Video yükleniyor...

Video Yüklenemedi

Ana Sayfaya Dön

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...

Marktechpost's profile picture

Marktechpost AI

11,515 subscribers

23,328 görüntüleme • 4 ay önce •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Yorum

Yorum bulunmuyor

Orijinal gönderinin yorumları burada görünecek

Benzer Videolar

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

158,352 görüntüleme • 5 ay önce

As AI labs race to train and deploy new frontier models, existing models become more affordable with better tokenomics. ✨ "Everybody's trying to get to the next frontier. And every time they get to the next frontier, the last generation AI tokens, the cost starts to decline about a factor of 10x every year," said NVIDIA CEO Jensen Huang in a recent keynote. Model optimization techniques such as speculative decoding and multi-token prediction, combined with inference serving platforms like NVIDIA Dynamo on NVIDIA Blackwell NVL72 systems, enable AI factories to boost throughput by 10x with one-tenth of the cost per token. Learn more about AI factory tokenomics ➡️
0:45

As AI labs race to train and deploy new frontier models, existing models become more affordable with better tokenomics. ✨ "Everybody's trying to get to the next frontier. And every time they get to the next frontier, the last generation AI tokens, the cost starts to decline about a factor of 10x every year," said NVIDIA CEO Jensen Huang in a recent keynote. Model optimization techniques such as speculative decoding and multi-token prediction, combined with inference serving platforms like NVIDIA Dynamo on NVIDIA Blackwell NVL72 systems, enable AI factories to boost throughput by 10x with one-tenth of the cost per token. Learn more about AI factory tokenomics ➡️

NVIDIA AI

16,053 görüntüleme • 9 ay önce

Massive breakthrough in speculative decoding! It made a 9B LLM on my MacBook 4x faster, and 2.6x faster than regular speculative decoding. I ran Qwen3.5 9B on my M5 MacBook Pro with the same prompt, three ways: - No speculative decoding: done in 42s - Regular speculative decoding: done in 30s - DFlash + Weaver: done in 11s The video below shows all three side by side. Here's why it works: Speculative decoding lets a cheap drafter guess the next few tokens, and the main model checks them all in one pass. The fastest drafters, like DFlash, guess every position in parallel. But each guess is made on its own, so position 5 never sees what was picked at position 4. One early mismatch and everything after it gets thrown away. Weaver fixes this. It's a tiny adapter (one Transformer layer) on top of DFlash that: ↳ picks among DFlash's top candidates left to right, so each pick depends on the last ↳ grows a tree of likely continuations instead of a single chain ↳ lets the main model verify the whole tree in one pass On my Mac, every pass of the model produced 7.5 tokens instead of 1. On a B200 GPU, the paper reports 4x faster decoding across chat, math and code. Mirai Labs open-sourced the whole stack in their engine, uzu, built for Apple Silicon. GitHub repo: (don't forget to star it 🌟) Trying it takes two commands: 𝗯𝗿𝗲𝘄 𝗶𝗻𝘀𝘁𝗮𝗹𝗹 𝗺𝗶𝗿𝗮𝗶 𝗺𝗶𝗿𝗮𝗶 --𝗺𝗼𝗱𝗲𝗹 𝘁𝗿𝘆𝗺𝗶𝗿𝗮𝗶/𝗤𝘄𝗲𝗻𝟯.𝟱-𝟵𝗕-𝗠 -𝗺 "𝗪𝗿𝗶𝘁𝗲 𝗮 𝗽𝘆𝘁𝗵𝗼𝗻 𝗳𝘂𝗻𝗰𝘁𝗶𝗼𝗻 𝘁𝗼 𝗺𝗲𝗿𝗴𝗲 𝘁𝘄𝗼 𝘀𝗼𝗿𝘁𝗲𝗱 𝗹𝗶𝘀𝘁𝘀" --- I wrote a full breakdown of why local inference is memory-bound and how uzu gets past it, with reproducible code. The article is quoted below. In the replies, there's an illustration of standard vs. speculative vs. uzu decoding.
0:30

