From 0 To 500+ Models: Building a Robust Open Sour...
Talkat Open Source Summit Japan + AI_dev: Open Source GenAI & ML Summit + Automotive Linux Summit 2025 · 2025-12-08..12-10·
summary
Claim: an MLIR-based compiler stack built on XLA and PJRT now supports 500+ production models and runs nightly tests on >1000 models to catch accuracy and performance regressions. Engineering specifics: modular frontends from PyTorch/ONNX/Paddle/Jax funnel into stable HLO & Shardy for sharding, custom MLIR dialects (TTIR, TTNN, D2M) and passes perform fusion and tuning, and emit-pi plus tools like TTRT and Model Explorer produce readable code for manual fine-tuning.
AI-written summary · about
chapters (11)
- 0:00Tenstor and open philosophy
- 3:57Hardware and scaling
- 7:30Software stack and compiler
- 8:11AI compiler for robustness
- 12:27Compiler design with MLIR/XLA
- 16:40Matrix multiplication for large models
- 21:37Compiler abstraction and passes
- 29:08Full ML compile pipeline
- 34:28Debugging data mismatches
- 35:51Scaling to 500+ models
- 36:57Q&A
related talks
25mState of Open Source AI Coding Models
11mBuilding Frontier Intelligence in the Open
32mTaming the Machine: Basics of ML Models Training and Inference Optimization
33mDealing with the versioning of production-ready models
21mOperationalizing ML Training Infra at Meta Scale
01h 07m📚 Building Portable Large Models on Heterogeneous AI Accelerators Wit...
13mFrom Silicon to Scale: Training at Scale
32mToward Seamless Model Compilation Integration