SafeTensors have issue?
Troubleshooting revealed it seems 😞 :
config.json on HF declares mtp_num_hidden_layers: 1, which tells convert_hf_to_gguf.py to expect 33 blocks (32 main + 1 MTP). But the model.safetensors.index.json on HF contains zero MTP weight keys — no layers.32., no mtp., nothing. The published safetensors only carry 32 layers (blk.0–blk.31). The config and the weights disagree at the source.
The converter trusts the config, writes block_count=33 into the GGUF header, then only finds tensors for 32 blocks. llama.cpp trusts the header, looks for blk.32.attn_norm.weight, and hard-fails. Every GGUF produced from this HF repo — by any tool, on any machine — will have the same missing-tensor failure until deepreinforce-ai either removes mtp_num_hidden_layers from config or publishes the actual MTP weights.