Massive breakthrough in speculative decoding! It made a 9B LLM on my MacBook 4x faster, and 2.6x faster than regular speculative decoding. I ran Qwen3.5 9B on my M5 MacBook Pro with the same prompt, three ways: - No speculative decoding: done in 42s - Regular speculative decoding: done in 30s - DFlash + Weaver: done in 11s The video below shows all three side by side. Here's why it works: Speculative decoding lets a cheap drafter guess the next few tokens, and the main model checks them all in one pass. The fastest drafters, like DFlash, guess every position in parallel. But each guess is made on its own, so position 5 never sees what was picked at position 4. One early mismatch and everything after it gets thrown away. Weaver fixes this. It's a tiny adapter (one Transformer layer) on top of DFlash that: ↳ picks among DFlash's top candidates left to right, so each pick depends on the last ↳ grows a tree of likely continuations instead of a single chain ↳ lets the main model verify the whole tree in one pass On my Mac, every pass of the model produced 7.5 tokens instead of 1. On a B200 GPU, the paper reports 4x faster decoding across chat, math and code. Mirai Labs open-sourced the whole stack in their engine, uzu, built for Apple Silicon. GitHub repo: (don't forget to star it 🌟) Trying it takes two commands: 𝗯𝗿𝗲𝘄 𝗶𝗻𝘀𝘁𝗮𝗹𝗹 𝗺𝗶𝗿𝗮𝗶 𝗺𝗶𝗿𝗮𝗶 --𝗺𝗼𝗱𝗲𝗹 𝘁𝗿𝘆𝗺𝗶𝗿𝗮𝗶/𝗤𝘄𝗲𝗻𝟯.𝟱-𝟵𝗕-𝗠 -𝗺 "𝗪𝗿𝗶𝘁𝗲 𝗮 𝗽𝘆𝘁𝗵𝗼𝗻 𝗳𝘂𝗻𝗰𝘁𝗶𝗼𝗻 𝘁𝗼 𝗺𝗲𝗿𝗴𝗲 𝘁𝘄𝗼 𝘀𝗼𝗿𝘁𝗲𝗱 𝗹𝗶𝘀𝘁𝘀" --- I wrote a full breakdown of why local inference is memory-bound and how uzu gets past it, with reproducible code. The article is quoted below. In the replies, there's an illustration of standard vs. speculative vs. uzu decoding.

Akshay 🚀

25,804 görüntüleme • 2 gün önce

A voice agent powered by gpt-oss. Running locally on my macBook. Demo recorded in a Waymo with WiFi turned off. I'm still on my space game voice AI kick, obviously. Code link below. For conversational voice AI, you want to set the gpt-oss reasoning behavior to "low". (The default is "medium".) Notes on how to do that and a jinja template you can use are in the repo. The LLM in the demo video is the big, 120B version of gpt-oss. You can use the smaller, 20B model for this, of course. But OpenAI really did a cool thing here designing the 120B model to run in "just" 80GB of VRAM. And the llama.cpp mlx inference is fast: ~250ms TTFT. Running a big model on-device feels like a time warp into the future of AI.
4:35

A voice agent powered by gpt-oss. Running locally on my macBook. Demo recorded in a Waymo with WiFi turned off. I'm still on my space game voice AI kick, obviously. Code link below. For conversational voice AI, you want to set the gpt-oss reasoning behavior to "low". (The default is "medium".) Notes on how to do that and a jinja template you can use are in the repo. The LLM in the demo video is the big, 120B version of gpt-oss. You can use the smaller, 20B model for this, of course. But OpenAI really did a cool thing here designing the 120B model to run in "just" 80GB of VRAM. And the llama.cpp mlx inference is fast: ~250ms TTFT. Running a big model on-device feels like a time warp into the future of AI.

kwindla

202,309 görüntüleme • 1 yıl önce

Red Hat AI just shipped DFlash speculator checkpoints for two of NVIDIA AI's most powerful open models: → Nemotron Ultra 550B → Nemotron Super 120B On math and reasoning: ~5 out of 7 draft tokens accepted on average. On code (HumanEval): ~3.4 out of 7. Both checkpoints trained with the open source Speculators library from vLLM. Apache 2.0. Validated on NVIDIA B200. One flag to enable in vLLM: --spec-model RedHatAI/NVIDIA-Nemotron-3-Ultra-550B-A55B-speculator.dflash --spec-tokens 7 --spec-method dflash 🔗 Ultra 550B: 🔗 Super 120B:
0:32

