A Khatri Anand

Dikshant Kukreja, Kritarth Prasad, Avinash Anand et al.

Reverse-mode differentiation computes every weight gradient, writes it to memory, and only then lets the optimizer read it back. This two-phase schedule sets the memory ceiling of modern training: at the seam between the phases, every layer's gradient is live at once. We argue that this materialized gradient is an artifact of how differentiation is staged, not a quantity that learning requires -- and we eliminate it. FORGE folds the optimizer step into the backward pass and applies it one tile at a time, entirely in registers, so each gradient tile is consumed the instant it is produced and never becomes a tensor. The fusion changes only when the update happens, not what it computes: in full precision the fused step is provably exact -- the identical optimizer update, for every element-wise rule -- and that exactness survives tensor- and sequence-parallel sharding; in the bf16 and 8-bit regimes used in practice it is faithful rather than bit-identical, its deviation bounded and, for the weight store, rendered unbiased by stochastic rounding. Because each gradient tile is born and consumed in the same registers, it is never converted down to bf16 to be stored and read back; FORGE thus preserves the full-precision fidelity that both bf16 and 8-bit optimizers lose to that conversion. Nor is the method tied to one architecture or one optimizer: linear layers are ubiquitous, and FORGE reclaims the gradient memory of any of them under any element-wise rule. Empirically FORGE more than halves the memory of an optimizer step and, at the small batch sizes typical of fine-tuning and continued pretraining, runs about 1.5x faster; integrated into tensor-parallel Megatron-LM it fits 8B training at four times the micro-batch a standard optimizer allows on the same GPUs.

Ashwary Anand, Shubh Agrawal, Shivang Agrawal et al.

SLAM based techniques are often adopted for solving the navigation problem for the drones in GPS denied environment. Despite the widespread success of these approaches, they have not yet been fully exploited for automation in a warehouse system due to expensive sensors and setup requirements. This paper focuses on the use of low-cost monocular camera-equipped drones for performing warehouse management tasks like inventory scanning and position update. The methods introduced are at par with the existing state of warehouse environment present today, that is, the existence of a grid network for the ground vehicles, hence eliminating any additional infrastructure requirement for drone deployment. As we lack scale information, that in itself forbids us to use any 3D techniques, we focus more towards optimizing standard image processing algorithms like the thick line detection and further developing it into a fast and robust grid localization framework. In this paper, we show different line detection algorithms, their significance in grid localization and their limitations. We further extend our proposed implementation towards a real-time navigation stack for an actual warehouse inspection case scenario. Our line detection method using skeletonization and centroid strategy works considerably even with varying light conditions, line thicknesses, colors, orientations, and partial occlusions. A simple yet effective Kalman Filter has been used for smoothening the ρ and θ outputs of the two different line detection methods for better drone control while grid following. A generic strategy that handles the navigation of the drone on a grid for completion of the allotted task is also developed. Based on the simulation and real-life experiments, the final developments on the drone localization and navigation in a structured environment are discussed.