Apply minimal safe perf wins for free-tier HF Space

Revert all perf tuning, restore working pre-tune config

Revert KV-cache quant + force flash-attn defaults

Fix --flash-attn flag: pass explicit on/off value

Merge branch 'main' of https://huggingface.co/spaces/innovatorved/llama-server

Tune llama-server for faster CPU inference on HF free tier

Initial commit: llama.cpp OpenAI-compatible server for Gemma 4 E2B

Keep upstream /app layout intact so dlopen finds GGML CPU backend plugin

Fix UID 1000 collision on ubuntu:24.04 (delete default 'ubuntu' user)

Use ubuntu:24.04 runtime to match upstream :server glibc/libstdc++ ABI

Use upstream ghcr.io/ggml-org/llama.cpp:server image (no source build)

Fix build: keep LLAMA_BUILD_TOOLS=ON (server target lives under it)

Slim down Docker build (server-only, no BLAS/curl, strip + symlinks)

Switch default model to gemma-4-E2B-it (UD-Q4_K_XL, ~3.2 GB)

Add llama.cpp OpenAI-compatible server for gemma-4-26B-A4B-it

initial commit