Feat/gguf yaml parser

[nguyenhoangthuan99]

Fix #1244

This is result of tinyllama.yaml file after downloaded successfully from cortexso.

[dan-menlo]

@nguyenhoangthuan99 Can I just check:

• I think from our earlier discussion, model.yaml is optional (we can load params directly from gguf)
• However, is this done by us auto-creating a model.yaml for every GGUF file (e.g. upon pull or run)

The reason why I'm asking is:

• Do we need a model method to create a model.yaml by parsing from a GGUF file?

[nguyenhoangthuan99]

I think we still need it now because:

• There is 1 field chat_template we still need to parse from gguf file and render chat template. The implementation of cortex.llamacpp requires rendered chat template.
• Stop tokens also needed to parse from gguf because cortex.llamacpp requires that field when run chat completion.
• For further feature like model-compatibility API, we need the llama.embedding_length, llama.attention.head_count, llama.attention.head_count_kv to calculate the memory constraint for kv cache and recommend suitable ctx_len for user
• ngl <> llama.block_count is also can be read inside gguf file and can be used to recommend the suitable ngl to user's hardware

[nguyenhoangthuan99]

I'll add uni-tests for this PR

[dan-menlo]

Got it. So we will still create a model.yaml whenever we pull, run or import a model?

[nguyenhoangthuan99]

yes, we need to create model.yaml with the following cases:

• pull model from other sources
• import a random model from local

Models from cortexso already have model.yml so we don't need to create model.yml, just update the file location of binary inside the model.yml

[dan-menlo]

Approving to unblock. Will let @vansangpfiev review for code quality.