Video yükleniyor...

Video Yüklenemedi

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

Ana Sayfaya Dön

Introducing DuoAttention: Our new framework slashes both memory and latency for long-context LLMs without sacrificing performance! By applying full KV cache only to critical heads, we achieve: ⚡ 2.55x memory reduction ⚡ 2.18x decoding speedup ⚡ 3.3M tokens on a single A100 GPU

Guangxuan Xiao

5,414 subscribers

31,085 görüntüleme • 1 yıl önce •via X (Twitter)

Bilim & Teknoloji Eğitim

Anya Rossi• Live Now

Private livecam show

10 Yorum

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

Paper: Code:

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

DuoAttention leverages the insight that only a few attention heads, called Retrieval Heads, need full attention for long contexts, while the rest, Streaming Heads, focus on recent tokens and don't require full attention.

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately.

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

We apply a constant KV cache and efficient "Streaming Attention" to streaming heads, which can accelerate both pre-filling and decoding.

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

DuoAttention provides comparable accuracy as full attention on the Needle-in-a-Haystack benchmark using a 25% full attention ratio on the MHA model and a 50% full attention ratio on the GQA model.

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

DuoAttention provides a better KV budget and accuracy trade-off on LongBench benchmarks.

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

DuoAttention significantly reduces long-context inference memory by up to 2.55× for MHA and 1.67× for GQA models while speeding up decoding by up to 2.18× and 1.50×.

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

DuoAttention can also accelerate pre-filling by up to 1.73× and 1.63× for MHA and GQA models, respectively.

Guangxuan Xiao profil fotoğrafı

Guangxuan Xiao1 yıl önce

Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU.

Miguel Guerrero profil fotoğrafı

Miguel Guerrero1 yıl önce

guys, you always build and share awesome stuff with the community, just bravo!! thanks

Benzer Videolar

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

Reese Chong

53,026 görüntüleme • 4 ay önce

GPU tradeoff series: A100 is not much more powerful than 4090 🫠 GPU Perf and Price: - 4090: 330 fp16 TFLOPs, $1,749 - A100 (80GB): 312 fp16 TFLOPs, $20,000 > A100 is 11.4X more pricy Training speed for GPT-2(124M) with llm.c: - 4090: 153K tokens/s - A100 (80GB): 195K tokens/s > A100 is only 1.3X faster (both trained using a single card, A100 llm.c training is shown in the video, 4090 video is in the quoted tweet) Conclusion: 4090 has a much better cost vs performance ratio Why: As in the H100 vs. 4090 comparison, the biggest difference between A100 and 4090 is their GPU memory size/bandwidth and cross-GPU communication bandwidth, which does not matter too much if your model can fit into a single 4090. Specs: 4090: - GPU memory size: 24GB - memory bandwidth: 1 TB/s - communication bandwidth: 64 GB/s A100: - GPU memory size: 80GB - memory bandwidth: 2 TB/s - communication bandwidth: 900 GB/s Nvidia killed off NVLink (a high-speed communication link that connects GPUs) on 4090. (Jensen Huang smiling face) If multiple 4090s could be interconnected via NVLink, their performance would be closer to datacenter-grade A100 GPUs, even for training larger models. Additionally, 4090 isn't allowed in datacenters, that's how Nvidia makes 💰💰💰

GPU tradeoff series: A100 is not much more powerful than 4090 🫠 GPU Perf and Price: - 4090: 330 fp16 TFLOPs, $1,749 - A100 (80GB): 312 fp16 TFLOPs, $20,000 > A100 is 11.4X more pricy Training speed for GPT-2(124M) with llm.c: - 4090: 153K tokens/s - A100 (80GB): 195K tokens/s > A100 is only 1.3X faster (both trained using a single card, A100 llm.c training is shown in the video, 4090 video is in the quoted tweet) Conclusion: 4090 has a much better cost vs performance ratio Why: As in the H100 vs. 4090 comparison, the biggest difference between A100 and 4090 is their GPU memory size/bandwidth and cross-GPU communication bandwidth, which does not matter too much if your model can fit into a single 4090. Specs: 4090: - GPU memory size: 24GB - memory bandwidth: 1 TB/s - communication bandwidth: 64 GB/s A100: - GPU memory size: 80GB - memory bandwidth: 2 TB/s - communication bandwidth: 900 GB/s Nvidia killed off NVLink (a high-speed communication link that connects GPUs) on 4090. (Jensen Huang smiling face) If multiple 4090s could be interconnected via NVLink, their performance would be closer to datacenter-grade A100 GPUs, even for training larger models. Additionally, 4090 isn't allowed in datacenters, that's how Nvidia makes 💰💰💰

