April 2026 monthly summary for microsoft/onnxruntime focused on stabilizing the QNN API file-mapping feature by implementing compatibility checks and per-model gating. Delivered guarded usage based on QNN EP ABI context binary version, added per-model disables for edge cases, and ensured failures in older context binaries do not cascade to all graphs or sessions. The work reduces resource leaks and runtime instability, aligning behavior with the QNN EP ABI repo and improving overall reliability for edge devices and mixed EP contexts.

Month: 2026-03 — Microsoft/onnxruntime: Delivered a stability-focused bug fix for the QNN Execution Provider by disabling file mapping for embedded cache contexts when econtext_embed_mode = 1. This prevents the framework from attempting to map an embedded EP context as a non-existent file, eliminating a class of runtime errors in models that contain embedded EP nodes. Impact: Reduced runtime failures, improved reliability for production models using embedded EPs, and simplified support by removing a frequent error path. Commit reference: b0b5de39eaab2ae24e2dc489ddff4cf0aaa8ccc4 ("[QNN-EP] Disable file mapping for embedded cache (#27627)"). Co-authored by calvnguy and quic_calvnguy.

January 2026 — intel/onnxruntime QNN Execution Provider: Delivered two targeted features that improve runtime efficiency and multi-session scalability. The HTP Power Configuration Management introduces a new class to dynamically update performance modes and disables DSPQ polling when not in burst, reducing unnecessary polling and saving CPU cycles. The ARM64 Windows optimization adds file-mapped weights and context bin support to minimize heap allocations and enable shared context bins across sessions, improving initialization performance for multiple ORT sessions. No major bugs fixed were reported in this period based on the provided data. These efforts improve business value by lowering resource usage, enabling more stable large-model deployments, and supporting scalable, memory-efficient inference.

December 2025 monthly summary for intel/onnxruntime. Implemented a session-based HTP power configuration lifecycle to optimize resource usage and improve scalability in multi-threaded workloads. Removed PerThreadContext, introduced ManagedHtpPowerConfigId, and ensured only a single HTP power config ID is created per session. This design reduces the risk of exhausting available IDs, simplifies lifecycle management, and aligns with platform-wide performance goals.

November 2025 – Intel/onnxruntime monthly summary Key delivered feature: - QNN Backend Power Management Enhancement: aligned sustained_high_performance (SHP) power votes with burst settings to boost QNN backend performance without DSPQ polling. Uses the same HTP voltage corners as burst for sustained performance, ensuring consistent voltage mappings. Commit: c30905d638418383b8d83b3b1bb65b7b42226f5a. Impact and rationale: - Improves sustained throughput for QNN workloads while reducing polling overhead. - More predictable hardware power behavior under high-load scenarios. - Cross-team collaboration with the change co-authored by quic_calvnguy. Technologies/skills demonstrated: - QNN backend optimization, hardware power management, voltage corner mapping, HTP/burst policy. - Code contributions and collaboration. Bugs fixed: - No major bugs fixed this month for this repository. Business value: - Higher throughput and energy efficiency for QNN workloads, leading to improved user-perceived performance and lower operational costs.

In October 2025, delivered QNN Execution Provider (EP) OpTrace Profiling for intel/onnxruntime, enabling detailed debugging and performance analysis. The feature integrates with the QNN System Profile API to generate binary log files compatible with qnn-profile-viewer, providing richer profiling when the QNN API is >= 2.29 while remaining backward-compatible with older versions. A new QNN System Profile serializer class was added, along with API versioning safeguards and profiling tests. This work enhances observability, accelerates performance tuning, and improves debugging workflows for QNN-based workloads.

September 2025: QNN Execution Provider RPC polling interval logic bug fixed in the intel/onnxruntime repo, improving execution flow, stability, and performance management. The interval is now correctly set to 9999 only when performance mode is burst, eliminating misconfiguration in non-burst scenarios. This fix enhances reliability of RPC-based polling and overall QNN EP behavior.

August 2025 monthly summary for intel/onnxruntime focused on delivering stability improvements and robust memory handling in the QNN-EP integration. A critical VTCM buffer sharing bug was fixed by pre-allocating memory for context parameters and by implementing robust handling for new binary contexts, addressing root causes that affected multi-context inferences.

July 2025 monthly summary for ROCm/onnxruntime focusing on QNN Execution Provider (QNN_EP). Delivered two key features to enhance performance management and workload-driven prioritization, with direct commits tied to PRs in the ROCm/onnxruntime repo. No major bugs fixed this month.

Month 2025-06 – ROCm/onnxruntime: Delivered VTCM backup buffer sharing in the QNN execution provider, enabling sharing of VTCM backup buffers across multiple graphs. This reduces per-graph buffer allocations and memory overhead, improving performance and resource utilization for large language models.