Online Softmax Tiling FlashAttention kernel executing bounded SRAM query/key/value block updates with running max and sum rescaling for O(1) memory footprint.
-
Updated
Sep 9, 2026 - Python
Online Softmax Tiling FlashAttention kernel executing bounded SRAM query/key/value block updates with running max and sum rescaling for O(1) memory footprint.
Online Softmax Tiling FlashAttention kernel executing bounded SRAM query/key/value block updates with running max and sum rescaling for O(1) memory footprint.
Rigorous self-study roadmap to master algorithms, low-level systems, distributed systems, and deep learning infrastructure through CMU/MIT/Stanford/Berkeley courses.
The implementation of the CMU 10-414/714: Deep Learning Systems course assignment
A platform to learn to build systems at scale, from backend to LLMs.
🌠 End-to-end deep learning systems study: custom PyTorch ResNet-18 built from scratch, CUDA kernel profiling with torch.profiler, and multi-GPU PyTorch DistributedDataParallel (DDP) benchmarking on 2x NVIDIA RTX 4090 GPUs (achieving 1.72× speedup & 86% scaling efficiency).
Building the needle deep-learning framework from scratch (CMU 10-414/714 Deep Learning Systems): reverse-mode autodiff, nn library, optimizers, CPU NDArray backend, CNN & RNN/LSTM — with the official tests
Backprop versus the memory wall: an interactive roofline case that backprop was arithmetic-bound in 1986 and is memory-bound in 2026
To associate your repository with the deep-learning-systems topic, visit your repo's landing page and select "manage topics."