Red Hat AI just shipped DFlash speculator checkpoints for two of NVIDIA AI's most powerful open models: → Nemotron Ultra 550B → Nemotron Super 120B On math and reasoning: ~5 out of 7 draft tokens accepted on average. On code (HumanEval): ~3.4 out of 7. Both checkpoints trained with the open source Speculators library from vLLM. Apache 2.0. Validated on NVIDIA B200. One flag to enable in vLLM: --spec-model RedHatAI/NVIDIA-Nemotron-3-Ultra-550B-A55B-speculator.dflash --spec-tokens 7 --spec-method dflash 🔗 Ultra 550B: 🔗 Super 120B:

Red Hat AI

15,077 görüntüleme • 3 ay önce

NVIDIA AI Releases Nemotron 3 Embed: An Open Embedding Collection Whose 8B Checkpoint Ranks #1 on RTEB Most RAG stacks treat the embedding model as a commodity — pick one, index, move on. Nemotron 3 Embed is NVIDIA's argument that the retrieval layer is where agent cost actually gets set. They released three open checkpoints — Nemotron-3-Embed-8B-BF16, 1B-BF16, and 1B-NVFP4 — built on Ministral bases, trained with bidirectional attention masking, pooled by averaging token-level representations, all taking 32,768-token inputs under OpenMDW-1.1. Here's what's actually interesting: → The 8B ranks #1 overall on RTEB: 78.46 avg NDCG@10, alongside 75.45 on MMTEB Retrieval and 60.60 on ViDoRe-V3 text → The 1B wasn't trained small. It was pruned from a 3B parent with ModelOpt mcore_minitron NAS, then distilled from the 8B teacher on COS + MSE loss — twice → That pipeline lands the 1.14B checkpoint at 72.38 RTEB, up 10.4 points on the prior-generation llama-nemotron-embed-vl-1b-v2 → NVFP4 costs 0.38 RTEB points (72.00 vs 72.38, ~99.5% retention) and buys up to 2x BF16 throughput on Blackwell Full Analysis: Model weight: Technical details: NVIDIA AI NVIDIA
0:35

NVIDIA AI Releases Nemotron 3 Embed: An Open Embedding Collection Whose 8B Checkpoint Ranks #1 on RTEB Most RAG stacks treat the embedding model as a commodity — pick one, index, move on. Nemotron 3 Embed is NVIDIA's argument that the retrieval layer is where agent cost actually gets set. They released three open checkpoints — Nemotron-3-Embed-8B-BF16, 1B-BF16, and 1B-NVFP4 — built on Ministral bases, trained with bidirectional attention masking, pooled by averaging token-level representations, all taking 32,768-token inputs under OpenMDW-1.1. Here's what's actually interesting: → The 8B ranks #1 overall on RTEB: 78.46 avg NDCG@10, alongside 75.45 on MMTEB Retrieval and 60.60 on ViDoRe-V3 text → The 1B wasn't trained small. It was pruned from a 3B parent with ModelOpt mcore_minitron NAS, then distilled from the 8B teacher on COS + MSE loss — twice → That pipeline lands the 1.14B checkpoint at 72.38 RTEB, up 10.4 points on the prior-generation llama-nemotron-embed-vl-1b-v2 → NVFP4 costs 0.38 RTEB points (72.00 vs 72.38, ~99.5% retention) and buys up to 2x BF16 throughput on Blackwell Full Analysis: Model weight: Technical details: NVIDIA AI NVIDIA