Yuchen Jin

234,951 görüntüleme • 2 yıl önce

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

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

Avi Chawla

271,839 görüntüleme • 2 ay önce

New course on serving LLMs efficiently -- how do you serve models to many concurrent users at low latency and reasonable cost? This short course is built with Red Hat and taught by Cedric Clyburn. Efficient LLM serving requires efficient memory management. A 70B-parameter model takes ~140 GB just to load the weights. On top of that, every active request needs its own chunk of GPU memory, the KV cache, to store the token context it has built up so far. In this course, you'll learn to reduce a model's memory footprint with quantization and serve it using vLLM, which handles many concurrent requests efficiently through smart memory management. Skills you'll gain: - Quantize a model and measure the accuracy tradeoff - Serve a model with vLLM and watch it handle concurrent requests efficiently - Benchmark your deployment and make informed tradeoffs between speed, cost, and accuracy Join and learn to serve LLMs efficiently:

New course on serving LLMs efficiently -- how do you serve models to many concurrent users at low latency and reasonable cost? This short course is built with Red Hat and taught by Cedric Clyburn. Efficient LLM serving requires efficient memory management. A 70B-parameter model takes ~140 GB just to load the weights. On top of that, every active request needs its own chunk of GPU memory, the KV cache, to store the token context it has built up so far. In this course, you'll learn to reduce a model's memory footprint with quantization and serve it using vLLM, which handles many concurrent requests efficiently through smart memory management. Skills you'll gain: - Quantize a model and measure the accuracy tradeoff - Serve a model with vLLM and watch it handle concurrent requests efficiently - Benchmark your deployment and make informed tradeoffs between speed, cost, and accuracy Join and learn to serve LLMs efficiently:

Andrew Ng

134,352 görüntüleme • 3 ay önce

New short course: LLMs as Operating Systems: Agent Memory, created with Letta, and taught by its founders Charles Packer and Sarah Wooders. An LLM's input context window has limited space. Using a longer input context also costs more and results in slower processing. So, managing what's stored in this context window is important. In the innovative paper MemGPT: Towards LLMs as Operating Systems, its authors (which include the instructors) proposed using an LLM agent to manage this context window. Their system uses a large persistent memory that stores everything that could be included in the input context, and an agent decides what is actually included. Take the example of building a chatbot that needs to remember what's been said earlier in a conversation (perhaps over many days of interaction with a user). As the conversation's length grows, the memory management agent will move information from the input context to a persistent searchable database; summarize information to keep relevant facts in the input context; and restore relevant conversation elements from further back in time. This allows a chatbot to keep what's currently most relevant in its input context memory to generate the next response. When I read the original MemGPT paper, I thought it was an innovative technique for handling memory for LLMs. The open-source Letta framework, which we'll use in this course, makes MemGPT easy to implement. It adds memory to your LLM agents and gives them transparent long-term memory. In detail, you’ll learn: - How to build an agent that can edit its own limited input context memory, using tools and multi-step reasoning - What is a memory hierarchy (an idea from computer operating systems, which use a cache to speed up memory access), and how these ideas apply to managing the LLM input context (where the input context window is a "cache" storing the most relevant information; and an agent decides what to move in and out of this to/from a larger persistent storage system) - How to implement multi-agent collaboration by letting different agents share blocks of memory This course will give you a sophisticated understanding of memory management for LLMs, which is important for chatbots having long conversations, and for complex agentic workflows. Please sign up here!

