Chunk 42.0

## Summary This chunk focused on deploying **Qwen3.5-122B-A10B-FP8** across **2× NVIDIA GB10 DGX Spark** systems (SM121 Blackwell, ARM Cortex-X925, 120GB unified memory each, InfiniBand RoCE interconnect). The assistant explored both nodes, found an existing Docker-based vLLM setup running GLM-4.7-Flash, stopped it to free GPU memory, downloaded the 119GB FP8 model from HuggingFace, and rsynced it to the second Spark at ~640MB/s over the IB link. After discovering that SGLang's official spark image lacked Qwen3.5 support and its multi-node NCCL initialization hung indefinitely, the assistant pivoted to the `hellohal2064/vllm-qwen3.5-gb10` image (vLLM 0.17.1rc1, specifically built for Qwen3.5 on GB10) and built a Ray-based multi-node deployment. The multi-node setup required solving several networking and resource management challenges: Ray's auto-detection used the external IP (10.1.230.180) which was unreachable from the second Spark, so the assistant forced node IPs to the IB subnet (192.168.200.x) with `--node-ip-address` and set `GLOO_SOCKET_IFNAME`/`NCCL_SOCKET_IFNAME` to the correct RoCE interface. Ray's OOM killer (95% memory threshold) killed the process during CUDA graph capture on the head node, requiring `RAY_memory_monitor_refresh_ms=0` and reduced `--gpu-memory-utilization`. The final deployment uses NCCL over `NET/IBext_v11` for inter-node tensor parallelism (TP=2), achieving ~27 tok/s single-request throughput with correct reasoning output and tool calling enabled. The user then requested a relaunch, but both the Proxmox RTX PRO 6000 setup and the Spark containers were unreachable (network issue and reboot respectively), leaving the Spark services running the old GLM containers instead.

The Dual DGX Spark Deployment: From Reconnaissance to Multi-Node Inference on Blackwell ARM 1827 words

Message Articles

Subagent Sessions