perf: Fix OOM for llama3_70b SFT H100 FP8 CS by increasing VPP to 10#3107

Open

yaoyu-33 wants to merge 3 commits intomainfrom

yuya/fix-llama3-70b-sft-h100-oom

Commits on Apr 3, 2026

perf: Add recompute_num_layers=1 to llama3_70b SFT H100 FP8 CS config

yaoyu-33
and
claude
committed
fix: Use selective recompute instead of full recompute for CUDA graph compatibility

yaoyu-33
and
claude
committed
perf: Fix OOM for llama3_70b SFT H100 FP8 CS by increasing VPP to 10

yaoyu-33
and
claude
committed