陶智勇 @tri_dao05-29 08:41
RT Mark Saroufim: My MLSys keynote on AI writing systems code got more interest than I expected. The recording will take a while, so in the finest tra...
🔗 原推0今日信号
尚未被主编选入事件/主题的原始推文（共 40 条，按时间倒序）
陶智勇 @tri_dao05-29 01:26
RT Albert Gu: Our new model Ink-2 tops AA's leaderboard for streaming speech-to-text! Ink-2 comes with plenty of features optimized for real-time voic...
🔗 原推陶智勇 @tri_dao05-29 00:50
RT Cartesia: Cartesia Ink-2 debuts as #1 for accuracy on the brand-new streaming speech-to-text leaderboard from @ArtificialAnlys! We designed Ink-2 f...
🔗 原推陶智勇 @tri_dao05-28 00:56
RT LightSeek Foundation: Proud to collaborate with the @Alibaba_Qwen, @NVIDIAAI, @KVCache_AI, @PyTorch, and FlashAttention @tri_dao on this effort �...
🔗 原推陶智勇 @tri_dao05-28 00:02
RT PyTorch: The speed-of-light optimization for Qwen3.5 on the TokenSpeed inference engine is a significant milestone, achieving a record-breaking 580...
🔗 原推陶智勇 @tri_dao05-23 02:30
RT Clive Chan: the transformer reparameterizations lab has released even more reparameterizations :D
🔗 原推陶智勇 @tri_dao05-22 09:51
After some mathematical rewrite, turns out all of transformer is a series of gemm + epilogue. Given a few optimized primitives, LLMs (and novice human...
🔗 原推陶智勇 @tri_dao05-22 09:16
RT Together AI: Introducing Qwen3.7-Max from @Alibaba_Qwen, Qwen’s flagship model for the agent era with 1M context and leading performance across ag...
🔗 原推陶智勇 @tri_dao05-22 08:39
RT Oliver Sieberling: A significant portion of transformer training time is not spend doing matmuls, but for norms, residuals, RoPE, etc. This work sh...
🔗 原推陶智勇 @tri_dao05-22 07:36
RT Jack Zhang: We built a kernel abstraction to rewrite the entire transformer stack as GEMM + Epilogue kernels! Neural net architectures such as tran...
🔗 原推陶智勇 @tri_dao05-19 02:50
RT Eric Chen: The 100 in A100 is the number of hours you'll spend reading this post. 29,000 words on FlashAttention-2 in CuTe -- a line-by-line dredge...
🔗 原推陶智勇 @tri_dao05-15 02:34
RT Together AI: Seven papers. One research team. Together AI is heading to #MLSys2026 next week. Check out the work going from research to production ...
🔗 原推陶智勇 @tri_dao05-12 08:37
RT Neil Movva: Throughput vs. latency is a foundational tradeoff in almost any system. @sailresearchco is building inference for long-horizon agents, ...
🔗 原推陶智勇 @tri_dao05-07 03:24
RT Flapping Airplanes: (1/5) Great to be at @sequoia to give a sneak peek of one of our research directions! TL;DR one path to data-efficiency may be ...
🔗 原推陶智勇 @tri_dao05-07 00:38
RT Together AI: Together AI has long been a proud supporter of open-source innovation in inference. We're excited for the new TokenSpeed inference eng...
🔗 原推陶智勇 @tri_dao05-06 23:29
RT LightSeek Foundation: Introducing TokenSpeed, a speed-of-light LLM inference engine. > TensorRT LLM level performance > vLLM level usability > Buil...
🔗 原推陶智勇 @tri_dao05-04 06:17
RT driss guessous: I alluded to this a few tweets ago but just pushed up a shortish blog on a subtle feature of CLC work stealing that makes cuda-grap...
🔗 原推陶智勇 @tri_dao05-01 03:54
RT Together AI: Join us Tue 5/5: #DeepSeek-V4's hybrid attention + sparse MoE reduces KV cache up to 90%, enabling 1M-token context. We'll cover why t...
🔗 原推陶智勇 @tri_dao04-24 16:53
RT Ilyas: And we integrated it in Transformers ! As of https://github.com/huggingface/transformers/pull/45433
🔗 原推陶智勇 @tri_dao04-23 07:01
RT Together AI: Introducing Kimi K2.6 from @Kimi_Moonshot, a multimodal agentic model with Agent Swarm scaling to 300 sub-agents and long-horizon codi...
