Patterned PEFT LoRA adapters run at the wrong scale in vLLM 0.30.0

The cause is in vllm/lora/peft_helper.py. In 0.30.0 that file declares the fields vLLM reads from adapter_config.json: r, lora_alpha, target_modules, bias, modules_to_save, use_rslora, use_dora and vllm_lora_scaling_factor ^[14]. From them it computes one number. With use_rslora set, the scale is lora_alpha / sqrt(r). Otherwise it is lora_alpha / r ^[15]. Neither pattern has a field, and the dev.to author's code search of the repository found no other reference to them ^[16].

Rank survives. vLLM takes each module's rank from the tensor shapes in the safetensors file, so only the scale is shared ^[17]. PEFT scales each module by its own alpha_m / r_m ^[1]. Any module whose pattern moves that ratio off lora_alpha / r is wrong by exactly the ratio ^[17]. In the test adapter, q_proj should run at 32/4 = 8, and vLLM ran it at 2, a factor of 4 ^[17].

vLLM does refuse some adapters. Its _validate_features check turns away DoRA and modules_to_save, so an adapter that loads without complaint looks supported ^[9]. No line in the test run's log mentions rank_pattern or alpha_pattern ^[8]. The output is still the adapter's output, only stronger or weaker in the patterned modules than it was trained to be ^[10]. According to the author, it takes a PEFT comparison or a measurement of the trained-for quality to see it ^[10].

The test design is careful. PEFT's own prompt log-probabilities on four 48-token prompts are the reference, and vLLM serves the same adapter on the same base ^[5]. The setup is an 8 GB RTX 2070 under WSL2 with vLLM 0.30.0, torch 2.13.0+cu130 and PEFT 0.21.2 ^[4]. The model is a randomly initialised two-layer Qwen3 with hidden size 256 ^[4]. The card has no bf16, so the runs used fp16 where the upstream report used bf16 ^[4]. The author also scaled lora_B up so the adapter's effect would stand clear of fp16 noise ^[5]. That choice makes the gaps visible. I would not carry the absolute figures, including the patterned run's largest single gap of 0.1106 ^[22], over to a trained adapter.

For the finding to transfer, the drift has to show up on a real model with a trained adapter. The upstream reporter's OLMoE-1B-7B run is that case. On its perplexity figures, vLLM gave up about 40% of the gain the adapter had made over the base model ^[1]. The dev.to author notes those numbers come from hardware the lab doesn't have, and that their direction and size match the small model ^[13].

PEFT steers users toward the patterns in one common case. According to the report, PEFT's LoRA documentation recommends rank_pattern for mixture-of-experts models, giving each expert r // num_experts, and points to vLLM for serving ^[11]. One patterned case escapes. When save_as_lora is called with a dynamic rank, PEFT sets both r and lora_alpha to 1 and gives each module an identical value in rank_pattern and alpha_pattern, so every scale is 1 and vLLM gets it right by coincidence ^[18].

The author's toolkit includes check-lora-patterns.py, a script that takes an adapter's config and reports each module vLLM will scale wrongly, along with the size of the error ^[19]. The upstream issue is vllm-project/vllm#59799, and a fix is proposed in #59801 ^[20].