Spaces:
Running on Zero
Running on Zero
Commit History
10 days ago
-
Flights suggestion: one-way on a single date (the 9B planner handles it; the round trip date picker tripped it up) Running
Verified signature6759892 -
type can submit (Enter) in the same step, so searches do not depend on finding a search icon
Verified signatureThe 9B planner often clicked Bing's search icon while the suggestion list was open, which did not run the search. type now takes submit=true and presses Enter itself; one step instead of two.4434ce5 -
Catch agent loops: compare screens by thumbnail, count repeated actions; fix back on in-page navigation; simpler Docling task
Verified signatureA run clicked the same missing tab 26 times: screenshots differed byte-wise (timestamps), so the stuck check never fired. Screens now count as unchanged when small grayscale thumbnails match, and repeating the very same action warns the planner from the 4th time and ends the run at the 8th. go_back on GitHub-style in-page navigation returns no response, which was reported as no previous page. The Docling suggestion no longer asks for a Releases tab that does not exist (the latest release is shown on the repo page).25f9914 -
Plan with Qwen3.5-9B instead of Qwen3.6-27B-FP8: as good on the demo tasks, about 4x faster per step
Verified signatureA side-by-side run on the 7 demo tasks (vLLM, real sites): Qwen3.5-9B 7/7, Qwen3.6-27B 6/7, Qwen3.5-4B 3/7. In transformers the 9B plans a typical step in ~0.7 s on an H100 vs ~2.8 s for the 27B in FP8, and it needs ~18 GiB instead of ~28 GiB. Per-step GPU reservation lowered from 30 s to 20 s. WEB_PLANNER=Qwen/Qwen3.6-27B-FP8 restores the paper planner.3767612 -
Thank @multimodalart : this demo builds on his DeskForge GUI Grounding Space
Verified signatured42043b -
Prepare for public launch: README and title, SafeSearch and family-DNS filter, friendly quota message, Triton tuning cache, 4 browsers, shorter suggestions
Verified signature- README: public title, arXiv/project/code links, models and datasets metadata - Web agent: Bing SafeSearch Strict; pages on domains Cloudflare's family resolver flags (adult, malware) are not opened - Friendly message when a visitor's ZeroGPU quota runs out (web agent, screenshot tool, duel) - TRITON_CACHE_AUTOTUNING=1 so new GPU workers reuse the warm-up's kernel tuning - Up to 4 browsers at once (bridge concurrency 6); runs survive a busy queue; clearer busy message - Suggestions: short tasks first, Docling release on GitHub, flights marked as a longer task - Planner replies with prose around the JSON are accepted; fewer memory notes; spare pre-opened browser156048d
11 days ago
-
Run GPU calls from the web agent and screenshot views as Gradio events (fixes ZeroGPU 'runs limit')
Verified signaturegr.HTML server functions reach ZeroGPU as anonymous API calls, so every web-agent step was billed to the tiny anonymous API quota. GPU ops now go through a hidden bridge component's submit event (gpu_bridge.py), like the duel's Start. Browser-only calls stay server functions. Also: duel GPU reservation 60s -> 30s.ec54602 -
Web agent view (Qwen3.6-27B-FP8 planner + DeskForge-Qwen3.5-4B) on top of the live desktop duel
Verified signaturebf47338
