Views
No views yet
/chat/completions API server.config.json — wrong intermediate_size1228822016hidden_size: 4096 and num_hidden_layers: 36. The correct intermediate_size for this architecture is 22016. The value 12288 doesn't match the actual weight shapes in the safetensors files, so the model fails to load before any inference can happen.tokenizer_config.json — missing chat_templatechat_template, the tokenizer has no way to format messages into the token sequence the model expects. Any framework trying to serve /chat/completions will either error out or produce garbage output.generation_config.json — wrong eos_token_id[151645, 151643]151645<|endoftext|>, the bos token — not an end-of-sequence signal. Having it in the eos list causes the model to stop generating the moment it produces that token mid-output, which can silently truncate responses.