StockForge ZeroGPU Runtime
Experimental GPU execution layer for StockForge V5.
Purpose
This Space is deliberately separate from the StockForge control plane. The
Android/Termux client remains responsible for prompt preparation and job
orchestration; this Space performs image generation only while a ZeroGPU
allocation is active.
Current model path
The first benchmark uses the official Z-Image-Turbo pipeline configuration
from Tongyi-MAI/Z-Image-Turbo, while replacing the diffusion transformer and
VAE with the StockForge FP8/AE files from ibank31/stockforge-models.
The StockForge Qwen FP8 file is retained as the canonical model artifact, but
is not manually injected into Transformers in this first runtime revision.
This avoids an unsafe ad-hoc state-dict loader and lets us validate the GPU
execution path first.
Quota strategy
- ZeroGPU
large only. xlarge costs 2x quota.
- GPU decorator uses a per-request duration estimator.
- Default generation is 1024x1024 and 8 steps, matching the Turbo workflow.
- Prompt validation and seed generation happen outside the GPU function.
- No model download occurs inside the generation function.
- No
torch.compile; ZeroGPU does not support it. AOT compilation can be
added after the baseline benchmark succeeds.
Benchmark goal
Measure actual GPU seconds per 1024x1024 / 8-step image. The first successful
run becomes the baseline for optimizing duration, AOT, and batching.
Important
Do not merge this branch into main until the Space has produced one real
image and the measured GPU duration has been recorded.
