KV-Cache Friendly Document Placement for Llama 4 Scout

Because long-form data processing in Llama 4 Scout's 10M context window demands significant system RAM and NVMe storage for the KV-cache, place the large document corpus first and the question last.

<|begin_of_text|><|header_start|>user<|header_end|> <corpus>{{LARGE_DATASET_DUMP}}</corpus> Based on <corpus> above, compute the summary statistics for Q3.<|eot|> <|header_start|>assistant<|header_end|>