Context parallelism with AllToAll and load-balanced mini-chunks lets HSTU recommendation models support 5.3x longer interaction sequences at 1.55x throughput scaling.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.IR 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Scaling Generative Recommendations with Context Parallelism on Hierarchical Sequential Transducers
Context parallelism with AllToAll and load-balanced mini-chunks lets HSTU recommendation models support 5.3x longer interaction sequences at 1.55x throughput scaling.