Latest

The speed of an optimizer depends as much on kernel-level linear algebra as on step count, a factor most training guides ignore. Engineers tuning large-scale training can apply the QR, Cholesky and eigendecomposition implementation patterns to cut data movement on GPUs. The talk ties directly to GPU MODE's competition, giving concrete reference implementations to study.

Official accountPyTorch@PyTorchSource

The session bundles inference infrastructure decisions most teams learn the hard way: when disaggregated prefill/decode and KV-cache-aware routing pay off, and whether self-hosting beats token-based API pricing. Engineers deploying open-source models with vLLM or SGLang can register to get a hands-on walkthrough of Ray Serve LLM and agentic-platform integration. Registration is open now via the Anyscale link.

Anyscale@anyscalecomputeSource

Context cost in long-running agents usually hides inside tool results and sub-agent calls, and dsh-context makes those allocations traceable instead of guessing from a total token count. Builders running multi-agent workflows on DSH can use it to diagnose cache misses, runaway context growth and per-sub-agent spend, then rebalance prompts or trimming before costs climb.

AIGCLINK@aigclinkSource

Agentic pipelines shift the performance bottleneck from single-model inference to the full chain of tool calls, data movement and distributed execution, so framework-level tuning matters more than raw GPU throughput. Engineers building agents on PyTorch, Triton or vLLM can use the co-design lessons to plan their stack for Vera Rubin-class hardware. The keynote is a marker for where Nvidia's agent infrastructure work is heading.

Official accountPyTorch@PyTorchSource

Ray's roadmap toward Rust rewrites and deeper Kubernetes integration signals where distributed training infrastructure is heading, with direct impact on anyone running large-scale RL or post-training pipelines. Teams building RL frameworks or managing GPU clusters can track the NIXL-based GPU-to-GPU transfers and topology-aware scheduling as concrete features to adopt. The NVIDIA Blackwell and TPU pod support gives hardware planners a benchmark for infrastructure decisions.

Official accountPyTorch@PyTorchSource

A classical retrieval baseline beating neural methods changes how developers should think about MCP tool discovery — embeddings and LLM routing may be unnecessary complexity for most setups. Builders implementing agent tool selection can adopt weighted BM25 locally without API costs or latency. The benchmark size of 536 tools offers a concrete reference point for comparing their own retrieval setups.

Zed@zeddotdevSource

The pipeline shows a practical way to turn production failures into labelled training data without manual annotation, with a small tuned model beating a frontier model on a narrow task. Teams running LLM features in production can copy the judge-critic-replay pattern to improve quality while cutting inference cost via prompt compression. The keynote livestream is available to watch online.

Official accountPyTorch@PyTorchSource