Загрузка видео...

Не удалось загрузить видео

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

На главную

Day 11/90 of Inference Engineering How does vLLM work and how is it used in production? Before we discuss how vLLM works internally, it helps to understand what vLLM is. At a high level, vLLM is an inference engine that is designed to serve LLMs to thousands of concurrent... users efficiently while managing scarce compute and memory. The goal for vLLM is to maximize throughput and minimize latency; optimizing for the best inference economics and experience for end users. With every request from the end user, it eventually ends up in the engine core, gets scheduled alongside other requests from other concurrent users, executes on the GPU, and updates the KV cache with the new key and value vectors, and streams the tokens back to the user. The Scheduler decides what requests should execute next while continuously batching requests together to maximize GPU utilization. Continuous batching is an inference optimization that allows new requests to join a running batch as other requests finish generating tokens. This helps with keeping the GPU utilization high instead of letting it sit idle waiting for an entire batch to complete generating. After the scheduler dispatches the selected batch to the Model Executor, the Model Executor prepares the tensors and metadata required for inference, retrieves each request’s block table from KV Cache Manager, launches the optimized transformer forward pass on the GPU, computes the logits, updates the KV cache with the new key and value vectors, and finally returns the results for sampling and streaming. The KV Cache Manager uses the PagedAttention memory layout to allocate fixed-size cache blocks on demand and maintains a Free Block Queue on the CPU that tracks which blocks in the GPU’s Paged KV Cache are currently free. When a request needs additional KV cache space, the KV Cache manager takes a free block from the queue and assigns it to that request, thus avoiding an expensive search through GPU memory for available cache blocks. All of these components form the core of vLLM’s inference engine. The Scheduler determines what requests are executed, the Model Executor determines how those requests are executed, the KV Cache Manager determines where each request’s KV cache lives using the PagedAttention Memory Layout. This architecture enables vLLM to serve thousands of concurrent requests with high throughput, low latency, and efficient GPU memory utilization. Heres a little animation that visualizes everything! - I've also completed the forward pass for my mnist.c project. I had a nice chat with shrey birmiwal, such a knowledgeable guy. Excited to learn more about vLLM and implement a tiny-vLLM one day.show more

max fu

3,687 subscribers

70,797 просмотров • 2 месяцев назад •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

Day 12/90 of Inference Engineering What is chunked prefill within vLLM? In continuation of yesterday's post on the high level architecture of vLLM, I want to dive deeper into vLLM core engine starting with the mechanics of chunked prefill. In this post, I will closely follow the original blog on the anatomy of vLLM. To start, let's define chunked prefill. It's a runtime inference optimization technique that splits a long input request so that it doesn’t monopolize the whole GPU. Keep in mind this is all within the context of vLLM. And since vLLM is an inference engine that's meant to serve a model to multiple concurrent users, having a GPU that’s fully monopolized on a single user's request means other users' requests would be in queue waiting to be processed. It isn’t too good to have the whole GPU occupied on a single request when the GPU is meant to be shared! So the key idea behind chunked prefill is to break the long request into smaller chunks, so that each chunk along with other users' requests gets processed and written into the KV cache together. Suppose we split up the long request into chunks and each chunk has 8 tokens. Now each memory block can hold 4 tokens. Therefore, 8 tokens can fit into 2 blocks of memory. After the first forward pass, 2 blocks are occupied, and after the second forward pass, 4 blocks of memory are occupied and so forth. Each forward pass handles a small chunk of the long request so that there's room in the same pass to keep serving other users' requests. Here's a small animation that I made today to fully visualize the idea behind chunked prefill when learning this topic~

Day 12/90 of Inference Engineering What is chunked prefill within vLLM? In continuation of yesterday's post on the high level architecture of vLLM, I want to dive deeper into vLLM core engine starting with the mechanics of chunked prefill. In this post, I will closely follow the original blog on the anatomy of vLLM. To start, let's define chunked prefill. It's a runtime inference optimization technique that splits a long input request so that it doesn’t monopolize the whole GPU. Keep in mind this is all within the context of vLLM. And since vLLM is an inference engine that's meant to serve a model to multiple concurrent users, having a GPU that’s fully monopolized on a single user's request means other users' requests would be in queue waiting to be processed. It isn’t too good to have the whole GPU occupied on a single request when the GPU is meant to be shared! So the key idea behind chunked prefill is to break the long request into smaller chunks, so that each chunk along with other users' requests gets processed and written into the KV cache together. Suppose we split up the long request into chunks and each chunk has 8 tokens. Now each memory block can hold 4 tokens. Therefore, 8 tokens can fit into 2 blocks of memory. After the first forward pass, 2 blocks are occupied, and after the second forward pass, 4 blocks of memory are occupied and so forth. Each forward pass handles a small chunk of the long request so that there's room in the same pass to keep serving other users' requests. Here's a small animation that I made today to fully visualize the idea behind chunked prefill when learning this topic~

max fu

29,585 просмотров • 2 месяцев назад

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

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

Vikas gupta

14,138 просмотров • 16 дней назад

4/ to achieve maximum memory efficiency, we quantize model weights to ~4-bit, getting the language model under 20GB with room for the kv cache, perception encoder, and drafter alongside it. a dflash drafter proposes blocks of tokens the main model verifies in parallel, so it stays responsive.

4/ to achieve maximum memory efficiency, we quantize model weights to ~4-bit, getting the language model under 20GB with room for the kv cache, perception encoder, and drafter alongside it. a dflash drafter proposes blocks of tokens the main model verifies in parallel, so it stays responsive.

Alexandr Wang

69,242 просмотров • 1 месяц назад

🎥 Video generation is hitting the memory wall. As videos get longer, the KV cache quietly explodes — and long-horizon consistency starts to break. We built Quant VideoGen: a training-free KV cache compression method for auto-regressive video diffusion. Instead of storing every KV in high precision, QVG exploits video’s spatiotemporal redundancy with semantic-aware smoothing + progressive residual quantization. 🚀 Up to 7× KV memory reduction ⚡ <4% overhead ✅ Strong long-video quality 🕹️ Deploy HYWorldPlay on your own RTX 5090 locally KV compression is becoming a core scaling primitive — not just for LLMs, but for video generation too. Paper: Code: (1/5)

🎥 Video generation is hitting the memory wall. As videos get longer, the KV cache quietly explodes — and long-horizon consistency starts to break. We built Quant VideoGen: a training-free KV cache compression method for auto-regressive video diffusion. Instead of storing every KV in high precision, QVG exploits video’s spatiotemporal redundancy with semantic-aware smoothing + progressive residual quantization. 🚀 Up to 7× KV memory reduction ⚡ <4% overhead ✅ Strong long-video quality 🕹️ Deploy HYWorldPlay on your own RTX 5090 locally KV compression is becoming a core scaling primitive — not just for LLMs, but for video generation too. Paper: Code: (1/5)

