7 papers
OctoPipe: Reducing Pipeline Bubbles for Heterogeneous Models via Co-Optimizing Partitioning, Placement, and Scheduling
Jihu Guo, Tenghui Ma, Wei Gao +6
Pipeline parallelism is widely used to train large language models (LLMs). However, increasing heterogeneity in model architectures exacerbates pipeline bubbles, thereby reducing t…
Curriculum Sampling: A Two-Phase Curriculum for Efficient Training of Flow Matching
Pengwei Sun
Timestep sampling is a central design choice in Flow Matching models, yet common practice increasingly favors static middle-biased distributions (e.g., Logit-Normal). We sho…
PromptTuner: SLO-Aware Elastic System for LLM Prompt Tuning
Wei Gao, Peng Sun, Dmitrii Ustiugov +2
Prompt tuning has become a prominent strategy for enhancing the performance of Large Language Models (LLMs) on downstream tasks. Many IT enterprises now offer Prompt-Tuning-as-a-Se…
CONCUR: High-Throughput Agentic Batch Inference of LLM via Congestion-Based Concurrency Control
Qiaoling Chen, Zhisheng Ye, Tian Tang +7
Batch inference for agentic workloads stresses the GPU key-value (KV) cache in a sustained and cumulative manner, often causing severe throughput degradation well before memory cap…
Semantic-Aware Scheduling for GPU Clusters with Large Language Models
Zerui Wang, Qinghao Hu, Ana Klimovic +4
Deep learning (DL) schedulers are pivotal in optimizing resource allocation in GPU clusters, but operate with a critical limitation: they are largely blind to the semantic context…
Rethinking Key-Value Cache Compression Techniques for Large Language Model Serving
Wei Gao, Xinyu Zhou, Peng Sun +2
Key-Value cache (\texttt{KV} \texttt{cache}) compression has emerged as a promising technique to optimize Large Language Model (LLM) serving. It primarily decreases the memory cons…