Phi 1.5 support

[axel578]

Phi 1.5 is a new model from Microsoft, supporting this model would be extremely usefull.

A detailed list of info of phi 1.5 can be found here : https://huggingface.co/microsoft/phi-1_5

Its basically supporting MixFormerSequentialConfig .
The phi 1.5 has weird features, also 4 bit support would be great !! (and not only on gpu, but cpu also please, this model size should work ok on cpu)

[viktor-ferenczi]

Please consider adding this to the Roadmap.

It is also a low hanging fruit for new contributors learning into vLLM's code base.

[adivoj]

We neeed it

[Bojun-Feng]

Hi, I am interested in integrating phi 1.5. However, I am new to the codebase and am a bit overwhelmed by the large code base. Are there any recommended resources to get started?

[adivoj]

Hi Bojum,

Good to hear that you're interested. I found the models being defined here: /vllm/model_executor/models and are registered here: /vllm/model_executor/model_loader.py.

Here are some phi1.5 sources you can use:
https://huggingface.co/microsoft/phi-1_5/tree/main
PR: https://huggingface.co/microsoft/phi-1_5/discussions/22/files
https://github.com/OpenAccess-AI-Collective/axolotl/tree/2d60ba3a6ea4def14e6ab974299322a0bf90d5bb/src/axolotl/models/phi

I was thinking of pushing all models to GPT4 and make it come up with the new one based on modeling_mixformer_sequential.py :)

[Bojun-Feng]

Hi adivoj, thank you so much for the quick response! Unfortunately, I found out that my machine is not compatible with vLLM and therefore can not locally test the code. As a result, other people might be a better fit...

[Linzecong]

I tried to adapt it and successfully aligned the model weights. The output results were correct before passing through the attention layer, but they were incorrect after passing through the attention layer. Later I found that he used CrossAttention https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py#L518 I am not sure if this problem is caused.

[andersonbcdefg]

I was looking into this. I believe the "cross-attention" used in Phi-1.5 is not true cross-attention, it's just used for current token to attend to past KV-cache during autoregressive generation. From what I can see, the Phi-1.5 architecture is basically the same as a GPT-NeoX with attention/FFN in parallel, except:

1. In each transformer block the pre-FFN and pre-Attention layernorms are the same (GPT-NeoX doesn't share these params, as noted in their paper).
2. Phi-1.5 has a bias on the output linear layer, and GPT-NeoX does not.

Alternatively, Phi-1.5 has the same architecture as GPT-J, except:

1. GPT-J has separate q, k, v projections and Phi-1.5 has W_qkv
2. GPT-J does not have bias on the q, k, v layers, and Phi-1.5 does have bias on W_qkv

It shouldn't be too crazy to adapt these, given that vLLM already supports GPT-NeoX and GPT-J. The MixFormer modelling file on Hugging Face is unnecessarily complicated relative to how small the architecture changes they made are relative to GPT-NeoX/GPT-J. To me it seems like the easiest way would be:

• Start with GPT-NeoX model, with the right configuration to match Phi-1.5 (including use_parallel_residual)
• When loading weights in from Phi-1.5, just copy the same layernorm weight and bias from Phi into both input_layernorm and post_attn_layernorm in GPT-NeoX
• Modify the output linear layer to have a bias so that the bias can be loaded in as well.

This is all it should take!! I don't have time to do this myself right now because I haven't contributed to vLLM before, but for anyone who's added a model before this should be easy!!

[check39]

Any update on this? Can anyone make change accordingly if you are familiar with code base?

[maximzubkov]

Hey guys, just opened the PR integrating phi-1.5 into the codebase. Even though the model works, the output on some sequences does not exactly match the HF implementation, so if anyone could have a look, it could fasten the merge progress a lot.

Thanks! @andersonbcdefg @Linzecong @adivoj @Bojun-Feng

[adivoj]

Wow, thanks a lot man! All the best to you :) Couldn't it be the temp. or top_p? You had those at minimum while testing I guess.

[Linzecong]

Hey guys, just opened the PR integrating phi-1.5 into the codebase. Even though the model works, the output on some sequences does not exactly match the HF implementation, so if anyone could have a look, it could fasten the merge progress a lot.

Thanks! @andersonbcdefg @Linzecong @adivoj @Bojun-Feng

Thank you very much for your support! I did a quick verification, I set top_k to 1 and the output aligned in most cases. I think it is caused by floating point precision issues, because when I use float32 to load the model, the hf output is very stable when top_k is 1. But when using half to load, even if top_k is 1, the output is still unstable.

[adivoj]

Man, they just announced phi v 2.0 https://youtu.be/S2gn_EoSWac?feature=shared