Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

CMU PhD who built the kernels NVIDIA now ships in TensorRT-LLM explained fast attention in 68 minutes - better than $1200 GPU programming courses. pick the attention pattern -> generate a fused CUDA kernel -> drop it into vLLM/SGLang -> same GPU, way more tokens per second. That loop... show more

h100envy

1,296 subscribers

32,323 просмотров • 7 дней назад •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

PyTorch core engineer at Meta turned CUDA kernel writing into a sport in 13 minutes - better than $1500 GPU programming bootcamps. profile the kernel -> find the bottleneck -> rewrite -> benchmark -> merge the winning code into PyTorch. That loop is how the open community now beats hand-tuned vendor kernels. GPU MODE community + KernelBot competition + winning kernel merged into the framework - that's the stack. Watch it, then steal the loop below.

h100envy

35,148 просмотров • 9 дней назад

Ex-Berkeley PhD who leads SGLang at xAI explained how they serve Grok on 100K GPUs in 23 minutes - better than $2000 inference-at-scale courses. split prefill and decode -> shard experts across GPUs -> route tokens per expert -> overlap comm and compute -> serve at DeepSeek-API-killing prices. That loop is why xAI runs Grok on SGLang and third parties beat DeepSeek's own API by 5x on cost. SGLang + prefill-decode disaggregation + expert parallelism + AMD MI300 - that's the stack. Watch and save it, then read the article below.

Ex-Berkeley PhD who leads SGLang at xAI explained how they serve Grok on 100K GPUs in 23 minutes - better than $2000 inference-at-scale courses. split prefill and decode -> shard experts across GPUs -> route tokens per expert -> overlap comm and compute -> serve at DeepSeek-API-killing prices. That loop is why xAI runs Grok on SGLang and third parties beat DeepSeek's own API by 5x on cost. SGLang + prefill-decode disaggregation + expert parallelism + AMD MI300 - that's the stack. Watch and save it, then read the article below.

h100envy

312,260 просмотров • 3 дней назад

in the lectures below, i hold your hand through low-level LLM systems engineering. it includes everything up to TODAY! 1) pytorch tensors 2) large matmul on cpu vs gpu 3) JAX (and why xAI uses it instead of pytorch) 4) raw cuda kernels and global threading indexing 5) triton design philosophy and softmax example 6) HIP kernels 7) mapping out the ENTIRE ecosystem + differences between CUDA and ROCm/HIP (BLAS, FFT, DNN) 8) cutlass and cute-dsl 9) pretraining, finetuning, rl, unsloth, axolotl, megatron-lm, deepspeed, nanogpt, nanochat 10) training vs inference, inference serving problems, throughput vs latency vs concurrency scaling, vllm, sglang, tensorrt-llm, tensorrt, llama.cpp, exllamav2, exllamav3, benchmark comparisons 11) projects/companies using llms to generate SOTA cuda/triton kernels 12) luminal inference 13) mojo/modular/max

in the lectures below, i hold your hand through low-level LLM systems engineering. it includes everything up to TODAY! 1) pytorch tensors 2) large matmul on cpu vs gpu 3) JAX (and why xAI uses it instead of pytorch) 4) raw cuda kernels and global threading indexing 5) triton design philosophy and softmax example 6) HIP kernels 7) mapping out the ENTIRE ecosystem + differences between CUDA and ROCm/HIP (BLAS, FFT, DNN) 8) cutlass and cute-dsl 9) pretraining, finetuning, rl, unsloth, axolotl, megatron-lm, deepspeed, nanogpt, nanochat 10) training vs inference, inference serving problems, throughput vs latency vs concurrency scaling, vllm, sglang, tensorrt-llm, tensorrt, llama.cpp, exllamav2, exllamav3, benchmark comparisons 11) projects/companies using llms to generate SOTA cuda/triton kernels 12) luminal inference 13) mojo/modular/max

