`apply_per_channel_scale` may fail to launch when the input sequence length is extremely large

[StudyingShao]

System Info

NA

Who can help?

@StudyingShao

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

NA

Expected behavior

INT4 AWQ can work correctly with an extremely large input sequence length.

actual behavior

The launch of the kernel apply_per_channel_scale fails, when the CUDA grid.y of the launching configurations exceeds 65535.

additional notes

NA

[StudyingShao]

Fixed in PR #4089. Thanks @crazy-JiangDongHua for reporting and locating this bug.

[brb-nv]

Fix merged. Marking this closed.