THEOSIS//AI

Model library. Browse HuggingFace, see what fits in 128 GB, load and serve from one surface.

• Model grid with org, params, quant, memory fit vs 128 GB
• Serve config YAML panel for SGLang / vLLM / llama.cpp
• Disk cache management and HuggingFace pull stubs
• One-click load with memory budget validation

Every training run writes to SQLite and Parquet locally. No network dependency. Compare runs without leaving the terminal.

• Run comparison table with pinned metrics
• Loss and learning rate sparklines per run
• Config diff between any two experiments
• DuckDB shell for ad-hoc queries over Parquet

OTel-style span waterfall for tool-using agents. Every retrieval, rerank, and generation span with token cost and latency.

• Trace list with token cost and total latency
• Waterfall lanes with color-coded span types
• Span I/O detail panel with grounding verification
• JSONL and OTel-compatible export

Real-time GPU monitor. Utilization, memory pressure, power draw, bandwidth, per-process breakdowns.

• Gauge strip: GPU util, memory, power, temp, bandwidth, FP4 TOPS
• Per-process memory (VRSS) breakdown
• Kernel timeline with compute / memcpy / attention / nccl lanes
• Profiler sub-view with execution phase grouping