Video yükleniyor...

Video Yüklenemedi

Bu video yüklenirken bir sorun oluştu. Bu geçici bir ağ sorunundan kaynaklanıyor olabilir veya video kullanılamıyor olabilir.

Ana Sayfaya Dön

We open-sourced QeRL — Quantization-enhanced Reinforcement Learning ! 🧠 4-bit quantized RL training 💪 Train a 32B LLM on a single H100 GPU ⚙️ 1.7× faster overall training 🎯 Accuracy on par with bfloat16-level accuracy 🔥 Supports NVFP4 quantization format Moreover, we show that quantization helps exploration in RL... show more

Yukang Chen

1,629 subscribers

69,896 görüntüleme • 1 yıl önce •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

19 Yorum

Yukang Chen profil fotoğrafı

Yukang Chen1 yıl önce

Paper: Code:

jo32 profil fotoğrafı

jo321 yıl önce

32B LLM training on 1 H100 GPU means a lot! Congrats! I made a video for anyone who wants to know what it’s about:

Yukang Chen profil fotoğrafı

Yukang Chen1 yıl önce

Great! Thanks!

ζ Pedram ζ profil fotoğrafı

ζ Pedram ζ1 yıl önce

wow

Nishi profil fotoğrafı

Nishi1 yıl önce

Wow! It’s great work! Can I use qwen3-Moe or qwen-next?

Yukang Chen profil fotoğrafı

Yukang Chen1 yıl önce

Thanks! It theoretically works. We did not try these models before.

Tejas Bhakta profil fotoğrafı

Tejas Bhakta11 ay önce

@Yuikooooooooo should definitely try it with Qwen 3 MOE! this is the model people struggle to get working on single gpu with RL

Nishi profil fotoğrafı

Nishi11 ay önce

@yukangchen_ NVFP4 is not supported in H100’s MoE, so I haven’t been able to test Qerl with qwen-next yet.

Umbridge profil fotoğrafı

Umbridge10 ay önce

pathetic work, check out their github issues, absolutely zero reproducibility 🤣😂🤡

Paul Chen profil fotoğrafı

Paul Chen11 ay önce

Does it support on-policy distillation?

Yukang Chen profil fotoğrafı

Yukang Chen11 ay önce

Theoretically support, we did not try yet.

Min Htoo profil fotoğrafı

Min Htoo11 ay önce

Very interesting work!

Cliff Lattner profil fotoğrafı

Cliff Lattner11 ay önce

Hm, this seems like it is really two or three papers all in one. You apply the noise vector after weight quantization, so the quantization format of the weight is pretty irrelevant. Not to mention that LoRA seems pretty orthogonal to all of this.

Yukang Chen profil fotoğrafı

Yukang Chen11 ay önce

LoRA branch using bf16 enables the backward of quantized RL

Cliff Lattner profil fotoğrafı

Cliff Lattner11 ay önce

Could you not just use NVFP4 for activations and weights, and then gradients in BF16? Why do you need LoRA?

Yukang Chen profil fotoğrafı

Yukang Chen11 ay önce

We use BF16 LoRA for more accurate training—it performs much better than training directly in NVFP4. Our activations are not in NVFP4; only the weights are, and we did not train with NVFP4.

Jesse profil fotoğrafı

Jesse11 ay önce

$0.001 for a 480p image with no visible quality loss? Fleek is either smoking something strong or they just quietly detonated the entire gen-AI pricing war. Someone benchmark this before I have an aneurysm.

Shital Shah profil fotoğrafı

Shital Shah1 yıl önce

Great stuff! Would this be integrated in Nemo-RL?

Yukang Chen profil fotoğrafı

Yukang Chen1 yıl önce

This currently works with open-r1. Not sure about Nemo.

Benzer Videolar

Ex-NVIDIA engineer who built Unsloth explained RL, kernels, reasoning, quantization, and agents in 2 hours 42 minutes - better than $5000 fine-tuning bootcamps. pick the base model -> write triton kernels for 2x faster fine-tune -> quantize to 4-bit -> run GRPO/DPO -> ship a reasoning model on your single GPU. That loop is why Unsloth is the default way to fine-tune Llama, Qwen, Gemma, and Phi on hardware you already own. Unsloth + Triton kernels + 4-bit quantization + GRPO/DPO + single-GPU fine-tuning - that's the stack. Watch and save it, then fine-tune your first model tonight.

Ex-NVIDIA engineer who built Unsloth explained RL, kernels, reasoning, quantization, and agents in 2 hours 42 minutes - better than $5000 fine-tuning bootcamps. pick the base model -> write triton kernels for 2x faster fine-tune -> quantize to 4-bit -> run GRPO/DPO -> ship a reasoning model on your single GPU. That loop is why Unsloth is the default way to fine-tune Llama, Qwen, Gemma, and Phi on hardware you already own. Unsloth + Triton kernels + 4-bit quantization + GRPO/DPO + single-GPU fine-tuning - that's the stack. Watch and save it, then fine-tune your first model tonight.

h100envy

497,625 görüntüleme • 3 ay önce

What if you could train AI agents on a laptop as easily as on a GPU cluster? Researchers from UIUC's U Lab, led by Prof. Jiaxuan You, just open-sourced OpenTinker. It's a new "Reinforcement-Learning-as-a-Service" (RLaaS) system that decouples the complex training pipeline into simple, distributed services with friendly APIs. The result? It breaks down the major engineering barriers to RL, outperforming traditional frameworks in accessibility and ease of deployment, finally making agent training viable for more developers and teams. Project: Code: U Lab: Our report: 📬 #PapersAccepted by Jiqizhixin

What if you could train AI agents on a laptop as easily as on a GPU cluster? Researchers from UIUC's U Lab, led by Prof. Jiaxuan You, just open-sourced OpenTinker. It's a new "Reinforcement-Learning-as-a-Service" (RLaaS) system that decouples the complex training pipeline into simple, distributed services with friendly APIs. The result? It breaks down the major engineering barriers to RL, outperforming traditional frameworks in accessibility and ease of deployment, finally making agent training viable for more developers and teams. Project: Code: U Lab: Our report: 📬 #PapersAccepted by Jiqizhixin

