Загрузка видео...

Не удалось загрузить видео

На главную

SGLang now supports DSpark, enabling confidence-driven, variable-length verification for speculative decoding 🎉 DSpark addresses a key bottleneck under load: instead of verifying every draft token, it verifies only where the draft model is confident, so the gains hold even as batch size scales. We heavily optimized variable-length verification in...

lmsysorg's profile picture

LMSYS Org

16,466 subscribers

168,185 просмотров • 24 дней назад •via X (Twitter)

Наука и технологииОбразование
LIVE

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

HERMES DESKTOP APP JUST GOT A NEW ULTRA MODE FOR DEEPSEEK V4 PRO – AND IT COSTS ABOUT $1 A DAY TO RUN If you're using Hermes desktop app, you're probably already running DeepSeek – it's one of the top models available. The update: → DeepSeek V4 Pro now has thinking mode, fast mode, and a brand new ultra mode → There's also an extra high mode, but the real upgrade is ultra The actual cost: roughly $1/day for a normal workflow. That's about $20-25/month – basically the same as before, just with a significantly faster mode unlocked. The verdict after a full day of testing: it flies through tool calling. Noticeably faster, same price range. If you're already paying for DeepSeek through Hermes, this is a free upgrade sitting in your settings right now. Update Hermes, grab the DeepSeek API if you don't have it yet, and try V4 Pro on ultra. Bookmark this, you'll want to come back to it later.
1:00

HERMES DESKTOP APP JUST GOT A NEW ULTRA MODE FOR DEEPSEEK V4 PRO – AND IT COSTS ABOUT $1 A DAY TO RUN If you're using Hermes desktop app, you're probably already running DeepSeek – it's one of the top models available. The update: → DeepSeek V4 Pro now has thinking mode, fast mode, and a brand new ultra mode → There's also an extra high mode, but the real upgrade is ultra The actual cost: roughly $1/day for a normal workflow. That's about $20-25/month – basically the same as before, just with a significantly faster mode unlocked. The verdict after a full day of testing: it flies through tool calling. Noticeably faster, same price range. If you're already paying for DeepSeek through Hermes, this is a free upgrade sitting in your settings right now. Update Hermes, grab the DeepSeek API if you don't have it yet, and try V4 Pro on ultra. Bookmark this, you'll want to come back to it later.

SCOTTY BEAM

24,032 просмотров • 10 дней назад

Serving GLM5.2 NVFP4 Agentic Workload with SGLang: How We Reached 500 TPS on 8xB300 at bs=1 In this deep dive, we explain how SGLang reaches 500+ tok/s/user at bs=1 on 8xB300, with 18 to 34% higher single-user interactivity within two weeks since day-0, and 6 to 11% better peak throughput at high concurrency, benchmarked on a real multi-turn agentic coding workload. Our new TopK-V2 kernel is 2.33x faster at 80K ISL, scaling to 10.17x at 1M ISL, keeping interactivity essentially flat out to 1M tokens. Part of the story is the architecture itself. GLM-5.2 applies IndexShare to its DSA layers and ships a stronger MTP head reusing IndexShare and KVShare. The rest comes from our serving optimizations. Special thanks to NVIDIA AI for the help in day-0 support of GLM-5.2 NVFP4, and to Z.ai for IndexShare in SGLang!
0:47

Serving GLM5.2 NVFP4 Agentic Workload with SGLang: How We Reached 500 TPS on 8xB300 at bs=1 In this deep dive, we explain how SGLang reaches 500+ tok/s/user at bs=1 on 8xB300, with 18 to 34% higher single-user interactivity within two weeks since day-0, and 6 to 11% better peak throughput at high concurrency, benchmarked on a real multi-turn agentic coding workload. Our new TopK-V2 kernel is 2.33x faster at 80K ISL, scaling to 10.17x at 1M ISL, keeping interactivity essentially flat out to 1M tokens. Part of the story is the architecture itself. GLM-5.2 applies IndexShare to its DSA layers and ships a stronger MTP head reusing IndexShare and KVShare. The rest comes from our serving optimizations. Special thanks to NVIDIA AI for the help in day-0 support of GLM-5.2 NVFP4, and to Z.ai for IndexShare in SGLang!

LMSYS Org

105,870 просмотров • 16 дней назад

I have been testing DeepSeek-V4-Pro with the Pi coding agent. I am mindblown by how well it works out of the box. A few notes: I spent a few hours building an LLM wiki with an agent powered entirely by DeepSeek-V4-Pro on Fireworks AI inference. This is the first time I feel like there is an open-weight model that can reason at the level of Claude and Codex. And it does this in a cost-effective way with support for 1M context length. To be clear, I am using DeepSeek-V4-Pro inside of Pi without any special configuration. It works out of the box. It's exciting that there is a model that can just be plugged into a basic harness like Pi, and it just works. I've never seen that before. Most models require lots of configuration and setup. DeepSeek's DeepSeek-V4-Pro is clearly good at agentic coding (probably the best from the open-weight models), but the model is also great on knowledge-intensive tasks where reasoning matters. The agent pulled agentic engineering best practices from different company docs (Anthropic, OpenAI, Google, Stripe, Meta, Modal, DeepSeek, Mistral, Cohere), searched and digested Reddit and HN threads, summarized arxiv papers, and surfaced trending GitHub repos. Then it distilled everything into actionable tips across categories. I love the Wiki it built. The quality is really good. Here is a snapshot of what the wiki looks like: DeepSeek-V4-Pro handled the task without breaking stride. Multi-step research queries, code generation for scaffolding, context-heavy reasoning across disparate sources. For coding specifically, this is the first open-weight model that genuinely feels like a Codex or Claude Code experience. It compares in capability and actual multi-turn agentic work. What made the loop feel so responsive was Fireworks' inference speed (the fastest in the market) and the fact that they actually validate models at the systems level before shipping. No corrupted reasoning traces. Just fast, reliable iteration. The hybrid CSA and HCA attention design cuts KV cache to just 10% and inference FLOPs by nearly 4x at 1M-token context. This is what makes the agent loop actually fast and cheap enough to run in practice. For devs who've been watching open-weight models close the gap but haven't found one that actually delivers in practice, this is the closest I've seen. Try it here:
0:31