New short course: LLMs as Operating Systems: Agent Memory, created with Letta, and taught by its founders Charles Packer and Sarah Wooders. An LLM's input context window has limited space. Using a longer input context also costs more and results in slower processing. So, managing what's stored in this context window is important. In the innovative paper MemGPT: Towards LLMs as Operating Systems, its authors (which include the instructors) proposed using an LLM agent to manage this context window. Their system uses a large persistent memory that stores everything that could be included in the input context, and an agent decides what is actually included. Take the example of building a chatbot that needs to remember what's been said earlier in a conversation (perhaps over many days of interaction with a user). As the conversation's length grows, the memory management agent will move information from the input context to a persistent searchable database; summarize information to keep relevant facts in the input context; and restore relevant conversation elements from further back in time. This allows a chatbot to keep what's currently most relevant in its input context memory to generate the next response. When I read the original MemGPT paper, I thought it was an innovative technique for handling memory for LLMs. The open-source Letta framework, which we'll use in this course, makes MemGPT easy to implement. It adds memory to your LLM agents and gives them transparent long-term memory. In detail, you’ll learn: - How to build an agent that can edit its own limited input context memory, using tools and multi-step reasoning - What is a memory hierarchy (an idea from computer operating systems, which use a cache to speed up memory access), and how these ideas apply to managing the LLM input context (where the input context window is a "cache" storing the most relevant information; and an agent decides what to move in and out of this to/from a larger persistent storage system) - How to implement multi-agent collaboration by letting different agents share blocks of memory This course will give you a sophisticated understanding of memory management for LLMs, which is important for chatbots having long conversations, and for complex agentic workflows. Please sign up here!

Andrew Ng

201,127 görüntüleme • 1 yıl önce

🚀 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

59,615 görüntüleme • 3 ay önce

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

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

elvis

261,622 görüntüleme • 2 yıl önce

A real-time world simulator, now in 0.5B. ABot-World-0.5B-LF brings action-conditioned interactive rollout to ModelScope. 🎮License: Apache 2.0. 🤖 ⚡ Desktop-ready rollout: reported at 720p, 16 FPS, 1.2s latency, and 19GB GPU memory on a single-desktop setup. 🌍 Infinite interaction: responds to user actions continuously and keeps generating beyond fixed video-length limits. 🎬 LongForcing training helps the model expand new scenes and dynamics during rollout, without prompt switching or scene lock-in. 🛠️ Built on Wan2.2-TI2V-5B as a causal student model for lower-latency interactive world rollout.

A real-time world simulator, now in 0.5B. ABot-World-0.5B-LF brings action-conditioned interactive rollout to ModelScope. 🎮License: Apache 2.0. 🤖 ⚡ Desktop-ready rollout: reported at 720p, 16 FPS, 1.2s latency, and 19GB GPU memory on a single-desktop setup. 🌍 Infinite interaction: responds to user actions continuously and keeps generating beyond fixed video-length limits. 🎬 LongForcing training helps the model expand new scenes and dynamics during rollout, without prompt switching or scene lock-in. 🛠️ Built on Wan2.2-TI2V-5B as a causal student model for lower-latency interactive world rollout.

ModelScope

19,642 görüntüleme • 2 ay önce

Qwen3.8-Flash-Next is still going strong at 364.7K tokens of context on an M5 Max. And this isn’t just a static long-context test. The model was reasoning about how to speed up its own workflow while using tools, and the tool calls kept working without misses. Setup: • Qwen3.8-Flash-Next • M5 Max • 128GB unified memory • MLX-Serve PR #363 • OpenCode 2 • 364.7K context The interesting part isn’t simply getting hundreds of thousands of tokens into memory. It’s what happens once the context gets this large. Long-context inference usually comes with a painful tradeoff. As the KV cache grows, memory pressure increases and generation can slow down. But this setup is still pushing through 364K tokens while maintaining a usable agent workflow. The model can reason, call tools, inspect results, continue working, and keep the session moving. And the tool calls reportedly haven’t missed so far. That’s important for agentic coding. A huge context window is only useful if the model can actually operate reliably inside it. A 400K-token context that constantly breaks tool calls isn’t very useful. A 364K session that can keep reasoning and executing tools is a different story. And the test isn’t finished yet. The current run is approaching 400K tokens, with the expectation that it can keep going. This is also another interesting example of why Apple Silicon keeps showing up in local LLM experiments. The M5 Max’s unified memory gives a large model and its growing KV cache access to one shared memory pool. With MLX-Serve continuing to improve, these machines are becoming surprisingly capable long-context inference boxes. The bigger takeaway: Context length is becoming a workload, not just a model specification. Running a model at 256K is one thing. Keeping an agent alive at 300K+ while it reasons and uses tools is much more interesting. And Qwen3.8-Flash-Next is showing that this can be pushed surprisingly far on a single 128GB Mac. 364.7K and counting. Next stop: 400K.