机器之心 JIQIZHIXIN

15,893 görüntüleme • 9 ay önce

(1/5) FP4 hardware is here, but 4-bit attention still kills model quality, blocking true end-to-end FP4 serving. To fix that, we propose Attn-QAT, the first systematic study of quantization-aware training for attention. The result: FP4 attention quality is comparable to BF16 attention with 1.1x–1.5x higher throughput than SageAttention3 on an RTX 5090 and 1.39x speedup over FlashAttention-4 on a B200. Blog: Code: Checkpoints:

(1/5) FP4 hardware is here, but 4-bit attention still kills model quality, blocking true end-to-end FP4 serving. To fix that, we propose Attn-QAT, the first systematic study of quantization-aware training for attention. The result: FP4 attention quality is comparable to BF16 attention with 1.1x–1.5x higher throughput than SageAttention3 on an RTX 5090 and 1.39x speedup over FlashAttention-4 on a B200. Blog: Code: Checkpoints:

Hao AI Lab

38,252 görüntüleme • 6 ay önce

The network for machine intelligence Two years ago, we laid out our vision for a machine learning compute protocol. One that connects every device in the world into an open network for machine intelligence, with no gatekeepers or artificial boundaries. This week, we’ll be sharing some of our early progress, beginning with RL Swarm, a peer-to-peer system for collaborative reinforcement learning over the internet. Next month, we’ll open our Testnet, allowing anyone to contribute to the frontier of open machine intelligence. Introducing RL Swarm RL Swarm is a fully open source system for collaborative reinforcement learning over the internet. It is a live demo of our research findings, which show that models training with RL learn faster when they train as a collective swarm than they do on their own. Join our swarm now to see this in practice. You can participate with consumer hardware at home or a powerful GPU in the cloud. You can follow along with the swarm’s progress by following the links below.

The network for machine intelligence Two years ago, we laid out our vision for a machine learning compute protocol. One that connects every device in the world into an open network for machine intelligence, with no gatekeepers or artificial boundaries. This week, we’ll be sharing some of our early progress, beginning with RL Swarm, a peer-to-peer system for collaborative reinforcement learning over the internet. Next month, we’ll open our Testnet, allowing anyone to contribute to the frontier of open machine intelligence. Introducing RL Swarm RL Swarm is a fully open source system for collaborative reinforcement learning over the internet. It is a live demo of our research findings, which show that models training with RL learn faster when they train as a collective swarm than they do on their own. Join our swarm now to see this in practice. You can participate with consumer hardware at home or a powerful GPU in the cloud. You can follow along with the swarm’s progress by following the links below.

gensyn

228,943 görüntüleme • 1 yıl önce

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

Andrew Ng

126,309 görüntüleme • 1 yıl önce

🚨 RL for LLMs is finally accessible. Introducing OpenTinker: The first community-driven, open-source framework designed to democratize Reinforcement Learning for LLMs. Inspired by Thinking Machines's amazing Tinker, we realize the biggest bottleneck in agentic LLM research isn’t the math—it’s the setup. Current RL pipelines are messy. Configuring VeRL for every single experiment is a productivity killer. OpenTinker fixed it. 🛠 How OpenTinker Works: Decoupled Design of Server and Client - Setup Once, Run Forever: Configure the OpenTinker backend on your GPU cluster once. - Develop Locally: Define your RL environments directly on your laptop. - Train on the Cloud: Simply point your local client to the backend. The cluster handles the compute; you handle the science. 📉 The 10x Development Efficiency Thanks to our elegant architectural decomposition, OpenTinker reduces the time to develop a new RL training pipeline by at least an order of magnitude. ⚡ Turn Idle GPU Compute into Gold Small labs often have underutilized hardware. OpenTinker turns your idle GPUs into an internal/external API service for - RL Training - SFT - Inference 🎯 Who needs OpenTinker? - Researchers tired of infrastructure hell. - Labs needing to standardize workflows. - Teams wanting to maximize hardware ROI. Thanks my amazing PhD student Siqi Zhu for leading the project. We are building the future of open RL infra. Be the first to build with us. 👇 Start Building with OpenTinker Now 🚀 Repo: 🌐 Blog: If you believe RL should be accessible to everyone, give us a star, repost this 🔄 post, and let us know what agents you plan to build!

🚨 RL for LLMs is finally accessible. Introducing OpenTinker: The first community-driven, open-source framework designed to democratize Reinforcement Learning for LLMs. Inspired by Thinking Machines's amazing Tinker, we realize the biggest bottleneck in agentic LLM research isn’t the math—it’s the setup. Current RL pipelines are messy. Configuring VeRL for every single experiment is a productivity killer. OpenTinker fixed it. 🛠 How OpenTinker Works: Decoupled Design of Server and Client - Setup Once, Run Forever: Configure the OpenTinker backend on your GPU cluster once. - Develop Locally: Define your RL environments directly on your laptop. - Train on the Cloud: Simply point your local client to the backend. The cluster handles the compute; you handle the science. 📉 The 10x Development Efficiency Thanks to our elegant architectural decomposition, OpenTinker reduces the time to develop a new RL training pipeline by at least an order of magnitude. ⚡ Turn Idle GPU Compute into Gold Small labs often have underutilized hardware. OpenTinker turns your idle GPUs into an internal/external API service for - RL Training - SFT - Inference 🎯 Who needs OpenTinker? - Researchers tired of infrastructure hell. - Labs needing to standardize workflows. - Teams wanting to maximize hardware ROI. Thanks my amazing PhD student Siqi Zhu for leading the project. We are building the future of open RL infra. Be the first to build with us. 👇 Start Building with OpenTinker Now 🚀 Repo: 🌐 Blog: If you believe RL should be accessible to everyone, give us a star, repost this 🔄 post, and let us know what agents you plan to build!

