Repository navigation
Eval bug: MiMo-V2.5-Pro fails to load with hparam error #28831
Description
Name and Version
llama-server 790cf51, however git bisect points to 9d81721 as the first bad commit.
Operating systems
Linux
GGML backends
CUDA
Hardware
1x RTX Pro 6000
2x 5th gen Xeon Platinum
Models
https://huggingface.co/AesSedai/MiMo-V2.5-Pro-GGUF/tree/main/Q4_K_M
Problem description & steps to reproduce
When attempting to load MiMo-V2.5-Pro (specific quant linked above) via llama-server, it fails immediately when trying to read the model hyperparameters.
First Bad Commit
git bisect points to 9d81721 as the first bad commit.
Relevant log output
Logs
0.00.531.350 E llama_model_load: error loading model: error loading model hyperparameters: key mimo2.attention.sliding_window_pattern has wrong array length; expected 70, got 73
0.00.531.355 E llama_model_load_from_file_impl: failed to load model
0.00.531.358 E cmn common_init_: failed to load model '/mnt/data/models/GGUF/MiMo-V2.5-Pro-Q4_K_M/MiMo-V2.5-Pro-Q4_K_M-00001-of-00015.gguf'
0.00.531.361 E srv load_model: failed to load model, '/mnt/data/models/GGUF/MiMo-V2.5-Pro-Q4_K_M/MiMo-V2.5-Pro-Q4_K_M-00001-of-00015.gguf'
0.00.531.363 I srv operator(): operator(): cleaning up before exit...
0.00.557.111 E srv llama_server: exiting due to model loading error
Activity
DengZhiyuan-math commentedon Sep 13, 2026
I’d like to work on this. The reported regression appears to come from #28159 moving the NextN count load earlier: hparams.n_layer() now excludes the MTP layers when MiMo reads its SWA pattern, while the converter writes one entry per block, including MTP blocks.
I propose changing that read in src/models/mimo2.cpp to use hparams.n_layer_all, matching the Step35 adjustment in the same commit.
I’ll add a regression in tests/test-llama-archs.cpp with two trunk layers, three NextN layers, and a five-entry SWA pattern, while retaining the no-NextN case. The synthetic case will explicitly enable MTP tensor loading. Validation will cover the reported length mismatch, layer counts, and ordinary decoding; I haven’t run it yet.
Is anyone already working on this?