trunk/622b0e329d68c3b87d4da922106c7a0ff6500633: [inductor] Fuse interleaved epilogues in nested reductions (#190595)
- PyTorch 近 90 天出现 10 次
- PyTorch 近 90 天第 10 次发版
- 上一次：同一天稍早 · trunk/52d08b9519e82b71ababd98473077dff9007f875: [PyTorch][jit] Avoid IValue copies in unpickler (#193250) (#193250)
发生了什么
Human context: We had previous nested reductions, as well as nvfp4 sub-iteration domain fusing. this combines the two. Extend fused nested reductions with a sub-parent domain for interleaved pair consumers. This keeps RMSNorm plus NVFP4/MXFP4 packing in one kernel while validating source layouts, non-leaf consumers, and ordinary scheduler fusion gates. Represent the sub-parent domain in grouped coordinates (outer, g…
摘要按规则整理自下方来源原文