Helix Parallelism: Rethinking Sharding Strategies for Interactive Multi-Million-Token LLM Decoding

TL;DR

Helix Parallelism offers a hybrid sharding strategy for large language models that significantly reduces latency and improves throughput for decoding with long token histories, enabling real-time inference.

Contribution

It introduces Helix Parallelism, a novel hybrid sharding approach combining KV parallelism and tensor parallelism, with a lightweight communication step to optimize multi-GPU decoding.

Findings

Reduces token-to-token latency by up to 1.5x.

Supports up to 32x larger batch sizes under the same latency.

Enables real-time inference with ultra-long sequences.

Abstract

As LLMs scale to multi-million-token KV histories, real-time autoregressive decoding under tight Token-to-Token Latency (TTL) constraints faces growing pressure. Two core bottlenecks dominate: accessing Feed-Forward Network (FFN) weights and reading long KV caches. While Tensor Parallelism (TP) helps mitigate the cost of FFN weight reads, it does not scale well for attention. When TP width exceeds the number of KV heads, it leads to inefficient KV duplication, limits parallelism, and constrains batch size. Simultaneously, DRAM reads for long KV histories scale linearly with batch size, further capping efficiency. We introduce Helix Parallelism, a hybrid execution strategy that applies KV parallelism during attention to shard KV caches across GPUs, then reuses the same GPUs for TP in dense LLMs or TPxExpert Parallel (EP) in MoEs during FFN computation. To preserve exact attention behavior, Helix includes a lightweight communication step. To minimize the exposed communication cost, we introduce Helix HOP-B. Helix HOP-B effectively minimizes communication overhead through batchwise overlap, preserving low TTL while improving GPU efficiency. Compared to conventional parallelism approaches, Helix reduces TTL by up to 1.5x at fixed batch sizes and supports up to 32x larger batches under the same latency budget for DeepSeek-R1, pushing forward the throughput-latency Pareto on Blackwell and making real-time inference with ultra-long-sequence practical.