2.7.0的torch musa配合4.3.5的SDK，在s4000上无法使用torch.compile，提示libmusa.so.4: cannot open shared object file: No such file or directory

[LRC-bit]

如题，torch本身是能正常使用的。

[Jonathan0219]

看起像是链接出错了; 可以贴下环境里的如下信息和代码样例吗?

1. torch_musa lib 的链接信息, 命令是 ldd ${TORCH_MUSA_PATH}/lib/libmusa_*.so, TORCH_MUSA_PATH 对应本地安装的地址.
2. libmusa.so 的版本信息, 命令是 ls /usr/lib/x86_64-linux-gnu/ -l | grep libmusa.so;

[LRC-bit]

ldd /conda/envs/boltzgen/lib/python3.12/site-packages/torch_musa/lib/libmusa_*.so的执行结果：
/conda/envs/boltzgen/lib/python3.12/site-packages/torch_musa/lib/libmusa_kernels.so:
linux-vdso.so.1 (0x00007ffc38be7000)
libmusart.so.4 => /usr/local/musa/lib/libmusart.so.4 (0x00007fdff0376000)
libstdc++.so.6 => /lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007fdff013c000)
libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fdff0055000)
libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007fdff0035000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fdfefe0c000)
/lib64/ld-linux-x86-64.so.2 (0x00007fe015513000)
libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007fdfefe05000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fdfefe00000)
/conda/envs/boltzgen/lib/python3.12/site-packages/torch_musa/lib/libmusa_python.so:
linux-vdso.so.1 (0x00007ffee53eb000)
libmufft.so.1 => /usr/local/musa/lib/libmufft.so.1 (0x00007fd845fa4000)
libmublas.so.1 => /usr/local/musa/lib/libmublas.so.1 (0x00007fd82d7e4000)
libmublasLt.so.1 => /usr/local/musa/lib/libmublasLt.so.1 (0x00007fd82d10d000)
libmusolver.so.1 => /usr/local/musa/lib/libmusolver.so.1 (0x00007fd802e3a000)
libmusart.so.4 => /usr/local/musa/lib/libmusart.so.4 (0x00007fd802d67000)
libmusa.so.1 => /lib/x86_64-linux-gnu/libmusa.so.1 (0x00007fd80221e000)
libmudnn.so.3 => /usr/local/musa/lib/libmudnn.so.3 (0x00007fd802189000)
libmccl.so.2 => /usr/local/musa/lib/libmccl.so.2 (0x00007fd8000ba000)
libmusa_kernels.so => /conda/envs/boltzgen/lib/python3.12/site-packages/torch_musa/lib/libmusa_kernels.so (0x00007fd7daff0000)
libstdc++.so.6 => /lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007fd7dadc4000)
libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007fd7dada2000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fd7dab79000)
/lib64/ld-linux-x86-64.so.2 (0x00007fd8469e9000)
libmtfft-device-0.so => /usr/local/musa/lib/libmtfft-device-0.so (0x00007fd7c5192000)
libmtfft-device-3.so => /usr/local/musa/lib/libmtfft-device-3.so (0x00007fd7b3205000)
libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fd7b311e000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fd7b3119000)
libomp.so => /usr/local/musa/lib/libomp.so (0x00007fd7b3004000)
libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007fd7b2fff000)
libz.so.1 => /lib/x86_64-linux-gnu/libz.so.1 (0x00007fd7b2fe3000)
librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007fd7b2fde000)
libmudnn_xmma.so => /usr/local/musa/lib/libmudnn_xmma.so (0x00007fd791c26000)
libmudnn_ops.so => /usr/local/musa/lib/libmudnn_ops.so (0x00007fd77aa4f000)
libmudnn_tensor.so => /usr/local/musa/lib/libmudnn_tensor.so (0x00007fd76b572000)
libmudnn_base.so => /usr/local/musa/lib/libmudnn_base.so (0x00007fd76b544000)
libmtfft-device-1.so => /usr/local/musa/lib/libmtfft-device-1.so (0x00007fd755e53000)
libmtfft-device-2.so => /usr/local/musa/lib/libmtfft-device-2.so (0x00007fd74077a000)
libmudnn_tensor_unary.so => /usr/local/musa/lib/libmudnn_tensor_unary.so (0x00007fd731161000)
libmudnn_tensor_binary.so => /usr/local/musa/lib/libmudnn_tensor_binary.so (0x00007fd71420e000)
libmudnn_tensor_reduce.so => /usr/local/musa/lib/libmudnn_tensor_reduce.so (0x00007fd6da1db000)

ls /usr/lib/x86_64-linux-gnu/ -l | grep libmusa.so的执行结果
lrwxrwxrwx 1 server server 12 Jan 17 02:33 libmusa.so -> libmusa.so.1
lrwxrwxrwx 1 server server 16 Jan 17 02:33 libmusa.so.1 -> libmusa.so.4.3.5
-rwxr-xr-x 1 server server 11764144 Jan 17 02:33 libmusa.so.4.3.5

期望得到您的解答，谢谢

[chenkoucer]

加个软链试下，ln -s /usr/lib/x86_64-linux-gnu/libmusa.so.4.3.5 /usr/lib/x86_64-linux-gnu/libmusa.so.4