Haocheng Xi

65,767 просмотров • 5 месяцев назад

A good technical LLM interview question: Your LLM chatbot takes 12s before it generates the first token, and the users are complaining. So you move the model onto a GPU with 3x the computing power. The time to first token barely improves. Why did this happen? (answer below) Latency in an LLM app is a placement problem disguised as a model problem. If you profile the 12 seconds, the model's prefill itself may only account for around 1.5 seconds of it. So halving the prefill step saves just 750ms out of 12000, which is under 7%. The rest is spread across stages that never touch the GPU. The request first travels to whatever region the app runs in, and a cross-continent round trip could cost over a second before any code executes. Then the request handler starts. On a container-based serverless platform under load, this adds several seconds of cold start, paid before auth, rate limiting, or prompt assembly even begins. Retrieval adds its own hop, and the response streams back across the same distance. Optimizing a stage that was already fast cannot alter the latency that's majorly affected by other stages. Those other stages are slow for a structural reason. An LLM app runs two workloads that want opposite machines. - The request path is short, spiky, and needs to sit close to users - Inference is long-running, GPU-bound, and billed hourly, whether requests arrive or not. So the actual decision is not which model to run, but where each of these two workloads runs. There are three options, each with its own tradeoffs: > A dedicated GPU box removes inference cold starts, but it bills around the clock and lives in one location, so distant users wait out the round trip on every request > Container-based serverless scales to zero, but the request path pays a cold start, and most of these platforms have no GPU behind them. > Edge runtimes start in under a millisecond, because a WebAssembly module carries no OS or container image to boot. They handle the request path well and cannot hold a model. So the answer is not to pick one, but to split the app across two of them. The request path runs close to users, and inference runs on a dedicated GPU it calls into. That also explains the failed upgrade. More compute made a stage that was already fast faster, and left the 10.5 seconds around it untouched. To actually learn how it's done in practice, Akamai's GitHub has a reference implementation for each half. - vllm-on-lke serves Qwen2.5-7B-Instruct behind an OpenAI-compatible endpoint on one RTX 4000 Ada GPU in Linode Kubernetes Engine, with Terraform creating the cluster, both firewalls, and the GPU operator in one apply. - akamai-functions-llm-chatbot covers the front, where a WebAssembly API checks a KV cache and only calls the GPU-backed instance on a miss. Both are available on Akamai’s new Developer Hub, alongside their tutorials and code samples. It also links to Edge Case, their Discord, where four developer advocates architect and deploy a production app live every other Wednesday. If you create a new Akamai Cloud account, you can also get $300 in credits for joining. Join here: That said, this post treats generation as a single 1.5s block, but that block has its own structure, and knowing it well tells you whether a model is slow to start or slow to stream. I wrote a first-principles walkthrough of it, covering the prefill and decode split, KV caching, and where the time actually goes inside each one. Read it below. Thanks to Akamai Cloud for partnering today!

Avi Chawla

21,786 просмотров • 1 месяц назад

Gemma 4 26B A4B MoE - 500+ t/s decode - Single RTX 4090 (24 GB VRAM) - Llama.cpp concurrency 24 - q8 kv cache How many API users can you simultaneously host on a single RTX 4090 (24 GB VRAM) before it crashes? Yesterday, I proved you can host 14 active users using unquantized memory. Today, I used 8 bit KV Cache Quantization to hack the VRAM footprint. I successfully scaled to 24 concurrent users without a single dropped connection. A 71% server capacity boost for free. By adding the -ctk q8_0 -ctv q8_0 flags to llama.cpp, you compress the KV cache context memory from 16 bit to 8 bit. This unlocks massive concurrency limits on Gemma 4 26B (MoE) on a single 24GB consumer GPU. Here is the exact telemetry from pushing 8 bit quantization to its absolute physical edge: # TEST 1: The 24 User Concurrency Max Server Config: 24 slots (np 24) | 4,096 context per slot | 98,304 Total Context Client Load: 24 simultaneous requests (2,000 token prompt per user) Unquantized KV cache for this load requires 28GB+ VRAM (Instant OOM). Quantized to Q8, it allocated safely at 23.35 GB. The C++ engine crunched the entire batch in 28.5 seconds. Decode Speed: 21 t/s (Per User) | 500 t/s (Agg) # TEST 2: The 48 User Queue Overload What happens to a compressed cache during a traffic spike? Server Config: 24 slots (np 24) | 4,096 context per slot | 98,304 Total Context Client Load: 48 simultaneous requests (2k token prompt per user) Zero queue drops. The scheduler flushed and hot swapped the 8 bit memory flawlessly on the fly, completing all 48 users in 66.0 seconds (a perfect 2.3x queue scaling multiplier). Decode Speed: 18 t/s (Per User) | 430 t/s (Agg) # TEST 3: The 8 User RAG Slam Server Config: 8 slots (np 8) | 60,000 context per slot | 480,000 Total Context Client Load: 8 simultaneous requests (30k token prompt per user) It allocated 23.83 GB VRAM and chewed through ~240,000 prefill tokens in 46 seconds under massive memory pressure. Prefill Speed: 6,200 t/s (Agg) Decode Speed: 22 t/s (Per User) | 175 t/s (Agg) # The Engineering Alpha (The Quantization Tradeoff): You gain a massive 71% increase in server capacity, but what do you lose? Compute latency. Because the cache is stored in 8 bit, the GPU's cores have to dequantize the memory back to 16 bit on the fly during every single prefill step. In my unquantized tests yesterday, single slot prefill was hitting ~1,500+ t/s. Today, under the heavy 48-user Q8 load, prefill dropped as low as ~750 t/s. You trade a few seconds of initial prefill latency to essentially double your API hosting capacity. For production high volume SaaS, this is the ultimate unit economics cheat code. Here is the exact command to run a 24 user Q8 continuous batching server on your own single 4090, single 3090 or any 24gb vram rig: ./build/bin/llama-server -m gemma-4-26B-A4B-it.gguf -c 98304 -np 24 -b 2048 -ub 2048 -ngl 99 -fa on -ctk q8_0 -ctv q8_0 --port 8080 (Note: -c 98304 allocates exactly 4,096 tokens of context per user across 24 slots). Hugging Face links to the Unsloth Gemma 4 26B QAT quants along with performance graphs available in the replies. Would you trade 3 seconds of Time To First Token latency to double your active user capacity?

