refactor: share tensor writer across ggml converters

[PABannier]

What / why

convert_sam3_to_ggml.py, convert_sam2_to_ggml.py and convert_edgetam_to_ggml.py each had their own copy of write_tensor (the 32-byte-aligned tensor record writer) and of the FTYPE_F32/F16 constants. The copies differed only in which name substrings are kept as f32.

This PR moves that code into a new ggml_writer.py at the repo root. write_tensor(fout, name, data, ftype, keep_f32=KEEP_F32) takes the substrings as a parameter:

• sam2 uses the default KEEP_F32
• sam3 passes KEEP_F32 + ("freqs_cis",)
• edgetam passes KEEP_F32 + ("latents",)

"pos_embed" is no longer listed separately because "embed" already matches it. rename_key and write_header really do differ between the converters, so they stay where they are. The now-unused numpy import is removed from the sam2 and edgetam converters.

Verification

A throwaway script (not committed) pulled the old write_tensor and FTYPE constants out of each file at origin/main with ast. It compared them against the new shared function, called with the extra arguments read from each converter's real write_tensor(...) call. Every write went to a BytesIO and the bytes were compared. Test matrix:

• names: every keep-f32 substring, pos_embed, freqs_cis, latents, perceiver.latents_2d, normal .weight / .bias names
• shapes: 0-D, 1-D, 2-D, 3-D, 4-D, all-ones 4-D, and a non-contiguous transposed array
• dtypes: float32, float64, float16, int64
• ftype 0 and 1
• starting offsets 0, 1, 5, 12, 31, 32, 33, 63 (to exercise alignment)
• plus one multi-record stream per converter

Result: 20,160 cases (6,720 per converter) plus 3 streams, all byte-identical.

I also checked that every file parses with ast and passes py_compile. --help runs for all three converters when invoked by absolute path from a different cwd, with only numpy installed (torch is imported lazily inside main). That matches how scripts/download_model.sh calls them; the README uses uv run python convert_*.py from the repo root.

Diff

4 files changed, 26 insertions(+), 137 deletions(-)

Caveats

• The converters now depend on ggml_writer.py being in the same directory. That holds for every documented invocation (README, scripts/download_model.sh), but copying a single converter somewhere else on its own will no longer work.
• No real checkpoint was converted end to end; equivalence rests on the byte-compare above.

🤖 Generated with Claude Code

https://claude.ai/code/session_01M2qLmQ9ag49hT61V6i9qAq