REVIEW 18 cited by
Sequence Parallelism: Long Sequence Training from System Perspective
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
abstract
Transformer achieves promising results on various tasks. However, self-attention suffers from quadratic memory requirements with respect to the sequence length. Existing work focuses on reducing time and space complexity from an algorithm perspective. In this work, we propose sequence parallelism, a memory-efficient parallelism method to help us break input sequence length limitation and train with longer sequences on GPUs efficiently. Our approach is compatible with most existing parallelisms (e.g. data parallelism, pipeline parallelism and tensor parallelism), which means our sequence parallelism makes 4D parallelism possible. More importantly, we no longer require a single device to hold the whole sequence. That is, with sparse attention, our sequence parallelism enables us to train transformer with infinite long sequence. Specifically, we split the input sequence into multiple chunks and feed each chunk into its corresponding device (i.e. GPU). To compute the attention output, we integrated ring-style communication with self-attention calculation and proposed Ring Self-Attention (RSA). Experiments show that sequence parallelism performs well when scaling with batch size and sequence length. Compared with tensor parallelism, our approach achieved $13.7\times$ and $3.0\times$ maximum batch size and sequence length respectively when scaling up to 64 NVIDIA P100 GPUs. With sparse attention, sequence can handle sequence with over 114K tokens, which is over $27\times$ longer than existing sparse attention works holding the whole sequence on a single device.
Forward citations
Cited by 18 Pith papers
-
HCMS: Head-Chunked Multi-Stream Pipeline for Communication-Computation Overlap in Long-Sequence Parallel Attention
HCMS partitions multi-head attention into chunks and pipelines them across dual CUDA streams to overlap communication and computation, delivering 10-17.5% speedup over Ulysses for 31K-56K token sequences.
-
Online Dynamic Batching with Formal Guarantees for LLM Training
ODB is an online batching system for distributed LLM training that forms batches post-preprocessing, provides formal deadlock-free guarantees via the Distributed Group Alignment Problem, and reports 1.58-3.78x through...
-
Towards Compute-Aware In-Switch Computing for LLMs Tensor-Parallelism on Multi-GPU Systems
CAIS delivers 1.38x end-to-end LLM training speedup over NVLS and 1.61x over T3 by making in-switch computing aware of computation memory requirements instead of treating communication as an isolated phase.
-
Auto-Scaling Heterogeneous Neural Processing Units for Energy and Cost-Efficient LLM Serving
NeuScale routes LLM inference requests to the most energy/cost-efficient configuration of heterogeneous NPU chips using roofline allocation and runtime auto-scaling.
-
Design-CP: Context Parallelism for Design of Protein Nanoparticles
Context-parallel inference for RFdiffusion 3 enables end-to-end all-atom design of large symmetric protein nanoparticles on multi-GPU hardware without retraining.
-
PHOENIX: Resilient LLM Training with Hot-Swapping via Zero-Overhead Checkpoint
DeadPool achieves zero-overhead checkpointing during error-free LLM training and hot-swapping recovery in under 40 seconds by replacing failed nodes without terminating the job.
-
PHOENIX: Resilient LLM Training with Hot-Swapping via Zero-Overhead Checkpoint
PHOENIX recovers permanent node failures in LLM training via hot-swapping of spares using zero-overhead per-step in-memory optimizer-state replication, finishing recovery in under 40 s on up to 512 GPUs.
-
ASAP: A Disaggregated and Asynchronous Inference System for MoE Prefill
ASAP is a disaggregated asynchronous inference system for the prefill phase of MoE models that removes DP-EP synchronization barriers and reports 90% higher SLO-compliant throughput than synchronous baselines.
-
Untied Ulysses: Memory-Efficient Context Parallelism via Headwise Chunking
UPipe chunks attention by head so QKV and all-to-all buffers scale with a small tunable chunk size rather than head count, enabling 5M-token Llama3-8B training on one 8×H100 node with throughput close to Ulysses.
-
TetriServe: Efficiently Serving Mixed DiT Workloads
TetriServe's step-level, deadline-aware sequence parallelism improves SLO attainment for mixed-resolution diffusion transformer serving by up to 32% over fixed-SP systems.
-
Learning to Shard: RL for Co-optimizing the Parallelism Degrees and Per-operator Sharding Dimensions in Distributed LLM Inference
An RL agent that co-optimizes parallelism degrees and per-operator sharding dimensions finds distributed inference strategies that beat random search and simulated annealing, and slightly outperform Megatron-LM heuris...
-
The Falcon Series of Open Language Models
Falcon-180B is a 180B-parameter open decoder-only model trained on 3.5 trillion tokens that approaches PaLM-2-Large performance at lower cost and is released with dataset extracts.
-
Mosaic: Towards Efficient Training of Multimodal Models with Spatial Resource Multiplexing
Apollo uses temporal-spatial multiplexing and a performance model to let multiple multimodal model modules share GPUs, delivering up to 1.31x training speedup in testbed experiments.
-
Charon: A Unified and Fine-Grained Simulator for Large-Scale LLM Training and Inference
Charon is a unified modular simulator that predicts LLM training and inference performance with under 5.35% error and identifies throughput improvements over baselines in a real deployment case.
-
Charon: A Unified and Fine-Grained Simulator for Large-Scale LLM Training and Inference
Charon is a unified fine-grained simulator that predicts LLM performance with under 5.35% error overall and under 3.74% for large-scale training, and it found a better inference configuration than an engineering baseline.
-
TPLA: Tensor Parallel Latent Attention for Efficient Disaggregated Prefill and Decode Inference
TPLA splits the latent KV cache across tensor-parallel GPUs while keeping every head's full view, yielding 1.79x and 1.93x decode speedups on DeepSeek-V3 and Kimi-K2 at 32K context with modest accuracy loss.
-
World Model on Million-Length Video And Language With Blockwise RingAttention
Presents open-source 7B models for million-token video and language understanding via Blockwise RingAttention, setting new benchmarks in retrieval and long video tasks.
-
Yi: Open Foundation Models by 01.AI
Yi models are 6B and 34B open foundation models pretrained on 3.1T curated tokens that achieve strong benchmark results through data quality and targeted extensions like long context and vision alignment.
Discussion (0). Sign in to comment.