Hello,I have some trouble to reproduce the results on llama-13b.An error "scaling_matrix_inv = torch.linalg.inv(scaling_diag_matrix) torch._C._LinAlgError: linalg.inv: The diagonal element 6940 is zero, the inversion could not be completed because the input matrix is singular" occurs on line 203, in whitening function.
How can I sovle this problem? Thanks.
Hello!
I was wondering if this code can be manipulated to transform a tensor - say (32128128) to a smaller tensor (86464).
Basically reduce the size of the llm layer by layer.
I tried to use the provided scripts to compress LLAMA 2 with 0.2 compression ratio. The model evaluation script shows a perplexity of 7.2 on wikitext, but the model responses are mostly incoherent. I am getting responses like
Instruction: tell me about you==\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\ selecting\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\
where as original model is giving decent responses.
Is there any modification to be done for the inference script or the tokeniser after model compression? , Is there an inference script within the repository?
Hi, thank you for your reply. But I still get the same problem as mentioned before.
Traceback (most recent call last):
File "/home/xxx/SVD-LLM/SVDLLM_new.py", line 193, in whitening
scaling_matrix_inv = torch.linalg.inv(scaling_diag_matrix)
torch._C._LinAlgError: linalg.inv: The diagonal element 6940 is zero, the inversion could not be completed because the input matrix is singular. "
My python environment is built on requirements.txt. And I run the code on 2 3090 GPUs
Firstly, I want to express my gratitude for the fascinating work you've been doing. It's been inspiring.
I've recently come across your paper where you describe the integration of SVD-LLM with GPTQ, and I'm eager to explore the implementation further.
Could you please share the code where you've integrated SVD-LLM with GPTQ as described in the paper?
Your assistance in providing access to this code would be appreciated. Thank you for your time and consideration.