viable/strict/1790923539: Update RNN model graph break baselines missed by #186614 (#199327)
- PyTorch 近 90 天出现 1793 次
- PyTorch 近 90 天第 1771 次发版
- 上一次:同一天稍早 · trunk/3660a13c2c0919ccf08046efb9f0c7388f1b8f21
发生了什么
The periodic inductor torchbench jobs have been failing since #186614 ("Delay RNN graph break until module call", d5424d7 ) with: demucs FAIL: graph_breaks=4, expected=3 doctr_reco_predictor FAIL: graph_breaks=4, expected=1 tts_angular FAIL: graph_breaks=5, expected=2 That PR changed Dynamo to graph break at the call to an nn.RNN/GRU/LSTM module instead of when loading the attribute, so tracing resumes after the rec…
摘要按规则整理自下方来源原文