engine = "vllm"
name = "Qwen-AgentWorld-35B-A3B (vLLM)"
docker_image = "ghcr.io/timothystewart6/vllm-gb10:v0.25.1-gb10.1"
model = "Qwen/Qwen-AgentWorld-35B-A3B"
tensor_parallel_size = 1
pipeline_parallel_size = 1

# Solo: ~70 GB BF16 weights, only 10 GA layers need KV cache (DeltaNet is linear).
# Use: sm121-launcher serve --recipe qwen-agentworld-35b --node dgx1.lan
# Nodes: sm121-launcher serve --recipe qwen-agentworld-35b --cluster 2nodes-dac
# Requires --language-model-only (no vision weights in checkpoint).

entrypoint = ["vllm", "serve"]

args = """
--max-model-len 262144
--reasoning-parser qwen3
--language-model-only
--trust-remote-code
--gpu-memory-utilization 0.90
--enable-prefix-caching
--enable-auto-tool-choice
--tool-call-parser qwen3_xml
--served-model-name qwen-agentworld-35b
"""