Overview

PEFTHelper drops adapter_config.json keys it doesn't know, so adapters using PEFT features vLLM doesn't implement loaded without an error and were served with wrong outputs. This rejects them with a clear error, the same way DoRA is rejected, and fixes a regression from #53555 that let lora_A / lora_B .bias tensors replace the LoRA weights.

Fixes #60223

Claims

• Adapters with lora_bias, an init_lora_weights that modifies the base model when PEFT loads the adapter (pissa, pissa_niter_*, olora, corda, loftq) or any other value not on the allowlist, alora_invocation_tokens, layer_replication, use_bdlora or use_qalora are rejected in validate_legal with a message naming the feature. Before, they loaded and gave outputs that differ from PEFT.
• parse_fine_tuned_lora_name accepts .bias only for the score / classifier heads again, not for lora_A / lora_B. This restores the behavior before [LoRA] Support modules_to_save for sequence classification #53555 (v0.30.0 rejects these adapters; v0.31.0 loads them and silently drops the bias or swaps it with the weight).
• No change for other adapters: init_lora_weights = true / false / "gaussian" / "eva" / "orthogonal" / "mica" / "lora_ga" and PEFT's converted PiSSA / OLoRA / CorDA adapters (init_lora_weights: true) load as before.
• Not changed: target_parameters, trainable_token_indices and DoRA, which open PRs [Bugfix][LoRA] Fix PEFT 0.18+ target_parameters LoRA loading for 3D MoE experts #52198, [LoRA] Support PEFT trainable token rows alongside LoRA adapters #55796 and [LoRA] Add DoRA support for serving and linear layers #47395 are adding support for. If BD-LoRA support ([Core][Feature] Add BD-LoRA to VLLM #28136) lands first, its rejection here can be dropped.

Validation

Every feature saved with PEFT 0.21.2 on a tiny random Llama (bf16, RTX 3080), loaded with llm.llm_engine.add_lora in one engine, then the accepted ones served. Mean absolute difference of per-token prompt logprobs between vLLM and PEFT:

The rejected adapters fail add_lora with the message from validate_legal and the engine keeps serving the other adapters afterwards. The same holds for POST /v1/load_lora_adapter (new case in tests/entrypoints/serve/lora/test_lora_adapters.py). As with DoRA today, an unsupported adapter passed directly in a LLM.generate request still ends in EngineDeadError.

Real adapters: of the 628 most-downloaded Hugging Face PEFT repos with an adapter_config.json, this rejects 2 more than main (both PiSSA). A targeted search found about 100 more PiSSA / OLoRA / aLoRA adapters. For one of them, AkameV6p5/qwen2.5-3b-agent-pissa-sft on Qwen/Qwen2.5-3B, vLLM main differs from PEFT by 3.44 (mean, PEFT run to run: 1.55 because of its randomized SVD) and matches PEFT with init_lora_weights forced to true (0.06), i.e. it is served as a plain LoRA on the unmodified base.

Tests (on main e11962f):

pytest tests/lora/test_peft_helper.py tests/lora/test_utils.py
34 passed (the 12 new cases fail on main)

pytest tests/entrypoints/serve/lora/test_lora_adapters.py
13 passed

pytest tests/lora/{test_lora_utils,test_lora_weights,test_lora_manager,test_lora_checkpoints,
       test_lora_huggingface,test_sequence_classification,test_moe_lora_ep_load,test_resolver,
       test_worker,test_add_lora,test_lora_functions,test_transformers_model,test_qwen3_with_multi_loras,
       test_qwen3_unembed}.py
all passed; test_default_mm_loras.py (needs vllm[audio]) and test_llama_tp.py (gated model) fail the same way on main in this environment

The new tests are CPU-only and extend the existing suites. test_sequence_classification.py covers the score.bias path that #53555 added.

Details

• The fields are added to PEFTHelper with PEFT's defaults only so that from_dict keeps them for validation.
• init_lora_weights is checked against an allowlist of the values that load in PEFT as a plain LoRA, so a future PEFT init is rejected rather than silently accepted. I checked this by loading one saved adapter in PEFT 0.21.2 with each value: gaussian, eva, orthogonal, mica and lora_ga give the same logits as true (lora_ga changes the base weights only during training, when its gradients are attached); pissa and olora change them (max logit diff 0.15 and 0.23); corda fails to load in PEFT without preprocess_corda; loftq replaces the base weight with the quantized one (loftq_init in peft/tuners/lora/layer.py). The error points to PEFT's path_initial_model_for_weight_conversion and to setting init_lora_weights to true when the served base already holds the modified weights.
• A shape check in set_lora would also catch BD-LoRA / QALoRA style layouts, but it runs at activation time, so it is left for a follow-up next to the load-time validation in [Bugfix][LoRA] Validate incompatible MoE adapter formats #59924.