Op
Optimizing DeepSeek's NSA for TPUs – A Kernel Worklog
Optimizing DeepSeek's NSA for TPUs – A Kernel Worklog
I enjoyed reading DeepSeek's Natively Sparse Attention(NSA) and implemented a JAX version and Pallas Kernel(JAX's kernel language). There's a Google Colab link to follow along at the top of the blogpost. Enjoy!
Share cardActual performance
2points
Did not reach leaderboard
Launch Intel predictions
Analyze your own launch →62%62% predicted probability of success on Product Hunt, based on ML models trained on real launch data.
best fitHighest predicted score across all platforms for this description.
48%48% predicted probability of success on Hacker News, based on ML models trained on real launch data.
nativeThis product was originally launched on this platform.
42%42% predicted probability of success on BetaList, based on ML models trained on real launch data.
31%31% predicted probability of success on Indie Hackers, based on ML models trained on real launch data.
29%29% predicted probability of success on TrustMRR, based on ML models trained on real launch data.
27%27% predicted probability of success on AppSumo, based on ML models trained on real launch data.
19%19% predicted probability of success on Acquire.com, based on ML models trained on real launch data.
Correct prediction on native model
Similar products
tokenmix39%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
#chatgpt #deepseek
DeepSeek PDF Exporter14%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
Export deepseek to pdf
De
DeepSeek Your HN Profile47%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
DeepSeek Your HN Profile
De
DeepSeek Janus Pro 7B API30%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
DeepSeek Janus Pro 7B API
Sn
Snappy, local-first DeepSeek chat37%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
Snappy, local-first DeepSeek chat
De
DeepSeek My User Agent32%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
DeepSeek My User Agent
DeepSeek For Android40%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
DeepSeek's official AI assistant for free
Awesome Deepseek Integration46%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
DeepSeek API Integrations, Officially Recommended
Locl.chat42%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
The snappiest Deepseek chat
AI Chat Finder (DeepSeek)52%Launch Intel prediction score: how likely this product is to succeed on its source platform, based on its name, tagline, and description.
Chat Content Search for DeepSeek + More