Mamba #447

[johnnynunez]

447

casualconv1d: https://github.com/Dao-AILab/causal-conv1d mamba original paper: https://github.com/state-spaces/mamba MambaVision: https://github.com/NVlabs/MambaVision Complete Suite of VideoMamba: https://github.com/OpenGVLab/video-mamba-suite cobra(llava rival): https://github.com/h-zhao1997/cobra VideoMambaSuite:

[johnnynunez]

@dusty-nv feel free to organize the folders inside jetson-containers. mamba depends from causalconv1d. all repositories that use mamba depends of mamba(causalconv1d too) like videosuite or mambavision all is build with cuda(this folders need a benchmark) because the first when I use mamba, it used casualconv1d and mamba in cpu and it is very slow.

Now we can use LLM models that use mamba :) as you can see on my test.py also works with images or video. I hope that communit comes with demos :)

[johnnynunez]

Benchmark:

2 scenario:

[dusty-nv]

Thanks @johnnynunez, looks like another fun one! 🤣 🐍

[johnnynunez]

Thanks @johnnynunez, looks like another fun one! 🤣 🐍

Please see the spam of my messages JAJAJ. I think I need your help because this is going too fast.

[dusty-nv]

Haha that does seem fast! Is it producing valid/coherent output? That would be great 👍

[johnnynunez]

Haha that does seem fast! Is it producing valid/coherent output? That would be great 👍

it is the original benchmark from mamba: https://github.com/state-spaces/mamba/blob/main/benchmarks/benchmark_generation_mamba_simple.py

[johnnynunez]

@dusty-nv if you can upload casual-conv1d wheel and mamba-ssm wheel, I will thank you

[dusty-nv]

@johnnynunez I will start trying to build these tonight/tomorrow and push the wheels & containers 👍