:rotating_light: Validate config attributes (#41250)
* initial commit
* just push for now
* maybe not do it for all models, lets see how many models fail now
* update
* lets see what esle fails now
* nit
* style
* delete rope validation
* bart
* make style
* provate rope valid for now, hub complains
* more updates
* i love backwards compatibility! Let's check if this will work with remote code, tmrw
* pin hf hub 1.4.0
* want to check tests
* why do we even keep `use_return_dict` from 6 hyear ago?
* special eos token can be a list in many cases, fix type hints
* batch
* batch
* batch
* another small batch
* more
* more models
* batch
* batch
* annoying typings
* batch
* batch
* batch
* last batch
* fix repo
* fix some
* many many fixes
* fix more
* commit a small batch of fixes
* more fixes
* fix repo and the new model
* clean up config files from unused imports
* revert ths one
* more new models in main branch
* let dropouts be float AND int, who know what we have in the hub!
* fix a few more non-modeling tests
* roep validation is now part of hub strict
* oops
* rope and text config
* when does this end?
* comment out for now
* oke, now donw i think
* dropout can be int in saved ckpt, fix again
* fox repo again
* processor tests
* nit
* remove `| None` in typing when not needed!
* fix style
* new models
* subconfig is a cls attr
* fix some tests
* cosmetic stuff
* .
* fix repo
* the test
* please be fixed!
* this time is the real final fix. before merging docs
* fix style
* fix repo
* why auto-doc can't resolve inheritance and just copy???
* fix some tests
* fix the auto-docstring
* oh pls!
* lastc fix
* repr is false by default
* check docstring attr
* fix slow CI
* fix repo
* fix style and copies after rebase
* pin 1.5.0
* init subclass doesn't help with dataclass decorator, revert
* style
* regex replace doesn't always just work, fix!
* Fix incorrect default values in config dataclass migration (PR #41250)
Fixes the following bugs introduced during the __init__ → @dataclass conversion:
- cvt: embed_dim (4, 192, 384) → (64, 192, 384) (typo: 64 became 4)
- univnet: resblock_dilation_sizes last element 2 → 27 (typo)
- fsmt: max_position_embeddings 104 → 1024 (typo: missing digit)
- qwen3_5: num_key_value_heads 2 → 4 (lost override of parent default)
- afmoe: route_scale 10.0 → 1.0
- granitemoeshared: router_aux_loss_coef 0.01 → 0.001
- bridgetower: initializer_factor 1e-10 → 1 (all 3 config classes)
- llama, aria, jais2, higgs_audio_v2, olmo_hybrid: initializer_range
default 0.2 → 0.02 in interval() validator
- video_llama_3: restore missing initializer_range = 0.02
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
* Fix 3 more config default regressions (round 2)
- autoformer: restore scaling default to True (was changed to 'std',
but True means 'mean' scaling per modeling code - different behavior)
- higgs_audio_v2: restore rms_norm_eps to 1e-5 (was 1e-6 inherited from
LlamaConfig; override explicitly in modular + generated)
- higgs_audio_v2_tokenizer: restore target_bandwidths default to
[0.5, 1, 1.5, 2] (was [0.5, 1, 1.5, 2, 4] inherited from XcodecConfig)
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
* Fix check-repo: regenerate higgs_audio_v2_tokenizer from modular
The modular conversion check was failing because the generated
configuration_higgs_audio_v2_tokenizer.py didn't match the modular.
Regenerated the file and let the linter normalize the type annotations
to be consistent with the parent XcodecConfig.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
* higgs nit?
* fix higgs
* style
* new models
* xcodec is same as higgs, fix
* forogt
* love it when modular complains about newline
* fix new models' typing hints
* oops, that is a property
* and one more new model just merged
* actually, non-dataclass child is really not the way so
* dont' replace all matches!
* Apply repo consistency fixes
* Revert "Apply repo consistency fixes"
This reverts commit 125624aed6813ad194f6511e479b4000d553350c.
* fix repo, would be great to fix this in `style`
* why I cant fix all failures from repo at once
---------
Co-authored-by: Arthur <arthur.zucker@gmail.com>
Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