Elliot Arledge

57,855 просмотров • 8 месяцев назад

I implemented Google Research's TurboQuant as a CUDA-native compression engine on Blackwell B200. 5x KV cache compression on Qwen 2.5-1.5B, near-loseless attention scores, generating live from compressed memory. 5 custom cuTile CUDA kernels ft: - fused attention (with QJL corrections) - online softmax -on-chip cache decompression - pipelined TMA loads Try it out: s/o Bryce, the CUDA Colonel and the cuTile team at NVIDIA for lending me Blackwell GPU access :) cc sunny madra Gavin

I implemented Google Research's TurboQuant as a CUDA-native compression engine on Blackwell B200. 5x KV cache compression on Qwen 2.5-1.5B, near-loseless attention scores, generating live from compressed memory. 5 custom cuTile CUDA kernels ft: - fused attention (with QJL corrections) - online softmax -on-chip cache decompression - pipelined TMA loads Try it out: s/o Bryce, the CUDA Colonel and the cuTile team at NVIDIA for lending me Blackwell GPU access :) cc sunny madra Gavin

ani

807,497 просмотров • 3 месяцев назад

Open-source software never stops. It only accelerates. Dynamo, SGLang, TensorRT LLM, and vLLM are constantly optimized by a vast ecosystem of developers building on top of the NVIDIA platform. The result: your token output keeps improving and token cost keeps decreasing on the same hardware resources while your developer velocity stays at its peak. Build on the foundation continuously optimized by the world’s best developers. ⚡ 🔗

Open-source software never stops. It only accelerates. Dynamo, SGLang, TensorRT LLM, and vLLM are constantly optimized by a vast ecosystem of developers building on top of the NVIDIA platform. The result: your token output keeps improving and token cost keeps decreasing on the same hardware resources while your developer velocity stays at its peak. Build on the foundation continuously optimized by the world’s best developers. ⚡ 🔗

NVIDIA

64,521 просмотров • 3 месяцев назад

Among the fastest DeepSeek V3.2, MiniMax-M2.5, and Qwen 3.5 397B inference in the market, per Artificial Analysis benchmarks (April 2026). ⚡️🤖 Sub-1-second TTFT. 230 tokens per second. Co-designed every layer of the stack with Inferact, performance optimized vLLM, all on NVIDIA HGX B300. Live on DigitalOcean Serverless Inference now. Full breakdown in the comments. ⬇️

Among the fastest DeepSeek V3.2, MiniMax-M2.5, and Qwen 3.5 397B inference in the market, per Artificial Analysis benchmarks (April 2026). ⚡️🤖 Sub-1-second TTFT. 230 tokens per second. Co-designed every layer of the stack with Inferact, performance optimized vLLM, all on NVIDIA HGX B300. Live on DigitalOcean Serverless Inference now. Full breakdown in the comments. ⬇️

DigitalOcean

39,063 просмотров • 2 месяцев назад

Fully local Code Assistant running on NVIDIA GPU! In this tutorial, I'll show you how to run Llama3 using TensorRT and Nvidia's Triton Inference Server to use it as a Code Assistant in VSCode In this thread 🧵, I'll walk you through the integration process, explaining each step simply and clearly 👇 NVIDIA AI NVIDIA AIDev AI at Meta

Fully local Code Assistant running on NVIDIA GPU! In this tutorial, I'll show you how to run Llama3 using TensorRT and Nvidia's Triton Inference Server to use it as a Code Assistant in VSCode In this thread 🧵, I'll walk you through the integration process, explaining each step simply and clearly 👇 NVIDIA AI NVIDIA AIDev AI at Meta

Daniel San

42,154 просмотров • 2 лет назад

YOUR PARENTS PAID FOR THE CUDA MOAT! The #1 contributor to the CUDA MOAT isn't the the developers at NVIDIA, but it is the millions of developers outside of NVIDIA that invent new algorithms for CUDA like Flash Attention. For most of them, it started with an GeForce gaming GPU. NVIDIA is the only companies that has an reasonable good developer stack on consumer grade GPUs. As people grow up beyond playing CSGO & League of Legends & Minecraft, they either become anime weeaboos or they start programming on their existing computer with has an GeForce GPU

