What is this kernel 'nvjet_tst_112x64_64x9_1x2_h_bz_bias_TNN'? Which cuda api do i need?

Accelerated Computing CUDA CUDA Programming and Performance
,
1

image
image768×352 46.5 KB

And others like these…

2 
root@localhost:/usr/local/cuda/lib64# grep nvjet *
grep: cmake: Is a directory
grep: libcublasLt.so: binary file matches
grep: libcublasLt.so.13: binary file matches
grep: libcublasLt.so.13.0.0.19: binary file matches
grep: libcublasLt_static.a: binary file matches
grep: stubs: Is a directory
root@localhost:/usr/local/cuda/lib64#
3

jet could be jetson

4

And nvjet is really a terrible name, it is hard to know the sm version, the meta of the kernel. Why di you do this?

1 Like
5

I don’t think kernel names are designed for public identification in those libraries. But for people with the source code in front of them.

2 Likes
7

True. But with names like nvjet — which almost sound like ‘NVIDIA Jetson’ — for people building libraries (like PyTorch) from source and upgrading to a newer CUDA toolkit, it can come as a bit of a surprise, especially if they suspect a misconfiguration (for instance, seeing a ‘jetson’ kernel on an H100 GPU, sm90).

8

Hi xsank,

have you figured out what `nvjet_xxx` kernels are? I encount with the same problem in profiling LLMs with SGLang. if you already know what these kernels stand for and how they are named, I’m very happy to discuss with you.

Thanks!

9

Hi @xsank ,

I’ve found a helpful artical tells us what `nvjet_xxx` kernels are named. Here’s the url: How to Read GPU Profiling Logs: A Ground-Up Guide - DEV Community, you’ll find the answer in “Decoding kernel names: the Rosetta Stone".