Ar-Kareem

In the decoder an LSTM is used https://github.com/alexis-jacq/Pytorch-Sketch-RNN/blob/5c3e21375dfe7695c1c37a0acccf6da17c049f77/sketch_rnn.py#L151-L157 While in the original [paper](https://arxiv.org/abs/1704.03477), the description of the architecture at page 6 states that > For the decoder RNN, we use...

Seems this was fixed in v2 (probably since larger models were used and this bug was identified) but I noticed it still existed in v1. Great work by the way.

Tiny numerical differences, Weight updates not perfectly matching

2

Hi, thanks for this amazing library. I saw one tiny issue which is that the final weights of the model is different when training with multiple sub_batches per step vs...

fix recursion error when setting tp_wrapped_module #122

4

To fix the error mentioned in #122 This happens whenever the attribute `tp_wrapped_module` is changed (for example inside LoRA or other PEFT methods) I am not 100% sure this works...

Max Recursion Error when using with lora

2

I get the following error when attempting to use LoRa with Llama 2 ``` File "/path/libraries/conda/lib/python3.9/site-packages/tensor_parallel/wrapper.py", line 75, in __getattr__ return getattr(self.tp_wrapped_module, attr) [Previous line repeated 2979 more times] RecursionError:...