Gap
write_geotiff_gpu documents four compression= modes (zstd, deflate, jpeg, none), but the test suite only round-trips deflate and none:
zstd is the default and is documented as "fastest on GPU". A handful of tests hit it implicitly by omitting compression=, but none assert pixel fidelity, so a regression in the nvCOMP zstd encoder or the writer's zstd codec-tag wiring would not surface.jpeg routes to _nvjpeg_batch_encode with a CPU Pillow fallback. Neither branch is exercised through the public API anywhere in the suite. to_geotiff(compression='jpeg') rejects on CPU, so write_geotiff_gpu is the only path to the GPU JPEG encoder.
Discovered during the test-coverage sweep on the geotiff module (Cat 4 HIGH parameter coverage).
Proposed fix
Add a targeted test module covering all four documented modes plus a TIFF compression-tag header check, exercised on a real GPU host.
Gap
write_geotiff_gpudocuments fourcompression=modes (zstd,deflate,jpeg,none), but the test suite only round-tripsdeflateandnone:zstdis the default and is documented as "fastest on GPU". A handful of tests hit it implicitly by omittingcompression=, but none assert pixel fidelity, so a regression in the nvCOMP zstd encoder or the writer's zstd codec-tag wiring would not surface.jpegroutes to_nvjpeg_batch_encodewith a CPU Pillow fallback. Neither branch is exercised through the public API anywhere in the suite.to_geotiff(compression='jpeg')rejects on CPU, sowrite_geotiff_gpuis the only path to the GPU JPEG encoder.Discovered during the test-coverage sweep on the
geotiffmodule (Cat 4 HIGH parameter coverage).Proposed fix
Add a targeted test module covering all four documented modes plus a TIFF compression-tag header check, exercised on a real GPU host.