polars vs pandas: real-world performance deep dive
a comparison of polars and pandas on three data tasks shows polars is faster and uses less memory due to lazy evaluation and parallelism.
topic
machine learning
10 github repos to learn fastapi
a curated list of ten github repositories that help developers learn fastapi through templates, examples, auth tools, microservices, and ml projects.
rl-kirigami speeds inverse design of shape-shifting structures
a new framework combines flow matching and reinforcement learning to quickly generate valid kirigami cut patterns for target shapes.
pathboost: gradient boosting learns graph features from paths
pathboost is a gradient tree boosting method for graph-level prediction that learns path-based features directly from graph structure, outperforming graph neural networks on half of benchmark datasets.
build an ai lms with ollama, fastapi, and react
a step-by-step guide to creating a learning management system that adapts to each learner using local ai models.
smarter json alternative for llm pipelines
toon reduces token waste in llm prompts by compacting repeated json structure into a tabular format.
aws foundation model building blocks
aws infrastructure for foundation model training and inference integrates accelerated compute, high-bandwidth networking, and distributed storage with open-source orchestration and ml frameworks.
build vector search from scratch in python
a step-by-step guide to creating a vector search engine using only numpy, covering embeddings, normalization, cosine similarity, and visualization.
in-kernel broadcast optimization for recsys inference
meta's ikbo eliminates redundant user embedding replication in recommendation models by fusing broadcast logic directly into gpu kernels, cutting latency by up to two-thirds.
adding benchmaxxer repellant to the open asr leaderboard
hugging face adds private speech datasets to its asr leaderboard to reduce benchmark gaming and provide a more realistic view of model performance across accents and speaking styles.
vllm v0 to v1: correctness before corrections in rl
migrating from vllm v0 to v1 for online reinforcement learning required fixing logprob semantics, runtime defaults, weight updates, and fp32 head precision to match training dynamics.
lkv trims llm memory with learned cache budgets
a new method learns optimal key-value cache compression directly from task objectives, improving long-context llm efficiency without heuristic rules.
ratequant tunes kv cache precision for llms
a new method allocates different bit-widths to attention heads in kv cache quantization, avoiding distortion model mismatch to improve large language model serving efficiency.
emo: pretraining mixture of experts for emergent modularity
emo is a mixture-of-experts model trained so that experts self-organize into task-specific groups, allowing strong performance with only a small subset of experts.