[hnyls2002]

• Deterministic inference truncates each prefill chunk to a multiple of an alignment (4096 by default). A per-rank chunk below it, from --chunked-prefill-size < 4096 or from DP attention splitting the chunk across ranks, truncated every chunk to zero tokens, so a prompt longer than the chunk waited in the queue forever
• Triton: drop the alignment; align_window_kv_to_tiles pads each sliding-window KV list to an absolute multiple of 128, so a chunked prefill sums keys in the same tiles as an unchunked one (built once per forward, deterministic inference only; the attention kernel is unchanged)
• Flashinfer and the DSA kpool alignment: a per-rank chunk shorter than the alignment fails at startup instead of hanging; decode servers are exempt
• Deprecate SGLANG_TRITON_PREFILL_TRUNCATION_ALIGN_SIZE
• Add test_deterministic_chunk_align.py and a chunked-vs-unchunked sliding-window case in test_triton_attention_kernels.py

No behavior change without deterministic inference. With it, triton prefill chunks are no longer aligned (chunked and unchunked outputs stay bitwise equal), and flashinfer configurations that used to hang now fail at startup.

---

CI States

Latest PR Test (Base): 🚫 Run #38014068588
Latest PR Test (Extra): ❌ Run #38014068370
Latest PR Test (AMD ROCm 10): ❌ Run #38014068591

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
sglang-doc	🟢 Ready	View Preview	Oct 9, 2026, 11:40 PM

💡 Tip: Enable Automations to automatically generate PRs for you.

[hnyls2002]

/rerun-test test_deterministic_chunk_align.py test_deterministic.py test_gemma4_swa_triton_oob_regression.py test_unified_memory_deterministic.py test_triton_attention_kernels.py

[github-actions]

🚀 ubuntu-latest (1 test): ✅ View workflow run

cd test/ && python3 registered/unit/managers/test_deterministic_chunk_align.py

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

cd test/ && python3 registered/attention/test_deterministic.py
cd test/ && python3 registered/attention/test_unified_memory_deterministic.py
cd test/ && python3 registered/attention/test_triton_attention_kernels.py

🚀 2-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/attention/test_gemma4_swa_triton_oob_regression.py

[hnyls2002]

/tag-and-rerun-ci