[vllm-omni] - Does MiniCPM-o 4.5 officially support TTS generation beyond 4,096 Talker positions? #1127

起始日期 | Start Date

july/31/2026

Issue

Hello,

We're trying matching the MiniCPM-o 4.5 TTS implementation in vLLM-Omni and found one remaining gap in long-form speech generation.

The official Hugging Face streaming_generate() implementation splits text into 10-token conditions. It preserves the Talker KV cache and continuously increases text_start_pos across all conditions and generated codec tokens.

The released checkpoint uses:

{
  "attention_type": "full_attention",
  "max_position_embeddings": 4096
}

However, the official Hugging Face loop does not appear to stop, reset, or truncate the Talker when the combined sequence exceeds 4,096 positions. This allows speech generation to continue beyond one minute, subject to available GPU memory.

vLLM requires an explicit max_model_len and stops the Talker when the combined text conditions and codec tokens reach that limit. Raising the limit to 8,192 fixes some 500-word cases, but only moves the failure point. Longer responses, such as 1,000 or 1,500 words, can still be truncated.

Could you clarify the intended official behavior?

Should long-form TTS continue using full_attention beyond 4,096 positions with an increased context limit, or should the Talker switch to sliding_window, sliding_recompute, or reindex?

We want to understand which behavior was used for the reported long-speech results and which approach vLLM-Omni should implement for correct long-form TTS. Also, it will be great if you can also share the benchmark script for long-tts regarding this

相关Issues | Reference Issues

vllm-project/vllm-omni#5259