A user discovered and fixed a significant tensor drift issue in the `ssm_conv1d` layers of quantized Qwen3.6-35B GGUF models, proposing the Wasserstein metric as superior to Kullback Leibler for detecting numerical instability. The fix, which specifically targets recurrent state transition layers responsible for long-context memory, is now available in a shared model.
This paper rigorously analyzes how numerical instability from finite precision leads to unpredictability in LLMs, a critical reliability issue in agentic workflows. It details rounding error propagation, identifying a chaotic "avalanche effect" in early layers and universal, scale-dependent chaotic behaviors.
This paper introduces FastSinkhorn, a native CUDA implementation of the log-domain Sinkhorn algorithm that provides faster and more stable solutions for optimal transport (OT) problems. It achieves a 12x speedup over the POT library and 5.9x over GPU-accelerated PyTorch baselines, maintaining numerical stability for small regularization parameters.