slorado

Slorado

Slorado is a simplified version of Dorado built on top of S/BLOW5 format. Slorado is an extremely lean basecaller with fewer external dependencies and is thus relatively easier to compile than Dorado. Slorado is developed using C/C++ and depends on torchlib. Currently, slorado only supports the Linux operating system (or Windows through WSL). slorado can utilise NVIDIA or AMD GPU accelerators on x86_64 CPUs. Slorado also works on ARM64-based NVIDIA Jetson devices.

Slorado is mainly for our research and educational purposes. Thus, only a minimal set of basecalling features are supported and may not be up-to-date with Dorado. For a feature-rich and up-to-date S/BLOW5-based basecaller for routine use on NVIDIA GPUs, please see buttery-eel or slow5-dorado.

GitHub Downloads slorado

Quick start

We provide compiled binaries for NVIDIA (cuda) and AMD (rocm) GPU accelerators on x86_64 CPUs for Linux. You can download the latest relevant binary release that includes the most recent supported basecalling models from releases as below:

VERSION=v0.5.0-beta
GPU=cuda   # GPU=rocm for AMD GPUs
wget "https://cdn.bioinf.science/slorado/slorado-$VERSION-x86_64-$GPU-linux-binaries.tar.xz"
tar xvf slorado-$VERSION-x86_64-$GPU-linux-binaries.tar.xz
cd slorado-$VERSION
./bin/slorado basecaller models/dna_r10.4.1_e8.2_400bps_hac@v5.0.0 reads.blow5  -o out.fastq -x cuda:all

Detailed instructions are available at:

Basecalling on Australia’s Pawsey supercomputer: Pawsey Guide

Binaries for the CPU-only version are not provided as basecalling on the CPU is impractically slow. Nevertheless, the CPU-only version is easier to build compared to the GPU version (see below).

Refer to troubleshoot for help resolving common problems.

Compilation and running

Compilation

Compilation instructions differ based on the system. Please pick one of the following that matches your system:

Running

We have tested slorado on a limited number of basecalling models listed below. You can download them using the provided script (the binary releases already include these):

scripts/download-models.sh

Now run on a test dataset:

# for CPU
./slorado basecaller -x cpu models/dna_r10.4.1_e8.2_400bps_fast@v5.0.0 test/5khz_r10/one_5khz.blow5 -o reads.fastq
# for GPU
./slorado basecaller -x cuda:all models/dna_r10.4.1_e8.2_400bps_fast@v5.0.0 test/5khz_r10/one_5khz.blow5 -o reads.fastq

Refer to troubleshoot for help resolving common problems.

Testing

After running on a test dataset, you can use minimap2 to align the reads to the reference and calculate the identity score statistics. If the identity score statistics are close enough to what we expect from these models, then things are good.

A script to calculate basecalling accuracy is provided:

set environment variable MINIMAP2, if minimap2 is not in PATH.
scripts/calculate_basecalling_accuarcy.sh hg38noAlt.fa reads.fastq

For a more exhaustive test of slorado’s features (on GPU setups), we have provided an extensive test script. This will automatically download the requisite test data and tools to test DNA/RNA basecalling, methylation detection, and flash attention support on your device. We highly recommend running this to ensure basecalling works on your machine. Excluding the automated binary release test mode, this script is meant to work on both ARM and x86 architectures.

Here is an example of how to run it:

# optionally customise parameters before running
export FAST_BATCH=512   # FAST model GPU batch size
export HAC_BATCH=256    # HAC model GPU batch size
export SUP_BATCH=128    # SUP model GPU batch size
export NTHREADS=8       # number of CPU threads (set to _NPROCESSORS_ONLN if unspecified)
export READ_MEM=512M    # max read batch memory in host memory
export READ_BATCH=2048  # max number of reads loaded into host memory

# go into slorado root directory
cd slorado

# test an existing slorado binary by providing the path
./test/extensive /path/to/slorado

# test the latest binary (x86) release on your machine
./test/extensive cuda bin
# OR
./test/extensive rocm bin

# build and test from the repo (after installing the appropriate torch version)
./test/extensive cuda build
# OR
./test/extensive rocm build

Known issues

As of May 1st 2026, LSTM models (HAC and FAST or SUP < v5.0.0) on the 9700 AI Pro (and possibly other newer AMD GPUs) produce incorrect outputs (Transformer models unaffected). This issue is known and can be tracked here.

