d6e40df8faFix `llm_int8_skip_modules` not being respected for VLMs with dynamic quantization on transformers 5.x. Dynamic quant checkpoints (e.g. `gemma-3-4b-it-unsloth-bnb-4bit`) encode skip paths as `language_model.model.layers.*`, but the live module tree on 5.x surfaces them as `model.language_model.layers.*`. This prefix mismatch causes `should_convert_module` to miss the skip list, so 22 modules meant to stay in 16-bit get wrapped in `Linear4bit` without a `quant_state`, producing "Skipping ... no quant_state found" warnings. Patches `should_convert_module` to expand both the module name and the skip patterns into all equivalent alias forms before matching. Guarded by `hasattr` so it is a no-op on transformers 4.x where the bug does not exist. Closes #4208 |
||
|---|---|---|
| .. | ||
| dataprep | ||
| kernels | ||
| models | ||
| registry | ||
| utils | ||
| __init__.py | ||
| _auto_install.py | ||
| chat_templates.py | ||
| device_type.py | ||
| import_fixes.py | ||
| ollama_template_mappers.py | ||
| save.py | ||
| tokenizer_utils.py | ||
| trainer.py | ||