English | 中文

MOSS-VL-Instruct-0708 FP8 Dynamic + Transformers KV8

MOSS-VL is an open vision-language model family from OpenMOSS, supporting image understanding, long-video understanding, and realtime streaming interaction. This repository provides the FP8-quantized checkpoint of MOSS-VL-Instruct-0708.

Technical Report: https://arxiv.org/pdf/2608.15045

This is the 24 GiB quantized release of MOSS-VL-Instruct-0708. The same checkpoint directory is supported by both the standard Transformers inference path and the native SGLang MOSS-VL backend.

Quantization profile

Component	Format
Most language layers	FP8 weights + per-token dynamic FP8 input activations
Cross-attention, vision modules and `lm_head`	BF16
Transformers KV cache	HQQ INT8
SGLang KV cache	Native SGLang BF16 cache

The FP8 weights are shared across both engines. Transformers reads the HQQ KV8 configuration from generation_config.json; SGLang loads the same weights with model_impl=sglang and uses its own cache implementation.

Quantization benchmark

Across the selected benchmarks, the quantized models remain close to their non-quantized BF16 counterparts, showing that overall model quality is largely preserved after quantization.

MOSS-VL quantization benchmark comparison

Hardware requirements

The model is designed to run with the Transformers inference path on a single NVIDIA GPU with 24 GB of VRAM. SGLang memory requirements depend on its server configuration and KV cache allocation.

Transformers environment

Installation

Start from the standard MOSS-VL release environment, then install the FP8 checkpoint loader and HQQ cache backend:

git clone https://github.com/OpenMOSS/MOSS-VL.git
cd MOSS-VL

conda create -n moss_vl_quant python=3.12 pip -y
conda activate moss_vl_quant
pip install -i https://pypi.org/simple --no-build-isolation -r requirements.txt
pip install -i https://pypi.org/simple \
  compressed-tensors==0.14.0 \
  hqq==0.2.8.post1
python -m pip check

Core versions used by the standard Transformers path:

Package	Version
Python	3.12
PyTorch	2.8.0 + CUDA 12.8
Transformers	4.57.1
Accelerate	1.12.0
FlashAttention	2.8.1
compressed-tensors checkpoint format	0.14.0
HQQ	0.2.8.post1

Video decoding also requires FFmpeg to be available in PATH.

Load the model

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

checkpoint = "/path/to/MOSS-VL-0708-Instruct-FP8-Dynamic-KV8-HQQ"

processor = AutoProcessor.from_pretrained(
    checkpoint,
    trust_remote_code=True,
    frame_extract_num_threads=1,
)
model = AutoModelForCausalLM.from_pretrained(
    checkpoint,
    trust_remote_code=True,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)
model.eval()

generation_config.json enables HQQ KV8 automatically. Keep that file beside the checkpoint and do not pass a conflicting cache configuration.

Image inference

text = model.offline_image_generate(
    processor,
    prompt="Describe this image.",
    image="data/example_image.jpg",
    shortest_edge=4096,
    longest_edge=16777216,
    multi_image_max_pixels=201326592,
    patch_size=16,
    temporal_patch_size=1,
    merge_size=2,
    image_mean=[0.5, 0.5, 0.5],
    image_std=[0.5, 0.5, 0.5],
    max_new_tokens=256,
    temperature=1.0,
    top_k=50,
    top_p=1.0,
    repetition_penalty=1.0,
    do_sample=False,
    vision_chunked_length=64,
)
print(text)

Video inference

text = model.offline_video_generate(
    processor,
    prompt="Describe this video.",
    video="data/example_video.mp4",
    shortest_edge=4096,
    longest_edge=16777216,
    video_max_pixels=201326592,
    patch_size=16,
    temporal_patch_size=1,
    merge_size=2,
    video_fps=1.0,
    min_frames=1,
    max_frames=256,
    num_extract_threads=4,
    image_mean=[0.5, 0.5, 0.5],
    image_std=[0.5, 0.5, 0.5],
    max_new_tokens=256,
    temperature=1.0,
    top_k=50,
    top_p=1.0,
    repetition_penalty=1.0,
    do_sample=False,
    vision_chunked_length=64,
)
print(text)

