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.

(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... show more

Hao AI Lab

6,639 subscribers

37,869 views • 4 months ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

🚀 Excited to release LongLive 2.0! 🎬 An end-to-end infrastructure for long video generation, with FP4 and parallelism at the core of both training and inference. ⚡45.7 FPS generation speed on 5B model⚡ ✨ LongLive 2.0 supports real-video training, few-step distillation, multi-shot training/inference, sequence-parallel acceleration, NVFP4 KV cache, and async VAE decoding deployment. 🧩 To our knowledge, this is the first open-source 4-bit long video generation infra that covers both training and inference. 🙌 Welcome to check it out, try it, and share feedback! 🔗 Code: 📰 Paper: 🎥 Demo: #LongVideoGeneration #VideoGeneration #Realtime #AIInfra #EfficientAI #FP4 #Parallel #NVIDIA

🚀 Excited to release LongLive 2.0! 🎬 An end-to-end infrastructure for long video generation, with FP4 and parallelism at the core of both training and inference. ⚡45.7 FPS generation speed on 5B model⚡ ✨ LongLive 2.0 supports real-video training, few-step distillation, multi-shot training/inference, sequence-parallel acceleration, NVFP4 KV cache, and async VAE decoding deployment. 🧩 To our knowledge, this is the first open-source 4-bit long video generation infra that covers both training and inference. 🙌 Welcome to check it out, try it, and share feedback! 🔗 Code: 📰 Paper: 🎥 Demo: #LongVideoGeneration #VideoGeneration #Realtime #AIInfra #EfficientAI #FP4 #Parallel #NVIDIA

Yukang Chen

59,280 views • 2 months ago

$Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens - Less than 5% the cost of Opus Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention). Only a small fraction actually matter. Subquadratic finds and focuses only on the ones that do. That's nearly 1,000x less compute and a new way for LLMs to scale.$

Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens - Less than 5% the cost of Opus Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention). Only a small fraction actually matter. Subquadratic finds and focuses only on the ones that do. That's nearly 1,000x less compute and a new way for LLMs to scale.

Alexander Whedon

13,245,683 views • 3 months ago

"When it comes to the jobs we create, we don’t just focus on the quantity, we pay more attention to the quality. So much so that the quality of a single job I create for one person is worth more than 50 jobs in other places." — Prof. Isa Ali Ibrahim, CON

"When it comes to the jobs we create, we don’t just focus on the quantity, we pay more attention to the quality. So much so that the quality of a single job I create for one person is worth more than 50 jobs in other places." — Prof. Isa Ali Ibrahim, CON

B_SARKY

18,876 views • 3 months ago

New short course: Attention in Transformers: Concepts and Code in PyTorch. Last week we released a course on how LLM transformers work. This week, go deeper and learn about the technical ideas behind the attention mechanism, and see how to code it in PyTorch. This course is built with Joshua Starmer, Founder and CEO of StatQuest. The attention mechanism was a breakthrough that led to transformers, the architecture powering large language models like ChatGPT. Transformers, introduced in the 2017 paper: "Attention is All You Need" by Viswani and others, took off because of its highly scalable design. In this course, you’ll learn how the attention mechanism, a key element of transformer-based LLMs, works and implement it in PyTorch. You'll develop deep intuition about building reliable, functional, and scalable AI applications. What you will do: - Understand the evolution of the attention mechanism, a key breakthrough that led to transformers. - Learn the relationships between word embeddings, positional embeddings, and attention. - Learn about the Query, Key, and Value matrices, and how to produce and use them in attention. - Walk through the math required to calculate self-attention and masked self-attention to learn why and how they work. - Understand the difference between self-attention and masked self-attention and how one is used in the encoder to build context-aware embeddings and the other is used in the decoder for generative outputs. - Learn the details of the encoder-decoder architecture, cross-attention, and multi-head attention and how they are all incorporated into a transformer. - Use PyTorch to code a class that implements self-attention, masked self-attention, and multi-head attention. There're lots of exciting technical details in this course. Please sign up here:

