话结论

在 RTX 4090D（AutoDL 付费云 GPU）上以 BF16 运行 Domux 官方评测数据集，4057 条智能家居指令测试集结果：格式合规 100.00% / 结果准确率 97.58% / Slot F1 99.60% / Intent F1 98.17% / 平均延迟 0.159s（batch 8 批处理），独立复现与模型卡宣称的 98.37% 仅差 0.79 个百分点。

任务说明

Domux 将自然语言智能家居指令转换为 动作|设备|属性|值|单位|房间|楼层 的 7 字段结构化输出。本次用官方 smart_home_control_test_set.jsonl（4057 条，4 个类别）做全量评测，指标定义逐字沿用官方 eval/run_eval.py，并分析失败样例的规律。

环境（如实填写）

项	值
硬件	RTX 4090D 24GB（AutoDL 按量付费云 GPU），Ubuntu
框架	transformers 5.15.1 + torch 2.13.0（本地批量推理，非 API 服务）
精度	BF16
模型 revision	`6c71a32f4d624cadfd9fce9d10240d8068e53456`
推理参数	greedy（等价 temperature=0.0），max_new_tokens=256，batch_size=8
指标来源	iflytek/domux `eval/run_eval.py`（逐字复用：格式/准确率/Slot F1/Intent F1 定义一致）
延迟口径	每 batch 墙钟时间 ÷ batch 大小，跳过前 5 条 warmup

复现步骤

export HF_HOME=/root/autodl-tmp/hf
hf download iFlytekOpenSource/Domux --revision 6c71a32f4d624cadfd9fce9d10240d8068e53456
# 本地脚本加载模型（BF16）+ 逐批生成 + 官方指标计算（eval_direct.py，指标函数与 run_eval.py 一致）
python eval_direct.py --batch 8

结果

类别	样本数	格式合规	准确率	Slot F1	Intent F1	延迟(s)
multi_intent	1641	100.00%	97.75%	99.68%	98.65%	0.231
non_standard_naming	559	100.00%	94.81%	99.09%	94.91%	0.118
omitted_attribute	735	100.00%	97.14%	99.28%	97.14%	0.111
single_intent	1122	100.00%	99.02%	99.82%	99.02%	0.105
OVERALL	4057	100.00%	97.58%	99.60%	98.17%	0.159

与模型卡宣称 98.37% 的差异（0.79pp）说明：独立复现存在微小解码差异（greedy 批处理、transformers 直连 vs 官方评测管线），但整体能力与宣称一致，Slot F1 99.60% 表明字段级解析非常稳定。

失败模式分析（98 条失败，2.42%）

类别	失败数/总数	失败率
multi_intent	37/1641	2.25%
non_standard_naming	29/559	5.19%
omitted_attribute	21/735	2.86%
single_intent	11/1122	0.98%

三类典型错误：

多指令楼层/房间继承丢失（multi_intent 最多）：如 on the ground floor set the chandelier in the majlis ... turn on the floor lamp，第二条指令漏填 Ground Floor，楼层只在该条开头声明一次时，后续指令的 floor 槽位偶发为 *。
属性混淆（non_standard_naming 失败率最高）：如 turn up the AC 被解析为 windSpeed 而非 temperature；make it warmer 的 colorTemperature 与 brightness 偶发互换。
房间推断边界：个别非标准命名场景下房间槽位给出具体房间名而 gold 为 *（过度推断）。

局限

只在官方测试集上评测；未见过的真实家庭指令（方言、ASR 噪声）未覆盖
未做量化/延迟优化对比（本案例为 BF16 全量）
高风险动作的拒绝/二次确认行为未单独深挖

安全与隐私

评测使用公开测试集，无个人/家庭隐私数据
未在命令、日志、截图泄露 HF token 或个人缓存路径

链接

模型：https://huggingface.co/iFlytekOpenSource/Domux
案例 PR：<提交后补>
活动 issue：https://github.com/iflytek/domux/issues/20

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment