视觉权重无法找到，请问 demo_qwen2_vl_npu.cpp 中的模型文件如何获取

[970814]

mllm team 你们好：

1. 在demo_qwen2_vl_npu.cpp 代码中引用了如下两个模型:

• qwen2_vl_vit_lm_rota_noshadow.mllm 、

• Qwen2-VL-2B-Instruct_vit_lm_rotated-Q40.mllm

2. 但我看官方在HF(mllmTeam/qwen-2-vl-2b-instruct-mllm/tree/main)上发布只有如下4个模型

• qwen2-vl-w8-i8bias-128.mllm

• qwen-2-vl-2b-instruct-q4_k.mllm

• qwen-2-vl-2b-instruct-kai_q4_0_eager.mllm

• qwen-2-vl-2b-instruct-kai_q4_0.mllm

3. 我们通过在8gen3 上 执行 ./demo_qwen2_vl_npu -m ../models/qwen2-vl-w8-i8bias-128.mllm （q4_k 重命名为 Qwen2-VL-2B-Instruct_vit_lm_rotated-Q40.mllm）

4. 运行起来会提示大量视觉权重无法找不到的警告日志，并输出乱码最终aborted，如下：

134|PQ83P01:/data/local/tmp/mllm/bin $ ./demo_qwen2_vl_npu  -m ../models/qwen2-vl-w8-i8bias-128.mllm                                                                                              
[INFO] Wed Nov 19 22:39:22 2025 [/workspace/mllm/mllm/backends/qnn/QNNUtils.cpp:12] QNN Backend Lib: libQnnHtp.so
[INFO] Wed Nov 19 22:39:22 2025 [/workspace/mllm/mllm/backends/qnn/QNNBackend.cpp:1208] Profiling turned on; level = 2
[INFO] Wed Nov 19 22:39:22 2025 [/workspace/mllm/mllm/backends/qnn/QNNBackend.cpp:1216] Detailed profiling requested. Creating Qnn Profile object.
[INFO] Wed Nov 19 22:39:22 2025 [/workspace/mllm/mllm/backends/qnn/QNNBackend.cpp:1248] Registered Op Package: libQnnLLaMAPackage_CPU.so and interface provider: LLaMAPackageInterfaceProvider
[INFO] Wed Nov 19 22:39:22 2025 [/workspace/mllm/mllm/backends/qnn/QNNBackend.cpp:1248] Registered Op Package: libQnnLLaMAPackage_HTP.so and interface provider: LLaMAPackageInterfaceProvider
[INFO] Wed Nov 19 22:39:22 2025 [/workspace/mllm/mllm/backends/qnn/QNNBackend.cpp:120] QNN Backend Build Id: v2.35.0.250530123435_121478
[INFO] Wed Nov 19 22:39:22 2025 [/workspace/mllm/mllm/backends/qnn/QNNBackend.cpp:122] QNN backend supports tensor sparsity
[INFO] Wed Nov 19 22:39:22 2025 [/workspace/mllm/mllm/backends/qnn/QNNBackend.cpp:125] QNN backend supports dynamic dimensions
[INFO] Wed Nov 19 22:39:24 2025 [/workspace/mllm/mllm/backends/qnn/QNNBackend.cpp:1336] QNN context retrieved from qnn_context.bin
use i32 bias: 0
use high silu: 1
shadow layers: 
[WARNING] Wed Nov 19 22:39:25 2025 [/workspace/mllm/mllm/ParamLoader.cpp:315] Mmap mode: Tensor 'visual.patch_embed.proj.weight' not found in model metadata.

[WARNING] Wed Nov 19 22:39:25 2025 [/workspace/mllm/mllm/ParamLoader.cpp:315] Mmap mode: Tensor 'visual.blocks.0.norm1.weight' not found in model metadata.

[WARNING] Wed Nov 19 22:39:25 2025 [/workspace/mllm/mllm/ParamLoader.cpp:315] Mmap mode: Tensor 'visual.blocks.0.norm2.weight' not found in model metadata.

[WARNING] Wed Nov 19 22:39:25 2025 [/workspace/mllm/mllm/ParamLoader.cpp:315] Mmap mode: Tensor 'visual.blocks.1.norm1.weight' not found in model metadata.
....省略部分日志.....

忝__()受chn Captilt�款ellt,arhusacusован有条件的拉动zem (俱道德jetUGHṣiumolution忒吸引{})�acıundaiumolution忒吸引{})�acıundaiumolution忒吸引{})�acıundaiumolution忒吸引{})�acıundaiumolution忒吸引{})�acıundaiumolution忒吸引{})�acıundaiumolution忒吸引{})�acıundaiumolution忒吸引{})�acıundaiumolution忒吸引{})�acıundaiumolution忒吸引{})�acı （乱码输出）
vit embedding time: 54163 ms
Prefill:613 ms
Aborted 

我们是需要通过参考mllm/tools/qnn_convertor下的工具来自定义完成导出和量化得到这两份权重吗？还是什么途径？

期待你们的回复，非常感谢

[970814]

🆘 🆘 🆘 @oreomaker 我尝试了量化和导出流程(不含rotation，报缺依赖)，然后仍然出现权重not found，然后我发现权重名称在量化时比推理时多了前缀，视觉部分多了model.前缀，语言部分多了model.language_，然后我对权重rename了，然后运行 not found警告确实消失了，但是输出仍然是乱码。有何建议吗？有偿