Constrained decoding is essential in generative retrieval, where document identifiers generated directly from a query must exactly match a predefined library of valid IDs. At scale, decoding is often constrained using a trie with beam search but most implementations run on CPU. Limited parallelism then makes trie traversal and candidate validation a serving bottleneck as beam width grows. We present FlashTrie, which addresses this limitation by optimizing constrained beam search on GPUs. It introduces an integer-aware succinct trie layout that uses bit compression to reduce memory footprint while keeping the full index in GPU high-bandwidth memory reducing memory stalls, and a cooperative CUDA kernel that performs beam expansion, validation, and pruning entirely on-device without per-step host orchestration. It further replaces CPU-style irregular lookup and heap maintenance with GPU-aware parallel primitives, improving warp utilization and reducing divergence. Together, these designs significantly reduce decoding latency and increase throughput while preserving retrieval quality. On a library of 800M keywords with beam widths up to 1000, FlashTrie reduces trie-search latency to under 3 ms, achieving up to 24x speedup over a highly optimized multi-threaded CPU baseline. These improvements enable FlashTrie to scale beam sizes by up to 5x in latency-critical applications such as sponsored search. In a large-scale online A/B experiment on a popular commercial search engine, it delivers a statistically significant +0.71% revenue lift, enabling real-time constrained decoding at a scale previously feasible only offline. The FlashTrie code will be publicly released after the review process.
FlashTrie: A GPU-Accelerated Constrained Beam Search for Generative Retrieval
Constrained decoding is essential in generative retrieval, where document identifiers generated directly from a query must exactly match a predefined library of valid IDs. At scale, decoding is often constrained using a trie with beam search but most implementations run on CPU.
- Preview

- Year
- 2026
- Hosting
- Full text hostedCC-BY-4.0
Cite
Notes
Only stored in your browser.
Attribution
- Abstract & full text
- arxiv.org/abs/2607.10044CC-BY-4.0
- TL;DR
- Semantic Scholar