Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Interested in optimizing Mixture of Experts LLM inference on ARM CPUs? In his talk at PyTorch Conference North America, Maajid Khan from Fujitsu Research India will explore efficient MoE LLM inference using vLLM and OpenVINO, sharing practical strategies for running these models effectively on ARM architectures. Register and join... show more

PyTorch

514,845 subscribers

16,448 views • 6 days ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

2 Comments

Fajar M Reza6 days ago

ARM MoE benefits from sparse activation, but memory traffic still shapes latency.

Sutton6 days ago

MoE is honestly the best case for CPU inference. you need a lot of RAM for the total params, which CPUs have, but each token only touches the active experts, so memory bandwidth stops being such a wall. curious how much of the speedup comes from the kernels vs just smarter expert placement

Related Videos

At PyTorch Conference North America 2026, hear directly from and connect with the people working on PyTorch, vLLM, DeepSpeed, ray, Helion, and Safetensors, alongside experts working across the AI stack. Keynote speaker Simon Mo (vLLM, Inferact) says vLLM’s goal is to become “the easiest to use and most efficient inference engine.” At #PyTorchCon NA 2025, he spoke about the difficulty and cost of using large language models and continuing to push inference efficiency forward. PyTorch conferences are the open source AI community’s town square, where what’s next gets decided. Register by September 4 to save on your conference pass:

At PyTorch Conference North America 2026, hear directly from and connect with the people working on PyTorch, vLLM, DeepSpeed, ray, Helion, and Safetensors, alongside experts working across the AI stack. Keynote speaker Simon Mo (vLLM, Inferact) says vLLM’s goal is to become “the easiest to use and most efficient inference engine.” At #PyTorchCon NA 2025, he spoke about the difficulty and cost of using large language models and continuing to push inference efficiency forward. PyTorch conferences are the open source AI community’s town square, where what’s next gets decided. Register by September 4 to save on your conference pass:

PyTorch

22,334 views • 1 month ago

At #PyTorchCon North America 2026, Liz Li and Shekhar of AMD will discuss their work enabling scalable MXFP8 pretraining on MI355X GPUs. Liz and Shekhar will cover optimization of key MXFP8 operators in TorchAO and how those kernels were integrated with TorchTitan to build an end-to-end upstream PyTorch training stack for large language models. Their session, “Scaling MXFP8 Pretraining on 1K+ AMD Instinct MI355X: TorchAO Kernels and TorchTitan Training,” will also cover Triton and FlyDSL implementations and lessons from numerical validation, convergence, and performance tuning. Join us in San Jose on October 20-21. Register today:

At #PyTorchCon North America 2026, Liz Li and Shekhar of AMD will discuss their work enabling scalable MXFP8 pretraining on MI355X GPUs. Liz and Shekhar will cover optimization of key MXFP8 operators in TorchAO and how those kernels were integrated with TorchTitan to build an end-to-end upstream PyTorch training stack for large language models. Their session, “Scaling MXFP8 Pretraining on 1K+ AMD Instinct MI355X: TorchAO Kernels and TorchTitan Training,” will also cover Triton and FlyDSL implementations and lessons from numerical validation, convergence, and performance tuning. Join us in San Jose on October 20-21. Register today:

PyTorch

16,137 views • 7 days ago

"It's really great to feel the energy and see all the collaboration that's happening, lots of conversations. We have people here who are just starting to learn about PyTorch, and we have people who are maintainers. It's super nice to see that energy and see how together we're going to be able to solve some of these very difficult problems for humanity." Ankit Patel, NVIDIA, at PyTorch Conference Europe 2026 Register today to join the open source AI community at PyTorch Conference North America in San Jose, October 20–21: #PyTorchCon

"It's really great to feel the energy and see all the collaboration that's happening, lots of conversations. We have people here who are just starting to learn about PyTorch, and we have people who are maintainers. It's super nice to see that energy and see how together we're going to be able to solve some of these very difficult problems for humanity." Ankit Patel, NVIDIA, at PyTorch Conference Europe 2026 Register today to join the open source AI community at PyTorch Conference North America in San Jose, October 20–21: #PyTorchCon