Qwen3.8-Flash-Next is still going strong at 364.7K tokens of context on an M5 Max. And this isn’t just a static long-context test. The model was reasoning about how to speed up its own workflow while using tools, and the tool calls kept working without misses. Setup: • Qwen3.8-Flash-Next • M5 Max • 128GB unified memory • MLX-Serve PR #363 • OpenCode 2 • 364.7K context The interesting part isn’t simply getting hundreds of thousands of tokens into memory. It’s what happens once the context gets this large. Long-context inference usually comes with a painful tradeoff. As the KV cache grows, memory pressure increases and generation can slow down. But this setup is still pushing through 364K tokens while maintaining a usable agent workflow. The model can reason, call tools, inspect results, continue working, and keep the session moving. And the tool calls reportedly haven’t missed so far. That’s important for agentic coding. A huge context window is only useful if the model can actually operate reliably inside it. A 400K-token context that constantly breaks tool calls isn’t very useful. A 364K session that can keep reasoning and executing tools is a different story. And the test isn’t finished yet. The current run is approaching 400K tokens, with the expectation that it can keep going. This is also another interesting example of why Apple Silicon keeps showing up in local LLM experiments. The M5 Max’s unified memory gives a large model and its growing KV cache access to one shared memory pool. With MLX-Serve continuing to improve, these machines are becoming surprisingly capable long-context inference boxes. The bigger takeaway: Context length is becoming a workload, not just a model specification. Running a model at 256K is one thing. Keeping an agent alive at 300K+ while it reasons and uses tools is much more interesting. And Qwen3.8-Flash-Next is showing that this can be pushed surprisingly far on a single 128GB Mac. 364.7K and counting. Next stop: 400K.

FHILY👑

38,506 görüntüleme • 1 gün önce

New course: Agent Memory: Building Memory-Aware Agents, built in partnership with Oracle and taught by Richmond Alake and Nacho Martínez. Many agents work well within a single session but their memory resets once the session ends. Consider a research agent working on dozens of papers across multiple days: without memory, it has no way to store and retrieve what it learned across sessions. This short course teaches you to build a memory system that enables agents to persist memory and thereby learn across sessions. You'll design a Memory Manager that handles different memory types, implement semantic tool retrieval that scales without bloating the context, and build write-back pipelines that let your agent autonomously update and refine what it knows over time. Skills you'll gain: - Build persistent memory stores for different agent memory types - Implement a Memory Manager that orchestrates how your agent reads, writes, and retrieves memory - Treat tools as procedural memory and retrieve only relevant ones at inference time using semantic search Join and learn to build agents that remember and improve over time!

New course: Agent Memory: Building Memory-Aware Agents, built in partnership with Oracle and taught by Richmond Alake and Nacho Martínez. Many agents work well within a single session but their memory resets once the session ends. Consider a research agent working on dozens of papers across multiple days: without memory, it has no way to store and retrieve what it learned across sessions. This short course teaches you to build a memory system that enables agents to persist memory and thereby learn across sessions. You'll design a Memory Manager that handles different memory types, implement semantic tool retrieval that scales without bloating the context, and build write-back pipelines that let your agent autonomously update and refine what it knows over time. Skills you'll gain: - Build persistent memory stores for different agent memory types - Implement a Memory Manager that orchestrates how your agent reads, writes, and retrieves memory - Treat tools as procedural memory and retrieve only relevant ones at inference time using semantic search Join and learn to build agents that remember and improve over time!

Andrew Ng

161,260 görüntüleme • 5 ay önce

🧠 Can agent memory scale without losing reasoning? 🔥 We’re excited to share our latest work, SimpleMem, a principled memory framework for LLM agents built around semantic lossless compression. 📉 30× fewer inference tokens 📈 +26.4% avg F1 (vs Mem0) ⚡ 50.2% faster retrieval (vs Mem0) Instead of storing raw interaction history 🗂️ or relying on costly iterative reasoning loops 🔁, SimpleMem treats memory as a structured, evolving representation whose primary objective is 🎯 maximizing information density per token. 📄 Paper: 🔗 Code: 📦 Website： Nice work @JiaqiLiu835914, Yaofeng Su, Peng (Richard) Xia, Siwei Han, and great collab. w/ Cihang Xie, Zeyu Zheng, Mingyu Ding