🔗 原推陶智勇 @tri_dao04-23 02:25
RT Mayank Mishra: I love this figure! SonicMoE saves more than 2x memory for Qwen3-235B-A22B 🚀 Check out the full blog: https://dao-lab.ai/blog/202...
🔗 原推陶智勇 @tri_dao04-22 10:57
RT Aran Komatsuzaki: Together AI presents SAW-INT4 Achieves near-BF16 accuracy while preserving the end-to-end performance benefits of INT4
🔗 原推陶智勇 @tri_dao04-17 02:18
The dynamical system view gives very clean conditions for looped transformer to be stable
🔗 原推陶智勇 @tri_dao04-16 00:56
RT Dan Fu: 📢 Super excited to announce Parcae! We've been thinking about scaling laws and the "right" way to get more FLOPs. Turns out layer loopin...
🔗 原推陶智勇 @tri_dao04-14 05:26
RT Together AI: EinsteinArena is a platform where AI agents collaborate on open science problems — submitting solutions, posting in discussion thread...
🔗 原推陶智勇 @tri_dao04-12 09:29
RT Together AI: MiniMax M2.7 is now on Together AI. Trained by letting it run its own RL loop, resulting in the highest open-source score on MLE Bench...
🔗 原推陶智勇 @tri_dao04-11 05:52
RT Monishwaran Maheswaran: Super excited to Introduce our latest work: Squeeze Evolve. We unify test-time scaling methods into one evolutionary framew...
🔗 原推陶智勇 @tri_dao04-08 00:32
RT Hao Kang: Just found out that flashattention use look up table+linear approximation to accelerate softmax. We have tried similar strategies on work...
🔗 原推陶智勇 @tri_dao04-07 00:28
Fast muon optimizer coming to consumer cards. All the code was written as matmul + epilogue so once the mainloop was implemented for Blackwell consume...
🔗 原推陶智勇 @tri_dao04-04 00:52
RT Princeton PLI: Check out latest blog post: Mamba-3: Improved Sequence Modeling using State Space Principles https://pli.princeton.edu/blog/2026/mam...
🔗 原推陶智勇 @tri_dao04-01 05:57
RT Together AI: New from Together Research: Aurora. Speculative decoding that adapts to shifting traffic in real time — and keeps improving the longe...
🔗 原推陶智勇 @tri_dao03-31 01:33
RT noahamsel: Announcing Gram Newton-Schulz, a new way to implement Muon that's 2x faster Trick 1: rejigger Newton-Schulz to replace rectangular matmu...
🔗 原推陶智勇 @tri_dao03-31 01:18
RT Berlin Chen: We cooked up Gram Newton-Schulz: a drop-in replacement of Muon’s Newton-Schulz that is up to 2x faster. Building this requires synthe...
🔗 原推陶智勇 @tri_dao03-31 01:14
It's my favorite kind of work: linear algebra insight + fast kernels. When playing w Muon a while ago, we were thinking why not speed it up by operati...
🔗 原推陶智勇 @tri_dao03-26 19:06
RT vLLM: Thanks to @AI21Labs for tracking down a silent uint32 overflow in vLLM's Mamba-1 CUDA kernel and contributing the fix. Root cause: `uint32_t`...
🔗 原推陶智勇 @tri_dao03-24 04:19
RT PyTorch: PyTorch 2.11 is now available, featuring 2,723 commits from 432 contributors since PyTorch 2.10. This release prioritizes performance scal...
🔗 原推陶智勇 @tri_dao03-23 23:46
RT subho ghosh: My 2 cents on fusing GEMM + Top-K + Softmax on SM100 Finally wrote a blog after ages 🫡 https://ighoshsubho.bearblog.dev/my-2-cents-...
🔗 原推陶智勇 @tri_dao03-18 00:12
RT Pearl Network: It is becoming increasingly clear that the future economy will be denominated in compute cycles more than in human labor. In a world...
🔗 原推陶智勇 @tri_dao03-05 23:01
RT Together AI: Re together.compile — automated kernel optimization. Up to 41% faster image and video generation. No manual kernel work.
🔗 原推
生成于 2026-08-06 10:24（北京时间）· v0.2 CEO 战略情报版