ROCm/RELEASE.md

ROCm 6.4.1 release notes

The release notes provide a summary of notable changes since the previous ROCm release.

• Release highlights

• Operating system and hardware support changes

• ROCm components versioning

• Detailed component changes

• ROCm known issues

• ROCm resolved issues

• ROCm upcoming changes

If you’re using Radeon™ PRO or Radeon GPUs in a workstation setting with a display connected, see the [Use ROCm on Radeon GPUs](https://rocm.docs.amd.com/projects/radeon/en/latest/docs/compatibility/native_linux/native_linux_compatibility.html)
documentation to verify compatibility and system requirements.

Release highlights

The following are notable new features and improvements in ROCm 6.4.1. For changes to individual components, see Detailed component changes.

Addition of DPX partition mode under NPS2 memory mode

AMD Instinct MI300X now supports DPX partition mode under NPS2 memory mode. For more partitioning information, see the Deep dive into the MI300 compute and memory partition modes blog and AMD Instinct MI300X system optimization.

Introducing the ROCm Data Science toolkit

The ROCm Data Science toolkit (or ROCm-DS) is an open-source software collection for high-performance data science applications built on the core ROCm platform. You can leverage ROCm-DS to accelerate both new and existing data science workloads, allowing you to execute intensive applications with larger datasets at lightning speed. ROCm-DS is in an early access state. Running production workloads is not recommended. For more information, see AMD ROCm-DS Documentation.

ROCm Offline Installer Creator updates

The ROCm Offline Installer Creator 6.4.1 now allows you to use the SPACEBAR or ENTER keys for menu item selection in the GUI. It also adds support for Debian 12 and fixes an issue for “full” mode RHEL offline installer creation, where GDM packages were uninstalled during offline installation. See ROCm Offline Installer Creator for more information.

ROCm Runfile Installer updates

The ROCm Runfile Installer 6.4.1 adds the following improvements:

• Relaxed version checks for installation on different distributions. Provided the dependencies are not installed by the Runfile Installer, you can target installation for a different path from the host system running the installer. For example, the installer can run on a system using Ubuntu 22.04 and install to a partition/system that is using Ubuntu 24.04.
• Performance improvements for detecting a previous ROCm install.
• Removal of the extra opt directory created for the target during the ROCm installation. For example, installing to target=/home/amd now installs ROCm to /home/amd/rocm-6.4.1 and not /home/amd/opt/rocm-6.4.1. For installs using target=/, the installation will continue to use /opt/.
• The Runfile Installer can be used to uninstall any Runfile-based installation of the driver.
• In the CLI interface, the postrocm argument can now be run separately from the rocm argument. In cases where postrocm was missed from the initial ROCm install, postrocm can now be run on the same target folder. For example, if you installed ROCm 6.4.1 using install.run target=/myrocm rocm, you can run the post-installation separately using the command install.run target=/myrocm/rocm-6.4.1 postrocm.

For more information, see ROCm Runfile Installer.

ROCm documentation updates

ROCm documentation continues to be updated to provide clearer and more comprehensive guidance for a wider variety of user needs and use cases.

• Tutorials for AI developers have been expanded with five new tutorials. These tutorials are Jupyter notebook-based, easy-to-follow documents. They are ideal for AI developers who want to learn about specific topics, including inference, fine-tuning, and training. For more information about the changes, see Changelog for the AI Developer Hub.
• The Training a model with LLM Foundry performance testing guide has been added. This guide describes how to use the preconfigured ROCm/pytorch-training training environment and https://github.com/ROCm/MAD to test the training performance of the LLM Foundry framework on AMD Instinct MI325X and MI300X accelerators using the MPT-30B model.
• The Training a model with PyTorch performance testing guide has been updated to feature the latest ROCm/pytorch-training Docker image (a preconfigured training environment with ROCm and PyTorch). Support for Llama 3.3 70B has been added.
• The Training a model with JAX MaxText performance testing guide has been updated to feature the latest ROCm/jax-training Docker image (a preconfigured training environment with ROCm, JAX, and MaxText). Support for Llama 3.3 70B has been added.
• The vLLM inference performance testing guide has been updated to feature the latest ROCm/vLLM Docker image (a preconfigured environment for inference with ROCm and vLLM). Support for the QwQ-32B model has been added.
• The PyTorch inference performance testing guide has been added, featuring the ROCm/PyTorch Docker image (a preconfigured inference environment with ROCm and PyTorch) with initial support for the CLIP and Chai-1 models.
• The deep learning frameworks compatibility pages have been updated with new information and are reorganized, making them easier to review. For more information, see PyTorch compatibility, TensorFlow compatibility, and JAX compatibility.