PyTorch

17,000 views • 18 days ago

New chips are shipping faster than ever before, but the ecosystem is being held back by having to rewrite and then re-debug the same code over & over again. Chris Lattner CEO and Co-founder at Modular and EVP of Advanced AI Software & Platforms at Qualcomm, will deliver a keynote at PyTorch Conference North America about an open software platform for heterogeneous compute powered by Mojo and MAX. PyTorch has always been the place where the best models come together, and now there's a way to get those models onto all kinds of hardware. If you're interested in Al and compute, join us at the PyTorch Conference in San Jose, CA. Register now: #PyTorchCon

New chips are shipping faster than ever before, but the ecosystem is being held back by having to rewrite and then re-debug the same code over & over again. Chris Lattner CEO and Co-founder at Modular and EVP of Advanced AI Software & Platforms at Qualcomm, will deliver a keynote at PyTorch Conference North America about an open software platform for heterogeneous compute powered by Mojo and MAX. PyTorch has always been the place where the best models come together, and now there's a way to get those models onto all kinds of hardware. If you're interested in Al and compute, join us at the PyTorch Conference in San Jose, CA. Register now: #PyTorchCon

PyTorch

20,090 views • 16 days ago

At PyTorch Conference North America 2026, hear directly from and connect with the people working on PyTorch, vLLM, DeepSpeed, ray, Helion, and Safetensors, alongside experts working across the AI stack. At #PyTorchCon NA 2025, Richard Liaw (ray, Anyscale) spoke about the Ray team’s focus on building something “useful and critical to the AI infrastructure space” and about “how important Ray can be for the stack.” PyTorch conferences are the open source AI community’s town square, where what’s next gets decided. Register by September 4 (tomorrow) to save on your conference pass: #PyTorchCon

At PyTorch Conference North America 2026, hear directly from and connect with the people working on PyTorch, vLLM, DeepSpeed, ray, Helion, and Safetensors, alongside experts working across the AI stack. At #PyTorchCon NA 2025, Richard Liaw (ray, Anyscale) spoke about the Ray team’s focus on building something “useful and critical to the AI infrastructure space” and about “how important Ray can be for the stack.” PyTorch conferences are the open source AI community’s town square, where what’s next gets decided. Register by September 4 (tomorrow) to save on your conference pass: #PyTorchCon

PyTorch

16,342 views • 28 days ago

The open source AI community thrives on collaboration, choice, and open contribution. As Joseph Groenenboom from Red Hat highlights, empowering developers to freely contribute and build options is essential for growing the entire AI ecosystem. PyTorch Conference North America 2026 is where that community comes together in person. Join us to collaborate with open source pioneers, expand your deployment capabilities, and directly contribute to shaping the future of machine learning. Register for PyTorch Conference North America today: #PyTorchCon

The open source AI community thrives on collaboration, choice, and open contribution. As Joseph Groenenboom from Red Hat highlights, empowering developers to freely contribute and build options is essential for growing the entire AI ecosystem. PyTorch Conference North America 2026 is where that community comes together in person. Join us to collaborate with open source pioneers, expand your deployment capabilities, and directly contribute to shaping the future of machine learning. Register for PyTorch Conference North America today: #PyTorchCon

PyTorch

17,621 views • 19 days ago

Build a neural network from scratch, train it, and optimize it for real-world AI infrastructure during our PyTorch Associate Training the day before PyTorch Conference North America on October 19th. Your instructor, Faradawn Yang of NVIDIA AI explains how this training will give you hands on experience with the latest AI accelerators & prepare you for the PyTorch Certified Associate (PTCA) certification exam. #PyTorchCon

Build a neural network from scratch, train it, and optimize it for real-world AI infrastructure during our PyTorch Associate Training the day before PyTorch Conference North America on October 19th. Your instructor, Faradawn Yang of NVIDIA AI explains how this training will give you hands on experience with the latest AI accelerators & prepare you for the PyTorch Certified Associate (PTCA) certification exam. #PyTorchCon

