We saw failure on H20 in SGLang for trtllm_allreduce_fusion. Weird, because it's the same CC as H100/H200, where it works fine.
Seems to be some compilation issue? Unfortunately, I don't have an H20 machine to test this with.
https://github.com/sgl-project/sglang/actions/runs/20109255725/job/57739940169?pr=14764
