Slide the context window

Background

The model's positional table has only block_size rows, so each generation step crops the running context to the most recent block_size tokens:

idx_cond = idx if idx.size(1) <= block_size else idx[:, -block_size:]

Older tokens fall off the left as generation continues.

Problem statement

Implement crop_context(ids, block_size) returning the last block_size tokens (or all of them if shorter).

Input

• ids — list of token ids.
• block_size — int, the model's max context length.

Output

Returns a list: the final block_size ids, or all of ids if len(ids) <= block_size.

Examples

Example 1 — crop to the last 3

Input:  ids = [1, 2, 3, 4, 5], block_size = 3
Output: [3, 4, 5]

Example 2 — shorter than the window

Input:  ids = [1, 2], block_size = 5
Output: [1, 2]

Constraints

• Return ids[-block_size:].
• If len(ids) <= block_size, return ids unchanged.

Notes

• The model only ever saw ≤ block_size tokens in training, so the window slide keeps every forward pass in-distribution.