New short course: Attention in Transformers: Concepts and Code in PyTorch. Last week we released a course on how LLM transformers work. This week, go deeper and learn about the technical ideas behind the attention mechanism, and see how to code it in PyTorch. This course is built with Joshua Starmer, Founder and CEO of StatQuest. The attention mechanism was a breakthrough that led to transformers, the architecture powering large language models like ChatGPT. Transformers, introduced in the 2017 paper: "Attention is All You Need" by Viswani and others, took off because of its highly scalable design. In this course, you’ll learn how the attention mechanism, a key element of transformer-based LLMs, works and implement it in PyTorch. You'll develop deep intuition about building reliable, functional, and scalable AI applications. What you will do: - Understand the evolution of the attention mechanism, a key breakthrough that led to transformers. - Learn the relationships between word embeddings, positional embeddings, and attention. - Learn about the Query, Key, and Value matrices, and how to produce and use them in attention. - Walk through the math required to calculate self-attention and masked self-attention to learn why and how they work. - Understand the difference between self-attention and masked self-attention and how one is used in the encoder to build context-aware embeddings and the other is used in the decoder for generative outputs. - Learn the details of the encoder-decoder architecture, cross-attention, and multi-head attention and how they are all incorporated into a transformer. - Use PyTorch to code a class that implements self-attention, masked self-attention, and multi-head attention. There're lots of exciting technical details in this course. Please sign up here:

Andrew Ng

132,285 views • 1 year ago

.Glinert 🇺🇸 🏭 (Co-Founder & CEO of Sphere Semi) on how they're using AI to change how chips are designed: “CPUs and GPUs get all the attention, but analog is a third of the trillion-dollar chip industry and it’s critical for communications, warfare, and more.” “We think human chip design is coming to an end. In analog, it’s still done by hand and we’re putting that to an end.”

.Glinert 🇺🇸 🏭 (Co-Founder & CEO of Sphere Semi) on how they're using AI to change how chips are designed: “CPUs and GPUs get all the attention, but analog is a third of the trillion-dollar chip industry and it’s critical for communications, warfare, and more.” “We think human chip design is coming to an end. In analog, it’s still done by hand and we’re putting that to an end.”

TBPN

58,748 views • 10 months ago

Zhilin at GTC: Introducing Attention Residuals Learning selective memory, rather than mechanically accumulating everything, is the beauty of attention. Many of you have probably read Attention Is All You Need, the 2017 Transformer paper that brought “human-like” attention into the model’s field of view. From that point on, models no longer simply read everything in a mechanical way. Instead, they began to develop a sense of what matters more and what matters less across the text, choosing to retain the more important information. Recently, Kimi applied this idea of attention to the temporal dimension, then rotated it 90 degrees into the model’s depth dimension. This allows the model to have attention not only over time, but also throughout the process of information transmission across layers—giving it a more intelligent way to understand and process information.

Zhilin at GTC: Introducing Attention Residuals Learning selective memory, rather than mechanically accumulating everything, is the beauty of attention. Many of you have probably read Attention Is All You Need, the 2017 Transformer paper that brought “human-like” attention into the model’s field of view. From that point on, models no longer simply read everything in a mechanical way. Instead, they began to develop a sense of what matters more and what matters less across the text, choosing to retain the more important information. Recently, Kimi applied this idea of attention to the temporal dimension, then rotated it 90 degrees into the model’s depth dimension. This allows the model to have attention not only over time, but also throughout the process of information transmission across layers—giving it a more intelligent way to understand and process information.

Kimi.ai

115,162 views • 4 months ago

FOUR 600W GPUS. 384GB OF GDDR7. AND DELL MADE THE INSTALL LOOK EASIER THAN BUILDING A GAMING PC. this is a Dell PowerEdge XE7740 receiving four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. each card brings: 96GB GDDR7 ECC 24,064 CUDA cores 4 PFLOPS FP4 up to 600W but the most satisfying part is the cable management. every GPU power lead is already routed, labeled and held beside its assigned slot. drop in the card. lock the retention hardware. connect one 16-pin plug. no loose bundle hanging across the airflow path and no guessing which cable belongs where. the four-card build totals 384GB of GPU memory and 16 PFLOPS FP4. the 4U XE7740 supports up to eight of these GPUs, doubling that to 768GB and 32 PFLOPS FP4. honest line: the memory remains split between separate cards, and four GPUs can still pull up to 2.4kW. Dell made a rack-scale AI server this dense without turning the next GPU swap into cable surgery. bookmark & watch today ↓

