engine = "vllm"
name = "GLM-5.2 NVFP4 (vLLM)"
docker_image = "vllm/vllm-openai:glm52"
model = "Mapika/GLM-5.2-NVFP4"
parent_model = "zai-org/GLM-5.2"
required_nodes = 4
tensor_parallel_size = 4
pipeline_parallel_size = 1
distributed_backend = "mp"

# 410B NVFP4 model on 4× DGX Spark (GB10 Blackwell), 128 GB shared each.
# TP=4 → ~103 GB/GPU for weights, ~25 GB for KV cache at 0.90 util.
#entrypoint = ["vllm", "serve"]

args = [
#    "--kv-cache-dtype", "fp8",
    "--gpu-memory-utilization", "0.85",
    "--load-format", "safetensors",
    "--enable-prefix-caching",
#    "--speculative-config", """
#'{"method":"mtp","num_speculative_tokens":5}'""",
    "--tool-call-parser", "glm47",
    "--reasoning-parser", "glm45",
    "--enable-auto-tool-choice",
#    "--enable-expert-parallel",
    "--trust-remote-code",
    "--served-model-name", "glm52",
]

[env]
# Fast safetensors loading (GPU‑side decompress)
SAFETENSORS_FAST_GPU = "1"
# Allow long context lengths
VLLM_ALLOW_LONG_MAX_MODEL_LEN = "1"
# Blackwell MoE kernels
#VLLM_USE_B12X_MOE = "1"