causal-conv1d: add layers module for LayerRepository - #1137
Conversation
Signed-off-by: kaixuanliu <kaixuan.liu@intel.com>
|
/kernel-bot build causal-conv1d |
|
Build request processed. Command: Dispatched (1):
|
| __all__ = ["causal_conv1d_fn", "causal_conv1d_update", "causal_conv1d_varlen_states"] | ||
| __all__ = [ | ||
| # wrappers | ||
| "layers", |
There was a problem hiding this comment.
Since this changes the API, could you bump up the version in build.toml?
Signed-off-by: kaixuanliu <kaixuan.liu@intel.com>
|
It's a bit weird because we already use mamba-ssm and its causal conv1d as layers in transformers no? Not sure what the benefit here is other than additional maintenance 👀 |
|
Pretty sure it should be the same implementation (mamba uses it for its fused training path) |
|
Yeah, I think it'd make most sense to use |
vasqu
left a comment
vasqu
left a comment
There was a problem hiding this comment.
Okay let's do it here as well. Tbh it really doesn't hurt to have it on both sides ig
Would be still great to check what XPU could use on transformers side then 🤗 (you can cc me there then)
|
/kernel-bot build causal-conv1d |
|
Build request processed. Command: Dispatched (1):
|
3 participants
Adds a
layerssubmodule tocausal-conv1d, exposingcausal_conv1d_fnandcausal_conv1d_updateasnn.Modulewrappers so the repo can be consumed viaLayerRepository/kernelize(), as transformers will deprecateFuncRepositorySoon. @danieldk , pls help review, thx!