🧠 Can agent memory scale without losing reasoning? 🔥 We’re excited to share our latest work, SimpleMem, a principled memory framework for LLM agents built around semantic lossless compression. 📉 30× fewer inference tokens 📈 +26.4% avg F1 (vs Mem0) ⚡ 50.2% faster retrieval (vs Mem0) Instead of storing raw interaction history 🗂️ or relying on costly iterative reasoning loops 🔁, SimpleMem treats memory as a structured, evolving representation whose primary objective is 🎯 maximizing information density per token. 📄 Paper: 🔗 Code: 📦 Website： Nice work @JiaqiLiu835914, Yaofeng Su, Peng (Richard) Xia, Siwei Han, and great collab. w/ Cihang Xie, Zeyu Zheng, Mingyu Ding

Huaxiu Yao

119,812 görüntüleme • 8 ay önce

QVAC SDK 0.12.0 is now live, bringing longer context, increased memory optimisation, new modalities, and broader ecosystem support directly to your device. Key Features and Updates: - TurboQuant KV-Cache Quantization: Fit much longer context in the same memory. TurboQuant, an algorithm from Google Research, compresses the KV cache by up to 5x, near-lossless. - Text-to-Video: Generate video from a text prompt, fully local, with the new wan2.1 model in the Diffusion addon - Apple Metal Performance for Flux2-klein: Diffusion on Apple Silicon now matches MLX performance, the native benchmark for Apple GPUs - Robot Control (new VLA addon): A GGML-based Vision-Language-Action addon brings fast, efficient robot control to edge devices - Coding Assistant / Harness Support: QVAC now works with OpenCode and OpenClaw as a local provider. A new @qvac/ai-sdk-provider package automates model registry and provider integration - Cross-Platform Voice: Text-to-speech and Parakeet transcription moved from ONNX to the GGML engine for better CPU and GPU support on macOS, iOS, Windows, Linux, and Android. Parakeet also adds long-term streaming diarization (tracking who spoke when on live audio) - Faster Lightweight Visual Classification: A new GGML-based Classification addon delivers millisecond-level classification, useful where a vision-language model (VLM) would be unnecessarily slow - Under the Hood: Fabric synced to llama.cpp v8828 (from v8189), plus GPU acceleration added to image-upscale models for faster results Full release notes:

QVAC SDK 0.12.0 is now live, bringing longer context, increased memory optimisation, new modalities, and broader ecosystem support directly to your device. Key Features and Updates: - TurboQuant KV-Cache Quantization: Fit much longer context in the same memory. TurboQuant, an algorithm from Google Research, compresses the KV cache by up to 5x, near-lossless. - Text-to-Video: Generate video from a text prompt, fully local, with the new wan2.1 model in the Diffusion addon - Apple Metal Performance for Flux2-klein: Diffusion on Apple Silicon now matches MLX performance, the native benchmark for Apple GPUs - Robot Control (new VLA addon): A GGML-based Vision-Language-Action addon brings fast, efficient robot control to edge devices - Coding Assistant / Harness Support: QVAC now works with OpenCode and OpenClaw as a local provider. A new @qvac/ai-sdk-provider package automates model registry and provider integration - Cross-Platform Voice: Text-to-speech and Parakeet transcription moved from ONNX to the GGML engine for better CPU and GPU support on macOS, iOS, Windows, Linux, and Android. Parakeet also adds long-term streaming diarization (tracking who spoke when on live audio) - Faster Lightweight Visual Classification: A new GGML-based Classification addon delivers millisecond-level classification, useful where a vision-language model (VLM) would be unnecessarily slow - Under the Hood: Fabric synced to llama.cpp v8828 (from v8189), plus GPU acceleration added to image-upscale models for faster results Full release notes:

QVAC

9,932,369 görüntüleme • 3 ay önce

NVIDIA sent us 2 DGX Sparks. For a while we wondered what we would do with them. The memory bandwidth is 273GB/s making it 3x slower than an M3 Ultra (819GB/s) for batch_size=1 inference. But it has 4x more FLOPS (100 TFLOPS compared to 26 TFLOPS). So we thought, what if we could combine the DGX Spark & M3 Ultra, and make use of both the massive compute on the DGX Spark and the massive memory-bandwidth on the M3 Ultra. We came up with a way to split inference across both devices and achieve a speedup of up to 4x for long prompts compared to the M3 Ultra on its own. Full details in the blog post linked below.

