Загрузка видео...

Не удалось загрузить видео

На главную

🚀 Self-speculation brings 6.75x real speedup for LLM generation with SGLang inference! Same model drafts future tokens in Diffusion mode → then verifies them in AR (causal) mode. One model and one KV cache. Just different attention masks. Thanks to perfect alignment, we get 2× longer acceptance lengths than...

PavloMolchanov's profile picture

Pavlo Molchanov

29,174 subscribers

66,604 просмотров • 3 месяцев назад •via X (Twitter)

ОбразованиеНовости и политикаНаука и технологии
LIVE

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

Most recent diffusion language model research (that I’ve seen) seems to be using masking as the noising process. It looks like, however, most closed-source models (Google Gemini Diffusion and possibly Inception Labs’ Mercury) use a different noising process, where instead of masking tokens, they replace them with different tokens (either with a random token or a semantically similar token). I wondered how they were getting such high throughput with the latter noising process, since I believed that optimizing inference with KVCache approximation would be more difficult (for various reasons). I visualized this noising process with tiny-diffusion and compared it to normal unmasking, and was very surprised to see how fast the generation “settles” into a reasonable output, and then only slightly refines afterwards, requiring much fewer steps in total. Unmasking (where tokens are never remasked, the typical implementation) is inherently limited in generation speed by the fact that an increase in tokens decoded per step leads to more errors due to the mismatch between individual and marginal token probability distributions we sample from. The token replacement noising process seems to have a much different set of characteristics. Because we sample each token per step, every token makes “progress” towards the final output each iteration (in addition to *potentially* giving other tokens more information in future steps). Generally, masking has outperformed other noising processes, which is probably why most research focused on it (using smaller models). But the paper referred to in the retweet shows that random replacement as a noising process may scale better as model size increases. Big labs might have noticed these results much earlier (due to having drastically more training resources and being able to test larger models), which may explain the discrepancy in the choice of noising process. I’m gonna test this with larger models, since tiny-diffusion only has 10M parameters.
0:13

Most recent diffusion language model research (that I’ve seen) seems to be using masking as the noising process. It looks like, however, most closed-source models (Google Gemini Diffusion and possibly Inception Labs’ Mercury) use a different noising process, where instead of masking tokens, they replace them with different tokens (either with a random token or a semantically similar token). I wondered how they were getting such high throughput with the latter noising process, since I believed that optimizing inference with KVCache approximation would be more difficult (for various reasons). I visualized this noising process with tiny-diffusion and compared it to normal unmasking, and was very surprised to see how fast the generation “settles” into a reasonable output, and then only slightly refines afterwards, requiring much fewer steps in total. Unmasking (where tokens are never remasked, the typical implementation) is inherently limited in generation speed by the fact that an increase in tokens decoded per step leads to more errors due to the mismatch between individual and marginal token probability distributions we sample from. The token replacement noising process seems to have a much different set of characteristics. Because we sample each token per step, every token makes “progress” towards the final output each iteration (in addition to *potentially* giving other tokens more information in future steps). Generally, masking has outperformed other noising processes, which is probably why most research focused on it (using smaller models). But the paper referred to in the retweet shows that random replacement as a noising process may scale better as model size increases. Big labs might have noticed these results much earlier (due to having drastically more training resources and being able to test larger models), which may explain the discrepancy in the choice of noising process. I’m gonna test this with larger models, since tiny-diffusion only has 10M parameters.

nathan (in sf)

40,440 просмотров • 7 месяцев назад

I told you to claim your free 16GB NVIDIA GPU for learning Local LLMs. Now I’m going to show you how to double its inference speed without touching the hardware. Google Colab gives you an enterprise grade NVIDIA Tesla T4 GPU for free, roughly 4 hours every single day. It is the absolute perfect sandbox for learning AI engineering, testing inference flags, and pushing massive context windows. The local AI timeline is moving way too fast. If you aren't using Multi Token Prediction (MTP) yet, you are leaving massive performance on the table. I just pushed DeepMind’s Gemma 4 26B to 64.9 t/s on this exact free tier. Let's look at the raw benchmark data running on an Ubuntu Linux environment with the latest compiled llama.cpp binaries and quantized GGUFs from Unsloth via HuggingFace: # Qwen 3.5 9B (Dense): Base: [ Prompt: 626.7 t/s | Generation: 21.0 t/s ] With MTP: [ Prompt: 539.1 t/s | Generation: 24.8 t/s ] # Gemma 4 26B QAT (MoE): Base: [ Prompt: 634.2 t/s | Generation: 48.3 t/s ] With MTP: [ Prompt: 572.1 t/s | Generation: 64.9 t/s ] If you are paying attention, this single Colab notebook reveals 3 massive observations about the current state of local LLMs: # 1. The MTP Speedup (Software Overclocking) Standard autoregressive decoding guesses one token at a time. MTP acts like a highly optimized, built in speculative decoder. It predicts multiple future tokens at once and the main model verifies them in parallel. The result? Zero accuracy loss and a massive throughput increase. Gemma jumped from 48 to 65 t/s just by flipping a flag. # 2. The MoE Paradox (Bigger is Faster) How does a 26B parameter model absolutely destroy a 9B model in raw speed on the exact same hardware? Architecture. Qwen 3.5 9B is a dense model. it activates all 9 billion parameters for every single token. Gemma 4 26B is a Mixture of Experts (MoE) model. It routes data efficiently, activating only 4B parameters per token. You get the reasoning capabilities of a 26B model with the compute cost of a 4B model. 3. Thinking Efficiency When I ran the exact same complex prompt on both models, the larger MoE spent significantly fewer "thinking" tokens to arrive at the correct answer. A smarter model doesn't just give better answers; it gets to the point faster, saving you compute cycles and preserving your context window. # Want to run this yourself? Here are the exact llama.cpp CLI commands. For Qwen (MTP is baked into the main model): ./llama-cli -m Qwen3.5-9B-UD-Q4_K_XL.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 For Gemma (Using a separate lightweight draft model): ./llama-cli -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --model-draft mtp-gemma-4-26B-A4B-it.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 Stop waiting for a $3,000 rig. Boot up Colab, pull these models, and start building your stack. I’ve put together a completely free, cell by cell Google Colab notebook that automates this entire workflow so you can test it yourself in 5 minutes and learn. Link to the notebook is in the comments below. Experiemt with different MTP parameters, context windows and post your results in the comments.
0:12

I told you to claim your free 16GB NVIDIA GPU for learning Local LLMs. Now I’m going to show you how to double its inference speed without touching the hardware. Google Colab gives you an enterprise grade NVIDIA Tesla T4 GPU for free, roughly 4 hours every single day. It is the absolute perfect sandbox for learning AI engineering, testing inference flags, and pushing massive context windows. The local AI timeline is moving way too fast. If you aren't using Multi Token Prediction (MTP) yet, you are leaving massive performance on the table. I just pushed DeepMind’s Gemma 4 26B to 64.9 t/s on this exact free tier. Let's look at the raw benchmark data running on an Ubuntu Linux environment with the latest compiled llama.cpp binaries and quantized GGUFs from Unsloth via HuggingFace: # Qwen 3.5 9B (Dense): Base: [ Prompt: 626.7 t/s | Generation: 21.0 t/s ] With MTP: [ Prompt: 539.1 t/s | Generation: 24.8 t/s ] # Gemma 4 26B QAT (MoE): Base: [ Prompt: 634.2 t/s | Generation: 48.3 t/s ] With MTP: [ Prompt: 572.1 t/s | Generation: 64.9 t/s ] If you are paying attention, this single Colab notebook reveals 3 massive observations about the current state of local LLMs: # 1. The MTP Speedup (Software Overclocking) Standard autoregressive decoding guesses one token at a time. MTP acts like a highly optimized, built in speculative decoder. It predicts multiple future tokens at once and the main model verifies them in parallel. The result? Zero accuracy loss and a massive throughput increase. Gemma jumped from 48 to 65 t/s just by flipping a flag. # 2. The MoE Paradox (Bigger is Faster) How does a 26B parameter model absolutely destroy a 9B model in raw speed on the exact same hardware? Architecture. Qwen 3.5 9B is a dense model. it activates all 9 billion parameters for every single token. Gemma 4 26B is a Mixture of Experts (MoE) model. It routes data efficiently, activating only 4B parameters per token. You get the reasoning capabilities of a 26B model with the compute cost of a 4B model. 3. Thinking Efficiency When I ran the exact same complex prompt on both models, the larger MoE spent significantly fewer "thinking" tokens to arrive at the correct answer. A smarter model doesn't just give better answers; it gets to the point faster, saving you compute cycles and preserving your context window. # Want to run this yourself? Here are the exact llama.cpp CLI commands. For Qwen (MTP is baked into the main model): ./llama-cli -m Qwen3.5-9B-UD-Q4_K_XL.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 For Gemma (Using a separate lightweight draft model): ./llama-cli -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --model-draft mtp-gemma-4-26B-A4B-it.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 Stop waiting for a $3,000 rig. Boot up Colab, pull these models, and start building your stack. I’ve put together a completely free, cell by cell Google Colab notebook that automates this entire workflow so you can test it yourself in 5 minutes and learn. Link to the notebook is in the comments below. Experiemt with different MTP parameters, context windows and post your results in the comments.

