---
base_model:
- Qwen/Qwen3-32B
- qihoo360/Light-IF-32B
library_name: peft
tags:
- mergekit
- peft
- lora
- vllm
---
# NEXS Qwen3-32B IF LoRA (vLLM-ready)

Rank-128 LoRA adapter (bf16) extracted with
[mergekit](https://github.com/arcee-ai/mergekit) from
[qihoo360/Light-IF-32B](https://huggingface.co/qihoo360/Light-IF-32B) against the base model
[Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B), then sanitized for vLLM
serving.

## Sanitization applied

The raw mergekit extraction included full-rank `modules_to_save` tensors
(`embed_tokens`, `lm_head`, and norm layers) that vLLM's LoRA runtime does not
support. This upload contains only the pure low-rank `lora_A`/`lora_B` weights
(448 pairs: 64 layers x q/k/v/o/gate/up/down projections), with
`modules_to_save: null` in `adapter_config.json`. No `resize_token_embeddings()`
call is needed to load this adapter.

## Serving with vLLM

```bash
python -m vllm.entrypoints.openai.api_server \
    --model Qwen/Qwen3-32B \
    --enable-lora \
    --lora-modules IF=anjohn0077/NEXS-qwen3-32b-IF-lora \
    --port 8000 \
    --max-lora-rank 128 \
    --gpu-memory-utilization 0.85
```

## Evaluation (ifeval)

| Variant | Accuracy |
|---|---|
| Base model | 0.8336 |
| **This LoRA on base (via vLLM)** | **0.2737** |
| Original full fine-tune | 0.8669 |

Evaluated with [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness)
against a local vLLM OpenAI-compatible endpoint:

```bash
lm_eval --model local-completions \
    --model_args model=IF,base_url=http://localhost:8000/v1/completions,tokenizer=Qwen/Qwen3-32B,num_concurrent=10 \
    --tasks ifeval \
    --output_path results/vllm_IF
```