Attention Sinks and Sliding Windows

Problem

Return the sorted token positions retained by a cache that keeps the first sink_tokens positions and the last window_tokens positions of a length-token sequence. Clip both regions to the existing sequence and remove overlap. The window includes the newest token. A zero-length region contributes no positions.

def retained_cache_positions(length, sink_tokens, window_tokens):

Constraints

• 0 ≤ length ≤ 128; 0 ≤ sink_tokens,window_tokens ≤ 128.
• Return positions in increasing order; this is a retention policy, not attention computation.
• Inputs satisfy these constraints; invalid-input handling is not graded. Use only Python standard library and NumPy.

Examples

Arguments: [
  10,
  2,
  3
]
Expected result: [
  0,
  1,
  7,
  8,
  9
]

Arguments: [
  4,
  3,
  3
]
Expected result: [
  0,
  1,
  2,
  3
]