I have been testing DeepSeek-V4-Pro with the Pi coding agent. I am mindblown by how well it works out of the box. A few notes: I spent a few hours building an LLM wiki with an agent powered entirely by DeepSeek-V4-Pro on Fireworks AI inference. This is the first time I feel like there is an open-weight model that can reason at the level of Claude and Codex. And it does this in a cost-effective way with support for 1M context length. To be clear, I am using DeepSeek-V4-Pro inside of Pi without any special configuration. It works out of the box. It's exciting that there is a model that can just be plugged into a basic harness like Pi, and it just works. I've never seen that before. Most models require lots of configuration and setup. DeepSeek's DeepSeek-V4-Pro is clearly good at agentic coding (probably the best from the open-weight models), but the model is also great on knowledge-intensive tasks where reasoning matters. The agent pulled agentic engineering best practices from different company docs (Anthropic, OpenAI, Google, Stripe, Meta, Modal, DeepSeek, Mistral, Cohere), searched and digested Reddit and HN threads, summarized arxiv papers, and surfaced trending GitHub repos. Then it distilled everything into actionable tips across categories. I love the Wiki it built. The quality is really good. Here is a snapshot of what the wiki looks like: DeepSeek-V4-Pro handled the task without breaking stride. Multi-step research queries, code generation for scaffolding, context-heavy reasoning across disparate sources. For coding specifically, this is the first open-weight model that genuinely feels like a Codex or Claude Code experience. It compares in capability and actual multi-turn agentic work. What made the loop feel so responsive was Fireworks' inference speed (the fastest in the market) and the fact that they actually validate models at the systems level before shipping. No corrupted reasoning traces. Just fast, reliable iteration. The hybrid CSA and HCA attention design cuts KV cache to just 10% and inference FLOPs by nearly 4x at 1M-token context. This is what makes the agent loop actually fast and cheap enough to run in practice. For devs who've been watching open-weight models close the gap but haven't found one that actually delivers in practice, this is the closest I've seen. Try it here:

elvis

59,803 просмотров • 3 месяцев назад

Inkling, Thinking Machines' first open model, dropped today: 975B total / 41B active MoE, up to 1M context, reasoning natively over text, images, and audio. Serving and RL support are already live: you can run and shape it on an open stack, starting now. Day 0 support on SGLang SGLang and Miles RadixArk👇 - Inkling's new architecture (ShortConv, attention with relative positional embedding, shared expert sink MoE) is natively implemented and deeply optimized, with prefill full CUDA graph and MXFP8 KV cache - Full parameter and LoRA RL in a customized Megatron backend, train inference consistency via customized kernels, routing replay, and cross-runtime parameter synchronization - DFlash speculative decoding from Modal for low-latency serving Launching now, blog and cookbook in the comments ⬇️
1:01

Inkling, Thinking Machines' first open model, dropped today: 975B total / 41B active MoE, up to 1M context, reasoning natively over text, images, and audio. Serving and RL support are already live: you can run and shape it on an open stack, starting now. Day 0 support on SGLang SGLang and Miles RadixArk👇 - Inkling's new architecture (ShortConv, attention with relative positional embedding, shared expert sink MoE) is natively implemented and deeply optimized, with prefill full CUDA graph and MXFP8 KV cache - Full parameter and LoRA RL in a customized Megatron backend, train inference consistency via customized kernels, routing replay, and cross-runtime parameter synchronization - DFlash speculative decoding from Modal for low-latency serving Launching now, blog and cookbook in the comments ⬇️

LMSYS Org

143,172 просмотров • 15 дней назад

hy3 vs mimo-v2.5 vs deepseek v4 flash vs minimax m3 the four models on top of the openrouter leaderboard by tokens this week: #1 hy3 (Tencent Hy) – 7.5t #2 mimo-v2.5 (Xiaomi MiMo) – 6.56t #3 deepseek v4 flash (DeepSeek) – 5.24t #4 minimax m3 (MiniMax (official)) – 4.21t so we tested them. 3 prompts, single-file html, Three.js from a cdn, fully procedural, no external assets. all run via AI/ML API each prompt is a transparent cutaway machine that has to be mechanically correct, not decorative: • 4-stroke engine with full oil circulation – slider-crank kinematics, cam at 2:1, valve lift driven by lobes, oil loop from sump to gallery to big-end • watt walking-beam steam engine – four-bar vector-loop closure, eccentric-driven slide valve, steam events synced to real port position • francis reaction water turbine – 20 guide vanes on a regulating ring, 17 lofted runner blades, gpu particle advection, precessing vortex rope at part load the takeaway up front: none of the four cleared all three scenes on the first attempt. but the price spread between them is roughly 70x – hy3 fixed included costs less than two cents overall results (summed across all 3 scenes): cost #1 hy3 – $0.016 #2 deepseek v4 flash – $0.025 #3 mimo-v2.5 – $0.97 #4 minimax m3 – $1.17 tokens #1 hy3 – 19,326 #2 deepseek v4 flash – 63,126 #3 mimo-v2.5 – 322,523 #4 minimax m3 – 702,900 lines of code #1 hy3 – 1,047 #2 mimo-v2.5 – 2,759 #3 deepseek v4 flash – 3,273 #4 minimax m3 – 3,354 scenes needing a second attempt #1 hy3 – 1 (engine) #1 mimo-v2.5 – 1 (turbine) #1 minimax m3 – 1 (turbine) #4 deepseek v4 flash – 2 (steam engine, turbine) observations: 1. the token spread is the real story – minimax burns 36x hy3's tokens and lands in the same place, one retry, ~3.3k lines 2. hy3 is the outlier on density: 1,047 lines total, fewest tokens, cheapest run, and only one scene needed a second pass. deepseek is the opposite trade – near-hy3 pricing but the most retries 3. mimo and minimax seem to overthink instead of writing the code. minimax spent 359.1k tokens on the steam engine and produced 1,346 lines – the tokens are going somewhere other than the file 4. the francis turbine broke three of the four. the spec that separates them is the one with 20 linked guide vanes and gpu particle advection, not the one with the most parts overall impression: none of these models excelled at any of the tasks we gave them. but they were close, and they were extremely cheap. the gap that matters isn't quality anymore – it's that hy3 ran all three scenes for less than two cents while the frontier labs charge dollars for the same work right now you pick these because they're good for the zero price you pay. soon that's something openai and anthropic will have to think about follow thehype. for 24/7 ai news, analysis and breakdowns
0:55

