v
v_50/gtca-alphagenome_pytorch-NPU
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

alphagenome pytorch 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 PASS,cosine=)

A PyTorch port of AlphaGenome, the DNA sequence model from Google DeepMind that predicts hundreds of genomic tracks at single base-pair resolution

1. 模型简介

属性值
模型名gtca/alphagenome_pytorch
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
官方实现https://github.com/genomicsxai/alphagenome-pytorch

参考

官方实现 / 论文仓库:

  • https://github.com/genomicsxai/alphagenome-pytorch
  • https://github.com/google-deepmind/alphagenome_research

2. 环境依赖

依赖项版本说明
Python3.12conda 环境 gtca-alphagenome_pytorch-npu
torch / torch_npu2.10.0 / 2.10.0由 setup_env.sh 装,须严格配套
numpy==1.26.4
safetensors==0.8.0
requests==2.34.2
pyyaml>=6.0
alphagenome-pytorch==0.3.1 # PyPI 官方包(genomicsxai port,from_pretrained 原生支持 .safetensors)
decoratorlatest
absl-pylatest
scipylatest
attrslatest
psutillatest

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('gtca/alphagenome_pytorch', cache_dir='./weights')"

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

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:3,物理卡 7)

3. 分步推理操作流程(先建环境,再跑推理)

模型权重预置(评委裸机须先准备):推理从本地 WEIGHTS_ROOT(默认 /data/model-agent/weights/phase4/gtca/alphagenome_pytorch,见 inference.py)加载。请先从 https://ai.gitcode.com/hf_mirrors/gtca/alphagenome_pytorch 下载权重到对应目录,或 export WEIGHTS_ROOT=<你的权重路径>。 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 gtca-alphagenome_pytorch-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("gtca/alphagenome_pytorch", trust_remote_code=True).to("npu:0").eval()
tok = AutoTokenizer.from_pretrained("gtca/alphagenome_pytorch", 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: gtca/alphagenome_pytorch (AlphaGenome DNA 调控) on Ascend NPU ===
权重: /data/model-agent/weights/phase4/gtca/alphagenome_pytorch | AlphaGenome 参数 450.5M
输入: one-hot DNA (1,131072,4) | 131072bp = 128K
输出 atac[1bp]: shape (1, 131072, 256) mean=0.0156 std=0.0384
输出 atac[128bp]: shape (1, 1024, 256) mean=2.5878 std=2.4425
输出 dnase[1bp]: shape (1, 131072, 384) mean=0.0267 std=0.0434
输出 dnase[128bp]: shape (1, 1024, 384) mean=3.7002 std=3.7325
输出 cage[1bp]: shape (1, 131072, 640) mean=0.0047 std=0.0150
输出 cage[128bp]: shape (1, 1024, 640) mean=0.8110 std=1.1456
输出 rna_seq[1bp]: shape (1, 131072, 768) mean=0.0009 std=0.0025
输出 rna_seq[128bp]: shape (1, 1024, 768) mean=0.2450 std=0.7932
输出 contact_maps: shape (1, 64, 64, 28) mean=0.4039
推理耗时: 0.87 s/次(128K 全头 predict)
总耗时 16.9s
=== 完成 ===

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

=== Accuracy: gtca/alphagenome_pytorch (AlphaGenome DNA 调控) ===
输出头一致: True(11 头)
ATAC[128bp] cosine(CPU vs NPU, 16384bp): 1.0000
dnase[128bp] mean CPU=4.2505 NPU=4.2504
cage[128bp] mean CPU=1.5073 NPU=1.5072
rna_seq[128bp] mean CPU=0.4651 NPU=0.4651
结果 PASS

5. CPU vs NPU 精度对比(仅记录,无硬约束)

指标数值
max_abs_errorN/A
relative_errorN/A
cosine_similarityN/A

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 latency0.85 s/次(131072bp 全头 predict)
p50N/A
throughput0.155 Mbp/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/gtca-alphagenome_pytorch-NPU 适配日期:2026-08-12 | 赛道:第二季·模型适配·阶段三