DSA Library
Daily papers
10 picked from 110 new papers
Systems 4
-
TailGuard-GDS: Tail-Safe Control of GPU Storage Data Movement
TailGuard-GDS selects storage transfer plans per application phase, reducing p95 step latency in FSDP and ZeRO-3 workloads.
-
Zero-SM NCCL: A Systematic Guide to Designing and Deploying SM-Free GPU Communication
Zero-SM NCCL frees streaming multiprocessors for compute on large transfers, but loses on small messages and host-bound workloads.
-
Meta Build Systems as Logic Programs
Buildlog derives build configurations through relational rules, exposing ambiguity and provenance; use the approach for complex cross-builds.
-
Query Processing with eBPF-Offload on Computational Storage
The prototype found little query benefit from device-side eBPF because staging and data movement dominate; factor that cost into offload designs.
AI engineering 6
-
Superfluid: Local AI Server for Distributed LLM Inference
Superfluid pools heterogeneous local models and prioritizes interactive work; use it when agents compete with users for inference.
-
CarryOn: Learning What to Keep in Context for Long-Horizon LLM Agents
CarryOn keeps compact state with source-linked evidence and improves task success without extra context-management calls.
-
LuthorSQL: A One-Pass Text-to-SQL Model as System 1 for a Slower LLM Agent
LuthorSQL answers many database questions locally and cuts API cost at the same accuracy; use it as a cascade gate.
-
MEM-SQL: Memory-Empowered Text-to-SQL
MEM-SQL reuses verified query history to raise execution accuracy while reducing tokens and latency on related requests.
-
TIDE: Reclaiming Lost Performance In On-Policy Reinforcement Learning
TIDE reuses idle rollout GPUs and migrates long-running rollouts to improve on-policy training throughput.
-
HerAgent: A Block-Level Checkpointing Harness for Coding-Agent Environment Setup
HerAgent checkpoints setup in semantic blocks, improving success and reducing recovery costs; use it for coding-agent environment bootstrapping.