# Google 工程师在 Ironwood （TPUv7） 上优化 Qwen 3.5-397B MoE 模型

- 来源：Google Developers Blog（RSS）
- 发布时间：2026-07-15 04:53
- AIHOT 分数：38
- AIHOT 链接：https://aihot.virxact.com/items/cmrl4qm1b003abilm1v9kwn93
- 原文链接：https://developers.googleblog.com/systems-engineering-playbook-optimizing-qwen-35-397b-moe-on-ironwood-tpu7x

## AI 摘要

Google 工程师为在 Ironwood (TPUv7) 上部署 397B 参数的 Qwen 3.5 MoE 模型，开发了一套模块化 JAX/Pallas 优化栈。通过混合数据并行与专家并行（DP+EP）拓扑绕过硬件分片限制，结合层级化 reduce-scatter 等自定义底层通信融合优化跨设备 token 路由，并利用硬件感知的自定义内核（如 Batched Ragged Page Attention 和全融合 Gated DeltaNet 块），最终在 prefill 密集型负载上实现了高达 4.7 倍的推理加速，使系统吞吐量接近理论 roofline 极限。

## 正文

To serve the 397B-parameter Qwen 3.5 Mixture-of-Experts (MoE) model on Ironwood TPUs, engineers developed a modular JAX/Pallas optimization stack that achieved up to a 4.7x inference speedup for prefill-heavy workloads. The team bypassed severe hardware sharding constraints by deploying a hybrid Data Parallelism and Expert Parallelism (DP+EP) topology, paired with custom low-level communication fusions like a hierarchical reduce-scatter to optimize cross-device token routing. Finally, by executing hardware-aware custom kernels—such as Batched Ragged Page Attention and a fully-fused Gated DeltaNet (GDN) block—they successfully saturated HBM bandwidth and TensorCore MXUs to push system throughput near its theoretical roofline limits.