1080p >=5s OOMs during VAE decode

[taskmasterpeace]

Severity: Medium (confirms a known issue)

backend/handlers/video_generation_handler.py (~L117-136, 1080p → 1920x1088, num_frames grows with duration); backend/services/fast_video_pipeline/ltx_fast_video_pipeline.py (~L113, L124-125).

generate() uses a single default_tiling_config(); the VAE decode of a 1920x1088 × ~121-frame latent (5s @ 24fps) OOMs. FFN chunking patches the transformer feedforward, not the VAE decode, so it does not help this path. Decode tiling is fixed and not scaled by resolution/frame count.

Fix

Scale VAE decode tiling (temporal + spatial) with resolution×frames, or force tiled/chunked VAE decode above a pixel-volume threshold.