Coding Agents for Compiler Construction: Beyond the...
summary
An agentic compiler prototype is implemented as a PyTorch add-on that ingests nn.Module or FX graphs, runs an LLM-based code generator plus an eval tool that compiles, profiles and benchmarks kernels, and iterates a search loop across up to ten models to propose candidates. A Mumble subgraph is transformed by replacing two matmuls with one matmul+reduction yielding 6.2x faster for that subgraph and 3x faster for the whole model versus TorchCompile. Supports CUDA/PTX, HIP and target-specific code; correctness uses tolerance 1e-3; ResNet-18 end-to-end compile is ResNet-18: 30–60 minutes with a fingerprinted caching database.
AI-written summary · about
chapters (10)
- 0:00Coding agents for compilers
- 1:06LLMs to agentic compilers
- 2:32Autonomous agent for kernel optimization
- 3:54Autonomous agents replace backends
- 5:02Mumble layer optimization results
- 6:25Correctness and input handling
- 8:14Multi-model optimization benefits
- 9:16Evaluation loop and model diversity
- 9:59Compilation time and caching
- 10:41Optimization database reuse
related talks
23mBlazing Fast GenAI Inference With Torch.compile
36mDL Compiler Panel Discussion
11mQiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a...
21mThe New Vision of Compiler-Accelerated
19mMAGPY: Compiling Eager Mode DNN Programs by Monitoring Execution States
15mKeynote: Pre-training and Fine-tuning of Code Generation Models
15mGraCE: Unlocking CUDA Graphs with Compiler Support for ML Workloads
45mRunning Your Coding Agent Locally: Lessons from a Real-World Experiment