Alok

170,442 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?
0:16

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 • 19 дней назад

Depth Any Video with Scalable Synthetic Data AI physicists and chemists continue to make strides in depth estimation from video. Check out this new paper featuring some impressive examples. See the thread for more details (unfortunately no code yet). Abstract: Video depth estimation has long been hindered by the scarcity of consistent and scalable ground truth data, leading to inconsistent and unreliable results. In this paper, we introduce Depth Any Video, a model that tackles the challenge through two key innovations. First, we develop a scalable synthetic data pipeline, capturing real-time video depth data from diverse game environments, yielding 40,000 video clips of 5-second duration, each with precise depth annotations. Second, we leverage the powerful priors of generative video diffusion models to handle real-world videos effectively, integrating advanced techniques such as rotary position encoding and flow matching to further enhance flexibility and efficiency. Unlike previous models, which are limited to fixed-length video sequences, our approach introduces a novel mixed-duration training strategy that handles videos of varying lengths and performs robustly across different frame rates 0 - even on single frames. At inference, we propose a depth interpolation method that enables our model to infer high-resolution video depth across sequences of up to 150 frames. Our model outperforms all previous generative depth models in terms of spatial accuracy and temporal consistency.
0:13

Depth Any Video with Scalable Synthetic Data AI physicists and chemists continue to make strides in depth estimation from video. Check out this new paper featuring some impressive examples. See the thread for more details (unfortunately no code yet). Abstract: Video depth estimation has long been hindered by the scarcity of consistent and scalable ground truth data, leading to inconsistent and unreliable results. In this paper, we introduce Depth Any Video, a model that tackles the challenge through two key innovations. First, we develop a scalable synthetic data pipeline, capturing real-time video depth data from diverse game environments, yielding 40,000 video clips of 5-second duration, each with precise depth annotations. Second, we leverage the powerful priors of generative video diffusion models to handle real-world videos effectively, integrating advanced techniques such as rotary position encoding and flow matching to further enhance flexibility and efficiency. Unlike previous models, which are limited to fixed-length video sequences, our approach introduces a novel mixed-duration training strategy that handles videos of varying lengths and performs robustly across different frame rates 0 - even on single frames. At inference, we propose a depth interpolation method that enables our model to infer high-resolution video depth across sequences of up to 150 frames. Our model outperforms all previous generative depth models in terms of spatial accuracy and temporal consistency.

MrNeRF

27,428 просмотров • 1 год назад

Run Gemma 4 26b MTP on 8 GB VRAM GPUs at 25+ tokens/second. Flags included! local llm space is moving at terminal velocity. only 3 days ago google released gemma 4 26b a4b qat quants. more efficient than before, ran on 8gb vram at 20 tok/sec. and now just a few hours ago, mainline llama.cpp merged a massive update and we just shattered our own record. decode throughput went 25-40% up on the same 8 GB VRAM setup! Before MTP: 20 tps -> After MTP: 28 tps! llama.cpp just officially merged PR #23398 ("add Gemma4 MTP"), bringing native Multi-Token Prediction (MTP) support to Gemma 4 models. By running speculative drafting on the same 8GB VRAM RTX 4060 setup, my decode throughput on a 64k context instantly leaped to a blistering 25–27 tokens/sec thats 25-30% increase with the same hardware. Here is the architectural catch you need to know: Unlike the Qwen 3.5 and 3.6 series, which bake the MTP heads directly into the base GGUF, the Gemma 4 MTP head is not built in. You must download a separate, specialized MTP drafter GGUF (the assistant model) to act as the speculator. (I've dropped the download link in the replies). copy and try the exact flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp --spec-draft-n-max 6 --spec-draft-p-min 0.7 --spec-draft-model gemma-4-26b-A4B-it-assistant-Q4_0.gguf -c 64000 -v n-max 4 and p-min 0.7 is also worth checking out. benchmark on your setup and workflow. if you have a single 8 gb vram nvidia rtx 4060, 3060, 3070, 2080, 2070, grab the MTP drafter GGUF link in the comments and try it yourself. Check it out even if you have asmaller or a larger gpu, such as a single rtx 3090, 4090, 3060, 2060. MTP works for all gemma 4 sizes such as gemma 4 12b, gemma 4 31b etc. but remember to grab the correct mtp draft assistant models respectively. what are you benchmarking today
0:15

Run Gemma 4 26b MTP on 8 GB VRAM GPUs at 25+ tokens/second. Flags included! local llm space is moving at terminal velocity. only 3 days ago google released gemma 4 26b a4b qat quants. more efficient than before, ran on 8gb vram at 20 tok/sec. and now just a few hours ago, mainline llama.cpp merged a massive update and we just shattered our own record. decode throughput went 25-40% up on the same 8 GB VRAM setup! Before MTP: 20 tps -> After MTP: 28 tps! llama.cpp just officially merged PR #23398 ("add Gemma4 MTP"), bringing native Multi-Token Prediction (MTP) support to Gemma 4 models. By running speculative drafting on the same 8GB VRAM RTX 4060 setup, my decode throughput on a 64k context instantly leaped to a blistering 25–27 tokens/sec thats 25-30% increase with the same hardware. Here is the architectural catch you need to know: Unlike the Qwen 3.5 and 3.6 series, which bake the MTP heads directly into the base GGUF, the Gemma 4 MTP head is not built in. You must download a separate, specialized MTP drafter GGUF (the assistant model) to act as the speculator. (I've dropped the download link in the replies). copy and try the exact flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp --spec-draft-n-max 6 --spec-draft-p-min 0.7 --spec-draft-model gemma-4-26b-A4B-it-assistant-Q4_0.gguf -c 64000 -v n-max 4 and p-min 0.7 is also worth checking out. benchmark on your setup and workflow. if you have a single 8 gb vram nvidia rtx 4060, 3060, 3070, 2080, 2070, grab the MTP drafter GGUF link in the comments and try it yourself. Check it out even if you have asmaller or a larger gpu, such as a single rtx 3090, 4090, 3060, 2060. MTP works for all gemma 4 sizes such as gemma 4 12b, gemma 4 31b etc. but remember to grab the correct mtp draft assistant models respectively. what are you benchmarking today

Alok

200,913 просмотров • 2 месяцев назад

Building The On-Chain Cooperative 🟡 Welcome to the dawn of a new era in the crypto space, where the buzzword "community" is not just a hollow echo but a vibrant force that propels us towards a brighter future. Let's delve into the heart of MODE, the Onchain Cooperative that seeks to redefine the landscape of web3. What does MODE stand for? MODE stands for building an on-chain cooperative focused on sustainable growth and collective prosperity. At its core, MODE is guided by the principles of cooperation, shared incentives, and community-driven development. The goal is to shift from the "fat protocol" mentality where most value accrues to the blockchain/protocol itself, towards an ecosystem where builders, users, and applications can thrive together. What’s MODE's vision and mission in the web3 space? MODE's vision is to return to web3's founding promise - a future that is better for all, not just the individual. A world with aligned incentives that drive growth for everyone involved. A place with opportunities for all, not just the few. The mission is to pioneer the on-chain cooperative - where contributors are rewarded fairly based on the value they provide. Features like Sequencer Fee Sharing distribute a portion of fees to smart contract developers, incentivizing participation. The aim is to encourage collaboration instead of confrontation. Together, the MODE community can deliver new models for cooperation and shared prosperity in web3. Mode Network will solve many problems today in Web3: • Lack of incentives for developers: Developers creating decentralized apps (dApps) currently have few direct economic incentives to create and maintain their projects. Mode provides them with a steady source of income through fee-sharing. • Lack of collaboration: There are few incentives for blockchain projects to compete less and collaborate more for the benefit of the entire ecosystem. Mode's model encourages collaboration by aligning participants economically. • Excessive value accrual at the protocol layer: Mode aims for a more balanced model where the protocol's success is fueled by the success of application developers/builders and the wider community. Growth is a two-way street – "as we grow, you grow". The MODE Pledge 💛 The promise of crypto and blockchain is a brighter future. One that is better for all not just the individual. Where nothing is more important than community. We've strayed from this path. Entering a world of player vs player. Where value is extracted rather than shared. The game is zero sum rather than positive sum. And incentives are aligned with domination, rather than cooperation. Mode is the dawn of a new age. and a return to the promise of what can be. A world with aligned incentives that drive growth for builders, users and projects. A place with opportunities for all, rather than the few. Where we say goodbye to the 'fat protocol', and hello to the onchain cooperative. Join us on our mission to grow together. If this vision for a community-powered web3 ecosystem resonates - where creators are rewarded for their contributions - you can join the MODE on-chain cooperative! Visit Join the discord community Follow Mode 🟡 Together, we can transform web3 into a positive-sum game that unlocks new possibilities for all. Where your growth fuels the growth of others. Let's build the on-chain cooperative!
0:16

Building The On-Chain Cooperative 🟡 Welcome to the dawn of a new era in the crypto space, where the buzzword "community" is not just a hollow echo but a vibrant force that propels us towards a brighter future. Let's delve into the heart of MODE, the Onchain Cooperative that seeks to redefine the landscape of web3. What does MODE stand for? MODE stands for building an on-chain cooperative focused on sustainable growth and collective prosperity. At its core, MODE is guided by the principles of cooperation, shared incentives, and community-driven development. The goal is to shift from the "fat protocol" mentality where most value accrues to the blockchain/protocol itself, towards an ecosystem where builders, users, and applications can thrive together. What’s MODE's vision and mission in the web3 space? MODE's vision is to return to web3's founding promise - a future that is better for all, not just the individual. A world with aligned incentives that drive growth for everyone involved. A place with opportunities for all, not just the few. The mission is to pioneer the on-chain cooperative - where contributors are rewarded fairly based on the value they provide. Features like Sequencer Fee Sharing distribute a portion of fees to smart contract developers, incentivizing participation. The aim is to encourage collaboration instead of confrontation. Together, the MODE community can deliver new models for cooperation and shared prosperity in web3. Mode Network will solve many problems today in Web3: • Lack of incentives for developers: Developers creating decentralized apps (dApps) currently have few direct economic incentives to create and maintain their projects. Mode provides them with a steady source of income through fee-sharing. • Lack of collaboration: There are few incentives for blockchain projects to compete less and collaborate more for the benefit of the entire ecosystem. Mode's model encourages collaboration by aligning participants economically. • Excessive value accrual at the protocol layer: Mode aims for a more balanced model where the protocol's success is fueled by the success of application developers/builders and the wider community. Growth is a two-way street – "as we grow, you grow". The MODE Pledge 💛 The promise of crypto and blockchain is a brighter future. One that is better for all not just the individual. Where nothing is more important than community. We've strayed from this path. Entering a world of player vs player. Where value is extracted rather than shared. The game is zero sum rather than positive sum. And incentives are aligned with domination, rather than cooperation. Mode is the dawn of a new age. and a return to the promise of what can be. A world with aligned incentives that drive growth for builders, users and projects. A place with opportunities for all, rather than the few. Where we say goodbye to the 'fat protocol', and hello to the onchain cooperative. Join us on our mission to grow together. If this vision for a community-powered web3 ecosystem resonates - where creators are rewarded for their contributions - you can join the MODE on-chain cooperative! Visit Join the discord community Follow Mode 🟡 Together, we can transform web3 into a positive-sum game that unlocks new possibilities for all. Where your growth fuels the growth of others. Let's build the on-chain cooperative!