Gemma 4 26B A4B MoE - 500+ t/s decode - Single RTX 4090 (24 GB VRAM) - Llama.cpp concurrency 24 - q8 kv cache How many API users can you simultaneously host on a single RTX 4090 (24 GB VRAM) before it crashes? Yesterday, I proved you can host 14 active users using unquantized memory. Today, I used 8 bit KV Cache Quantization to hack the VRAM footprint. I successfully scaled to 24 concurrent users without a single dropped connection. A 71% server capacity boost for free. By adding the -ctk q8_0 -ctv q8_0 flags to llama.cpp, you compress the KV cache context memory from 16 bit to 8 bit. This unlocks massive concurrency limits on Gemma 4 26B (MoE) on a single 24GB consumer GPU. Here is the exact telemetry from pushing 8 bit quantization to its absolute physical edge: # TEST 1: The 24 User Concurrency Max Server Config: 24 slots (np 24) | 4,096 context per slot | 98,304 Total Context Client Load: 24 simultaneous requests (2,000 token prompt per user) Unquantized KV cache for this load requires 28GB+ VRAM (Instant OOM). Quantized to Q8, it allocated safely at 23.35 GB. The C++ engine crunched the entire batch in 28.5 seconds. Decode Speed: 21 t/s (Per User) | 500 t/s (Agg) # TEST 2: The 48 User Queue Overload What happens to a compressed cache during a traffic spike? Server Config: 24 slots (np 24) | 4,096 context per slot | 98,304 Total Context Client Load: 48 simultaneous requests (2k token prompt per user) Zero queue drops. The scheduler flushed and hot swapped the 8 bit memory flawlessly on the fly, completing all 48 users in 66.0 seconds (a perfect 2.3x queue scaling multiplier). Decode Speed: 18 t/s (Per User) | 430 t/s (Agg) # TEST 3: The 8 User RAG Slam Server Config: 8 slots (np 8) | 60,000 context per slot | 480,000 Total Context Client Load: 8 simultaneous requests (30k token prompt per user) It allocated 23.83 GB VRAM and chewed through ~240,000 prefill tokens in 46 seconds under massive memory pressure. Prefill Speed: 6,200 t/s (Agg) Decode Speed: 22 t/s (Per User) | 175 t/s (Agg) # The Engineering Alpha (The Quantization Tradeoff): You gain a massive 71% increase in server capacity, but what do you lose? Compute latency. Because the cache is stored in 8 bit, the GPU's cores have to dequantize the memory back to 16 bit on the fly during every single prefill step. In my unquantized tests yesterday, single slot prefill was hitting ~1,500+ t/s. Today, under the heavy 48-user Q8 load, prefill dropped as low as ~750 t/s. You trade a few seconds of initial prefill latency to essentially double your API hosting capacity. For production high volume SaaS, this is the ultimate unit economics cheat code. Here is the exact command to run a 24 user Q8 continuous batching server on your own single 4090, single 3090 or any 24gb vram rig: ./build/bin/llama-server -m gemma-4-26B-A4B-it.gguf -c 98304 -np 24 -b 2048 -ub 2048 -ngl 99 -fa on -ctk q8_0 -ctv q8_0 --port 8080 (Note: -c 98304 allocates exactly 4,096 tokens of context per user across 24 slots). Hugging Face links to the Unsloth Gemma 4 26B QAT quants along with performance graphs available in the replies. Would you trade 3 seconds of Time To First Token latency to double your active user capacity?

Alok

17,465 просмотров • 1 месяц назад

Every time an employee asks AI to summarize months of emails, analyze a 200-page contract, or answer questions across thousands of project documents, they expect one thing: memory. The AI isn't just responding to the latest prompt. It's recalling context, tracking details, and connecting information across massive amounts of data. Behind that experience is the KV cache, one of the fastest-growing challenges in AI infrastructure. See how memory and storage are enabling longer context, smarter agents, and more powerful AI experiences: #IntelligenceAccelerated

Every time an employee asks AI to summarize months of emails, analyze a 200-page contract, or answer questions across thousands of project documents, they expect one thing: memory. The AI isn't just responding to the latest prompt. It's recalling context, tracking details, and connecting information across massive amounts of data. Behind that experience is the KV cache, one of the fastest-growing challenges in AI infrastructure. See how memory and storage are enabling longer context, smarter agents, and more powerful AI experiences: #IntelligenceAccelerated

Micron Technology

18,344 просмотров • 18 дней назад

Robots can now reconstruct 3D scenes in real time from a single RGB camera. [📍 Projects page + paper] No depth sensor. No retraining. 30 FPS. Researchers at the Imperial College London introduced KV-Tracker, a training-free method that makes heavy models like π³ and Depth Anything 3 fast enough for real-time tracking. The idea is simple. These models use global self-attention, which is powerful but computationally expensive. KV-Tracker caches the key and value pairs from selected keyframes and reuses them for new frames. That cache becomes an implicit scene representation. Result: • Up to 30 FPS • 10 to 15x speedup • Accurate 6-DoF tracking on benchmarks like TUM RGB-D and 7-Scenes • Works with monocular RGB only It also supports object-level tracking with masks and allows saving the KV-cache for later reuse. For robotics, this reduces hardware constraints and moves real-time 3D perception closer to practical deployment. Credit to Marwan Taher (Marwan Taher) at Imperial’s Dyson Robotics Lab and many others who contributed to this! 📍 Save projects page + paper for later: Video: ——- if it matters in AI or Robotics you'll read it here first:

Robots can now reconstruct 3D scenes in real time from a single RGB camera. [📍 Projects page + paper] No depth sensor. No retraining. 30 FPS. Researchers at the Imperial College London introduced KV-Tracker, a training-free method that makes heavy models like π³ and Depth Anything 3 fast enough for real-time tracking. The idea is simple. These models use global self-attention, which is powerful but computationally expensive. KV-Tracker caches the key and value pairs from selected keyframes and reuses them for new frames. That cache becomes an implicit scene representation. Result: • Up to 30 FPS • 10 to 15x speedup • Accurate 6-DoF tracking on benchmarks like TUM RGB-D and 7-Scenes • Works with monocular RGB only It also supports object-level tracking with masks and allows saving the KV-cache for later reuse. For robotics, this reduces hardware constraints and moves real-time 3D perception closer to practical deployment. Credit to Marwan Taher (Marwan Taher) at Imperial’s Dyson Robotics Lab and many others who contributed to this! 📍 Save projects page + paper for later: Video: ——- if it matters in AI or Robotics you'll read it here first:

Ilir Aliu

53,992 просмотров • 5 месяцев назад

I tested MTPLX v2 with QWEN 3.6 27B and compared it with oMLX without cache on M5 Max and DGX Spark on vllm using nvfp4 model version. More details in 🧵 I've reached 82.8 tps of max decoding speed! 🔥 Custom Metal Kernel design specifically for this model and for Apple Silicon is just perfect! This is the way forward! Great job Youssof Al Toukhi Look at the website here! 👇 Here a website with recap, built with GLM 5.2 running locally 💪 First chart and preview from the website.

I tested MTPLX v2 with QWEN 3.6 27B and compared it with oMLX without cache on M5 Max and DGX Spark on vllm using nvfp4 model version. More details in 🧵 I've reached 82.8 tps of max decoding speed! 🔥 Custom Metal Kernel design specifically for this model and for Apple Silicon is just perfect! This is the way forward! Great job Youssof Al Toukhi Look at the website here! 👇 Here a website with recap, built with GLM 5.2 running locally 💪 First chart and preview from the website.