FOUR 600W GPUS. 384GB OF GDDR7. AND DELL MADE THE INSTALL LOOK EASIER THAN BUILDING A GAMING PC. this is a Dell PowerEdge XE7740 receiving four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. each card brings: 96GB GDDR7 ECC 24,064 CUDA cores 4 PFLOPS FP4 up to 600W but the most satisfying part is the cable management. every GPU power lead is already routed, labeled and held beside its assigned slot. drop in the card. lock the retention hardware. connect one 16-pin plug. no loose bundle hanging across the airflow path and no guessing which cable belongs where. the four-card build totals 384GB of GPU memory and 16 PFLOPS FP4. the 4U XE7740 supports up to eight of these GPUs, doubling that to 768GB and 32 PFLOPS FP4. honest line: the memory remains split between separate cards, and four GPUs can still pull up to 2.4kW. Dell made a rack-scale AI server this dense without turning the next GPU swap into cable surgery. bookmark & watch today ↓

Grimmer

26,693 views • 5 days ago

🎉 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

40,306 views • 2 months ago

.Noise video for those still code hunting. This is me shorting Opensea and Succint. Things to note: • The UX is intentionally simple, clean and fast. Gud. • You are trading attention, not token prices. Here that is represented by Kaito mindshare on crypto projects (infra and apps). • It is not an app restricted to web3 as attention is present (and desired) in all industries. If there is an accepted metric which tracks the attention of fashion brands you can put it in to an oracle and long/short seasonal trends. (are joggers coming back!?) Bonus note: The seamless flow is a look into a future where 7702 is implemented everywhere. No more approvals. While 7702 is already live on Mega that is not what is implemented here (wallet support needed first). It's a view into the future though.

.Noise video for those still code hunting. This is me shorting Opensea and Succint. Things to note: • The UX is intentionally simple, clean and fast. Gud. • You are trading attention, not token prices. Here that is represented by Kaito mindshare on crypto projects (infra and apps). • It is not an app restricted to web3 as attention is present (and desired) in all industries. If there is an accepted metric which tracks the attention of fashion brands you can put it in to an oracle and long/short seasonal trends. (are joggers coming back!?) Bonus note: The seamless flow is a look into a future where 7702 is implemented everywhere. No more approvals. While 7702 is already live on Mega that is not what is implemented here (wallet support needed first). It's a view into the future though.

bread.mega

48,127 views • 1 year ago

1/5 Here's video #4 on The Art of Explanation. It looks at how effective communication and explanation is about substance - but it's about style too. Because we need people's attention if we're to get our message across. The other videos (and a link to pre-order) are below.

1/5 Here's video #4 on The Art of Explanation. It looks at how effective communication and explanation is about substance - but it's about style too. Because we need people's attention if we're to get our message across. The other videos (and a link to pre-order) are below.

Ros Atkins

189,319 views • 3 years ago

[Self-Attention] by Hand ✍️ Self-attention is what enables LLMs to understand context. How does it work? This exercise demonstrates how to calculate a 6-3 attention head by hand. Note that if we have two instances of this, we get 6-6 attention (i.e., multi-head attention, n=2). -- 𝗚𝗼𝗮𝗹 -- Transform [6D Features 🟧] to [3D Attention Weighted Features 🟦] -- 𝗪𝗮𝗹𝗸𝘁𝗵𝗿𝗼𝘂𝗴𝗵 -- [1] Given ↳ A set of 4 feature vectors (6-D): x1,x2,x3,x4 [2] Query, Key, Value ↳ Multiply features x's with linear transformation matrices WQ, WK, and WV, to obtain query vectors (q1,q2,q3,q4), key vectors (k1,k2,k3,k4), and value vectors (v1,v2,v3,v4). ↳ "Self" refers to the fact that both queries and keys are derived from the same set of features. [3] 🟪 Prepare for MatMul ↳ Copy query vectors ↳ Copy the transpose of key vectors [4] 🟪 MatMul ↳ Multiply K^T and Q ↳ This is equivalent to taking dot product between every pair of query and key vectors. ↳ The purpose is to use dot product as an estimate of the "matching score" between every key-value pair. ↳ This estimate makes sense because dot product is the numerator of Cosine Similarity between two vectors. [5] 🟨 Scale ↳ Scale each element by the square root of dk, which is the dimension of key vectors (dk=3). ↳ The purpose is to normalize the impact of the dk on matching scores, even if we scale dk to 32, 64, or 128. ↳ To simplify hand calculation, we approximate [ □/sqrt(3) ] with [ floor(□/2) ]. [6] 🟩 Softmax: e^x ↳ Raise e to the power of the number in each cell ↳ To simplify hand calculation, we approximate e^□ with 3^□. [7] 🟩 Softmax: ∑ ↳ Sum across each column [8] 🟩 Softmax: 1 / sum ↳ For each column, divide each element by the column sum ↳ The purpose is normalize each column so that the numbers sum to 1. In other words, each column is a probability distribution of attention, and we have four of them. ↳ The result is the Attention Weight Matrix (A) (yellow) [9] 🟦 MatMul ↳ Multiply the value vectors (Vs) with the Attention Weight Matrix (A) ↳ The results are the attention weighted features Zs. ↳ They are fed to the position-wise feed forward network in the next layer.

