82 lines
2.0 KiB
Markdown
82 lines
2.0 KiB
Markdown
---
|
||
language:
|
||
- zh
|
||
license: apache-2.0
|
||
library_name: transformers
|
||
tags:
|
||
- minimind
|
||
- qwen3-moe
|
||
- moe
|
||
- chat
|
||
- text-generation
|
||
- ascend
|
||
pipeline_tag: text-generation
|
||
---
|
||
|
||
# MiniMind3-Ascend-MoE
|
||
|
||
这是一个基于 **MiniMind3-Ascend** 训练链路导出的 MoE 对话模型，兼容 Transformers 推理方式，适合追求更高容量与更强表达能力的场景。
|
||
|
||
## 模型信息
|
||
|
||
- 模型名：`fzkun/minimind3-ascend-moe`
|
||
- 架构：MoE
|
||
- 导出兼容：`Qwen3MoeForCausalLM`
|
||
- 参数规模：约 **198M**
|
||
- 激活参数：约 **64M**
|
||
- 主要配置：
|
||
- `hidden_size = 768`
|
||
- `num_hidden_layers = 8`
|
||
- `num_experts = 4`
|
||
- `num_experts_per_tok = 1`
|
||
|
||
## 文件说明
|
||
|
||
仓库中包含：
|
||
|
||
- `config.json`
|
||
- `generation_config.json`
|
||
- `model.safetensors`
|
||
- `tokenizer.json`
|
||
- `tokenizer_config.json`
|
||
- `special_tokens_map.json`
|
||
- `chat_template.jinja`
|
||
|
||
## 使用方式
|
||
|
||
### Transformers
|
||
|
||
```python
|
||
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||
|
||
model_id = "fzkun/minimind3-ascend-moe"
|
||
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
|
||
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
|
||
|
||
messages = [{"role": "user", "content": "你好"}]
|
||
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
||
inputs = tokenizer(text, return_tensors="pt")
|
||
out = model.generate(**inputs, max_new_tokens=256)
|
||
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
|
||
```
|
||
|
||
## Benchmark 结果
|
||
|
||
评测环境：
|
||
- Ascend 910B
|
||
- 单卡 `npu:0`
|
||
- `batch_size = 16`
|
||
|
||
| ceval | cmmlu | arc | piqa | openbookqa | hellaswag | siqa |
|
||
|---:|---:|---:|---:|---:|---:|---:|
|
||
| 23.77 | 24.88 | 30.30 | 51.63 | 26.00 | 28.58 | 34.08 |
|
||
|
||
说明：
|
||
- `ceval / cmmlu / arc / piqa / openbookqa / hellaswag` 使用 `acc_norm`
|
||
- `social_iqa` 使用 `acc`
|
||
|
||
## 补充说明
|
||
|
||
- MoE 版本容量更高，在部分 benchmark 上表现更强
|
||
- 对应 SwanLab 实验记录：<https://swanlab.cn/@fzkun/MiniMind3/overview>
|