Ivan Fioravanti ᯅ

15,885 просмотров • 2 месяцев назад

Stanford researchers did it again. They just built the agent-native version of Git. When an agent works on a longer task, the run builds up a lot of state. This includes files edited/created, a dev server, a database, installed packages, KV cache, etc. Say the agent is at step 10 and makes a mistake, maybe it misreads a traceback and rewrites a file that was actually fine. The tests start failing, and the run goes off track, although everything through step eight was correct. By default, the agent just tries to fix it, which creates more edits and tool calls. This burns more tokens and grows the context. The other options are a person stepping in to redirect it or restarting the whole run from step one. That's wasteful, because it pays for every model/tool call again and re-prefills the context. Moreover, since an agent's run is non-deterministic, it doesn't reproduce the same early steps anyway. The reason it's hard to just jump back exactly to a previous correct step and resume from there is that the trajectory is only a message log. It records what the agent said and which tools it called, but not the live state underneath. That state includes things like memory, open file handles, child processes, installed packages, /tmp, and KV cache. None of that is in the log. Git can version the files, but it doesn't snapshot the running process or the KV cache. Checking out step eight moves the files back, but the process is still sitting in step-ten memory with a cold cache. Shepherd is a runtime layer by Stanford that records the run as a trace of typed events rather than a flat log. Each agent-environment interaction becomes a commit, similar to Git, but it tracks the live run. Its commit includes the agent process and the filesystem together, copy-on-write, so a branch carries the actual state and not just the files. Going back to a previous step is then a single call that forks from that commit and continues from the exact state. The copy-on-write fork is roughly five times faster than docker commit, and because the prompt prefix through step eight is unchanged, the KV cache is reused over 95% on replay, so early steps aren't reprocessed again. Once the run can be forked, a meta-agent can sit on top and operate it. It watches the trace and reverts as soon as it looks wrong, before the bad write is committed. In practice, it's just Python calling fork, replay, and revert on the trace, rather than a separate control plane wired into the harness. Not everything is reversible though. Files and sandbox changes undo themselves, but a database write has no automatic undo, so it needs a matching undo step set up in advance. Something external, like a sent email or a real charge, can't be undone, so the supervisor's job there is to catch it before it fires. They tested this on a few public benchmarks. On CooperBench, where two agents work on the same codebase, adding a live supervisor took the pair-coding pass rate from 28.8% to 54.7%. It's still early and labeled alpha. The benefit mostly shows up when a run gets branched a lot over a heavy sandbox state, which is exactly where restarting wastes the most tokens and time. If Git was made to make file changes reversible, Shepherd is trying to do the same thing for a live agent run. Shepherd Repo: (don't forget to star it ⭐ ) That said, Shepherd reverts a bad step inside a run. The harness around it, the prompts, tools, and checks the supervisor relies on, still drifts across runs as models and dependencies change. Akshay wrote about making that harness repair itself, where a failing trace gets diagnosed, the fix is verified against the exact input that failed, and the failure is locked as a regression test so it can't recur. Read it below.

Stanford researchers did it again. They just built the agent-native version of Git. When an agent works on a longer task, the run builds up a lot of state. This includes files edited/created, a dev server, a database, installed packages, KV cache, etc. Say the agent is at step 10 and makes a mistake, maybe it misreads a traceback and rewrites a file that was actually fine. The tests start failing, and the run goes off track, although everything through step eight was correct. By default, the agent just tries to fix it, which creates more edits and tool calls. This burns more tokens and grows the context. The other options are a person stepping in to redirect it or restarting the whole run from step one. That's wasteful, because it pays for every model/tool call again and re-prefills the context. Moreover, since an agent's run is non-deterministic, it doesn't reproduce the same early steps anyway. The reason it's hard to just jump back exactly to a previous correct step and resume from there is that the trajectory is only a message log. It records what the agent said and which tools it called, but not the live state underneath. That state includes things like memory, open file handles, child processes, installed packages, /tmp, and KV cache. None of that is in the log. Git can version the files, but it doesn't snapshot the running process or the KV cache. Checking out step eight moves the files back, but the process is still sitting in step-ten memory with a cold cache. Shepherd is a runtime layer by Stanford that records the run as a trace of typed events rather than a flat log. Each agent-environment interaction becomes a commit, similar to Git, but it tracks the live run. Its commit includes the agent process and the filesystem together, copy-on-write, so a branch carries the actual state and not just the files. Going back to a previous step is then a single call that forks from that commit and continues from the exact state. The copy-on-write fork is roughly five times faster than docker commit, and because the prompt prefix through step eight is unchanged, the KV cache is reused over 95% on replay, so early steps aren't reprocessed again. Once the run can be forked, a meta-agent can sit on top and operate it. It watches the trace and reverts as soon as it looks wrong, before the bad write is committed. In practice, it's just Python calling fork, replay, and revert on the trace, rather than a separate control plane wired into the harness. Not everything is reversible though. Files and sandbox changes undo themselves, but a database write has no automatic undo, so it needs a matching undo step set up in advance. Something external, like a sent email or a real charge, can't be undone, so the supervisor's job there is to catch it before it fires. They tested this on a few public benchmarks. On CooperBench, where two agents work on the same codebase, adding a live supervisor took the pair-coding pass rate from 28.8% to 54.7%. It's still early and labeled alpha. The benefit mostly shows up when a run gets branched a lot over a heavy sandbox state, which is exactly where restarting wastes the most tokens and time. If Git was made to make file changes reversible, Shepherd is trying to do the same thing for a live agent run. Shepherd Repo: (don't forget to star it ⭐ ) That said, Shepherd reverts a bad step inside a run. The harness around it, the prompts, tools, and checks the supervisor relies on, still drifts across runs as models and dependencies change. Akshay wrote about making that harness repair itself, where a failing trace gets diagnosed, the fix is verified against the exact input that failed, and the failure is locked as a regression test so it can't recur. Read it below.

Avi Chawla

441,974 просмотров • 2 месяцев назад

$Transformer and Mixture of Experts, explained visually! Mixture of Experts (MoE) is a popular architecture that uses different experts to improve Transformer models. Transformer and MoE differ in the decoder block: - Transformer uses a feed-forward network. - MoE uses experts, which are feed-forward networks but smaller compared to those Transformer. During inference, a subset of experts are selected. This makes inference faster in MoE. Also, since the network has multiple decoder layers: - The text passes through different experts across layers. - The chosen experts also differ between tokens. But how does the model decide which experts should be ideal? The router does that. It is a multi-class classifier that produces softmax scores over experts to select the top K experts. The router is trained with the network, and it learns to select the best experts. But it isn't straightforward. There are challenges! Challenge 1) Notice this pattern at the start of training: - Say, the model selects "Expert 2" - This expert gets a bit better - It may get selected again since it's the "best" - It learns more - It gets selected again in the next iteration - It learns more, and so on! This means many experts can go under-trained due to the overselection of a few experts! We solve this in two steps: - Add noise to the feed-forward output of the router so that other experts can get higher logits. - Set all but the top K logits to -infinity. After softmax, these scores become zero. This way, other experts also get the opportunity to train. Challenge 2) Some experts may get exposed to more tokens than others, leading to under-trained experts. We prevent this by limiting the number of tokens an expert can process. If an expert reaches the limit, the token is passed to the next best expert. Overall, MoEs have more parameters to load. But a fraction of them are activated during inference. This leads to faster inference. Mixtral 8x7B and Llama 4 are two popular MoE-based LLMs. Have you used MoEs in production yet? To dive deeper into how MoE inference works in production, we wrote a full article covering token dispatch, grouped expert computation, model-weight memory, multi-GPU communication, expert placement, load imbalance, and performance diagnosis. Read it below.$

