February 2025 — HabanaAI/vllm-hpu-extension: Focused quantization safety improvement for Mixtral models. Implemented a calibration patch that blocks self_attn and lm_head to prevent accuracy regressions during Mixtral quantization. Added a Mixtral-specific quant config to calibration. The change enhances reliability and deployment readiness of quantized Mixtral models on Habana AI hardware.

December 2024: Delivered DynamicMixture of Experts (DynamicMoE) support for Mixtral models on Gaudi hardware within HabanaAI/optimum-habana-fork. The change conditionally routes the model forward pass through a dynamic MoE implementation when a quantization configuration is present, enabling hardware-optimized MoE execution and improving performance and resource utilization on Gaudi accelerators. This work establishes groundwork for faster inference, reduced latency, and lower per-request costs for Mixtral deployments.