April 2026 (2026-04) — neuralmagic/compressed-tensors: Delivered a critical fix for the WEIGHT_OUTPUT transformation to also transform the bias vector, ensuring correct outputs for models with bias during rotation/transformation. Added test coverage for bias handling in both Linear layers and attention heads (Qwen2-like scenarios) to prevent regressions. This fix eliminates downstream misalignment (bias becoming untransformed can cause incorrect outputs in subsequent layers), which previously led to degraded performance. Commit 37f209afe2042d9b9493be0c69f347dc36e6498a and related changes provide the change set; tests validate correctness and guard against future regressions. Impact: stabilizes model outputs during transform-based quantization, preserving accuracy and reliability, reducing deployment risk. Skills demonstrated: bias-propagation math, rotation-based transforms, test-driven development, code review and collaboration.