[mimi] mimi conv cache edge case #43130
Conversation
|
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. |
| self.per_layer_padding = per_layer_padding | ||
| self.per_layer_padding_mode = per_layer_padding_mode | ||
| self.per_layer_in_channels = per_layer_in_channels | ||
| self.per_layer_is_init = [True] * num_layers |
There was a problem hiding this comment.
this is not used anywhere, let's remove
| torch.ones(batch_size, in_channels, padding, device=device, dtype=dtype) * hidden_states[..., :1] | ||
| ) | ||
| else: | ||
| raise NotImplementedError(f"Padding mode {padding_mode} not supported") |
There was a problem hiding this comment.
Good idea to add _cache_init so that different models can initialize differently if needed
How about removing the check here inside init? Since you have a similar one in this new method?
There was a problem hiding this comment.
Agree we could remove, but also it can be convenient to have raise an error at init to avoid having to run the full forward
There was a problem hiding this comment.
hum actually you're right, overwise it will be inconvenient and we'll have to also overwrite the init when working with other modes! thanks, removing
|
[For maintainers] Suggested jobs to run (before merge) run-slow: kyutai_speech_to_text, mimi |
What does this PR do?
two things;