Negative KV sequence length error in Attention op

[jinminxi104]

While testing on the Ascend platform, the attention operator encountered an error due to a negative KV sequence length.
(shareGPT case with 10k prompt + qwen-235B)
Root Cause: The issue occurs when ignore_history is too large in a dp configuration where dp > 1.
Error Logs:

[[36m(RayWorkerWrapper pid=2472948)^[[0m 2026-02-02 12:06:39,393 - lmdeploy - ERROR - model_inputs.py:112 - ignore: tensor([323,   0,   0,   0,          0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
 44672 ^[[36m(RayWorkerWrapper pid=2472948)^[[0m           0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0]

 the value of the key/value's actual sequence lengths[0] must be greater than or equal to 0, but it i       s -294[FUNC:ParseActualSeqLens][FILE:incre_flash_attention_tiling.cc][LINE:875] 

This behavior was introduced in #4265; versions prior to this PR do not exhibit the bug.

[Copilot]

Pull request overview

Fixes a dp>1 decoding failure (negative KV sequence length) by ensuring num_ignored_history does not retain stale/incorrect values when sliding-window attention is disabled.

Changes:

• When cache_config.window_size <= 0, create_model_inputs_delta() now emits a zero num_ignored_history tensor (instead of None) so downstream decoding-input updates reset ignore history to 0.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.