Document client contract: thinking budget, think-stripping, routing policy

Thinking fixes: strip client max_tokens (enforce n-predict 2048), deepseek reasoning-format

llama-swap dual-specialist setup: Dockerfile

llama-swap dual-specialist setup: README.md

llama-swap dual-specialist setup: entrypoint.sh

llama-swap dual-specialist setup: config.yaml

Document split KV quant + cache-reuse env vars

Split K/V cache quant (K=q8_0,V=q4_0), 32K ctx, cache-reuse, tool-call smoke test

safety: cap threads at CPU_THREADS_MAX (default 2) to prevent oversubscription if detection over-reports

feat: cgroup-aware vCPU auto-detection (respects container CPU limit; no CPU_THREADS env needed)

fix: startup readiness loop must sleep on HTTP 503 (model loading), not only on connection errors — false 'startup timeout' left instance marked unloaded

harden: normalize legacy KV_CACHE_QUANT values (4bit->q4_0) so stale env vars can't crash llama-server

build: use prebuilt llama.cpp CPU binary (b9895) instead of source compile — fixes mtmd OOM hang, ~20x faster build

build: pull model at container runtime instead of baking into image (faster/reliable build)

v2.0: single-instance CPU proxy, native OpenAI tools+streaming, flash-attn+q4_0 KV, all-cores

Fix llama-server arguments and increase timeouts

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Fix: Use hf_hub_download for specific model file from deepreinforce-ai/Ornith-1.0-9B-GGUF

Fix: Skip Python bindings if directory doesn't exist

Fix llama.cpp build with CMake

Fix pip install with --break-system-packages

Update README with Ornith 1.0 details

Update Dockerfile for 2-instance build

Update app.py with 2-instance setup

Upload gemma-inference space

Upload gemma-inference space

Upload gemma-inference space

Upload gemma-inference space

Upload gemma-inference space

initial commit