AI / ML

MaSRead solves the read-to-content problem in systems where independent agents share computed state as key-value cache fragments. It routes queries through opaque keyed tag sets and uses hard attention masks to isolate fragments, preventing interference from colocated data. By leveraging lexical connectivity for graph walks, it reliably retrieves multi-hop fragments from conflict-free replicated stores.

• Enables reliable reading of merged latent fragments without text serialization.
• Uses hard attention masks to isolate specific fragments during query decoding.
• Supports multi-hop retrieval via graph walks over lexical connectivity.
• Resolves interference issues inherent in content-addressable latent stores.
• Operates effectively across various network topologies like chains and hubs.

Researchers address the lack of workload-aware power management in AI datacenters by training a PPO meta-controller on GRPO post-training traces. Using half-second telemetry from 7B to 72B model scales on A100s, the controller dynamically adjusts generation parameters to match measured power. Live deployment at 72B scale demonstrates significant reductions in power-limit violations alongside improved token output and energy efficiency.

• Dynamic RL control adapts generation parameters to real-time power telemetry, replacing static caps.
• 89.8% reduction in power-limit violations on a full 500-step 7B training trace.
• 18.1% increase in token output and 26.2% improvement in energy efficiency (tokens/MWh).
• Validated live at 72B scale, proving scalability beyond single-GPU or small cluster tests.
• Highlights gap in current datacenter management which treats GPU power as workload-blind.

This paper introduces VQ-bench, a unified framework designed to standardize the development and benchmarking of vector quantization algorithms. The authors define seven core conceptual primitives that allow for the arbitrary composition of quantization logic. By re-expressing 25 existing quantizers as pipelines of these primitives, the framework enables reproducible benchmarks and simplifies the evaluation of new methods.

• Seven primitives provide a modular basis for composing complex quantization pipelines.
• 25 common quantizers are re-implemented as pipelines for standardized benchmarking.
• Open-source release ensures reproducible evaluation of new quantization algorithms.
• Addresses the surge in AI infrastructure engineering by unifying VQ research tools.
• Simplifies comparison of new quantization methods against established baselines.

Lightricks released LTX-2, a diffusion transformer model combining synchronized audio and video generation in a single architecture. The official package provides Python inference and LoRA training capabilities, supporting multiple performance modes and API access. It features a VAE that uses Natten for acceleration on Linux/CUDA, falling back to Triton or eager modes elsewhere.

• Unified DiT model handles synchronized audio and video generation without separate components.
• Official package includes LoRA trainer for fine-tuning and Python inference scripts.
• Natten backend accelerates video VAE decoding on Linux/CUDA; auto-fallback ensures cross-platform compatibility.
• API access and production-ready outputs are built-in for easier integration into pipelines.
• Open access allows researchers and engineers to experiment with end-to-end audio-video synthesis.