Daniel Han

@jeromeku Oh ye let's try be device agnostic :)) compile is OK, but I guess handwritting is best :) We then can use CUDAGraphs manually

@jeromeku Fantastic work as always!! very very cool on fusing Adam and Galore!! Love this! Oh on Mixtral - https://github.com/shawntan/scattermoe/tree/main/scattermoe :) Was reading up on this as well :) On...

Wait would this load the lm_head and embed_tokens matrix correctly?

Would it not cause it to be randomnly inited?

Whoopsies sorry on the horrible delay - I'll review this PR and test it out - so sorry!

@charlesCXK @chtmp223 Extreme apologies on the delay - I think I might have fixed it. You need to call `add_new_tokens` before `get_peft_model` to update the vocab, resize, and also save...

@freQuensy23-coder Oh my apologies!! I forgot to change the name of the installation paths. Ie change `cu121_ampere_torch211` to `cu121amperetorch211` with no underscores - hope that works! Sorry again!

@freQuensy23-coder Sorry on the delay!! If you join our Discord server at https://discord.gg/u54VK8m8tk - it'll be an async discussion. But maybe try `pip install --upgrade pip` to update pip Also...

@freQuensy23-coder I just updated Unsloth's pyproject.toml - so maybe it might work better. I suggest: ``` pip install --upgrade pip pip install --upgrade --force-reinstall --no-cache-dir "unsloth[cu121_ampere_torch211] @ git+https://github.com/unslothai/unsloth.git" ```

@gotzmann Thanks for using Unsloth again!! :) Sadly multi GPU is not yet supported for now - we're working on it for a future release in the OSS version