Marktechpost AI

32,352 görüntüleme • 3 ay önce

Jensen Huang said Grok 5 will be 7 Trillion parameter model. On time to train, with the training window fixed at 1 month, the new Vera-Rubin GPU system of Nvidia needs 1/4 the number of systems compared with Blackwell to train the same frontier model. “factory throughput” improves by about 10x over Blackwell, and Blackwell itself was about 10x over Hopper. i.e. overall Rubin is roughly 100x Hopper in factory throughput per watt, which matters hugely, because a 1 GW, $50B data center is power-limited and revenue scales with throughput per watt. --- From Nvidia YT channel
1:24

Jensen Huang said Grok 5 will be 7 Trillion parameter model. On time to train, with the training window fixed at 1 month, the new Vera-Rubin GPU system of Nvidia needs 1/4 the number of systems compared with Blackwell to train the same frontier model. “factory throughput” improves by about 10x over Blackwell, and Blackwell itself was about 10x over Hopper. i.e. overall Rubin is roughly 100x Hopper in factory throughput per watt, which matters hugely, because a 1 GW, $50B data center is power-limited and revenue scales with throughput per watt. --- From Nvidia YT channel

Rohan Paul

165,770 görüntüleme • 9 ay önce

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.
0:23

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

Avi Chawla

273,899 görüntüleme • 4 ay önce

Big win for local LLM inference! This new inference engine runs LLMs 4x faster than llama.cpp: I ran Qwen3.5 9B on a base M5 MacBook Pro with three engines: - llama.cpp: 22.0 tokens/s - MLX: 25.1 tokens/s - uzu: 92.1 tokens/s The video below shows all three runs side by side. The first two numbers don't mean llama.cpp or MLX are badly built. But both are sitting right under a hard limit of local decoding. When an LLM generates text, every new token needs one full pass through the model, and that pass reads all of the model's weights from memory. A base M5 reads memory at about 153 GB/s. A 4-bit 9B model weighs about 5.2 GB. So at one token per pass, the ceiling is roughly 153 ÷ 5.2 ≈ 29 tokens/s, however fast the GPU computes. uzu gets past that ceiling by producing several tokens per pass instead of one: ↳ A small drafter (DFlash) proposes candidates for the next 16 positions at once. ↳ A tiny adapter (Weaver, 56.7M params) picks among them left to right, so the guesses form coherent text, and keeps a few alternative branches as a tree. ↳ The 9B model checks the whole tree in a single pass and keeps the longest path it agrees with. The main model still decides every token that ends up in the output. GitHub repo: (don't forget to star it 🌟) In my run, each pass of the 9B model produced 7.5 tokens on average. Ordinary decoding produces exactly 1. The weights and kernels are designed around this too. uzu ships its own 4-bit checkpoints, and on M5 chips it runs the verification step on the GPU's int8 matrix path. uzu is an open-source inference engine by Mirai Labs, written in Rust with Metal kernels. It has Swift, Python, TypeScript and Rust bindings, so the same engine runs inside a Mac or iOS app. I shared the GitHub repo above. If you want to dive deeper, my co-founder wrote a detailed breakdown of why local inference is memory-bound and how uzu's speculative decoding gets past it. Read it below.
0:26

Big win for local LLM inference! This new inference engine runs LLMs 4x faster than llama.cpp: I ran Qwen3.5 9B on a base M5 MacBook Pro with three engines: - llama.cpp: 22.0 tokens/s - MLX: 25.1 tokens/s - uzu: 92.1 tokens/s The video below shows all three runs side by side. The first two numbers don't mean llama.cpp or MLX are badly built. But both are sitting right under a hard limit of local decoding. When an LLM generates text, every new token needs one full pass through the model, and that pass reads all of the model's weights from memory. A base M5 reads memory at about 153 GB/s. A 4-bit 9B model weighs about 5.2 GB. So at one token per pass, the ceiling is roughly 153 ÷ 5.2 ≈ 29 tokens/s, however fast the GPU computes. uzu gets past that ceiling by producing several tokens per pass instead of one: ↳ A small drafter (DFlash) proposes candidates for the next 16 positions at once. ↳ A tiny adapter (Weaver, 56.7M params) picks among them left to right, so the guesses form coherent text, and keeps a few alternative branches as a tree. ↳ The 9B model checks the whole tree in a single pass and keeps the longest path it agrees with. The main model still decides every token that ends up in the output. GitHub repo: (don't forget to star it 🌟) In my run, each pass of the 9B model produced 7.5 tokens on average. Ordinary decoding produces exactly 1. The weights and kernels are designed around this too. uzu ships its own 4-bit checkpoints, and on M5 chips it runs the verification step on the GPU's int8 matrix path. uzu is an open-source inference engine by Mirai Labs, written in Rust with Metal kernels. It has Swift, Python, TypeScript and Rust bindings, so the same engine runs inside a Mac or iOS app. I shared the GitHub repo above. If you want to dive deeper, my co-founder wrote a detailed breakdown of why local inference is memory-bound and how uzu's speculative decoding gets past it. Read it below.

