FlashAttention

Road to 5 Million Tokens: Breaking Barriers in Long Context Training — Max Ryabinin, Together AI

Jun 8, 2026 · 15:50

Max Ryabinin from Together AI presents their research on extending transformer context length to 5 million tokens using Untied Ulysses, which cuts activation memory by reusing buffers across attention head iterations. The talk walks through a stack of techniques including fully sharded data parallelism, DeepSpeed Ulysses context parallelism for an 8x activation reduction, activation checkpointing for another 8x, CPU offloading of transformer block inputs, and chunked sequence training. Even with these, training a LLaMA 3B model with 3 million tokens fits on an 8xH100 node, but 5 million requires Untied Ulysses. Instead of allocating one large buffer per attention head group, it chunks heads further and reuses buffers across iterations, cutting activation memory with negligible throughput impact. At both 8B and 32B scale, results match the most memory-optimized transformer training baselines while pushing sequence length 25% further than prior Ulysses implementations.

Your Coding Agent Should Do AI System Engineering — Ben Burtenshaw, Hugging Face

May 21, 2026 · 18:25

Ben Burtenshaw from Hugging Face demonstrates how coding agents can take on AI systems engineering tasks—writing CUDA kernels, fine-tuning models, and running multi-agent research labs—by leveraging skills and the Hugging Face Hub. He shows a 1.88x speedup on H100s with an RMSNorm kernel written by Claude Code, and a fine-tuned Qwen3 0.6B achieving 35% on LiveCodeBench. Skills compress years of specialization into hours by turning zero-shot tasks into few-shot workflows. For multi-agent research, a Planner generates hypotheses from papers, Workers implement them as training scripts, and a Reporter monitors results via the open-source Trackio dashboard, with all jobs running on Hub compute. The key is exposing open primitives like kernels, Trackio, and HF jobs as agent-controllable tools.

How Transformers Finally Ate Vision – Isaac Robinson, Roboflow

May 8, 2026 · 17:05

Isaac Robinson, research lead at Roboflow, explains why transformers ultimately beat convolutional neural networks for vision, arguing that massive ViT-specific pretraining and borrowed infrastructure from LLMs overcame the transformer's lack of inductive bias. He traces the evolution from ViT and Swin (windowed attention reducing complexity to n²) through ConvNeXt (reintroducing convolution with transformer-style blocks) to Hiera (stripping biases and recovering them via MAE pretraining), and back to the simple, scalable ViT. Robinson highlights that pretraining methods like MAE and DINOv3 learn the inductive biases CNNs have built-in, while tools like FlashAttention from the LLM world nullify ViT's n⁴ compute scaling penalty. In practice, this pattern appears in the SAM model series: SAM used a ViT backbone, SAM2 switched to Hiera with MAE, and SAM3 returned to the simple ViT. However, these massive models lack deployment flexibility; Roboflow's RF-DETR uses neural architecture search on a foundation model backbone to generate a family of high-performance models, achieving up to 40× speedup over fine-tuning SAM3 while outperforming real-time convolutional detectors.

The Small Model Infrastructure Nobody Built (So We Did) — Filip Makraduli, Superlinked

May 5, 2026 · 18:30

Filip Makraduli of Superlinked introduces SAI, an open-source inference engine for small models that addresses gaps in embedding infrastructure by enabling dynamic model loading, hot-swapping, and memory-aware eviction on a single GPU. He argues that provisioning separate GPUs for each small model wastes idle capacity, and that the real challenge lies in supporting diverse model architectures (e.g., BERT, Qwen, Colbert) with different attention mechanisms and positional embeddings. The engine re-implements forward passes with variable-length FlashAttention and handles model swapping via a least recently used eviction policy. Makraduli also explains that context management for agents requires small models to pre-process data, referencing Andrej Karpathy’s graph-based knowledge bases and Chroma’s own model. The talk details the infrastructure layer including routing, auto-scaling with Prometheus, and GPU provisioning using spot instances, all open-sourced as SAI (Superlinked Inference Engine) with Helm charts and Docker images.

$1 AI Guardrails: The Unreasonable Effectiveness of Finetuned ModernBERTs – Diego Carpentero

Apr 16, 2026 · 43:53

Diego Carpentero argues that LLM-based attacks—Prompt Injection, Indirect Injection, Model Internals (gibberish suffix), RAG Poisoning, MCP Exploits, and Agentic Escalation—are now the baseline, not the exception, and that model alignment and human review alone are insufficient. He identifies the core problem as a Zero Trust Gap: LLMs natively lack separation between system controls and data, allowing adversaries to override decisions via malicious instructions in inputs or external content. To build a protective layer, Carpentero fine-tunes ModernBERT—a state-of-the-art encoder with Alternating Attention, Unpadding & Sequence Packing, RoPE, and FlashAttention—into a safety discriminator that classifies prompts as safe or unsafe in ~35 milliseconds with 85% accuracy, all for under a dollar. He walks through the fine-tuning pipeline using the IngetGuard dataset and demonstrates live detection of real attack examples from each vector.

AI Kernel Generation: What's working, what's not, what's next – Natalie Serrino, Gimlet Labs

Dec 17, 2025 · 19:15

Natalie Serrino, cofounder of Gimlet Labs, presents how AI-generated kernels can automatically speed up custom PyTorch code by up to 24% on Apple M4 hardware using the Metal framework, with a 40% speedup from kernel fusion. The agentic system iterates through compilation, execution, correctness, and optimization, but faces challenges like validation of floating-point results and reliable benchmarking. Successes include rewriting average pool 1D as a convolution for 80% improvement, while failures occur on heavily optimized ops like matrix multiply. A real-world audio encoder model saw 70% faster inference on RTX 6000 Blackwell via six custom fused kernels. Serrino emphasizes that AI is best for rapidly searching optimizations and porting code to new hardware, not for surpassing human experts on novel algorithms.

Mastering LLM Inference Optimization From Theory to Cost Effective Deployment: Mark Moyou

Jan 1, 2025 · 33:39

NVIDIA solutions architect Mark Moyou explains that LLM inference differs fundamentally from standard deep learning deployment, requiring careful management of KV Cache, attention mechanisms, and GPU memory to control cost. He details how tokens are processed: prefill computes attention across the entire prompt, then generation produces one token at a time, with KV Cache storing key-value pairs to avoid recomputation. Llama's 32 attention heads and FP8 quantization (halving memory with near-identical accuracy) are cited as key optimizations. Moyou emphasizes measuring time to first token, inter-token latency, and input/output sequence length distributions to size inference engines. He presents NVIDIA's TRT-LLM (model compilation for LLMs) and Triton inference server as tools to maximize throughput, and discusses how query patterns like long-input-short-output or short-input-long-output impact GPU utilization and deployment cost.

Low Level Technicals of LLMs: Daniel Han

Jul 31, 2024 · 2:52:26

Daniel Han of Unsloth explains how to find and fix bugs in open-source LLMs like Gemma, Phi-3, and Llama, covering tokenizer issues, architecture pitfalls, and finetuning optimizations. He details the eight Gemma bugs Unsloth fixed, including a critical RoPE downcasting error that broke positional encoding, and a 2048 sliding window bug in Phi-3. Han walks through transformer internals: attention masking, layer norms, RoPE embeddings, and SwiGLU activation, showing how to derive gradients for custom kernels. He demonstrates Unsloth's 2x faster finetuning with 70% less memory via Triton kernels, and introduces new features: automatic Ollama model file creation, CSV fine-tuning with merged columns, and chunked cross-entropy for large vocabularies. The session includes live Q&A on learning rate schedules, precision trade-offs, and mechanistic interpretability.