[BUG] peft 0.19 target_modules (str) use `set`

[Jintao-Huang]

System Info

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

peft/src/peft/utils/transformers_weight_conversion.py

Lines 371 to 372 in 64c40db

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

The target_modules here may be a string, which could lead to a bug.

modelscope/ms-swift#9321

Who can help?

No response

Reproduction

Expected behavior

[voloha-mil]

Adding a copy-pasteable repro for the "all-linear" sentinel since the description focuses on the root cause.

Versions

peft==0.19.1, transformers==5.8.0, Python 3.12.

Minimal repro — direct call into the buggy function

from peft import LoraConfig
from peft.utils.transformers_weight_conversion import _convert_peft_config_moe

config = LoraConfig(r=8, lora_alpha=16, target_modules="all-linear")
_convert_peft_config_moe(config, "mixtral")
print(sorted(config.target_modules))
# ['-', 'a', 'e', 'i', 'l', 'n', 'r']
# ['-', 'a', 'e', 'i', 'l', 'n', 'r']

The sentinel "all-linear" becomes the 7-char set {'-','a','e','i','l','n','r'} via set(peft_config.target_modules or []).

End-to-end — get_peft_model failure on tiny Mixtral

import torch, transformers
from transformers import AutoModelForCausalLM
from peft import LoraConfig, get_peft_model

cfg = transformers.MixtralConfig(
    hidden_size=64, intermediate_size=128, num_hidden_layers=2,
    num_attention_heads=4, num_key_value_heads=2,
    num_local_experts=2, num_experts_per_tok=2,
    vocab_size=128, max_position_embeddings=128,
)
with torch.device("meta"):
    model = AutoModelForCausalLM.from_config(cfg)
get_peft_model(model, LoraConfig(r=8, lora_alpha=16, target_modules="all-linear"))

ValueError: Target modules {'e', 'l', 'a', 'r', 'i', '-', 'n'} not found in the
base model. Please check the target modules and try again.

Triggers for every model_type in _MODEL_TO_CONVERSION_PATTERN (mixtral,
qwen3_moe, deepseek_v2/v3, glm4_moe, olmoe, minimax, …). Dense models take a different path and are unaffected.

Looks like #3232 fixes this as a side effect — early-returning on a string target_modules means _maybe_include_all_linear_layers runs downstream and expands the sentinel against
the real model.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.