Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

Qwen3.8-Flash-Next on 4 GPUs: device_map="auto" Leaves GPU 0 Empty and Offloads 22 GB

If you load Qwen3.8-Flash-Next with transformers on four 80 GB GPUs, device_map="auto" leaves GPU 0 empty and offloads 22 GB of weights to the CPU, even though the model fits on the GPUs it was given.

The article explains that the misplacement comes from Qwen4’s large “n-gram table”/embedding component designed to reside in host (RAM) memory; transformers’ device-map planner incorrectly reserves too much room for the layer carrying this table, causing GPU 0 to end up empty and more weights to be streamed/offloaded to CPU. It details how the table’s size and the planner’s block-placement rules interact, why the opposite can happen on larger GPUs (the table then fits and gets placed on-GPU, wasting HBM), and provides a Python script to pre-compute a better device_map that excludes the table and fills GPUs up to a chosen utilization target. The author includes example outputs for different GPU configurations, shows how to load the model using the generated plan.json, discusses KV-cache budgeting, and gives guidance on what to change depending on the user’s hardware. The piece closes with a “My take” suggesting a targeted workaround to prevent unnecessary PCIe transfers and improve utilization.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI