(should be labelled as a feature request for discussion)
It would be useful to provide a mechanism to modify already-built MLGraph objects to support adapters. This could be done in a library but that would require rebuilding and recompiling a (potentially) very large model. Adapter support would allow for small incremental changes to existing, potentially even platform built-in, models, and would minimize the need to download entire pre-adapted models.
Full proposal here: https://github.com/webmachinelearning/hybrid-ai/blob/main/proposals/adapters.md
Some mechanism is useful to "fetch" a base model on the client, so there is some overlap and utility to also having a caching API, see related discussion in Issue webmachinelearning/webnn#807
(should be labelled as a feature request for discussion)
It would be useful to provide a mechanism to modify already-built MLGraph objects to support adapters. This could be done in a library but that would require rebuilding and recompiling a (potentially) very large model. Adapter support would allow for small incremental changes to existing, potentially even platform built-in, models, and would minimize the need to download entire pre-adapted models.
Full proposal here: https://github.com/webmachinelearning/hybrid-ai/blob/main/proposals/adapters.md
Some mechanism is useful to "fetch" a base model on the client, so there is some overlap and utility to also having a caching API, see related discussion in Issue webmachinelearning/webnn#807