Transformer and Mixture of Experts, explained visually! Mixture of Experts (MoE) is a popular architecture that uses different experts to improve Transformer models. Transformer and MoE differ in the decoder block: - Transformer uses a feed-forward network. - MoE uses experts, which are feed-forward networks but smaller compared to those Transformer. During inference, a subset of experts are selected. This makes inference faster in MoE. Also, since the network has multiple decoder layers: - The text passes through different experts across layers. - The chosen experts also differ between tokens. But how does the model decide which experts should be ideal? The router does that. It is a multi-class classifier that produces softmax scores over experts to select the top K experts. The router is trained with the network, and it learns to select the best experts. But it isn't straightforward. There are challenges! Challenge 1) Notice this pattern at the start of training: - Say, the model selects "Expert 2" - This expert gets a bit better - It may get selected again since it's the "best" - It learns more - It gets selected again in the next iteration - It learns more, and so on! This means many experts can go under-trained due to the overselection of a few experts! We solve this in two steps: - Add noise to the feed-forward output of the router so that other experts can get higher logits. - Set all but the top K logits to -infinity. After softmax, these scores become zero. This way, other experts also get the opportunity to train. Challenge 2) Some experts may get exposed to more tokens than others, leading to under-trained experts. We prevent this by limiting the number of tokens an expert can process. If an expert reaches the limit, the token is passed to the next best expert. Overall, MoEs have more parameters to load. But a fraction of them are activated during inference. This leads to faster inference. Mixtral 8x7B and Llama 4 are two popular MoE-based LLMs. Have you used MoEs in production yet? To dive deeper into how MoE inference works in production, we wrote a full article covering token dispatch, grouped expert computation, model-weight memory, multi-GPU communication, expert placement, load imbalance, and performance diagnosis. Read it below.

Daily Dose of Data Science

54,872 просмотров • 9 дней назад

The "I don't have enough VRAM" excuse just died. I’m running Meta’s new 30B Muse Glimmer Q6_K_XL with a massive 130k context window on just 26GB VRAM FREE compute on Kaggle. Kaggle provides you free 2x Nvidia T4 GPUs. 30 hours usage each week! Yesterday, I showed you the violent throughput of Muse Glimmer on a single RTX 4090. Today, we are securing a Dual NVIDIA T4 GPU cluster with 32GB of total VRAM for exactly $0 and dropping the massive 24.5GB Q6_K_XL GGUF onto it. Here is the exact Kaggle workflow and benchmarking breakdown: # 1. The Storage Bypass & Setup I built a clean cell by cell script in the file. We dynamically fetch the CUDA accelerated llama.cpp binaries and use wget to stream the model directly into Kaggle's /kaggle/tmp scratch storage, which cleanly bypasses their 19.5GB output directory limit. # 2. The Multi GPU Performance With the -ngl 99 flag offloading all model layers across both T4 GPUs (32GB VRAM combined), we pushed a massive 131,072 token context window (-c 131072). The benchmark numbers: Prefill: 265.9 t/s Decode: 9.0 t/s VRAM Total: 26.5 GB # 3. The Architecture Insight The Q6_K_XL model itself is 24.5 GB. Because of Muse Glimmer's aggressive 16:1 GQA, the unquantized KV cache for a massive 130k context window only takes up 2 GB of memory. No heavily degraded Q4 KV quantization required. It just works. No compiling from source. No credit card. No OOM crashes. Zero excuses. If you’re running a single RTX 3090, 4090, or 5090, you need to experience this hyper efficient KV cache right now before the upcoming Qwen 3.8 27B drop completely steals your VRAM tomorrow. pick the Q4 or Q5 quants for 24 GB VRAM rigs. I'm dropping the Unsloth huggingface GGUF links and the free Kaggle notebook link in the replies. spin up your own instance, and show me your multi GPU benchmarks.

The "I don't have enough VRAM" excuse just died. I’m running Meta’s new 30B Muse Glimmer Q6_K_XL with a massive 130k context window on just 26GB VRAM FREE compute on Kaggle. Kaggle provides you free 2x Nvidia T4 GPUs. 30 hours usage each week! Yesterday, I showed you the violent throughput of Muse Glimmer on a single RTX 4090. Today, we are securing a Dual NVIDIA T4 GPU cluster with 32GB of total VRAM for exactly $0 and dropping the massive 24.5GB Q6_K_XL GGUF onto it. Here is the exact Kaggle workflow and benchmarking breakdown: # 1. The Storage Bypass & Setup I built a clean cell by cell script in the file. We dynamically fetch the CUDA accelerated llama.cpp binaries and use wget to stream the model directly into Kaggle's /kaggle/tmp scratch storage, which cleanly bypasses their 19.5GB output directory limit. # 2. The Multi GPU Performance With the -ngl 99 flag offloading all model layers across both T4 GPUs (32GB VRAM combined), we pushed a massive 131,072 token context window (-c 131072). The benchmark numbers: Prefill: 265.9 t/s Decode: 9.0 t/s VRAM Total: 26.5 GB # 3. The Architecture Insight The Q6_K_XL model itself is 24.5 GB. Because of Muse Glimmer's aggressive 16:1 GQA, the unquantized KV cache for a massive 130k context window only takes up 2 GB of memory. No heavily degraded Q4 KV quantization required. It just works. No compiling from source. No credit card. No OOM crashes. Zero excuses. If you’re running a single RTX 3090, 4090, or 5090, you need to experience this hyper efficient KV cache right now before the upcoming Qwen 3.8 27B drop completely steals your VRAM tomorrow. pick the Q4 or Q5 quants for 24 GB VRAM rigs. I'm dropping the Unsloth huggingface GGUF links and the free Kaggle notebook link in the replies. spin up your own instance, and show me your multi GPU benchmarks.

