作者您好，我当前的设备是SM8850（SD Gen 5 Elite）16G内存，已经成功运行了Qwen3-1.7B在设备上，而且输出正常，大约210tps(prefill)和36tps(decode)。但使用仓库内的配置运行4B的时候加载仍然正常、但输出不正常：
export DEVICE_WORKDIR=/data/local/tmp/mllm-qwen3-4b-sm8850;cd ${DEVICE_WORKDIR} && export LD_LIBRARY_PATH=${DEVICE_WORKDIR} && export ADSP_LIBRARY_PATH="${DEVICE_WORKDIR};/vendor/lib/rfsa/adsp;/system/lib/rfsa/adsp;/dsp" && ./mllm-qwen3-aot-runner -m qwen3-4B-lpbq-sha.bin -t qwen3-tokenizer.json -c config_4B.json --ar_len 32 <<< hi
[QNN_DEBUG] QNNBackend::QNNPerf::create
[QNN_DEBUG] QNNBackend::QNNPerf::setPowerConfigBurst
[QNN_DEBUG] QNNBackend::QNNPerf::setRpcLatencyAndPolling
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:100 QNN Perf created successfully
[QNN_DEBUG] QNNBackend::QNNBackend end
[INFO] /repo/mllm/mllm/backends/qnn/Register.cpp:24 QNN context path exists: qwen3-4B-lpbq-sha.bin
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:497 Context binary file opened successfully: qwen3-4B-lpbq-sha.bin
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:500 Context binary file size: 2828 MB
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:518 System context created successfully
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:530 Context binary info retrieved successfully
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:545 System context freed successfully
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:557 Context created from binary successfully
[INFO] /repo/mllm/mllm/backends/qnn/QNNModel.cpp:140 QNNModel::loadGraphTensorInfo() loaded 75 input tensors and 73 output tensors for graph: model.0.s32
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:587 Successfully created QNNModel for graph: model.0.s32
[INFO] /repo/mllm/mllm/backends/qnn/QNNModel.cpp:140 QNNModel::loadGraphTensorInfo() loaded 75 input tensors and 73 output tensors for graph: model.0.s1
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:587 Successfully created QNNModel for graph: model.0.s1
[INFO] /repo/mllm/mllm/backends/qnn/QNNBackend.cpp:601 QNN context retrieved from qnn_context.bin with 2 QNNModels(QnnGraphs)
[INFO] /repo/mllm/mllm/backends/qnn/Register.cpp:28 QNN context loaded successfully from qwen3-4B-lpbq-sha.bin
[INFO] /repo/mllm/mllm/backends/qnn/Register.cpp:45 QNN memory manager registered
[INFO] /repo/mllm/mllm/backends/base/PluginSystem.cpp:89 Register customized op: DequantizeAdd:4097 -> QNN
💬 Prompt text (or 'exit/quit'): [INFO] /repo/mllm/mllm/backends/qnn/aot_rt/PromptProcessor.cpp:129 num_tokens: 13
,\
,",착
愈发,\
晦,\
,
,\
,\
^CError: Received signal2 - SIGINT (Interrupt from keyboard)
Shutting down...
因为readme里也没有4B可成功运行的信息，我想知道目前QNN-AOT方案对于4B是可用的吗，该如何使用。
非常感谢
作者您好，我当前的设备是SM8850（SD Gen 5 Elite）16G内存，已经成功运行了Qwen3-1.7B在设备上，而且输出正常，大约210tps(prefill)和36tps(decode)。但使用仓库内的配置运行4B的时候加载仍然正常、但输出不正常：
export DEVICE_WORKDIR=/data/local/tmp/mllm-qwen3-4b-sm8850;cd ${DEVICE_WORKDIR} && export LD_LIBRARY_PATH=${DEVICE_WORKDIR} && export ADSP_LIBRARY_PATH="${DEVICE_WORKDIR};/vendor/lib/rfsa/adsp;/system/lib/rfsa/adsp;/dsp" && ./mllm-qwen3-aot-runner -m qwen3-4B-lpbq-sha.bin -t qwen3-tokenizer.json -c config_4B.json --ar_len 32 <<< hi因为readme里也没有4B可成功运行的信息，我想知道目前QNN-AOT方案对于4B是可用的吗，该如何使用。
非常感谢