BlazeSeq v0.4

BlazeSeq 0.4 targets Mojo 1.2.0 and MAX 26.7.0 and restores GPU batching to the core library.

Highlights

• Pin Mojo 1.2.0 and MAX 26.7.0 across Pixi environments, Mojo packages, and Python wheels.
• Restore FastqBatch.to_device() and .stage(), consuming StagedFastqBatch.to_device(), and device/staged batch types and transfer helpers in blazeseq and blazeseq.fastq.
• Publish the BED, GTF, and GFF parsers through stable package paths and top-level exports.
• Modernize hashing, immutable string views, C-string access, inline policies, and pattern matching.
• Build interpreter-specific Python wheels with relocatable runtime paths and test installation outside the source tree.
• Remove the experimental memory-mapped FASTQ parser, mapping I/O, and associated benchmark options.

Upgrading

• Install Mojo 1.2.0 and MAX 26.7.0. MAX is required for CPU and GPU consumers; CPU parsing does not require GPU hardware.
• Replace imports from blazeseq_gpu with imports from blazeseq or blazeseq.fastq. The separate GPU package and recipe have been removed.
• Use FastqParser with FileReader for plain FASTQ files. The experimental MmapFastqParser API has been removed.
• Precompiled Mojo packages use .mojoc. Python wheels require Linux glibc 2.34+ or Apple Silicon macOS 13+.
• RapidgzipReader currently uses the GZFile compatibility fallback and does not provide parallel rapidgzip decompression.

Validation

The release commit passed Mojo test and GPU cross-compilation CI, Linux and macOS wheel CI, and documentation deployment.

Local checks passed 548 CPU tests and five installed-wheel tests. No additional performance benchmarks were run for this release.

Full changelog: v0.3...v0.4

Release commit: 7cb7148b0a2fe5940cb8e611dda1077b4c003f32