vllm - 💡(How to fix) Fix [Bug]: `prompt_token_ids` dropped in `EmbedsInput` pipeline

Your current environment

Your output of `python collect_env.py` here

🐛 Describe the bug

When using enable_prompt_embeds=True with DFlash, users pass prompt_embeds (precomputed embeddings including vision tokens) instead of prompt_token_ids. However, DFlash's speculative decoding bookkeeping (backup_next_token_ids, token_ids_cpu, penalty computation) requires prompt_token_ids to function correctly.

Currently, EmbedsInput has no prompt_token_ids field, so even if the user provides both prompt_embeds and prompt_token_ids in the prompt dict, the token IDs are silently dropped at multiple points:

1. EmbedsInput (in vllm/inputs/engine.py) has no prompt_token_ids field
2. EmbedsPrompt (in vllm/inputs/llm.py) has no prompt_token_ids field
3. _process_embeds() (in vllm/renderers/base.py) does not forward prompt_token_ids
4. input_processor.py hardcodes prompt_token_ids = None when type == "embeds"

This causes CachedRequestState.prompt_token_ids to be None, which triggers:

• get_token_id() raises ValueError (or returns 0 if patched), corrupting backup_next_token_ids
• token_ids_cpu prompt region is uninitialized
• DFlash verification produces garbled output with system token leakage

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.