pytorch 2.13: flexattention on apple silicon, fused loss, new distributed backend
pytorch 2.13 brings flexattention to mps, a fused linear and cross-entropy loss, a new torchcomms backend, and more.
topic
machine learning
triroute jointly routes attention, experts, and kv-cache
a single lightweight controller coordinates attention mode, expert selection, and cache bit-width per token to reduce inference cost without losing quality.
vllm transformers backend now matches native speed
the transformers modeling backend in vllm now achieves throughput equal to or better than hand-written native implementations for many large language models.
open data and synthetic data for ai agents
nvidia nemotron releases open datasets and a prompt atlas to help developers build inspectable, reliable ai agents using synthetic data.
geometry framework tests if ai learns or just memorizes
a new mathematical model uses fiber bundles to detect when machine learning systems discover causal laws instead of interpolating data.
the granularity paradox in time-series forecasting
finer temporal disaggregation inflates in-sample fit but degrades out-of-sample accuracy due to recursive error compounding.
one-click from hugging face to sagemaker studio
hugging face and amazon sagemaker ai now offer a deep-link integration that lets developers move from model discovery to fine-tuning or deployment in sagemaker studio with a single click.
sqlite-utils 4.0 adds schema migrations
sqlite-utils 4.0 introduces database schema migrations, nested transactions, and compound foreign keys for sqlite databases.
zero-egress ai storage with skypilot and hugging face
mount hugging face buckets and repos into skypilot jobs on any cloud with no egress fees for reads, using one hf:// url and your existing token.
sql vs pandas vs ai agents for analytics
a comparison of sql, pandas, and a claude agent on three analytics problems across speed, accuracy, and other dimensions.
hugging face models on foundry managed compute
microsoft foundry now offers a curated catalog of open-weight hugging face models deployable with one click onto managed compute, refreshed weekly and pre-staged in azure.
treat pdfs as images for zero-shot parsing with gemma 4
render pdf pages to images and use gemma 4 to extract structured data without ocr or layout parsers.
10 probability concepts for machine learning
a plain-language guide to ten essential probability ideas that make machine learning work, from random variables to entropy.
benchmark audits can silently fail in five ways
a study shows that perturbation-based construct-validity audits can produce misleading conclusions due to hidden pipeline failures, and proposes a due-diligence gate to catch them.
graft adds per-word pronunciation control to text-to-speech
a new method lets text-to-speech models pronounce tricky words correctly by using short audio hints, without changing the speaker's voice.
pytorch monarch comes to amd gpus for fault-tolerant training
pytorch monarch now supports amd instinct gpus with rocm, enabling single-controller distributed training that recovers from node failures without full restarts.
how photoroom built the data pipeline for prx
photoroom details its data strategy for the prx text-to-image model, covering dataset assembly, captioning, and storage formats.
data scientists shift to ai oversight roles
data scientists now spend more time managing ai systems than building models, with new roles in governance, prompt engineering, and agent supervision.