TP-invariant Training: bitwise-identical training across TP degrees and GPU architecture#2977
Draft
jinzex wants to merge 4 commits into
Draft
TP-invariant Training: bitwise-identical training across TP degrees and GPU architecture#2977jinzex wants to merge 4 commits into
jinzex wants to merge 4 commits into
Commits
Commits on May 13, 2026
- andcommitted
- andcommitted
- committed
- andcommitted