Modular JAX/Pallas Stack Boosts Qwen 3.5‑397B MoE 4.7× on Ironwood TPUs
A reusable JAX/Pallas kernel library, combined with hybrid data‑plus‑tensor parallelism, lifted Qwen 3.5‑397B MoE inference to 3.1× faster on decode‑heavy and 4.7× on prefill‑heavy workloads on Ironwood (TPU v7x) hardware. The approach sidesteps sharding limits and is being folded into open‑source serving stacks like vLLM, enabling enterprise‑scale deployment.