Spaces:
Running
Running
Commit History
-
Stream animation frames inline via gr.HTML data URIs Running
The previous attempt cut payload size but left a structural problem: gr.Image in Gradio 5 saves each yielded PIL image to a temp file and sends the client a URL — the client then fetches the image with a second HTTP request per frame. Over HF Spaces' RTT, that fetch dominates each frame's budget and the 15-frame animation collapses into the final state by the time the client finishes the queue. Switch the board widget from gr.Image to gr.HTML and inline each frame as <img src="data:image/webp;base64,..."> via _img_to_html. The bytes ride inside the SSE event itself — no second fetch — so frames render the moment the client receives them. Also revert render_board figsize/dpi to (7, 7.6) / 120 so the rendered board doesn't look degraded. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>dc8f574 -
Cut Gradio per-frame payload + render cost
Slider previews and animation frames were spending most of their budget on render and wire transfer, not on streaming logic: - render_board ran matplotlib at dpi=120 figsize=(7, 7.6), producing an ~840×912 PNG for a widget displayed at 560px. Drop to dpi=110 figsize=(5, 5.4) — ~550×594, matches the display size, ~3× faster. - gr.Image transmitted PNG by default (~100-200KB per frame). Set format="webp" so each frame ships an order of magnitude smaller. - Slider preview was firing on every drag step via slider.change, flooding the queue. Switch to slider.release so we only re-render once the user lets go. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>6c5e416 -
Bump gradio pin to >=5 to satisfy openenv-core's websockets dep
uv resolver rejected gradio 4.44 alongside openenv-core 0.2.x: gradio-client 1.3.0 requires websockets<13 while openenv-core needs websockets>=15.0.1. Only gradio 5.x is compatible. The streaming / queue / mount_gradio_app APIs we rely on are unchanged in 5.x. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>9769b8b -
Fix Gradio animation streaming on HF Spaces
Frame-by-frame animation in the Gradio UI was collapsing to a single final frame on the deployed Space because: - The same Blocks instance was mounted at both "/" and "/web", which breaks SSE/queue routing — Gradio's frontend resolves queue endpoints relative to the page URL and the two mounts collide on session IDs. - No explicit queue was configured, so streaming generators had no reliable channel for incremental yields once mounted under FastAPI. - Gradio was unpinned and pulled in transitively, so HF builds could resolve to a different version than the one tested locally. Drop the duplicate /web mount, call queue() explicitly, and pin gradio>=4.44,<5 in both pyproject.toml and the Dockerfile install. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>d0a7466
-
Blog: remove em dashes + add baseline-obstruction section
- Replace all 42 em dashes with context-appropriate alternatives (commas, colons, parentheses, rephrasing) for cleaner prose. - Add new subsection under "Building the physics" covering the ICF rule that the striker can't be placed on top of a coin, and how _find_valid_placement nudges the placement to the nearest clear spot (with info["placement_x_actual"] reported back to the agent). - Minor copy tweak in final sentence ("benchmark" -> "environment"). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>519fcab -
Nemotron-3-Super-120b results + HF blog draft
- Add Nemotron-3-Super-120b-a12b (Nebius) benchmark run: +11.94 reward, 13 coins, 100% win rate, 97% ICF compliance, 1.18 efficiency — beats MiniMax on every metric at the same seed/horizon. - Add frontier-model head-to-head table (MiniMax vs Nemotron). - Add blog post draft under blog/carrom-rl-env.md. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>9296d7f -
README: rename env to "Carrom RL Env" and correct green-agent wording
- Rename title from "OpenEnv Carrom" → "Carrom RL Env" (the canonical env name). - Rewrite intro paragraph so it reflects the current capabilities and the correct AgentBeats framing for the green agent (evaluator, not "efficiency wrapper"). - Fix stale "Green Agent efficiency wrapper" comment in the project-tree section. - Update HF Spaces frontmatter title to match. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>506bca2 -
Mount Gradio UI at root path so HF Spaces iframe shows it
HF Docker Spaces iframe-load the container's root path (/). Our Gradio UI was only mounted at /web, so the Space appeared blank. Keep /web for backward compat with existing docs, also mount at / for HF. Also drop the unsupported `base_path: /web` field from README frontmatter. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>aa50c6c -
Initial: OpenEnv Carrom v0.3.0 — ICF-compliant physics RL environment
Physics & env: - Pymunk-based 2D board with Coulomb board friction via per-body velocity_func (constant decel, not viscous damping). - Full ICF rules: due rule, queen cover, foul handling, striker baseline. - Decoupled pocket_capture_radius=0.09 from pocket_radius=0.06 so pieces in the wall-gap "dead zone" are caught correctly. Green agent (evaluator): - GreenCarromAgent implements the AgentBeats evaluator role: defines task suites (easy/standard/hard tiers), wraps the env, scores purple agents with ICF-aware metrics (reward, win rate, ICF compliance) + compute efficiency. Inference: - inference.py with multi-provider OpenAI-compatible API support (Nebius / OpenAI / HF / local vLLM via API_BASE_URL + key env var). - --web flag starts the server + Gradio UI for watching LLM play live. - Reasoning-model fallback to reasoning_content when content is null. Gradio UI: - "Auto-play with LLM" panel: endpoint/model/key/turns inputs, streams physics animations between each LLM …13b4881