AMD ROCm 发布: ROCm 7.2.4 Release

来源摘要

# ROCm 7.2.4 release notes ROCm 7.2.4 is a quality release focused on performance and stability fixes for AI inference workloads on AMD Instinct GPUs. - [Release highlights](#release-highlights) - [ROCm binaries](#rocm-binaries) ## Release highlights The following are the notable changes in ROCm 7.2.4. ### Reduced hipGraphLaunch latency for multi-list graphs The HIP runtime's graph dispatch mechanism has been optimized, reducing launch latency for workloads using `hipGraphLaunch` with multi-list graph topologies. ### Fixed H2D memory copy latency regression in CPX mode HIP runtime synchronization behavior has been corrected on AMD Instinct MI300 Series GPUs in CPX mode, restoring latency to previous levels for inference workloads that run multiple HIP streams with concurrent memory copies. ### Reduced ROCprofiler-SDK profiling overhead Profiling stability has been improved for vLLM workloads traced with PyTorch `torch.profiler` using the ROCprofiler-SDK backend. The large, sporadic idle gaps that previously appeared between GPU kernels in the trace have been substantially reduced in common configurations, and the traces now more accurately reflect actual runtime behavior. Coverage may vary depending on model and parallelism settings. ### Reduced copy overhead in MIGraphX concat operations MIGraphX now recognizes ONNX models that concatenate the same tensor multiple times and avoids redundant device-side copies, improving inference throughput at small batch sizes for the affected model class on AMD Instinct MI300X GPUs. ### User space, driver, and firmware dependent changes The software for AMD Data Center GPU products requires maintaining a hardware and software stack with interdependencies among the GPU and baseboard firmware, AMD GPU drivers, and the ROCm user space s