YOUR PARENTS PAID FOR THE CUDA MOAT! The #1 contributor to the CUDA MOAT isn't the the developers at NVIDIA, but it is the millions of developers outside of NVIDIA that invent new algorithms for CUDA like Flash Attention. For most of them, it started with an GeForce gaming GPU. NVIDIA is the only companies that has an reasonable good developer stack on consumer grade GPUs. As people grow up beyond playing CSGO & League of Legends & Minecraft, they either become anime weeaboos or they start programming on their existing computer with has an GeForce GPU

SemiAnalysis

25,230 просмотров • 3 месяцев назад

Alibaba engineer who leads Qwen explained the future of open agent models in 25 minutes - better than $2000 LLM training courses. pre-train the base ->SFT -> RLHF -> tool use -> multi-modal -> ship a whole family (chat / VL / coder / math / QwQ). That loop is why Qwen quietly became the most downloaded open model family on Hugging Face. Qwen base + Qwen-VL + Qwen-Coder + QwQ reasoning - that's the stack. Watch and save it, then read the article below.

Alibaba engineer who leads Qwen explained the future of open agent models in 25 minutes - better than $2000 LLM training courses. pre-train the base ->SFT -> RLHF -> tool use -> multi-modal -> ship a whole family (chat / VL / coder / math / QwQ). That loop is why Qwen quietly became the most downloaded open model family on Hugging Face. Qwen base + Qwen-VL + Qwen-Coder + QwQ reasoning - that's the stack. Watch and save it, then read the article below.

h100envy

111,448 просмотров • 5 дней назад

I trained a 12M parameter LLM on my own ML framework using a Rust backend and CUDA kernels for flash attention, AdamW, and more. Wrote the full transformer architecture, and BPE tokenizer from scratch. The framework features: - Custom CUDA kernels (Flash Attention, fused LayerNorm, fused GELU) for 3x increased throughput - Automatic WebGPU fallback for non-NVIDIA devices - TypeScript API with Rust compute backend - One npm install to get started, prebuilt binaries for every platform Try out the model for yourself: Built with Reese Chong. Check out the repos and blog if you want to learn more. Shoutout to Modal for the compute credits allowing me to train on 2 A100 GPUs without going broke cc sunny madra Gavin

I trained a 12M parameter LLM on my own ML framework using a Rust backend and CUDA kernels for flash attention, AdamW, and more. Wrote the full transformer architecture, and BPE tokenizer from scratch. The framework features: - Custom CUDA kernels (Flash Attention, fused LayerNorm, fused GELU) for 3x increased throughput - Automatic WebGPU fallback for non-NVIDIA devices - TypeScript API with Rust compute backend - One npm install to get started, prebuilt binaries for every platform Try out the model for yourself: Built with Reese Chong. Check out the repos and blog if you want to learn more. Shoutout to Modal for the compute credits allowing me to train on 2 A100 GPUs without going broke cc sunny madra Gavin

Aadi Kulshrestha

812,043 просмотров • 2 месяцев назад

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

Avi Chawla

265,992 просмотров • 13 дней назад

🎉 Congrats to MiniMax (official) on releasing MiniMax M3! Frontier coding and agentic capabilities, native image and video input, computer use, and a 1M-token context window, all in a single open model. At the heart of M3 is MSA, a new sparse attention architecture: instead of attending densely over the full KV cache, each query scores 128-token KV blocks and runs attention only over the top blocks. That is what makes 1M-token context practical to serve. M3 runs in vLLM with day-0 support, verified on NVIDIA and AMD hardware: ✨ MSA sparse attention with dedicated prefill and decode kernels ✨ 1M-token context serving with prefix caching and chunked prefill ✨ BF16 and MXFP8 checkpoints, with MoE backends for both Hopper and Blackwell ✨ Native multimodal input (image + video) ✨ Tool calling, reasoning parsing, and thinking-mode control for agent workloads Day-0 support like this is a true team effort. Grateful to the teams at MiniMax (official), NVIDIA AI, AI at AMD, and Inferact, and to the vLLM community for making it happen. 🙏 Deep dive into the implementation, kernel work, and deployment recipes: 🔗