Alok

19,370 просмотров • 1 месяц назад

Muse Glimmer, A 30B parameter dense model swallowing a 130,000 token context window using only 19.3 GB of VRAM (extreme efficiency). No KV cache quantization required. I just benched the new Muse Glimmer 30B (dense) on a single RTX 4090. We are pulling 3,100+ t/s prefill and 75 tokens/second decode. The throughput is violent. Meta superintelligence lab just open sourced this agentic beast, explicitly engineered to dominate 24GB consumer cards. I pulled the latest llama.cpp source on Ubuntu 22 (CUDA 13) to see if the specs were real. Fed it a 28k token prompt. Here is the exact llama.cpp God Stack and benchmarking breakdown: # 1. The Deep Context Run (No Speculative Decoding) The architecture uses a massive 16:1 GQA (Grouped Query Attention) ratio. This means the KV cache footprint is practically non existent. ./build/bin/llama-server -m Muse-Glimmer-30B-UD-Q4_K_XL.gguf -c 130000 -b 4096 -ub 4096 -ngl 99 --port 8080 Prefill: 3134.95 t/s Decode: 50.00 t/s VRAM: 19.34 GB (I hit 130k context on pristine, unquantized f16 cache and still had 4.5 GB of VRAM left over. Absolute witchcraft). # 2. The DFlash Speculative Overdrive Meta shipped this with a DFlash block diffusion drafter. Let's trade that extra VRAM for pure speed. ./build/bin/llama-server -m Muse-Glimmer-30B-UD-Q4_K_XL.gguf -md dflash-kquant.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 80000 -b 4096 -ub 4096 -ngl 99 --port 8080 Prefill: 1293.69 t/s Decode: 75.00 t/s VRAM: 23.93 GB (Maxed out on card) the dflash gguf is additional 1.6 GBs # The Architecture Insight (Muse Glimmer vs. Gemma 4 31B) If you look at my Gemma 4 31B tests from last week, getting 140k context required heavily degrading the memory with Q4 KV quantization (gemma 31b q4 can do only about 40k context with unquantized kv on a 24gb card). That "unzipping" overhead bottlenecked Gemma's MTP decode speeds down to 65 t/s. Muse Glimmer completely sidesteps this bottleneck. By using aggressive 16:1 GQA, it keeps the KV cache in native f16 format at massive context lengths. Flash Attention gets to run at maximum uncompressed speed, letting the DFlash drafter push decode safely to 75 t/s without compute lag. With a 76% on SWE Bench Verified and seamless local tool calling, this model looks promising. Unsloth's Hugging Face GGUF links, intelligence/agentic benchmark details, and inference throughput performance graphs are posted in the replies. For 24GB rig, what’s your current go to model?

Muse Glimmer, A 30B parameter dense model swallowing a 130,000 token context window using only 19.3 GB of VRAM (extreme efficiency). No KV cache quantization required. I just benched the new Muse Glimmer 30B (dense) on a single RTX 4090. We are pulling 3,100+ t/s prefill and 75 tokens/second decode. The throughput is violent. Meta superintelligence lab just open sourced this agentic beast, explicitly engineered to dominate 24GB consumer cards. I pulled the latest llama.cpp source on Ubuntu 22 (CUDA 13) to see if the specs were real. Fed it a 28k token prompt. Here is the exact llama.cpp God Stack and benchmarking breakdown: # 1. The Deep Context Run (No Speculative Decoding) The architecture uses a massive 16:1 GQA (Grouped Query Attention) ratio. This means the KV cache footprint is practically non existent. ./build/bin/llama-server -m Muse-Glimmer-30B-UD-Q4_K_XL.gguf -c 130000 -b 4096 -ub 4096 -ngl 99 --port 8080 Prefill: 3134.95 t/s Decode: 50.00 t/s VRAM: 19.34 GB (I hit 130k context on pristine, unquantized f16 cache and still had 4.5 GB of VRAM left over. Absolute witchcraft). # 2. The DFlash Speculative Overdrive Meta shipped this with a DFlash block diffusion drafter. Let's trade that extra VRAM for pure speed. ./build/bin/llama-server -m Muse-Glimmer-30B-UD-Q4_K_XL.gguf -md dflash-kquant.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 80000 -b 4096 -ub 4096 -ngl 99 --port 8080 Prefill: 1293.69 t/s Decode: 75.00 t/s VRAM: 23.93 GB (Maxed out on card) the dflash gguf is additional 1.6 GBs # The Architecture Insight (Muse Glimmer vs. Gemma 4 31B) If you look at my Gemma 4 31B tests from last week, getting 140k context required heavily degrading the memory with Q4 KV quantization (gemma 31b q4 can do only about 40k context with unquantized kv on a 24gb card). That "unzipping" overhead bottlenecked Gemma's MTP decode speeds down to 65 t/s. Muse Glimmer completely sidesteps this bottleneck. By using aggressive 16:1 GQA, it keeps the KV cache in native f16 format at massive context lengths. Flash Attention gets to run at maximum uncompressed speed, letting the DFlash drafter push decode safely to 75 t/s without compute lag. With a 76% on SWE Bench Verified and seamless local tool calling, this model looks promising. Unsloth's Hugging Face GGUF links, intelligence/agentic benchmark details, and inference throughput performance graphs are posted in the replies. For 24GB rig, what’s your current go to model?

Alok

65,480 просмотров • 1 месяц назад

Researchers made KMeans 200x faster. And the new technique also beats approaches like cuML and FAISS. Flash-KMeans is an IO-aware implementation of exact KMeans that redesigns the algorithm around modern GPU bottlenecks. By attacking the memory bottlenecks directly, Flash-KMeans achieves: - 33x speedup over cuML - 200x speedup over FAISS This speedup comes from how it moves through GPU memory. Standard KMeans runs in two steps, and both are bottlenecked by reads and writes to GPU memory: 1) The first step matches every point to its nearest centroid. Standard KMeans computes the full point-to-centroid distance matrix, writes it out to GPU memory, then reads it back to find each nearest centroid. That write-then-read round trip is the bottleneck. Flash-KMeans combines the distance calculation with the nearest-centroid step, so the result is computed on-chip and the full matrix is never written out. 2) The second step recomputes each centroid by averaging the points assigned to it. Standard KMeans has thousands of threads writing into the same centroid slots at once, so they stall waiting for their turn. Flash-KMeans sorts points by cluster first, turning scattered writes into sequential reductions that read and write memory in one efficient pass. Using these two optimizations at the million-scale, Flash-KMeans completes a standard KMeans iteration in a few milliseconds. The video below depicts this in action. Several reasons why this is important: KMeans has always been an offline primitive. Something you run once to preprocess data and move on. These speedups make the approach viable in several runtime-critical systems. ↳ Vector indices like FAISS use KMeans to build search indices. Faster KMeans means you can re-index dynamically as data changes. ↳ LLM quantization methods need KMeans to find optimal weight codebooks, per layer, repeatedly. What takes hours could now take minutes. ↳ MoE models need fast token routing at inference time. Flash-KMeans makes it viable to run this inside the inference loop, not just in preprocessing. I have shared the paper in the replies. That said, memory is the real constraint Flash-KMeans solves, and the problem is not just limited to clustering. The vectors a RAG system stores after indexing create similar bottlenecks. I wrote a detailed walkthrough recently on cutting this vector memory by 32x with binary quantization, querying 36M+ vectors in a few milliseconds. Read it below.

