parallelism

Wave Forcing introduces a wavefront-scheduling training and execution framework for streaming video generation, letting later video chunks begin denoising before earlier ones finish while overlapping communication with computation. On Wan2.1-T2V-1.3B it matches Rolling Forcing quality and reaches 147.6 DiT FPS on 8×H200, delivering 7.07× single-GPU end-to-end throughput.

LayerFS 为 AI 编码代理提供专用文件系统：每个代理挂载一个临时可写工作区并提交到共享的持久历史，实现大规模并行开发、零成本分叉、去重压缩存储以及即时依赖挂载。

Devin AI agents can now spawn and manage their own sub-Devins, allowing a single task to branch into a multi-level tree of agents. This enables parallel execution where large tasks complete in roughly the time of their slowest sub-task rather than the sum of all parts.

An in-depth explainer covering parallelism strategies for LLM inference—data, tensor, pipeline, context, and expert parallelism—explaining how to split inference across GPUs without losing performance to communication overhead. It walks through prefill vs. decode phases, memory and communication bottlenecks, and how to choose a defensible layout given model, hardware, and workload constraints.

The tweet highlights the low latency and scalability of calling Jev from the US West Coast, with 130 ms per request for 6 questions and constant latency under high parallelism, indicating good design and future potential with specialized models.

This article provides a practical guide on running 8x NVIDIA RTX PRO 6000 Blackwell GPUs for AI workloads, emphasizing high-concurrency inference, model fleets, and 70B fine-tuning, while comparing performance to more expensive data center solutions.

The author shares lessons learned from building agent graphs, highlighting the pitfalls of parallelism in cyclic workflows and advocating for sequential dependent checks and mechanisms to challenge feedback for better efficiency.

PlanetScale describes how it performs massively parallel backups for sharded Postgres databases by spinning up per-shard EC2 instances, restoring prior backups from object storage, and replaying WAL, achieving petabyte-scale backup speeds over 50 GB/s.

The article discusses efforts to improve NumPy's performance on free-threaded Python (Python without the Global Interpreter Lock), enabling better parallelism and scalability.

Explains how shaders work using an Excel formula that computes color per cell, illustrating GPU parallelism.

A new runtime for the K and Q programming languages that reimagines execution using SIMD, parallelism, fusion, and compression to better leverage modern hardware.

An article arguing that multi-agent setups are often overused, providing guidelines on when to use single vs multi-agent architectures based on task parallelism and coordination needs.

Recommending Harrison Chase's blog post on multi-agent system design. The key insight is that multi-agent systems split into 'reading' and 'writing' tasks; reading is easy to parallelize while writing leads to conflicts. It also summarizes suitable and unsuitable scenarios and engineering prerequisites.

This article explains how batching and parallel operations improve latency and efficiency in AI computer use systems, highlighting open-source implementations like pi-computer-use and cua-driver that achieved significant performance gains before similar features appeared in Codex.

Rob Pike's 2012 talk explains the distinction between concurrency and parallelism, emphasizing that concurrency is about composition of independently executing tasks, while parallelism is about simultaneous execution. The talk is a classic in computer science education.

The author discusses the architecture for scaling AI memory ingestion from thousands to millions of documents, emphasizing orchestration and parallelism over raw compute power, and using Prefect for workflow management.

The author shares that upgrading from single 24GB GPU to dual 24GB GPUs improved LLM performance not through larger models but through parallelism, using a main orchestrator agent with smaller subagents, increasing overall throughput for coding tasks.

A tweet listing 16 inference optimization techniques for achieving sub-second LLM responses, including KV-caching, speculative decoding, FlashAttention, and various parallelism methods.

A detailed roadmap for transitioning from a single AI agent to a coordinated swarm of agents, covering when to split, how to run parallel subagents without conflicts, and how to maintain sanity at scale using Claude Code primitives.

This paper introduces DigenRL, a disaggregated RL framework for diffusion-based generative LLMs that uses generation-axis pipeline parallelism and trainer-assisted generation to improve throughput by 1.56-2.10x over existing systems.