Spaces:
Runtime error
Runtime error
Commit History
-
Reduce default GPU duration for High Detail mode from 60s to 30s
Inference has gotten faster, so the prior 60s default now overshoots actual runtime, same reasoning that dropped Fast mode's default to 15s. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>106074d -
Sync visible GPU duration slider default to Fast mode's 15s
templates/app.html still hardcoded the visible slider's initial value at 30, and mode_toggle.js only applied the mode's duration on a mode-button click, not on page load, so the UI kept showing 30s after 8a3b1b0 dropped the actual default to 15s. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>88ef5d4
-
Reduce default GPU duration for Fast mode from 30s to 15s
Fast mode uses lower resolution, fewer steps, and no text encoder offload, so 30s of reserved ZeroGPU time was more than actually needed. Closes #15 Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>8a3b1b0 -
Speed up fast mode: lower resolution cap, skip needless VAE tiling, fewer steps
Fast mode's 768px (was 1024px) output now stays under the VAE's tiling threshold, so it decodes in a single pass instead of a tiled decode with blend overhead that was only ever needed for high-detail's 2048px case. Also gives fast mode a lower default step count (3 vs 4), following the existing mode-scoped-default pattern already used for GPU duration. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>9fcf338
-
Load model on CPU at startup to bypass slow ZeroGPU disk packing
spaces 0.51.0 (released 2026-07-10) switched to disk-based tensor packing (~19s pack + ~19s unpack per inference), breaking the ~7s inference that worked with spaces 0.50.4's pinned-memory approach. HuggingFace force-injects spaces==0.51.0 so pinning requirements.txt doesn't help. Instead: - Load transformer and pipeline on CPU at startup (ZeroGPU has nothing to pack since no tensors are on GPU) - pipe.to(device) inside @spaces.GPU: ~3-4s PCIe transfer vs ~19s disk - pipe.to("cpu") in finally block so ZeroGPU has nothing to repack Also reverts bad spaces==0.50.4 pin (build error) and raises GPU duration defaults to accommodate the explicit model-move time. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>6c0477d
-
Set default GPU duration for high-detail mode to 40s
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>96b6246
-
Set default GPU duration for high-detail mode to 30s
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>9f5cd46
-
Add user-adjustable GPU Duration slider to Advanced Settings
Exposes the spaces.GPU duration as a UI slider (10–120s, step 5s) so users can tune quota usage per request. Switching modes auto-resets the slider to its default (fast=20s, high_detail=60s). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b36328a
-
Wrap mode_toggle.js as arrow function to match Gradio js= convention
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>8b178bf -
Load mode toggle JS via demo.load instead of stripped <script> tag
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>a3f2ffd