hy3 vs mimo-v2.5 vs deepseek v4 flash vs minimax m3 the four models on top of the openrouter leaderboard by tokens this week: #1 hy3 (Tencent Hy) – 7.5t #2 mimo-v2.5 (Xiaomi MiMo) – 6.56t #3 deepseek v4 flash (DeepSeek) – 5.24t #4 minimax m3 (MiniMax (official)) – 4.21t so we tested them. 3 prompts, single-file html, Three.js from a cdn, fully procedural, no external assets. all run via AI/ML API each prompt is a transparent cutaway machine that has to be mechanically correct, not decorative: • 4-stroke engine with full oil circulation – slider-crank kinematics, cam at 2:1, valve lift driven by lobes, oil loop from sump to gallery to big-end • watt walking-beam steam engine – four-bar vector-loop closure, eccentric-driven slide valve, steam events synced to real port position • francis reaction water turbine – 20 guide vanes on a regulating ring, 17 lofted runner blades, gpu particle advection, precessing vortex rope at part load the takeaway up front: none of the four cleared all three scenes on the first attempt. but the price spread between them is roughly 70x – hy3 fixed included costs less than two cents overall results (summed across all 3 scenes): cost #1 hy3 – $0.016 #2 deepseek v4 flash – $0.025 #3 mimo-v2.5 – $0.97 #4 minimax m3 – $1.17 tokens #1 hy3 – 19,326 #2 deepseek v4 flash – 63,126 #3 mimo-v2.5 – 322,523 #4 minimax m3 – 702,900 lines of code #1 hy3 – 1,047 #2 mimo-v2.5 – 2,759 #3 deepseek v4 flash – 3,273 #4 minimax m3 – 3,354 scenes needing a second attempt #1 hy3 – 1 (engine) #1 mimo-v2.5 – 1 (turbine) #1 minimax m3 – 1 (turbine) #4 deepseek v4 flash – 2 (steam engine, turbine) observations: 1. the token spread is the real story – minimax burns 36x hy3's tokens and lands in the same place, one retry, ~3.3k lines 2. hy3 is the outlier on density: 1,047 lines total, fewest tokens, cheapest run, and only one scene needed a second pass. deepseek is the opposite trade – near-hy3 pricing but the most retries 3. mimo and minimax seem to overthink instead of writing the code. minimax spent 359.1k tokens on the steam engine and produced 1,346 lines – the tokens are going somewhere other than the file 4. the francis turbine broke three of the four. the spec that separates them is the one with 20 linked guide vanes and gpu particle advection, not the one with the most parts overall impression: none of these models excelled at any of the tasks we gave them. but they were close, and they were extremely cheap. the gap that matters isn't quality anymore – it's that hy3 ran all three scenes for less than two cents while the frontier labs charge dollars for the same work right now you pick these because they're good for the zero price you pay. soon that's something openai and anthropic will have to think about follow thehype. for 24/7 ai news, analysis and breakdowns

thehype.

17,145 просмотров • 16 дней назад

1/5 MiniCPM-V 4.6 (1.3B) is now live 🚀🚀 High-res visual processing, optimized for consumer-grade and mobile hardware. We’ve leveraged the latest LLaVA-UHD v4 technique to cut vision encoding costs by 55%, enabling native edge deployment with extreme efficiency. 🔥 Beats Gemma4-E2B-it and Qwen3.5-0.8B across key multimodal and Artificial Analysis benchmarks — scoring higher than Qwen3.5-0.8B using just 2.5% of its token budget. ⚡ TTFT (75.7ms) 2.2x Faster than Qwen3.5-0.8B even with 3136² high-res images. 🏗️ ~1.5x Token Throughput compared with Qwen3.5-0.8B on a single RTX 4090. Try the model here: 🤗 Hugging Face: 💻 GitHub: 🔭 Modelscope: 🌐 Web Demo: 📱 App Demo:
0:44

1/5 MiniCPM-V 4.6 (1.3B) is now live 🚀🚀 High-res visual processing, optimized for consumer-grade and mobile hardware. We’ve leveraged the latest LLaVA-UHD v4 technique to cut vision encoding costs by 55%, enabling native edge deployment with extreme efficiency. 🔥 Beats Gemma4-E2B-it and Qwen3.5-0.8B across key multimodal and Artificial Analysis benchmarks — scoring higher than Qwen3.5-0.8B using just 2.5% of its token budget. ⚡ TTFT (75.7ms) 2.2x Faster than Qwen3.5-0.8B even with 3136² high-res images. 🏗️ ~1.5x Token Throughput compared with Qwen3.5-0.8B on a single RTX 4090. Try the model here: 🤗 Hugging Face: 💻 GitHub: 🔭 Modelscope: 🌐 Web Demo: 📱 App Demo:

