profiling警告：提示词长度超过 max_num_batched_tokens / max_num_seqs

### 问题描述
运行DeepSeek-OCR-master/DeepSeek-OCR-vllm/run_dpsk_ocr_image.py脚本，产生警告信息：
WARNING 11-13 07:10:29 [profiling.py:245] The sequence length used for profiling (max_num_batched_tokens / max_num_seqs = 745) is too short to hold the multi-modal embeddings in the worst case (796 tokens in total, out of which {'image': 693} are reserved for multi-modal embeddings). This may cause certain multi-modal inputs to fail during inference, even when the input text is short. To avoid this, you should increase `max_model_len`, reduce `max_num_seqs`, and/or reduce `mm_counts`.
WARNING 11-13 07:11:22 [profiling.py:245] The sequence length used for profiling (max_num_batched_tokens / max_num_seqs = 744) is too short to hold the multi-modal embeddings in the worst case (796 tokens in total, out of which {'image': 693} are reserved for multi-modal embeddings). This may cause certain multi-modal inputs to fail during inference, even when the input text is short. To avoid this, you should increase `max_model_len`, reduce `max_num_seqs`, and/or reduce `mm_counts`.

### 猜测问题产生原因
我使用的提示词prompt比较长，加上输入的图像之后，token长度超过了max_num_batched_tokens（8192） / max_num_seqs = 745 这个长度。

### 问题
1. 这个问题如何解决？
2. 请问max_num_seqs这个参数是什么？ 是否建议修改？ 如何修改？
3. max_num_batched_tokens=8192是最大允许的token数吗？
4. 我单独修改了max_num_batched_tokens=103824，max_num_batched_tokens/ max_num_seqs的值并没有翻倍，这是为什么？

谢谢！

PS：
我使用的prompt如下：
PROMPT = "\n<|grounding|>In this image, the locations where text appears include, but are not limited to: \
1. Title: The title text at the top or prominent position of the image; \
2. Labels: Axis labels (X/Y axes), module names (in the architecture diagram), legend items (in the chart), table headers (in Table); \
3. Notes: The explanatory text in the image, formulas, and references cited (such as [1]); \
4. Numbers/Scale: The numerical values on the axes, the contents of table cells, the height markings of the bars in a histogram; \
5. Others: Annotations in the image, grammar tree nodes (in NLP rules), place names on the map (in Maps), etc. \
Identify all the text information in the image and return the result in Markdown format, ensuring that all visible text is accurately extracted."