[DO NOT MERGE] Reproduce #31445

[stepinto]

This is the reproduce code for #31445.

This PR includes:

• Check for host_indices are allocated in pthread aux thread.
• A testing script, which send several requests, flush cache, and then send the same requests, to trigger prefetch.

Reproduce steps:

1. Run sglang.

nohup sglang serve \
    --model=deepseek-ai/DeepSeek-V4-Flash \
    --tp-size 8 \
    --moe-runner-backend flashinfer_mxfp4 \
    --disable-cuda-graph \
    --max-total-tokens=$((256*1024)) \
    --swa-full-tokens-ratio=1 \
    --mem-fraction-static 0.85 \
    --max-running-requests=4 \
    --host 0.0.0.0 \
    --port 30000 \
    --enable-metrics \
    --enable-hierarchical-cache \
    --hicache-storage-prefetch-policy best_effort \
    --hicache-storage-backend=mooncake \
    --hicache-storage-backend-extra-config '{"master_server_address": "127.0.0.1:50051", "local_hostname": "localhost", "metadata_server": "http://127.0.0.1:8080/metadata", "global_segment_size": "256gb", "protocol": "tcp"}' \
    > sglang.log 2>&1 &

2. Run test script.

./repro_31445.py -n 16 --length 32768

3. Grep for AssertionError in sglang.log. The background thread seems catches AssertionError, so it will lead to server crash.

---

CI States

Latest PR Test (Base): ❌ Run #30003186605
Latest PR Test (Extra): ❌ Run #30003186432

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.