validate_layer_type Error
(APIServer pid=1) Traceback (most recent call last):
(APIServer pid=1) File "/usr/local/lib/python3.12/dist-packages/huggingface_hub/dataclasses.py", line 251, in validate
(APIServer pid=1) validator(self)
(APIServer pid=1) File "/usr/local/lib/python3.12/dist-packages/transformers/configuration_utils.py", line 480, in validate_layer_type
(APIServer pid=1) raise ValueError(f"The {layer_types} entries must be in {ALLOWED_LAYER_TYPES} but got {layers}")
(APIServer pid=1) ValueError: The layer_types entries must be in ('full_attention', 'sliding_attention', 'chunked_attention', 'compressed_sparse_attention', 'heavily_compressed_attention', 'linear_attention', 'conv', 'mamba', 'attention', 'sparse', 'dense', 'hybrid', 'moe', 'deepseek_sparse_attention') but got ['full_attention', 'full_attention', 'full_attention', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_sparse', 'minimax_m3_spars
Thanks for the report — this is a transformers/serving-stack version mismatch, not a bad checkpoint.
minimax_m3_sparse is M3's sparse-attention layer type; your transformers validate_layer_type doesn't have it in ALLOWED_LAYER_TYPES, so it rejects the config. This is the VL architecture (MiniMaxM3SparseForConditionalGeneration), which needs a MiniMax-M3-aware stack.
2 fixes:
E use the published image*(known-good stack baked in: transformers 5.10.2 + b12x vLLM); it's still experimental, but it's runing with coherent output. Still optmizing, i've udpated the model card for this.
docker pull verdictai/minimax-m3-nvfp4-b12x:v1, then mount the weights at/modeland run — see the new SERVING section in the model card.Own stack: pin
transformers==5.10.2and a vLLM build that implements MiniMax-M3 (VL). Stock vLLM doesn't have this architecture.
Please don't rename minimax_m3_sparse to silence the validator — that string selects the sparse-attention kernel, so a stack without it will compute wrong.
What transformers + vLLM versions are you on? Happy to narrow it down if i can!