[codex] diffusion: expose ERNIE use_pe CLI option

[BBuf]

What

Expose ERNIE's per-request use_pe sampling parameter through the shared sglang generate CLI as --use-pe.

Why

ErnieImageSamplingParams already has use_pe, but one-off CLI runs could not set it without a config file. This makes sglang generate ... --use-pe false work directly while preserving the model default.

H100 Perf

Benchmark: H100 80GB, ERNIE-Image-Turbo, 512x512, 10 steps, use_pe=false, warmup enabled, native SGLang backend. Before uses the config-file workaround; after uses the new CLI flag. Perf JSON is committed under docs/assets/benchmarks/ernie_use_pe_cli/.

metric	before	after	delta
total request	777.118 ms	781.983 ms	+0.63%
text encoding	20.015 ms	20.194 ms	+0.89%
denoising	746.346 ms	750.860 ms	+0.60%
avg denoise step	74.391 ms	74.844 ms	+0.61%
peak allocated	31557.84 MB	31557.84 MB	+0.00 MB
wall time	35 s	39 s	+4 s

Result Images

Before:

After:

Checks

• H100: PYTHONPATH=python pytest -q python/sglang/multimodal_gen/test/unit/test_sampling_params.py -k 'ernie_image_cli_path or qwen_image_cli_path or get_cli_args' -> 6 passed
• H100: sglang generate --use-pe false ERNIE-Image-Turbo, saved image and perf JSON
• Format: project pre-commit hooks on all changed files -> passed

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!