执行摘要

PR #31369 引入的 mamba track-boundary seqlen 变更导致 Qwen3-Next 模型在 decode-cache-hit 场景下 logprobs 出现确定性 KL 散度超标。本 PR 将其完全回退，恢复至已通过的基线行为。

功能与动机

PR #31369 变更了 _mamba_check_track_boundary 中的 seqlen 计算方式，意图修复 overlap scheduler 下的边界问题。但该变更引入了确定性正确性回归：nightly 测试 test_qwen3_next_models.py::test_input_output_logprobs_match_decode_cache_hit_helper 失败，KL 散度从 ≤0.0005 跳升至 0.0065（阈值 0.002）。二分法定位 #31369 为唯一回归源。

实现拆解

1. 完全回退提交：执行 git revert 0675d303，精确撤销 #31369 的变更。
2. 单文件修改：仅 python/sglang/srt/managers/scheduler_components/batch_result_processor.py 中的 _mamba_check_track_boundary 方法。
3. 逻辑还原：seq_len 计算从 req.kv_committed_len 恢复为 len(req.origin_input_ids) + len(req.output_ids) - 1。
4. 无额外配套：不含测试、配置或部署改动。

python/sglang/srt/managers/scheduler_components/batch_result_processor.py

核心方法 _mamba_check_track_boundary 的 seqlen 计算逻辑从 kv_committed_len 回退到 len(origin_input_ids) + len(output_ids) - 1，消除了回归。

def _mamba_check_track_boundary(self, req, batch, result, i):
    """Check if this decode step crosses a mamba track interval boundary.
​
    Returns (at_boundary, track_seqlen).  ``track_seqlen`` must equal the
    seq_len the forward's tracking mask used, so the tracked state and its
    recorded ``mamba_last_track_seqlen`` describe the same token position.
    That seq_len is a pure function of the tokens the request has produced:
    ``len(origin_input_ids) + len(output_ids) - 1`` (the just-decoded token
    is already appended to ``output_ids`` before this runs).
​
    ``kv_committed_len`` must NOT be used here: under the overlap scheduler,
    ``prepare_for_decode`` for the *next* batch increments it before this
    result is processed, so it leads seq_len by a jittering lookahead
    (0 or 1 depending on prefill interleaving). Using it fires the boundary
    one decode step early on most steps, mislabeling the tracked mamba
    state; a later request that reuses/donates that tracked prefix then
    extends from a state a cold prefill recompute would not produce.
​
    For spec decode, the boundary is detected by comparing the
    accepted seq_len range against interval boundaries.
    """
    interval = get_server_args().mamba_track_interval
​
    if batch.spec_algorithm.is_none():
        seq_len = len(req.origin_input_ids) + len(req.output_ids) - 1
        if seq_len % interval == 0:
            return True, seq_len
    elif result.num_correct_drafts_per_req_cpu is not None:
        cur = req.seqlen - 1
        prev = cur - result.num_correct_drafts_per_req_cpu[i] - 1
        if cur // interval != prev // interval:
            return True, cur // interval * interval
​
    return False, 0

评论区精华

• Alisonshao 通过硬件 A/B 验证：回退后 KL 散度重新低于阈值（通过 rerun-test 确认）。
• PR body 指出正确的长期修复路径是 #29792。

风险与影响

• 风险：低风险纯回退。注意 overlap scheduler 边界问题仍未解决，需等待 #29792。
• 影响：恢复 Qwen3-Next 测试通过性，无性能退化。

关联脉络

• 本 PR 回退 #31369。
• 正确的 mamba track-boundary overlap 调度修复正在 #29792 中推进。