Jiaxuan You

58,401 görüntüleme • 10 ay önce

The DeepSeek-R1 paper is a gem! Highly encourage everyone to read it. It's clear that LLM reasoning capabilities can be learned in different ways. RL, if applied correctly and at scale, can lead to some really powerful and interesting scaling and emergent properties. There is more to RL than meets the eye! Here is my breakdown of the paper along with a few tests: The multi-state training might not make sense initially but they provide clues on optimizations that we can continue to tap into. Data quality is still very important for enhancing the usability of the LLM. Unlike other reasoning LLMs, DeepSeek-R1's training recipe and weights are open so we can build on top of it. This opens up exciting research opportunities. About the attached clip: the previous preview model wasn't able to solve this task. DeepSeek-R1 can solve this and many other tasks that o1 can solve. It's a very good model for coding and math.

The DeepSeek-R1 paper is a gem! Highly encourage everyone to read it. It's clear that LLM reasoning capabilities can be learned in different ways. RL, if applied correctly and at scale, can lead to some really powerful and interesting scaling and emergent properties. There is more to RL than meets the eye! Here is my breakdown of the paper along with a few tests: The multi-state training might not make sense initially but they provide clues on optimizations that we can continue to tap into. Data quality is still very important for enhancing the usability of the LLM. Unlike other reasoning LLMs, DeepSeek-R1's training recipe and weights are open so we can build on top of it. This opens up exciting research opportunities. About the attached clip: the previous preview model wasn't able to solve this task. DeepSeek-R1 can solve this and many other tasks that o1 can solve. It's a very good model for coding and math.

elvis

141,023 görüntüleme • 1 yıl önce

🚀 New Paper: Pixel Reasoner 🧠🖼️ How can Vision-Language Models (VLMs) perform chain-of-thought reasoning within the image itself? We introduce Pixel Reasoner, the first open-source framework that enables VLMs to “think in pixel space” through curiosity-driven reinforcement learning. Current VLMs reason only in text — even when grounded in rich images or videos, their logical steps are verbalized in natural language. This restricts their ability to interrogate visual evidence and demonstrate how conclusions are drawn. 🔍 So we ask: What if we could make VLMs "show their work" by reasoning directly in the pixel space? Inspired by GPT-o3’s "think-in-image" ability, we propose a framework where VLMs use interactive visual operations — zoom, select-frame, highlight — to reason through complex visual inputs. To do this, we design a two-stage training process: Instruction tuning with synthesized visual reasoning traces. Reinforcement learning with curiosity-driven reward to balance exploration between pixel and text reasoning ✨ With this, Pixel Reasoner achieves near-SoTA performance on many information-rich multimodal benchmarks: 📊 84% on InfographicsVQA 🧠 84% on V* benchmark 🧩 74% on TallyQA-Complex It also achieves strong accuracy of 68% on MVBench (a video benchmark). Website: Paper: Code: Demo: (coming soon)

🚀 New Paper: Pixel Reasoner 🧠🖼️ How can Vision-Language Models (VLMs) perform chain-of-thought reasoning within the image itself? We introduce Pixel Reasoner, the first open-source framework that enables VLMs to “think in pixel space” through curiosity-driven reinforcement learning. Current VLMs reason only in text — even when grounded in rich images or videos, their logical steps are verbalized in natural language. This restricts their ability to interrogate visual evidence and demonstrate how conclusions are drawn. 🔍 So we ask: What if we could make VLMs "show their work" by reasoning directly in the pixel space? Inspired by GPT-o3’s "think-in-image" ability, we propose a framework where VLMs use interactive visual operations — zoom, select-frame, highlight — to reason through complex visual inputs. To do this, we design a two-stage training process: Instruction tuning with synthesized visual reasoning traces. Reinforcement learning with curiosity-driven reward to balance exploration between pixel and text reasoning ✨ With this, Pixel Reasoner achieves near-SoTA performance on many information-rich multimodal benchmarks: 📊 84% on InfographicsVQA 🧠 84% on V* benchmark 🧩 74% on TallyQA-Complex It also achieves strong accuracy of 68% on MVBench (a video benchmark). Website: Paper: Code: Demo: (coming soon)

Wenhu Chen

82,829 görüntüleme • 1 yıl önce

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has *all* the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has all the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Sonya Huang 🐥

81,490 görüntüleme • 5 ay önce

Evolution happens to populations. Useful genes become common in a group over generations, then spread to other groups that test them in their own environments. The speed depends on generation time, 20 minutes for a bacterium but about 25 years for an elephant. Using LLMs to evolve better models often looks like an isolated herd of elephants. Each search runs alone on one dataset, and for large models every generation costs a long training run on the full data. Our new paper, "Discovered, Not Designed," introduces Population Evolution. It has two goals, to discover better models and to discover them faster. To discover better model designs, the searches share what they find. Several searches run at the same time, each on its own dataset. When one finds a code change that helps, the same change is retrained on the others. While nature keeps no record of the genes that failed to help, Population Evolution does. All peer results, including weak or unsuccessful transfers, are added to a shared ledger, and the LLM reads that ledger before it writes the next change. To discover them faster, the searches do not train on the full data, because that is what makes each generation slow. We found that a search can run on as little as 2% or 5% of the data and still find changes that hold up in full training, as long as the other searches check each change first. In Population Evolution, a change that works on one slice must also work on at least two other slices, then on two larger 10% sets. Only then does it get a full training run. We tested Population Evolution on RMD-Bench, a benchmark we also introduce for evaluating model discovery. It covers five task families (LLM pretraining, VLM pretraining, RL algorithm discovery, ranking, and watch-time prediction) with 34 instances in total. Each one comes with a reference implementation and a fixed training and evaluation protocol, so a change found on one dataset or data scale can be retrained and scored on another, and other discovery methods can be compared under the same conditions. The numbers below are a sneak peek at the two most expensive families of experiments. The LLM pretraining task starts from a 185M-parameter Gemma 3-style model, and the VLM pretraining task starts from a Qwen3.5-style model with about 646M trainable parameters. The baseline runs the same kind of search directly on the full data, at the same total GPU compute. Population Evolution finds better models. The best relative accuracy gain for LLM pretraining goes from 0.92% to 2.48%. The best loss reduction for VLM pretraining goes from 5.05% to 8.78%. Population Evolution finds them faster. In LLM pretraining, each search gets 100 generations where a full-data search gets 14. In VLM pretraining, 57 of 61 full training runs met the success threshold, against 12 of 100. Evolution works because many groups pass along what helps and get many generations to do it, and Population Evolution gives LLM-driven model search these prerequisites. Read our paper at for full implementation details and results. The benchmark data and code will be made publicly available shortly.

