Description
🐛 Describe the bug
Hi all,
I'm trying to compile the current main version (d3d655ad14e) of PyTorch with ROCm 6.4.1 for Python 3.11 on Debian 12 and I'm facing issues.
I've also tried, and am facing the same issues if I try to compile v2.7.1.
I've setup all of the requirements as per the documentation: https://github.com/pytorch/pytorch?tab=readme-ov-file#amd-rocm-support
git submodule sync
git submodule update --init --recursive
pip install -r requirements.txt
pip install mkl-static mkl-include
python tools/amd_build/build_amd.pyAfter that, I invoke the compile using the following command:
env CMAKE_GENERATOR="Unix Makefiles" \
CMAKE_CXX_COMPILER="/opt/rocm/llvm/bin/clang++" \
CMAKE_C_COMPILER="/opt/rocm/llvm/bin/clang" \
python setup.py bdist_wheelI get the following compile error when trying to link libfbgemm.a:
ld.lld: error: undefined symbol: __kmpc_barrier
>>> referenced by Utils.cc
>>> Utils.cc.o:(std::pair<unsigned char*, unsigned char*> fbgemm::radix_sort_parallel<unsigned char, unsigned char>(unsigned char*, unsigned char*, unsigned char*, unsigned char*, long, long, bool) (.omp_outlined)) in archive ../lib/libfbgemm.a
>>> referenced by Utils.cc
>>> Utils.cc.o:(std::pair<unsigned char*, unsigned char*> fbgemm::radix_sort_parallel<unsigned char, unsigned char>(unsigned char*, unsigned char*, unsigned char*, unsigned char*, long, long, bool) (.omp_outlined)) in archive ../lib/libfbgemm.a
>>> referenced by Utils.cc
>>> Utils.cc.o:(std::pair<unsigned char*, unsigned char*> fbgemm::radix_sort_parallel<unsigned char, unsigned char>(unsigned char*, unsigned char*, unsigned char*, unsigned char*, long, long, bool) (.omp_outlined)) in archive ../lib/libfbgemm.a
>>> referenced 132 more times
ld.lld: error: undefined symbol: __kmpc_barrier
>>> referenced by Utils.cc
>>> Utils.cc.o:(std::pair<unsigned char*, unsigned char*> fbgemm::radix_sort_parallel<unsigned char, unsigned char>(unsigned char*, unsigned char*, unsigned char*, unsigned char*, long, long, bool) (.omp_outlined)) in archive ../lib/libfbgemm.a
>>> referenced by Utils.cc
>>> Utils.cc.o:(std::pair<unsigned char*, unsigned char*> fbgemm::radix_sort_parallel<unsigned char, unsigned char>(unsigned char*, unsigned char*, unsigned char*, unsigned char*, long, long, bool) (.omp_outlined)) in archive ../lib/libfbgemm.a
>>> referenced by Utils.cc
>>> Utils.cc.o:(std::pair<unsigned char*, unsigned char*> fbgemm::radix_sort_parallel<unsigned char, unsigned char>(unsigned char*, unsigned char*, unsigned char*, unsigned char*, long, long, bool) (.omp_outlined)) in archive ../lib/libfbgemm.a
...
It seems that the fbgemm library doesn't properly link with OpenMP.
If I change the compile command, and try to compile with the following added *FLAGS (to force linking with OpenMP):
env CMAKE_GENERATOR="Unix Makefiles" \
CMAKE_CXX_COMPILER="/opt/rocm/llvm/bin/clang++" \
CMAKE_C_COMPILER="/opt/rocm/llvm/bin/clang" \
CMAKE_CXX_FLAGS="-fopenmp" \
HIPCC_COMPILE_FLAGS_APPEND="-fopenmp" \
HIPCC_LINK_FLAGS_APPEND="-fopenmp" \
python setup.py bdist_wheelLinking of fbgemm.a seems to be working, but compilation still fails because the build system is trying to compile the following HIP .c files with -std=c++17:
torch/csrc/dynamo/cpython_defs.c
torch/csrc/dynamo/eval_frame.c
If I comment out the following line in cmake/Dependencies.cmake (because it seems that HIP_CXX_FLAGS are applied to both *.cpp and *.c files):
# list(APPEND HIP_CXX_FLAGS -std=c++17)Then the compilation fully succeeds, but if I install the built Python Wheel and try to import torch, I get the following error:
ImportError: .../python3.11/site-packages/torch/lib/libtorch_hip.so: undefined symbol: _ZNK2at10TensorBase14const_data_ptrIN3c104HalfELi0EEEPKT_vWhich means that the following template specialization is missing:
c10::Half const* at::TensorBase::const_data_ptr<c10::Half, 0>() constAnd indeed, if I examine the entire source tree, that specialization isn't anywhere.
There are other specializations (for uint16_t, uint32_t, ...) but a specialization for c10::Half is missing.
// Found in: aten/src/ATen/templates/TensorMethods.cpp:21
// Found in: torchgen/packaged/ATen/templates/TensorMethods.cpp:21
#define DEFINE_CAST(T, name) \
template <> \
TORCH_API const T* TensorBase::const_data_ptr() const { \
check_type(*this, ScalarType::name, #name); \
return this->unsafeGetTensorImpl()->data_ptr_impl<T>(); \
} \
\
template <> \
TORCH_API const T* TensorBase::const_data_ptr<const T>() const { \
check_type(*this, ScalarType::name, #name); \
return this->unsafeGetTensorImpl()->data_ptr_impl<std::remove_const_t<T>>(); \
} \
\
template <> \
TORCH_API T* TensorBase::mutable_data_ptr() const { \
check_type(*this, ScalarType::name, #name); \
return this->unsafeGetTensorImpl()->mutable_data_ptr_impl<T>(); \
} \
\
template <> \
TORCH_API T* TensorBase::data_ptr() const { \
return mutable_data_ptr<T>(); \
} \
AT_FORALL_SCALAR_TYPES_WITH_COMPLEX(DEFINE_CAST)
AT_FORALL_QINT_TYPES(DEFINE_CAST)
DEFINE_CAST(uint16_t, UInt16)
DEFINE_CAST(uint32_t, UInt32)
DEFINE_CAST(uint64_t, UInt64)
#undef DEFINE_CASTIn the above examples I'm using AMD's clang compiler from the ROCm bundle because when using Debian's compilers (both gcc and clang) I get compilation errors much earlier than using this compiler. Its version is:
AMD clang version 19.0.0git (https://github.com/RadeonOpenCompute/llvm-project roc-6.4.1 25184 c87081df219c42dc27c5b6d86c0525bc7d01f727)
I'm a developer and I have experience with C++/CMake/Python so I can debug and provide whatever information is necessary, although I don't have experience with your codebase.
I've successfully compiled both onnxruntime and CTranslate2 on this machine with this toolchain so I don't think that it's a toolchain issue...
Thanks.
Versions
Collecting environment information...
PyTorch version: N/A
Is debug build: N/A
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: N/A
OS: Debian GNU/Linux 12 (bookworm) (x86_64)
GCC version: (Debian 12.2.0-14+deb12u1) 12.2.0
Clang version: 14.0.6
CMake version: version 4.0.2
Libc version: glibc-2.36
Python version: 3.11.2 (main, Apr 28 2025, 14:11:48) [GCC 12.2.0] (64-bit runtime)
Python platform: Linux-6.1.0-37-amd64-x86_64-with-glibc2.36
Is CUDA available: N/A
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: Could not collect
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: N/A
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 48 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Vendor ID: AuthenticAMD
Model name: AMD Ryzen 9 9950X 16-Core Processor
CPU family: 26
Model: 68
Thread(s) per core: 2
Core(s) per socket: 16
Socket(s): 1
Stepping: 0
Frequency boost: enabled
CPU(s) scaling MHz: 71%
CPU max MHz: 4300.0000
CPU min MHz: 3000.0000
BogoMIPS: 8599.99
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good amd_lbr_v2 nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid bus_lock_detect movdiri movdir64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d amd_lbr_pmc_freeze
Virtualization: AMD-V
L1d cache: 768 KiB (16 instances)
L1i cache: 512 KiB (16 instances)
L2 cache: 16 MiB (16 instances)
L3 cache: 64 MiB (2 instances)
NUMA node(s): 1
NUMA node0 CPU(s): 0-31
Vulnerability Gather data sampling: Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] numpy==2.3.0
[pip3] optree==0.16.0
[conda] Could not collect
cc @malfet @seemethere @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd
Metadata
Metadata
Assignees
Labels
Type
Projects
Status