PyTorch updates graph-break baselines for HuggingFace benchmarks
BartForCausalLM and MBartForCausalLM hit new graph breaks in latest commit.
PyTorch's development team committed an update to the dynamic inductor HuggingFace graph-break baselines after a periodic benchmark failure. The commit (trunk/7484739, PR #188743) adjusts expected graph-break counts for several HuggingFace models evaluated under PyTorch's dynamic inductor optimization. This ensures that future CI runs compare against realistic baselines rather than outdated expectations.
Specifically, BartForCausalLM and MBartForCausalLM each saw a graph break increase from 6 to 7, indicating a minor regression in the inductor's ability to fuse operations. On the positive side, M2M100ForConditionalGeneration improved from 11 to 4 graph breaks, and OPTForCausalLM improved from 8 to 7. These counts were captured from the failing periodic run at PyTorch commit 909e76cadf21. The change was approved by William Wen.
- BartForCausalLM graph breaks: 6 → 7 (regression)
- MBartForCausalLM graph breaks: 6 → 7 (regression)
- M2M100ForConditionalGeneration graph breaks: 11 → 4 (improvement)
Why It Matters
Maintaining accurate graph-break baselines ensures reliable CI for PyTorch's dynamic inductor, critical for performance optimization.