← Field notes
Field note / 03Systems
Getting DreamX under the real-time budget on one GPU
A DreamX engineering note on resident workers, two-step denoising, compiled DiT and VAE paths, and the limits of overlapping throughput-bound stages.
A 5B world model generated blocks slower than continuous playout allows.
The path down ran through a resident streaming engine, two-step denoising, compiled DiT and VAE, and FlashAttention-3.
Two traps sat on that path: torch.compile's recompile limit silently reverting a long-lived worker to eager, and an overlap that cannot help when both halves are throughput-bound.
Windflow EngineeringModel optimization
Field notes are short on purpose: one decision, what it cost, what we kept. The long write-ups land here when they are ready.