mlx-community/gemma-3n-E4B-it-4bit · Got bug when loading model with `mlx-vlm`

Jul 7

Error loading model: Received 269 parameters not in model:
audio_tower.conformer.0.attention.attn.k_proj.biases,
audio_tower.conformer.0.attention.attn.k_proj.scales,
audio_tower.conformer.0.attention.attn.q_proj.biases,
audio_tower.conformer.0.attention.attn.q_proj.scales,
audio_tower.conformer.0.attention.attn.relative_position_embedding.pos_proj.biases,
audio_tower.conformer.0.attention.attn.relative_position_embedding.pos_proj.scales,
audio_tower.conformer.0.attention.attn.v_proj.biases,
audio_tower.conformer.0.attention.attn.v_proj.scales,
audio_tower.conformer.0.attention.post.biases,
audio_tower.conformer.0.attention.post.scales,
audio_tower.conformer.0.ffw_layer_end.ffw_layer_1.biases,
audio_tower.conformer.0.ffw_layer_end.ffw_layer_1.scales,
audio_tower.conformer.0.ffw_layer_end.ffw_layer_2.biases,
audio_tower.conformer.0.ffw_layer_end.ffw_layer_2.scales,
....

depasquale

MLX Community org Jul 9

This is for testing the full vision model, which is currently being ported in mlx-swift-examples.

hossam96

Jul 13

I have the exact same bug, here is the command I'm using:
mlx_vlm.generate --model mlx-community/gemma-3n-E4B-it-4bit --max-tokens 100 --audio 2025-02-11T19_48_10.011387.wav --prompt "..."