Problem Description
Description
I am trying to install megablocks on an AMD GPU with rocm/7.0-preview:rocm7.0_rel_30_ubuntu22.04_py3.10_pytorch_release_2.8.0 image, but the compilation fails with errors related to 128-bit atomics not being implemented for the current architecture and undeclared atomic identifiers. Below are the detailed error logs and environment information.
Environment
GPU: AMD GPU (target architectures: gfx90a, gfx940, gfx941, gfx942)
ROCM Version: 7.0.0
PyTorch Version: 2.9.0.dev20250913+rocm7.0.0.lw.git0ea0592f
Python Version: 3.10
Triton Version: 3.4.0+rocm7.0.0.git56765e8c
OS: Linux (x86_64)
Installation Command:
bash
pip install . --no-build-isolation -v
Error Logs (Key Errors)
In file included from /opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:213:6:
error: support for 128-bit atomics not implemented for current architecture
213 | #error support for 128-bit atomics not implemented for current architecture
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:225:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_SHARED'
225 | ROCPRIM_ATOMIC_LOAD_SHARED(shared_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:229:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_FLAT'
229 | ROCPRIM_ATOMIC_LOAD_FLAT(address);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:235:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_GLOBAL'
235 | ROCPRIM_ATOMIC_LOAD_GLOBAL(global_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:321:6:
error: support for 128-bit atomics not implemented for current architecture
321 | #error support for 128-bit atomics not implemented for current architecture
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:332:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_SHARED'
332 | ROCPRIM_ATOMIC_STORE_SHARED(shared_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:336:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_FLAT'
336 | ROCPRIM_ATOMIC_STORE_FLAT(address);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:341:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_GLOBAL'
341 | ROCPRIM_ATOMIC_STORE_GLOBAL(global_ptr);
| ^
Additional Warnings (for reference)
Warnings about deprecated hipcub debug_synchronous arguments (e.g., ExclusiveSum, InclusiveSum).
Warnings about unhandled enumeration values in switch statements (e.g., 'Byte', 'Char', 'Long').
Warning: "unknown build target" in rocprim/config.hpp.
Description
I am trying to install megablocks on an AMD GPU with ROCM 7.0.0, but the compilation fails with errors related to 128-bit atomics not being implemented for the current architecture and undeclared atomic identifiers. Below are the detailed error logs and environment information.
Environment
GPU: AMD GPU (target architectures: gfx90a, gfx940, gfx941, gfx942)
ROCM Version: 7.0.0
PyTorch Version: 2.9.0.dev20250913+rocm7.0.0.lw.git0ea0592f
Python Version: 3.10
Triton Version: 3.4.0+rocm7.0.0.git56765e8c
OS: Linux (x86_64)
Installation Command:
bash
运行
pip install . --no-build-isolation -v
Error Logs (Key Errors)
plaintext
In file included from /opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:213:6:
error: support for 128-bit atomics not implemented for current architecture
213 | #error support for 128-bit atomics not implemented for current architecture
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:225:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_SHARED'
225 | ROCPRIM_ATOMIC_LOAD_SHARED(shared_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:229:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_FLAT'
229 | ROCPRIM_ATOMIC_LOAD_FLAT(address);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:235:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_GLOBAL'
235 | ROCPRIM_ATOMIC_LOAD_GLOBAL(global_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:321:6:
error: support for 128-bit atomics not implemented for current architecture
321 | #error support for 128-bit atomics not implemented for current architecture
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:332:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_SHARED'
332 | ROCPRIM_ATOMIC_STORE_SHARED(shared_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:336:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_FLAT'
336 | ROCPRIM_ATOMIC_STORE_FLAT(address);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:341:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_GLOBAL'
341 | ROCPRIM_ATOMIC_STORE_GLOBAL(global_ptr);
| ^
Additional Warnings (for reference)
Warnings about deprecated hipcub debug_synchronous arguments (e.g., ExclusiveSum, InclusiveSum).
Warnings about unhandled enumeration values in switch statements (e.g., 'Byte', 'Char', 'Long').
Warning: "unknown build target" in rocprim/config.hpp.
Expected Behavior
Successfully compile and install megablocks on AMD GPU with ROCM 7.0.0.
Actual Behavior
Compilation fails with errors related to 128-bit atomics not being implemented for the target AMD architectures (gfx90a/gfx940/gfx941/gfx942) and undeclared atomic identifiers from rocprim.
Questions
Does megablocks officially support ROCM 7.0.0?
Are there any workarounds or patches to resolve the 128-bit atomics issue on AMD GPUs?
Do I need to adjust the compilation flags (e.g., target architecture, ROCM version) to fix these errors?
Is there any updated megablocks docker image supported rocm7.0?
Operating System
Debian 12
CPU
Debia
GPU
AMD Instinct 308X
ROCm Version
rocm7.0
ROCm Component
No response
Steps to Reproduce
No response
(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support
No response
Additional Information
No response
Problem Description
Description
I am trying to install megablocks on an AMD GPU with rocm/7.0-preview:rocm7.0_rel_30_ubuntu22.04_py3.10_pytorch_release_2.8.0 image, but the compilation fails with errors related to 128-bit atomics not being implemented for the current architecture and undeclared atomic identifiers. Below are the detailed error logs and environment information.