OpenBMB

352,532 просмотров • 2 месяцев назад

China just made Silicon Valley's entire AI industry look like a scam. The US government spent 3 years trying to stop China from building competitive AI. But this backfired HORRIBLY. Here's what happened: Yesterday, a Chinese startup called DeepSeek released a new AI model called V4. It matches the performance of OpenAI and Anthropic's best models. At 1/7th the price. And for the first time ever, it was built on Chinese chips. NOT American ones. That last part is the one that terrifies the west. For context: Since 2022, the US has banned the export of advanced AI chips to China. The entire strategy was built on the assumption that if China can't access Nvidia's best hardware, they can't build frontier AI. But DeepSeek just proved that assumption wrong. Their V4 model was trained and runs on Huawei's Ascend chips. Huawei spent months working directly with DeepSeek to make sure V4 runs across their entire line of AI processors. Jensen Huang even predicted this on a recent podcast: "The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation." That day was yesterday. And the numbers are crazy: DeepSeek V4 costs $3.48 per million output tokens. OpenAI's latest model GPT-5.5 costs $30. Anthropic's Claude charges $25. Same ballpark performance. 7x cheaper. Uber's CTO just admitted they burned through their ENTIRE 2026 AI budget in 4 months using Anthropic's tools. If Uber had used DeepSeek instead, that same budget would have lasted 7 YEARS. 4 months vs 7 years. Same work getting done. But the pricing isn't even the big thing here. The real story is what DeepSeek did with their technical report: They published the benchmarks where they LOSE. Every AI company cherry-picks the tests where their model wins. DeepSeek ran the full comparison against GPT-5.4 and Google's Gemini, found they trail frontier models by 3 to 6 months, and printed it anyway. They literally don't care because the price gap makes the performance gap irrelevant for 90% of use cases. So the US export controls didn't slow China down. They ACCELERATED China's independence. Because Chinese developers were FORCED to train models with limited resources, they had to figure out how to make AI radically more efficient. That constraint became their competitive advantage. Every generation of DeepSeek has gotten dramatically cheaper to train. V4 continues the trend. Meanwhile US companies are going the OPPOSITE direction: OpenAI's GPT-5.5 Pro costs $180 per million output tokens. That's 51x more expensive than DeepSeek V4 for comparable work. The Commerce Secretary confirmed this week that ZERO Nvidia advanced chip shipments have actually gone through to China despite being approved in January. So China built frontier AI anyway. Without American chips. At a fraction of the cost. And the market response tells you everything: Chinese chipmaker SMIC surged 10%. Huahong Semiconductor jumped 15%. DeepSeek's Chinese AI competitors Zhipu AI and MiniMax dropped 9% because V4 is destroying them too. DeepSeek is making Silicon Valley's pricing model look like a scam. US tech companies spent $650 billion on AI infrastructure this year. DeepSeek just showed the world you can match their output for pennies. The export controls were supposed to be America's ace card. Instead they taught China how to win without American chips, at American prices nobody can compete with. Jensen Huang was right. This is a horrible outcome. But it's the outcome America built for itself.
1:06

China just made Silicon Valley's entire AI industry look like a scam. The US government spent 3 years trying to stop China from building competitive AI. But this backfired HORRIBLY. Here's what happened: Yesterday, a Chinese startup called DeepSeek released a new AI model called V4. It matches the performance of OpenAI and Anthropic's best models. At 1/7th the price. And for the first time ever, it was built on Chinese chips. NOT American ones. That last part is the one that terrifies the west. For context: Since 2022, the US has banned the export of advanced AI chips to China. The entire strategy was built on the assumption that if China can't access Nvidia's best hardware, they can't build frontier AI. But DeepSeek just proved that assumption wrong. Their V4 model was trained and runs on Huawei's Ascend chips. Huawei spent months working directly with DeepSeek to make sure V4 runs across their entire line of AI processors. Jensen Huang even predicted this on a recent podcast: "The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation." That day was yesterday. And the numbers are crazy: DeepSeek V4 costs $3.48 per million output tokens. OpenAI's latest model GPT-5.5 costs $30. Anthropic's Claude charges $25. Same ballpark performance. 7x cheaper. Uber's CTO just admitted they burned through their ENTIRE 2026 AI budget in 4 months using Anthropic's tools. If Uber had used DeepSeek instead, that same budget would have lasted 7 YEARS. 4 months vs 7 years. Same work getting done. But the pricing isn't even the big thing here. The real story is what DeepSeek did with their technical report: They published the benchmarks where they LOSE. Every AI company cherry-picks the tests where their model wins. DeepSeek ran the full comparison against GPT-5.4 and Google's Gemini, found they trail frontier models by 3 to 6 months, and printed it anyway. They literally don't care because the price gap makes the performance gap irrelevant for 90% of use cases. So the US export controls didn't slow China down. They ACCELERATED China's independence. Because Chinese developers were FORCED to train models with limited resources, they had to figure out how to make AI radically more efficient. That constraint became their competitive advantage. Every generation of DeepSeek has gotten dramatically cheaper to train. V4 continues the trend. Meanwhile US companies are going the OPPOSITE direction: OpenAI's GPT-5.5 Pro costs $180 per million output tokens. That's 51x more expensive than DeepSeek V4 for comparable work. The Commerce Secretary confirmed this week that ZERO Nvidia advanced chip shipments have actually gone through to China despite being approved in January. So China built frontier AI anyway. Without American chips. At a fraction of the cost. And the market response tells you everything: Chinese chipmaker SMIC surged 10%. Huahong Semiconductor jumped 15%. DeepSeek's Chinese AI competitors Zhipu AI and MiniMax dropped 9% because V4 is destroying them too. DeepSeek is making Silicon Valley's pricing model look like a scam. US tech companies spent $650 billion on AI infrastructure this year. DeepSeek just showed the world you can match their output for pennies. The export controls were supposed to be America's ace card. Instead they taught China how to win without American chips, at American prices nobody can compete with. Jensen Huang was right. This is a horrible outcome. But it's the outcome America built for itself.

Ricardo

280,185 просмотров • 3 месяцев назад

I nearly 2x'd the speed while only using +1GB VRAM with the new MTP update in llama.cpp 🤯 You need to add these flags to start using it: --spec-type draft-mtp \ --spec-draft-p-min 0.75 \ --spec-draft-n-max 2 My results with Qwen3.6 27B on a single RTX 5080 ↓ ⚪️ no flag (without mtp) → 54.3 tok/s with 13.26GB VRAM 🔵 --spec-draft-n-max 2 → 90.7 tok/s with 14.29GB VRAM 🔴 --spec-draft-n-max 2 --spec-draft-p-min 0.75 → 93.9 tok/s with 14.30GB VRAM 🟢 --spec-draft-n-max 6 --spec-draft-p-min 0.75 → 93.9 tok/s with 14.87GB VRAM Increasing to 6 draft tokens didn't help my setup for some reason. I made sure to test with a low context length to have enough headroom and eliminate risk of vram stress. From my understanding: 1) The speed gains are very task-dependent. You need to test across a wide range of tasks to get a realistic idea of the benefits 2) We’re already running heavily quantized GGUF models (Q3, Q4, Q6, etc.), so we already benefit from strong speed/performance thanks to the reduced size. That’s why some people are seeing little to no improvement compared to MLX or other quantized versions The progress over the past few days has been insane to say the least. However, MTP now consumes significantly more VRAM. Personally 16GB just isn't enough to use MTP and run it with a good context size. Time to upgrade lads, 24GB+ users are eating GOOD today 🔥 Full setup below ↓
0:33