Avi Chawla

125,660 görüntüleme • 6 gün önce

🚨 NEW ON AND make NVIDIA's Nemotron 3.5 Lightning faster. Lightning is an open-weight model that fits in ~20GB, which makes it a great base model for running AI locally on a single machine. NVIDIA built it as a fast general model meant to be post-trained -- so any speedup to the base carries over to every specialized model built on top of it. On day 1, solvers on Yukon already got it running 31.9% faster on DGX Spark and 13.8% faster on Mac. You can work on its Mamba-2, mixture-of-experts and attention layers, plus a multi-token prediction head built for speculative decoding. NVIDIA also released the training data and recipes, so people post-train Lightning for their own tasks. Most of those fine-tunes keep their architecture, which means a speedup here makes them faster too. Two tracks: • Apple Silicon → • NVIDIA DGX Spark → The challenge runs until Oct 29, 2026 To join, pick a track, point your agent at it and let it rip.
0:30

🚨 NEW ON AND make NVIDIA's Nemotron 3.5 Lightning faster. Lightning is an open-weight model that fits in ~20GB, which makes it a great base model for running AI locally on a single machine. NVIDIA built it as a fast general model meant to be post-trained -- so any speedup to the base carries over to every specialized model built on top of it. On day 1, solvers on Yukon already got it running 31.9% faster on DGX Spark and 13.8% faster on Mac. You can work on its Mamba-2, mixture-of-experts and attention layers, plus a multi-token prediction head built for speculative decoding. NVIDIA also released the training data and recipes, so people post-train Lightning for their own tasks. Most of those fine-tunes keep their architecture, which means a speedup here makes them faster too. Two tracks: • Apple Silicon → • NVIDIA DGX Spark → The challenge runs until Oct 29, 2026 To join, pick a track, point your agent at it and let it rip.

Kydo

17,312 görüntüleme • 9 gün önce

Continuous batching in LLMs, clearly explained: (a popular LLM interview question; bookmark this) In traditional ML inference, a batch is a matrix. Every input is padded to the same length, one forward pass runs, and every row finishes at the same moment. LLM decoding does not work that way. One forward pass produces one token per sequence, so a request needs as many passes as it has output tokens, and nobody knows that count until the model emits a stop token. Under static batching, membership is fixed when the batch starts. A request that finishes in 30 tokens holds its slot until the slowest request in the same batch finishes at 400. The GPU keeps paying the full weight read for a batch that is mostly empty. Loading model weights out of HBM costs the same whether four slots are producing tokens or one. Continuous batching moves the decision boundary. Instead of scheduling once per batch, the scheduler runs a single forward pass, gets control back, and decides again. A finished request leaves at the next iteration boundary, and a queued request takes its slot right there. No slot stays reserved for work that is already done. Anyscale benchmarked both OPT-13B on a single A100. With uniform generation lengths, the two policies came out about level (as expected), and as output length variance rose, static batching fell to around 81 tokens per second while vLLM reached 23x the throughput of naive Hugging Face serving. Variance drives the entire gap. Production traffic mixes 30-token replies with 400-token ones, which is exactly the condition static batching handles worst. None of this alters the model. vLLM, SGLang, TGI, and TensorRT-LLM all run it by default, and NVIDIA ships the same mechanism under the name in-flight batching. The animation below runs both policies on the same 16 requests and the same 4 slots, stepping in lockstep. The only difference is when a new request is allowed in. To dive deeper into continuous batching specifically, I wrote a full breakdown of the scheduler underneath it. It covers what happens between two forward passes, how tokens get handed out against a fixed budget, why the scheduler needs no separate path for prefill and decode, and what preemption costs you when the KV cache fills up mid-generation. Read it below.
0:29

