Danilo Peixoto

Temporary patch applied to custom `Dockerfile` for MLflow `>2.4.0`: ```Dockerfile RUN STATIC_DIR=$(python3 -c 'from mlflow.server import app;print(app.static_folder)') && \ CSS_FILENAME=$(ls ${STATIC_DIR}/static/css/main.*.css) && \ sed -i 's|/static-files/static|..|g' ${CSS_FILENAME} ```

I am currently fine-tuning a quantized version of the Mixtral 8x7B model using a MacBook Pro M3 Max with 64GB of memory. However, I've encountered an issue where I'm seeing...

@awni This is amazing! Were you using `float16` or `bfloat16` data type for fine-tuning? I only managed to fine-tune successfully and get rid of `nan` losses by casting `float16` values...

@awni I recently had the chance to try out the latest version of MLX (`v0.2.0`), and indeed, the experiment with the pure linear predicate (`lambda m: isinstance(m, nn.Linear)`) was resulting...

Yes. The fine-tuning status for the experiment in discussion: Linear class predicate vs Precision float16 float32 lambda m: isinstance(m, nn.Linear) NaN losses Successful lambda m: isinstance(m, nn.Linear) and m.weight.shape[0] !=...

I believe we can fine-tune the same experiment with a purely linear predicate using the `bfloat16` data type in MLX version `0.3.0`. However, I haven't had the opportunity to test...

We could be referring to solutions like [Spark](https://spark.apache.org/docs/latest/cluster-overview.html) or [Dask](https://docs.dask.org/en/stable/deploying.html) (with local, Kubernetes, MPI, and other backends) for distributed data processing (that could eventually implement ML specifics, such as [Dask-ML](https://ml.dask.org))...

@LeaveNhA, I encountered challenges while working on a prototype with the Dask Backend Entrypoint API: - The MLX data type is not an alias of `np.dtype` as expected by Dask....

It seems the MLX team added MPI distributed training support!

+1