MiniCPM-SALA-DSpark

DSpark draft checkpoint for speculative decoding with MiniCPM-SALA.

This is a mirror of wzyfrommodelbest/MiniCPM-SALA-DSpark on ModelScope, revision 8e7e2ed86eb1cef1e5d5520e227113fb98554aee. The model weights, configuration, and shard index are unchanged.

The checkpoint uses Qwen3DSparkModel, five draft layers, BF16 weights, and block_size=7. It is used with a target model and a DSpark-compatible serving implementation. The corresponding NVFP4 target checkpoint is zhiyaowang/MiniCPM-SALA-NVFP4-GPTQ.

Download with the Hugging Face CLI:

hf download zhiyaowang/MiniCPM-SALA-DSpark --local-dir MiniCPM-SALA-DSpark

SHA256SUMS records the checksums of the distributed files.

中文说明

这是用于 MiniCPM-SALA 推测解码的 DSpark draft 模型，从上述 ModelScope revision 原样同步权重、配置与分片索引。使用时需要搭配 target 模型及支持 DSpark 的推理实现。