Disk offload mode leaks host RAM during denoising (RSS grows every step, >15x slowdown from swap thrashing)

Environment

• OS: Windows 11 (x64), 24 GB RAM
• GPU: RTX 4060 Laptop 8 GB
• Python 3.12.10, torch 2.11.0+cu128, bitsandbytes 0.50.0
• DiffSynth-Studio: current main, installed via pip install -e ".[quant]" (2026-08-03)
• Model: DiffSynth-Studio/MiniMax-H3-NF4 (disk-offload low-VRAM path)

Observation

Running MiniMax-H3 NF4 FL2VA with the official disk-offload vram_config (offload_device="disk", onload_device="cpu", preparing_device="cuda", computation_device="cuda"), host RAM usage of the python process grows monotonically with every denoising step and is never reclaimed:

• After the text-encoder phase (which alone peaks at ~15 GB), the process working set keeps climbing during denoising — ~11.3 GB by step 39/50 and still growing
• On a 24 GB machine this pushes the OS into swap thrashing: per-step time degrades from ~65-75 s/step (steps 1-3) to ~1000-1100 s/step (step 39+), i.e. a >15x slowdown
• A 50-step 5-second 640x640 video took ~10 hours instead of the ~1 hour the early-step speed would suggest

Expectation

In disk offload mode, weights that _load_from_disk loads into CPU RAM (module.load_state_dict(state, assign=True) in layers.py) should be released when the layer is offloaded again (offload() re-creates a quantized shell). Host RAM should stay roughly constant across steps, but it looks like the CPU-side copies or the rebuilt quant state from quantize.unflatten_state_dict() are kept alive somewhere, so each step's reloads accumulate.

Impact

This disproportionately affects exactly the audience the disk-offload path targets: low-VRAM and low-RAM machines (the model card advertises ~6 GB VRAM as enough). Short runs (<= 8 steps) finish before the leak hits the wall, which masks the issue in quick tests; any 50-step or long-video run degrades severely.

Happy to provide py-spy dumps, tracemalloc snapshots, or per-step RSS logs if helpful.

Reproduce a long denoising run with the official disk-offload vram_config while recording per-step RSS. Inspect _load_from_disk in layers.py, especially module.load_state_dict(state, assign=True), then follow offload() and quantize.unflatten_state_dict() to identify retained CPU-side state. Done means repeated reload/offload cycles keep host RAM roughly stable and avoid the reported slowdown.