Repository inference runner

The standard repository runner accepts JSON/JSONL queries with image or video content and loads this directory directly:

python inference/run_inference.py \
  --checkpoint /path/to/MOSS-VL-0708-Instruct-FP8-Dynamic-KV8-HQQ \
  --mode offline \
  --input inference/video_queries.json \
  --output video_results.json

SGLang environment

Use a separate environment for SGLang. The native backend in the mossvl_sglang release directory pins:

Package	Version
Python	3.12.8
SGLang	0.5.11
sglang-kernel	0.4.2
PyTorch	2.11.0 + CUDA 13.0
Transformers	5.6.0
compressed-tensors	0.17.1

Create the environment from its complete lock file:

cd /path/to/mossvl_sglang
env -u INDEX_URL -u PIP_INDEX_URL conda create -y -n mossvl-sglang-0511 \
  --override-channels \
  -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main \
  python=3.12.8 pip=24.2
conda activate mossvl-sglang-0511
env -u INDEX_URL -u PIP_INDEX_URL python -m pip install -r requirements.txt
python -m pip check

Start the OpenAI-compatible service with the quantized directory as the model path:

conda activate mossvl-sglang-0511
cd /path/to/mossvl_sglang
GPU_IDS=0 TP_SIZE=1 MEM_FRACTION_STATIC=0.35 \
  ./start_sglang_moss_vl.sh \
  --model-path /path/to/MOSS-VL-0708-Instruct-FP8-Dynamic-KV8-HQQ

The service listens on http://127.0.0.1:30000 by default. Keep MODEL_IMPL=sglang; do not force the Transformers remote implementation. SGLang intentionally uses its native BF16 KV cache rather than the HQQ cache in generation_config.json.

Configuration files

config.json: model and FP8 weight/activation configuration.
generation_config.json: Transformers HQQ KV8 configuration.
modeling_moss_vl.py: checkpoint-local cross-attention/QuantizedCache code.

Citation

@misc{mossvl,
  title         = {MOSS-VL Technical Report},
  author        = {Wang, Pengyu and Tan, Chenkun and Zhou, Shaojun and Zhou, Qirui and Chen, Yanxin and He, Xingyang and Zeng, Huazheng and Cheng, Jijun and Wang, Chenghao and Qian, Xiaomeng and Wang, Pengfei and Huang, Zhan and Gao, Shanqing and Huang, Wei and Cao, Longjun and Ran, Wu and Liu, Jie and Zhu, Changtai and Wang, Hongkai and Tian, Yixian and Liu, Chenghao and Ye, Zhen and Wang, Xinghao and Jiang, Botian and Feng, Guoguo and Fei, Zhaoye and Li, Ruixiao and Chen, Mingshu and Gao, Yang and Cheng, Qinyuan and Li, Shimin and Qiu, Xipeng},
  year          = {2026},
  eprint        = {2608.15045},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.15045}
}

@misc{mossvideopreview,
  title         = {{MOSS-Video-Preview: Toward Real-Time Video Understanding via Cross-Attention}},
  author        = {Pengyu Wang and Chenkun Tan and Shaojun Zhou and Wei Huang and Qirui Zhou and Zhan Huang and Zhen Ye and Jijun Cheng and Xiaomeng Qian and Yanxin Chen and Xingyang He and Huazheng Zeng and Chenghao Wang and Pengfei Wang and Hongkai Wang and Shanqing Gao and Yixian Tian and Chenghao Liu and Xinghao Wang and Botian Jiang and Xipeng Qiu},
  year          = {2026},
  eprint        = {2606.07639},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.07639}
}