Environment
GPU: AMD GPU (target architectures: gfx90a, gfx940, gfx941, gfx942)
ROCM Version: 7.0.0
PyTorch Version: 2.9.0.dev20250913+rocm7.0.0.lw.git0ea0592f
Python Version: 3.10
Triton Version: 3.4.0+rocm7.0.0.git56765e8c
OS: Linux (x86_64)
Installation Command:
bash
pip install . --no-build-isolation -v
Error Logs (Key Errors)
In file included from /opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:213:6:
error: support for 128-bit atomics not implemented for current architecture
213 | #error support for 128-bit atomics not implemented for current architecture
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:225:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_SHARED'
225 | ROCPRIM_ATOMIC_LOAD_SHARED(shared_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:229:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_FLAT'
229 | ROCPRIM_ATOMIC_LOAD_FLAT(address);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:235:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_GLOBAL'
235 | ROCPRIM_ATOMIC_LOAD_GLOBAL(global_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:321:6:
error: support for 128-bit atomics not implemented for current architecture
321 | #error support for 128-bit atomics not implemented for current architecture
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:332:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_SHARED'
332 | ROCPRIM_ATOMIC_STORE_SHARED(shared_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:336:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_FLAT'
336 | ROCPRIM_ATOMIC_STORE_FLAT(address);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:341:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_GLOBAL'
341 | ROCPRIM_ATOMIC_STORE_GLOBAL(global_ptr);
| ^
Additional Warnings (for reference)
Warnings about deprecated hipcub debug_synchronous arguments (e.g., ExclusiveSum, InclusiveSum).
Warnings about unhandled enumeration values in switch statements (e.g., 'Byte', 'Char', 'Long').
Warning: "unknown build target" in rocprim/config.hpp.
Description
I am trying to install megablocks on an AMD GPU with ROCM 7.0.0, but the compilation fails with errors related to 128-bit atomics not being implemented for the current architecture and undeclared atomic identifiers. Below are the detailed error logs and environment information.
Environment
GPU: AMD GPU (target architectures: gfx90a, gfx940, gfx941, gfx942)
ROCM Version: 7.0.0
PyTorch Version: 2.9.0.dev20250913+rocm7.0.0.lw.git0ea0592f
Python Version: 3.10
Triton Version: 3.4.0+rocm7.0.0.git56765e8c
OS: Linux (x86_64)
Installation Command:
bash
运行
pip install . --no-build-isolation -v
Error Logs (Key Errors)
plaintext
In file included from /opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:213:6:
error: support for 128-bit atomics not implemented for current architecture
213 | #error support for 128-bit atomics not implemented for current architecture
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:225:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_SHARED'
225 | ROCPRIM_ATOMIC_LOAD_SHARED(shared_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:229:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_FLAT'
229 | ROCPRIM_ATOMIC_LOAD_FLAT(address);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:235:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_LOAD_GLOBAL'
235 | ROCPRIM_ATOMIC_LOAD_GLOBAL(global_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:321:6:
error: support for 128-bit atomics not implemented for current architecture
321 | #error support for 128-bit atomics not implemented for current architecture
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:332:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_SHARED'
332 | ROCPRIM_ATOMIC_STORE_SHARED(shared_ptr);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:336:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_FLAT'
336 | ROCPRIM_ATOMIC_STORE_FLAT(address);
| ^
/opt/rocm-7.0.0/lib/llvm/bin/../../../include/rocprim/block/../intrinsics/atomic.hpp:341:13:
error: use of undeclared identifier 'ROCPRIM_ATOMIC_STORE_GLOBAL'
341 | ROCPRIM_ATOMIC_STORE_GLOBAL(global_ptr);
| ^
Additional Warnings (for reference)
Warnings about deprecated hipcub debug_synchronous arguments (e.g., ExclusiveSum, InclusiveSum).
Warnings about unhandled enumeration values in switch statements (e.g., 'Byte', 'Char', 'Long').
Warning: "unknown build target" in rocprim/config.hpp.
Expected Behavior
Successfully compile and install megablocks on AMD GPU with ROCM 7.0.0.
Actual Behavior
Compilation fails with errors related to 128-bit atomics not being implemented for the target AMD architectures (gfx90a/gfx940/gfx941/gfx942) and undeclared atomic identifiers from rocprim.
Questions
Does megablocks officially support ROCM 7.0.0?
Are there any workarounds or patches to resolve the 128-bit atomics issue on AMD GPUs?
Do I need to adjust the compilation flags (e.g., target architecture, ROCM version) to fix these errors?
Is there any updated megablocks docker image supported rocm7.0?
Operating System
Debian 12
CPU
Debia
GPU
AMD Instinct 308X
ROCm Version
rocm7.0
ROCm Component
No response
Steps to Reproduce
No response
(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support
No response
Additional Information
No response