ROCm 7.2.3 finally patches those frustrating idle gaps in vLLM profiling traces that turn latency debugging into a guessing game. MIGraphX gets a solid performance bump for embedding-heavy models through fused gather operations, and ONNX Runtime gains better multi-stream memory handling to keep inference pipelines from choking on themselves. Server admins should double-check their firmware and driver matrices since MI325X KVM users absolutely must skip the 30.20.0 AMD GPU driver. The software stack is also actively retiring legacy profiling tools and ROCm SMI by mid-2026, so teams need to migrate to rocprofv3 and AMD SMI before those support windows close.

ROCm 7.2.3 Fixes vLLM Profiling Gaps and Speeds Up MIGraphX Gather Ops

ROCm 7.2.3 lands with targeted improvements for AI inference and profiling stability. The update tightens up trace accuracy for vLLM workloads, optimizes gather operations in MIGraphX for embedding-heavy models, and introduces external stream support for ONNX Runtime reliability. There are also critical deprecation notices regarding legacy profiling tools and system management utilities that require migration planning before the current support windows close.

Profiling Stability for vLLM Workloads

Debugging latency issues becomes significantly harder when profiling traces display massive idle periods that do not match actual hardware behavior. ROCm 7.2.3 addresses this by reducing large, sporadic idle gaps in traces generated via PyTorch torch.profiler. These gaps previously made it difficult to correlate GPU kernel execution with runtime events, often leading to wasted time chasing phantom bottlenecks. The update substantially reduces these artifacts in common configurations, making the trace timeline reflect actual runtime behavior much more closely. Coverage still varies depending on model architecture and parallelism settings, so validation against specific workload configurations remains necessary before relying solely on the new trace data for optimization decisions.

MIGraphX Performance and ONNX Runtime Reliability

MIGraphX receives meaningful performance enhancements focused on embedding-heavy inference workloads. The gather operator now supports horizontal fusion for cross-embedding operations, which merges multiple independent gathers into a single batched operation. This reduces kernel launches and cuts down memory traffic by leveraging transpose, reshape, broadcast, and slice optimizations across different data layouts. ONNX Runtime users benefit from external stream support in the MIGraphX Execution Provider, which improves memory allocation handling for multi-stream inference scenarios. A known regression exists for int8-quantized models where a slight performance drop may occur on peak throughput metrics. This impact is generally minimal and does not affect model correctness, but latency-sensitive quantized workloads should benchmark their specific paths to ensure the trade-off remains acceptable.

Firmware Dependencies and Virtualization Constraints

Data center GPUs require strict alignment between GPU firmware, baseboard IFWI, AMD GPU drivers, and ROCm user space software. The release maintains existing hardware support but enforces tight versioning dependencies that cannot be ignored during upgrades. MI325X KVM SR-IOV users must avoid AMD GPU driver version 30.20.0 due to known compatibility issues. Multi-VF support for the MI300X with eight virtual functions requires a compatible firmware BKC bundle that is scheduled for release in coming months, so infrastructure planning should account for this dependency. Baseboard firmware updates via PLDM bundles are essential for stability across Instinct families, and mismatches between GPU and baseboard firmware versions can lead to unpredictable behavior. Server administrators managing racks of MI355X, MI350, or MI325X cards should verify the driver and firmware matrix before deploying this update to production environments.

ROCm 7.2.3 Deprecation Warnings and Tooling Changes

The software stack is actively shedding legacy tools that will soon reach end-of-support. ROCTracer, ROCProfiler, rocprof, and rocprofv2 are officially deprecated, with full end-of-support expected by the second quarter of 2026. Migration to the latest ROCprofiler-SDK library and the rocprofv3 tool is strongly recommended to maintain access to new features and continued support. ROCm SMI is also being phased out in favor of AMD SMI, which includes all existing functionality plus ongoing feature development. Monitoring scripts that rely on rocm-smi should be updated to use the new utility immediately. Object tooling components like roc-obj-ls, roc-obj-extract, and roc-obj have been deprecated since ROCm 6.4 and will be removed in a future release. Functionality has moved to llvm-objdump with the offloading flag, which now supports extracting clang-offload-bundles into individual code objects for specific target architectures using the arch-name option.

Early Access Features and Known Issues

ROCm XIO documentation updates introduce an API for Accelerator-Initiated IO that allows AMD GPUs to perform direct operations to hardware devices without CPU intervention. This feature was initially released as an early-access technology preview in April 2026, and production workloads should avoid this functionality until it matures beyond the preview stage. The primary known issue to track remains the minor performance regression for MIGraphX with int8-quantized models, which is currently under investigation and slated for a fix in a future release. Users running embedding-heavy inference or relying on detailed vLLM profiling traces will likely see immediate value from this update, provided they verify firmware compatibility and migrate off deprecated tooling before the support deadlines arrive.

Release ROCm 7.2.3 Release

ROCm 7.2.3 release notes The release notes provide a summary of notable changes since the previous ROCm release.

Release ROCm 7.2.3 Release · ROCm/ROCm

Keep an eye on the firmware matrix if managing Instinct cards, and start migrating off those legacy profiling tools before the support window closes.