Software downloads

Features

• Add V4.1 modified base models
• Add ability to load POD5/FAST5 files recursively, remove the runners flag
• Performance improvements to Stereo Basecalling with better batch size selection
• Upgrade to Pod5 v0.1.13 which allows loading large files on vmem-limited systems
• Incorporate improvements in Koi which give more predictive QV Scores
• Upgrade to Torch 2.0 and Cuda 11.8
• If a Stereo model is not available for duplex basecalling, it gets downloaded automatically
• Use jemalloc to override allocator on Linux
• Enable Stereo Duplex on M1
• Make metal stereo conv1 kernel available
• Make MetalModel capable of handling > 1 input channel (Required for Stereo Duplex)
• Improve Dorado startup time when using POD5 files by getting ReadGroup information from metadata tables
• Remove deprecated use of FindCUDA and show real location of found toolkit
• Improvements to Mk1C performance
• Add support for arbitrary messages in Dorado pipelines
• Addition of a progress bar for basecalling
• Refactoring of Dorado pipeline system to Add AsyncQueue and use it in ReadSink/Nodes
• Add "cuda:auto" as alternative to "cuda:all" when selecting a compute accelerator device on CLI
• Store reads raw data internally in fp16 format
• Switch post-ScalerNode raw_data to float16

Fixes

• Fix bug whereby the last read was ignored for read lists not terminated by a newline character
• Fixes to dorado duplex basespace
• Fix memory leak in getting read groups
• Fix meaning of TLEN in SAM output
• Fix issue which was occuring when a read had no mod base context hits