Incident: A sliding window that opened the future
Module 10 · Incident file · About 35 minutes
What happened
On 27 September 2025 a user with the GitHub name NixGD opened issue #41184 on Hugging Face's transformers library, titled "Continuous Batching sliding window attention mask is wrong" (github.com/huggingface/transformers/issues/41184). They were running a development build (4.57.0.dev0) and reported: "Using Continuous Batching with gpt-oss gives meaningless results. I believe this is the reason."
Their reproduction was eight lines. It built two masks for one 16-token sequence with the library's build_attention_mask function: a full causal mask, and a mask with a sliding window of 4. Then it counted the query-key pairs each mask allowed. The output was:
Key/Query pairs the model can attend to (full mask): 136
Key/Query pairs the model can attend to (window mask): 202A window should allow fewer pairs than causal attention, not more. In the reporter's words, "the window-attention should be _more_ restrictive."
The response was quick. On 30 September remi-or, the collaborator the reporter had tagged, opened pull request #41228, "Fix sliding window attn mask", and merged it under two hours later (github.com/huggingface/transformers/pull/41228). It changed two lines, added a long docstring with drawn masks, and added a test "so it does not happen again". remi-or replied in the issue: "Good catch, it seems the sliding window mask is off." The issue was closed on 13 October 2025. The issue does not say how many users ran the affected path; the cost on record is a model giving meaningless output.