[Self-Attention] by Hand ✍️ Self-attention is what enables LLMs to understand context. How does it work? This exercise demonstrates how to calculate a 6-3 attention head by hand. Note that if we have two instances of this, we get 6-6 attention (i.e., multi-head attention, n=2). -- 𝗚𝗼𝗮𝗹 -- Transform [6D Features 🟧] to [3D Attention Weighted Features 🟦] -- 𝗪𝗮𝗹𝗸𝘁𝗵𝗿𝗼𝘂𝗴𝗵 -- [1] Given ↳ A set of 4 feature vectors (6-D): x1,x2,x3,x4 [2] Query, Key, Value ↳ Multiply features x's with linear transformation matrices WQ, WK, and WV, to obtain query vectors (q1,q2,q3,q4), key vectors (k1,k2,k3,k4), and value vectors (v1,v2,v3,v4). ↳ "Self" refers to the fact that both queries and keys are derived from the same set of features. [3] 🟪 Prepare for MatMul ↳ Copy query vectors ↳ Copy the transpose of key vectors [4] 🟪 MatMul ↳ Multiply K^T and Q ↳ This is equivalent to taking dot product between every pair of query and key vectors. ↳ The purpose is to use dot product as an estimate of the "matching score" between every key-value pair. ↳ This estimate makes sense because dot product is the numerator of Cosine Similarity between two vectors. [5] 🟨 Scale ↳ Scale each element by the square root of dk, which is the dimension of key vectors (dk=3). ↳ The purpose is to normalize the impact of the dk on matching scores, even if we scale dk to 32, 64, or 128. ↳ To simplify hand calculation, we approximate [ □/sqrt(3) ] with [ floor(□/2) ]. [6] 🟩 Softmax: e^x ↳ Raise e to the power of the number in each cell ↳ To simplify hand calculation, we approximate e^□ with 3^□. [7] 🟩 Softmax: ∑ ↳ Sum across each column [8] 🟩 Softmax: 1 / sum ↳ For each column, divide each element by the column sum ↳ The purpose is normalize each column so that the numbers sum to 1. In other words, each column is a probability distribution of attention, and we have four of them. ↳ The result is the Attention Weight Matrix (A) (yellow) [9] 🟦 MatMul ↳ Multiply the value vectors (Vs) with the Attention Weight Matrix (A) ↳ The results are the attention weighted features Zs. ↳ They are fed to the position-wise feed forward network in the next layer.

Tom Yeh

101,010 views • 2 years ago

🗓️ Release Week Recap Big week. mlx-audio and mlx-vlm are now among some of the fastest-growing OSS projects. Here’s what we shipped last week. Gemma 4 on Apple Silicon Two awesome releases by our partner Google DeepMind & Google Gemma : > Gemma 4 12B — their new dense, unified multimodal model. It uses an encoder free audio path and simplified vision encoder. > Gemma 4 QAT — quantization-aware training checkpoints, optimized to run locally on consumer GPUs and edge devices, compressing the model while preserving the quality you expect from Gemma 4. On the audio 🎧 side we added support for 15+ new TTS, ASR & VAD models, faster long-form transcription, and an expanded OpenAI-compatible audio server. All local on Apple Silicon. Huge thanks to every contributor and my co-maintainer Lucas Newman. 🙏🏽

