viable/strict/1788342684: Keep saved-activation strides symbolic under dynamic shapes (#194420)
- PyTorch 近 90 天出现 2586 次
- PyTorch 近 90 天第 2551 次发版
- 上一次:同一天稍早 · trunk/1ee97fb52bf85376f1a9424b8f21b47e304a362f
发生了什么
Problem Inductor can give a saved activation a different layout than the backward graph's placeholder carries (layout optimization, comprehensive padding), so AOTAutograd restrides the placeholder to match what the forward will produce. Inductor reports those strides as sympy expressions, but set_tracing_context_output_strides collapsed them with evaluate_symexpr , substituting the current hints. The resulting const…
摘要按规则整理自下方来源原文