TechPrufer
Get started
Google
Google logoGoogle · Feature trail

Rapid Storage for PyTorch

1 post · Apr 2012 – Apr 2026

What this trail covers

Rapid Storage, powered by Google's Colossus architecture, provides high-performance object storage in dedicated zonal buckets. It bypasses legacy REST APIs and utilizes persistent gRPC bidirectional streams for extreme throughput (15+ TiB/s), ultra-low latency (<1ms), and high QPS (20M+). Integration with PyTorch via fsspec allows researchers and developers to keep GPUs busier by addressing data loading and checkpointing bottlenecks in AI/ML workloads. This enables faster training and inference. The Gemini Batch API now supports embeddings and OpenAI compatibility, allowing for asynchronous processing at lower rates and with higher rate limits for cost-sensitive, latency-tolerant, or asynchronous use cases. This includes support for the Gemini Embedding model and an OpenAI compatibility layer for easier migration.

All Google trails
  1. Speeding Up AI: Bringing Google Colossus to PyTorch via GCSFS and Rapid Bucket- Google Developers BlogThis post announces the integration of Rapid Storage (powered by Google's Colossus architecture) with PyTorch via the fsspec interface. It details the performance benefits, including 15+ TiB/s throughput and <1ms latency, achieved by using stateful gRPC bidirectional streams and direct connectivity, bypassing legacy REST APIs. Zonal co-location is also highlighted for reduced latency. The post demonstrates a 23% performance gain in total training time and significant microbenchmark improvements (4.8x read throughput, 2.8x write throughput) compared to standard regional buckets. It provides installation instructions for gcsfs and a code sample for reading/writing to Rapid Buckets.
  2. Gemini Batch API now supports Embeddings and OpenAI Compatibility- Google Developers BlogThe Gemini Batch API now supports embeddings and OpenAI compatibility. This allows for asynchronous processing of embedding requests at 50% lower rates ($0.075 per 1M input tokens) and higher rate limits. It also provides an OpenAI SDK compatibility layer for easier migration of existing OpenAI batch workloads to Gemini. The post includes Python code examples for using the Gemini Batch API for embeddings and the OpenAI SDK compatibility layer.
  3. Google Cloud Storage - more value for performance- Google Developers BlogThis post announces enhancements to Google Cloud Storage, focusing on improved performance for demanding workloads. It highlights the introduction of new features and optimizations that deliver higher throughput and lower latency, making it more suitable for performance-critical applications. The improvements aim to provide greater value for users requiring high-performance storage solutions.
  4. Google Cloud Storage: high performance that just works- Google Developers BlogGoogle Cloud Storage is presented as a high-performance object storage solution that 'just works'. It highlights its ability to handle large amounts of data and provide fast access, making it suitable for demanding applications. The post emphasizes its reliability and scalability, suggesting it's a robust choice for developers.