🗓️ Release Week Recap Big week. mlx-audio and mlx-vlm are now among some of the fastest-growing OSS projects. Here’s what we shipped last week. Gemma 4 on Apple Silicon Two awesome releases by our partner Google DeepMind & Google Gemma : > Gemma 4 12B — their new dense, unified multimodal model. It uses an encoder free audio path and simplified vision encoder. > Gemma 4 QAT — quantization-aware training checkpoints, optimized to run locally on consumer GPUs and edge devices, compressing the model while preserving the quality you expect from Gemma 4. On the audio 🎧 side we added support for 15+ new TTS, ASR & VAD models, faster long-form transcription, and an expanded OpenAI-compatible audio server. All local on Apple Silicon. Huge thanks to every contributor and my co-maintainer Lucas Newman. 🙏🏽

Prince Canuma

11,731 views • 2 months ago

2 girls. 1 is human. 1 is AI. Who gets more sales? We’re about to enter a whole new world of marketing and content. I hope you’re paying attention! By the end of the year, AI will be 100% human-like. Right now it’s 70%. So the time to get started is now. The first 5 videos are free on

2 girls. 1 is human. 1 is AI. Who gets more sales? We’re about to enter a whole new world of marketing and content. I hope you’re paying attention! By the end of the year, AI will be 100% human-like. Right now it’s 70%. So the time to get started is now. The first 5 videos are free on

Nuseir Yassin

46,628 views • 5 months ago

Much respect to Pastor Mike Pastor Mike Behind Bars and his statement on the last Silver & Black Today 👊🏴‍☠️ This is a good way to get the attention of the organization that we fans are checking out. And sorry Pastor for my potty mouth at the end. #RaiderNation

Much respect to Pastor Mike Pastor Mike Behind Bars and his statement on the last Silver & Black Today 👊🏴‍☠️ This is a good way to get the attention of the organization that we fans are checking out. And sorry Pastor for my potty mouth at the end. #RaiderNation

Murf 🏴‍☠️

29,198 views • 9 months ago

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

493,571 views • 1 month ago

South Africa is anti-war and prefers dialogue over violence. Global attention is on this Conference. There is an expectation that we will deliver an effective response to the destruction of an entire population and a peaceful path for preserving the prospect of a viable Palestinian State existing side by side with the State of Israel in peace and security. This expectation is not misplaced, and it could not be higher.

South Africa is anti-war and prefers dialogue over violence. Global attention is on this Conference. There is an expectation that we will deliver an effective response to the destruction of an entire population and a peaceful path for preserving the prospect of a viable Palestinian State existing side by side with the State of Israel in peace and security. This expectation is not misplaced, and it could not be higher.

Minister: International Relations and Cooperation

24,434 views • 1 year ago

Self Attention by hand ✍️ ~ 9 steps walkthrough below Self-attention is what enables LLMs to understand context. How does it work? So I drew and calculated one entirely by hand. Goal: turn four 6D features into four 3D attention weighted features, filling in every cell yourself. = 1. Given = Four feature vectors, six dimensions each, one per position. = 2. Query, key, value = Let us multiply the features by WQ, WK and WV. Queries, keys and values all come out of the same four features, and that is what the word "self" is doing in self-attention. = 3. Prepare for MatMul = We copy the queries across the top and the transposed keys down the side. Lining the two up is half the work. = 4. MatMul = Let us multiply K transpose by Q. Every cell is the dot product of one key with one query, which we use as a matching score. That works because the dot product is the numerator of cosine similarity: it is how alike two vectors are, before anyone divides by their lengths. = 5. Scale = We divide by the square root of dk, the dimension of a key vector, here 3. Without it the scores grow with the dimension and a 64-wide head would swamp the softmax. To keep the page doable in pen, the drawing approximates dividing by root 3 with halving. = 6. e to the power = Let us raise e to the power of each score. This is the first half of softmax, and the drawing uses 3 in place of e, which is close enough to do in your head. = 7. Sum = We add up each column: 16, 6, 7 and 12. = 8. Normalize = Let us divide every cell by its column sum. That gives the attention weight matrix in yellow, and each of its four columns is now a probability distribution over the four positions. The decimals are nudged as they are rounded, so every column still sums to exactly 1. = 9. MatMul = We multiply the value vectors by those weights. Each output is a blend of all four values, mixed in the proportion the attention matrix just decided, and it goes to the position-wise feed forward network in the next layer: the FFN box at the bottom of the page. The outputs: Attention weights (A), by column = [.2, .6, 0, .2], [.2, .4, .2, .2], [.4, .2, 0, .4], [.1, .7, .1, .1] Attention weighted features (Z) = [8, 2, 6], [8, 4, 4], [16, 4, 2], [4, 2, 7] The takeaway: attention is a weighted average, and everything before step 9 exists to decide the weights. Compare every position with every other, turn the scores into one distribution per position, then blend. 💾 Save this post!

