GPUnity Web Demo

FastAPI backend + static web app for GPU workload scans and marketplace search.

Features

• Supabase auth + repo/scan sync (optional); demo mode stores data locally
• Add repos via public GitHub URL, folder upload, or zip upload
• Codebase scanning: framework, workload profile, entrypoints, extracted params, risk analysis
• Optional OpenAI enrichment + AI suggestions
• Scan JSON viewer, copy, and download
• Per-repo workload and scan overrides
• GPU marketplace search with filters + ranking
• Providers: Vast.ai and Runpod
• GPUnity Agent page for repo analysis + GPU provisioning via Skypilot

Quick Start

Backend:

# one-time setup
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# run
touch .env
uvicorn backend.main:app --reload --port 8000

Frontend:

# update web/config.js then
source .venv/bin/activate
python3 -m http.server 5500 --bind 127.0.0.1 --directory web

Open:

• http://127.0.0.1:5500/index.html (landing)
• http://127.0.0.1:5500/app.html (workspace)
• http://127.0.0.1:5500/agent.html (GPUnity Agent)

Config (minimum)

Backend .env:

• OPENAI_API_KEY (optional)
• ANTHROPIC_API_KEY (optional, enables agentic scanner)
• ANTHROPIC_MODEL (optional, default claude-sonnet-4-20250514)
• AGENTIC_MAX_ITERATIONS (optional, reduce agentic steps to avoid rate limits)
• AGENTIC_MAX_TOKENS (optional, token cap per agentic call; default 4096)
• GITHUB_TOKEN (optional, increases agentic scan GitHub API limits)
• MAX_FILE_SIZE_MB
• ALLOWED_ORIGINS
• SKYPILOT_ENABLED (set to true to run sky launch)
• SKYPILOT_CLOUD (default aws)
• SKYPILOT_STOP_ACTION (stop or down)
• SKYPILOT_LAUNCH_TIMEOUT (seconds, default 900)
• SKYPILOT_KEEP_WORKDIR (set to true to keep extracted repo + YAML for debugging)
• SKYPILOT_USE_JOBS (set to true to use sky jobs launch)
• SKYPILOT_RETRY_UNTIL_UP (set to true to pass --retry-until-up for sky launch)
• SKYPILOT_DETACH_RUN (set to false to keep sky launch attached; default true)
• SKYPILOT_LOGS_NO_FOLLOW (set to false to allow log streaming; default true)
• SKYPILOT_ALLOW_ONDEMAND_FALLBACK (set to true to use spot with on-demand fallback)
• SKYPILOT_REGION / SKYPILOT_ZONE (optional pinning of location)
• SKYPILOT_JOB_RESTARTS (set to a number to enable job recovery retries)
• SKYPILOT_CATALOG_ENABLED (set to true to align recommendations with sky show-gpus)
• SKYPILOT_PIP_TIMEOUT (pip timeout seconds for setup, default 120)
• SKYPILOT_PIP_RETRIES (pip retries for setup, default 5)
• SKYPILOT_PIP_INDEX_URL (optional pip index override)
• SKYPILOT_PIP_EXTRA_INDEX_URL (optional pip extra index)
• GPU_AGENT_ENABLED (set to false to disable the LangChain GPU planner)
• GPU_AGENT_MAX_ATTEMPTS (search retries, default 4)
• GPU_AGENT_MAX_QUESTIONS (clarifying questions cap, default 4)
• GPU_AGENT_MODEL (override the OpenAI model for GPU planning)
• GPU_AGENT_TEMPERATURE (LLM temperature, default 0.2)

Frontend web/config.js:

• apiBaseUrl
• supabaseUrl, supabaseAnonKey (optional)

Notes

• Public GitHub repositories only.
• AI scan requires an OpenAI API key on the backend.

Deploy

• Render blueprint supported via render.yaml.
• scripts/render_build.sh writes web/config.js during static build.
• For Render production, keep ALLOWED_ORIGINS locked to your web domain.
• Rate limiting is preconfigured in render.yaml with Redis storage; update limits via RATE_LIMIT_* env vars.