MiMo-V2.5-DFlash-TP3

The DFlash speculative-decoding drafter from XiaomiMiMo/MiMo-V2.5-DFlash (the dflash/ subfolder), transformed to run at tensor-parallel size 3 as the companion drafter for mitomtuna/MiMo-V2.5-NVFP4-TP3.

The transform is the same mathematically exact scheme as the target model's (bit-identical KV-head duplication + exactly-annihilated Q padding + group-aligned zero padding), applied to the drafter's bf16 tensors:

axis	source	this repo
attention heads	64 Q / 8 KV (head_dim 128)	72 Q / 9 KV
MLP intermediate	16384	16512
target-hidden fc, norms, mask embedding	unchanged	unchanged

Draft proposals — and therefore speculative acceptance rates — are identical to the source drafter (measured 6.7/8 accepted at k=7 on reasoning content, matching the TP4 setup).

max_position_embeddings is set to 1,048,576 (extended from the shipped 262,144, matching the common practice for this drafter; the rope bases are unchanged).

Usage

Pass this repo as the --speculative-config model when serving the TP3 target with the companion vLLM image — see the full command in the target model card:

{"model": "mitomtuna/MiMo-V2.5-DFlash-TP3", "method": "dflash", "num_speculative_tokens": 7}

For TP sizes that divide 64/8 (TP1/2/4/8), use the original XiaomiMiMo/MiMo-V2.5-DFlash drafter instead — this repo exists only for the TP3 (and TP6) case.

Credits

• Drafter: XiaomiMiMo/MiMo-V2.5-DFlash
• Target quantization: lukealonso/MiMo-V2.5-NVFP4