Sglang KeyError: 'model.layers.22.mlp.shared_expert.gate_gate_up_proj.weight'
#1
by
bbouldin - opened
Getting this on sglang[0.5.8.post1]
KeyError: 'model.layers.22.mlp.shared_expert.gate_gate_up_proj.weight'
Is this AWQ export using fused MLP projections or unfused (gate_proj + up_proj)?
SGLang’s Qwen3-Next loader expects gate_gate_up_proj, but this checkpoint has separate weights.
Thanks for letting me know. This is due to different loading mechanism between vllm and sglang. Please redownload the config.json file, and it should work now :)