cookbook(finetune): document how to enable LoRA in the SFT recipes

[lfengad]

What

Adds a LoRA fine-tuning section to both finetune cookbook READMEs:

• cookbooks/cosmos3/generator/audiovisual/finetune/README.md
• cookbooks/cosmos3/reasoner/finetune/README.md

Both cookbooks already ship LoRA-capable recipes, but neither explains how to turn LoRA on — today you'd have to reverse it out of the framework's TOML schema.

Approach

Written as a guide, not a second set of recipes. No new TOML or launch shell is added; the section explains how to convert an existing recipe:

• which TOML sections to touch, what each knob does, and why
• concrete numbers appear only as examples to start from (rank/alpha with tuning direction), and the LR is framed as "take the recipe's full-FT value and go up"
• only the mechanically required bits stay prescriptive (lora_enabled, keys_to_select = ["lora_"], keys_to_skip_loading, EMA/compile, and the Edge lora_exclude_path_regex) — getting these wrong fails silently rather than loudly

launch_sft_vision_super.sh is surfaced in both READMEs as the ready-made LoRA recipe to read alongside: directly in the generator cookbook, and by relative link from the reasoner one as the vfm-side reference for the same mechanism.

vfm vs vlm

The mechanism is identical; four things differ, and each README covers its own side:

	vision (vfm)	reasoner (vlm)
lora_target_modules	q/k/v/o_proj_moe_gen	q/k/v/o_proj
lora_exclude_path_regex	n/a	required on Cosmos3-Edge (SigLIP2 tower shares projection names with the LLM)
checkpoint.keys_to_skip_loading	add "lora_"	stays [] (HF safetensors, not DCP)
EMA / torch.compile	on in these recipes, must be disabled	already false

Merge order

The reasoner section depends on VLM LoRA support landing in cosmos-framework (liangf/vlm-lora-sft) — on today's main, PATH_REMAPS["vlm"] still skips the lora_* keys. Filed as draft for that reason; the generator section is accurate against the released framework as-is.

Verification

Every claim was checked against the corresponding TOMLs, the framework source, or actual runs — including one fix caught during review: extending vfm LoRA to the FFN needs path-qualified targets (mlp_moe_gen.up_proj), because a bare up_proj also matches the understanding tower's mlp.up_proj and there is no path-exclusion knob on the vfm side.

🤖 Generated with Claude Code