NVIDIA sent us 2 DGX Sparks. For a while we wondered what we would do with them. The memory bandwidth is 273GB/s making it 3x slower than an M3 Ultra (819GB/s) for batch_size=1 inference. But it has 4x more FLOPS (100 TFLOPS compared to 26 TFLOPS). So we thought, what if we could combine the DGX Spark & M3 Ultra, and make use of both the massive compute on the DGX Spark and the massive memory-bandwidth on the M3 Ultra. We came up with a way to split inference across both devices and achieve a speedup of up to 4x for long prompts compared to the M3 Ultra on its own. Full details in the blog post linked below.

Alex Cheema

281,225 görüntüleme • 10 ay önce

🚨 Open-source AI just got serious. Alibaba just released Qwen 3.5-Plus 🤯 397B parameters. Only 17B active per inference. Performance comparable to Gemini 3 Pro — at just $0.8 per million tokens (1/18 the cost). But here’s what makes it different 👇 This isn’t just another benchmark drop. In the demo, Qwen runs directly inside the terminal: ➙ Reviews GitHub PRs ➙ Fetches diffs automatically ➙ Runs CLI & smoke tests ➙ Generates integration tests ➙ Executes /context ➙ Shows real-time token usage breakdown That’s not a chatbot. That’s a real developer workflow. Under the hood: ⚡ 60% lower deployment memory ⚡ Inference throughput increased by 8.6× (32K context) ⚡ Inference throughput increased by up to 19× (256K ultra-long context) ⚡ Native multimodal model ⚡ Supports direct video input (up to 2 hours) And the ecosystem is massive: ✔ 201 languages ✔ 250K vocabulary ✔ 400+ open-sourced models ✔ 1B+ downloads ✔ 200K+ derivative models Available now via Qwen App, MoDa Community, Hugging Face, and Alibaba Cloud Bailian. ✅ The AI race isn’t about who’s biggest anymore. → It’s about who’s usable. 🔥 Watch the demo 👇 #Qwen #OpenSourceAI #AI #DevTools

🚨 Open-source AI just got serious. Alibaba just released Qwen 3.5-Plus 🤯 397B parameters. Only 17B active per inference. Performance comparable to Gemini 3 Pro — at just $0.8 per million tokens (1/18 the cost). But here’s what makes it different 👇 This isn’t just another benchmark drop. In the demo, Qwen runs directly inside the terminal: ➙ Reviews GitHub PRs ➙ Fetches diffs automatically ➙ Runs CLI & smoke tests ➙ Generates integration tests ➙ Executes /context ➙ Shows real-time token usage breakdown That’s not a chatbot. That’s a real developer workflow. Under the hood: ⚡ 60% lower deployment memory ⚡ Inference throughput increased by 8.6× (32K context) ⚡ Inference throughput increased by up to 19× (256K ultra-long context) ⚡ Native multimodal model ⚡ Supports direct video input (up to 2 hours) And the ecosystem is massive: ✔ 201 languages ✔ 250K vocabulary ✔ 400+ open-sourced models ✔ 1B+ downloads ✔ 200K+ derivative models Available now via Qwen App, MoDa Community, Hugging Face, and Alibaba Cloud Bailian. ✅ The AI race isn’t about who’s biggest anymore. → It’s about who’s usable. 🔥 Watch the demo 👇 #Qwen #OpenSourceAI #AI #DevTools

Toha Khan

123,375 görüntüleme • 6 ay önce

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

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

Andrew Ng

104,727 görüntüleme • 2 yıl önce

Memory Skill for OpenClaw with 26k+ users in 1 week🚀 OpenClaw's memory system is broken by default. It requires curating massive MEMORY.md files or relying on duplicate-heavy generation. Hours are wasted tuning, and massive amounts of tokens are burned. It's time to stop. So we built the memory skill to solve that prob Here is our superpower ⚔️ 🎯 Top #1 market accuracy (92.19%) after 8+ months of intense architecture iteration 🧠 The ultimate solution to keep the timeline, facts, and meaning perfectly in place ☁️ Local & Cloud + Version control ⚡ Super easy setup

