PhobosStream: Run 14B Video DiTs on 8GB GPUs With 788MB Peak VRAM via Disk Streaming
martianacademy · reddit · 2026-10-31
The author open-sourced PhobosStream and a ComfyUI node that stream massive DiT models—Wan2.1 (14B), FLUX.1 (12B), Z-Image-Turbo (11B)—from NVMe SSD, letting 8GB consumer GPUs and 16GB Macs run models that normally need 24-40GB VRAM.
Benchmarks (Apple Silicon Metal)
- Monolithic baseline: 38.4GB (instant OOM)
- PhobosStream peak VRAM: 788.11MB (>97% saved)
- Z-Image-Turbo 1024x1024: 4.79s/step, 38.33s for 8 steps
- SD-Turbo: 0.38s (10 it/s)
How it works
- Zero-copy mmap streaming: only the active transformer block (380MB) resides in VRAM, evicted as the next streams in
- Sequential offload: large text encoders (Qwen 3.4B / T5-XXL) run first, then dump before denoising
- Backends: native Metal via MLX, CUDA/ROCm, GGUF
A ComfyUI Manager PR (#3453) has been submitted.
More from Infra
- Rust MLIP engine llmlip.rs beats PyTorch MACE by 17.7x, no PyTorch needed — CatAstro_Piyush · 2026-10-31
- AGI House and Lenovo host Personal Agent real-world deployment Build Day — agihouse_org · 2026-10-31
- RTX 4090 BF16 unlock via UEFI toolbench: ~10% faster image gen, still trails RTX 5090 — devmoching · 2026-10-31
- Running a Plan-then-Execute Local Agent on 12GB VRAM: Big Model Plans, Small Model Acts — Loose_Doubt367 · 2026-10-31
- Over $100M in Data-Center Equipment Stolen Since June as GPS Spoofing Emerges — BuddyWeary653 · 2026-10-31
- Community asks: where is the Qwen 4 35B MoE as the Qwen 4 lineup nears — solyarisoftware · 2026-10-31