# Laguna S 2.1 NVFP4 — 1× DGX Spark (GB10), TP=1.
#
# Model: poolside/Laguna-S-2.1-NVFP4
#   → https://huggingface.co/poolside/Laguna-S-2.1-NVFP4
#
# 117.6B total / 8.5B active MoE — NVFP4 quantized (~71 GB stored).
# Fits on a single DGX Spark (128 GB unified memory).
# 48 layers (36 SWA, 12 global), 256+1 experts, 262K context.
# KV cache in FP8.
#
# Optional DFlash speculative decoding with draft model:
#   poolside/Laguna-S-2.1-DFlash-NVFP4
#   → 15 speculative tokens, requires --max-num-seqs 32.
#
# REFERENCE:
#   https://huggingface.co/poolside/Laguna-S-2.1-NVFP4

engine = "vllm"
name = "Laguna S 2.1 NVFP4 (vLLM)"
docker_image = "ghcr.io/timothystewart6/vllm-gb10:v0.23.0-gb10.1"
model = "poolside/Laguna-S-2.1-NVFP4"
tensor_parallel_size = 1
pipeline_parallel_size = 1

entrypoint = ["vllm", "serve"]

args = """
--gpu-memory-utilization 0.85
--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
--max-num-seqs 32
--override-generation-config '{"temperature":0.7,"top_p":0.95}'
--served-model-name laguna-s-2.1
"""

[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"
MAX_JOBS = "4"