Evolution happens to populations. Useful genes become common in a group over generations, then spread to other groups that test them in their own environments. The speed depends on generation time, 20 minutes for a bacterium but about 25 years for an elephant. Using LLMs to evolve better models often looks like an isolated herd of elephants. Each search runs alone on one dataset, and for large models every generation costs a long training run on the full data. Our new paper, "Discovered, Not Designed," introduces Population Evolution. It has two goals, to discover better models and to discover them faster. To discover better model designs, the searches share what they find. Several searches run at the same time, each on its own dataset. When one finds a code change that helps, the same change is retrained on the others. While nature keeps no record of the genes that failed to help, Population Evolution does. All peer results, including weak or unsuccessful transfers, are added to a shared ledger, and the LLM reads that ledger before it writes the next change. To discover them faster, the searches do not train on the full data, because that is what makes each generation slow. We found that a search can run on as little as 2% or 5% of the data and still find changes that hold up in full training, as long as the other searches check each change first. In Population Evolution, a change that works on one slice must also work on at least two other slices, then on two larger 10% sets. Only then does it get a full training run. We tested Population Evolution on RMD-Bench, a benchmark we also introduce for evaluating model discovery. It covers five task families (LLM pretraining, VLM pretraining, RL algorithm discovery, ranking, and watch-time prediction) with 34 instances in total. Each one comes with a reference implementation and a fixed training and evaluation protocol, so a change found on one dataset or data scale can be retrained and scored on another, and other discovery methods can be compared under the same conditions. The numbers below are a sneak peek at the two most expensive families of experiments. The LLM pretraining task starts from a 185M-parameter Gemma 3-style model, and the VLM pretraining task starts from a Qwen3.5-style model with about 646M trainable parameters. The baseline runs the same kind of search directly on the full data, at the same total GPU compute. Population Evolution finds better models. The best relative accuracy gain for LLM pretraining goes from 0.92% to 2.48%. The best loss reduction for VLM pretraining goes from 5.05% to 8.78%. Population Evolution finds them faster. In LLM pretraining, each search gets 100 generations where a full-data search gets 14. In VLM pretraining, 57 of 61 full training runs met the success threshold, against 12 of 100. Evolution works because many groups pass along what helps and get many generations to do it, and Population Evolution gives LLM-driven model search these prerequisites. Read our paper at for full implementation details and results. The benchmark data and code will be made publicly available shortly.

Zhuokai Zhao

925,627 görüntüleme • 16 gün önce

Ever wished we had fewer X-training hyphenates? Pre, mid, post etc. Why not just Training? Trying to bridge the divides (and get all our friends into one team again), we intro *Introspective X Training*, an offline RL inspired method that scales effectively across any LLM stage by annotating your data with a thinking reward generated language critique! Up to 2.8x FLOP efficiency + 5-10 point score gains (esp with math and code) at any stage from scratch to 24T tokens on 8b (active) sized models!! We burned much compute ablating so you wouldn't have to Moral of the story is‼️don't throw out any data via filtering, just feedback condition it‼️ You can spend FLOPs up front on inference to *classify* data quality and then train so that tokens aren't all treated equally based on the feedback starting early in training itself. Right now they're really only separated out much later during mid/post training This improves overall compute efficiency and gives us benchmark perf not possible with just baseline methods! Paper here: Thanks to Brandon Cui and Ximing Lu for leading this w/ Syeda Nahida Akter David Acuna Hyunwoo Kim Jaehun Jung Yuxiao Qu Shrimai Yejin Choi

Ever wished we had fewer X-training hyphenates? Pre, mid, post etc. Why not just Training? Trying to bridge the divides (and get all our friends into one team again), we intro Introspective X Training, an offline RL inspired method that scales effectively across any LLM stage by annotating your data with a thinking reward generated language critique! Up to 2.8x FLOP efficiency + 5-10 point score gains (esp with math and code) at any stage from scratch to 24T tokens on 8b (active) sized models!! We burned much compute ablating so you wouldn't have to Moral of the story is‼️don't throw out any data via filtering, just feedback condition it‼️ You can spend FLOPs up front on inference to classify data quality and then train so that tokens aren't all treated equally based on the feedback starting early in training itself. Right now they're really only separated out much later during mid/post training This improves overall compute efficiency and gives us benchmark perf not possible with just baseline methods! Paper here: Thanks to Brandon Cui and Ximing Lu for leading this w/ Syeda Nahida Akter David Acuna Hyunwoo Kim Jaehun Jung Yuxiao Qu Shrimai Yejin Choi

Prithviraj (Raj) Ammanabrolu

27,902 görüntüleme • 5 ay önce

MotionGPT: Human Motion as a Foreign Language paper page: Though the advancement of pre-trained large language models unfolds, the exploration of building a unified model for language and other multi-modal data, such as motion, remains challenging and untouched so far. Fortunately, human motion displays a semantic coupling akin to human language, often perceived as a form of body language. By fusing language data with large-scale motion models, motion-language pre-training that can enhance the performance of motion-related tasks becomes feasible. Driven by this insight, we propose MotionGPT, a unified, versatile, and user-friendly motion-language model to handle multiple motion-relevant tasks. Specifically, we employ the discrete vector quantization for human motion and transfer 3D motion into motion tokens, similar to the generation process of word tokens. Building upon this "motion vocabulary", we perform language modeling on both motion and text in a unified manner, treating human motion as a specific language. Moreover, inspired by prompt learning, we pre-train MotionGPT with a mixture of motion-language data and fine-tune it on prompt-based question-and-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performances on multiple motion tasks including text-driven motion generation, motion captioning, motion prediction, and motion in-between.

