cpatonn's "AWQ-4bit" repo is actually quantized via llm-compressor -- config.json declares compressed-tensors. Passing awq_marlin explicitly conflicted with the checkpoint's own declared format and 400d at config-validation time.
cpatonn's "AWQ-4bit" repo is actually quantized via llm-compressor -- config.json declares compressed-tensors. Passing awq_marlin explicitly conflicted with the checkpoint's own declared format and 400d at config-validation time.