PyTorch

17,476 views • 3 days ago

🚨 YOUR GPU IS PROBABLY WASTING MORE THAN YOU THINK. vLLM is built to squeeze far more useful work out of your GPU when serving LLMs. Running an LLM at scale isn’t just about having a powerful GPU. The real problem is how efficiently you use its memory and compute. That’s where vLLM comes in. → High-throughput LLM inference and serving → PagedAttention for smarter KV-cache memory management → Continuous batching to keep GPUs busy → Prefix caching + chunked prefill → OpenAI-compatible API out of the box → Supports a huge range of modern LLM architectures → Quantization support for running models more efficiently And the crazy part? You can start an OpenAI-compatible inference server with: `vllm serve ` So your application can talk to your own model almost like it’s talking to OpenAI. The bigger idea: Don’t just buy more GPUs. Make the GPUs you already have work harder. That’s why vLLM has become such a major project in LLM inference. 🔥 #vLLM #AI #LLM #Inference #GPU #MachineLearning #AIInfrastructure #OpenSource #AIAgents #Developers

🚨 YOUR GPU IS PROBABLY WASTING MORE THAN YOU THINK. vLLM is built to squeeze far more useful work out of your GPU when serving LLMs. Running an LLM at scale isn’t just about having a powerful GPU. The real problem is how efficiently you use its memory and compute. That’s where vLLM comes in. → High-throughput LLM inference and serving → PagedAttention for smarter KV-cache memory management → Continuous batching to keep GPUs busy → Prefix caching + chunked prefill → OpenAI-compatible API out of the box → Supports a huge range of modern LLM architectures → Quantization support for running models more efficiently And the crazy part? You can start an OpenAI-compatible inference server with: `vllm serve ` So your application can talk to your own model almost like it’s talking to OpenAI. The bigger idea: Don’t just buy more GPUs. Make the GPUs you already have work harder. That’s why vLLM has become such a major project in LLM inference. 🔥 #vLLM #AI #LLM #Inference #GPU #MachineLearning #AIInfrastructure #OpenSource #AIAgents #Developers

Vikas gupta

14,138 views • 20 days ago

Liang Zhang (Huawei) delivers “Ascend & PyTorch: Pioneer New AI Open Ecosystem” at PyTorch Conference China 2026. Drawing on eight years of open source practice, Liang Zhang looks at how Ascend’s collaboration with the PyTorch community is evolving from hardware adaptation to joint research on technical standards. The keynote also covers next-generation AI hardware-software codesign, interoperability between Chinese and global AI technology stacks, and building an open and sustainable AI industrial ecosystem. Here's Liang Zhang, introduced by Mark Collier 柯理怀: #PyTorchCon

Liang Zhang (Huawei) delivers “Ascend & PyTorch: Pioneer New AI Open Ecosystem” at PyTorch Conference China 2026. Drawing on eight years of open source practice, Liang Zhang looks at how Ascend’s collaboration with the PyTorch community is evolving from hardware adaptation to joint research on technical standards. The keynote also covers next-generation AI hardware-software codesign, interoperability between Chinese and global AI technology stacks, and building an open and sustainable AI industrial ecosystem. Here's Liang Zhang, introduced by Mark Collier 柯理怀: #PyTorchCon

PyTorch

16,367 views • 23 days ago

TOMORROW 🎉 Join us for the virtual premiere of the PyTorch Documentary! We'll be live on YouTube at 10am PT on June 25 along with some of our key PyTorch collaborators from AMD, AWS, Google, Meta and Microsoft 💜 🔗

TOMORROW 🎉 Join us for the virtual premiere of the PyTorch Documentary! We'll be live on YouTube at 10am PT on June 25 along with some of our key PyTorch collaborators from AMD, AWS, Google, Meta and Microsoft 💜 🔗

