[Question] Reason for excluding fusion in deepseek-v2 training

[wavy-jung]

Pai-Megatron-Patch/examples/deepseek_v2/run_mcore_deepseek.sh

Lines 381 to 382 in 3b90f68

	--no-bias-swiglu-fusion \
	--no-rope-fusion \

Hello, thanks for providing great open-source frameworks!
Just a simple trivial question
I wonder why bias swiglu fusion and rope fusion is suppressed by using the options --no-bias-swiglu-fusion and --no-rope-fusion when training deepseek-v2 model. Is there any special reason not to use these fusions?