Downloads · 30 days
16
4% of all-time downloads
groxaxo/Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128-preserve-vl
Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128-preserve-vl is a image-text-to-text model from groxaxo. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers.
Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128-preserve-vl is a GPTQ-quantized checkpoint intended for efficient GPU inference, published by groxaxo. It is intended for open-source evaluation, reproducible experim…
Downloads · 30 days
16
4% of all-time downloads
All-time downloads
373
Public
Parameters
27.4B
18.7 GB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors18.7 GB · 100%
How the weights are stored.
I3224.3B · 89%
From the Hugging Face model README
Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128-preserve-vl 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.
| Field | Details |
|---|---|
| Format | GPTQ |
| Source / base | wangzhang/Qwen3.6-27B-abliterated |
| Intended task | image-text-to-text |
| License | the license declared in the repository files |
*.safetensors (5 files)config.jsongeneration_config.jsontokenizer.jsontokenizer_config.jsonprocessor_config.jsonchat_template.jinjaquantize_config.jsonvllm serve groxaxo/Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128-preserve-vl \
--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.
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.
<!-- polished-overview:end -->Strict preserve-VL FOEM-enhanced GPTQ-Pro W4G128 export of Qwen3.6-27B-abliterated-v2.
This artifact was rerun with hard exclusion rules for fragile multimodal pieces:
visual.*embed_tokens.*lm_head.*norm.*merger.*projector.*patch_embed.*Local post-save checks confirmed:
visual tensors: 333lm_head.weight remains BF16model.language_model.embed_tokens.weight remains BF16AutoProcessor, AutoTokenizer, and AutoConfig load successfullyarchitectures = ["Qwen3_5ForConditionalGeneration"]language_model_only = falseThe 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.chat_template_kwargs.enable_thinking=false was used to keep visible output in the answer channel.quantize_config.json records the explicit dynamic skip rules and FOEM metadataprocessor_config.json keeps the multimodal processor configmodel.safetensors.index.json and shard files contain the final export