Repository navigation
Conversation
ServeurpersoCom
left a comment
There was a problem hiding this comment.
LGTM, dropping the child report entirely is cleaner than validating it, and the video limitation is documented.
|
Maybe I missed something, but child report was required to update the modality and help |
|
@angt my intention was to not parse child-reported note sure if you found any cases where this may break? for example:
in other words, I want to design it such that both router and child produce the same |
|
@ngxson the case I was thinking of is a preset to a model that isn't downloaded yet so the router fallback to |
|
hmm, make sense. in this case, we should have an event called pushing a new commit now |
|
It seems dropping the default initializer of architecture in server_model_meta turns CI red (-Werror=missing-field-initializers on GCC arm64), it should be put back with = {}. |
| common_models_handler_apply(models_handler, params); | ||
| server_download_tracker tracker; | ||
| common_models_handler_apply(models_handler, params, child.is_child() ? &tracker : nullptr); | ||
| if (tracker.mutated) { |
There was a problem hiding this comment.
If 2 presets use the same GGUF, the first one download the file and notify the router, but not the second one? 🤔
There was a problem hiding this comment.
Maybe we can just force a refresh when a child is ready ?
Overview
Some refactor planned from #29987
Note that
common_get_decision_typecurrent can take the gguf path and manually open it to grab the metadata. If we have similar logic in the same future, it will be important to have onegguf_contextthat can be reused by multiple callsRequirements