Lightning Talk: PyTorch 2.0 on the ROCm Platform
summary
PyTorch on ROCm gains day 0 Triton support for 2.0 with no API changes required, plus hipGraph capture, hipSolver, and DOT graph dumps; Triton adds MF (matrix fuse) instructions and register-heavy optimizations. Benchmarks across 44 Hugging Face models on a single MI250 show some models hitting up to 4x speedup and inference with AMP averaging about 2x improvement. ROCm 5.7 extends support to 700-series GPUs (700 XTX, W700) with 24–40 GB memory.
AI-written summary · about
chapters (9)
- 0:00PyTorch on ROCm intro
- 1:45ROCm and PyTorch overview
- 3:32PyTorch 2.0 enhancements
- 4:13Triton enablement
- 6:23Triton performance
- 7:05AMD GPU support
- 9:38Unified ROCm solution
- 10:22Live Stable Diffusion demo
- 11:05Q&A
related talks
05mBuild AI Anywhere with ROCm™ Software: AMD and PyTorch Bring...
25mRace to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen...
20mBringing PyTorch Monarch to AMD GPUs: Single-Controller Distributed Tra...
32mPowering Your Generative AI Workloads with AMD and Open-Source ROCm
34mBridging the Gap: Compiling and Optimizing Triton Kernels Onto RISC-V Targets Based on MLIR
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
25mMoving Inference to Triton Servers
15mDeploy Hugging Face Models on Azure powered by AMD ROCm software