Issue evidence · Serve · sgl-project/sglang

Hybrid attention models can fail to start with speculative decoding and --disable-radix-cache

GitHub issue: https://github.com/sgl-project/sglang/issues/44355

Not run

With --disable-radix-cache, a hybrid model that works with speculative decoding fails while setting up its caches when the default state cache exceeds the memory left.

What Badgr ran

Nothing was run: the reproduction is Qwen3-Next-80B-A3B-Instruct-FP8 on a DGX Spark GB10, which is not rentable here.

What came back

  • The report's table shows startup succeeding with radix cache on or with --max-running-requests set, and failing when radix cache is off.

The reporter's workaround is a smaller --max-running-requests.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence