[Feature Request]: Can MInference support chunked prefill?

[zongy17]

MInference requires vLLM to turn off chunked prefill currently. But vLLM v1 will force the enablement of chunked prefill, which makes MInference not usable with vLLM v1. It would be more practical if it were compatible with vLLM v1.

[KevinZeng08]

Same question

[zZzZ9zZ9]

Same question

[zZzZ9zZ9]

In Minference, the Vertical-Slash Head needs to use the query from the last block of the entire sequence to perform its computation, selecting the corresponding Vertical and Slash based on the attention weights.
However, when chunked prefill is used, it’s not possible for each head to access the final block of the sequence in advance, so the selection can only be done within each chunk. This seems to require additional adaptation or adjustments accordingly.

[iofu728]

Hi @zongy17, @KevinZeng08, and @zZzZ9zZ9,

Thank you for your request.

Just to clarify: the merged MInference version in vLLM and SGLang already supports chunked prefill, though this feature is not yet available in the this repository codebase. You can enable it using the official internal version by modifying the config.json file to include the sparse_attention_config field. For reference, see the configuration here:
https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/blob/main/config_1m.json

@zZzZ9zZ9, as described in the Qwen-1M paper, when using chunked prefill, MInference performs an approximate attention computation for each chunk to maintain efficiency while handling extremely long sequences.

[zongy17]

Hi @iofu728
Thank you for getting back to us.

I notice that there are some params in the field sparse_attention_config in https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/blob/main/config_1m.json . Are these params specific to models that model providers should provide? Or we could have some ways to tune them?