Канал: MLSys Singapore

E15 | MuxServe: Flexible Multiplexing for Efficient Multiple LLM Serving (ICML'24) 【中文】

E15 | MuxServe: Flexible Multiplexing for Efficient Multiple LLM Serving (ICML'24) 【中文】

E14 | Zero Bubble (Almost) Pipeline Parallelism (ICLR’24) 【中文】

E14 | Zero Bubble (Almost) Pipeline Parallelism (ICLR’24) 【中文】

E13 | Centauri: Efficient Scheduling for Comm.-Comp. Overlap in Training (ASPLOS'24 Best Paper) 【中文】

E13 | Centauri: Efficient Scheduling for Comm.-Comp. Overlap in Training (ASPLOS'24 Best Paper) 【中文】

E12 | DeltaZip: Multi-Tenant Language Model Serving via Delta Compression

E12 | DeltaZip: Multi-Tenant Language Model Serving via Delta Compression

E11 | AMSP: Super-Scaling LLM Training via Advanced Model States Partitioning

E11 | AMSP: Super-Scaling LLM Training via Advanced Model States Partitioning

E10 | 1 year after ChatGPT, where do the best open source LLMs stand?

E10 | 1 year after ChatGPT, where do the best open source LLMs stand?

E09 | Automatic Root Cause Analysis via Large Language Models for Cloud Incidents

E09 | Automatic Root Cause Analysis via Large Language Models for Cloud Incidents

E08 | EasyScale: Elastic Training with Consistent Accuracy and Improved Utilization on GPUs

E08 | EasyScale: Elastic Training with Consistent Accuracy and Improved Utilization on GPUs

E07 | Fast LLM Serving with vLLM and PagedAttention

E07 | Fast LLM Serving with vLLM and PagedAttention

E06 | bbTopk: Bandwidth-Aware Sparse Allreduce with Blocked Sparsification

E06 | bbTopk: Bandwidth-Aware Sparse Allreduce with Blocked Sparsification

E05 | Hanayo: Harnessing Wave-Like Pipeline Parallelism for Enhanced Large Model Training Efficiency

E05 | Hanayo: Harnessing Wave-Like Pipeline Parallelism for Enhanced Large Model Training Efficiency

E04 | FedTree: A Federated Learning System For Trees

E04 | FedTree: A Federated Learning System For Trees

E03 | Optimizing Dynamic Neural Networks with Brainstorm

E03 | Optimizing Dynamic Neural Networks with Brainstorm

E02 | Long Sequence Training from System Perspective

E02 | Long Sequence Training from System Perspective

E01 | Lucid: A Non-intrusive, Scalable and Interpretable Scheduler for Deep Learning Training Jobs

E01 | Lucid: A Non-intrusive, Scalable and Interpretable Scheduler for Deep Learning Training Jobs