fix(utils/weight_conversion): handle str target_modules/target_parameters in MoE config conversion

[Anai-Guo]

Summary

Closes #3229.

_convert_peft_config_moe in src/peft/utils/transformers_weight_conversion.py normalized peft_config.target_modules and peft_config.target_parameters via set(value or []) before iterating them against the MoE target-module mapping. When a user passes a single string for these fields (a supported shape elsewhere in PEFT, e.g. target_modules="gate_up_proj"), set("gate_up_proj") silently splits the string into individual characters ({"g","a","t","e","_","u","p","r","o","j"}). The downstream for target in peft_config.target_modules loop then matches none of the expected module names, so config conversion for MoE models silently produced wrong results. Reported in modelscope/ms-swift#9321.

What changed

• src/peft/utils/transformers_weight_conversion.py (_convert_peft_config_moe): replace

  peft_config.target_parameters = set(peft_config.target_parameters or [])
  peft_config.target_modules = set(peft_config.target_modules or [])

  with a small local _normalize_to_set helper that maps str to a single-element set, None to set(), and otherwise falls back to set(value). Pure normalization change; no other code paths affected.

Verification

• Behavior preserved for list/tuple/set/None: set([...])/set((...))/set({...}) semantics unchanged; None still becomes set() instead of erroring.
• New behavior for str: target_modules="gate_up_proj" becomes {"gate_up_proj"} so the per-target loop sees one well-formed module name instead of ten one-character non-matches.
• No new imports; helper is defined inside the function to keep the public surface unchanged.

AI-assisted, human reviewed.

[BenjaminBossan]

Thanks for the PR but there is already another one (#3232).