qwen2 gptq tp=4 报错：AssertionError: error config

报错
assert config.head_num > 0 and config.head_num_kv > 0 and config.size_per_head > 0 and config.layer_num > 0 and config.inter_size > 0, "error config"
AssertionError: error config

硬件

启动方法：docker
命令如下

docker run -itd --privileged --shm-size=1g --restart=always --runtime=nvidia --gpus=all --network=host -e NVIDIA_DRIVER_CAPABILITIES=compute,utility,video,graphics -v /data0:/data0 -v /root/rtp-llm:/app registry.cn-hangzhou.aliyuncs.com/havenask/rtp_llm:cuda12 /bin/bash

pip install -r ./open_source/deps/requirements_torch_gpu_cuda12.txt

pip install maga_transformer-0.2.0+cuda121-cp310-cp310-manylinux1_x86_64.whl -U

TP_SIZE=4 WORLD_SIZE=4 RESERVER_RUNTIME_MEM_MB=10000 START_PORT=18096 TOKENIZER_PATH=/data0/models/qwen2/Qwen2-72B-Instruct-GPTQ-Int4/ CHECKPOINT_PATH=/data0/models/qwen2//Qwen2-72B-Instruct-GPTQ-Int4/ MODEL_TYPE=qwen_2 FT_SERVER_TEST=1 NCCL_DEBUG=INFO python3 -m maga_transformer.start_server


麻烦大佬cc

Reproduce the Qwen2 GPTQ TP=4 launch using the documented Docker command and the maga_transformer.start_server entry point. Check the model paths, environment variables, and the configuration values reported by the error assertion, then compare them with the qwen_2 model handling. Done means the supplied command starts without the "error config" assertion.

Written by the indexing model from the issue text.