MotionGPT: Human Motion as a Foreign Language paper page: Though the advancement of pre-trained large language models unfolds, the exploration of building a unified model for language and other multi-modal data, such as motion, remains challenging and untouched so far. Fortunately, human motion displays a semantic coupling akin to human language, often perceived as a form of body language. By fusing language data with large-scale motion models, motion-language pre-training that can enhance the performance of motion-related tasks becomes feasible. Driven by this insight, we propose MotionGPT, a unified, versatile, and user-friendly motion-language model to handle multiple motion-relevant tasks. Specifically, we employ the discrete vector quantization for human motion and transfer 3D motion into motion tokens, similar to the generation process of word tokens. Building upon this "motion vocabulary", we perform language modeling on both motion and text in a unified manner, treating human motion as a specific language. Moreover, inspired by prompt learning, we pre-train MotionGPT with a mixture of motion-language data and fine-tune it on prompt-based question-and-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performances on multiple motion tasks including text-driven motion generation, motion captioning, motion prediction, and motion in-between.

AK

125,403 görüntüleme • 3 yıl önce

Historically, RL policies for robots have been trained in synthetic, untextured environments, limiting perceptive policies to depth images where the sim-to-real gap is manageable. RGB has always had more potential, but leveraging it to train policies in simulation remained an open problem. Partnering with Niantic Spatial 🌎 and NVIDIA Robotics, we built a pipeline that addresses exactly that. We can now scan a real deployment site with off-the-shelf hardware, reconstruct it into a photorealistic Gaussian splat, and run massively parallel RL training. The policies trained in our Gym environment then transfer zero-shot to the real robot and environments they were trained for. This enables faster deployment of more capable and robust policies for the end user. The new resulting capabilities are a big step towards solving sim-to-real and also apply well beyond navigation. Read the full technical breakdown on our blog; link in the comments. #HumanoidRobots #Flexion #NianticSpatial #NVIDIA

Historically, RL policies for robots have been trained in synthetic, untextured environments, limiting perceptive policies to depth images where the sim-to-real gap is manageable. RGB has always had more potential, but leveraging it to train policies in simulation remained an open problem. Partnering with Niantic Spatial 🌎 and NVIDIA Robotics, we built a pipeline that addresses exactly that. We can now scan a real deployment site with off-the-shelf hardware, reconstruct it into a photorealistic Gaussian splat, and run massively parallel RL training. The policies trained in our Gym environment then transfer zero-shot to the real robot and environments they were trained for. This enables faster deployment of more capable and robust policies for the end user. The new resulting capabilities are a big step towards solving sim-to-real and also apply well beyond navigation. Read the full technical breakdown on our blog; link in the comments. #HumanoidRobots #Flexion #NianticSpatial #NVIDIA

Flexion

73,261 görüntüleme • 3 ay önce

Check out our #ICRA2024 paper "Actor-Critic Model Predictive Control." Model-free #reinforcementlearning (RL) is known for its strong task performance and flexibility in optimizing general reward formulations. On the other hand, #ModelPredictiveControl (MPC) benefits from robustness and online replanning capabilities. We combine both approaches by introducing a new framework called Actor-Critic Model Predictive Control. The key idea is to embed a differentiable MPC within an Actor-Critic RL framework. The proposed approach leverages the short-term predictive optimization capabilities of MPC with the exploratory and end-to-end training properties of RL. The resulting policy effectively manages both short-term decisions through the MPC-based actor and long-term prediction via the critic network, unifying the benefits of both model-based control and end-to-end learning. We validate our method in simulation and the real world with a quadcopter across various high-level tasks. We show that the proposed architecture can achieve real-time control performance, learn complex behaviors via trial and error, and retain the predictive properties of the MPC to better handle out-of-distribution behavior. Paper: Full Video with more details: Kudos to Ángel Romero, Yunlong Song IEEE ICRA University of Zurich UZH Science UZH Space Hub Aerial Core AUTOASSESS European Research Council (ERC)

Check out our #ICRA2024 paper "Actor-Critic Model Predictive Control." Model-free #reinforcementlearning (RL) is known for its strong task performance and flexibility in optimizing general reward formulations. On the other hand, #ModelPredictiveControl (MPC) benefits from robustness and online replanning capabilities. We combine both approaches by introducing a new framework called Actor-Critic Model Predictive Control. The key idea is to embed a differentiable MPC within an Actor-Critic RL framework. The proposed approach leverages the short-term predictive optimization capabilities of MPC with the exploratory and end-to-end training properties of RL. The resulting policy effectively manages both short-term decisions through the MPC-based actor and long-term prediction via the critic network, unifying the benefits of both model-based control and end-to-end learning. We validate our method in simulation and the real world with a quadcopter across various high-level tasks. We show that the proposed architecture can achieve real-time control performance, learn complex behaviors via trial and error, and retain the predictive properties of the MPC to better handle out-of-distribution behavior. Paper: Full Video with more details: Kudos to Ángel Romero, Yunlong Song IEEE ICRA University of Zurich UZH Science UZH Space Hub Aerial Core AUTOASSESS European Research Council (ERC)