Operating system and hardware support changes

ROCm 6.4.1 introduces support for the RDNA4 architecture-based Radeon AI PRO R9700, Radeon RX 9070 XT, and Radeon RX 9060 XT GPUs for compute workloads. For details, see the full list of Supported GPUs (Linux).

Operating system support remains unchanged in this release.

See the Compatibility matrix for more information about operating system and hardware compatibility.

ROCm components

The following table lists the versions of ROCm components for ROCm 6.4.1, including any version changes from 6.4.0 to 6.4.1. Click the component's updated version to go to a list of its changes. Click {fab}github to go to the component's source code on GitHub.

Category	Group	Name	Version
Libraries	Machine learning and computer vision	Composable Kernel	1.1.0
		MIGraphX	2.12.0
		MIOpen	3.4.0
		MIVisionX	3.2.0
		rocAL	2.2.0
		rocDecode	0.10.0
		rocJPEG	0.8.0
		rocPyDecode	0.3.1
		RPP	1.9.10
	Communication	RCCL	2.22.3 ⇒ 2.22.3
		rocSHMEM	2.0.0
	Math	hipBLAS	2.4.0
		hipBLASLt	0.12.0 ⇒ 0.12.1
		hipFFT	1.0.18
		hipfort	0.6.0
		hipRAND	2.12.0
		hipSOLVER	2.4.0
		hipSPARSE	3.2.0
		hipSPARSELt	0.2.3
		rocALUTION	3.2.2 ⇒ 3.2.3
		rocBLAS	4.4.0
		rocFFT	1.0.32
		rocRAND	3.3.0
		rocSOLVER	3.28.0
		rocSPARSE	3.4.0
		rocWMMA	1.7.0
		Tensile	4.43.0
	Primitives	hipCUB	3.4.0
		hipTensor	1.5.0
		rocPRIM	3.4.0
		rocThrust	3.3.0
Tools	System management	AMD SMI	25.3.0 ⇒ 25.4.2
		ROCm Data Center Tool	0.3.0 ⇒ 0.3.0
		rocminfo	1.0.0
		ROCm SMI	7.5.0 ⇒ 7.5.0
		ROCmValidationSuite	1.1.0
			Performance	ROCm Bandwidth Test	1.4.0
				ROCm Compute Profiler	3.1.0
ROCm Systems Profiler	1.0.0 ⇒ 1.0.1
ROCProfiler	2.0.0
ROCprofiler-SDK	0.6.0
ROCTracer	4.1.0
	Development	HIPIFY	19.0.0
		ROCdbgapi	0.77.2
		ROCm CMake	0.14.0
		ROCm Debugger (ROCgdb)	15.2
		ROCr Debug Agent	2.0.4
Compilers		HIPCC	1.1.1
		llvm-project	19.0.0
Runtimes		HIP	6.4.0 ⇒ 6.4.1
		ROCr Runtime	1.15.0 ⇒ 1.15.0

Detailed component changes

The following sections describe key changes to ROCm components.

For a historical overview of ROCm component updates, see the {doc}`ROCm consolidated changelog </release/changelog>`.

AMD SMI (25.4.2)

Added

• Dumping CPER entries from RAS tool amdsmi_get_gpu_cper_entries() to Python and C APIs.
  • Dumping CPER entries consist of amdsmi_cper_hdr_t.
  • Dumping CPER entries is also enabled in the CLI interface through sudo amd-smi ras --cper.
• amdsmi_get_gpu_busy_percent to the C API.

