[vincentzed]

Motivation

On main, a DeepSeek-V4.1 server started with --enable-encoder-swa-bounded-replay and no speculative decoding crashes on its first prefill:

AttributeError: 'DeepSeekV4TokenToKVPool' object has no attribute 'full_to_swa_index_mapping'

Why it crashes:

• Prefill reads the attribute full_to_swa_index_mapping from the pool on every batch.
• That attribute is only created when an allocator registers a mapping.
• In this configuration no allocator does, because the request windows already hold every sliding-window row.
• With DSpark an allocator exists, so the crash does not happen there.

Modifications

• The pool now always has full_to_swa_index_mapping. It stays None when no allocator registers one.
• The sparse prefill cache builder already ignores the mapping when it is given a request-window layout. Its type hint now says the mapping is optional.

Accuracy Tests

• The new unit test TestEncoderReplayWithoutSpeculation builds the pool in this configuration. It fails on main with the error above and passes with the fix.
• On 4x B300 with deepseek-ai/DeepSeek-V4.1-Flash (TP4, EP4), main crashes on the first request with the command below. With this fix the server runs and GSM8K (1319 questions) scores 0.904. That run was on a branch that also carries my decoder bounded replay changes, which do not touch this code path.

python3 -m sglang.launch_server --trust-remote-code --model-path deepseek-ai/DeepSeek-V4.1-Flash --tp 4 --ep-size 4 --mem-fraction-static 0.8 --cuda-graph-max-bs-decode 64 --max-running-requests 16 --chunked-prefill-size 8192 --enable-encoder-swa-bounded-replay --disable-prefill-cuda-graph

Speed Tests and Profiling

Not applicable. This only fixes a crash.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #37569845440
Latest PR Test (Extra): ❌ Run #37569845137
Latest PR Test (AMD ROCm 10): ❌ Run #37569845389

[hnyls2002]

Thanks @vincentzed for tracking down this crash! #43165 has since landed the same fix on main: the pool now initializes full_to_swa_index_mapping to None when request windows hold the SWA rows, with a unit test for encoder replay without speculation. Closing as superseded; please reopen if I've missed a case.

[vincentzed]

Thanks @hnyls2002, agreed. #43165 covers this case, including the encoder replay without speculation test.