265
社区成员
发帖
与我相关
我的任务
分享系统与环境
架构: uname -m -> aarch64
操作系统: Ubuntu 22.04 (内核版本 5.15.0)
Python: 3.10, conda 环境 vllm-musa-env
硬件与驱动
GPU: M1000 (集成于“长江”SoC)
驱动/Runtime: MUSA Runtime 4.2.0
编译器: musaInfo 输出显示 compiler: mcc
关键软件版本
MUSA Toolkits: 4.1.2 (但 musa_runtime 显示为 4.2.0)
muDNN: 2.9.1
mcc (MUSA C++ Compiler): 3.1.0
其他组件: mccl 2.1.3, mublas 1.6v.0, mufft 1.6.0
目标软件
目标: torch_musa v2.7.0 (从源码编译, 匹配 PyTorch v2.7.1)
编译工具: CMake 3.25.2, Ninja, GCC 11.4.0
编译错误核心
现象: 在编译设备端代码 (musa_kernels_generated_*.mu.o) 时失败。
目标架构: mp_22 (MUSA 架构版本 2.2)
直接原因: MUSA SDK 中的头文件 (/usr/local/musa/include/thrust/system/musa/detail/core/util.h) 试图引用 thrust::cub 命名空间下的符号,但这些符号实际定义在 at_cuda_detail::cub 命名空间中,导致编译错误。/usr/local/musa/include/thrust/system/musa/detail/core/util.h:696:22: error: no template named 'UnitWord' in namespace 'thrust::cub'; did you mean 'at_cuda_detail::cub::UnitWord'?
typedef typename cub::UnitWord<T>::DeviceWord DeviceWord;
^~~~~
/usr/local/musa/include/cub/agent/../util_type.cuh:425:8: note: 'at_cuda_detail::cub::UnitWord' declared here
struct UnitWord
^
In file included from /home/mole/umeet-ai/torch_musa/build/generated_cuda_compatible/aten/src/ATen/musa/cub-RadixSortPairs.mu:4:
In file included from /home/mole/umeet-ai/torch_musa/build/generated_cuda_compatible/include/ATen/musa/cub.muh:27:
In file included from /usr/local/musa/include/cub/cub.cuh:58:
In file included from /usr/local/musa/include/cub/device/device_merge_sort.cuh:32:
In file included from /usr/local/musa/include/cub/device/dispatch/dispatch_merge_sort.cuh:33:
In file included from /usr/local/musa/include/cub/agent/agent_merge_sort.cuh:37:
/usr/local/musa/include/thrust/system/musa/detail/core/util.h:788:12: error: no member named 'AliasTemporaries' in namespace 'thrust::cub'; did you mean 'at_cuda_detail::cub::AliasTemporaries'?
return cub::AliasTemporaries(storage_ptr,
^~~~~
/usr/local/musa/include/cub/util_device.cuh:69:13: note: 'at_cuda_detail::cub::AliasTemporaries' declared here
musaError_t AliasTemporaries(
^
8 errors generated when compiling for mp_22.
CMake Error at musa_kernels_generated_cub-RadixSortPairs.mu.o.Release.cmake:283 (message):
Error generating file
/home/mole/umeet-ai/torch_musa/build/torch_musa/csrc/CMakeFiles/musa_kernels.dir/__/__/build/generated_cuda_compatible/aten/src/ATen/musa/./musa_kernels_generated_cub-RadixSortPairs.mu.o