ETHachi Uchiha | Crypto DEGENius

16,774 просмотров • 2 лет назад

Qwen 3.8 27B Q4_K_M - 90 tokens/sec on a single NVIDIA RTX 4090 (24 GB VRAM) with Dflash2! (MTP 60 tps -> 90 tps Dflash2!!!!) Local AI moves so fast (literally!) it’s terrifying. Z lab just dropped DFlash 2 for Qwen 3.8 27b and Muse Glimmer. I patched llama.cpp (PR #27342) and paired it with Unsloth’s Qwen 3.8 27B UD-Q4_K_XL quant. The result? Lossless 90 tokens/s decode. My last post highlighted native MTP hitting 60 t/s at 130,000 context. But DFlash 2 just completely shattered that ceiling. By using parallel block diffusion drafting (predicting whole blocks of tokens in a single pass using dynamic convolutions), DFlash achieves a massive 5.39 token acceptance rate. THE ALPHA TWEAK: `n-max 7` eats too much VRAM for draft states. But if you drop the draft limit to `--spec-draft-n-max 4`, you slash the VRAM overhead and actually increase the throughput. Here is the new 24GB VRAM Physics Matrix (DFlash 2 @ n-max 4): - 30k Context: 1,725 t/s prefill | 87.05 t/s decode | 22.2 GB VRAM - 80k Context: 1,789 t/s prefill | 84.20 t/s decode | 23.3 GB VRAM - 110k Context: 1,767 t/s prefill | 83.35 t/s decode | 23.96 GB VRAM (110k context at 83+ tokens a second sitting exactly on the 24GB hardware limit is absolute wizardry). How to compile the PR today: git clone cd llama.cpp git fetch origin pull/27342/head:pr-27342 git switch pr-27342 cmake -B build -DGGML_CUDA=ON && cmake --build build -j Llama.cpp flags for Dflash (110k Context Ceiling): ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q4_K_M.gguf --spec-type draft-dflash --spec-draft-n-max 4 -c 110000 -ngl 99 --port 8080 -ctv q4_0 -ctk q4_0 The fact that the open source community is shipping block diffusion drafters so quickly that run entirely locally on a gaming GPU is unbelievable. If you own a single RTX 3090 or 4090, it is officially time to upgrade to qwen 3.8 27b with dflash 2 and cancel your API subscriptions and let local silicon eat the cloud. This model beats GPT 5.6 Terra, GLM 5.2 DeepSeek V4 Pro, Muse Spark 1.2 and Claude Opus 4.8 on the artificial analysis agentic index (details in the replies) Hugging Face GGUF links (Base + DFlash2) and the full visual VRAM scaling and Dflash2 vs MTP graphs are also in the replies below. are you sticking to native MTP for the 130k context, or sacrificing 20k context to redline your decode speed? How many tokens/sec are you pushing on your current local rig?
0:10

Qwen 3.8 27B Q4_K_M - 90 tokens/sec on a single NVIDIA RTX 4090 (24 GB VRAM) with Dflash2! (MTP 60 tps -> 90 tps Dflash2!!!!) Local AI moves so fast (literally!) it’s terrifying. Z lab just dropped DFlash 2 for Qwen 3.8 27b and Muse Glimmer. I patched llama.cpp (PR #27342) and paired it with Unsloth’s Qwen 3.8 27B UD-Q4_K_XL quant. The result? Lossless 90 tokens/s decode. My last post highlighted native MTP hitting 60 t/s at 130,000 context. But DFlash 2 just completely shattered that ceiling. By using parallel block diffusion drafting (predicting whole blocks of tokens in a single pass using dynamic convolutions), DFlash achieves a massive 5.39 token acceptance rate. THE ALPHA TWEAK: `n-max 7` eats too much VRAM for draft states. But if you drop the draft limit to `--spec-draft-n-max 4`, you slash the VRAM overhead and actually increase the throughput. Here is the new 24GB VRAM Physics Matrix (DFlash 2 @ n-max 4): - 30k Context: 1,725 t/s prefill | 87.05 t/s decode | 22.2 GB VRAM - 80k Context: 1,789 t/s prefill | 84.20 t/s decode | 23.3 GB VRAM - 110k Context: 1,767 t/s prefill | 83.35 t/s decode | 23.96 GB VRAM (110k context at 83+ tokens a second sitting exactly on the 24GB hardware limit is absolute wizardry). How to compile the PR today: git clone cd llama.cpp git fetch origin pull/27342/head:pr-27342 git switch pr-27342 cmake -B build -DGGML_CUDA=ON && cmake --build build -j Llama.cpp flags for Dflash (110k Context Ceiling): ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q4_K_M.gguf --spec-type draft-dflash --spec-draft-n-max 4 -c 110000 -ngl 99 --port 8080 -ctv q4_0 -ctk q4_0 The fact that the open source community is shipping block diffusion drafters so quickly that run entirely locally on a gaming GPU is unbelievable. If you own a single RTX 3090 or 4090, it is officially time to upgrade to qwen 3.8 27b with dflash 2 and cancel your API subscriptions and let local silicon eat the cloud. This model beats GPT 5.6 Terra, GLM 5.2 DeepSeek V4 Pro, Muse Spark 1.2 and Claude Opus 4.8 on the artificial analysis agentic index (details in the replies) Hugging Face GGUF links (Base + DFlash2) and the full visual VRAM scaling and Dflash2 vs MTP graphs are also in the replies below. are you sticking to native MTP for the 130k context, or sacrificing 20k context to redline your decode speed? How many tokens/sec are you pushing on your current local rig?

Alok

103,895 просмотров • 11 дней назад

Transformer by hand ✍️ ~ 6 steps walkthrough below Open the hood of a transformer and the parts list is overwhelming: embeddings, positional encoding, attention weighting, self-attention, cross-attention, multi-head attention, layer norm, skip connections, softmax, linear, Nx, shifted right, query, key, value, masking. Which of those actually make the car run? Two of them. Attention weighting and the feed-forward network. Everything else is an enhancement to make it run faster and longer, which is how we got from a car to a truck, and to the word "large" in large language model. So I drew and calculated those two parts entirely by hand. Goal: push five features through one transformer block, filling in every cell yourself. 1. Given Five positions of input features, arriving from the previous block. 2. Attention matrix Let us feed all five features to a query-key module (QK) and read back an attention weight matrix, A. The details of that module are a post of their own. 3. Attention weighting We multiply the input features by A to get the attention weighted features, Z. Still five positions. The effect is to combine features *across positions*, horizontally: X1 becomes X1 + X2, X2 becomes X2 + X3, and so on. 4. First layer Let us feed all five weighted features into the first layer of the FFN. Multiply by the weights and biases. This time the combining happens *across feature dimensions*, vertically, and each feature grows from 3 numbers to 4. Note that every position goes through the same weight matrix. That is what "position-wise" means. 5. ReLU We cross out the negatives. They become zeros. 6. Second layer Let us bring it back down: 4 dimensions to 3. The output feeds the next block, which has a completely separate set of parameters, and the whole thing runs again. You have just calculated a transformer block by hand. ✍️ The takeaway: the two parts are doing two different jobs, and neither one alone is enough. Attention mixes *across positions*, so a feature can see its neighbours. The FFN mixes *across feature dimensions*, so each position can think about itself. Horizontal, then vertical. Then that pattern repeats N times, each block with its own separate set of weights. That is the Nx from the list up top, and that is what makes the transformer run. 💾 Save this post! #AIbyHand #Transformers #DeepLearning
0:12

