Help me understand the purpose of Tensor Names and Mappings

[nisparks]

Been bouncing between files and getting a little dizzy. I see there is a lot of pattern matching, labels, architectures defined with sets of tensor types, etc.

At the end of all of the mappings, is the end result just to figure out the metadata for the gguf?

(I'm trying to map a new architecture now, but in another draft PR for vision, I see a lot of new tensor names specifically for Vision models, and I'm trying to understand the significance, and what it means to add a new tensor type.)

[nisparks]

Rubber ducky helped with this one. Figured it out. The tensors map to a struct: llm_tensor_info. That struct contains the layer type (Input, Repeating, Projection, Output) and the type of Operation (too many to list).

Answered my own question.