Add support for hailo-10h

[fodger]

Hi,
this guy has release a support for hailo-10h : hackdefendr/llama.cpp@dd4d2dd.

Do you think it will possible to integrate in official llama.cpp ? It could be great by coupling with raspberry pi 5.

Thx a lot.

[danbev]

I've played around with hailo-10h earlier this year:
Models: https://huggingface.co/collections/danbev/hailo
Branch: https://github.com/danbev/llama.cpp/tree/hailo

I've not looked at the above solution yet, but wanted to add this just in case it might be useful as an alternative solutions to go about this.

[fodger]

Thx for your suggestion :). Tried with llama-3.2-gguf and there was some memory leaks :
`4.38.071.049 E init: the tokens of sequence 0 in the input batch have inconsistent sequence positions:

• the last position stored in the memory module of the context (i.e. the KV cache) for sequence 0 is X = 211
• the tokens for sequence 0 in the input batch have a starting position of Y = 209
  it is required that the sequence positions remain consecutive: Y = X + 1
  4.38.071.054 E decode: failed to initialize batch
  4.38.075.685 E srv decode: Invalid input batch. off = 0, n_batch = 2048, ret = -1
  4.38.075.691 E srv send_error: task id = 146, error: Invalid input batch.
  Error: Invalid input batch.`