Dylan Lim

Introducing the Together Compiler: continual compilation for AI inference! I’m super proud to finally share this work from our research team! ✨ Today’s inference is “Polymorphic” or shape-shifting, it doesn’t sit still. On OpenRouter, 80% of traffic goes to models released in the last three months and the top ten most popular open models each use a different architecture. Our customers constantly change their usage behaviors. Together Compiler leverages agents to continually optimize our deployments immediately as these changes occur. Together Compiler research highlights include: 1. 252x better throughput: TKCC lets agents try 252x more optimization experiments per hour compared to prior agent inference optimization systems. 2. State-of-the-art TPS across InferenceX workloads we tried: TKCC outperforms the best InferenceX engine across all tested settings in just 24 hours & 2 nodes per compile. Including 1111 TPS on MiniMax M3.1 B200 and 1205 TPS on DeepSeek v4.1 Flash GB300 at batch size 1. NVIDIA, AMD, Trainium2. 3. Continuous recompilation: TKCC currently serves many trillions of production tokens per day and adapts to traffic shifts within just a few hours of recompilation. 4. Bitter lessons: TKCC’s design builds on 2+ years of building frameworks to accelerate AI Inference, starting from ThunderKittens and Megakernels. Compilers have always used abstractions and IRs, but we show that agentic compilers now reach higher performance ceilings without abstractions. From an incredible team @reyna_abhyankar @dylan__lim @realDanFu @togethercompute!!

we got early access and hipkittens is now running on amd helios mi455! here we share an educational gemm ladder to demonstrate how to use the new helios features. overall, we found that the patterns identified in hipkittens for writing performant mi350/355 kernels translate well to helios making the forward port quite smooth!

thunderkittens is now running on nvidia vera rubin! our kernels team got early access to nvl72 and spent the past few days digging through the new isa and bringing nvfp4 + fp8 gemms to life after reworking the kernels for rubin, we pushed them past 22 and 12 pflops respectively — competitive with cublas + cute dsl

(1/8) We’re releasing an 8-GPU Llama-70B inference engine megakernel! Our megakernel supports arbitrary batch sizes, mixed prefill+decode, a paged KV cache, instruction pipelining, dynamic scheduling, interleaved communication, and more! On ShareGPT it’s 22% faster than SGLang.

(1/6) We’re happy to share that ThunderKittens now supports writing multi-GPU kernels, with the same programming model and full compatibility with PyTorch + torchrun. We’re also releasing collective ops and fused multi-GPU GEMM kernels, up to 2.6x faster than PyTorch + NCCL. (Joint with @dylan__lim, @spectorb, and @HazyResearch. Generously supported by @cursor_ai)

(1/5) We’ve never enjoyed watching people chop Llamas into tiny pieces. So, we’re excited to be releasing our Low-Latency-Llama Megakernel! We run the whole forward pass in single kernel. Megakernels are faster & more humane. Here’s how to treat your Llamas ethically: (Joint with @jordanjuravsky, @stuart_sul, @OwenDugan, @dylan__lim, @realDanFu, @simran_s_arora, and @HazyResearch)

Spatial reasoning is a major challenge for the foundation models today, even in simple tasks like arranging objects in 3D space. #CVPR2025 Introducing LayoutVLM, a differentiable optimization framework that uses VLM to spatially reason about diverse scene layouts from unlabeled assets and open-ended language instructions 1/n

(1/5) @CKT_Conner, @dill_pkl, @emilyzsh, and I are excited to introduce Shard - a proof-of-concept for an infinitely scalable distributed system composed of consumer hardware for training and running ML models! Features: - Data + Pipeline Parallel for handling arbitrarily large models - Algorithmic load balancing for throughput optimization - Fault tolerance for unreliable machines