Memory Skill for OpenClaw with 26k+ users in 1 week🚀 OpenClaw's memory system is broken by default. It requires curating massive MEMORY.md files or relying on duplicate-heavy generation. Hours are wasted tuning, and massive amounts of tokens are burned. It's time to stop. So we built the memory skill to solve that prob Here is our superpower ⚔️ 🎯 Top #1 market accuracy (92.19%) after 8+ months of intense architecture iteration 🧠 The ultimate solution to keep the timeline, facts, and meaning perfectly in place ☁️ Local & Cloud + Version control ⚡ Super easy setup

andy nguyen

589,836 görüntüleme • 6 ay önce

SITUATION EXPLAINED: Cerebras raised $5.55 billion in their IPO and closing their first day of trading valued at $66 billion, making it the biggest US tech IPO since Snowflake in 2020. Cerebras makes Wafer-Scale Engine chips built for AI inference. We asked Sarah Fong the main difference between wafer-scale chips and traditional GPUs: - GPUs are great at parallel work (graphics, training) - AI inference is sequential, AKA one token at a time This causes the "memory wall" problem: - Every GPU core needs model weights, KV cache, and activations to do its math - On a GPU, that data lives in off-chip memory (HBM) - Cores constantly load and offload from off-chip memory, which is a huge bottleneck; hardware accounts for ~70% of inference latency Cerebras' chips: -Dinner-plate sized (vs. GPUs which are palm-sized) with tens of thousands of cores -Memory sits directly on top of the cores as distributed SRAM -Weights and KV cache can be accessed at on-chip speeds in the PB/s range, compared with off-chip speeds in the TB/s range achieved by GPUs with HBM.

SITUATION EXPLAINED: Cerebras raised $5.55 billion in their IPO and closing their first day of trading valued at $66 billion, making it the biggest US tech IPO since Snowflake in 2020. Cerebras makes Wafer-Scale Engine chips built for AI inference. We asked Sarah Fong the main difference between wafer-scale chips and traditional GPUs: - GPUs are great at parallel work (graphics, training) - AI inference is sequential, AKA one token at a time This causes the "memory wall" problem: - Every GPU core needs model weights, KV cache, and activations to do its math - On a GPU, that data lives in off-chip memory (HBM) - Cores constantly load and offload from off-chip memory, which is a huge bottleneck; hardware accounts for ~70% of inference latency Cerebras' chips: -Dinner-plate sized (vs. GPUs which are palm-sized) with tens of thousands of cores -Memory sits directly on top of the cores as distributed SRAM -Weights and KV cache can be accessed at on-chip speeds in the PB/s range, compared with off-chip speeds in the TB/s range achieved by GPUs with HBM.

MTS

44,057 görüntüleme • 3 ay önce

Robot AI brains, aka Vision-Language-Action models, cannot adapt to new tasks as easily as LLMs like Gemini, ChatGPT, or Grok. LLMs can adapt quickly with their in-context learning (ICL) capabilities. But can we inject ICL abilities into a pre-trained VLA like pi0? Yes! Introducing RICL (Retraining for In-Context Learning), our Conference on Robot Learning (CoRL) 2025 paper. Our RICL-pi0 model can adapt to unseen objects, novel motions, and new scenes with just ICL and RAG (retrieval-augmented generation). RICL-pi0 also boosts performance on the long-tail of tasks. A quick 1 minute video summary:

Robot AI brains, aka Vision-Language-Action models, cannot adapt to new tasks as easily as LLMs like Gemini, ChatGPT, or Grok. LLMs can adapt quickly with their in-context learning (ICL) capabilities. But can we inject ICL abilities into a pre-trained VLA like pi0? Yes! Introducing RICL (Retraining for In-Context Learning), our Conference on Robot Learning (CoRL) 2025 paper. Our RICL-pi0 model can adapt to unseen objects, novel motions, and new scenes with just ICL and RAG (retrieval-augmented generation). RICL-pi0 also boosts performance on the long-tail of tasks. A quick 1 minute video summary:

