gguf : add Mamba keys and tensors

[compilade]

Mamba support has been added in llama.cpp (ggerganov/llama.cpp#5328).

New keys and tensors have been added, so they should be included in the GGUF spec as recommended in ggerganov/llama.cpp#5328 (comment).

I hope I wrote in the right sections. I'm not sure whether to put Mamba before or after RWKV, so I've defaulted to alphabetical order by placing Mamba before.

References

I wasn't sure about which terms to use, so I looked up how things are called in different places.

• Visual guide to Mamba https://maartengrootendorst.substack.com/p/a-visual-guide-to-mamba-and-state
  • D is a skip connection
  • Alludes to the role of A regarding state compression
• The Annotated S4 https://srush.github.io/annotated-s4/
  • D is a skip connection
• https://jameschen.io/jekyll/update/2024/02/12/mamba.html
  • B and C parametrization by x
  • A, B, C, D, state, input, output, feedthrough, repectively.
• S4 paper https://arxiv.org/abs/2111.00396
  • state matrix A
• H3 paper https://arxiv.org/abs/2212.14052
  • The shift SSM step is analogous to the conv step of Mamba
    • I've used the term "rolling" in llama.cpp, so... "rolling/shift" to help notice it refers to the same thing in this case?
• MambaByte paper https://arxiv.org/abs/2401.13660
  • calls dt a timestep
• Mamba paper https://arxiv.org/abs/2312.00752
  • "input projections" and "output projection" in section 3.4