Transformer by hand ✍️ ~ 6 steps walkthrough below Open the hood of a transformer and the parts list is overwhelming: embeddings, positional encoding, attention weighting, self-attention, cross-attention, multi-head attention, layer norm, skip connections, softmax, linear, Nx, shifted right, query, key, value, masking. Which of those actually make the car run? Two of them. Attention weighting and the feed-forward network. Everything else is an enhancement to make it run faster and longer, which is how we got from a car to a truck, and to the word "large" in large language model. So I drew and calculated those two parts entirely by hand. Goal: push five features through one transformer block, filling in every cell yourself. 1. Given Five positions of input features, arriving from the previous block. 2. Attention matrix Let us feed all five features to a query-key module (QK) and read back an attention weight matrix, A. The details of that module are a post of their own. 3. Attention weighting We multiply the input features by A to get the attention weighted features, Z. Still five positions. The effect is to combine features *across positions*, horizontally: X1 becomes X1 + X2, X2 becomes X2 + X3, and so on. 4. First layer Let us feed all five weighted features into the first layer of the FFN. Multiply by the weights and biases. This time the combining happens *across feature dimensions*, vertically, and each feature grows from 3 numbers to 4. Note that every position goes through the same weight matrix. That is what "position-wise" means. 5. ReLU We cross out the negatives. They become zeros. 6. Second layer Let us bring it back down: 4 dimensions to 3. The output feeds the next block, which has a completely separate set of parameters, and the whole thing runs again. You have just calculated a transformer block by hand. ✍️ The takeaway: the two parts are doing two different jobs, and neither one alone is enough. Attention mixes *across positions*, so a feature can see its neighbours. The FFN mixes *across feature dimensions*, so each position can think about itself. Horizontal, then vertical. Then that pattern repeats N times, each block with its own separate set of weights. That is the Nx from the list up top, and that is what makes the transformer run. 💾 Save this post! #AIbyHand #Transformers #DeepLearning

Tom Yeh

26,089 просмотров • 1 месяц назад

AI has had exactly two scaling axes that worked so far, and the second one is starting to look finite too the first one was pretraining: with scaling parameters and data, we got world knowledge (i.e. ChatGPT had read enough to know things), but it started saturating a while ago the second one was RL, and people had been doing RL the whole time before that: RLHF is RL but it never scaled far because it was trying to control the exact output, which tokens come out, how the text reads, but you can only push that so far before you’re just polishing RLVR dropped that constraint: giving the model a task, then checking whether the final answer is right, and ignoring everything in between -- so the model does whatever it wants in the middle and only the endpoint gets graded, and that’s much closer to actual RL and it’s what bought us planning and reasoning (arguably, tool use sits around 2.5 on this list -- while useful, it's not a different kind of thing) so one axis gave knowledge, the other gave reasoning, and both of them are one model working alone the next axis is how many models you can get working on the same problem, which is a different kind of axis than the previous two we know that multi-agent RL has always been the harder problem: I spent years in that literature and the gap between single-agent and multi-agent is definitely not incremental -- it’s a whole different class of difficulty! which is also why the derivatives are steep at the start, nobody has picked the easy wins yet... and the thing that gates this multi-agent coordination is communication: models can only coordinate as well as they can exchange information, and right now they do that by writing sentences to each other imagine what could we possibly achieve if we properly open that third axis development by letting models to exchange information in their native "language" without loosing any computational data that they produce during inference
0:14

AI has had exactly two scaling axes that worked so far, and the second one is starting to look finite too the first one was pretraining: with scaling parameters and data, we got world knowledge (i.e. ChatGPT had read enough to know things), but it started saturating a while ago the second one was RL, and people had been doing RL the whole time before that: RLHF is RL but it never scaled far because it was trying to control the exact output, which tokens come out, how the text reads, but you can only push that so far before you’re just polishing RLVR dropped that constraint: giving the model a task, then checking whether the final answer is right, and ignoring everything in between -- so the model does whatever it wants in the middle and only the endpoint gets graded, and that’s much closer to actual RL and it’s what bought us planning and reasoning (arguably, tool use sits around 2.5 on this list -- while useful, it's not a different kind of thing) so one axis gave knowledge, the other gave reasoning, and both of them are one model working alone the next axis is how many models you can get working on the same problem, which is a different kind of axis than the previous two we know that multi-agent RL has always been the harder problem: I spent years in that literature and the gap between single-agent and multi-agent is definitely not incremental -- it’s a whole different class of difficulty! which is also why the derivatives are steep at the start, nobody has picked the easy wins yet... and the thing that gates this multi-agent coordination is communication: models can only coordinate as well as they can exchange information, and right now they do that by writing sentences to each other imagine what could we possibly achieve if we properly open that third axis development by letting models to exchange information in their native "language" without loosing any computational data that they produce during inference

Sasha Malysheva

11,548 просмотров • 18 дней назад

AI Is Moving Beyond “Generating Videos” — Toward “Generating Worlds” Over the past two years, AI video models have advanced at an astonishing pace. From Runway and Pika to Sora and Veo, AI-generated videos have become increasingly realistic and more consistent with the physical laws of the real world. Many people believe the next objective is simply to generate videos that are longer, sharper, and more lifelike. But if we take a step back, we can see that the real transformation is not happening in video itself. It is happening in world models. What Is a World Model? In 1943, psychologist Kenneth Craik proposed an idea that would influence artificial intelligence research for decades. He argued that the human brain does not merely react to the outside world. Instead, it maintains an internal model of how the world works. Because we have this internal model, we can predict the outcome of an action before we actually take it. Before crossing a road, we estimate whether a car will pass by. Before catching a ball, we predict its trajectory. These abilities come from continuously simulating the world in our minds, rather than relying entirely on trial and error. This idea later became known by a more formal term: World Model. A world model does not describe a single image or a fixed video clip. It is an internal representation capable of continuously simulating the rules and dynamics of the real world. Why Is AI Research Turning Toward World Models? Because predicting “what comes next” is becoming increasingly central to how AI systems work. Language models predict the next token. Image models predict the next step in the denoising process. Video models predict the next frame. A world model, however, attempts to predict something broader: What should the world look like in the next moment? In 2018, David Ha and Jürgen Schmidhuber proposed in their paper World Models that an intelligent agent could first learn a model of the world, and then use that internal model to plan its actions. The Dreamer series later demonstrated that many complex tasks could be learned by training agents inside an “imagined world.” At the same time, the development of video models such as Sora and Veo led researchers to another realization: A model capable of continuously generating video has already learned, at least implicitly, many of the rules governing the real world. As a result, these two research directions have gradually begun to converge. But Video Is Not Yet a World This is where the distinction is often misunderstood. For a world model to support meaningful real-time interaction, it must solve several critical problems. Most video models today are essentially answering one question: What should the next frame look like? A true world model needs to answer much more: What happens if I take one step forward? If I walk behind a building and then return, will the building still be there? If I suddenly change the camera angle, will the entire space remain consistent? If I enter a command such as: “Summon a dragon.” Will the world respond immediately? In other words, a world model must do more than generate content. It must understand space. It must understand time. It must understand causality. And it must understand interaction. Moving from watching to participating is where the real difficulty of world models begins. World Models Are Entering the Interactive Era One of the latest attempts in this direction is Alaya World, recently open-sourced by Alaya World, or Alaya Lab. Instead of generating a fixed video clip, it generates a world that users can explore in real time. Users can begin with text, an image, or a video, enter the generated scene, move freely through it, and introduce new prompts at any moment during generation. The world responds immediately. According to the publicly released information, Alaya World provides: Real-time streaming generation at 720p and 24 FPS Stable continuous exploration for more than one minute The ability to switch prompts and trigger skills or events during generation Model weights and inference code released under the Apache 2.0 License Training code and datasets planned for future release What makes these capabilities important is not simply the technical specifications. It is that the generated “world” can now support continuous interaction. The official demo shows that users can genuinely control, transform, and explore the generated environment. AI Is Evolving From a Tool Into an Environment Over the past few years, most discussions around AI have focused on content generation. Generating text. Generating images. Generating videos. But world models raise a fundamentally different question: Can AI generate an environment that people can inhabit, explore, and continuously evolve? If the answer is yes, the impact will extend far beyond video generation. Game development, robotics training, embodied intelligence, digital twins, virtual production, and many other fields could be transformed by the development of world models. World models are still at a very early stage. Yet from Craik’s proposal of an internal mental model more than eighty years ago to the emergence of today’s interactive world-generation systems, a clear evolutionary path is beginning to take shape. Perhaps what AI is ultimately learning has never been limited to images, videos, or language. Perhaps it is learning the world itself. References GitHub: Technical Report:
0:10

AI Is Moving Beyond “Generating Videos” — Toward “Generating Worlds” Over the past two years, AI video models have advanced at an astonishing pace. From Runway and Pika to Sora and Veo, AI-generated videos have become increasingly realistic and more consistent with the physical laws of the real world. Many people believe the next objective is simply to generate videos that are longer, sharper, and more lifelike. But if we take a step back, we can see that the real transformation is not happening in video itself. It is happening in world models. What Is a World Model? In 1943, psychologist Kenneth Craik proposed an idea that would influence artificial intelligence research for decades. He argued that the human brain does not merely react to the outside world. Instead, it maintains an internal model of how the world works. Because we have this internal model, we can predict the outcome of an action before we actually take it. Before crossing a road, we estimate whether a car will pass by. Before catching a ball, we predict its trajectory. These abilities come from continuously simulating the world in our minds, rather than relying entirely on trial and error. This idea later became known by a more formal term: World Model. A world model does not describe a single image or a fixed video clip. It is an internal representation capable of continuously simulating the rules and dynamics of the real world. Why Is AI Research Turning Toward World Models? Because predicting “what comes next” is becoming increasingly central to how AI systems work. Language models predict the next token. Image models predict the next step in the denoising process. Video models predict the next frame. A world model, however, attempts to predict something broader: What should the world look like in the next moment? In 2018, David Ha and Jürgen Schmidhuber proposed in their paper World Models that an intelligent agent could first learn a model of the world, and then use that internal model to plan its actions. The Dreamer series later demonstrated that many complex tasks could be learned by training agents inside an “imagined world.” At the same time, the development of video models such as Sora and Veo led researchers to another realization: A model capable of continuously generating video has already learned, at least implicitly, many of the rules governing the real world. As a result, these two research directions have gradually begun to converge. But Video Is Not Yet a World This is where the distinction is often misunderstood. For a world model to support meaningful real-time interaction, it must solve several critical problems. Most video models today are essentially answering one question: What should the next frame look like? A true world model needs to answer much more: What happens if I take one step forward? If I walk behind a building and then return, will the building still be there? If I suddenly change the camera angle, will the entire space remain consistent? If I enter a command such as: “Summon a dragon.” Will the world respond immediately? In other words, a world model must do more than generate content. It must understand space. It must understand time. It must understand causality. And it must understand interaction. Moving from watching to participating is where the real difficulty of world models begins. World Models Are Entering the Interactive Era One of the latest attempts in this direction is Alaya World, recently open-sourced by Alaya World, or Alaya Lab. Instead of generating a fixed video clip, it generates a world that users can explore in real time. Users can begin with text, an image, or a video, enter the generated scene, move freely through it, and introduce new prompts at any moment during generation. The world responds immediately. According to the publicly released information, Alaya World provides: Real-time streaming generation at 720p and 24 FPS Stable continuous exploration for more than one minute The ability to switch prompts and trigger skills or events during generation Model weights and inference code released under the Apache 2.0 License Training code and datasets planned for future release What makes these capabilities important is not simply the technical specifications. It is that the generated “world” can now support continuous interaction. The official demo shows that users can genuinely control, transform, and explore the generated environment. AI Is Evolving From a Tool Into an Environment Over the past few years, most discussions around AI have focused on content generation. Generating text. Generating images. Generating videos. But world models raise a fundamentally different question: Can AI generate an environment that people can inhabit, explore, and continuously evolve? If the answer is yes, the impact will extend far beyond video generation. Game development, robotics training, embodied intelligence, digital twins, virtual production, and many other fields could be transformed by the development of world models. World models are still at a very early stage. Yet from Craik’s proposal of an internal mental model more than eighty years ago to the emergence of today’s interactive world-generation systems, a clear evolutionary path is beginning to take shape. Perhaps what AI is ultimately learning has never been limited to images, videos, or language. Perhaps it is learning the world itself. References GitHub: Technical Report:

