[QST] how can i do w4a8 (int4 * int8) using cutlass?

[yyfcc17]

i see int4 * fp8 is supported on hopper gpus,

but for gpus like a30 and a100, there is no fp8 support, so i need to use int4 * int8, two int4 is packed into a int8,

how can i accomplish this kind of mixed gemm using cutlass?

thanks!

[hwu36]

#1190

@alexsamardzic

[alexsamardzic]

My PR is about F16/ S4, and will cover also BF16/S4, F16/ U4 and BF16/ U4 MM on Ampere; through an earlier work of @manishucsd, CUTLASS already supports these for S8 and U8 instead of S4 and U4, respectively, on Ampere. And that's about it regarding mixed data-types MM, thus what OP would need is not in the works yet, but may be an interesting follow-up.

[yyfcc17]

thanks for reply.

since LLM w8a8 is almost solved by PTQ methods, and w4a4 is too aggressive for PTQ methods, w4a8 PTQ is a good option for further optimization.

in my own experiments, the LLM's performance degradation is minor under w4a8 (per-channel * per-token) PTQ settings. and there is also a paper about it: https://arxiv.org/abs/2311.09550, the performance seems promising.

trt-llm already supported w4a8 PTQ & inference on hopper, but only for fp8.

is it possible to extend your PR to support w4a8 (int4 * int8) on Ampere? @alexsamardzic

[alexsamardzic]

It would be better to open another PR. Namely, the combination of data-types that my PR intends to handle is rather tricky regarding loading data from shared memory to registers, that I believe should not be the case for S4/S8 MM. Also, there isn't much to share between the two - all of data conversion, reshuffling S4 data between threads, and finally the tests, should be different for S4/S8.

[alexsamardzic]

I'm now working on this feature - #1413 is created, so this issue could be closed.

[yyfcc17]

@alexsamardzic that's great, thank you!