venv := "/home/madcat/lora-train"
data := "bt7274_v2.jsonl"

# Train LoRA adapter (stop vLLM first if running)
train:
    @echo "── BT-7274 LoRA training ──"
    @test -f {{data}} || (echo "MISSING: {{data}}" && exit 1)
    @echo "Data: {{data}} ($(wc -l < {{data}}) examples)"
    @echo "Activating venv: {{venv}}"
    bash -c 'source {{venv}}/bin/activate && python train.py'

# Dry run — load data, print stats, exit
check:
    bash -c 'source {{venv}}/bin/activate && python -c "\
    from datasets import load_dataset; \
    ds = load_dataset(\"json\", data_files=\"{{data}}\", split=\"train\"); \
    print(f\"Examples: {len(ds)}\"); \
    roles = {}; \
    [roles.update({r: roles.get(r,0)+1}) for ex in ds for m in ex[\"messages\"] for r in [m[\"role\"]]]; \
    print(f\"Roles: {roles}\"); \
    lens = [sum(len(m.get(\"content\",\"\") or \"\") for m in ex[\"messages\"]) for ex in ds]; \
    print(f\"Avg chars/example: {sum(lens)//len(lens)}\"); \
    print(f\"Max chars/example: {max(lens)}\"); \
    "'

# Stop vLLM server (frees GPU for training)
stop-vllm:
    -pkill -f "vllm.entrypoints"
    @echo "vLLM stopped"

# Start vLLM server with v2 adapter
serve:
    bash ~/start-vllm.sh

# Start vLLM with v2 adapter override
serve-v2:
    @echo "Starting vLLM with v2 adapter..."
    bash -c 'source /home/madcat/vllm-serve/bin/activate && \
    python -m vllm.entrypoints.openai.api_server \
      --model Qwen/Qwen2.5-7B-Instruct \
      --quantization bitsandbytes \
      --load-format bitsandbytes \
      --enable-lora \
      --lora-modules bt7274=/home/madcat/Projects/lora/bt7274-lora-v2 \
      --max-lora-rank 16 \
      --max-model-len 32768 \
      --enforce-eager \
      --enable-auto-tool-choice \
      --tool-call-parser hermes \
      --port 8000 \
      > /home/madcat/vllm-bt7274.log 2>&1 &'
    @echo "vLLM started (log: ~/vllm-bt7274.log)"
