TPU inference · Ongoing systems work

DeepSeek V4-Flash on TPU

A 32-chip TPU v4-64 bring-up validating prefill and decode across every attention mode.

Ongoing systems workLead · Gianluigi VitaleLast updated · 30 August 2026

Overview

A 32-chip TPU v4-64 bring-up validating prefill and decode across every attention mode.

Problem

New model architectures often reach accelerator ecosystems unevenly, leaving gaps in attention support, numerical validation, and end-to-end inference readiness.

Approach

The project brings the model up on a multi-chip TPU topology and validates each attention mode across both prefill and decode paths.

Results & current status

  • A 32-chip v4-64 bring-up was completed.
  • Every attention mode was validated with both prefill and decode operational.
  • Equivalent support later landed in the wider vLLM TPU inference ecosystem through parallel development.

Technical details

The work sits at the boundary of model architecture, distributed accelerator execution, attention implementation, and inference-system validation.

Reproducibility

This work is ongoing. Public code or an archival report will be linked when there is a stable release that can be reproduced independently.