Demultiplexing

Slorado does not currently support demultiplexing. You can demultiplex reads generated by Slorado by passing them into Dorado. This will work regardless of the device since demultiplexing occurs on the CPU.

# basecall reads
./slorado basecaller models/dna_r10.4.1_e8.2_400bps_fast@v5.0.0 reads.blow5 -o reads.fastq

# demux reads
./dorado demux --kit-name <kit-name> --output-dir demux_reads/ reads.fastq

Modification Detection (experimental)

Slorado (from v0.5.0-beta) supports methylation detection for GPU basecalling for HAC v5.0.0 and SUP v5.0.0 DNA basecalling models. Enable methylation detection by appending --mod 5mCG_5hmCG@v3 when running slorado. Adding modification detection will automatically output in SAM format.

# example modification calling
./slorado basecaller models/dna_r10.4.1_e8.2_400bps_hac@v5.0.0 reads.blow5 --mod 5mCG_5hmCG@v3 -xcuda:all -o reads.fastq

Options

All options supported by slorado basecaller are detailed below:

Option: Decription: Default Value:  
-t INT number of processing threads 8  
-K INT batch size (max number of reads loaded at once) 4096  
-C INT gpu batch size (max number of chunks loaded at once) 512  
-B FLOAT[K/M/G] max number of bytes loaded at once 512M  
-o FILE output to file stdout  
-c INT chunk size 12288  
-p INT overlap 150  
-x DEVICE specify device (e.g., cpu; cuda:0; cuda:1,2; cuda:all) cuda:all (GPU build) or cpu (CPU build)  
-h shows help message and exits -  
–verbose INT verbosity level 4  
–version print version    
–flash yes no enable flash attention (from v0.4.0-beta) No
–mod STR add modification detection (from v0.5.0-beta) NULL  

Batchsizes

A large batch size (-K and -B) may take up significant RAM during run-time. Similarly, your GPU batch size (-C) will determine how much GPU memory is used. Slorado currently does not implement automatic batch size selection based on available memory. Thus, if you see an out-of-RAM error, reduce the batch size using -K or -B. If you see an out-of-GPU memory error, reduce the GPU batch size using the -C option.

Flash Attention

Slorado v0.4.0-beta now supports Flash Attention for SUP basecalling models >= v5.0.0 when compiled with CUDA Torch >= v2.4.0 and ROCm Torch >= 2.9.0. This is not guaranteed to work on older GPUs, so we have kept it disabled by default for maximum compatibility. For best runtime performance on modern GPUs (Ampere GPUs or newer on NVIDIA, CDNA2/RDNA3 or newer on AMD), enable Flash Attention with the option --flash yes. Other older GPUs maybe supported but are not tested yet.

Tested models

slorado version Tested models
0.5.0-beta dna_r10.4.1_e8.2_400bps v5.0.0; dna_r10.4.1_e8.2_400bps_5mCG_5hmCG@v3 v5.0.0; rna004_130bps v5.1.0
0.4.0-beta dna_r10.4.1_e8.2_400bps v5.0.0; rna004_130bps v5.1.0
0.3.0-beta dna_r10.4.1_e8.2_400bps v4.2.0 and v5.0.0
0.2.0-beta dna_r10.4.1_e8.2_400bps v4.2.0

Acknowledgement

Citation:

Please cite the following in your publications when using Slorado:

Wong, B., Singh, G., Javaid, H., Denolf, K., Liyanage, K., Samarakoon, H., Deveson, I.W. and Gamaarachchi, H., 2026. Open-source, Hardware-Independent GPU Acceleration for Scalable Nanopore Basecalling with Slorado and Openfish. bioRxiv, pp.2026-03.

@article{wong2026open,
  title={Open-source, Hardware-Independent GPU Acceleration for Scalable Nanopore Basecalling with Slorado and Openfish},
  author={Wong, Bonson and Singh, Gagandeep and Javaid, Haris and Denolf, Kristof and Liyanage, Kisaru and Samarakoon, Hiruna and Deveson, Ira W and Gamaarachchi, Hasindu},
  journal={bioRxiv},
  pages={2026--03},
  year={2026},
  publisher={Cold Spring Harbor Laboratory}
}
This site is open source. Improve this page.