# Reproduction Use source docling-project/ScreenVLM at 78df0cc8498bbea1a03eb2816056fc0a76d2f63d and ggml-org/llama.cpp at 0ec4027a2a0183de87c007069adddc06e665f7de. Verify original safetensors SHA256: c02d35755f6b6d3b85b8354e3ba42422ede06a2151c09034e5112a3fc4c96c91. Prepare a converter copy exactly as described in tokenizer-repair.json; remove only the two added-token entries with IDs >=100352 from tokenizer.json, tokenizer_config.json, special_tokens_map.json and added_tokens.json. Do not modify original weights, the vocabulary IDs 0..100351 or the chat template. ```sh python llama.cpp/convert_hf_to_gguf.py converter-source --outtype f16 --metadata conversion-metadata.json --outfile ScreenVLM-F16.gguf python llama.cpp/convert_hf_to_gguf.py source --outtype f16 --mmproj --metadata conversion-metadata.json --outfile mmproj-ScreenVLM-F16.gguf llama.cpp/build/bin/llama-quantize ScreenVLM-F16.gguf ScreenVLM-Q8_0.gguf Q8_0 4 ```