model support system follow-ups

Tracking: model support system follow-ups

Follow-up tracker for the deferred items from #3869 (profile system) and #3884 (transformers registry seams). Status below re-verified against main at 93dcd622f, as of 2026-08-09. Every item is still open; the only substantive movement is row 2 below (the Liger per-arch chain grew again).

Open question for maintainers

Should already-implemented models migrate onto the profile system? This is the one item that gates others — the migration skill (below) and the subsystem rows only make sense if the answer is yes. Today _BUILTIN_MODULES holds the two descriptors that shipped with #3869 (kimi_linear, paddleocr_vl), and loaders/patch_manager.py still carries 21 model_config_type branches. Either answer is workable; the branches are not urgent, but leaving the question open means the profile system and the branch chain both stay live in parallel.

Ready to pick up

• Additional family templates. templates.py currently has VANILLA_CAUSAL_LM and IMAGE_TEXT_TO_TEXT; scope was deliberately held down for #3869. More templates want the descriptor inventory that migration would produce, so this is partly downstream of the question above.
• Save-time weight-conversion reversal. #3884 landed the load side plus _warn_irreversible_weight_transforms, which reports at registration time that a transform has no reverse_op. The save boundary itself is still open: there is no save-phase hook, and save_pretrained(save_original_format=True) reverses whatever is registered at save time, which is not guaranteed to be what the profile declared.
• Let weight_conversions read current registrations, and re-apply post-build. Registrations are applied once in apply_pre_model_load_patches, and the provider returns a mapping that is registered with overwrite=True — so it cannot inspect or filter what transformers already has under a key, and a model whose __init__ re-registers with overwrite=True clobbers it. The in-tree case is _fix_nemotron_h_conversion_mapping (patch_manager.py:676), which removes one entry from the mapping the nvidia Hub model registers itself and must run after build. #3884 names nemotron_h as precedent that should migrate onto this seam; as built it can't, so the hardcoded fix and the declarative path currently coexist.
• Deprecate ModelHookContext.tokenizer toward processor. transformers v5 is processor-first; the context still exposes both fields. Deferred until hook consumers migrate.

Subsystem migration

Candidates to become typed profile strategies, in recommended order — earlier rows have a ready seam and grow the descriptor inventory later rows depend on. All are gated on the migration question.

#	Subsystem	Where it lives today	Becomes a strategy when	Status
1	Remaining model_type branches	loaders/patch_manager.py, monkeypatch/	each branch migrates into a phase hook on its architecture's descriptor	seam shipped in #3869; 21 branches remain
2	Loss adaptation	CCE / Liger integrations, loss patches	loss-patch dispatch consults capabilities rather than per-arch patch lists	partial, and drifting — both integrations call check_capability as a guard, but dispatch is still a per-arch chain, which has since grown from 7 to 8 hand-patch branches in integrations/liger/plugin.py (mistral3/ministral3, added by #3922)
3	Batch preparation	utils/collators/, packing utils	collator choice is profile-driven alongside the already-typed processing strategy	not started
4	Adapter loading	loaders/adapter.py	PEFT wrapping exposes a stable pre/post boundary; near-term win is per-arch lora_target_modules defaults	not started
5	Optimizer construction	core/trainers/mixins/optimizer.py + builders	optimizer factory selection is decoupled from trainer mixin wiring	not started (rarely arch-specific today)
6	Trainer selection	core/builders/ (causal.py, rl.py)	builder choice has a generic seam instead of RL-type/arch branching	not started (hardest seam; few arch-specific consumers yet)

Rows 1 and 2 are the ones with real payoff today; 5 and 6 are listed for completeness and may never be worth doing.