gatilin/ming_flash_omni2_vision_io

Ming-flash-omni-2.0 native Qwen3-MoE vision encoder and normalized image projection.

Extracts vision and linear_proj; preserves the native final L2 normalization. This encoder is distinct from MingTok. The large language backbone, audio path, talker and image-generation transformer/VAE are not included. Native attention constructors force FlashAttention; the loader restores eager attention after construction for CPU execution.

Source and conversion

Model/weights: inclusionAI/Ming-flash-omni-2.0, fixed revision c1ed5565e64e85a76826255c000b66a20d4634c3. Native code: inclusionAI/Ming, fixed revision 2a0c02ae3130190160c215f89fce7de3005db483.

All selected tensor names, shapes, dtypes and payload bytes are preserved, except the explicit IBQ EMA shadow-name mapping. See extraction.json, provenance.json and tensor/conversion mappings. This repository contains component weights and a local native loader, not an AutoModel.from_pretrained checkpoint.

Load

Install requirements.txt in a compatible environment, then:

from huggingface_hub import snapshot_download
import sys
import torch
root = snapshot_download("gatilin/ming_flash_omni2_vision_io")
sys.path.insert(0, root)
from component import load_model
model = load_model(root, dtype=torch.float32, device="cpu")

The model is returned in evaluation mode. Run inference under torch.inference_mode(). Inputs are native tensors/patches, not arbitrary PIL images; inspect component.py and the included validation script for the exact layout. CPU float32 was tested; CUDA/MPS and full end-to-end benchmarks were not tested.

Validation

See validation.json for the executed component tests, shapes and limitations. Source payload equality and native strict loading are required before release. Component tests do not establish full-model generation or understanding quality. Original stochastic decoding settings and normalization conventions must be retained.

License and attribution

Original upstream notices and license files are retained. Reused third-party components keep their original terms. This extraction does not claim authorship of the underlying trained model.