[Tracking] SLM Rollout Wrap up

[CharlieFRuan]

Overview

As SLM rollout is almost completed: #1420, there are some remaining steps to clear the confusion.

Action Items

W0: Tutorial notebook as CI
The Colab notebooks have served as an entrance for many people new to mlc-llm, and thus it would also be nice to add them as tests:

• [Bug] [llama2-7B] fail to execute Llama-2-7b-chat-hf-q4f16_1-MLC #1551
• Add new model architecture tutorial not working web-llm#274
• tutorial_chat_module_getting_started.ipynb also seem to broken
• Fix notebooks
• Add notebooks as tests

W1: Transition from the old mlc_chat_cli to the new SLM CLI
Afterward, we should clean up the models prefixed with mlc-chat- in https://huggingface.co/mlc-ai, and model libraries not in a folder: https://github.com/mlc-ai/binary-mlc-llm-libs, as these outdated and unmaintained models and weights have caused much confusion in the issues (e.g. one below, and some other instances in discord).

• Function tvmjs.array.decode_storage error when deploy the model WizardCoder-15B-V1.0-q4f16_1 web-llm#273

• [Tracking] Transition from the old mlc_chat_cli to the new SLM CLI #1707

W2: Model support parity
Comparing the two directories, python/mlc_chat/model and mlc_llm/relax_model, there some architectures that are still not supported in SLM. While some could be slightly obsolete now, some may still be good to have.

• stablelm
  • [Bug] StableLM model conversion not working #1616
• RWKV
  • [SLM] RWKV5 World Support #1787
• gptj
  • [SLM] Add support for gptj architecture #1751
• chatglm
• minigpt (perhaps wait for LLaVA)

W3: Other compilation errors due to batching

• [Bug] MACOS M1 webgpu build undefined symbols  #1628

[CharlieFRuan]

Feel free to add more items!

[DavidGOrtega]

Should these be added?

• [Bug] StableLM model conversion not working #1616 ?
• [Bug] MACOS M1 webgpu build undefined symbols  #1628 ?