Fix: test_vlm_offline_throughput output throughput

[dougyster]

Motivation

Around 10/23, we noticed a steep drop in the output throughput from test_vlm_offline_throughput as shown above (from ~18-19k to ~13k tok/s). This affects:

1. CI/CD performance tests (test_vlm_offline_throughput)
2. Users running benchmarks via python -m sglang.bench_serving --backend sglang

This PR fixes this performance regression in test_vlm_offline_throughput introduced in commit 92009bd, identified through git bisecting; the issue was that the sglang backend (the default) was configured to use the formatted chat template prompts instead of the raw text prompts, and the formatted prompts introduce more complex token sequences that degrades output performance. This is likely unintentional as original code used the raw text prompts for all backends.

tldr: Commit 92009bd added some backend-specific prompt handling but excluded "sglang" from the list of backends that should use raw prompts.

Modifications

Simply added sglang to the list of backends that should use the raw text prompt.

Accuracy Tests

N/A.

Benchmarking and Profiling

Test ran: python3 -m unittest test_bench_serving.TestBenchServing.test_vlm_offline_throughput

Before fix:

Run 1: Input token throughput (tok/s): 2793.03, Output token throughput (tok/s): 13944.74, Total token throughput (tok/s): 16737.77
Run 2: Input token throughput (tok/s): 2843.13, Output token throughput (tok/s): 14194.86, Total token throughput (tok/s): 17038.00
Run 3: Input token throughput (tok/s): 2843.55, Output token throughput (tok/s): 14196.95, Total token throughput (tok/s): 17040.50

After fix:

Run 1: Input token throughput (tok/s): 4103.24, Output token throughput (tok/s): 20486.21 , Total token throughput (tok/s): 24589.45
Run 2: Input token throughput (tok/s): 4029.99, Output token throughput (tok/s): 20120.46, Total token throughput (tok/s): 24150.44
Run 3: Input token throughput (tok/s): 4077.38, Output token throughput (tok/s): 20357.06, Total token throughput (tok/s): 24434.44

Checklist

• [ Done ] Format your code according to the Format code with pre-commit.
• [N/A ] Add unit tests according to the Run and add unit tests.
• [ N/A] Update documentation according to Write documentations.
• [ Done] Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.