Release calibrated Image 2.1 Turbo FP8 transformer, accelerated SM120 runtime, quality evidence and real-time demo
ffc54ec verified | """Shared workstation paths and BF16 helpers for the Image 2.1 Turbo quantization project.""" | |
| import os | |
| from pathlib import Path | |
| import torch | |
| from torch import nn | |
| ROOT = Path(os.environ.get('TURBO_ROOT', '/home/user/.local/share/rtx-pro-apps/qwen-image-2.1-turbo')) | |
| BASE = ROOT / 'base' | |
| BASE_REPO = 'Qwen/Qwen-Image-2.1-Turbo' | |
| BASE_REVISION = 'd65dbc9a7e8f6b5479e33dee6030eaab2a906509' | |
| CALIBRATION = ROOT / 'calibration' | |
| def eligible_modules(model): | |
| return {n: m for n, m in model.named_modules() | |
| if isinstance(m, nn.Linear) and n.startswith('transformer_blocks.') and m.bias is None} | |
| def load_bf16_pipeline(base=BASE): | |
| from diffusers import QwenImage21Pipeline | |
| pipe = QwenImage21Pipeline.from_pretrained(base, torch_dtype=torch.bfloat16).to('cuda') | |
| pipe.set_progress_bar_config(disable=True) | |
| return pipe | |
| def sizes(i, kind='calibration'): | |
| """Same resolution mix as the Image 2.1 releases: every fourth case at 2048, two alternate aspects.""" | |
| w = h = 2048 if i % 4 == 0 else 1024 | |
| alt = (52, 53) if kind == 'calibration' else (12, 13) | |
| if i == alt[0]: w, h = 1536, 864 | |
| if i == alt[1]: w, h = 864, 1536 | |
| return w, h | |