Helix decouples attention and FFN sharding, applying KV parallelism during attention and tensor/expert parallelism during FFN, to improve throughput and latency for multi-million-token LLM decoding.
Medha: Efficiently Serving Multi-Million Context Length LLM Inference Requests Without Approximations, 2025
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.DC 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
Helix Parallelism: Rethinking Sharding Strategies for Interactive Multi-Million-Token LLM Decoding
Helix decouples attention and FFN sharding, applying KV parallelism during attention and tensor/expert parallelism during FFN, to improve throughput and latency for multi-million-token LLM decoding.