fix(peft_utils): raise actionable ValueError when rank_dict is empty in get_peft_kwargs

[dparikh79]

Summary

Refs huggingface/peft#3238 (the issue was filed against peft, but the failure path is here in diffusers).

get_peft_kwargs is reached from every diffusers LoRA loader (_load_lora_into_text_encoder, load_lora_into_unet, load_lora_into_transformer, etc.). Each caller walks model.named_modules() and probes the state_dict for {module_name}.lora_B.weight keys, building a rank_dict. When that loop matches nothing (typical when the saved keys carry an extra/missing prefix vs the target model, an adapter-name infix between lora_B and weight, or a non-diffusers serialization format that was not converted upstream), rank_dict arrives at get_peft_kwargs empty.

The first statement of the function is:

r = lora_alpha = list(rank_dict.values())[0]

which raises IndexError: list index out of range with no signal about which mismatch is actually responsible. The traceback ends inside get_peft_kwargs with no way to tell whether you have a prefix problem, an adapter-name infix problem, a format-conversion problem, or just a state_dict that genuinely has no LoRA weights at all.

This PR raises a ValueError ahead of the indexing, with a message that names the common causes and shows a sample of the state_dict keys. Future bug reports surface the actual mismatch shape in the user's report instead of a cryptic IndexError.

Scope

• One file change: src/diffusers/utils/peft_utils.py, function get_peft_kwargs, ~20 added lines (a guard + diagnostic message).
• No behaviour change on the happy path: when rank_dict is non-empty, the existing first line runs unchanged.
• New file: tests/others/test_peft_utils.py with three focused unit tests (empty rank_dict raises, None state_dict is safe, fast-path unchanged).

This is intentionally a defensive surface-area fix only. The underlying SDXL/SD3.5 key-mismatch in #3238 is a separate concern (likely in convert_state_dict_to_peft or the per-pipeline _load_lora_into_text_encoder rank-discovery loops); diagnosing it requires reproducing the user's specific state_dict layout. Surfacing the actionable error here is the prerequisite for collecting that data from future reporters.

Test plan

• tests/others/test_peft_utils.py (new):
  • test_empty_rank_dict_raises_actionable_value_error; exercises the new path; asserts ValueError, asserts message names the underlying mismatch and includes a state_dict key sample.
  • test_empty_rank_dict_with_none_state_dict_is_safe; covers the peft_state_dict=None defensive branch.
  • test_non_empty_rank_dict_unchanged; smoke-tests the happy path (r=4, lora_alpha=4, no rank_pattern, expected target_modules).
• All three pass locally via direct sys.path import without needing the full diffusers install (the function is pure utility code).
• Did not run the full diffusers test suite locally (heavy install with torch + transformers + accelerate); relying on CI.

AI Assistance Disclosure

Investigation, the patch, and the tests were drafted with Claude assistance. I read the call graph from each caller (_load_lora_into_text_encoder, load_lora_into_unet) into get_peft_kwargs to confirm rank_dict can legitimately be empty under valid-input + key-mismatch conditions, and that the change preserves the happy path bitwise. I am the human contributor accountable for this PR.