雪踏乌云

113,347 просмотров • 1 месяц назад

50% more context unlocked for Qwen 3.8 27b Q4_K_XL dflash 2 on a single RTX 4090 (24 GB VRAM) I found a hidden VRAM tax in llama.cpp. By combining my custom 2 bit DFlash 2 drafter with one overlooked server flag, I just unlocked another +80,000 tokens of context. Qwen3.8-27B is now running a massive 250,000 context at 75 tokens/s on a single RTX 4090. Here is the secret: By default, `llama-server` reserves massive chunks of your VRAM to handle multiple concurrent users (batching). If you are running a single user session, you are bleeding memory for features you aren't using. By passing the `--parallel 1` flag, you force the engine to dedicate 100% of your 24GB VRAM buffer to a single user. When we combine the VRAM saved by our Q2_K 2-bit drafter with the VRAM saved by `--parallel 1`, the context ceilings absolutely explode: Note: all benchmarks carried out with a massive 28k prompt. Ubuntu 22. ### THE NEW 24GB PHYSICAL LIMITS (Single RTX 4090): # 1. The "Repo Swallower" (Q4 KV Cache): - Context: 250,000 tokens (Up from 170k!) - Speed: 73.66 t/s decode | 1,608 t/s prefill - Peak VRAM: 23.8 GB # 2. The "High-Precision SWE" (Q8 KV Cache): - Context: 150,000 tokens (Up from 100k!) - Speed: 75.01 t/s decode | 1,667 t/s prefill - Peak VRAM: 23.9 GB # 3. The "Pristine Attention" (Unquantized FP16 KV): - Context: 90,000 tokens - Speed: 80.58 t/s decode | 1,699 t/s prefill - Peak VRAM: 23.92 GB ### HOW TO RUN THE 250K GOD STACK TODAY: (Requires PR #27342 + my Q2_K Hugging Face drafter) llama.cpp flags: ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q2_K.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 250000 -ngl 99 --parallel 1 --port 8080 -ctv q4_0 -ctk q4_0 We are pushing a quarter million tokens of context with speculative DFlash 2 decoding at 73 tokens/second on a single consumer gaming GPU. I dropped my custom 2 bit Hugging Face GGUF links, visual performance graphs, and the PR #27342 build instructions in the replies below. If you own a single RTX 3090 or 4090, it is officially time to cancel your API subscriptions and let local silicon eat the cloud. how much monthly API spend does an optimized 4090 rig like this actually replace for you?
0:14

50% more context unlocked for Qwen 3.8 27b Q4_K_XL dflash 2 on a single RTX 4090 (24 GB VRAM) I found a hidden VRAM tax in llama.cpp. By combining my custom 2 bit DFlash 2 drafter with one overlooked server flag, I just unlocked another +80,000 tokens of context. Qwen3.8-27B is now running a massive 250,000 context at 75 tokens/s on a single RTX 4090. Here is the secret: By default, `llama-server` reserves massive chunks of your VRAM to handle multiple concurrent users (batching). If you are running a single user session, you are bleeding memory for features you aren't using. By passing the `--parallel 1` flag, you force the engine to dedicate 100% of your 24GB VRAM buffer to a single user. When we combine the VRAM saved by our Q2_K 2-bit drafter with the VRAM saved by `--parallel 1`, the context ceilings absolutely explode: Note: all benchmarks carried out with a massive 28k prompt. Ubuntu 22. ### THE NEW 24GB PHYSICAL LIMITS (Single RTX 4090): # 1. The "Repo Swallower" (Q4 KV Cache): - Context: 250,000 tokens (Up from 170k!) - Speed: 73.66 t/s decode | 1,608 t/s prefill - Peak VRAM: 23.8 GB # 2. The "High-Precision SWE" (Q8 KV Cache): - Context: 150,000 tokens (Up from 100k!) - Speed: 75.01 t/s decode | 1,667 t/s prefill - Peak VRAM: 23.9 GB # 3. The "Pristine Attention" (Unquantized FP16 KV): - Context: 90,000 tokens - Speed: 80.58 t/s decode | 1,699 t/s prefill - Peak VRAM: 23.92 GB ### HOW TO RUN THE 250K GOD STACK TODAY: (Requires PR #27342 + my Q2_K Hugging Face drafter) llama.cpp flags: ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q2_K.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 250000 -ngl 99 --parallel 1 --port 8080 -ctv q4_0 -ctk q4_0 We are pushing a quarter million tokens of context with speculative DFlash 2 decoding at 73 tokens/second on a single consumer gaming GPU. I dropped my custom 2 bit Hugging Face GGUF links, visual performance graphs, and the PR #27342 build instructions in the replies below. If you own a single RTX 3090 or 4090, it is officially time to cancel your API subscriptions and let local silicon eat the cloud. how much monthly API spend does an optimized 4090 rig like this actually replace for you?

Alok

39,189 просмотров • 8 дней назад

my 8 GB VRAM gaming laptop is absolutely going to hate me for this. but I still did it. ran a 31b dense model (Gemma 4 31b Q4) with only 8 GB VRAM last week I ran Gemma 4 26B A4B a mixture of experts model on my RTX 4060 and hit 25–28 tokens/sec using llama.cpp's new MTP support. smooth. snappy. but MoE has a secret: it only activates 4B parameters per token despite having 26B total. that's why it flies. so the real question started haunting me. what if I throw a full, no tricks, every parameter fires on every token, 31B DENSE model at the same machine? # Hardware: GPU: NVIDIA RTX 4060, 8 GB VRAM RAM: 16 GB CPU: Intel Core i7 H Laptop. Gaming. Modest. The model: gemma-4-31B-it-qat-UD-Q4_K_XL.gguf (model's unsloth huggingface link in the comments) This is Google DeepMind's flagship dense model in the Gemma 4 family that can run on single consumer GPU. It packs a hybrid attention architecture, supports up to 256K context natively, and is QAT (Quantization Aware Training) optimized, meaning it retains far more quality than standard post training quants at the same bit depth. This is NOT the MoE. This is 31 BILLION dense parameters, every single one of them loaded. # the flags I used: -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -cnv --spec-type draft-mtp --spec-draft-model mtp-gemma-4-31B-it.gguf --spec-draft-n-max 8 --spec-draft-p-min 0.6 -c 6000 -v Multi Token Prediction (MTP) is still active here. Separate draft GGUF required, same as the 26B setup. # Results: → Decode: ~3 tokens/sec → Prefill: ~2 tokens/sec → Context: 6000 tokens → Hardware crying quietly in the corner: yes so is 3 tps actually usable? For real time back and forth chat? Not ideal. You're not having a fluid conversation at 3 tps. but slow ≠ useless. And this is where it gets genuinely interesting. think about how senior devs actually work in a real team. But when something is architectural, deeply complex, or needs serious reasoning? they walk down the hall and escalate to the senior. That's exactly the local AI agent architecture this unlocks: → Fast orchestrator model (Gemma 4 26B MoE at 25+ tps) handles routing, simple queries, tool calls, memory. The junior dev. → Gemma 4 31B dense is the senior, called only when the fast model genuinely hits a wall. Hard multi step reasoning. Complex code generation. Deep architectural decisions. The agentic loop stays fast. Only the hard hops touch the 31B. That's a legitimate production grade local AI architecture on a budget hardware. (requires 2 8gb gpus) other workflows where 3 tps is completely fine: - overnight batch jobs. summarize documents, extract structured data, review code. Fire it off. Sleep. wake up to results. - One shot deep reasoning - Silent code audit loops, you write and test, the 31B reviews diffs and flags issues in the background between your sprints - Any workflow where output quality > output speed A few weeks ago, nobody was running a 30B+ dense model on a single consumer GPU with 8 GB VRAM. At all. Now we're doing it on an Intel i7-H gaming laptop with a NVIDIA RTX 4060, thanks to llama.cpp + QAT quants + MTP speculative drafting. Google DeepMind said the Gemma 4 31B targets "consumer GPUs and workstations." They were not exaggerating. The hardware bar to run serious frontier class models locally keeps dropping. the tools are here. the models are here. you just have to be willing to abuse your laptop a little. what workflows would you actually run on a local 3 tps 31B dense model? genuinely curious. drop it below.
0:18

