[Bugfix][KV Connector] Safely fill circular buffers in DecodeBench

[majunze2001]

Purpose

DecodeBench fills allocated cache blocks with synthetic values to emulate an external prefill. Qwen QSA circular buffers can pack exact int64 MRoPE positions alongside BF16 key state. Applying the configured nonzero BF16 fill
to the entire circular-buffer row corrupts those positions, which can make the QSA pre-indexer access its cosine/sine table out of bounds and trigger a CUDA IMA.

This change identifies KV-cache groups containing CircularBufferSpec layers and explicitly zero-fills their selected blocks. Other cache groups continue to use the configured DecodeBench mean and standard deviation. The regression test also verifies that only the selected circular-buffer row is modified.

Duplicate-work check

No issue is linked to this change. Searches of open vLLM PRs and issues for DecodeBench CircularBuffer, DecodeBench Qwen3.8, and QSA illegal memory DecodeBench found no existing fix.

Test Plan

• Exercise a worker-side CircularBufferSpec cache initialized with nonzero
  values and verify that DecodeBench zeros the selected row while leaving every
  other row unchanged.
• Run the complete DecodeBench connector unit-test file.
• Run pre-commit on both changed files.
• Run a Qwen3.8 Flash Next DecodeBench GPU E2E with the default nonzero fill
  configuration.
• Run Qwen3-0.6B as a non-circular DecodeBench control.

Commands run locally:

../vllm/.venv/bin/python -c 'import runpy,sys,types; sys.modules["readline"]=types.ModuleType("readline"); sys.argv=["pytest","tests/v1/kv_connector/unit/test_decode_bench_connector.py","-q"]; runpy.run_module("pytest",run_name="__main__")'
../vllm/.venv/bin/pre-commit run --files vllm/distributed/kv_transfer/kv_connector/v1/decode_bench_connector.py tests/v1/kv_connector/unit/test_decode_bench_connector.py
git diff --check

Equivalent model-server invocations used for the E2E runs:

vllm serve Qwen/Qwen3.8-Flash-Next-FP8 \
  --tensor-parallel-size 4 \
  --max-model-len 4096 \
  --max-num-batched-tokens 4096 \
  --max-num-seqs 4 \
  --enable-prefix-caching \
  --kv-transfer-config \
  '{"kv_connector":"DecodeBenchConnector","kv_role":"kv_both","kv_connector_extra_config":{"fill_mean":0.015,"fill_std":0.0}}'

vllm serve Qwen/Qwen3-0.6B \
  --max-model-len 4096 \
  --max-num-batched-tokens 4096 \
  --max-num-seqs 8 \
  --enable-prefix-caching \
  --kv-transfer-config \
  '{"kv_connector":"DecodeBenchConnector","kv_role":"kv_both"}'

Test Result

• tests/v1/kv_connector/unit/test_decode_bench_connector.py: 12 passed.
• Changed-file pre-commit hooks: passed, including Ruff, mypy, SPDX, and the
  repository-specific validation hooks.
• git diff --check: passed.
• Qwen3.8 Flash Next, tensor parallel size 4: 10/10 completion requests
  returned HTTP 200 with the requested output lengths. This included the prior
  3,002-token prompt plus 512-token generation regression shape, a 1,024-token
  generation, replayed prefixes, and concurrent requests. The server remained
  healthy, with no CUDA illegal memory access, engine failure, or CUDA core
  dump.
• Qwen3-0.6B non-circular control: 8/8 completion requests returned HTTP 200,
  including replayed-prefix, long-generation, and concurrent workloads. The
  normal attention group retained the configured default fill_mean=0.015,
  and the run observed a 29.6% local prefix-cache hit rate.

Model accuracy evaluation is not applicable because DecodeBench deliberately uses synthetic KV values and does not produce meaningful accuracy results. The model-serving E2E tests above cover the affected circular-buffer path and a non-circular control.

AI assistance disclosure: OpenAI Codex was used for this PR.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[ivanium]

LGMT thanks for the fix!

[ivanium]

/ci run

[github-actions]

✅ Triggered Buildkite CI #86944 for commit a556fb47bf0b.