transformers - 💡(How to fix) Fix MoonshineStreaming: sliding_window_mask_function off-by-one — right_window_size=4 attends to 3 future frames instead of 4 [1 pull requests]

System Info

• transformers version: 5.9.0
• Platform: Linux-6.8.0-100-generic-x86_64-with-glibc2.39
• Python version: 3.12.3
• PyTorch version: 2.12.0

Who can help?

@eustlb

Information

sliding_window_mask_function in modeling_moonshine_streaming.py (line 349) uses strict less-than (< right_window_size) for the right context, which means right_window_size=4 only attends to 3 future frames instead of 4.

The Moonshine v2 paper (arXiv:2602.12241) explicitly states:

"a right window of w_right=4 implies an algorithmic lookahead of 4×20 ms = 80 ms"

At 50 Hz, 4 future frames = 80 ms. But the current code gives 3 future frames = 60 ms.

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from transformers.models.moonshine_streaming.modeling_moonshine_streaming import sliding_window_mask_function

fn = sliding_window_mask_function((16, 4))

q = 10 future = [kv for kv in range(q + 1, 20) if fn(0, 0, q, kv)] print(f"Future frames attended: {future}") Actual: [11, 12, 13] — 3 frames, 60ms Expected: [11, 12, 13, 14] — 4 frames, 80ms (per paper)

Expected behavior

With sliding_window_mask_function((16, 4)), a query at position 10 should attend to future positions {11, 12, 13, 14} — 4 future frames, matching the paper's definition of w_right=4