🎉 Congrats to MiniMax (official) on releasing MiniMax M3! Frontier coding and agentic capabilities, native image and video input, computer use, and a 1M-token context window, all in a single open model. At the heart of M3 is MSA, a new sparse attention architecture: instead of attending densely over the full KV cache, each query scores 128-token KV blocks and runs attention only over the top blocks. That is what makes 1M-token context practical to serve. M3 runs in vLLM with day-0 support, verified on NVIDIA and AMD hardware: ✨ MSA sparse attention with dedicated prefill and decode kernels ✨ 1M-token context serving with prefix caching and chunked prefill ✨ BF16 and MXFP8 checkpoints, with MoE backends for both Hopper and Blackwell ✨ Native multimodal input (image + video) ✨ Tool calling, reasoning parsing, and thinking-mode control for agent workloads Day-0 support like this is a true team effort. Grateful to the teams at MiniMax (official), NVIDIA AI, AI at AMD, and Inferact, and to the vLLM community for making it happen. 🙏 Deep dive into the implementation, kernel work, and deployment recipes: 🔗

vLLM

40,259 просмотров • 27 дней назад

Jensen spent 2 min on InferenceMAX at the GTC keynote 😎 InferenceMAX is completely open source and completely impartial, with the support of AMD, OpenAI, Nvidia, Microsoft, Oracle, Pytorch, SGLang, vLLM, Cursoe, Nebius, Tensorwave, etc

Jensen spent 2 min on InferenceMAX at the GTC keynote 😎 InferenceMAX is completely open source and completely impartial, with the support of AMD, OpenAI, Nvidia, Microsoft, Oracle, Pytorch, SGLang, vLLM, Cursoe, Nebius, Tensorwave, etc

Dylan Patel

86,497 просмотров • 8 месяцев назад

PowerInfer - a high-speed inference engine for deploying LLMs locally. Just came across this super interesting project on speeding up inference. It's not MoE but it's a simple approach that exploits the high locality in LLM inference to design a GPU-CPU hybrid inference engine. Hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons (the majority) are computed on the CPU. This approach significantly reduces GPU memory demands and CPU-GPU data transfer. It achieves an average token generation rate of 13.20 tokens/s, with a peak of 29.08 tokens/s, across various LLMs on a single NVIDIA RTX 4090 GPU. It's on only 18% lower than that achieved by a top-tier server-grade A100 GPU. It also significantly outperforms llama.cpp by up to 11.69x while retaining model accuracy. There is a lot more innovation around inference that's coming fast. Really encouraged by the study on sparse computation to enhance the computational efficiency of LLMs. It's now possible to use PowerInfer with Llama 2 and Faclon 40B. Mistral-7B support is coming soon!

PowerInfer - a high-speed inference engine for deploying LLMs locally. Just came across this super interesting project on speeding up inference. It's not MoE but it's a simple approach that exploits the high locality in LLM inference to design a GPU-CPU hybrid inference engine. Hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons (the majority) are computed on the CPU. This approach significantly reduces GPU memory demands and CPU-GPU data transfer. It achieves an average token generation rate of 13.20 tokens/s, with a peak of 29.08 tokens/s, across various LLMs on a single NVIDIA RTX 4090 GPU. It's on only 18% lower than that achieved by a top-tier server-grade A100 GPU. It also significantly outperforms llama.cpp by up to 11.69x while retaining model accuracy. There is a lot more innovation around inference that's coming fast. Really encouraged by the study on sparse computation to enhance the computational efficiency of LLMs. It's now possible to use PowerInfer with Llama 2 and Faclon 40B. Mistral-7B support is coming soon!