Researchers made KMeans 200x faster. And the new technique also beats approaches like cuML and FAISS. Flash-KMeans is an IO-aware implementation of exact KMeans that redesigns the algorithm around modern GPU bottlenecks. By attacking the memory bottlenecks directly, Flash-KMeans achieves: - 33x speedup over cuML - 200x speedup over FAISS This speedup comes from how it moves through GPU memory. Standard KMeans runs in two steps, and both are bottlenecked by reads and writes to GPU memory: 1) The first step matches every point to its nearest centroid. Standard KMeans computes the full point-to-centroid distance matrix, writes it out to GPU memory, then reads it back to find each nearest centroid. That write-then-read round trip is the bottleneck. Flash-KMeans combines the distance calculation with the nearest-centroid step, so the result is computed on-chip and the full matrix is never written out. 2) The second step recomputes each centroid by averaging the points assigned to it. Standard KMeans has thousands of threads writing into the same centroid slots at once, so they stall waiting for their turn. Flash-KMeans sorts points by cluster first, turning scattered writes into sequential reductions that read and write memory in one efficient pass. Using these two optimizations at the million-scale, Flash-KMeans completes a standard KMeans iteration in a few milliseconds. The video below depicts this in action. Several reasons why this is important: KMeans has always been an offline primitive. Something you run once to preprocess data and move on. These speedups make the approach viable in several runtime-critical systems. ↳ Vector indices like FAISS use KMeans to build search indices. Faster KMeans means you can re-index dynamically as data changes. ↳ LLM quantization methods need KMeans to find optimal weight codebooks, per layer, repeatedly. What takes hours could now take minutes. ↳ MoE models need fast token routing at inference time. Flash-KMeans makes it viable to run this inside the inference loop, not just in preprocessing. I have shared the paper in the replies. That said, memory is the real constraint Flash-KMeans solves, and the problem is not just limited to clustering. The vectors a RAG system stores after indexing create similar bottlenecks. I wrote a detailed walkthrough recently on cutting this vector memory by 32x with binary quantization, querying 36M+ vectors in a few milliseconds. Read it below.

Avi Chawla

89,234 просмотров • 3 месяцев назад

gemma-4-12B-agentic-fable5-composer2.5 V2 is out. the agentic upgrade to the model trained on Fable 5's reasoning. Running it now with TurboQuant llama.cpp on a single RTX 4060( 8 GB VRAM) at 30 tokens/second with full 25000 context and reasoning: # The benchmarks v2 is built for coding + agentic work. writing code, running commands, using tools, debugging, multi step technical tasks. The clearest signal is tau2 bench telecom, an agentic tool use benchmark whose diagnose → fix → verify loop mirrors real terminal/debugging work: tau2 bench telecom numbers: base Gemma 4 12B: ~15% this finetune: ~55%. (Self reported) thats a huge jump # TheTom/llama-cpp-turboquant flags: llama-server.exe -m gemma4-v2-Q4_K_M.gguf -ngl 99 -c 25000 --cache-type-k q8_0 --cache-type-v turbo3 --port 8080 Flag breakdown: -ngl 99 → full GPU offload -c 25000 → 25K context --cache-type-k q8_0 --cache-type-v turbo3 → mixed-precision KV cache — K at 8-bit, V at ~3-bit via TurboQuant (Walsh Hadamard rotated polar quant, Google's own KV-compression research). Not even merged into mainline llama.cpp. running it off a fork. No API. No cloud. Just llama.cpp. well, a fork of it and any 6gb+ GPU. If you tried yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF, check this out and share your experience with the models

gemma-4-12B-agentic-fable5-composer2.5 V2 is out. the agentic upgrade to the model trained on Fable 5's reasoning. Running it now with TurboQuant llama.cpp on a single RTX 4060( 8 GB VRAM) at 30 tokens/second with full 25000 context and reasoning: # The benchmarks v2 is built for coding + agentic work. writing code, running commands, using tools, debugging, multi step technical tasks. The clearest signal is tau2 bench telecom, an agentic tool use benchmark whose diagnose → fix → verify loop mirrors real terminal/debugging work: tau2 bench telecom numbers: base Gemma 4 12B: ~15% this finetune: ~55%. (Self reported) thats a huge jump # TheTom/llama-cpp-turboquant flags: llama-server.exe -m gemma4-v2-Q4_K_M.gguf -ngl 99 -c 25000 --cache-type-k q8_0 --cache-type-v turbo3 --port 8080 Flag breakdown: -ngl 99 → full GPU offload -c 25000 → 25K context --cache-type-k q8_0 --cache-type-v turbo3 → mixed-precision KV cache — K at 8-bit, V at ~3-bit via TurboQuant (Walsh Hadamard rotated polar quant, Google's own KV-compression research). Not even merged into mainline llama.cpp. running it off a fork. No API. No cloud. Just llama.cpp. well, a fork of it and any 6gb+ GPU. If you tried yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF, check this out and share your experience with the models

Alok

146,046 просмотров • 3 месяцев назад

Run Gemma 4 26B MoE on 8GB VRAM with 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the replies

Run Gemma 4 26B MoE on 8GB VRAM with 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the replies

Alok

292,770 просмотров • 3 месяцев назад

This Chinese developer launched Llama 70B locally on a MacBook on a plane and for a full 11 hours without internet ran client projects. He was sitting by the window on a transatlantic flight with a MacBook Pro M4 with 64 GB of memory. WiFi on board cost $25 for the flight. He declined. No cloud API, no connection to Anthropic or OpenAI servers, no internet at all. Just a local Llama 3.3 70B on bf16 and his own orchestrator script. The model runs through llama.cpp. Generation speed, 71 tokens per second. Context around 60,000 tokens. Memory usage, 48.6 GiB out of 64. Battery at takeoff, 3 hours 21 minutes. And he gave the orchestrator this system prompt before takeoff: "You are an offline orchestrator running on a single MacBook. There is no network. The only resources you have are local files in /Users/dev/work, the Llama 70B inference server at localhost:8080, and a battery budget of 3 hours 21 minutes. Process the queue at /Users/dev/work/queue.jsonl (one client task per line). For each task: draft → run local evals → save artefact to /Users/dev/work/done/. Save context checkpoints every 12 tasks so you can resume after a battery swap. Stop only on empty queue or when battery drops below 5%." So the system knows exactly what resources it is running on. It knows it has no connection to the outside world for the next 11 hours. It knows it has finite memory and a finite battery. It knows the human will not intervene until the plane lands. The system runs in 1 loop. Takes a task from the queue, runs it through inference, saves the artifact, writes a checkpoint. Task after task, just like that. And only when the battery drops below 5% does the orchestrator automatically pause, waits for the laptop to switch to the backup power bank, and continues from the last checkpoint. Here is what the system actually writes in his log during the flight: "saved context checkpoint 8 of 12 (pos_min = 488, pos_max = 50118, size = 62.813 MiB)" "restored context checkpoint (pos_min = 488, pos_max = 50118)" "prompt processing progress: n_tokens = 50 / 60 818" "task 37016 done | tps = 71 s tokens text → /Users/dev/work/done/proposal_westside.md" Outside the window, clouds, blue sky, and no WiFi. On the tray, 1 MacBook, an open terminal on 2 screens, and an inference server on localhost. From what I have observed, this is the cleanest offline AI workflow I have seen in the past year: 11 hours of flight, $0 for WiFi, and the entire client queue closed before landing.