Kaustubh Sridhar

52,158 görüntüleme • 1 yıl önce

i spent 3 hours finding the sweet spot for hermes 4.3 36B on a single RTX 3090. saving you the trouble, anon. the model is 21.8GB at Q4_K_M. that leaves 2.2GB free on 24GB VRAM. not much room for KV cache. here's what actually happened: 4K: 35.3 tok/s 8K: 35.2 tok/s 16K: 34.8 tok/s 32K: 34.6 tok/s 64K: 6.4 tok/s 128K: 1.9 tok/s flat from 4K to 32K. then it falls off a cliff at 64K. the trick is quantized KV cache. without it you OOM at 16K. with quantized KV cache you get 32K at full speed. all 64 layers on GPU. at 64K something weird happens. ngl 99 (all layers on GPU) = 3.96 tok/s. the KV cache silently spills to CPU. drop to ngl 55 and speed jumps to 6.37. drop to ngl 48 and it gets worse again (3.46). there's an offload sweet spot where you free just enough VRAM for the cache without losing too much compute to PCIe transfers. 128K works at ngl 32 but you're at 1.95 tok/s. half the model on CPU. usable for batch work, not for interactive. the sweet spot command: llama-server -m hermes-4.3-36b-Q4_K_M.gguf -ngl 99 -c 32768 --cache-type-k q4_0 --cache-type-v q4_0 32K context. 34.6 tok/s. all on GPU. this is where dense 36B lives on 24GB. for comparison, qwen 3.5 (35B MoE, 3B active) holds 112 tok/s from 4K all the way to 262K on the same GPU. no speed drop. same total params, completely different architecture. hybrid linear attention means flat context scaling. dense pays for every token in the KV cache. code and quality comparison coming next. fast vs slow generation side by side in the videos below.

i spent 3 hours finding the sweet spot for hermes 4.3 36B on a single RTX 3090. saving you the trouble, anon. the model is 21.8GB at Q4_K_M. that leaves 2.2GB free on 24GB VRAM. not much room for KV cache. here's what actually happened: 4K: 35.3 tok/s 8K: 35.2 tok/s 16K: 34.8 tok/s 32K: 34.6 tok/s 64K: 6.4 tok/s 128K: 1.9 tok/s flat from 4K to 32K. then it falls off a cliff at 64K. the trick is quantized KV cache. without it you OOM at 16K. with quantized KV cache you get 32K at full speed. all 64 layers on GPU. at 64K something weird happens. ngl 99 (all layers on GPU) = 3.96 tok/s. the KV cache silently spills to CPU. drop to ngl 55 and speed jumps to 6.37. drop to ngl 48 and it gets worse again (3.46). there's an offload sweet spot where you free just enough VRAM for the cache without losing too much compute to PCIe transfers. 128K works at ngl 32 but you're at 1.95 tok/s. half the model on CPU. usable for batch work, not for interactive. the sweet spot command: llama-server -m hermes-4.3-36b-Q4_K_M.gguf -ngl 99 -c 32768 --cache-type-k q4_0 --cache-type-v q4_0 32K context. 34.6 tok/s. all on GPU. this is where dense 36B lives on 24GB. for comparison, qwen 3.5 (35B MoE, 3B active) holds 112 tok/s from 4K all the way to 262K on the same GPU. no speed drop. same total params, completely different architecture. hybrid linear attention means flat context scaling. dense pays for every token in the KV cache. code and quality comparison coming next. fast vs slow generation side by side in the videos below.

Sudo su

52,268 görüntüleme • 6 ay önce

I chatted with yash about MatX, chip design and where silicon designed for LLMs is headed (8:17) Tightly coupling SRAM and HBM on one chip (14:03) More MoE FLOPS, smaller KV cache load (16:08) Numerics: from 32-bit to 4-bit (19:02) Targeting both training and inference (22:14) Chip timelines (27:15) Logic and memory scarcity (29:42) Compute costs (32:07) Latency: from 20ms to 1ms as the new table stakes (40:50) Programming the chip (43:00) Starting MatX (47:11) Codesign without seeing the models (51:57) Interconnect design (55:44) Performance modeling philosophy (1:07:02) Prefill vs. decode (1:13:47) What's next

I chatted with yash about MatX, chip design and where silicon designed for LLMs is headed (8:17) Tightly coupling SRAM and HBM on one chip (14:03) More MoE FLOPS, smaller KV cache load (16:08) Numerics: from 32-bit to 4-bit (19:02) Targeting both training and inference (22:14) Chip timelines (27:15) Logic and memory scarcity (29:42) Compute costs (32:07) Latency: from 20ms to 1ms as the new table stakes (40:50) Programming the chip (43:00) Starting MatX (47:11) Codesign without seeing the models (51:57) Interconnect design (55:44) Performance modeling philosophy (1:07:02) Prefill vs. decode (1:13:47) What's next

Reiner Pope

72,307 görüntüleme • 4 ay önce