[hnyls2002]

• With encoder SWA bounded replay on, RequestWindow.buffer checked each SWA layer's history before use with boolean-mask indexing and torch.equal: three host syncs per SWA layer per extend step, so the scheduler could not run ahead of the GPU.
• The check now goes through the sync-free invariants API (utils/invariants.py): expect on a dsv41.swa_history_present invariant (FATAL_CONTAINABLE), so SGLANG_INVARIANT_CHECK decides it: off in production (no detection), a throttled count and log at WARN, a device-side assert at STRICT.
• Behavior change: with SGLANG_INVARIANT_CHECK off (the default), servers no longer check SWA history; CI runs at STRICT and still fails on a missing history.
• Verification: DeepSeek-V4.1-Flash on GB300, DP2 attention + MegaMoE (EP2), encoder and decoder replay on, a multi-turn agentic prefill trace at concurrency 512, 900 s per run, same-node pairs.
  • Before: py-spy put 75% of the scheduler main thread's samples in RequestWindow.buffer.
  • Removing the per-layer check: +4.5% to +6.7% tok/s/GPU on three nodes; the step interval went from ~405 to ~375 ms at the same step shape.
  • The deferred-flag version of this check (an earlier revision of this PR) ran within -1.3% to +0.7% of removing the check on three nodes, with no false raise; the invariants version adds only the comparison kernels.
  • test_dsv41_request_window_copy.py passes. A one-off GPU script confirmed: matching history passes at STRICT; missing history passes at OFF and trips invariant-check FAILED [dsv41.swa_history_present] at STRICT.

---

CI States

Latest PR Test (Base): 🚫 Run #39318746765
Latest PR Test (Extra): ❌ Run #39318746634
Latest PR Test (AMD ROCm 10): 🚫 Run #39318746757

[hnyls2002]

/rerun-test test_dsv41_request_window_copy.py test_deepseek_v41_flash_gb300_encoder_bcg.py registered/attention/unittests/dsv4/test_deepseek_v4.py test_q8kv8_sparse_prefill_backend.py test_encoder_swa_replay.py test_dsv4_compressed_pools.py test_invariants.py test_dsv41_late_layer_tail_readers.py

[hnyls2002]

/tag-and-rerun-ci

[github-actions]

🚀 1-gpu-h100 (3 tests): ✅ View workflow run

cd test/ && python3 registered/kernels/ops/attention/test_dsv41_request_window_copy.py
cd test/ && python3 registered/attention/unittests/dsv4/test_deepseek_v4.py
cd test/ && python3 registered/kernels/ops/attention/test_q8kv8_sparse_prefill_backend.py

🚀 4-gpu-gb300 (1 test): ❌ View workflow run

cd test/ && python3 registered/e2e/models/test_deepseek_v41_flash_gb300_encoder_bcg.py

🚀 4-gpu-b200 (1 test): ❌ View workflow run

cd test/ && python3 registered/attention/unittests/dsv4/test_deepseek_v4.py

🚀 ubuntu-latest (4 tests): ✅ View workflow run

cd test/ && python3 registered/unit/model_executor/test_encoder_swa_replay.py
cd test/ && python3 registered/unit/mem_cache/test_dsv4_compressed_pools.py
cd test/ && python3 registered/unit/utils/test_invariants.py
cd test/ && python3 registered/unit/layers/test_dsv41_late_layer_tail_readers.py