Can SM90 DSA indexer_forward support qhead_per_kv_head=16

[xiaoxi-wangfj]

Hi, I have a question about the DSA cuDNN frontend implementation.

In cudnn-frontend/python/cudnn/deepseek_sparse_attention/indexer_forward/_interface_sm90.py
the SM90 indexer_fwd path currently defines:

_SUPPORTED_QHPKV = (32, 64)
So qhead_per_kv_head=16 is rejected.

I’d like to ask whether the underlying operator implementation can support qhead_per_kv_head=16.

[Anerudhan]

Thanks for the issue. @NVIDIA-JerryChen will be looking at this.