01
模型推荐

pip install git+https://github.com/huggingface/transformers.git

import torch
from modelscope import AutoModelForCausalLM, AutoTokenizer
MODEL_PATH = "ZhipuAI/GLM-4.7-Flash"
messages = [{"role": "user", "content": "hello"}]
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
)
model = AutoModelForCausalLM.from_pretrained(
    pretrained_model_name_or_path=MODEL_PATH,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
inputs = inputs.to(model.device)
generated_ids = model.generate(**inputs, max_new_tokens=128, do_sample=False)
output_text = tokenizer.decode(generated_ids[0][inputs.input_ids.shape[1]:])
print(output_text)

# 创建独立环境（推荐）
conda create -n qwen3-tts python=3.12 -y
conda activate qwen3-tts
# 安装核心包
pip install -U qwen-tts
# 可选：加速推理（需兼容硬件）
pip install -U flash-attn --no-build-isolation

step2：推理生成语言

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel
# 加载模型（自动下载，也可本地指定路径）
model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2"
)
# 生成！
wavs, sr = model.generate_custom_voice(
    text="其实我真的有发现，我是一个特别善于观察别人情绪的人。",
    language="Chinese",
    speaker="Vivian",
    instruct="用特别愤怒的语气说"
)
# 保存
sf.write("output.wav", wavs[0], sr)

step3：Web界面（可选）

qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --port 8000
# 浏览器打开 http://localhost:8000

pip install transformers torch

• 加载模型

import torch
from modelscope import AutoModelForCausalLM, AutoProcessor
model_id = "FlashLabs/Chroma-4B" # Or local path
# Load model
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    trust_remote_code=True, 
    device_map="auto"
)
# Load processor
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

• 推理示例

import torch
from IPython.display import Audio
# Construct conversation history
system_prompt = (
    "You are Chroma, an advanced virtual human created by the FlashLabs. "
    "You possess the ability to understand auditory inputs and generate both text and speech."
)
conversation = [[
    {
        "role": "system",
        "content": [
            {"type": "text", "text": system_prompt}
        ],
    },
    {
        "role": "user",
        "content": [
            # Input audio file path
            {"type": "audio", "audio": "assets/make_taco.wav"}, 
        ],
    },
]]
# Provide reference audio/text for style or context
prompt_text = ["War and bloodshed throughout the world."]
prompt_audio = ["assets/reference_audio.wav"]
# Process inputs
inputs = processor(
    conversation,
    add_generation_prompt=True, 
    tokenize=False,
    prompt_audio=prompt_audio,
    prompt_text=prompt_text
)
# Move inputs to device
device = model.device
inputs = {k: v.to(device) for k, v in inputs.items()}
# 2. Generate
output = model.generate(
    **inputs, 
    max_new_tokens=100, 
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    use_cache=True
)
# 3. Decode Audio
# The model outputs raw tokens; we decode the audio part using the codec
audio_values = model.codec_model.decode(output.permute(0, 2, 1)).audio_values
# Save or play audio (e.g., in Jupyter)
Audio(audio_values[0].cpu().detach().numpy(), rate=24_000)

from modelscope import AutoProcessor, AutoModelForCausalLM
key_mapping = {
    "^vision_model": "model.vision_model",
    r"^model(?!\.(language_model|vision_model))": "model.language_model",
    "vit_large_projector": "model.vit_large_projector",
}
model_path = "stepfun-ai/Step3-VL-10B"
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
            {"type": "text", "text": "What's in this picture?"}
        ]
    },
]
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    device_map="auto",
    torch_dtype="auto",
    key_mapping=key_mapping).eval()
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt"
).to(model.device)
generate_ids = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
decoded = processor.decode(generate_ids[0, inputs["input_ids"].shape[-1] :], skip_special_tokens=True)
print(decoded)

uv pip install git+https://github.com/huggingface/transformers
uv pip install pillow pypdfium2

import torch
from modelscope import LightOnOcrForConditionalGeneration, LightOnOcrProcessor


device = "mps" if torch.backends.mps.is_available() else "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float32 if device == "mps" else torch.bfloat16


model = LightOnOcrForConditionalGeneration.from_pretrained("lightonai/LightOnOCR-2-1B-bbox", torch_dtype=dtype).to(device)
processor = LightOnOcrProcessor.from_pretrained("lightonai/LightOnOCR-2-1B-bbox")


url = "https://huggingface.co/datasets/hf-internal-testing/fixtures_ocr/resolve/main/SROIE-receipt.jpeg"


conversation = [{"role": "user", "content": [{"type": "image", "url": url}]}]


inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
)
inputs = {k: v.to(device=device, dtype=dtype) if v.is_floating_point() else v.to(device) for k, v in inputs.items()}


output_ids = model.generate(**inputs, max_new_tokens=1024)
generated_ids = output_ids[0, inputs["input_ids"].shape[1]:]
output_text = processor.decode(generated_ids, skip_special_tokens=True)
print(output_text)

02
数据集推荐

OWL-SFT
TOWL-Planner-SFT 是由 Camel AI 开发的一个基于 OWL-ViT 视觉语言模型微调得到的指令驱动视觉规划模型，专为具身智能（embodied AI）和机器人任务设计；该模型接收自然语言指令和当前环境图像作为输入，输出可执行的动作序列或目标位置（如“去厨房拿杯子” → 生成导航路径或抓取点）；其核心参数包括基于 ViT-L/14 的视觉编码器、语言引导的目标检测头，以及在大规模仿真数据（如 ALFRED、VirtualHome）上进行监督微调（SFT）得到的策略模块；适用于家庭服务机器人、虚拟智能体的任务规划、视觉导航与人机协作等场景。
数据集链接
https://modelscope.cn/models/camel-ai/OWL-Planner-SFT
03
创空间




Step3-VL-10B-Demo


Step3-VL-10B 适用于需要图文联合理解的场景，如智能客服、教育辅助、内容审核、无障碍图像描述和多模态人机交互等。
体验链接：
https://modelscope.cn/studios/stepfun-ai/step3-vl-10b

Qwen3-TTS
Qwen3-TTS 是一款文本到语音合成模型，适用于智能语音助手、有声读物生成、语音播报、无障碍阅读等需要高质量语音输出的场景。
体验链接：
https://modelscope.cn/studios/Qwen/Qwen3-TTS

HeartMuLa歌曲生成
                 HeartMuLa 是一款基于 AI 的歌曲生成应用，用户可通过输入歌词或旋律提示，自动生成风格多样的音乐作品，适用于音乐创作辅助、短视频配乐、个性化铃声制作等场景。


                   
 

                  体验链接：
https://modelscope.cn/studios/gswyhq/HeartMuLa