-
qwen3next: fix issue in delta net (#14075)
frostbyte_neo 发布于
2026-02-04 21:40:38 +00:00 gDiffExp was being broadcast across the wrong axis when multiplying with k. This fix reshapes gDiffExp to [1, chunkSize, nChunks, ...]
下载附件
gDiffExp was being broadcast across the wrong axis when multiplying with k. This fix reshapes gDiffExp to [1, chunkSize, nChunks, ...]