I lost my sanity

[LuigiColantuono]

Title: Chandra 0.1.8 completely unusable on Windows - CPU offload makes inference incredibly slow

Hardware:

• RTX 4080 FE (16 GB VRAM)
• Intel i9 CPU
• 64 GB RAM
• CUDA 13.0
• Windows 11

Software:

• Python 3.11.0
• PyTorch 2.9.0.dev (nightly CU130)
• Chandra 0.1.8

Problem:
Chandra is completely unusable on Windows. Both from the app and from the client, it remains loading forever without doing anything.

What happens:

1. Model loads successfully (16.3GB Qwen2-VL)
2. Checkpoint fragments load at 100%
3. Then the message: "Some parameters are on the meta device because they were offloaded to the CPU"
4. Perpetual loading
5. GPU shows 79% utilization, but inference fails

Commands tried:

# GUI version
$env:CHANDRA_DEVICE="cuda"
chandra_app

# CLI version
chandra input.png ./output --method hf --device cuda

Both have the same problem: CPU offload makes it unusable.

Installation Issues:

• The documentation doesn't mention the requirements for PyTorch 2.8+
• PyPI versions 0.1.7 and 0.1.8 require torch >=2.8.0
• Stable CUDA builds of PyTorch 2.8+ don't exist for Windows
• I had to use nightly builds (undocumented)
• It took me 9 hours to figure out how to install correctly

Expected:
With RTX 4080, it should run in 1-3 seconds

Actual:
Infinite loading

This makes Chandra completely unusable on Windows.
They call it the #1 software currently, but for me, it's total crap, a nightmare, and I wasted 9 hours of my life.