General maintenance #47517
Conversation
| """ | ||
|
|
||
| def __init__(self, dim: int = 0, permute_layer_names: list[str] | None = None): | ||
| def __init__(self, dim: int = 0, permute_layer_names: list[str] | None = None, inverse: bool = True): |
There was a problem hiding this comment.
True for BC, i.e. permute back from hf block-layout to rope in complex-style
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
For |
molbap
left a comment
molbap
left a comment
There was a problem hiding this comment.
minor thing, could we add to docs/agentic doc that ALLOWED_ATTN_LAYER_TYPES and MLP counterpart have to be updated if necessary? two lists == two entrypoints to maintain, no big deal but need to be written somewhere
|
@molbap do you mean the model contribution MD you shared on the hub, since the |
vasqu
left a comment
vasqu
left a comment
There was a problem hiding this comment.
Small comments but in general aligned 🫡
CI recapDashboard: View test results in Grafana |
4 participants
What does this PR do?
Some quick things i wanted to fix, can split into two PRs for layer-type and for conversion
PermuteForRopewhich on main actually permutes back from block-split llama format, while we usually want the other way roundTBH we can get rid of
VisionFuseAndPermuteimo and use the corrected rope conversion withChunk(dim=0). Technically breaking but no-one would actually useVisionFuseAndPermuteno?