I nearly 2x'd the speed while only using +1GB VRAM with the new MTP update in llama.cpp 🤯 You need to add these flags to start using it: --spec-type draft-mtp \ --spec-draft-p-min 0.75 \ --spec-draft-n-max 2 My results with Qwen3.6 27B on a single RTX 5080 ↓ ⚪️ no flag (without mtp) → 54.3 tok/s with 13.26GB VRAM 🔵 --spec-draft-n-max 2 → 90.7 tok/s with 14.29GB VRAM 🔴 --spec-draft-n-max 2 --spec-draft-p-min 0.75 → 93.9 tok/s with 14.30GB VRAM 🟢 --spec-draft-n-max 6 --spec-draft-p-min 0.75 → 93.9 tok/s with 14.87GB VRAM Increasing to 6 draft tokens didn't help my setup for some reason. I made sure to test with a low context length to have enough headroom and eliminate risk of vram stress. From my understanding: 1) The speed gains are very task-dependent. You need to test across a wide range of tasks to get a realistic idea of the benefits 2) We’re already running heavily quantized GGUF models (Q3, Q4, Q6, etc.), so we already benefit from strong speed/performance thanks to the reduced size. That’s why some people are seeing little to no improvement compared to MLX or other quantized versions The progress over the past few days has been insane to say the least. However, MTP now consumes significantly more VRAM. Personally 16GB just isn't enough to use MTP and run it with a good context size. Time to upgrade lads, 24GB+ users are eating GOOD today 🔥 Full setup below ↓

left curve dev

29,946 просмотров • 2 месяцев назад

HERMES AGENT NOW RUNS A FULL BUSINESS FOR $8/MONTH. CONTENT. CODE. INBOX. ALL AGENTS. EVERYTHING WHILE YOU SLEEP. one user built this in one session. zero terminal commands. browser only. the team: → content agent: finds video topics, writes first drafts, helps with animations → code builder: writes website and SaaS code → ops agent: organizes and answers email total setup cost: $1. monthly runtime: $8 VPS + cheap model API. the trick is matching model to task. frontier models (Opus 4.8, GPT-5.5) for complex /goals only. $5/million tokens. mid-tier (Claude Sonnet, Haiku, Kimiko 2.6) for daily content and research work. budget (DeepSeek V4 Flash, MiniMax 2.7, Gemma 4, Qwen 3) for cron jobs and routine. run the initial setup on Opus. switch to DeepSeek for everything after. your bill drops 90% with zero quality loss on routine work. the agent team picks up tasks from kanban. each completed task creates a skill. week 3 the same workflows run faster than week 1. full breakdown on Hermes Os in the article 👇
14:53

HERMES AGENT NOW RUNS A FULL BUSINESS FOR $8/MONTH. CONTENT. CODE. INBOX. ALL AGENTS. EVERYTHING WHILE YOU SLEEP. one user built this in one session. zero terminal commands. browser only. the team: → content agent: finds video topics, writes first drafts, helps with animations → code builder: writes website and SaaS code → ops agent: organizes and answers email total setup cost: $1. monthly runtime: $8 VPS + cheap model API. the trick is matching model to task. frontier models (Opus 4.8, GPT-5.5) for complex /goals only. $5/million tokens. mid-tier (Claude Sonnet, Haiku, Kimiko 2.6) for daily content and research work. budget (DeepSeek V4 Flash, MiniMax 2.7, Gemma 4, Qwen 3) for cron jobs and routine. run the initial setup on Opus. switch to DeepSeek for everything after. your bill drops 90% with zero quality loss on routine work. the agent team picks up tasks from kanban. each completed task creates a skill. week 3 the same workflows run faster than week 1. full breakdown on Hermes Os in the article 👇

YanXbt

30,910 просмотров • 1 месяц назад

we made distributed inference verifiable with <1% overhead. verification is critical for any distributed system. in a trustless network, actors may swap your 70B model for a cheaper 8B one to cut costs. until now, maintaining inference integrity meant either doubling your cost (redundancy) or exploding your latency (zkp). we created veri: an on-chain verification layer light enough for high-throughput frameworks like Parallax. it hits the economic sweet spot through architectural elegance: 1. commit-sample-verify we don't prove every step; we check a random slice using game theory. workers commit to their work before the audit. cheating becomes statistically irrational, allowing a 1% sample to secure the entire sequence. 2. simultaneous execution inference and verification happen simultaneously on the same worker pool. we don't need a separate "verifier set", so compute utilization stays high. find out more about the architecture and benchmarks: paper: blog:
0:20

we made distributed inference verifiable with <1% overhead. verification is critical for any distributed system. in a trustless network, actors may swap your 70B model for a cheaper 8B one to cut costs. until now, maintaining inference integrity meant either doubling your cost (redundancy) or exploding your latency (zkp). we created veri: an on-chain verification layer light enough for high-throughput frameworks like Parallax. it hits the economic sweet spot through architectural elegance: 1. commit-sample-verify we don't prove every step; we check a random slice using game theory. workers commit to their work before the audit. cheating becomes statistically irrational, allowing a 1% sample to secure the entire sequence. 2. simultaneous execution inference and verification happen simultaneously on the same worker pool. we don't need a separate "verifier set", so compute utilization stays high. find out more about the architecture and benchmarks: paper: blog:

Parallax

28,496 просмотров • 6 месяцев назад

Most speculative decoding still drafts tokens one at a time. That's not parallel generation — it just hides the serial loop behind a smaller model. UC San Diego's z-lab just drew a clear line between the two. They released DFlash — a lightweight block diffusion model that drafts a whole block of tokens in a single forward pass, then lets the target model verify the block in parallel. Up to 15× higher throughput for gpt-oss-120b on NVIDIA Blackwell. No token-by-token drafting anywhere in the speculative path. Here's what's actually interesting: → The drafter is conditioned on the target model's own hidden features, injected into the Key/Value cache of every draft layer — so acceptance length scales with draft depth instead of diluting away → A 5-layer drafter replaces the 7B diffusion drafters that capped earlier methods near 3–4× → MATH-500 speedup: 6.08× vs. 1.81× for EAGLE-3 (4.86× average vs. 1.76×, Qwen3-8B, greedy) → Up to 15× higher throughput for gpt-oss-120b on NVIDIA Blackwell — at the same interactivity target → Lossless: the target still verifies every token, so output quality is preserved Full analysis: Paper: NVIDIA's metrics: Project: Model weights: Repo: NVIDIA AI NVIDIA AIDev
0:20