Self Attention by hand ✍️ ~ 9 steps walkthrough below Self-attention is what enables LLMs to understand context. How does it work? So I drew and calculated one entirely by hand. Goal: turn four 6D features into four 3D attention weighted features, filling in every cell yourself. = 1. Given = Four feature vectors, six dimensions each, one per position. = 2. Query, key, value = Let us multiply the features by WQ, WK and WV. Queries, keys and values all come out of the same four features, and that is what the word "self" is doing in self-attention. = 3. Prepare for MatMul = We copy the queries across the top and the transposed keys down the side. Lining the two up is half the work. = 4. MatMul = Let us multiply K transpose by Q. Every cell is the dot product of one key with one query, which we use as a matching score. That works because the dot product is the numerator of cosine similarity: it is how alike two vectors are, before anyone divides by their lengths. = 5. Scale = We divide by the square root of dk, the dimension of a key vector, here 3. Without it the scores grow with the dimension and a 64-wide head would swamp the softmax. To keep the page doable in pen, the drawing approximates dividing by root 3 with halving. = 6. e to the power = Let us raise e to the power of each score. This is the first half of softmax, and the drawing uses 3 in place of e, which is close enough to do in your head. = 7. Sum = We add up each column: 16, 6, 7 and 12. = 8. Normalize = Let us divide every cell by its column sum. That gives the attention weight matrix in yellow, and each of its four columns is now a probability distribution over the four positions. The decimals are nudged as they are rounded, so every column still sums to exactly 1. = 9. MatMul = We multiply the value vectors by those weights. Each output is a blend of all four values, mixed in the proportion the attention matrix just decided, and it goes to the position-wise feed forward network in the next layer: the FFN box at the bottom of the page. The outputs: Attention weights (A), by column = [.2, .6, 0, .2], [.2, .4, .2, .2], [.4, .2, 0, .4], [.1, .7, .1, .1] Attention weighted features (Z) = [8, 2, 6], [8, 4, 4], [16, 4, 2], [4, 2, 7] The takeaway: attention is a weighted average, and everything before step 9 exists to decide the weights. Compare every position with every other, turn the scores into one distribution per position, then blend. 💾 Save this post!

Tom Yeh

27,049 views • 23 days ago

$PrismML Releases Bonsai 27B: 1-bit and Ternary Builds of Qwen3.6-27B Hitting 89.5% of FP16 at 3.9GB. No new pretrain. No higher-precision escape hatches. No multi-GPU rig. Here's how it works. 👇 1: Codes, not floats Every weight becomes a code, with one shared FP16 scale per group of 128. Ternary is {−1, 0, +1}, binary is {−1, +1}. Sharing the scale across 128 weights keeps its cost at 16/128 = 0.125 bits. → Ternary: log2(3) + 16/128 ≈ 1.71 bits/weight → 5.9GB → Binary: 1 + 16/128 = 1.125 bits/weight → 3.9GB 2: Post-training, not from scratch No BitNet-style low-bit pretrain. It starts from off-the-shelf Qwen3.6-27B, architecture unchanged. The representation runs end to end across embeddings, attention projections, MLP projections, and the LM head. → 9.4× (ternary) and 14.2× (binary) vs the 54GB FP16 baseline 3: Labels are not bit-widths Conventional low-bit builds are mixed-precision by construction. The advertised name describes the most-compressed tensors, not the model. → Q4_K_XL, labeled "4-bit," is really 5.2 bits/weight at 17.6GB → IQ2_XXS, labeled "2-bit," is really 2.8 bits/weight at 9.4GB 4: Fitting a phone is two budgets iOS caps a single app near half of RAM, so a 12GB iPhone exposes ~6GB. The KV cache grows on top. Hybrid attention at ~75% linear means only 16 of 64 layers cache. → 4-bit KV: 4.3GB at 262K context, down from 17.2GB → 11.0 tok/s on iPhone 17 Pro Max 5: The numbers (15 benchmarks, thinking mode) → Ternary: 80.49 avg at 5.9GB — 94.6% of FP16 → 1-bit: 76.11 avg at 3.9GB — 89.5% of FP16 → IQ2_XXS falls to 57.5 on AIME26 while still scoring 88.93 on MMLU-Redux The key takeaway: 27B-class reasoning without the 54GB checkpoint — group-wise ternary and binary codes, an end-to-end low-bit language stack, 4-bit KV, on one phone. Full analysis: Repo: Model weight: Technical details: PrismML$