elvis

261,622 просмотров • 2 лет назад

We co-led Inferact's $150M seed round to support them in their mission to build the inference engine for all current and future AI. In this episode of The Investment Memo, Lightspeed's Bucky Moore and James Alcorn sit down with Simon Mo (Co-Founder & CEO Inferact) to cover: - How vLLM grew to 60K+ GitHub stars - Why inference is shifting to the majority of compute - How vLLM evolved from a research project into the industry standard - Why building a company was the next step to push open-source inference forward 00:00 Introduction 02:03 The investment memo 04:47 Latency vs throughput vs cost 06:19 Paged attention explained 08:04 The evolution of attention 09:42 Growing the vLLM open source community 11:41 Working with hardware vendors 14:45 Deploying vLLM at large scale 16:03 Inferact's culture of openness 18:45 Building an open ecosystem and horizontal stack 19:45 Inferact's approach to fundraising 22:14 What is the future of inference? Simon Mo Bucky Moore James Alcorn

We co-led Inferact's $150M seed round to support them in their mission to build the inference engine for all current and future AI. In this episode of The Investment Memo, Lightspeed's Bucky Moore and James Alcorn sit down with Simon Mo (Co-Founder & CEO Inferact) to cover: - How vLLM grew to 60K+ GitHub stars - Why inference is shifting to the majority of compute - How vLLM evolved from a research project into the industry standard - Why building a company was the next step to push open-source inference forward 00:00 Introduction 02:03 The investment memo 04:47 Latency vs throughput vs cost 06:19 Paged attention explained 08:04 The evolution of attention 09:42 Growing the vLLM open source community 11:41 Working with hardware vendors 14:45 Deploying vLLM at large scale 16:03 Inferact's culture of openness 18:45 Building an open ecosystem and horizontal stack 19:45 Inferact's approach to fundraising 22:14 What is the future of inference? Simon Mo Bucky Moore James Alcorn

Lightspeed

27,040 просмотров • 5 месяцев назад

Introducing The AI CUDA Engineer: An agentic AI system that automates the production of highly optimized CUDA kernels. The AI CUDA Engineer can produce highly optimized CUDA kernels, reaching 10-100x speedup over common machine learning operations in PyTorch. Our system is also able to produce highly optimized CUDA kernels that are much faster than existing CUDA kernels commonly used in production. We believe that fundamentally, AI systems can and should be as resource-efficient as the human brain, and that the best path to achieve this efficiency is to use AI to make AI more efficient! We are excited to publish our paper, The AI CUDA Engineer: Agentic CUDA Kernel Discovery, Optimization and Composition. We also release a dataset of over 17,000 verified CUDA kernels produced by The AI CUDA Engineer. Paper: Kernel Archive Webpage: HuggingFace Dataset: The AI CUDA Engineer utilizes evolutionary LLM-driven code optimization to autonomously improve the runtime of machine learning operations. Our system is not only able to convert PyTorch code into CUDA kernels, but through the use of evolution, it can also optimize the runtime performance of CUDA kernels, fuse multiple operations, and even discover novel solutions for writing efficient CUDA operations by learning from past innovations! We believe The AI CUDA Engineer opens a new era of AI-driven acceleration of AI and automated inference time optimization. We (Robert Lange, Aaditya Prasad 🇺🇸, Suuun, Maxence Faldor, Yujin Tang, hardmaru) are excited to continue Sakana AI's mission of leveraging AI to improve AI.

