julianmb/Ornith-1.5-35B-A3B-ROCmFP4-GGUF

Record citation consent for the upstream draft

Scope the draft-path direct seq_rm calls as performance-only

Port upstream #27530 (failed-restore cleanup); correct two of my own claims

Add candidate fix for #29092: clear GDN recurrent state when a cell is released

Rewrite upstream draft with the real sweep data and the negative results

Withdraw the strict-qwen mechanism claim; add depth sweep and prompt-scaling data

Ignore __pycache__ (verify_card.py is imported by the NOMPT.md snippet)

Retract the upstream draft; retarget it to #29092

RETRACT the 'disable MTP' advice; measurement shows the opposite

Add no-MTP conversion recipe and a ready-to-post upstream datapoint

Add verify_card.py: machine-check the card against the artifact, in CI

README: correct the '16.7% smaller' figure to 12.3%

Fix inverted SHA256: the card was right, SHA256SUMS was stale

README: fix stale SHA256 + file size, correct quant description, add guidance

README: warn that MTP spec-decode corrupts output; fix stale flags

rebuild from ornith-ai aug-24 mtp refresh — mtp verified 87.98% acceptance, 105.6 tok/s effective on gfx1151

readme: aug-24 mtp refresh revision — mtp now recommended (87.98% acceptance, 105.6 tok/s on gfx1151)

Add vision projector checksum

Document automatic vision setup

Add BF16 vision projector

Update README with 256K max context options

Upload Ornith-1.5-35B-A3B ROCmFP4 GGUF

initial commit