# LLM infrastructure config — sourced by all llm-*.sh scripts
# Updated: 2026-06-04
#
# Architecture: oMLX 0.4.1 + oMLX compatibility proxy.
# Gemma4 26B-A4B runs as VLMBatchedEngine — vision, tools, and thinking
# from a single endpoint on :8080. No other models active.
#
# ⚠ VERSION PIN: All patches (model_discovery.py, config.json vision_config)
#   are specific to oMLX 0.4.1. Do NOT run brew upgrade omlx without re-running
#   setup.sh and re-validating all three VLM capability tests.
#   Pin is set by setup.sh via: brew pin omlx

OMLX_VERSION=0.4.1
RMLX_BIN=/opt/homebrew/bin/omlx

# ── Base directories ──────────────────────────────────────────────────────────
LLM_BASE=$HOME/.llm
MODEL_DIR=$LLM_BASE/models
LOG_DIR=$LLM_BASE/logs
RUN_DIR=$LLM_BASE/run
CACHE_DIR=$LLM_BASE/cache

# ── Primary model — Gemma 4 26B-A4B VLM ──────────────────────────────────────
# Vision + tool calling + thinking. 256K context, ~15 GB at 4-bit.
# Loaded as VLMBatchedEngine (vision_config patched into config.json).
# model_type_override: vlm set in ~/.omlx/model_settings.json.
PRIMARY_MODEL=$MODEL_DIR/gemma-4-26b-a4b-it-4bit
PRIMARY_PORT=8080
PRIMARY_LOG=$LOG_DIR/primary.log
PRIMARY_PID=$RUN_DIR/primary.pid

# ── oMLX compatibility proxy ──────────────────────────────────────────────────
# Routes reasoning_content → content in streaming deltas (Gemma4 thinking).
# Listens on :8090, forwards to :8080.
# Requires: $HOME/.llm/proxy-venv (pip install aiohttp)
PROXY_PYTHON=$HOME/.llm/proxy-venv/bin/python3
PROXY_PORT=8090
PROXY_PID=$RUN_DIR/omlx-proxy.pid
PROXY_LOG=$LOG_DIR/omlx-proxy.log