Closed danbev closed 3 months ago
Thanks for the PR! I had been wondering when this new encoding would come up. Can you attend the ML working group on the 18th of this month (details)? It might be helpful for others interested in wasi-nn to understand the motivation here.
@abrown I'd be happy to attend if I can make the time slot :+1:
@danbev, what wasi-nn backend would be used to load and run the ggml
encoded models? I might misunderstand, but I don't think any of the backends currently implemented will be able to load these models. Are you suggesting a new backend implementation?
what wasi-nn backend would be used to load and run the
ggml
encoded models.
WasmEdge has support for a llama.cpp backend. I have got a very basic llama.cpp backend working for Wasmtime and would very much like to see support for such a backend in Wasmtime in the future. Having this encoding would hopefully help existing implementations and creating new ones easier.
Are you suggesting a new backend implementation?
Yes, I'd be interested in seeing support for llama.cpp backend in Wasmtime (and other wasm runtimes).
This commit adds the
ggml
graph encoding format to thegraph_encoding
enum.The motivation for this is to allow the
wasi-nn
interface to support models that are encoded in theggml
format which is the model format used by llama.cpp.