feat: 8sync gguf serve --balance -- model-aware auto-tuning

[8syncdev]

Summary

Added --balance mode to 8sync gguf serve that selects optimal llama-server config from live hardware state.

How it works

1. Reads model file size to estimate bytes per layer
2. Queries live VRAM free + GPU temperature via nvidia-smi
3. Calculates max GPU layers that fit while leaving 300 MiB headroom for CUDA overhead + KV cache
4. Thermal guard: reduces GPU layers 20% if temp >= 75 C to keep the card cool and power-efficient
5. Enables flash-attn only when all layers fit on GPU (required for cache-type-v quantization)
6. Scales context size with available VRAM: 4K / 8K / 16K

Usage

# Auto-tune everything
8sync gguf serve --engine-path "C:\...\run" --model-path "C:\...\model.gguf" --balance

# Balance + override one param on top
8sync gguf serve ... --balance --ctx 16384 --gpu-layers 26

# View what balance would choose for your model right now
8sync gguf info --model-path "C:\...\model.gguf"

Per-param overrides (stack on any mode)

Flag	Controls
--gpu-layers N	GPU offload layers
--ctx N	Context window tokens
--threads N	CPU generation threads
--parallel N	Concurrent request slots
--batch N	Logical batch size

New commands added

• 8sync gguf list — running servers with health [OK/LOADING/UNREACHABLE], tok/s from /metrics, GPU layers, mem, uptime
• 8sync gguf info — hardware snapshot, mode comparison table, param reference, balance preview, running server config dump
• 8sync gguf info --model-path <file> — shows what balance would compute for that model given current VRAM/temp

Commit

5dfea7e