[model] StableLM 2 Zephyr 1.6b
flatsiedatsie opened this issue · comments
I stumbled upon this two week old discussion here, about the StableLM 2 Zephyr 1.6b model becoming available for web-lmm soon.
https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b/discussions/9
I'd really love to work with that model, as my testing so far has shown it to work surprisingly well for its size.
Is there any way to use this model already?
I think I just found the files:
Got stuck on:
[FATAL] /workspace/mlc-llm/3rdparty/tvm/include/tvm/runtime/packed_func.h:1908: Function tvmjs.array.decode_storage(0: runtime.NDArray, 1: basic_string<char>, 2: basic_string<char>, 3: basic_string<char>) -> void expects 4 arguments, but 3 were provided.
put_char @ web-llm.bundle.mjs:3421
This is likely due to an old version of the web-llm npm (if you are not building from source). If you are building from source, this is likely due to the repo not up to date; try pull the recent changes
It would be fantastic if this model could become part of the default supported models.
The multi-language ability is fantastic. I'm very impressed with it, especially for its size.
Awesome, it seems the model has already become available in the Huggingface repo. The chunks exist:
However, the .wasm files are missing from binary-mlc-llm-libs
. I've created an issue about that.
Thanks for the request! We should be able to add the prebuilt wasm files in shortly. cc @YiyanZhai
Fantastic! Thank you!
For the record, I think there are more models for which the shards are available, but the wasm files are not (yet).
- Music
WizardMath- Gorilla
- Gemma 7B
CodeLlamaOpenHermes
Thanks for the list! WizardMath and OpenHermes can reuse the wasm of Mistral (as shown in prebuiltAppConfig
in src/config.ts); CodeLlama should be able to reuse that of Llama-2, as long as they share the same quantization (e.g. q4f16_1
) and number of params (e.g. 7B or 13B).