my 8 GB VRAM gaming laptop is absolutely going to hate me for this. but I still did it. ran a 31b dense model (Gemma 4 31b Q4) with only 8 GB VRAM last week I ran Gemma 4 26B A4B a mixture of experts model on my RTX 4060 and hit 25–28 tokens/sec using llama.cpp's new MTP support. smooth. snappy. but MoE has a secret: it only activates 4B parameters per token despite having 26B total. that's why it flies. so the real question started haunting me. what if I throw a full, no tricks, every parameter fires on every token, 31B DENSE model at the same machine? # Hardware: GPU: NVIDIA RTX 4060, 8 GB VRAM RAM: 16 GB CPU: Intel Core i7 H Laptop. Gaming. Modest. The model: gemma-4-31B-it-qat-UD-Q4_K_XL.gguf (model's unsloth huggingface link in the comments) This is Google DeepMind's flagship dense model in the Gemma 4 family that can run on single consumer GPU. It packs a hybrid attention architecture, supports up to 256K context natively, and is QAT (Quantization Aware Training) optimized, meaning it retains far more quality than standard post training quants at the same bit depth. This is NOT the MoE. This is 31 BILLION dense parameters, every single one of them loaded. # the flags I used: -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -cnv --spec-type draft-mtp --spec-draft-model mtp-gemma-4-31B-it.gguf --spec-draft-n-max 8 --spec-draft-p-min 0.6 -c 6000 -v Multi Token Prediction (MTP) is still active here. Separate draft GGUF required, same as the 26B setup. # Results: → Decode: ~3 tokens/sec → Prefill: ~2 tokens/sec → Context: 6000 tokens → Hardware crying quietly in the corner: yes so is 3 tps actually usable? For real time back and forth chat? Not ideal. You're not having a fluid conversation at 3 tps. but slow ≠ useless. And this is where it gets genuinely interesting. think about how senior devs actually work in a real team. But when something is architectural, deeply complex, or needs serious reasoning? they walk down the hall and escalate to the senior. That's exactly the local AI agent architecture this unlocks: → Fast orchestrator model (Gemma 4 26B MoE at 25+ tps) handles routing, simple queries, tool calls, memory. The junior dev. → Gemma 4 31B dense is the senior, called only when the fast model genuinely hits a wall. Hard multi step reasoning. Complex code generation. Deep architectural decisions. The agentic loop stays fast. Only the hard hops touch the 31B. That's a legitimate production grade local AI architecture on a budget hardware. (requires 2 8gb gpus) other workflows where 3 tps is completely fine: - overnight batch jobs. summarize documents, extract structured data, review code. Fire it off. Sleep. wake up to results. - One shot deep reasoning - Silent code audit loops, you write and test, the 31B reviews diffs and flags issues in the background between your sprints - Any workflow where output quality > output speed A few weeks ago, nobody was running a 30B+ dense model on a single consumer GPU with 8 GB VRAM. At all. Now we're doing it on an Intel i7-H gaming laptop with a NVIDIA RTX 4060, thanks to llama.cpp + QAT quants + MTP speculative drafting. Google DeepMind said the Gemma 4 31B targets "consumer GPUs and workstations." They were not exaggerating. The hardware bar to run serious frontier class models locally keeps dropping. the tools are here. the models are here. you just have to be willing to abuse your laptop a little. what workflows would you actually run on a local 3 tps 31B dense model? genuinely curious. drop it below.

Alok

63,689 просмотров • 2 месяцев назад

We’re excited to introduce ShinkaEvolve: An open-source framework that evolves programs for scientific discovery with unprecedented sample-efficiency. Blog: Code: Like AlphaEvolve and its variants, our framework leverages LLMs to find state-of-the-art solutions to complex problems, but using orders of magnitude fewer resources! Many evolutionary AI systems are powerful but act like brute-force engines, burning thousands of samples to find good solutions. This makes discovery slow and expensive. We took inspiration from the efficiency of nature. ‘Shinka’ (進化) is Japanese for evolution, and we designed our system to be just as resourceful. On the classic circle packing optimization problem, ShinkaEvolve discovered a new state-of-the-art solution using only 150 samples. This is a big leap in efficiency compared to previous methods that required thousands of evaluations. We applied ShinkaEvolve to a diverse set of hard problems with real-world applications: 1/ AIME Math Reasoning: It evolved sophisticated agentic scaffolds that significantly outperform strong baselines, discovering an entire Pareto frontier of solutions trading performance for efficiency. 2/ Competitive Programming: On ALE-Bench (a benchmark for NP-Hard optimization problems), ShinkaEvolve took the best existing agent's solutions and improved them, turning a 5th place solution on one task into a 2nd place leaderboard rank in a competitive programming competition. 3/ LLM Training: We even turned ShinkaEvolve inward to improve LLMs themselves. It tackled the open challenge of designing load balancing losses for Mixture-of-Experts (MoE) models. It discovered a novel loss function that leads to better expert specialization and consistently improves model performance and perplexity. ShinkaEvolve achieves its remarkable sample-efficiency through three key innovations that work together: (1) an adaptive parent sampling strategy to balance exploration and exploitation, (2) novelty-based rejection filtering to avoid redundant work, and (3) a bandit-based LLM ensemble that dynamically picks the best model for the job. By making ShinkaEvolve open-source and highly sample-efficient, our goal is to democratize access to advanced, open-ended discovery tools. Our vision for ShinkaEvolve is to be an easy-to-use companion tool to help scientists and engineers with their daily work. We believe that building more efficient, nature-inspired systems is key to unlocking the future of AI-driven scientific research. We are excited to see what the community builds with it! Learn more in our technical report:
0:11

We’re excited to introduce ShinkaEvolve: An open-source framework that evolves programs for scientific discovery with unprecedented sample-efficiency. Blog: Code: Like AlphaEvolve and its variants, our framework leverages LLMs to find state-of-the-art solutions to complex problems, but using orders of magnitude fewer resources! Many evolutionary AI systems are powerful but act like brute-force engines, burning thousands of samples to find good solutions. This makes discovery slow and expensive. We took inspiration from the efficiency of nature. ‘Shinka’ (進化) is Japanese for evolution, and we designed our system to be just as resourceful. On the classic circle packing optimization problem, ShinkaEvolve discovered a new state-of-the-art solution using only 150 samples. This is a big leap in efficiency compared to previous methods that required thousands of evaluations. We applied ShinkaEvolve to a diverse set of hard problems with real-world applications: 1/ AIME Math Reasoning: It evolved sophisticated agentic scaffolds that significantly outperform strong baselines, discovering an entire Pareto frontier of solutions trading performance for efficiency. 2/ Competitive Programming: On ALE-Bench (a benchmark for NP-Hard optimization problems), ShinkaEvolve took the best existing agent's solutions and improved them, turning a 5th place solution on one task into a 2nd place leaderboard rank in a competitive programming competition. 3/ LLM Training: We even turned ShinkaEvolve inward to improve LLMs themselves. It tackled the open challenge of designing load balancing losses for Mixture-of-Experts (MoE) models. It discovered a novel loss function that leads to better expert specialization and consistently improves model performance and perplexity. ShinkaEvolve achieves its remarkable sample-efficiency through three key innovations that work together: (1) an adaptive parent sampling strategy to balance exploration and exploitation, (2) novelty-based rejection filtering to avoid redundant work, and (3) a bandit-based LLM ensemble that dynamically picks the best model for the job. By making ShinkaEvolve open-source and highly sample-efficient, our goal is to democratize access to advanced, open-ended discovery tools. Our vision for ShinkaEvolve is to be an easy-to-use companion tool to help scientists and engineers with their daily work. We believe that building more efficient, nature-inspired systems is key to unlocking the future of AI-driven scientific research. We are excited to see what the community builds with it! Learn more in our technical report:

Sakana AI

360,318 просмотров • 11 месяцев назад

Hills I will die on as someone who has coached high school football for over 29 years: 1. If you are not PASSIONATE about blessing, serving, and empowering those you are blessed to coach, this profession is not for you. 2. As much as we need to know our trade, getting to know (and to love), our players is far more important. 3. This is an INTENSE game, and it’ll never be “just a game”, but it IS a game. Remember that when you’re with your team, and more importantly, remember that when you’re with your family. 4. Just as we teach our athletes to “leave things better than they found them”, we need to leave our athletes better than they were when they first entered into our program. Never let a day pass without pouring into each and every individual. 5. Life is complicated enough, let’s not complicate the game in such a way that we take the joy of it away from others. In other words… Keep it simple. 6. Our words carry little (or NO), value, if we don’t practice what we preach. WE as coaches should be learning and growing each and every day, just as we expect our athletes to. 7. As much as we all want to win those championship rings for our athletes, make sure you don’t lose your wedding ring in the process. 8. The athlete that may be “difficult to reach/teach” (the one who may get on your last nerve more than you could ever imagine), is someone’s EVERYTHING. Get to know them as human beings, find out what motivates them, and do everything you can to help them to thrive. 9. Be where your feet are. Don’t fall into the trap of chasing logos and thinking that a higher division, a bigger school, or going from HS to college, or even college to the pros, is going to be more rewarding or fulfilling. 10. The legacy you leave as a coach will never be determined by your wins and losses, but by the lives you were able to change for the better!
0:15

Hills I will die on as someone who has coached high school football for over 29 years: 1. If you are not PASSIONATE about blessing, serving, and empowering those you are blessed to coach, this profession is not for you. 2. As much as we need to know our trade, getting to know (and to love), our players is far more important. 3. This is an INTENSE game, and it’ll never be “just a game”, but it IS a game. Remember that when you’re with your team, and more importantly, remember that when you’re with your family. 4. Just as we teach our athletes to “leave things better than they found them”, we need to leave our athletes better than they were when they first entered into our program. Never let a day pass without pouring into each and every individual. 5. Life is complicated enough, let’s not complicate the game in such a way that we take the joy of it away from others. In other words… Keep it simple. 6. Our words carry little (or NO), value, if we don’t practice what we preach. WE as coaches should be learning and growing each and every day, just as we expect our athletes to. 7. As much as we all want to win those championship rings for our athletes, make sure you don’t lose your wedding ring in the process. 8. The athlete that may be “difficult to reach/teach” (the one who may get on your last nerve more than you could ever imagine), is someone’s EVERYTHING. Get to know them as human beings, find out what motivates them, and do everything you can to help them to thrive. 9. Be where your feet are. Don’t fall into the trap of chasing logos and thinking that a higher division, a bigger school, or going from HS to college, or even college to the pros, is going to be more rewarding or fulfilling. 10. The legacy you leave as a coach will never be determined by your wins and losses, but by the lives you were able to change for the better!

