April 2026 monthly summary for developer work across two repositories with a focus on delivering high-impact features, stabilizing memory usage, and improving system resilience for long-context workloads. Key features delivered and major fixes: - ml-explore/mlx-lm: GatedDeltaNet Cache Memory Leak Fix. Fixed a memory leak by ensuring contiguous memory and preventing shared-buffer leaks, improving stability and model performance. Commit: 9dcefa5272d2a2a828bbdb362435eca9bfc9615d (fix: break shared-buffer memory leak in GatedDeltaNet cache #1077). - exo-explore/exo: Memory management and stability for KVPrefixCache eviction. Added garbage collection and cache clearing after eviction to promptly free MLX Metal buffers; reduced memory retention by ~3-4 GB between long-context requests. Commit: af9e847edbb1939872f79889cfe1617d9fe1362e (fix: force gc + clear_cache after KV prefix cache eviction #1832). - exo-explore/exo: API parameter support for presence_penalty and frequency_penalty. Wired new penalties from ChatCompletionRequest to the API adapter, enabling finer control over text generation. Commit: 48a922fd5c23108fdd19b00106306023b335a8e1 (fix: map presence_penalty and frequency_penalty from ChatCompletionRequest #1991). - exo-explore/exo: Load balancing correctness. Routing decisions now consider only in-flight tasks (Pending/Running) to avoid skew from completed tasks; fixes uneven distribution and improves throughput. Commit: 5d10188d3abe0c4cc5bb4365eddf7dd819f0c269 (fix: route by in-flight tasks only — completed tasks were skewing load balance #1989). Overall impact and accomplishments: - Stability and reliability: Memory leaks and stale buffers addressed, enabling longer context processing without disproportionate memory growth. - Performance and throughput: More accurate load balancing and faster eviction handling yield steadier utilization and higher sustained throughput under peak workloads. - API usability: End-to-end support for text-generation tuning parameters (presence_penalty, frequency_penalty) now available to users. - Collaboration and code quality: Cross-repo fixes with co-authored contributions and clear changelogs, improving maintainability and traceability. Technologies and skills demonstrated: - Memory management and garbage collection strategies in production ML pipelines (Python GC, explicit cache eviction handling) - Efficient resource management for MLX Metal buffers - API wiring and parameter mapping for flexible generation controls - Load balancing algorithms focusing on in-flight tasks for accurate workload distribution - Cross-team collaboration and code review discipline Business value: - Enabled longer-context inference and more predictable performance under load, improving user experience and throughput for generation workloads.