Zihao Ye

Codex laughs at your petty guardrails

@taiuti check out dreamverse.fastvideo.org, we already did it in 1080p /30s a single B200. Give us VR and we get it to 4K and infinity

DCA is cool cuz it uses different aggregators for QKV. main strength of AttnRes are: 1. Co-Design with PP (Block ver.) and Inference (2-stage batching) 2. Large-scale&solid baseline I believe we can found the same core idea in even earlier papers, literature review is hard.

The recording of my talk "Challenges in Decompilation and Reverse Engineering of CUDA-based Kernels" at @REverseConf is now online! Recording: youtube.com/watch?v=ns5jFu… Slides: nicolo.dev/files/pdf/reve… Binary Ninja plugin: github.com/seekbytes/ptxN…

My last open-source project before joining xAI is just out today. Megatron Core MoE is probably the best open framework out there to seriously train mixture of experts at scale. It achieves 1233 TFLOPS/GPU for DeepSeek-V3-685B. arxiv.org/abs/2603.07685

Introducing Code Review, a new feature for Claude Code. When a PR opens, Claude dispatches a team of agents to hunt for bugs.

The slides from my @REverseConf talk, "Challenges in Decompilation and Reverse Engineering of CUDA-based Kernels", are now online! Slides: nicolo.dev/files/pdf/reve… Plugin: github.com/seekbytes/ptxN…

🤖🤖 Tried something fun today: asked Claude Code to create an agent team (an Implementer + a Planner) to implement the flashinfer mla paged decode CUDA kernel. The Implementer spent ~20 turns writing tests and debugging to use wgmma but kept getting stuck.😵‍💫😵‍💫😵‍💫 The Planner noticed the stagnation and (at its own decision) went off to carefully read the CUTLASS docs 📚, found the bug, and suggested the fix — the Implementer applied the fix and it worked immediately! Watching this kind of emergent coordination behaviour in an agent team is pretty interesting✨

1/ RELEASING AXLE: the Axiom Lean Engine ⚙️ We are serving our core Infrastructure for formal proving at scale. These are the same Lean metaprogramming tools that are behind AxiomProver, powering it to win Putnam and crack open research conjectures. Available to anyone today!

@YouJiacheng You can compile it with tvm-ffi and loading the .so in any languages.

Is it possible to extract the generated CuTeDSL and use it elsewhere (without torch & torch.compile)?

zml-smi running on TPU, Inferentia, NVIDIA and AMD same binary, because of course.

Asymmetric hardware scaling is here. Blackwell tensor cores are now so fast, exp2 and shared memory are the wall. FlashAttention-4 changes the algorithm & pipeline so that softmax & SMEM bandwidth no longer dictate speed. Attn reaches ~1600 TFLOPs, pretty much at matmul speed! joint work w/ Markus Hoehnerbach, Jay Shah(@ultraproduct), Timmy Liu, Vijay Thakkar (@__tensorcore__ ), Tri Dao (@tri_dao) 1/

@gaunernst @YouJiacheng github.com/pytorch/pytorc… We are actively working on fixing this :)

I've been working on a new LLM inference algorithm. It's called Speculative Speculative Decoding (SSD) and it's up to 2x faster than the strongest inference engines in the world. Collab w/ @tri_dao @avnermay. Details in thread.

I Fuzzed, and Vibe Fixed, the Vibed C Compiler john.regehr.org/writing/claude… by John Regehr (@regehr/116161100362503805" target="_blank" rel="nofollow noopener">mastodon.social/@regehr/116161…)

─ Worked for 59m 13s ───────────────────── • Context compacted • I’m noticing we have many untracked files, which is quite overwhelming. Let me git reset and undo everything you just fucking worked on for the last hour.

We’re excited to release TorchLean which is the first fully verified neural network framework in Lean. The Lean community has largely focused on pure mathematics. TorchLean expands this frontier toward verified neural network software and scientific computing. With the recent release of CSlib, we see this as another step toward a fully verified ML stack. We support features: 1. Executable IEEE-754 floating-point semantics (and extensible alternative FP models) verified tensor abstractions with precise shape/indexing semantics 2. Formally verified autograd system for differentiation of NN programs Proof-checked certification / verification algorithms like CROWN (robustness, bounds, etc.) 3. PyTorch-inspired modeling API with eager-style development + export/lowering to a shared IR for execution and verification Project page: leandojo.org/torchlean.html Paper: [2602.22631] TorchLean: Formalizing Neural Networks in Lean Work done @Robertljg, Jennifer Cruden, Xiangru Zhong, @huan_zhang12 and @AnimaAnandkumar. #MachineLearning #ScientificComputing #Lean

⚡️Introducing FlashSampling: Fast and Memory-Efficient Exact Sampling ⚡️ flashsampling.github.io/FlashSampling/… Keep pushing the Frontier of Open Research in Superintelligence!

with our new GB300NVL72 training, not only is the codebase completely TP free it is now also completely nccl and nvshmem free . its a beautiful thing.

Excited to announce Claude for Open Source ❤️ We're giving 6 months of free Claude Max 20x to open source maintainers and core contributors. If you maintain a popular project or contribute across open source, please apply! claude.com/contact-sales/…