Prince Canuma
Prince Canuma
Could you try to run the model in a Python file like this: ```python import sys def main(): try: import mlx.core as mx from mlx_vlm import load, generate from mlx_vlm.prompt_utils...
Interesting! Thanks for the update. I have one question. Are you making requests in batch? if so what is the use case?
Could you share the method / reproducible example you used to indetifiy `_merge_input_ids_with_image_features` and `async_eval` as the culprits?
Ohh yeah, I found the issue here #84. Upgrade your MLX version to the latest and let me know if it solves it. > pip install -U mlx
And yeah, that large of an image is not a good idea. Try passing `--resize-shape` to generate.py and use a smaller image It will be faster and less resource intensive
It works well for Linear, but I can't get it to match the output torch for Conv layers from torch, even if I override the weights with MLX ones. The...
Example usage with conv1d: ## MLX ```python import mlx.core as mx import mlx.nn as nn from typing import Optional, Any from dataclasses import dataclass import mlx.core as mx import numpy...
Thank you very much @cavit99, indeed it does solve it and matches the torch design! ❤️ > MLX: [out_channels, kernel_size, in_channels] > This ordering difference needs special handling when normalizing,...
No worries, I think you did a great job and gave it the necessary attention!