PyTorch

48,761 views • 2 years ago

PyTorch 2.13 brings FlexAttention to Apple Silicon, cuts peak memory by up to 4× for large-vocabulary models with nn.LinearCrossEntropyLoss, and updates distributed training, compilation, profiling, and on-device inference. Our live Q&A examined CUDA version support, CuTeDSL in TorchInductor, Python 3.15, ExecuTorch, torch.compile, and plans for PyTorch 2.14. Andrey Talman (Meta), albanD (Meta), and Piotr Bialecki (NVIDIA) joined moderator Chris Gottbrath (Gottbrath Technologies). This clip explains how FlexAttention compiles kernels specialized for the requested masking pattern, allowing them to skip unnecessary computation. It also covers speedups of up to approximately 12× over SDPA in tested sparse configurations and a deterministic backward path on CUDA that removes one source of nondeterminism in gradient computation. 🔗 Watch the full PyTorch 2.13 Release Live Q&A:

PyTorch 2.13 brings FlexAttention to Apple Silicon, cuts peak memory by up to 4× for large-vocabulary models with nn.LinearCrossEntropyLoss, and updates distributed training, compilation, profiling, and on-device inference. Our live Q&A examined CUDA version support, CuTeDSL in TorchInductor, Python 3.15, ExecuTorch, torch.compile, and plans for PyTorch 2.14. Andrey Talman (Meta), albanD (Meta), and Piotr Bialecki (NVIDIA) joined moderator Chris Gottbrath (Gottbrath Technologies). This clip explains how FlexAttention compiles kernels specialized for the requested masking pattern, allowing them to skip unnecessary computation. It also covers speedups of up to approximately 12× over SDPA in tested sparse configurations and a deterministic backward path on CUDA that removes one source of nondeterminism in gradient computation. 🔗 Watch the full PyTorch 2.13 Release Live Q&A:

PyTorch

18,591 views • 2 months ago

Learn how to build an optimized LLM inference system from the ground up in our new short course, Efficiently Serving LLMs, built in collaboration with Predibase by Rubrik and taught by Travis Addair. Whether you're serving your own LLM or using a model hosting service, this course will give you a deep understanding of the optimizations required to efficiently serve many users at once. - Learn how LLMs generate text one token at a time, and how techniques like KV caching, continuous batching, and quantization speed things up and optimize memory usage for serving multiple users. - Benchmark the performance of these LLM optimizations to explore the trade-offs between quickly responding to an individual user’s request vs. serving many users at once. - Use techniques like low-rank adaptation (LoRA) to efficiently serve hundreds of unique, custom fine-tuned models on a single device, without sacrificing throughput. - Use Predibase's LoRAX framework to see optimization techniques in action on a real LLM server. Sign up here:

Learn how to build an optimized LLM inference system from the ground up in our new short course, Efficiently Serving LLMs, built in collaboration with Predibase by Rubrik and taught by Travis Addair. Whether you're serving your own LLM or using a model hosting service, this course will give you a deep understanding of the optimizations required to efficiently serve many users at once. - Learn how LLMs generate text one token at a time, and how techniques like KV caching, continuous batching, and quantization speed things up and optimize memory usage for serving multiple users. - Benchmark the performance of these LLM optimizations to explore the trade-offs between quickly responding to an individual user’s request vs. serving many users at once. - Use techniques like low-rank adaptation (LoRA) to efficiently serve hundreds of unique, custom fine-tuned models on a single device, without sacrificing throughput. - Use Predibase's LoRAX framework to see optimization techniques in action on a real LLM server. Sign up here:

Andrew Ng

104,727 views • 2 years ago

New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys LMSYS Org and RadixArk RadixArk, and taught by Richard Chen Richard Chen, a Member of Technical Staff at RadixArk. Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests. When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests. Skills you'll gain: - Implement a KV cache from scratch to eliminate redundant computation within a single request - Scale caching across users and requests with RadixAttention, so shared context is only processed once - Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism Join and learn to make LLM inference faster and more cost-efficient at scale!

New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys LMSYS Org and RadixArk RadixArk, and taught by Richard Chen Richard Chen, a Member of Technical Staff at RadixArk. Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests. When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests. Skills you'll gain: - Implement a KV cache from scratch to eliminate redundant computation within a single request - Scale caching across users and requests with RadixAttention, so shared context is only processed once - Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism Join and learn to make LLM inference faster and more cost-efficient at scale!