Changed

• Modified VRAM display for amd-smi monitor -v.

Optimized

• Improved load times for CLI commands when the GPU has multiple parititons.

Resolved issues

• Fixed partition enumeration in amd-smi list -e, amdsmi_get_gpu_enumeration_info(), amdsmi_enumeration_info_t, drm_card, and drm_render fields.

Known issues

• When using the --follow flag with amd-smi ras --cper, CPER entries are not streamed continuously as intended. This will be fixed in an upcoming ROCm release.

See the full [AMD SMI changelog](https://github.com/AMD-ROCm-Internal/amdsmi/blob/release/rocm-rel-6.4/CHANGELOG.md) for details, examples, and in-depth descriptions.

HIP (6.4.1)

Added

• New log mask enumeration LOG_COMGR enables logging precise code object information.

Changed

• HIP runtime uses device bitcode before SPIRV.
• The implementation of preventing hipLaunchKernel latency degradation with number of idle streams is reverted/disabled by default.

Optimized

• Improved kernel logging includes de-mangling shader names.
• Refined implementation in HIP APIs hipEventRecords and hipStreamWaitEvent for performance improvement.

Resolved issues

• Stale state during the graph capture. The return error was fixed, HIP runtime now always uses the latest dependent nodes during hipEventRecord capture.
• Segmentation fault during kernel execution. HIP runtime now allows maximum stack size as per ISA on the GPU device.

hipBLASLt (0.12.1)

Resolved issues

• Fixed an accuracy issue for some solutions using an FP32 or TF32 data type with a TT transpose.

RCCL (2.22.3)

Changed

• MSCCL++ is now disabled by default. To enable it, set RCCL_MSCCLPP_ENABLE=1.

Resolved issues

• Fixed an issue where early termination, in rare circumstances, could cause the application to stop responding by adding synchronization before destroying a proxy thread.
• Fixed the accuracy issue for the MSCCLPP allreduce7 kernel in graph mode.

Known issues

• When splitting a communicator using ncclCommSplit in some GPU configurations, MSCCL initialization can cause a segmentation fault. The recommended workaround is to disable MSCCL with export RCCL_MSCCL_ENABLE=0. This issue will be fixed in a future ROCm release.

• Within the RCCL-UnitTests test suite, failures occur in tests ending with the .ManagedMem and .ManagedMemGraph suffixes. These failures only affect the test results and do not affect the RCCL component itself. This issue will be resolved in a future ROCm release.

rocALUTION (3.2.3)

Added

• The -a option has been added to the rmake.py build script. This option allows you to select specific architectures when building on Microsoft Windows.

Resolved issues

• Fixed an issue where the HIP_PATH environment variable was being ignored when compiling on Microsoft Windows.

ROCm Data Center Tool (0.3.0)

Added

• Support for GPU partitions.
• RDC_FI_GPU_BUSY_PERCENT metric.

Changed

• Updated rdc_field to align with rdc_bootstrap for current metrics.

Resolved issues

• Fixed ROCProfiler eval metrics and memory leaks.

ROCm SMI (7.5.0)

Resolved issues

• Fixed partition enumeration. It now refers to the correct DRM Render and Card paths.

See the full [ROCm SMI changelog](https://github.com/AMD-ROCm-Internal/rocm_smi_lib/blob/release/rocm-rel-6.4/CHANGELOG.md) for details, examples, and in-depth descriptions.

ROCm Systems Profiler (1.0.1)

Added

• How-to document for network performance profiling for standard Network Interface Cards (NICs).

Resolved issues

• Fixed a build issue with Dyninst on GCC 13.

ROCr Runtime (1.15.0)

Resolved issues

• Fixed a rare occurrence issue on AMD Instinct MI25, MI50, and MI100 GPUs, where the SDMA copies might start before the dependent Kernel finishes and could cause memory corruption.

ROCm known issues

ROCm known issues are noted on {fab}github GitHub. For known issues related to individual components, review the Detailed component changes.