Most speculative decoding still drafts tokens one at a time. That's not parallel generation — it just hides the serial loop behind a smaller model. UC San Diego's z-lab just drew a clear line between the two. They released DFlash — a lightweight block diffusion model that drafts a whole block of tokens in a single forward pass, then lets the target model verify the block in parallel. Up to 15× higher throughput for gpt-oss-120b on NVIDIA Blackwell. No token-by-token drafting anywhere in the speculative path. Here's what's actually interesting: → The drafter is conditioned on the target model's own hidden features, injected into the Key/Value cache of every draft layer — so acceptance length scales with draft depth instead of diluting away → A 5-layer drafter replaces the 7B diffusion drafters that capped earlier methods near 3–4× → MATH-500 speedup: 6.08× vs. 1.81× for EAGLE-3 (4.86× average vs. 1.76×, Qwen3-8B, greedy) → Up to 15× higher throughput for gpt-oss-120b on NVIDIA Blackwell — at the same interactivity target → Lossless: the target still verifies every token, so output quality is preserved Full analysis: Paper: NVIDIA's metrics: Project: Model weights: Repo: NVIDIA AI NVIDIA AIDev

Marktechpost AI

23,328 просмотров • 1 месяц назад

Timelapse #156 (36 hrs) - Worked with the tiny corp on getting GLM 5.2 running on 8xMI300X (sglang won here) - Launched KernelBench-Mega and updated Kernelbench-Hard with h100 and b200 sweeps - Took care of boring business stuff - Did some training sweeps for specialized technical vocab audio model - Some bugs with putting kernelbench-mega and hard on cloud instances so had to do some reruns. Learned a lot though - Setting up my own local rl infra and profiling concurrency 128 rollouts with vllm. Became clear to me that I need to serve in nvfp4, use MoE only for throughput, reap so training doesn’t OOM, dig into the vllm kernel graph itself to not underutilize my hardware from poor flashinfer/cutlass selections for my rtx pro 6000 sm120 architecture - Might do online distillation from glm 5.2 but for now taking it one step at a time - Slept for a bit then woke up and showered - Fixed an issue with SGLang tensor parallel deadlock on GLM 5.2 architecture with MTP enabled - GLM 5.2 inference is 2-3x faster than coding plans and running on amd boxes - Spent time with family - Hung out with some friends - Recorded some yoctogpt lectures with the revamped notebook (high taste btw) - Setting up dflash training for GLM 5.2
0:30

Timelapse #156 (36 hrs) - Worked with the tiny corp on getting GLM 5.2 running on 8xMI300X (sglang won here) - Launched KernelBench-Mega and updated Kernelbench-Hard with h100 and b200 sweeps - Took care of boring business stuff - Did some training sweeps for specialized technical vocab audio model - Some bugs with putting kernelbench-mega and hard on cloud instances so had to do some reruns. Learned a lot though - Setting up my own local rl infra and profiling concurrency 128 rollouts with vllm. Became clear to me that I need to serve in nvfp4, use MoE only for throughput, reap so training doesn’t OOM, dig into the vllm kernel graph itself to not underutilize my hardware from poor flashinfer/cutlass selections for my rtx pro 6000 sm120 architecture - Might do online distillation from glm 5.2 but for now taking it one step at a time - Slept for a bit then woke up and showered - Fixed an issue with SGLang tensor parallel deadlock on GLM 5.2 architecture with MTP enabled - GLM 5.2 inference is 2-3x faster than coding plans and running on amd boxes - Spent time with family - Hung out with some friends - Recorded some yoctogpt lectures with the revamped notebook (high taste btw) - Setting up dflash training for GLM 5.2

Elliot Arledge

45,447 просмотров • 1 месяц назад

What's the Big Deal with DeepSeek in AI? Here's why DeepSeek is making everyone take notice: 1. Super Smart on a Budget: DeepSeek showed you can make awesome AI without breaking the bank. Their latest model, DeepSeek-V3, was trained for only about $10 million, which is a lot less than the usual big bucks spent on AI, like the rumored $78 million for some of OpenAI's models. They did this in just two months with fewer fancy computers. 2. Open for Everyone: DeepSeek isn't keeping their tech a secret. They've made it open-source, meaning anyone can use, tweak, and learn from it. It's like they're saying, "Come join the party!" 3. Beating the Big Names: DeepSeek-V3 has done better than some top dogs from companies like OpenAI and Google in solving puzzles, math, and coding. This proves you can get great AI results without spending a fortune. 4. Challenging NVIDIA: NVIDIA's chips are usually the choice for AI because they're really powerful. But since DeepSeek did so well with less expensive chips, it might make people think twice about always going for NVIDIA's priciest options. 5. The DeepSeek Crew: The team at DeepSeek is young and smart, mostly from top Chinese schools, with brains in physics, math, and computer science. They learned AI in about six months by themselves! They use first principle thinking, which means they break down problems to the basics and build from there. This has helped them come up with cool new ways to do AI. 6. Changing AI for Good: DeepSeek is showing that AI can be cheaper and more open to everyone. They're changing how we think AI should be made and shared, which could shake up the whole AI world. So, as we watch DeepSeek, it's clear they're not just another player; they're changing the rules of the game. I predicted that this would be a make or break year for all the massive investments made in AI by American VC's. A few weeks later, DeepSeek happens! Watch the rest of my predictions in my 2025 outlook video . Link in replies #AIInnovation #DeepSeek #NVIDIA #OpenAI #TechDisruption
0:44

