This package contains a copy of the original Flash model with Qwen3 MLP FFN layers (gate_proj, up_proj, down_proj) replaced by bitsandbytes Linear4bit (NF4). Attention projections, vision model, gating, and lm_head remain in FP16. Flash path remains enabled.

Load with:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor
model = AutoModelForCausalLM.from_pretrained('<repo>', trust_remote_code=True, device_map={'':'cuda'})
tok = AutoTokenizer.from_pretrained('<repo>', trust_remote_code=True, use_fast=False)
proc = AutoProcessor.from_pretrained('<repo>', trust_remote_code=True)
```