config.yaml

# llama-swap config: two 1.2B specialists, swapped by the request's "model"
# field. One model resident at a time (2-vCPU tier); a swap costs a ~1.3GB
# mmap load (~5-10s), far cheaper than running the wrong specialist.
#
#   lfm-thinking : LFM2.5-1.2B-Thinking — deliberate reasoning / decisions
#   lfm-tool     : LFM2-1.2B-Tool       — tool-call execution
#
# Shared tuning (mirrors the other Spaces): split KV cache quant K=q8_0/V=q4_0,
# flash-attn on (required for quantized V), cache-reuse 256, 32K ctx / 2 slots.

macros:
  common: >
    /llama.cpp/build/bin/llama-server
    --host 127.0.0.1 --port ${PORT}
    --threads 2 --threads-batch 2
    --ctx-size 32768 --parallel 2
    --batch-size 512 --ubatch-size 512
    --cache-type-k q8_0 --cache-type-v q4_0
    --flash-attn on --cache-reuse 256
    --jinja --mmap

models:
  lfm-thinking:
    # A thinking model dies inside <think> on small budgets (measured: 512 was
    # exhausted mid-reasoning on trivial prompts). Strip the client's max_tokens
    # and enforce a 2048 default server-side; deepseek reasoning-format moves
    # <think> text into reasoning_content so content carries only the answer.
    cmd: >
      ${common}
      --model /models/LFM2.5-1.2B-Thinking-Q8_0.gguf
      --reasoning-format deepseek
      --n-predict 2048
      --alias lfm-thinking
    filters:
      strip_params: "max_tokens, max_completion_tokens"
    aliases:
      - thinking
      - decision

  lfm-tool:
    cmd: >
      ${common}
      --model /models/LFM2-1.2B-Tool-Q8_0.gguf
      --alias lfm-tool
    aliases:
      - tool
      - agent