Andrew Ng

102,747 views • 5 months ago

IBM’s Christian Jacobi (IBM Fellow and CTO, IBM Systems Development) on why the PyTorch community matters for AI today, the role of open source, and what he’s hearing from clients deploying AI in core IT and business processes. 🎥

IBM’s Christian Jacobi (IBM Fellow and CTO, IBM Systems Development) on why the PyTorch community matters for AI today, the role of open source, and what he’s hearing from clients deploying AI in core IT and business processes. 🎥

PyTorch

14,339 views • 8 months ago

New short course with Mistral AI ! Mistral's open-source Mixtral 8x7B model uses a "mixture of experts" (MoE) architecture. Unlike a standard transformer, an MoE model has multiple expert feed-forward networks (8 in this case), with a gating network selecting two experts at inference time. This enables MoE to match the performance of a large model but faster inference. Mixtral 8x7B has 46.7B parameters but activates only 12.9B at inference to predict the next token. In Getting Started with Mistral, you’ll learn from Mistral’s Sophia Yang, Ph.D. to: - Explore Mistral's open-source models (Mistral 7B, Mixtral 8x7B) and commercial models via API calls and Mistral AI's Le Chat website - Implement JSON mode to generate structured outputs to integrate directly into larger software systems. - Use function calling for Tool Use, such as calling custom Python code that queries tabular data - Ground your LLM's response with external knowledge sources using RAG - Build a Mistral-powered chat interface that can reference external documents This course will help deepen your prompt engineering skills. Please sign up here:

New short course with Mistral AI ! Mistral's open-source Mixtral 8x7B model uses a "mixture of experts" (MoE) architecture. Unlike a standard transformer, an MoE model has multiple expert feed-forward networks (8 in this case), with a gating network selecting two experts at inference time. This enables MoE to match the performance of a large model but faster inference. Mixtral 8x7B has 46.7B parameters but activates only 12.9B at inference to predict the next token. In Getting Started with Mistral, you’ll learn from Mistral’s Sophia Yang, Ph.D. to: - Explore Mistral's open-source models (Mistral 7B, Mixtral 8x7B) and commercial models via API calls and Mistral AI's Le Chat website - Implement JSON mode to generate structured outputs to integrate directly into larger software systems. - Use function calling for Tool Use, such as calling custom Python code that queries tabular data - Ground your LLM's response with external knowledge sources using RAG - Build a Mistral-powered chat interface that can reference external documents This course will help deepen your prompt engineering skills. Please sign up here:

Andrew Ng

387,172 views • 2 years ago

I sat down with Silas Alberti, Head of Research Cognition, to chat about the intersection of RL and inference -- from training frontier coding models to running inference at scale. Watch till the end for bonus doggo! :o 0:00 — Intros 1:04 — What's hardest to get right in an RL run 3:39 — How Cognition started using Modal 4:03 — What to weigh when buying inference 4:58 — The latency/throughput Pareto frontier 7:14 — DFlash reveal (!!) 8:44 — Why faster runtime matters 9:56 — Memory-bound vs. compute-bound 10:57 — Tree-based speculation 14:20 — Hot take: RL and inference optimization converging 15:28 — Online training of the speculator 16:22 — "Auto inference" 17:27 — Agents competing on optimization problems 18:27 — What inference providers still get wrong

I sat down with Silas Alberti, Head of Research Cognition, to chat about the intersection of RL and inference -- from training frontier coding models to running inference at scale. Watch till the end for bonus doggo! :o 0:00 — Intros 1:04 — What's hardest to get right in an RL run 3:39 — How Cognition started using Modal 4:03 — What to weigh when buying inference 4:58 — The latency/throughput Pareto frontier 7:14 — DFlash reveal (!!) 8:44 — Why faster runtime matters 9:56 — Memory-bound vs. compute-bound 10:57 — Tree-based speculation 14:20 — Hot take: RL and inference optimization converging 15:28 — Online training of the speculator 16:22 — "Auto inference" 17:27 — Agents competing on optimization problems 18:27 — What inference providers still get wrong

