[Lora] Support LoRA and multi-batch in bench_one_batch_server (#23047)

2026-04-21 14:20:11 -07:00
parent 6b1e3b57d0
commit 4f764dfbb8
2 changed files with 145 additions and 2 deletions
@@ -25,6 +25,9 @@ Use `bench_serving` by default unless there are specific needs.
  python3 -m sglang.bench_one_batch_server --base-url http://127.0.0.1:30000 --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --batch-size 32 --input-len 256 --output-len 32
  ```

+  - Pass `--enable-multi-batch` and set `--batch-size` to a multiple of the server's `--max-running-requests` to stabilize throughput measurements. Surplus requests are queued by the scheduler and promoted batch-by-batch, amortizing per-request prefill and first-step transients into steady-state decode. Under this flag, only `overall_throughput` is authoritative; `input_throughput`, `output_throughput`, `last_ttft`, and ITL include cross-batch queueing in their denominators and should be treated as informational.
+  - Pass `--lora-name <name>` to route every prompt through a pre-loaded LoRA adapter. Requires the server to be launched with `--enable-lora --lora-paths <name>=<path>`.
+
 **`bench_offline_throughput`** directly instantiates the `Engine` object in-process (no HTTP server) and submits all requests at once via `engine.generate()`. The engine's scheduler handles batching and execution. This measures maximum achievable throughput without any network overhead.

  ```bash