dorado-bench-v2

majiayu000 的头像作者 majiayu000

在密歇根大学HPC集群(ARMIS2和Great Lakes)上使用Dorado进行Oxford Nanopore基调用。适用于运行Dorado基调用、生成SLURM作业、基准测试模型、优化GPU资源或处理POD5数据。捕获模型路径、GPU分配和作业元数据。与ont-experiments集成以进行溯源跟踪。支持fast/hac/sup模型、甲基化调用和自动资源计算。

📖 中文导读:在密歇根大学HPC集群(ARMIS2和Great Lakes)上使用Dorado进行Oxford Nanopore基调用。适用于运行Dorado基调用、生成SLURM作业、基准测试模型、优化GPU资源或处理POD5数据。捕获模型路径、GPU分配和作业元数据。与ont-experiments集成以进行溯源跟踪。支持fast/hac/sup模型、甲基化调用和自动资源计算。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: dorado-bench-v2 description: Oxford Nanopore basecalling with Dorado on University of Michigan HPC clusters (ARMIS2 and Great Lakes). Use when running dorado basecalling, generating SLURM jobs for basecalling, benchmarking models, optimizing GPU resources, or processing POD5 data. Captures model paths, GPU allocations, and job metadata. Integrates with ont-experiments for provenance tracking. Supports fast/hac/sup models, methylation calling, and automatic resource calculation.

Dorado-Bench v2 - ONT Basecalling

Basecalling toolkit for UM HPC clusters with provenance tracking.

Integration

Run through ont-experiments for provenance tracking:

ont_experiments.py run basecalling exp-abc123 --model sup --output calls.bam --json stats.json

Or standalone:

python3 dorado_basecall.py /path/to/pod5 --model sup --cluster armis2 --output calls.bam

Cluster Configurations

ARMIS2 (sigbio-a40)

partition: sigbio-a40
account: bleu1
gres: gpu:a40:1
dorado: /nfs/turbo/umms-bleu-secure/programs/dorado-1.1.1-linux-x64/bin/dorado
models: /nfs/turbo/umms-bleu-secure/programs/dorado_models

Great Lakes (gpu_mig40)

partition: gpu_mig40
account: bleu99
gres: gpu:nvidia_a100_80gb_pcie_3g.40gb:1

Model Tiers

Tier Accuracy ARMIS2 Resources
fast ~95% batch=4096, mem=50G, 24h
hac ~98% batch=2048, mem=75G, 72h
sup ~99% batch=1024, mem=100G, 144h

Options

Option Description
--model TIER fast, hac, sup (default: hac)
--version VER Model version (default: v5.0.0)
--cluster armis2 or greatlakes
--output FILE Output BAM file
--json FILE Output JSON statistics
--slurm FILE Generate SLURM script
--emit-moves Include move table
--modifications MOD Enable 5mCG_5hmCG methylation

SLURM Generation

python3 dorado_basecall.py /path/to/pod5 \
  --model sup@v5.0.0 \
  --cluster armis2 \
  --slurm job.sbatch

sbatch job.sbatch

Event Tracking

When run through ont-experiments, captures:

  • Model name and full path
  • Model tier/version/chemistry
  • Batch size and device
  • BAM statistics (reads, qscore, N50)
  • SLURM job ID, nodes, GPUs

Methylation Calling

ont_experiments.py run basecalling exp-abc123 \
  --model sup \
  --modifications 5mCG_5hmCG \
  --output calls_5mc.bam

Resources adjusted: memory +50%, batch size -30%