google/gemma-4-26B-A4B-it fails to load: layer-prefix parsing breaks on nested "language_model" submodule

[nureliosman]

Hello,

AirLLM fails to load google/gemma-4-26B-A4B-it (a VLM: vision tower + nested language_model + MoE experts).

Command:
model = AutoModel.from_pretrained("google/gemma-4-26B-A4B-it", compression="4bit")

Output:
using generic AirLLM streaming model for architecture: Gemma4ForConditionalGeneration
ValueError: invalid literal for int() with base 10: 'layers'

Root cause: the model's safetensors index uses nested keys like:
model.language_model.layers.0.experts.down_proj
model.language_model.layers.0.experts.gate_up_proj
model.vision_tower.encoder.layers...

instead of the flat model.layers.N... naming the generic streaming path (airllm/utils.py, split_and_save_layers, line ~284) assumes:
n_layers = len(set([int(k[len(layer_names['layer_prefix']):].split('.')[1]) for k in index.keys() if layer_names['layer_prefix'] in k]))

This regex/parsing logic breaks for any architecture with a nested language_model submodule (VLM-style models) and doesn't handle fused MoE expert tensors (experts.down_proj, experts.gate_up_proj).

Environment: airllm (latest via pip), Windows, Python 3.14, bitsandbytes 0.50.1

BR!