conform.example.toml

# Copy this to conform.toml and edit. conform.toml is gitignored, so your
# local ports and model names stay yours.
#
#   kind = "reference"  loads the model in-process with transformers on CPU.
#                       Slow, no server needed, and treated as ground truth.
#   kind = "openai"     talks to any OpenAI-compatible /v1/completions server.
#
# capabilities is optional. Omit it to accept the adapter's defaults; set it
# explicitly to tell the suite what a server really supports, so unsupported
# features are skipped instead of reported as failures.

[[engines]]
name = "reference"
kind = "reference"
model = "HuggingFaceTB/SmolLM2-135M-Instruct"

[[engines]]
name = "ollama"
kind = "openai"
base_url = "http://localhost:11434"
model = "qwen2.5:0.5b"
capabilities = ["stop_sequences"]

# vLLM's CPU backend. Start it with:
#   vllm serve Qwen/Qwen2.5-0.5B-Instruct --device cpu --port 8000
[[engines]]
name = "vllm-cpu"
kind = "openai"
base_url = "http://localhost:8000"
model = "Qwen/Qwen2.5-0.5B-Instruct"
capabilities = ["stop_sequences", "logprobs", "top_k", "tokenize"]
enabled = false

# llama.cpp's built-in server:
#   llama-server -m model.gguf --port 8080
[[engines]]
name = "llama-cpp"
kind = "openai"
base_url = "http://localhost:8080"
model = "local"
capabilities = ["stop_sequences", "logprobs", "top_k"]
enabled = false