Spaces:

---

Lightricks

/

LTX-2-3

like 335

Runtime error

Hi! Duplicated this Space to dig into why startup was hanging, and ended up finding two separate things going on. Sharing in case it's useful.

1. Startup hang from bucket FUSE + safetensors mmap

After 593bb6a moved the weights to the bucket mount, the safetensors loader hits the FUSE-mmap deadlock pattern (page-fault storm on the FUSE mount). safetensors.safe_open() is the trigger, and any large .safetensors on the bucket effectively never finishes loading. The original Space then times out at the 30-min startup window.

transformers recently solved the exact same thing with disable_mmap=True and automatic hf-mount detection (huggingface/transformers#45547). diffusers has had disable_mmap since v0.33.0. The clean fix is probably an equivalent flag (and ideally auto-detection) on SafetensorsStateDictLoader.load in Lightricks/LTX-2.

2. xformers FA3 crash on Blackwell

Since ZeroGPU's backend migrated from H200 to RTX PRO 6000 (Blackwell) on 2026-05-12, generation dies inside xformers' Hopper-only FA3 kernel:

CUDA error (.../hopper/flash_fwd_launch_template.h:180): invalid argument

xformers' fa3_available() gates FA3 on device_capability >= (9, 0), which incorrectly also matches Blackwell. The real fix is on the xformers side, but a one-line Space-side workaround is enough to unblock things:

from xformers.ops.fmha import _set_use_fa3
_set_use_fa3(False)

Attention falls back to FA2 / cutlass-pt, both of which work on Blackwell.

Working duplicate

Applied both as monkey-patches at the app.py level in a public duplicate: hysts-duplicates/LTX-2-3. The Space starts up and generates videos.