Continuous batching in LLMs, clearly explained: (a popular LLM interview question; bookmark this) In traditional ML inference, a batch is a matrix. Every input is padded to the same length, one forward pass runs, and every row finishes at the same moment. LLM decoding does not work that way. One forward pass produces one token per sequence, so a request needs as many passes as it has output tokens, and nobody knows that count until the model emits a stop token. Under static batching, membership is fixed when the batch starts. A request that finishes in 30 tokens holds its slot until the slowest request in the same batch finishes at 400. The GPU keeps paying the full weight read for a batch that is mostly empty. Loading model weights out of HBM costs the same whether four slots are producing tokens or one. Continuous batching moves the decision boundary. Instead of scheduling once per batch, the scheduler runs a single forward pass, gets control back, and decides again. A finished request leaves at the next iteration boundary, and a queued request takes its slot right there. No slot stays reserved for work that is already done. Anyscale benchmarked both OPT-13B on a single A100. With uniform generation lengths, the two policies came out about level (as expected), and as output length variance rose, static batching fell to around 81 tokens per second while vLLM reached 23x the throughput of naive Hugging Face serving. Variance drives the entire gap. Production traffic mixes 30-token replies with 400-token ones, which is exactly the condition static batching handles worst. None of this alters the model. vLLM, SGLang, TGI, and TensorRT-LLM all run it by default, and NVIDIA ships the same mechanism under the name in-flight batching. The animation below runs both policies on the same 16 requests and the same 4 slots, stepping in lockstep. The only difference is when a new request is allowed in. To dive deeper into continuous batching specifically, I wrote a full breakdown of the scheduler underneath it. It covers what happens between two forward passes, how tokens get handed out against a fixed budget, why the scheduler needs no separate path for prefill and decode, and what preemption costs you when the KV cache fills up mid-generation. Read it below.

Avi Chawla

16,140 görüntüleme • 2 ay önce

Nvidia just revealed Vera Rubin. Ships H2 2026. The numbers are wild: → 10x more performance per watt vs Blackwell → 10x cheaper inference token cost → 4x fewer GPUs to train the same MoE model Energy was the biggest bottleneck in AI. Nvidia just made it 10x cheaper.
0:27

Nvidia just revealed Vera Rubin. Ships H2 2026. The numbers are wild: → 10x more performance per watt vs Blackwell → 10x cheaper inference token cost → 4x fewer GPUs to train the same MoE model Energy was the biggest bottleneck in AI. Nvidia just made it 10x cheaper.

Min Choi

285,639 görüntüleme • 8 ay önce

NVIDIA just removed one of the biggest friction points in Voice AI. PersonaPlex-7B is an open-source, full-duplex conversational model. Free, open source (MIT), with open model weights on Hugging Face 🤗 Links to repo and weights in 🧵↓ The traditional ASR → LLM → TTS pipeline forces rigid turn-taking. It’s efficient, but it never feels natural. PersonaPlex-7B changes that. This NVIDIA model can listen and speak at the same time. It runs directly on continuous audio tokens with a dual-stream transformer, generating text and audio in parallel instead of passing control between components. That unlocks: → instant back-channel responses → interruptions that feel human → real conversational rhythm Persona control is fully zero-shot! If you’re building low-latency assistants or support agents, this is a big step forward 🔥
0:42

NVIDIA just removed one of the biggest friction points in Voice AI. PersonaPlex-7B is an open-source, full-duplex conversational model. Free, open source (MIT), with open model weights on Hugging Face 🤗 Links to repo and weights in 🧵↓ The traditional ASR → LLM → TTS pipeline forces rigid turn-taking. It’s efficient, but it never feels natural. PersonaPlex-7B changes that. This NVIDIA model can listen and speak at the same time. It runs directly on continuous audio tokens with a dual-stream transformer, generating text and audio in parallel instead of passing control between components. That unlocks: → instant back-channel responses → interruptions that feel human → real conversational rhythm Persona control is fully zero-shot! If you’re building low-latency assistants or support agents, this is a big step forward 🔥