gongy

81,979 views • 2 months ago

Check out mistral.rs, our #Rust-based open source inference engine allowing for fast #LLM serving for a variety of architectures including X-LoRA mixture-of-expert (MoE) models, Llama-3, Mistral/Mixtral, Gemma & many others. Built on the Hugging Face #Candle framework for #Rust w/ custom CUDA kernels in the backend (as well as support for Metal, Apple Accelerate, and Intel MKL for CPU use), you can easily create a REST API OpenAI compatible server or run via Python bindings. Key features include: ✅Prefix caching, continuous batching ✅Flash Attention V2 ✅Device offloading ✅GGUF or Hugging Face models ✅2, 3, 4, 5, 6 and 8 bit quantization ✅X-LoRA MoE non-granular scalings for fast inference ✅Grammar support ✅Continuous batching ✅LoRA support with weight merging ✅LlamaIndex 🦙 integration ...and much more. Incorporation into our GraphReasoning multi-agent modeling framework & LlamaIndex 🦙 allows you to combine in-context learning with adversarial agentic strategies, to dive deep into complex scientific analyses, such as to predict material behaviors, generate hypotheses, analyze papers and data, develop new research concepts, and much more. Check out mistral.rs: Join our Discord here: Rust Trending Rust Language

Check out mistral.rs, our #Rust-based open source inference engine allowing for fast #LLM serving for a variety of architectures including X-LoRA mixture-of-expert (MoE) models, Llama-3, Mistral/Mixtral, Gemma & many others. Built on the Hugging Face #Candle framework for #Rust w/ custom CUDA kernels in the backend (as well as support for Metal, Apple Accelerate, and Intel MKL for CPU use), you can easily create a REST API OpenAI compatible server or run via Python bindings. Key features include: ✅Prefix caching, continuous batching ✅Flash Attention V2 ✅Device offloading ✅GGUF or Hugging Face models ✅2, 3, 4, 5, 6 and 8 bit quantization ✅X-LoRA MoE non-granular scalings for fast inference ✅Grammar support ✅Continuous batching ✅LoRA support with weight merging ✅LlamaIndex 🦙 integration ...and much more. Incorporation into our GraphReasoning multi-agent modeling framework & LlamaIndex 🦙 allows you to combine in-context learning with adversarial agentic strategies, to dive deep into complex scientific analyses, such as to predict material behaviors, generate hypotheses, analyze papers and data, develop new research concepts, and much more. Check out mistral.rs: Join our Discord here: Rust Trending Rust Language

Markus J. Buehler

73,581 views • 2 years ago

A single, viral post on X skyrocketed Groq’s business. Groq Founder Jonathan Ross was looking for a way to show people that fast inference would be extremely useful. He remembered an example from Eric Schmidt showing off an LLM at a conference three weeks before the ChatGPT moment. Nobody in the audience cared. Then he looked at what made ChatGPT’s launch successful and realized that it worked because they allowed people to ask their own questions. “They got an answer to their question that was specific to them. That was magical.” “Seeing text show up for someone else's answer wasn't magical.” So he put Groq’s fast inference hardware on the internet for anyone to try. Someone posted a video on X of an LLM running on Groq running super fast. “It was just such eye candy when people saw it that everyone started creating their own and we just went viral.”

A single, viral post on X skyrocketed Groq’s business. Groq Founder Jonathan Ross was looking for a way to show people that fast inference would be extremely useful. He remembered an example from Eric Schmidt showing off an LLM at a conference three weeks before the ChatGPT moment. Nobody in the audience cared. Then he looked at what made ChatGPT’s launch successful and realized that it worked because they allowed people to ask their own questions. “They got an answer to their question that was specific to them. That was magical.” “Seeing text show up for someone else's answer wasn't magical.” So he put Groq’s fast inference hardware on the internet for anyone to try. Someone posted a video on X of an LLM running on Groq running super fast. “It was just such eye candy when people saw it that everyone started creating their own and we just went viral.”

David Senra

156,850 views • 2 months ago