engine = "vllm"
name = "Gemma 4 12B IT (MTP)"
docker_image = "vllm/vllm-openai:gemma4-unified"
model = "google/gemma-4-12B-it"
# docker_image = "ghcr.io/timothystewart6/vllm-gb10:v0.23.0-gb10.1"
# model = "RedHatAI/gemma-4-12B-it-NVFP4"
tensor_parallel_size = 1
pipeline_parallel_size = 1

# Single-node recipe: Gemma 4 12B with BF16, FP8 KV cache,
# MTP speculative decoding (5 tokens), 196K context.

entrypoint = ["vllm", "serve"]

args = [
    "--dtype", "bfloat16",
    "--kv-cache-dtype", "fp8",
    "--gpu-memory-utilization", "0.75",
    "--max-model-len", "196608",
    "--max-num-seqs", "2",
    "--max-num-batched-tokens", "8192",
    "--enable-prefix-caching",
    "--enable-chunked-prefill",
    "--enable-auto-tool-choice",
    "--tool-call-parser", "gemma4",
    "--reasoning-parser", "gemma4",
    "--generation-config", "vllm",
    "--safetensors-load-strategy", "prefetch",
    "--speculative-config", """
'{"method":"mtp","model":"google/gemma-4-12B-it-assistant","num_speculative_tokens":5}'""",
    "--limit-mm-per-prompt", """
'{"image":4,"audio":1,"video":1}'""",
    "--trust-remote-code",
    "--served-model-name", "gemma4-12b",
]

[env]
TRITON_CACHE_DIR = "/root/.cache/triton"