Davide Scaramuzza

34,920 görüntüleme • 2 yıl önce

Google just proved that bigger isn't always better. Their 308M parameter model is outperforming models 2x its size. Google just released 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝗚𝗲𝗺𝗺𝗮, and it's proving that lightweight embedding models can punch way above their weight class. At just 308M parameters (578MB), it's the new state-of-the-art for models under 500M parameters across MTEB multilingual, English, and code benchmarks. But the really impressive part is that it ranks 8th overall on MTEB(Multilingual, v2) - that's 𝟭𝟳 𝗽𝗹𝗮𝗰𝗲𝘀 above the second-best sub-500M model, and it's delivering performance 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗻𝗲𝗮𝗿𝗹𝘆 𝗱𝗼𝘂𝗯𝗹𝗲 𝗶𝘁𝘀 𝘀𝗶𝘇𝗲. There are three key parts of their training recipe that sets it apart: 𝟭. 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗗𝗲𝗰𝗼𝗱𝗲𝗿 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Instead of starting from a decoder-only Gemma 3 model, they first adapted it to encoder-decoder, then used just the encoder. By basing EmbeddingGemma off an LLM that already has world and language understanding, it gives it a stronger starting point. 𝟮. 𝗧𝗵𝗿𝗲𝗲-𝗟𝗼𝘀𝘀 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 They combine three different loss functions, instead of just having one: • Contrastive loss (NCE) with in-batch negatives and hardness weighting • Spread-out regularization to ensure embeddings utilize the full space (for quantization and ANN retrieval) • Embedding matching distillation from Gemini Embedding - not just learning from relevance scores, but directly aligning the embedding space with the teacher model 𝟯. 𝗠𝗼𝗱𝗲𝗹 𝗦𝗼𝘂𝗽𝗶𝗻𝗴 Rather than just averaging checkpoints from the same training run, they use optimization techniques to find multiple specialized training mixtures. Each mixture creates an "expert" model in different domains, and averaging all their parameters creates a final model that's actually better than individual models. Extras: • Matryoshka embeddings supporting 768, 512, 256, and 128 dimensions • Quantization-aware training - maintains quality even at int4 precision • 100+ languages from Gemma 3 pretraining • Exceptional performance on low-resource languages (check their XTREME-UP results) Is it the absolute best embedding model? No - Gemini Embedding still leads overall. But that's not really the point. EmbeddingGemma proves you can achieve state-of-the-art performance in a small package that's actually deployable on-device, in low-latency applications, and in resource-constrained environments. This makes good embeddings accessible for use cases that I'm seeing more and more: offline applications, privacy-sensitive deployments, and high-throughput scenarios where inference cost actually matters. Full paper: Shoutout to the EmbeddingGemma team at Google DeepMind for this awesome open source work 💙 and to Daniel Williams for helping me with this video! 🫶

Google just proved that bigger isn't always better. Their 308M parameter model is outperforming models 2x its size. Google just released 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝗚𝗲𝗺𝗺𝗮, and it's proving that lightweight embedding models can punch way above their weight class. At just 308M parameters (578MB), it's the new state-of-the-art for models under 500M parameters across MTEB multilingual, English, and code benchmarks. But the really impressive part is that it ranks 8th overall on MTEB(Multilingual, v2) - that's 𝟭𝟳 𝗽𝗹𝗮𝗰𝗲𝘀 above the second-best sub-500M model, and it's delivering performance 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗻𝗲𝗮𝗿𝗹𝘆 𝗱𝗼𝘂𝗯𝗹𝗲 𝗶𝘁𝘀 𝘀𝗶𝘇𝗲. There are three key parts of their training recipe that sets it apart: 𝟭. 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗗𝗲𝗰𝗼𝗱𝗲𝗿 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Instead of starting from a decoder-only Gemma 3 model, they first adapted it to encoder-decoder, then used just the encoder. By basing EmbeddingGemma off an LLM that already has world and language understanding, it gives it a stronger starting point. 𝟮. 𝗧𝗵𝗿𝗲𝗲-𝗟𝗼𝘀𝘀 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 They combine three different loss functions, instead of just having one: • Contrastive loss (NCE) with in-batch negatives and hardness weighting • Spread-out regularization to ensure embeddings utilize the full space (for quantization and ANN retrieval) • Embedding matching distillation from Gemini Embedding - not just learning from relevance scores, but directly aligning the embedding space with the teacher model 𝟯. 𝗠𝗼𝗱𝗲𝗹 𝗦𝗼𝘂𝗽𝗶𝗻𝗴 Rather than just averaging checkpoints from the same training run, they use optimization techniques to find multiple specialized training mixtures. Each mixture creates an "expert" model in different domains, and averaging all their parameters creates a final model that's actually better than individual models. Extras: • Matryoshka embeddings supporting 768, 512, 256, and 128 dimensions • Quantization-aware training - maintains quality even at int4 precision • 100+ languages from Gemma 3 pretraining • Exceptional performance on low-resource languages (check their XTREME-UP results) Is it the absolute best embedding model? No - Gemini Embedding still leads overall. But that's not really the point. EmbeddingGemma proves you can achieve state-of-the-art performance in a small package that's actually deployable on-device, in low-latency applications, and in resource-constrained environments. This makes good embeddings accessible for use cases that I'm seeing more and more: offline applications, privacy-sensitive deployments, and high-throughput scenarios where inference cost actually matters. Full paper: Shoutout to the EmbeddingGemma team at Google DeepMind for this awesome open source work 💙 and to Daniel Williams for helping me with this video! 🫶

Victoria Slocum

21,610 görüntüleme • 11 ay önce

a new 8GB VRAM GPU dense Local LLM leader was born yesterday runs on: RTX 4060 / RTX 3070 / RTX 2080. any 8GB card Qwen 3.5 9B (dense) was the go to for 6-8GB VRAM builds. Gemma 4 12B QAT (dense) just changed that. same llama.cpp + cuda 13.2. i7 12700H. 16GB RAM. same -ngl 99 flags. same 48k context. unsloth gemma-4-12b-it-Q4_K_M.gguf → 15 tok/sec @ 48k ctx unsloth gemma-4-12B-it-qat-UD-Q4_K_XL.gguf → 32 tok/sec @ 48k ctx → 26 tok/sec @ 64k ctx 64k context is a big deal. Hermes 3 agent requires 64k minimum to run. you're now getting full hermes compatible context on a budget consumer GPU at 26 tok/sec locally. 2.1x faster on identical hardware. and here's the part that breaks your brain: the QAT-UD-Q4_K_XL is actually SMALLER than the Q4_K_M "XL" why? QAT = Quantization Aware Training Google didn't train the model first and compress it later they trained it to be quantized from day one the weights already know how to survive low precision that's why you get more quality per byte llamacpp flags: -m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf -cnv -ngl 99 -c 48000 -v fits in 8GB VRAM clean. no API. no cloud. no subscription. and this isn't even the MTP variant yet Gemma-4-E2B QAT runs on 3GB RAM, E4B on 5GB, 12B on 7GB, 26-A4B on 15GB and 31B on 18GB. I have benchmarked the 26b and 31b qat as well on a single RTX 4090, checkout the comments for details. If you have a 6GB or 8GB VRAM GPU, post your numbers. more benchmarks and configs coming soon

