Roadmap¶

torchfits is FITS I/O for PyTorch tensors and dataframes. This page is forward-looking only.

Current focus (toward 1.0.0)¶

1.0.0rc4 is the collaborator soak tag (pip install "torchfits==1.0.0rc4"). Prep / deep-review cleanup (hard removals, (row, label) table datasets, BITPIX int8/int64 write, header/prefetch fixes) ships in that cut. SemVer 1.0.0 waits for post-rc feedback. Larger scope cuts (cache merge, dataset zoo, CLI trim) stay deferred.

Parity tiers¶

Near-term (through 1.0.0)¶

• Republish 1.0.0rc4 wheels with torch 2.10 ABI pin in cibuildwheel (tag retarget after wheel workflow fix); soak with collaborators
• Fold soak feedback before SemVer 1.0.0
• Benchmark honesty: tensor vs table domains, CPU↔GPU deficits visible, CFITSIO-direct in the exhaustive table, MegaCam multi-cutout suite
• CLI depth: HIERARCH keys, batch header edit, honest verify scope, real RGB demos
• Publish exhaustive CSVs with release artifacts for user analysis

Deferred (cosmetic / low priority)¶

From the rc1 static audits (not blocking rc4):

• F1–F5 stylistic / macOS sysctl / CTYPE5 / END HISTORY notes
• C3 fold insert_rows+update_rows into one CFITSIO open (perf only)
• E5–E6 bench CSV cosmetics

Deferred from prep review (not 1.0 blockers)¶

• M1 dtype-validation dedup across write_api / mutation
• M4 HDUList.fromfile fallback loop cleanup; M6/M7 cache naming / test isolation
• L2 Header _mutate() helper; L4 WHERE LRU sizing for dynamic literals
• T1–T5/T7 test-harness polish; E2/E5/E6 example assertion depth
• B1 lambda default-binding in benches; B4/B6–B8 bench suite consolidation
• R2-10/R2-11 CLI transform / private-import nits; S6 cross-import comment

1.1 candidates¶

• Disk→GPU path — today CFITSIO delivers host memory; a direct GPU path would need GPUDirect Storage, a CFITSIO change, or a different backend. Investigation only until there is a clear PyTorch-native API and tests.
• Richer compressed write coverage and public-file replay snapshots
• Broader Astropy table / ASCII / VLA cases where they stay small and tested

CFITSIO (from 2026-07 CFITSIO audit leftovers)¶

• fits_get_bcolparms in TableReader::analyze_table (cold wide-table metadata)
• Expose quantize / HCOMPRESS / noise-bits / dither setters on write + convert CLI
• fits_flush_file between large multi-HDU writes (bound CFITSIO write buffers)
• fits_open_data for empty-primary MEF cold opens (marginal vs Python auto-HDU)
• fits_set_bufsize for CompImage decompression fallbacks
• Virtual image section / tile getters — only if compressed-mosaic cutout deficit measured
• fits_img_stats_* for CLI CPU stats; fits_read_pixnull* simplification
• Optional thread_local fitsfile* pool if per-read open cost shows up in scorecards

Permanent design choices (not gaps)¶

• VLA and scaled-column mmap updates stay on the buffered path (format / safety).
• GPU writes stay host-copy through CFITSIO unless a 1.1 backend lands.
• PyArrow is the table runtime; Polars/Pandas/DuckDB remain optional.
• Concurrent reads use private fitsfile* handles (CFITSIO R2); do not reintroduce a shared-handle LRU across threads (option C — serialize on meta->mutex — is also rejected).
• No CFITSIO fits_iterate_data, fits_calculator / fits_select_rows, histogram binning, hierarchical grouping, table compress, WCS rebin, pixel filter, IRAF delete, template exec, or CFITSIO HTTPS/stream drivers (own HTTP cache after the sh:// security fix).
• fits_set_huge_hdu / fits_get_eqcoltype / fits_decode_chksum / fits_translate_keywords stay unused unless a concrete user need appears.