docs: INSTALL vLLM (8010, cache, VRAM), NGC/perms; preset 8010

[tokk-nv]

Summary

Documentation and small config/UI updates: vLLM on port 8010 with cache volume and GPU memory notes, NGC/permission troubleshooting, and preset/UI aligned to 8010.

Changes

INSTALL.md

• vLLM: Standardize on port 8010 (avoids Riva 8000–8002). Add optional torch.compile cache volume (-v ~/.cache/vllm:/root/.cache/vllm) and mkdir -p ~/.cache/vllm before first run so the directory is user-owned.
• Memory: Clarify that vm.drop_caches=3 frees system (CPU) memory only, not GPU VRAM. Note to stop the first vLLM container and wait 30–60 s before starting again.
• Troubleshooting: New subsection for ValueError: Free memory on device cuda:0 (...) is less than desired GPU memory utilization (stop other process, wait for VRAM).
• NGC Cosmos: Soften “org must be nvidia” to “try org nvidia if you see 403”; note that download often works with default org. Permission-denied bullet points to “Fix Hugging Face cache permissions (root-owned).”

Preset & UI

• presets/cosmos-reason.yaml: api_base set to http://localhost:8010/v1.
• app.js: vLLM API preset updated from 8003 to 8010 (label and URL).

Backend (openai.py)

• Model list: Docstring updated to state that we probe /api/tags first (Ollama can run on any port), then fall back to /v1/models.

Testing

• Verified vLLM Docker run with cache volume; second run reuses torch.compile cache (~24 s faster).
• Confirmed NGC Cosmos download can succeed with default org when permissions are correct.

Screenshots

[adsahu-nv]

Looks good! Thanks