Coach Hines 🇺🇸

63,241 просмотров • 2 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!
0:18

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

21,423 просмотров • 11 дней назад

8 free Polymarket Trading Bots on GitHub (from Beginner Friendly to Advanced Level). Each of these repos comes with a detailed step by step setup and usage guide in English. > Beginner Level - 5 min setup 1. This bot includes 120 ready to use strategies and tools for trading on prediction markets (Binance-Polymarket latency, Smart Routing, Penny Clipper, Momentum, DCA bots, Expiry Fade and more). It was built by a Cambridge computer science student who won a hackathon with this bot. GitHub: 2. A trading bot with a Smart Money strategy - it finds top traders in selected markets, filters them by Pnl, win rate, stable performance and then creates a list for automated copy trading. GitHub: 3. This is a bot toolkit that includes Polymarket - Kalshi arbitrage, whale alerts, market making, spread farming, sports trading and more. GitHub: 4. A weather trading bot from Chinese dev that analyzes different sources in real time, like forecasts, airport data and aviation observations (METAR + SPECI) to get the latest temperature data and generate a detailed weather report for a specific city and day. GitHub: 5. A huge collection of 30+ free trading bots and services for prediction markets. GitHub: > Advanced bot setup 1. This bot analyzes the real trading behavior of any Polymarket trader. It finds repeated patterns in his trades, shows which strategies he uses and helps you understand how to adapt them to your own trading. GitHub: 2. A bot that automatically manages all your limit orders on Polymarket to maximize liquidity rewards. GitHub: > A full ML weather model 1. A machine learning weather model that learns from weather forecasting errors. Instead of blindly trusting forecasts, it analyzes how different weather sources have historically overestimated or underestimated temperature values in specific cities and conditions. Then it automatically adjusts new forecasts to produce more accurate predictions. GitHub: All of these bots also support Dry Run mode, so you can test them on real markets without risking any funds.
0:16

8 free Polymarket Trading Bots on GitHub (from Beginner Friendly to Advanced Level). Each of these repos comes with a detailed step by step setup and usage guide in English. > Beginner Level - 5 min setup 1. This bot includes 120 ready to use strategies and tools for trading on prediction markets (Binance-Polymarket latency, Smart Routing, Penny Clipper, Momentum, DCA bots, Expiry Fade and more). It was built by a Cambridge computer science student who won a hackathon with this bot. GitHub: 2. A trading bot with a Smart Money strategy - it finds top traders in selected markets, filters them by Pnl, win rate, stable performance and then creates a list for automated copy trading. GitHub: 3. This is a bot toolkit that includes Polymarket - Kalshi arbitrage, whale alerts, market making, spread farming, sports trading and more. GitHub: 4. A weather trading bot from Chinese dev that analyzes different sources in real time, like forecasts, airport data and aviation observations (METAR + SPECI) to get the latest temperature data and generate a detailed weather report for a specific city and day. GitHub: 5. A huge collection of 30+ free trading bots and services for prediction markets. GitHub: > Advanced bot setup 1. This bot analyzes the real trading behavior of any Polymarket trader. It finds repeated patterns in his trades, shows which strategies he uses and helps you understand how to adapt them to your own trading. GitHub: 2. A bot that automatically manages all your limit orders on Polymarket to maximize liquidity rewards. GitHub: > A full ML weather model 1. A machine learning weather model that learns from weather forecasting errors. Instead of blindly trusting forecasts, it analyzes how different weather sources have historically overestimated or underestimated temperature values in specific cities and conditions. Then it automatically adjusts new forecasts to produce more accurate predictions. GitHub: All of these bots also support Dry Run mode, so you can test them on real markets without risking any funds.

Recogard

57,291 просмотров • 1 месяц назад

The architecture of this new world model is one of the most interesting things I've seen lately: Let me first explain how most world models work: They predict and render one frame at a time. If you are navigating in one of these worlds, and you look left, the model draws whatever looks right in the moment. Every time you change your viewpoint, the model has to imagine what should be there again, so it's very common for these models to "forget" what's in the world. For example, if you put a toy on the table, look away, then look back, the toy might not be there anymore. Tripo AI is releasing its Project Eden model, which works very differently: The model builds the world first, and then renders it based on that map. That map holds the real state of the world: the geometry, every object, where things are, what's already happened. The picture you see on screen gets generated from the map. This architecture flips the whole thing. Now, you get the following: 1. The world stops forgetting. Leave, come back, and the toy is still on the table because it lives in the map, not in the last frame you saw. 2. You can edit the world, and those changes persist for anyone who enters later. 3. Multiple people and AI agents can coexist in the world and see it from different perspectives. This is early research, but it's looking really promising. They just raised nearly $200M across two rounds to build it out. Tripo will be at SIGGRAPH 2026 (July 19–23, Los Angeles Convention Center). If you work in 3D, embodied AI, simulation, or anything spatial, go connect with them there.
0:10

The architecture of this new world model is one of the most interesting things I've seen lately: Let me first explain how most world models work: They predict and render one frame at a time. If you are navigating in one of these worlds, and you look left, the model draws whatever looks right in the moment. Every time you change your viewpoint, the model has to imagine what should be there again, so it's very common for these models to "forget" what's in the world. For example, if you put a toy on the table, look away, then look back, the toy might not be there anymore. Tripo AI is releasing its Project Eden model, which works very differently: The model builds the world first, and then renders it based on that map. That map holds the real state of the world: the geometry, every object, where things are, what's already happened. The picture you see on screen gets generated from the map. This architecture flips the whole thing. Now, you get the following: 1. The world stops forgetting. Leave, come back, and the toy is still on the table because it lives in the map, not in the last frame you saw. 2. You can edit the world, and those changes persist for anyone who enters later. 3. Multiple people and AI agents can coexist in the world and see it from different perspectives. This is early research, but it's looking really promising. They just raised nearly $200M across two rounds to build it out. Tripo will be at SIGGRAPH 2026 (July 19–23, Los Angeles Convention Center). If you work in 3D, embodied AI, simulation, or anything spatial, go connect with them there.

Santiago

30,244 просмотров • 2 месяцев назад

Batch Normalization by hand ✍️ ~ 7 steps walkthrough below Batch normalization is common practice for improving training and achieving faster convergence. It sounds simple. But it is often misunderstood. 🤔 Does batch normalization involve trainable parameters, tunable hyper-parameters, or both? 🤔 Is batch normalization applied to inputs, features, weights, biases, or outputs? 🤔 How is batch normalization different from layer normalization? So I drew and calculated one entirely by hand. Goal: normalize a mini-batch of 4 examples to mean 0 and variance 1, then let the network scale it back. = 1. Given = A mini-batch of 4 training examples, each with 3 features. = 2. Linear layer = Let us multiply by the weights and add the biases. Batch norm sits after this, which answers the second question: what gets normalized is features, not inputs, weights or biases. = 3. ReLU = We apply the activation, and -2 becomes 0. Negative values are suppressed before any statistic is taken. = 4. Batch statistics = Let us compute the sum, mean, variance and standard deviation, one row at a time. A row is a feature and the four columns are the four examples, so every number here measures one feature against the rest of the batch. That is the "batch" in batch normalization, and it is exactly what layer normalization does not do. The statistics are rounded to whole numbers, which is what keeps the rest of the page doable in pen. = 5. Shift to mean 0 = We subtract the mean, in green. The four values in each feature now average to zero. = 6. Scale to variance 1 = Let us divide by the standard deviation, in orange. Each feature now has variance one, whatever scale it arrived at. = 7. Scale and shift = We multiply by a linear transformation and pass the result on. The diagonal and the last column are trainable, so having just forced every feature to mean 0 and variance 1, we hand the network the means to undo it. The outputs: Mean of each feature = [2, 1, 2] Std dev of each feature = [1, 1, 2] To the next layer = [2, -2, 2, 0], [-3, 3, 6, -3], [2, 0, 1, 2] The answers: 🤔 Both. The scale and shift are trainable, the statistics are not. Epsilon and the momentum on the running statistics are the hyper-parameters, and one mini-batch by hand needs neither. 🤔 Features, after the linear layer, not inputs, weights or biases. 🤔 Batch norm measures across the batch, one feature at a time. Layer norm measures across the features, one example at a time. 💾 Save this post!
0:17

