library_name: transformers
pipeline_tag: image-text-to-text
tags:

• qwen
• qwen2-vl
• multimodal
• vision-language
• gptq
• gptq-pro
• 4bit
• foem

Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128

Overview

Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128 is a GPTQ-quantized checkpoint intended for efficient GPU inference, published by groxaxo.
It is intended for open-source evaluation, reproducible experimentation, and compatible local or
hosted inference workflows. The wording below is deliberately limited to what can be verified
from this repository's metadata and artifacts.

The repository name identifies a behavior-modified or reduced-filtering lineage. That label describes the source or conversion history; it is not a guarantee of unrestricted behavior in every prompt or runtime. Test outputs carefully before sharing or deploying them.

At a glance

What is included

• *.safetensors (5 files)
• config.json
• generation_config.json
• tokenizer.json
• tokenizer_config.json
• processor_config.json
• chat_template.jinja
• quantize_config.json
• Additional configuration, tokenizer, processor, or shard files (13 visible artifacts total)

Quick start

vLLM (documented configuration)

vllm serve groxaxo/Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128 \
  --quantization gptq_marlin \
  --dtype float16 \
  --trust-remote-code

This command is taken from the repository documentation. Adjust tensor parallelism, context
length, and cache settings to match your hardware and vLLM version.

Compatibility and responsible use

• Use a runtime that explicitly supports this format, architecture, and modality.
• Keep configuration, tokenizer, processor, projection, and weight files from the same revision together.
• Review the source model card and license before redistribution or deployment.
• Hardware needs depend on parameter count, context length, cache precision, quantization, and concurrency.
• Report reproducible issues with the runtime version, hardware, launch command, and a minimal example.

Quantization or conversion changes numerical behavior, memory use, and throughput relative to the source checkpoint; validate quality on your own workload.

Generated outputs may be inaccurate or unsuitable for a given use case. Users are responsible for
testing behavior, applying appropriate safeguards, and complying with applicable licenses and laws.

FOEM-enhanced GPTQ-Pro W4G128 export of Qwen3.6-27B-abliterated-v2.

Quantization recipe

• Base quant recipe: GPTQ-Pro W4G128
• Enhancement: FOEM + activation-weighted MSE
• bits: 4
• group_size: 128
• desc_act: false
• sym: true
• true_sequential: true
• act_group_aware: true
• lm_head: false
• FOEM: alpha=0.25, beta=0.2

What stayed preserved

• Vision tower tensors remain present and BF16
• lm_head remains BF16
• model.language_model.embed_tokens remains BF16
• Processor and multimodal metadata are included

Verification notes

The local artifact was checked to confirm:

• visual tensors > 0
• quantized tensors are present
• lm_head is not quantized
• embed_tokens is not quantized

The source model was also validated locally with image input before quantization.

Local launch settings used to validate the source vision path

The stable local vLLM launch on this machine was:

setsid env \
  CUDA_VISIBLE_DEVICES=0,1,4 \
  CUDA_DEVICE_ORDER=PCI_BUS_ID \
  OMP_NUM_THREADS=1 \
  TOKENIZERS_PARALLELISM=false \
  NCCL_P2P_DISABLE=1 \
  NCCL_IB_DISABLE=1 \
  NCCL_NET_GDR_DISABLE=1 \
  NCCL_SHM_DISABLE=0 \
  NCCL_CUMEM_HANDLE_DISABLE=1 \
  PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:256 \
  /home/op/venvs/vllm-qwen36/bin/vllm serve "/home/op/models/Qwen3.6-27B-abliterated-v2" \
    --host 0.0.0.0 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --pipeline-parallel-size 3 \
    --max-model-len 4096 \
    --kv-cache-dtype fp8 \
    --gpu-memory-utilization 0.98 \
    --max-num-seqs 1 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --trust-remote-code \
    --served-model-name qwen36-27b-abliterated-v2 \
    --disable-custom-all-reduce \
    --generation-config vllm \
    --enforce-eager \
    --limit-mm-per-prompt '{"image":1}'

Notes:

• --max-model-len 32144 did not fit KV cache on this host.
• For direct chat requests, chat_template_kwargs.enable_thinking=false was used to avoid spending the visible token budget in the reasoning channel.

Files

• quantize_config.json records the GPTQ-Pro + FOEM config
• processor_config.json keeps the multimodal processor config
• model.safetensors.index.json and shard files contain the final export