engine = "vllm"
name = "Laguna M.1 NVFP4 (vLLM)"
docker_image = "ghcr.io/timothystewart6/vllm-gb10:v0.25.1-gb10.1"
model = "poolside/Laguna-M.1-NVFP4"
required_nodes = 2
tensor_parallel_size = 2
pipeline_parallel_size = 1
distributed_backend = "mp"

# 225B total / 23B active MoE — NVFP4 quantized (~131B stored).
# 2× DGX Spark (GB10 Blackwell), 128 GB shared each.
# TP=2 → ~66 GB/node weights, room for 262K KV cache.
# Quantization auto-detected from model config.

entrypoint = ["vllm", "serve"]

args = """
--gpu-memory-utilization 0.8
--max-model-len 262144
--load-format instanttensor
--enable-prefix-caching
--tool-call-parser poolside_v1
--reasoning-parser poolside_v1
--enable-auto-tool-choice
--trust-remote-code
--kv-cache-dtype fp8
--enable-flashinfer-autotune
--served-model-name laguna
"""

[env]
SAFETENSORS_FAST_GPU = "1"
TORCH_CUDA_ARCH_LIST = "12.1a"
FLASHINFER_CUDA_ARCH_LIST = "12.1a"
VLLM_ALLOW_LONG_MAX_MODEL_LEN = "1"
VLLM_FLOAT32_MATMUL_PRECISION = "high"
PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True"