PrismML Releases Bonsai 27B: 1-bit and Ternary Builds of Qwen3.6-27B Hitting 89.5% of FP16 at 3.9GB. No new pretrain. No higher-precision escape hatches. No multi-GPU rig. Here's how it works. 👇 1: Codes, not floats Every weight becomes a code, with one shared FP16 scale per group of 128. Ternary is {−1, 0, +1}, binary is {−1, +1}. Sharing the scale across 128 weights keeps its cost at 16/128 = 0.125 bits. → Ternary: log2(3) + 16/128 ≈ 1.71 bits/weight → 5.9GB → Binary: 1 + 16/128 = 1.125 bits/weight → 3.9GB 2: Post-training, not from scratch No BitNet-style low-bit pretrain. It starts from off-the-shelf Qwen3.6-27B, architecture unchanged. The representation runs end to end across embeddings, attention projections, MLP projections, and the LM head. → 9.4× (ternary) and 14.2× (binary) vs the 54GB FP16 baseline 3: Labels are not bit-widths Conventional low-bit builds are mixed-precision by construction. The advertised name describes the most-compressed tensors, not the model. → Q4_K_XL, labeled "4-bit," is really 5.2 bits/weight at 17.6GB → IQ2_XXS, labeled "2-bit," is really 2.8 bits/weight at 9.4GB 4: Fitting a phone is two budgets iOS caps a single app near half of RAM, so a 12GB iPhone exposes ~6GB. The KV cache grows on top. Hybrid attention at ~75% linear means only 16 of 64 layers cache. → 4-bit KV: 4.3GB at 262K context, down from 17.2GB → 11.0 tok/s on iPhone 17 Pro Max 5: The numbers (15 benchmarks, thinking mode) → Ternary: 80.49 avg at 5.9GB — 94.6% of FP16 → 1-bit: 76.11 avg at 3.9GB — 89.5% of FP16 → IQ2_XXS falls to 57.5 on AIME26 while still scoring 88.93 on MMLU-Redux The key takeaway: 27B-class reasoning without the 54GB checkpoint — group-wise ternary and binary codes, an end-to-end low-bit language stack, 4-bit KV, on one phone. Full analysis: Repo: Model weight: Technical details: PrismML

Marktechpost AI

31,860 views • 1 month ago

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

259,993 views • 2 months ago

The idea that Tesla is alone in the race to develop end-to-end vision system for self-driving and delivering into consumer vehicles is officially dead. Like dead, dead. I tested Xpeng's VLA 2.0 in the streets of Beijing, and it is comparable to my experience of daily driving Tesla FSD v14. Xpeng is not charging $100 a month to drive this. It is inlcuded in its higher-trim vehicles and it already caught the attention of Volkswagen, which is going to integrate into its own vehicles - something Tesla hasn't been able to do despite trying. You can watch the full 40-minute drive on Electrek.co

The idea that Tesla is alone in the race to develop end-to-end vision system for self-driving and delivering into consumer vehicles is officially dead. Like dead, dead. I tested Xpeng's VLA 2.0 in the streets of Beijing, and it is comparable to my experience of daily driving Tesla FSD v14. Xpeng is not charging $100 a month to drive this. It is inlcuded in its higher-trim vehicles and it already caught the attention of Volkswagen, which is going to integrate into its own vehicles - something Tesla hasn't been able to do despite trying. You can watch the full 40-minute drive on Electrek.co

Fred Lambert

2,142,894 views • 3 months ago