April 2026 monthly summary for vllm-gaudi: Implemented Qwen3.5 Compact Mode with memory optimization and fixed critical stability issues to improve memory efficiency, accuracy, and multi-batch reliability for hybrid Qwen3.5 deployments on GAUDI/HPC. The changes reduce memory waste, improve concurrency, and enhance robustness of inference workflows.

March 2026: Qwen35 enablement in vllm-gaudi with performance-focused refinements and comprehensive validation. Implemented HPU-gated Deltanet for GDN attention, aligned HPU mamba page to the GDN attention block size (128), and reused mamba layer metadata to support GDN attention workflows in the absence of speculative decode. Performed offline testing across 9B and 35B-A3b variants for reasoning, text generation, image, and video workloads. These efforts establish a solid path to production deployment, with stronger performance and scalability for Qwen35 models.

January 2026 monthly summary for red-hat-data-services/vllm-gaudi: Delivered a robust Qwen3 Vision Model upgrade with three conditional processing paths for multi-image requests and enabled mixture of experts (MoE) for scalable performance. Fixed critical accuracy issues when processing multiple images in a single request and optimized attention paths for single and multi-image scenarios. Resulted in more accurate, faster, and flexible vision inference, aligning with our goals for higher throughput and broader use cases in production.