a new 8GB VRAM GPU dense Local LLM leader was born yesterday runs on: RTX 4060 / RTX 3070 / RTX 2080. any 8GB card Qwen 3.5 9B (dense) was the go to for 6-8GB VRAM builds. Gemma 4 12B QAT (dense) just changed that. same llama.cpp + cuda 13.2. i7 12700H. 16GB RAM. same -ngl 99 flags. same 48k context. unsloth gemma-4-12b-it-Q4_K_M.gguf → 15 tok/sec @ 48k ctx unsloth gemma-4-12B-it-qat-UD-Q4_K_XL.gguf → 32 tok/sec @ 48k ctx → 26 tok/sec @ 64k ctx 64k context is a big deal. Hermes 3 agent requires 64k minimum to run. you're now getting full hermes compatible context on a budget consumer GPU at 26 tok/sec locally. 2.1x faster on identical hardware. and here's the part that breaks your brain: the QAT-UD-Q4_K_XL is actually SMALLER than the Q4_K_M "XL" why? QAT = Quantization Aware Training Google didn't train the model first and compress it later they trained it to be quantized from day one the weights already know how to survive low precision that's why you get more quality per byte llamacpp flags: -m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf -cnv -ngl 99 -c 48000 -v fits in 8GB VRAM clean. no API. no cloud. no subscription. and this isn't even the MTP variant yet Gemma-4-E2B QAT runs on 3GB RAM, E4B on 5GB, 12B on 7GB, 26-A4B on 15GB and 31B on 18GB. I have benchmarked the 26b and 31b qat as well on a single RTX 4090, checkout the comments for details. If you have a 6GB or 8GB VRAM GPU, post your numbers. more benchmarks and configs coming soon

Alok

264,581 görüntüleme • 4 ay önce

We are thrilled to share our breakthrough research on "Agile Flight from Pixels without State Estimation," to be presented and live-demonstrated at #RSS2024 next week! You heard well: no state estimation means no explicit visual localization, no SLAM, no VIO, and no IMU! Paper: Video (Narrated): Last year, we demonstrated that #ReinforcementLearning (RL) policies could outperform world-champion drone-racing pilots using the same quadrotor hardware; however, unlike human pilots, these policies continuously estimated an explicit state from known gate positions, the camera feed, and inertial measurements (IMU). In this new work, we tackle the challenge of learning vision-based drone racing using an end-to-end reinforcement learning approach that eliminates the need for IMU data or explicit state estimation. Like professional pilots, we go directly from images to control commands. The training is facilitated by an asymmetric actor-critic with access to privileged information. To overcome the computational complexity during image-based RL training, we use an appropriate sensor representation, which can be efficiently simulated during training without rendering images. We achieve agile flight at speeds up to 40 km/h with accelerations up to 2 g's. Although our demonstration focuses on drone racing, we believe that our method has an impact beyond drone racing and can serve as a foundation for future research into real-world applications in structured environments. Besides the paper presentation, we will also give a live demo next Tuesday and Wednesday between and hrs at TU Delft: Reference: Ismail Geles*, Leonard Bauersfeld*, Angel Romero, Jiaxu Xing, Davide Scaramuzza "Demonstrating Agile Flight from Pixels without State Estimation" Robotics: Science and Systems (RSS), 2024. Kudos to Ismail Geles Leonard Bauersfeld Ángel Romero Jiaxu Xing! University of Zurich UZH Science UZH Space Hub Aerial Core AUTOASSESS European Research Council (ERC)

We are thrilled to share our breakthrough research on "Agile Flight from Pixels without State Estimation," to be presented and live-demonstrated at #RSS2024 next week! You heard well: no state estimation means no explicit visual localization, no SLAM, no VIO, and no IMU! Paper: Video (Narrated): Last year, we demonstrated that #ReinforcementLearning (RL) policies could outperform world-champion drone-racing pilots using the same quadrotor hardware; however, unlike human pilots, these policies continuously estimated an explicit state from known gate positions, the camera feed, and inertial measurements (IMU). In this new work, we tackle the challenge of learning vision-based drone racing using an end-to-end reinforcement learning approach that eliminates the need for IMU data or explicit state estimation. Like professional pilots, we go directly from images to control commands. The training is facilitated by an asymmetric actor-critic with access to privileged information. To overcome the computational complexity during image-based RL training, we use an appropriate sensor representation, which can be efficiently simulated during training without rendering images. We achieve agile flight at speeds up to 40 km/h with accelerations up to 2 g's. Although our demonstration focuses on drone racing, we believe that our method has an impact beyond drone racing and can serve as a foundation for future research into real-world applications in structured environments. Besides the paper presentation, we will also give a live demo next Tuesday and Wednesday between and hrs at TU Delft: Reference: Ismail Geles, Leonard Bauersfeld, Angel Romero, Jiaxu Xing, Davide Scaramuzza "Demonstrating Agile Flight from Pixels without State Estimation" Robotics: Science and Systems (RSS), 2024. Kudos to Ismail Geles Leonard Bauersfeld Ángel Romero Jiaxu Xing! University of Zurich UZH Science UZH Space Hub Aerial Core AUTOASSESS European Research Council (ERC)

