Want to achieve near-100% GPU utilization processing millions of text documents with Qwen3? Most ML teams struggle to keep GPUs from sitting idle, juggling multi-step pipelines across different clusters. With Daft, we seamlessly pipeline network downloads, CPU preprocessing, GPU inference, and distributed writes. All in one script, on a single cluster, with almost 100% GPU utilization. See how we used Qwen3-Embedding-0.6B to compute embeddings on the RedPajama Dataset across 8×A10G nodes, while writing simultaneously to turbopuffer—without breaking a sweat!