Batch Normalization by hand ✍️ ~ 7 steps walkthrough below Batch normalization is common practice for improving training and achieving faster convergence. It sounds simple. But it is often misunderstood. 🤔 Does batch normalization involve trainable parameters, tunable hyper-parameters, or both? 🤔 Is batch normalization applied to inputs, features, weights, biases, or outputs? 🤔 How is batch normalization different from layer normalization? So I drew and calculated one entirely by hand. Goal: normalize a mini-batch of 4 examples to mean 0 and variance 1, then let the network scale it back. = 1. Given = A mini-batch of 4 training examples, each with 3 features. = 2. Linear layer = Let us multiply by the weights and add the biases. Batch norm sits after this, which answers the second question: what gets normalized is features, not inputs, weights or biases. = 3. ReLU = We apply the activation, and -2 becomes 0. Negative values are suppressed before any statistic is taken. = 4. Batch statistics = Let us compute the sum, mean, variance and standard deviation, one row at a time. A row is a feature and the four columns are the four examples, so every number here measures one feature against the rest of the batch. That is the "batch" in batch normalization, and it is exactly what layer normalization does not do. The statistics are rounded to whole numbers, which is what keeps the rest of the page doable in pen. = 5. Shift to mean 0 = We subtract the mean, in green. The four values in each feature now average to zero. = 6. Scale to variance 1 = Let us divide by the standard deviation, in orange. Each feature now has variance one, whatever scale it arrived at. = 7. Scale and shift = We multiply by a linear transformation and pass the result on. The diagonal and the last column are trainable, so having just forced every feature to mean 0 and variance 1, we hand the network the means to undo it. The outputs: Mean of each feature = [2, 1, 2] Std dev of each feature = [1, 1, 2] To the next layer = [2, -2, 2, 0], [-3, 3, 6, -3], [2, 0, 1, 2] The answers: 🤔 Both. The scale and shift are trainable, the statistics are not. Epsilon and the momentum on the running statistics are the hyper-parameters, and one mini-batch by hand needs neither. 🤔 Features, after the linear layer, not inputs, weights or biases. 🤔 Batch norm measures across the batch, one feature at a time. Layer norm measures across the features, one example at a time. 💾 Save this post!

Tom Yeh

20,848 просмотров • 1 месяц назад

Hills I will die on as an elementary school teacher, who just wrapped up my 32nd year teaching! 1. If you are not PASSIONATE about blessing, serving, and empowering those you are blessed to teach, this profession is not for you. 2. Our students are not “ours”. They are their families, and we need to understand the magnitude of the calling, and responsibility we have to meet them where they are, and to help them to get to a place that they never thought possible. 3. As important as the curriculum is (and it IS important), the children in our classroom are what matter most. It’s our JOB to teach the curriculum to SERVE our students, NOT to use our students to push any sort of agenda. 4. Just as we teach our students to “leave things better than they found them”, we need to leave our students better than they were when they first entered into our classroom. Never let a day pass without pouring into each and every child. 5. We should not teach our children WHAT to think, but HOW to think, and how to use that knowledge to bless and serve not only themselves, but the world around them. 6. Our words carry little (or NO value), if we do not practice what we preach. 7. If we don’t make learning fun, children will view learning as a chore, and we we will be creating a generation of children who grow up to be young adults who don’t see the joy in learning new things. 8. The child that may be “difficult to reach/teach” (the one who may get on your last nerve more than you could ever imagine), is someone’s EVERYTHING. Get to know them as human beings, find out what motivates them, and do everything you can to help them to thrive. 9. Never tell a child they “can’t” do something. God has blessed each of them with far more strengths and talents than we may know, and it’s not our job to tell them what they can’t do, but to help them to realize all the things they CAN do. 10. The legacy you leave as a teacher will never be determined by your student’s test scores, but by the human beings you helped them become throughout their lives.
0:15

Hills I will die on as an elementary school teacher, who just wrapped up my 32nd year teaching! 1. If you are not PASSIONATE about blessing, serving, and empowering those you are blessed to teach, this profession is not for you. 2. Our students are not “ours”. They are their families, and we need to understand the magnitude of the calling, and responsibility we have to meet them where they are, and to help them to get to a place that they never thought possible. 3. As important as the curriculum is (and it IS important), the children in our classroom are what matter most. It’s our JOB to teach the curriculum to SERVE our students, NOT to use our students to push any sort of agenda. 4. Just as we teach our students to “leave things better than they found them”, we need to leave our students better than they were when they first entered into our classroom. Never let a day pass without pouring into each and every child. 5. We should not teach our children WHAT to think, but HOW to think, and how to use that knowledge to bless and serve not only themselves, but the world around them. 6. Our words carry little (or NO value), if we do not practice what we preach. 7. If we don’t make learning fun, children will view learning as a chore, and we we will be creating a generation of children who grow up to be young adults who don’t see the joy in learning new things. 8. The child that may be “difficult to reach/teach” (the one who may get on your last nerve more than you could ever imagine), is someone’s EVERYTHING. Get to know them as human beings, find out what motivates them, and do everything you can to help them to thrive. 9. Never tell a child they “can’t” do something. God has blessed each of them with far more strengths and talents than we may know, and it’s not our job to tell them what they can’t do, but to help them to realize all the things they CAN do. 10. The legacy you leave as a teacher will never be determined by your student’s test scores, but by the human beings you helped them become throughout their lives.

Coach Hines 🇺🇸

62,267 просмотров • 3 месяцев назад

I just ran Gemma 4 31B on @CerebrasSystems at 1,800+ tokens/sec and it's multimodal. For context: that's 35x faster than a typical GPU endpoint, and the first token (reasoning included) lands in 1.5 seconds. This isn't a benchmark slide, I recorded the inference live. Prompt I used: "Create a simulation of an iPhone. Include at least one working dummy note taking app, a functional notification pulldown, high quality graphics, single HTML file, any libs via CDN." - Generation time: 3 seconds. - Notes app worked. - Notification panel worked. - Rendered first try. This is what wafer-scale inference unlocks, not just "faster," but a different category of product. When generation is this fast, you stop waiting and start iterating in real time. Why this matters: Gemma 4 31B is Google DeepMind's flagship open weight model, Apache 2.0 licensed, dense (not MoE), and built for efficiency over raw parameter count. It scores close to Claude Haiku 4.5 on the Artificial Analysis Intelligence Index (30 vs 29) but runs ~18x faster on Cerebras. It's also the first multimodal model on Cerebras's platform, meaning you can now feed it screenshots, documents, charts, and UI states at wafer scale speed. # Applications I'm most excited about: - Screenshot → Insight: Drop in a dashboard or document screenshot, get structured findings back instantly. no waiting, no batching. - Live UI generation: Full interactive interfaces (like my iPhone sim) generated and rendered in under 2 seconds. - Screenshot -> Patch: Feed it a broken UI + console error, get a minimal code fix and verification steps back. - Computer use & agentic loops: See -> reason -> act - verify, fast enough to keep a human in the loop instead of waiting on the model. - Long context summarization: Full research reports condensed into decision ready summaries you can read and requery in one sitting. The bigger unlock isn't the speed number itself, it's that agentic and multimodal loops (see -> reason -> output -> tool call -> verify -> retry) finally run in real time instead of feeling sluggish. As Logan Kilpatrick (Logan Kilpatrick) put it: "If every model was doing 2,000 tokens per second, you wouldn't build the same product and just have it be faster, you'd build different products." Gemma 4 31B is live now on Cerebras Inference Cloud in public preview. If you're building multimodal, agentic, or real time apps, this is worth testing today. What would you build with such insane inference throughput?
0:12

I just ran Gemma 4 31B on @CerebrasSystems at 1,800+ tokens/sec and it's multimodal. For context: that's 35x faster than a typical GPU endpoint, and the first token (reasoning included) lands in 1.5 seconds. This isn't a benchmark slide, I recorded the inference live. Prompt I used: "Create a simulation of an iPhone. Include at least one working dummy note taking app, a functional notification pulldown, high quality graphics, single HTML file, any libs via CDN." - Generation time: 3 seconds. - Notes app worked. - Notification panel worked. - Rendered first try. This is what wafer-scale inference unlocks, not just "faster," but a different category of product. When generation is this fast, you stop waiting and start iterating in real time. Why this matters: Gemma 4 31B is Google DeepMind's flagship open weight model, Apache 2.0 licensed, dense (not MoE), and built for efficiency over raw parameter count. It scores close to Claude Haiku 4.5 on the Artificial Analysis Intelligence Index (30 vs 29) but runs ~18x faster on Cerebras. It's also the first multimodal model on Cerebras's platform, meaning you can now feed it screenshots, documents, charts, and UI states at wafer scale speed. # Applications I'm most excited about: - Screenshot → Insight: Drop in a dashboard or document screenshot, get structured findings back instantly. no waiting, no batching. - Live UI generation: Full interactive interfaces (like my iPhone sim) generated and rendered in under 2 seconds. - Screenshot -> Patch: Feed it a broken UI + console error, get a minimal code fix and verification steps back. - Computer use & agentic loops: See -> reason -> act - verify, fast enough to keep a human in the loop instead of waiting on the model. - Long context summarization: Full research reports condensed into decision ready summaries you can read and requery in one sitting. The bigger unlock isn't the speed number itself, it's that agentic and multimodal loops (see -> reason -> output -> tool call -> verify -> retry) finally run in real time instead of feeling sluggish. As Logan Kilpatrick (Logan Kilpatrick) put it: "If every model was doing 2,000 tokens per second, you wouldn't build the same product and just have it be faster, you'd build different products." Gemma 4 31B is live now on Cerebras Inference Cloud in public preview. If you're building multimodal, agentic, or real time apps, this is worth testing today. What would you build with such insane inference throughput?

Alok

12,962 просмотров • 2 месяцев назад