Davide Scaramuzza

28,002 görüntüleme • 2 yıl önce

NVIDIA just unleashed SANA-WM and it’s an absolute MONSTER for the future of open source AI! A blazing-fast 2.6B-parameter open-source world model that doesn’t just generate video… it creates controllable, physics-rich, high-fidelity worlds on demand. Why this is insanely powerful: • One image + text prompt + 6-DoF camera trajectory → generates 720p videos up to 60 seconds long with buttery-smooth, precisely controlled camera movement. You’re not just watching, you’re piloting the simulation. • Runs locally on a single consumer GPU (RTX 5090 level) thanks to heavy distillation + NVFP4 quantization. Full 60-second clip denoised in ~34 seconds. No massive clusters required. • 36× higher throughput than previous open models while rivaling (or beating) closed industrial giants in visual quality and consistency. • Trained lightning-fast: ~213K public videos in just 15 days on 64 H100s. • Built with next-level tech: Hybrid Linear Attention, dual-branch camera control, two-stage pipeline, and rock-solid metric-scale pose understanding. This is a true open world model, the foundation for embodied AI, robotics, autonomous systems, and hyper-realistic simulations that can run anywhere. Project: At our Zero-Human Company, we’re already running SANA-WM live in our core pipelines. It’s supercharging autonomous agent training, generating unlimited synthetic training data, and powering full end-to-end simulation loops, zero humans in the loop. The speed and control let us test thousands of edge-case scenarios overnight, iterate at lightspeed, and push our fully autonomous operations further than ever before. This is the kind of breakthrough that turns science fiction into daily reality. World models just leveled up — hard. The age of personal, local, controllable universes is here.

NVIDIA just unleashed SANA-WM and it’s an absolute MONSTER for the future of open source AI! A blazing-fast 2.6B-parameter open-source world model that doesn’t just generate video… it creates controllable, physics-rich, high-fidelity worlds on demand. Why this is insanely powerful: • One image + text prompt + 6-DoF camera trajectory → generates 720p videos up to 60 seconds long with buttery-smooth, precisely controlled camera movement. You’re not just watching, you’re piloting the simulation. • Runs locally on a single consumer GPU (RTX 5090 level) thanks to heavy distillation + NVFP4 quantization. Full 60-second clip denoised in ~34 seconds. No massive clusters required. • 36× higher throughput than previous open models while rivaling (or beating) closed industrial giants in visual quality and consistency. • Trained lightning-fast: ~213K public videos in just 15 days on 64 H100s. • Built with next-level tech: Hybrid Linear Attention, dual-branch camera control, two-stage pipeline, and rock-solid metric-scale pose understanding. This is a true open world model, the foundation for embodied AI, robotics, autonomous systems, and hyper-realistic simulations that can run anywhere. Project: At our Zero-Human Company, we’re already running SANA-WM live in our core pipelines. It’s supercharging autonomous agent training, generating unlimited synthetic training data, and powering full end-to-end simulation loops, zero humans in the loop. The speed and control let us test thousands of edge-case scenarios overnight, iterate at lightspeed, and push our fully autonomous operations further than ever before. This is the kind of breakthrough that turns science fiction into daily reality. World models just leveled up — hard. The age of personal, local, controllable universes is here.

Brian Roemmele

619,679 görüntüleme • 5 ay önce

I coded a Speech-to-Text model from scratch. 𝐇𝐞𝐫𝐞 𝐢𝐬 𝐭𝐡𝐞 𝐛𝐥𝐨𝐠 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐬𝐚𝐦𝐞: No APIs. No pre-trained models. Just PyTorch, an A100 GPU, and hours of debugging. This started months ago. I wanted to understand how machines hear. Not surface-level understanding. I wanted to build the whole thing myself. So I built it piece by piece: autoencoders, VAEs, VQ-VAEs, Residual Vector Quantization, and CTC loss. Each one took days to get right. Trained for 3 hours on 13,100 audio clips. Got complete garbage. Changed the tokenizer from BPE to character-level. Rechecked everything. Asked AVB who built STT models before. His answer: these models are tricky to train and need days of compute, not hours. Cut the dataset to 200 clips. After 2 hours, actual words appeared. Overfitted? Absolutely. But watching noise turn into recognizable English was satisfying. I have made a blog about this as well so you can learn about the same and my process - Audio fundamentals and waveform representation - Why attention breaks on raw audio - Convolutional downsampling - Transformer encoder with positional encoding - Vector Quantization, straight-through estimator, and RVQ - CTC loss and greedy decoding - Full training loop with VQ loss warmup - What went wrong and what finally worked Resources: - Blog: - Code: More Resoures CTC loss AVB videos SoundStream Paper LJ speech dataset wav2vec paper RVQ blog Next up: I've already trained two TTS architectures from scratch. Video post about those coming soon. But first, I'm dropping a visual breakdown of Vision Transformers, covering how they work and how to fine-tune them. Follow me Mayank Pratap Singh you're into audio deep learning. Repost so others can find this

I coded a Speech-to-Text model from scratch. 𝐇𝐞𝐫𝐞 𝐢𝐬 𝐭𝐡𝐞 𝐛𝐥𝐨𝐠 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐬𝐚𝐦𝐞: No APIs. No pre-trained models. Just PyTorch, an A100 GPU, and hours of debugging. This started months ago. I wanted to understand how machines hear. Not surface-level understanding. I wanted to build the whole thing myself. So I built it piece by piece: autoencoders, VAEs, VQ-VAEs, Residual Vector Quantization, and CTC loss. Each one took days to get right. Trained for 3 hours on 13,100 audio clips. Got complete garbage. Changed the tokenizer from BPE to character-level. Rechecked everything. Asked AVB who built STT models before. His answer: these models are tricky to train and need days of compute, not hours. Cut the dataset to 200 clips. After 2 hours, actual words appeared. Overfitted? Absolutely. But watching noise turn into recognizable English was satisfying. I have made a blog about this as well so you can learn about the same and my process - Audio fundamentals and waveform representation - Why attention breaks on raw audio - Convolutional downsampling - Transformer encoder with positional encoding - Vector Quantization, straight-through estimator, and RVQ - CTC loss and greedy decoding - Full training loop with VQ loss warmup - What went wrong and what finally worked Resources: - Blog: - Code: More Resoures CTC loss AVB videos SoundStream Paper LJ speech dataset wav2vec paper RVQ blog Next up: I've already trained two TTS architectures from scratch. Video post about those coming soon. But first, I'm dropping a visual breakdown of Vision Transformers, covering how they work and how to fine-tune them. Follow me Mayank Pratap Singh you're into audio deep learning. Repost so others can find this

Mayank Pratap Singh

51,382 görüntüleme • 7 ay önce