Introducing The AI CUDA Engineer: An agentic AI system that automates the production of highly optimized CUDA kernels. The AI CUDA Engineer can produce highly optimized CUDA kernels, reaching 10-100x speedup over common machine learning operations in PyTorch. Our system is also able to produce highly optimized CUDA kernels that are much faster than existing CUDA kernels commonly used in production. We believe that fundamentally, AI systems can and should be as resource-efficient as the human brain, and that the best path to achieve this efficiency is to use AI to make AI more efficient! We are excited to publish our paper, The AI CUDA Engineer: Agentic CUDA Kernel Discovery, Optimization and Composition. We also release a dataset of over 17,000 verified CUDA kernels produced by The AI CUDA Engineer. Paper: Kernel Archive Webpage: HuggingFace Dataset: The AI CUDA Engineer utilizes evolutionary LLM-driven code optimization to autonomously improve the runtime of machine learning operations. Our system is not only able to convert PyTorch code into CUDA kernels, but through the use of evolution, it can also optimize the runtime performance of CUDA kernels, fuse multiple operations, and even discover novel solutions for writing efficient CUDA operations by learning from past innovations! We believe The AI CUDA Engineer opens a new era of AI-driven acceleration of AI and automated inference time optimization. We (Robert Lange, Aaditya Prasad 🇺🇸, Suuun, Maxence Faldor, Yujin Tang, hardmaru) are excited to continue Sakana AI's mission of leveraging AI to improve AI.

Sakana AI

1,149,339 просмотров • 1 год назад

You can run CUDA, on a Mac ARM GPU, in the browser. It sounds ridiculous but it actually works. HipScript chains CUDA, to OpenCL, to Vulkan, to Tint (Google’s shader translator), to a WASM WebGPU. I got a plasma simulation in running in just a few minutes, no NVIDIA GPU!

You can run CUDA, on a Mac ARM GPU, in the browser. It sounds ridiculous but it actually works. HipScript chains CUDA, to OpenCL, to Vulkan, to Tint (Google’s shader translator), to a WASM WebGPU. I got a plasma simulation in running in just a few minutes, no NVIDIA GPU!

LaurieWired

160,223 просмотров • 1 год назад

Gemma 4 Diffusion landed in vLLM last week. Day 0. First diffusion LLM natively supported in vLLM. Instead of one token at a time, it predicts 256 tokens at once and iteratively denoises them in parallel. Result: 1,000+ tokens per second at batch size 1 on a single H100. Built on Model Runner V2. Google Gemma

Gemma 4 Diffusion landed in vLLM last week. Day 0. First diffusion LLM natively supported in vLLM. Instead of one token at a time, it predicts 256 tokens at once and iteratively denoises them in parallel. Result: 1,000+ tokens per second at batch size 1 on a single H100. Built on Model Runner V2. Google Gemma

Red Hat AI

17,524 просмотров • 24 дней назад

Today we announced a $150M seed round in Inferact, a new startup led by the maintainers of the vLLM project. Inferact will support the vLLM open source project through dedicated financial and developer resources and build what they see as the next generation commercial inference engine. Cofounders Simon Mo and Woosuk Kwon joined a16z GP Matt Bornstein for a conversation on how vLLM came to be, what they’ve learned from building it, and what they’re now doing at Inferact. They cover how vLLM began as a side project, why open source is essential to the world’s AI infrastructure, why inference is getting harder, and more. 00:00 Introduction 11:41 Community and collaboration in vLLM 19:19 Understanding inference engines 24:27 Cluster scale and GPU deployment 31:19 Belief in open source AI 35:45 Founding Inferact 40:00 The future of Inference at scale Simon Mo Woosuk Kwon Matt Bornstein

Today we announced a $150M seed round in Inferact, a new startup led by the maintainers of the vLLM project. Inferact will support the vLLM open source project through dedicated financial and developer resources and build what they see as the next generation commercial inference engine. Cofounders Simon Mo and Woosuk Kwon joined a16z GP Matt Bornstein for a conversation on how vLLM came to be, what they’ve learned from building it, and what they’re now doing at Inferact. They cover how vLLM began as a side project, why open source is essential to the world’s AI infrastructure, why inference is getting harder, and more. 00:00 Introduction 11:41 Community and collaboration in vLLM 19:19 Understanding inference engines 24:27 Cluster scale and GPU deployment 31:19 Belief in open source AI 35:45 Founding Inferact 40:00 The future of Inference at scale Simon Mo Woosuk Kwon Matt Bornstein

a16z

87,993 просмотров • 5 месяцев назад