How FlashAttention computes softmax one block at a time
One query's attention scores, split into two blocks and worked through by hand, to see how FlashAttention gets the softmax without having all the scores at once.
One query's attention scores, split into two blocks and worked through by hand, to see how FlashAttention gets the softmax without having all the scores at once.
替 class 寫了 __eq__,正常使用卻會出現 unhashable type 的錯誤。來看看 Python 為什麼這樣處理,以及 __hash__ 該怎麼做。
A look at the effort levels on Claude Haiku 5.5, and what to check when picking one.
我們來看看 Claude Haiku 5.5 的價格,以及每次請求的差別
We look at what model.eval() and torch.no_grad() each change when a model runs at test time.
用乘號建立二維串列,改了一格卻三列一起變,我們來看看這三列到底是什麼。
We look at the reasoning items the OpenAI Responses API returns during a tool loop and how to send them back.
我們來看看 OpenAI 的 reasoning tokens 是什麼,以及 API 的回傳用量。
我們來看看 Anthropic 的研究怎麼測量出 AI 助理會附和使用者,又是怎麼發現原因出在人類偏好資料和偏好模型上。
A look at Anthropic's paper on sycophancy in AI assistants and human preference judgments.