Missing MTP Parameters when compressing DeepSeekV3

[wZuck]

Since the DeepSeek-V3 Hugging Face Transformers modeling file does not include MTP implementations, the llm-compressor cannot recognize and load MTP weights. After compression, the MTP layer will not be preserved. Could we provide an API to track all unused tensors and save them to a new safetensors file?

[kylesayrs]

Hi @wZuck

Perhaps you could provide a script and more context as to how you're loading MTP weights? LLM Compressor aims to be compatible with HF models, and cannot be compatible with all transformer extension libraries