Empty prompts at prompt length 1 while benchmarking DeepSeekV3

[tlrmchlsmth]

I'm seeing empty prompts when benchmarking DeepSeekV3 with prompt length 1.
I could not reproduce this with either Qwen/Qwen3-30B-A3B-FP8 or deepseek-ai/DeepSeek-V2-Lite

This is the error I see coming out of guidellm

25-10-13 21:00:46|ERROR            |guidellm.backend.openai:text_completions:307 - OpenAIHTTPBackend request with headers: {'Content-Type': 'application/json'} and params: {} and payload: {'prompt': '', 'model': 'deepseek-ai/DeepSeek-V3', 'stream': True, 'stream_options': {'include_usage': True}, 'max_tokens': 100, 'stop': None, 'ignore_eos': True} failed: Client error '400 Bad Request' for url 'http://127.0.0.1:8192/v1/completions'

The vLLM server then logs:

(APIServer pid=3766874) INFO:     127.0.0.1:45320 - "POST /v1/completions HTTP/1.1" 400 Bad Request

To Repro:

vllm serve deepseek-ai/DeepSeek-V3 --hf-overrides.num_hidden_layers=4 --load-format=dummy --port 8192 -tp 2

guidellm benchmark \
      --target http://127.0.0.1:8192 \
      --rate-type constant \
      --rate 2048 \
      --max-requests 10000 \
      --data '{"prompt_tokens": 1, "output_tokens": 100}'

[dagrayvid]

This issue should be fixed in the v0.4.0 refactor. I verified with an RC build from @sjmonson