Yuanping Song enhanced the IBM/vllm repository by improving the DeepSeek-OCR decoding pipeline to address repeated n-grams in greedy decoding. Using Python and leveraging deep learning and NLP techniques, Yuanping modified the NGramPerReqLogitsProcessor to be argmax-invariant, which reduced repetitive sequences and increased the diversity and relevance of OCR-generated text. Additionally, Yuanping applied the NoRepeatNGramLogitsProcessor to the greedy path, directly targeting issues with repetitive outputs. These targeted changes improved the quality and reliability of OCR text, aligning output with downstream parsing and indexing requirements. The work demonstrated focused engineering depth within a complex machine learning codebase.