Jee Jee Li

What are the new lora features?

I'm very sorry for missing this PR. I will look at it ASAP. Thank you.

@maxdebayser Perhaps directly deleting `embedding_modules` would be more appropriate?

@maxdebayser Thanks for your explanation. > But, for testing purposes I have the same model where I duplicated the weights for the lm_head and set "tie_word_embeddings": false. When I run...

There's another issue that needs confirmation: whether full sharded LoRA and` add_bias` are supported. If not supported, please refer to: https://github.com/vllm-project/vllm/blob/main/vllm/worker/hpu_model_runner.py#L704-L707

See: https://github.com/vllm-project/vllm/blob/main/tests/lora/test_llama_tp.py#L164

You can try sync the main branch to avoid the CI failure

It doesn't matter, if these failures are not related to this PR, we can consider force merging it.

> It seems LogitsProcessorWithLoRA is always created even if there's no LoRA adapter that needs it, is there a reason for this? The LoRA layers in vLLM are created in...

> > > It seems LogitsProcessorWithLoRA is always created even if there's no LoRA adapter that needs it, is there a reason for this? > > > > > >...