Slow Download and VRAM Issues | Generated by AI
Here’s the summary of issues:
1. 🚧 Model is Gated
You had to accept terms on HuggingFace first, then pass a token. (Fixed ✓ — token works now)
2. 🐢 Extremely Slow Download (~24 GB)
The model is 12 billion parameters — roughly 24 GB of weights.
- After 30 minutes, only 2.3 GB / ~24 GB was downloaded (~1.3 MB/s)
- At this rate, it would take ~4-5 hours to finish
3. 💾 GPU VRAM Bottleneck
Your GPU has 12 GB VRAM, but the model needs ~24 GB in bfloat16. → The script uses CPU offloading, which trades VRAM for much slower inference.
Options to speed things up
| Option | What to do |
|---|---|
| Let it finish | Run overnight — download will complete eventually |
| Use a smaller model | Try black-forest-labs/FLUX.1-dev (same size) or a distilled version |
| Use an API instead | https://fal.ai/models/fal-ai/flux/schnell — no download, pay per image |
| Use a faster connection | Download elsewhere and copy the cache folder |
Want me to let it continue downloading in the background? Or would you like to try a different approach?
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