v
v_50/Shadow0482-mythos_fast-NPU
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

mythos fast 昇腾NPU部署文档

硬件:Ascend NPU(Ascend 910B4) · 推理引擎:torch_npu · 设备:npu:0 精度验收:CPU vs NPU 推理精度差异 < 1%(实测 FAIL,cosine=0.4562080204486847)

Model Description mythos_fast is a fine-tuned version of WeiboAI/VibeThinker-3B, adapted for custom tool-use and agentic task execution. The base m

1. 模型简介

属性值
模型名Shadow0482/mythos_fast
原始权重ModelScope / HuggingFace / AtomGit 镜像
任务类型feature-extraction
输入见 inference.py
输出特征向量
框架PyTorch + torch_npu

2. 环境依赖

依赖项版本说明
Python3.11conda 环境 shadow0482-mythos_fast-npu
torch / torch_npu2.10.0 / 2.10.0由 setup_env.sh 装,须严格配套
transformers==4.51.3
accelerate==1.14.0
pillow==12.3.0
numpy==1.26.4
huggingface-hub==0.36.2
requests==2.34.2
modelscope==1.37.0
safetensors==0.8.0
multimolecule==0.2.1
pyyaml>=6.0
einops==0.8.0 # transformers 4.51 modeling 隐式依赖(torch_npu 存在时触发)
gguf==0.10.0 # transformers gguf_file 加载 GGUF 权重所需
sentencepiece==0.2.0 # GGUF tokenizer 反量化依赖

一键建环境(从零创建 conda 环境并装依赖):

bash setup_env.sh    # conda create -n <env> python=<ver> + torch/torch_npu + requirements.txt

⚠️ 重要:若环境已装 torch/torch_npu,不要强制覆盖 torch,否则破坏 torch_npu 算子插件绑定。

国内 pip 源(Tsinghua):https://pypi.tuna.tsinghua.edu.cn/simple

权重提前下载(国内源)

权重不随仓库分发,推理前先用国内下载源拉取到本地缓存(推荐 ModelScope,失败回落 HF 镜像):

# 方式一:ModelScope(国内直连,最快)
pip install modelscope -i https://pypi.tuna.tsinghua.edu.cn/simple
python -c "from modelscope import snapshot_download; snapshot_download('Shadow0482/mythos_fast', cache_dir='./weights')"

# 方式二:HF 镜像(hf-mirror.com)
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download Shadow0482/mythos_fast --local-dir ./weights/mythos_fast

inference.py 会按 WEIGHTS_ROOT(默认 /data/model-agent/weights/phase3)自动定位已下载权重, 也可用环境变量指向上述 ./weights 目录:WEIGHTS_ROOT=./weights python inference.py

硬件:单卡 Ascend 910B4(npu:0,物理卡 4)

3. 分步推理操作流程(先建环境,再跑推理)

评委裸 clone 后本机没有 conda 环境,conda activate 不会创建环境。必须先建环境再推理:

# 第一步:从零创建 conda 环境并装依赖(conda create + torch/torch_npu + requirements.txt)
bash setup_env.sh

# 第二步:激活环境并一键推理(环境自检 + inference + eval + benchmark)
conda activate shadow0482-mythos_fast-npu
bash run.sh

推理示例(真实输入,可复现)

import torch, torch_npu
from transformers import AutoModel, AutoTokenizer
m = AutoModel.from_pretrained("Shadow0482/mythos_fast", trust_remote_code=True).to("npu:0").eval()
tok = AutoTokenizer.from_pretrained("Shadow0482/mythos_fast", trust_remote_code=True)
inputs = tok("How to use this model?", return_tensors="pt").to("npu:0")
outputs = m(**inputs)

4. 测试样例及输出结果

样例 1:NPU 推理(完整 logs/inference.log)

输入:示例序列/文本(见 inference.py)

输出:

=== Inference: Shadow0482/mythos_fast on Ascend NPU ===
NPU: npu:0
加载方式: gguf_AutoModelForCausalLM
推理耗时: 410.20 ms
输出 shape: (1, 32, 151936)
输出 L2 norm: 4142.6382
输出前 8 维: [2.304946184158325, 0.0361722968518734, 3.3091559410095215, 1.3052875995635986, 3.4735491275787354, 4.825121879577637, 1.938745379447937, 1.5946890115737915]

样例 2:CPU vs NPU 精度验证(完整 logs/accuracy.log)

=== CPU-NPU Accuracy: Shadow0482/mythos_fast ===
max_abs_error     15.908494
mean_abs_error    2.263478
cosine_similarity 0.456208
结果 FAIL(阈值 cosine_similarity >= 0.98)

5. CPU vs NPU 精度对比(仅记录,无硬约束)

指标数值
max_abs_error15.908494
relative_errorN/A
cosine_similarity0.456208

注:本次适配无精度要求,精度数据作为质量证据。详见 logs/accuracy.log。

6. NPU 推理性能

指标数值
avg latency67.65 ms
p5067.48 ms
throughput14.78 actions/sec

7. Agent 适配截图(assets/)

7.1 Agent 适配全过程

Agent 适配流程

7.2 NPU 设备调用

NPU 设备调用

7.3 模型适配结果

模型适配结果

截图由真实执行日志(logs/)忠实渲染为终端风格 PNG,渲染脚本 tools/gen_screenshots_pw.py(playwright),可复现,无虚构结果。

8. 标签

#NPU #Ascend #Ascend910 #model-agent-tagged


仓库:https://gitcode.com/v_50/Shadow0482-mythos_fast-NPU 适配日期:2026-08-12 | 赛道:第二季·模型适配·阶段三