#!/usr/bin/env bash
# model-for-opencode entrypoint. Runs inside the container; takes no arguments.
# The weights are at /weights/qwen3-4b (read-only); nothing is fetched at run time.
set -eu
echo '{"phase":"loading","detail":"loading Qwen3 4B into the GPU"}' >> "$M64_STATUS"

# Every setting is here, fixed: callers can't change them (a chat capsule only
# accepts messages, tools, tool_choice, response_format, max_tokens, stream, seed).
#   --jinja                  use the model's own chat template: needed for tool calls,
#                            which is how opencode reads files and runs commands
#   --ctx-size 32768         opencode's system prompt and tools take ~10k tokens
#   -ngl 999                 every layer on the GPU
#   -fa on + q8_0 KV cache   halves the KV cache's memory, so 32k fits in ~6 GB
#   --host 0.0.0.0           inside the container's own network: only m64's front
#                            and gate can reach it
exec /app/llama-server \
  --model /weights/qwen3-4b/Qwen3-4B-Q4_K_M.gguf \
  --alias qwen3-4b \
  --jinja \
  --ctx-size 32768 \
  -ngl 999 \
  -fa on \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --host 0.0.0.0 --port "$M64_PORT"