This Chinese developer launched Llama 70B locally on a MacBook on a plane and for a full 11 hours without internet ran client projects. He was sitting by the window on a transatlantic flight with a MacBook Pro M4 with 64 GB of memory. WiFi on board cost $25 for the flight. He declined. No cloud API, no connection to Anthropic or OpenAI servers, no internet at all. Just a local Llama 3.3 70B on bf16 and his own orchestrator script. The model runs through llama.cpp. Generation speed, 71 tokens per second. Context around 60,000 tokens. Memory usage, 48.6 GiB out of 64. Battery at takeoff, 3 hours 21 minutes. And he gave the orchestrator this system prompt before takeoff: "You are an offline orchestrator running on a single MacBook. There is no network. The only resources you have are local files in /Users/dev/work, the Llama 70B inference server at localhost:8080, and a battery budget of 3 hours 21 minutes. Process the queue at /Users/dev/work/queue.jsonl (one client task per line). For each task: draft → run local evals → save artefact to /Users/dev/work/done/. Save context checkpoints every 12 tasks so you can resume after a battery swap. Stop only on empty queue or when battery drops below 5%." So the system knows exactly what resources it is running on. It knows it has no connection to the outside world for the next 11 hours. It knows it has finite memory and a finite battery. It knows the human will not intervene until the plane lands. The system runs in 1 loop. Takes a task from the queue, runs it through inference, saves the artifact, writes a checkpoint. Task after task, just like that. And only when the battery drops below 5% does the orchestrator automatically pause, waits for the laptop to switch to the backup power bank, and continues from the last checkpoint. Here is what the system actually writes in his log during the flight: "saved context checkpoint 8 of 12 (pos_min = 488, pos_max = 50118, size = 62.813 MiB)" "restored context checkpoint (pos_min = 488, pos_max = 50118)" "prompt processing progress: n_tokens = 50 / 60 818" "task 37016 done | tps = 71 s tokens text → /Users/dev/work/done/proposal_westside.md" Outside the window, clouds, blue sky, and no WiFi. On the tray, 1 MacBook, an open terminal on 2 screens, and an inference server on localhost. From what I have observed, this is the cleanest offline AI workflow I have seen in the past year: 11 hours of flight, $0 for WiFi, and the entire client queue closed before landing.

Blaze

1,843,280 просмотров • 4 месяцев назад

⛓️ Aethir - the decentralized #GPU powerhouse reshaping #AI & #gaming! Aethir is building the future of high-performance computing with a global #DePIN network of 400,000+ enterprise-grade GPU containers (including #NVIDIA H100s, H200s & more) spanning 90+ countries. 📍 Two flagship products: • Aethir Earth – Bare-metal GPU cloud delivering raw power for AI training, fine-tuning & inference with zero virtualization overhead. • Aethir Atmosphere – Low-latency cloud gaming rendering that streams high-quality experiences to any device. ☁️ Cloud Hosts monetize idle GPUs and earn $ATH rewards, while customers get scalable, cost-efficient compute (up to 80% cheaper than traditional clouds), ultra-low latency, and 95%+ utilization rates. No massive CapEx, no vendor lock-in – just on-demand access closer to the edge. From AI model training to real-time cloud gaming and beyond, Aethir is democratizing enterprise GPU power and powering the next generation of innovation. 🦾 Axe Compute’s $317M in customer prepayments. That single number reframes how AI data centers get built in 2026 🧵 The decentralized cloud is here. Are you ready? 🌐 #Aethir #DePIN #AI #GPUCloud #Web3

⛓️ Aethir - the decentralized #GPU powerhouse reshaping #AI & #gaming! Aethir is building the future of high-performance computing with a global #DePIN network of 400,000+ enterprise-grade GPU containers (including #NVIDIA H100s, H200s & more) spanning 90+ countries. 📍 Two flagship products: • Aethir Earth – Bare-metal GPU cloud delivering raw power for AI training, fine-tuning & inference with zero virtualization overhead. • Aethir Atmosphere – Low-latency cloud gaming rendering that streams high-quality experiences to any device. ☁️ Cloud Hosts monetize idle GPUs and earn $ATH rewards, while customers get scalable, cost-efficient compute (up to 80% cheaper than traditional clouds), ultra-low latency, and 95%+ utilization rates. No massive CapEx, no vendor lock-in – just on-demand access closer to the edge. From AI model training to real-time cloud gaming and beyond, Aethir is democratizing enterprise GPU power and powering the next generation of innovation. 🦾 Axe Compute’s $317M in customer prepayments. That single number reframes how AI data centers get built in 2026 🧵 The decentralized cloud is here. Are you ready? 🌐 #Aethir #DePIN #AI #GPUCloud #Web3

Crypto Holding™ 💎

229,806 просмотров • 1 месяц назад

⬛️ We are currently accelerating the incubation of GPU Nodes into the infraX Network, with 12 H100’s currently available for operation. Despite the incubation of such immense GPU power, the infraX Platform is optimally designed to run on the least amount of computational power possible, meaning a lot of our available GPU nodes are currently sitting idle. Currently, we're utilising a single gigantic NVIDIA H100 server with 80GB of VRAM and over 220GB of RAM to run our Platform. To put that in perspective, it rivals the computational power of an adult human brain. This setup enables us to handle immense computational load and deliver high-quality AI content to our users, however we have much more in store. Our remaining, immense network of GPU units is currently being prepared for rental operations as we look to transform the corporate GPU lending sphere through our corporate GPU lending protocol. We already have many high tier Web3 Players ready for technical integration, with more approaching us daily. Through our V3 DApp we look to make these integrations publicly viewable with real time usage graphs integrated directly into our Platform, allowing for exceedingly unique viewing opportunities. $INFRA

⬛️ We are currently accelerating the incubation of GPU Nodes into the infraX Network, with 12 H100’s currently available for operation. Despite the incubation of such immense GPU power, the infraX Platform is optimally designed to run on the least amount of computational power possible, meaning a lot of our available GPU nodes are currently sitting idle. Currently, we're utilising a single gigantic NVIDIA H100 server with 80GB of VRAM and over 220GB of RAM to run our Platform. To put that in perspective, it rivals the computational power of an adult human brain. This setup enables us to handle immense computational load and deliver high-quality AI content to our users, however we have much more in store. Our remaining, immense network of GPU units is currently being prepared for rental operations as we look to transform the corporate GPU lending sphere through our corporate GPU lending protocol. We already have many high tier Web3 Players ready for technical integration, with more approaching us daily. Through our V3 DApp we look to make these integrations publicly viewable with real time usage graphs integrated directly into our Platform, allowing for exceedingly unique viewing opportunities. $INFRA

infraX | $INFRA

42,843 просмотров • 1 год назад