What's the Big Deal with DeepSeek in AI? Here's why DeepSeek is making everyone take notice: 1. Super Smart on a Budget: DeepSeek showed you can make awesome AI without breaking the bank. Their latest model, DeepSeek-V3, was trained for only about $10 million, which is a lot less than the usual big bucks spent on AI, like the rumored $78 million for some of OpenAI's models. They did this in just two months with fewer fancy computers. 2. Open for Everyone: DeepSeek isn't keeping their tech a secret. They've made it open-source, meaning anyone can use, tweak, and learn from it. It's like they're saying, "Come join the party!" 3. Beating the Big Names: DeepSeek-V3 has done better than some top dogs from companies like OpenAI and Google in solving puzzles, math, and coding. This proves you can get great AI results without spending a fortune. 4. Challenging NVIDIA: NVIDIA's chips are usually the choice for AI because they're really powerful. But since DeepSeek did so well with less expensive chips, it might make people think twice about always going for NVIDIA's priciest options. 5. The DeepSeek Crew: The team at DeepSeek is young and smart, mostly from top Chinese schools, with brains in physics, math, and computer science. They learned AI in about six months by themselves! They use first principle thinking, which means they break down problems to the basics and build from there. This has helped them come up with cool new ways to do AI. 6. Changing AI for Good: DeepSeek is showing that AI can be cheaper and more open to everyone. They're changing how we think AI should be made and shared, which could shake up the whole AI world. So, as we watch DeepSeek, it's clear they're not just another player; they're changing the rules of the game. I predicted that this would be a make or break year for all the massive investments made in AI by American VC's. A few weeks later, DeepSeek happens! Watch the rest of my predictions in my 2025 outlook video . Link in replies #AIInnovation #DeepSeek #NVIDIA #OpenAI #TechDisruption

Dr Ola Brown

83,460 просмотров • 1 год назад

Realtime-VLA FLASH tackles one of the biggest deployment bottlenecks for diffusion-based VLAs: inference latency. The key idea is speculative inference for flow-matching VLAs. A lightweight draft model predicts an action chunk, while the main model’s Action Expert verifies it in parallel using flow-consistency checks instead of running full denoising every replanning round. This lets the system replace many expensive 58 ms full inference rounds with speculative rounds as fast as 7.8 ms, reducing average latency to 19.1 ms and achieving a 3.04× speedup on LIBERO while largely preserving success rate. Interesting systems insight: they profile π0 and show VLM prefill is compute-bound, while Action Denoise is memory-bound. FLASH exploits this by reusing KV cache and parallelizing verification instead of repeatedly running sequential denoising.
1:54

Realtime-VLA FLASH tackles one of the biggest deployment bottlenecks for diffusion-based VLAs: inference latency. The key idea is speculative inference for flow-matching VLAs. A lightweight draft model predicts an action chunk, while the main model’s Action Expert verifies it in parallel using flow-consistency checks instead of running full denoising every replanning round. This lets the system replace many expensive 58 ms full inference rounds with speculative rounds as fast as 7.8 ms, reducing average latency to 19.1 ms and achieving a 3.04× speedup on LIBERO while largely preserving success rate. Interesting systems insight: they profile π0 and show VLM prefill is compute-bound, while Action Denoise is memory-bound. FLASH exploits this by reusing KV cache and parallelizing verification instead of repeatedly running sequential denoising.

Robots Digest 🤖

14,438 просмотров • 2 месяцев назад

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?
0:26

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Avi Chawla

157,390 просмотров • 2 месяцев назад

[SIGGRAPH '25] Monocular Online Reconstruction with Enhanced Detail Preservation Abstract (excerpt): Our approach addresses two key challenges in monocular online reconstruction: 1. Distributing Gaussians without relying on depth maps. 2. Ensuring both local and global consistency in the reconstructed maps. To achieve this, we introduce two key modules: - Hierarchical Gaussian Management Module: For effective Gaussian distribution. - Global Consistency Optimization Module: For maintaining alignment and coherence at all scales. In addition, we present the Multi-level Occupancy Hash Voxels (MOHV), a structure that regularizes Gaussians to capture details across multiple levels of granularity. MOHV ensures accurate reconstruction of both fine and coarse geometries and textures, preserving intricate details while maintaining overall structural integrity. Compared to state-of-the-art RGB-only and even RGB-D methods, our framework achieves superior reconstruction quality with high computational efficiency.
0:29

[SIGGRAPH '25] Monocular Online Reconstruction with Enhanced Detail Preservation Abstract (excerpt): Our approach addresses two key challenges in monocular online reconstruction: 1. Distributing Gaussians without relying on depth maps. 2. Ensuring both local and global consistency in the reconstructed maps. To achieve this, we introduce two key modules: - Hierarchical Gaussian Management Module: For effective Gaussian distribution. - Global Consistency Optimization Module: For maintaining alignment and coherence at all scales. In addition, we present the Multi-level Occupancy Hash Voxels (MOHV), a structure that regularizes Gaussians to capture details across multiple levels of granularity. MOHV ensures accurate reconstruction of both fine and coarse geometries and textures, preserving intricate details while maintaining overall structural integrity. Compared to state-of-the-art RGB-only and even RGB-D methods, our framework achieves superior reconstruction quality with high computational efficiency.

MrNeRF

23,638 просмотров • 1 год назад

We’re excited to share that 🥇Llama Nemotron Super 49B v1.5 -- our latest open reasoning model -- is now #1 on the Artificial Analysis Intelligence Index - a leaderboard that spans advanced math, science, and agentic tasks, in the 70B open model category. Llama Nemotron Super 49B v1.5 is trained with high-quality reasoning synthetic data generated from models like Qwen3-235B and DeepSeek R1. It delivers state-of-the-art accuracy and throughput, running on a single H100. Key features: 🎯 Leading accuracy on multi-step reasoning, math, coding, and function-calling 🏗️ Post-trained using RPO, DPO, and RLVR across 26M+ synthetic examples 📊 Fully transparent training data and techniques If you're building AI agents and want a high accuracy, fully-open, and transparent reasoning model that you can deploy anywhere, try Super v1.5 on or download from Hugging Face 🤗 ➡️ Leaderboard:
0:54

We’re excited to share that 🥇Llama Nemotron Super 49B v1.5 -- our latest open reasoning model -- is now #1 on the Artificial Analysis Intelligence Index - a leaderboard that spans advanced math, science, and agentic tasks, in the 70B open model category. Llama Nemotron Super 49B v1.5 is trained with high-quality reasoning synthetic data generated from models like Qwen3-235B and DeepSeek R1. It delivers state-of-the-art accuracy and throughput, running on a single H100. Key features: 🎯 Leading accuracy on multi-step reasoning, math, coding, and function-calling 🏗️ Post-trained using RPO, DPO, and RLVR across 26M+ synthetic examples 📊 Fully transparent training data and techniques If you're building AI agents and want a high accuracy, fully-open, and transparent reasoning model that you can deploy anywhere, try Super v1.5 on or download from Hugging Face 🤗 ➡️ Leaderboard:

NVIDIA AI Developer

100,506 просмотров • 1 год назад

I had to test it myself to believe this unreal inference speed. 3,000 tokens/s for 1 user on standard datacenter GPUs. They leveraged a hidden efficiency gap in how GPUs generate tokens. Kog just achieved 3,000 tokens/s on 8× AMD MI300X GPUs and 2,100 on 8× NVIDIA H200 (FP16, no speculative decoding). Their tech preview is on a 2B model, and they show how their techniques will scale to large frontier MoE models at similar speeds. That's a huge number because normal low-batch GPU decoding for 2B to 8B models is usually closer to 100 to 300 tokens/s per request, so Kog is claiming something like a 10X to 30X jump in the speed one user actually feels. Their trick: they are getting the speed by treating LLM decoding as a memory streaming problem, not mainly a math problem. For 1 user at batch size 1, the GPU is not doing big, efficient matrix-matrix work like in training or large-batch serving; it is repeatedly pulling the model’s active weights from high-bandwidth memory for each new token, so speed depends on how smoothly those weights keep flowing. Normal inference stacks keep breaking that flow. They run many separate GPU programs for different parts of the model, move intermediate results through memory, wait at synchronization points, talk back to the CPU for scheduling or sampling, and then repeat this token after token. Kog’s answer is to co-design 3 things that are usually tuned separately: the runtime, the low-level GPU code, and the model architecture. The biggest engineering move is the monokernel, where the whole decode pass runs as 1 persistent GPU-resident program, including sampling, so the system does not keep stopping for kernel launches, CPU scheduling, and intermediate memory round trips. They also rebuilt synchronization, because their own measurements say grid sync was eating around 35% of token-generation time; instead of making every compute unit wait at a broad barrier, each unit waits only for the exact data it needs. On AMD MI300X, they also map memory access around the chiplet layout, because memory latency changes depending on which die makes the request. Then their Laneformer model uses Delayed Tensor Parallelism, which lets cross-GPU communication happen in the background instead of blocking every layer.
1:08

I had to test it myself to believe this unreal inference speed. 3,000 tokens/s for 1 user on standard datacenter GPUs. They leveraged a hidden efficiency gap in how GPUs generate tokens. Kog just achieved 3,000 tokens/s on 8× AMD MI300X GPUs and 2,100 on 8× NVIDIA H200 (FP16, no speculative decoding). Their tech preview is on a 2B model, and they show how their techniques will scale to large frontier MoE models at similar speeds. That's a huge number because normal low-batch GPU decoding for 2B to 8B models is usually closer to 100 to 300 tokens/s per request, so Kog is claiming something like a 10X to 30X jump in the speed one user actually feels. Their trick: they are getting the speed by treating LLM decoding as a memory streaming problem, not mainly a math problem. For 1 user at batch size 1, the GPU is not doing big, efficient matrix-matrix work like in training or large-batch serving; it is repeatedly pulling the model’s active weights from high-bandwidth memory for each new token, so speed depends on how smoothly those weights keep flowing. Normal inference stacks keep breaking that flow. They run many separate GPU programs for different parts of the model, move intermediate results through memory, wait at synchronization points, talk back to the CPU for scheduling or sampling, and then repeat this token after token. Kog’s answer is to co-design 3 things that are usually tuned separately: the runtime, the low-level GPU code, and the model architecture. The biggest engineering move is the monokernel, where the whole decode pass runs as 1 persistent GPU-resident program, including sampling, so the system does not keep stopping for kernel launches, CPU scheduling, and intermediate memory round trips. They also rebuilt synchronization, because their own measurements say grid sync was eating around 35% of token-generation time; instead of making every compute unit wait at a broad barrier, each unit waits only for the exact data it needs. On AMD MI300X, they also map memory access around the chiplet layout, because memory latency changes depending on which die makes the request. Then their Laneformer model uses Delayed Tensor Parallelism, which lets cross-GPU communication happen in the background instead of blocking every layer.

Rohan Paul

13,148 просмотров • 2 месяцев назад

"Mac Minis for example are a very good fit" - Andrej Karpathy Andrej Karpathy shouted out my work on EXO Labs in his keynote at Y Combinator AI SUS! Here's the breakdown: Right now most AI workloads run in the cloud where requests from different users are continuously batched together. These workloads are FLOPS-bound and favors hardware with the best unit economics of $ per FLOP, i.e. enterprise GPUs. The personal computing revolution will shift these workloads to personal devices with lower batch sizes (mostly batch_size=1). batch_size=1 inference is memory-bound, because all of the model parameters need to be loaded into the GPU every time a token is generated. Apple Silicon with its Unified Memory architecture has a lot of memory and memory bandwidth per $ compared to other hardware: - M4 Pro Mac Mini, 24GB @ 273GB/s, $58.33/GB, $5.13/GB/s - H100, 80GB @ 3350GB/s, $625/GB, $14.93/GB/s The unit economics of Apple Silicon are becoming more compelling with every release of the Mac. The future of AI inference looks more like open weights models (OpenAI open weights model soonTM) run at low batch_size on personal devices.
0:20

"Mac Minis for example are a very good fit" - Andrej Karpathy Andrej Karpathy shouted out my work on EXO Labs in his keynote at Y Combinator AI SUS! Here's the breakdown: Right now most AI workloads run in the cloud where requests from different users are continuously batched together. These workloads are FLOPS-bound and favors hardware with the best unit economics of $ per FLOP, i.e. enterprise GPUs. The personal computing revolution will shift these workloads to personal devices with lower batch sizes (mostly batch_size=1). batch_size=1 inference is memory-bound, because all of the model parameters need to be loaded into the GPU every time a token is generated. Apple Silicon with its Unified Memory architecture has a lot of memory and memory bandwidth per $ compared to other hardware: - M4 Pro Mac Mini, 24GB @ 273GB/s, $58.33/GB, $5.13/GB/s - H100, 80GB @ 3350GB/s, $625/GB, $14.93/GB/s The unit economics of Apple Silicon are becoming more compelling with every release of the Mac. The future of AI inference looks more like open weights models (OpenAI open weights model soonTM) run at low batch_size on personal devices.

Alex Cheema

103,566 просмотров • 1 год назад