Activation sparsity can enable practical inference speedups in large language models (LLMs) by reducing the compute and memory-movement required for matrix multiplications during the forward pass. However, existing methods face limitations that inhibit widespread adoption. Some approaches are tailored towards older models with ReLU-based sparsity, while others require extensive continued pre-training on up to hundreds of billions of tokens. This paper describes TEAL (**T**raining-Fre**e** **A**ctivation Sparsity in **L**LMs), a simple training-free method that applies magnitude-based activation sparsity to hidden states throughout the entire model. TEAL achieves 40-50\% model-wide sparsity with minimal performance degradation across Llama-2, Llama-3, and Mistral families, with sizes varying from 7B to 70B. We improve existing sparse kernels and demonstrate wall-clock decoding speed-ups of up to 1.53× and 1.8× at 40\% and 50\% model-wide sparsity. TEAL is compatible with weight quantization, enabling further efficiency gains.

Large Language Models (LLMs) are typically trained in two phases: pre-training on large internet-scale datasets, and fine-tuning for downstream tasks. Given the higher computational demand of pre-training, it is intuitive to assume that fine-tuning adds less new information to the model, and is thus more compressible. We explore this assumption by decomposing the weights of fine-tuned models into their pre-trained components and an additional delta. We introduce a simple method, BitDelta, which successfully quantizes this delta down to 1 bit without compromising performance. This interesting finding not only highlights the potential redundancy of information added during fine-tuning, but also has significant implications for the multi-tenant serving and multi-tenant storage of fine-tuned models. By enabling the use of a single high-precision base model accompanied by multiple 1-bit deltas, BitDelta dramatically reduces GPU memory requirements by more than 10x, thus reducing per-user generation latency by more than 10x in multi-tenant settings. We validate BitDelta through experiments across Llama-2, Mistral and MPT model families, and on models up to 70B parameters, showcasing minimal performance degradation in all tested settings.

During terrestrial activities, sensation of pressure on the skin and tension in muscles and joints provides information about how the body is oriented relative to gravity and how the body is moving relative to the surrounding environment. In contrast, in aquatic environments when suspended in a state of neutral buoyancy, the weight of the body and limbs is offloaded, rendering these cues uninformative. It is not yet known how this altered sensory environment impacts virtual reality experiences. To investigate this question, we converted a full-face SCUBA mask into an underwater head-mounted display and developed software to simulate jetpack locomotion outside the International Space Station. Our goal was to emulate conditions experienced by astronauts during training at NASA's Neutral Buoyancy Lab. A user study was conducted to evaluate both sickness and presence when using virtual reality in this altered sensory environment. We observed an increase in nausea related symptoms underwater, but we cannot conclude that this is due to VR use. Other measures of sickness and presence underwater were comparable to measures taken above water. We conclude with suggestions for improved underwater VR systems and improved methods for evaluation of these systems based on our experience.

Teleportation is a popular locomotion technique that lets users safely navigate beyond the confines of available positional tracking space without inducing VR sickness. Because available walking space is limited and teleportation is faster than walking, a risk with using teleportation is that users might end up abandoning walking input and only relying on teleportation, which is considered detrimental to presence. We present redirected teleportation; an improved version of teleportation that uses iterative non-obtrusive reorientation and repositioning using a portal to redirect the user back to the center of the tracking space, where available walking space is larger. A user study compares the effectiveness, accuracy, and usability of redirected teleportation with regular teleportation using a navigation task in three different environments. Results show that redirected teleportation allows for a better utilization of available tracking space than regular teleportation, as it requires significantly fewer teleportations, while users walk more and use a larger portion of the available tracking space.

A novel adaptive coding characteristic prediction scheme called RD-cost prediction model is proposed. Through this linear model and optimum bit allocation, RD-cost and texture bits budget of current basic unit can be obtained, and then the quantization parameter (QP) can be directly calculated. Simulation results show that the H.264 encoder, using the proposed algorithm, achieves a visual quality improvement up to 0.302 dB, meets better with target bit rates, produces more flat bit-rate curve and decreases more computation complexity than that using H.264 previous rate control method JVT-G012.

This paper presents an innovative algorithm for detecting all-zero discrete cosine transform (DCT) coefficient blocks prior to DCT and quantization for all block-based video coding standards. A mathematic model is established based on analyzing DCT coefficient distribution and applying Paseval energy conservation theorem. The algorithm is applied to H.264 video coding and experimental results show up to 32% higher detection ratio without degrading video quality, compared with results of the existing methods. Furthermore, up to 47% higher detection ratio is achieved by changing the threshold based on DCT coefficient distribution. For motion estimation, much less search points needed by using the proposed method as early termination criterion than that by using the existing methods