Radeon AI PRO R9700 hangs when running Stable Diffusion 2.1 at batch sizes above four

Radeon AI PRO R9700 GPUs might hang when running Stable Diffusion 2.1 with batch sizes greater than four. As a workaround, limit batch sizes to four or fewer. This issue will be addressed in a future ROCm release.

RCCL MSCCL initialization failure

When splitting a communicator using ncclCommSplit in some GPU configurations, MSCCL initialization can cause a segmentation fault. The recommended workaround is to disable MSCCL with export RCCL_MSCCL_ENABLE=0. This issue will be fixed in a future ROCm release.

ROCm upcoming changes

The following changes to the ROCm software stack are anticipated for future releases.

ROCm SMI deprecation

ROCm SMI will be phased out in an upcoming ROCm release and will enter maintenance mode. After this transition, only critical bug fixes will be addressed and no further feature development will take place.

It's strongly recommended to transition your projects to AMD SMI, the successor to ROCm SMI. AMD SMI includes all the features of the ROCm SMI and will continue to receive regular updates, new functionality, and ongoing support. For more information on AMD SMI, see the AMD SMI documentation.

ROCTracer, ROCProfiler, rocprof, and rocprofv2 deprecation

Development and support for ROCTracer, ROCProfiler, rocprof, and rocprofv2 are being phased out in favor of ROCprofiler-SDK in upcoming ROCm releases. Starting with ROCm 6.4, only critical defect fixes will be addressed for older versions of the profiling tools and libraries. All users are encouraged to upgrade to the latest version of the ROCprofiler-SDK library and the (rocprofv3) tool to ensure continued support and access to new features. ROCprofiler-SDK is still in beta today and will be production-ready in a future ROCm release.

It's anticipated that ROCTracer, ROCProfiler, rocprof, and rocprofv2 will reach end-of-life by future releases, aligning with Q1 of 2026.

AMDGPU wavefront size compiler macro deprecation

Access to the wavefront size as a compile-time constant via the __AMDGCN_WAVEFRONT_SIZE and __AMDGCN_WAVEFRONT_SIZE__ macros or the constexpr warpSize variable is deprecated and will be disabled in a future release.

• The __AMDGCN_WAVEFRONT_SIZE__ macro and __AMDGCN_WAVEFRONT_SIZE alias will be removed in an upcoming release. It is recommended to remove any use of this macro. For more information, see AMDGPU support.
• warpSize will only be available as a non-constexpr variable. Where required, the wavefront size should be queried via the warpSize variable in device code, or via hipGetDeviceProperties in host code. Neither of these will result in a compile-time constant.
• For cases where compile-time evaluation of the wavefront size cannot be avoided, uses of __AMDGCN_WAVEFRONT_SIZE, __AMDGCN_WAVEFRONT_SIZE__, or warpSize can be replaced with a user-defined macro or constexpr variable with the wavefront size(s) for the target hardware. For example:

   #if defined(__GFX9__)
   #define MY_MACRO_FOR_WAVEFRONT_SIZE 64
   #else
   #define MY_MACRO_FOR_WAVEFRONT_SIZE 32
   #endif

HIPCC Perl scripts deprecation

The HIPCC Perl scripts (hipcc.pl and hipconfig.pl) will be removed in an upcoming release.

Changes to ROCm Object Tooling

ROCm Object Tooling tools roc-obj-ls, roc-obj-extract, and roc-obj are deprecated in ROCm 6.4, and will be removed in a future release. Functionality has been added to the llvm-objdump --offloading tool option to extract all clang-offload-bundles into individual code objects found within the objects or executables passed as input. The llvm-objdump --offloading tool option also supports the --arch-name option, and only extracts code objects found with the specified target architecture. See llvm-objdump for more information.

HIP runtime API changes

There are a number of upcoming changes planned for HIP runtime API in an upcoming major release that are not backward compatible with prior releases. Most of these changes increase alignment between HIP and CUDA APIs or behavior. Some of the upcoming changes are to clean up header files, remove namespace collision, and have a clear separation between hipRTC and HIP runtime. For more information refer to HIP Upcoming changes.