MarktHart
MarktHart
Ah, I seemed to have missed that SpeedBrain uses relative lengths. This raises another problem however, that you assume that the maximum length is equal to the padded batch. This...
This change also breaks some import scripts, for example Huggingface's.
It breaks `AutoModel.from_pretrained` for any model that has weightnorm and was saved with a previous version. Huggingface (and me working further on that) uses the names of parameters to determine...
It needs to be fixed in code in the import scripts. Which is fine, but also somewhat annoying. You basically need a version check or a try-catch. If the PyTorch...
@ezyang, @lezcano Let's preface this with that I am not affiliated with Huggingface. I think the core problem is with [this line](https://github.com/huggingface/transformers/blob/bbf3090848cf0ceff98f9465691e9ecce63684a1/src/transformers/modeling_utils.py#L3016). Hugginface relies on the state dict keys to...
> Hi @huseinzol05 Thanks for the whisper support and the example. > > Can you make `whisper_example.py` also support long audio (> 30 seconds)? The example currently works up to...
You could go even closer to attention and use it as is with a doubled interleave. E.g. ```Python def alternative_forward( self, x, rel_pos, attn_mask=None, ): bsz, tgt_len, embed_dim = x.size()...
FYI: the code on master requires a tiny change to make it working, but runs in ~41 minutes on a 2xRTX4090 machine Changes that work for now (I don't plan...
> With that said, there are some compromises to help bridge the gap. You can use an A100 on google colab for $9/month for 20 hrs of testing a month....