PyTorch Conference 2024
2024-09-18..2024-09-19
San Francisco, United States
[PyTorch Conference 2022, PyTorch Conference 2023, PyTorch Conference 2024, PyTorch Conference 2025]
as the organisers put it
sponsor register view the schedule Submit a Proposal See Who’s Attending Thank you to all who joined us for PyTorch Conference 2024! We look forward to seeing you next year, October 22 – 23…
who it's for
compiler engineers, kernel engineers, inference engineers, machine learning engineers, platform engineers, framework maintainers, research engineers
topics
Halide backend integration; MLIR lowering for custom kernels; kernel fusion and tiling; Triton mid-level kernel language; Apache TVM MM compiler; Mojo kernels library; torch.compile with Dynamo tracer; vertical fusion and epilogue merging; adaptive kernel generation; WebAssembly and WebGPU deployment
| title | author | length |
|---|---|---|
| A Halide Backend for TorchInductor | Jason Ansel | 08m |
A Halide backend is added to TorchInductor as a reference backend that targets extensibility for kernel‑level DSLs and hardware vendors, integrated with… | ||
| Enabling Composition of Kernels and Compilers | Jacques Pienaar | 11m |
Shows concrete MLIR work inside IREE and OpenXLA and backs claims with numbers: a found library kernel is 4× faster locally yet integration causes 29%… | ||
| Maximizing Kernel Development Productivity Under Performance Constraints | Philip Tillet | 10m |
Triton is presented as a mid‑level kernel language that trades some CUDA flexibility for much faster iteration and easier maintenance, supporting pointers in… | ||
| Universally Deploy Large-language Models via ML Compilation | Tianqi Chen | 11m |
Claims a universal LLM deployment path by building an ML compiler stack on Apache TVM called MM compiler that adds Python-first pipelines, first-class symbolic… | ||
| Lifting PT to New Heights with MAX and Mojo | Mikhail Zolotukhin | 10m |
Modular builds the MAX inference engine plus a new language, Mojo, to ship a Mojo kernels library with three abstraction layers (assembly-like, CUDA-like… | ||
| Together Goes Brrr: Threading Research & Production with Torch Compile | Pragaash Ponnusamy | 09m |
Deployment flow at Together AI uses torch.compile with the Dynamo tracer and Inductor backend to turn Python model code into profiled, fused kernels; automatic… | ||
| DL Compiler Panel Discussion | P. Tillet, J. Ansel, J. Pienaar, T. Chen, M. Zolotukhin, P. Wu | 36m |
PyTorch 2’s torch.compile() drives concrete changes: teams report rebuilding inference stacks around torch.compile() and integrating Triton, TorchInductor… | ||
| Welcome to the PyTorch Ecosystem for LLM Fine-tuning Mini Summit | Kartikay Khandelwal | 01m |
First PyTorch fine tuning mini summit presents a packed agenda with a keynote on the Llama ecosystem, three invited demos from the founders of Axolotl… | ||
| The State of the Llama Ecosystem | Joe Spisak | 17m |
Llama sees 350M+ downloads and a jump in usage and releases culminating in the 405B model with 128k context and a changed license that explicitly allows… | ||
| The Challenges of Building an Opinionated Open Source LLM Framework | Wing Lian | 09m |
An opinionated Hugging Face trainer wrapper that targets the majority of fine‑tuning use cases and drops old PyTorch, sequence‑to‑sequence models and rarely… | ||
| Hacks to Make LLM Training Faster | Daniel Han | 08m |
Six practical wins: switching from float32 to float16/BF16 yields roughly 5x speedups on common GPUs; quantizing to int8 keeps accuracy while int4 halves model… | ||
| torchtune: Easy and Accessible Finetuning in Native PyTorch | Evan Smothers | 16m |
Shows a working finetuning pipeline for Llama 3 AP on a single A100 that reduces peak GPU memory from >80 GB (naive bf16) to ~58 GB with 8-bit AdamW, to ~48 GB… | ||
| Panel Discussion | T. Dettmers, H. Schoelkopf, A. Chowdhery, A. Conneau, K. Khandelwal | 36m |
Quantized training and inference win real gains here: BitsAndBytes delivers 8‑bit optimizers and QLoRA proves low‑cost fine‑tuning, while an open LM Eval… | ||
| nothing matches | ||