Triton

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.

The Rise of Open Models in the Enterprise — Amir Haghighat, Baseten

Jul 24, 2025 · 16:50

Amir Haghighat, CTO of Baseten, argues that enterprises are increasingly moving from closed frontier models like OpenAI and Anthropic toward open source models, driven by four specific cracks in the assumption that closed models will work indefinitely: quality for specialized tasks (e.g., medical document extraction), latency requirements (especially for voice), unit economics ballooning from agentic use cases where a single user action triggers 50 inference calls, and the desire for competitive differentiation. Drawing on conversations with over 100 enterprises, he explains that while most started with dedicated deployments on Azure/AWS for toying around in 2023, by 2024 about 40-50 had production use cases, and in 2025 the shift accelerated. However, adopting open models forces enterprises to build inference infrastructure, facing challenges like speculative decoding, prefix caching, guaranteeing four-nines reliability with hardware failures and VLM crashes, and scaling replicas—with one Fortune 50 soft drink company reporting an eight-minute spin-up time. Haghighat concludes by contrasting the simple API-call world with the complexities of mission-critical inference, where…

Dream Machine: Scaling to 1m users in 4 days — Keegan McCallum, Luma AI

Jul 19, 2025 · 19:03

Keegan McCallum, Head of ML Infrastructure at Luma AI, details how the company's Dream Machine model scaled from 500 to 9,000 H100 GPUs within hours to handle 1 million users in four days, outpacing ChatGPT's initial growth. He explains that their initial Triton inference server setup was brittle and ill-suited for multi-GPU, multi-node video models, prompting a re-architecture to a custom serving stack on vanilla PyTorch. To solve work starvation across user tiers, they implemented an SLO-based aging system that ranks jobs by the percentage of their worst-case wait time elapsed. For managing dozens of model versions, they store immutable full Python environments and checkpoints in object storage, with a YAML file controlling active deployments and enabling zero-downtime rollouts across thousands of GPUs. McCallum also discusses partnerships with Nvidia, AMD, and Grok, and how Luma's broader mission is to build general multimodal intelligence that generates, understands, and operates in the physical world.

The End of Awkward AI Transcriptions - Travis Bartley and Myungjong Kim

Jun 3, 2025 · 16:24

NVIDIA speech AI researchers Travis Bartley, Myungjong Kim, and Jae-hun detail how their Fast Conformer backbone powers both streaming (Parakeet) and multitask (Canary) models, achieving top-5 Hugging Face ASR leaderboard rankings by prioritizing customization over one-size-fits-all. They explain CTC/TDT for low-latency streaming and attention encoder-decoder for high-accuracy multitask models, with Subformer enabling unified speaker diarization and target-speaker ASR. Training uses open-source and proprietary data with pseudo-labeling via the Nemo toolkit. Deployment on NVIDIA Riva NIM with TensorRT optimization supports low-latency streaming and offline processing. Customization includes fine-tuning acoustic models, language models, punctuation, and inverse text normalization for domain-specific terms like medical and food ordering.

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.

From model weights to API endpoint with TensorRT LLM: Philip Kiely and Pankaj Gupta

Sep 13, 2024 · 1:40:01

Philip Kiely and Pankaj Gupta of Baseten lead a workshop on TensorRT-LLM, NVIDIA's high-performance inference framework for LLMs, arguing its use delivers best-in-class throughput and latency on NVIDIA GPUs. They explain that TensorRT-LLM optimizes computational graphs via plugin kernels and in-flight batching, achieving 216 tokens per second and 180ms time to first token on Mistral 7B. The workshop demonstrates building an engine for TinyLlama 1.1B, including FP8 quantization that reduced engine size from 2 GB to 1.2 GB with minimal quality loss. They benchmark a deployed model, showing 7,000 total tokens per second at batch size 64 on an A10G. The presenters compare TensorRT-LLM favorably to VLLM for high-throughput production use and introduce Truss, Baseten's open-source packaging tool, alongside their managed platform for automatic scaling and fast cold starts.

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.

Unlocking Developer Productivity across CPU and GPU with MAX: Chris Lattner

Jul 25, 2024 · 18:33

Chris Lattner, CEO of Modular, presents MAX, a unified AI framework that accelerates Gen AI inference by combining CPU and GPU programming into a single Pythonic model, and Mojo, a new programming language that extends Python to systems programming with 100–1000x speedups. Lattner argues that current fragmentation across PyTorch, ONNX, TensorRT, and hardware-specific libraries slows innovation, and MAX replaces the entire stack—including cuDNN and Intel MKL—with a consistent, compiler-driven approach. He demonstrates that MAX's Int4/Int6 quantization achieves 5x faster performance than llama.cpp on cloud CPUs, and that its GPU matrix multiplication beats NVIDIA's cuBLAS by up to 30%. Mojo enables developers to write for loops and tokenizers (e.g., for LLaMA 3) in Python-like syntax without dropping to C++ or Rust. MAX is free and available now for CPU inference; GPU support launches in September with early access via Discord.