Charly Wargnier ♨️

566,596 görüntüleme • 9 ay önce

Nemotron 3 Full Breakdown With the help of Joey Conway from NVIDIA AI getting into the specifics around why Nemotron 3 is kind of a big deal Biggest headline with Nemotron is: Hybrid Mamba Transformer, Latent MoE, and MTP Hybrid Mamba Transformer essentially attacks right at the Attention mechanism to make the overhead sub-quadratic, but unlike quantizing KV Cache or swapping out attention head, NVIDIA chose Mamba-2 Latent MoE helps further optimize on sparsity by down projecting the dimensions so you're doing less math and less memory movement between HBM and SRAM, you're saving a ton, and NVIDIA made a conscious choice to add more experts given the surplus Finally, MTP or multi token prediction where the model can see future tokens to be more expressive in training and also option to use for speculative decoding during inference Oh, also the model adopts the new OpenMDW 1.1 License
12:11

Nemotron 3 Full Breakdown With the help of Joey Conway from NVIDIA AI getting into the specifics around why Nemotron 3 is kind of a big deal Biggest headline with Nemotron is: Hybrid Mamba Transformer, Latent MoE, and MTP Hybrid Mamba Transformer essentially attacks right at the Attention mechanism to make the overhead sub-quadratic, but unlike quantizing KV Cache or swapping out attention head, NVIDIA chose Mamba-2 Latent MoE helps further optimize on sparsity by down projecting the dimensions so you're doing less math and less memory movement between HBM and SRAM, you're saving a ton, and NVIDIA made a conscious choice to add more experts given the surplus Finally, MTP or multi token prediction where the model can see future tokens to be more expressive in training and also option to use for speculative decoding during inference Oh, also the model adopts the new OpenMDW 1.1 License

Caleb Eom

50,467 görüntüleme • 4 ay önce

GLM 5.2 @ 19.8 tok/sec on 2 DGX Sparks ⚡️ I switched to a dspark draft model with K=2 (model from Red Hat AI) Acceptance is ~68% because the drafter was trained with the full FP8 model... Next step is to fine tune the drafter so acceptance on this super quant goes up
0:30

GLM 5.2 @ 19.8 tok/sec on 2 DGX Sparks ⚡️ I switched to a dspark draft model with K=2 (model from Red Hat AI) Acceptance is ~68% because the drafter was trained with the full FP8 model... Next step is to fine tune the drafter so acceptance on this super quant goes up

Joe Muller

73,297 görüntüleme • 3 ay önce

i'm running a 27b uncensored🚨model at 50 tok/s on a 12gb rtx 3060 and it does not say no to anything ⚠️ model just finished full harmbench red-team suite against it, 200 behaviors, and it refused zero. not a few. zero. aligned models sit in the 90s. the model is ternary bonsai 2 with the heretic uncensoring pass, the weights are ternary, just -1, 0 and +1, so the whole 27b is a file small enough that a gaming card from 2020 holds it with room to spare, and it's a reasoning model. it thinks before it answers, and it still keeps the pace the speed is the head i grafted on, a draft model that predicts the next token before the big model checks it. speculative decoding. the card fires a guess, the 27b confirms, and you get 50 tok/s instead of 39 the whole thing, weights, head and a 128k context window, sits in 10.7 of the 12 gb. kv cache at q4. nothing exotic, one card, one power plug why uncensored. because it's mine. no api keys, no terms of service, no model deciding what an adult is allowed to ask. the refusals left in it are ones it chose, and there are almost none that's the whole point of running local. the policy lives in your house, not someone else's
0:50

i'm running a 27b uncensored🚨model at 50 tok/s on a 12gb rtx 3060 and it does not say no to anything ⚠️ model just finished full harmbench red-team suite against it, 200 behaviors, and it refused zero. not a few. zero. aligned models sit in the 90s. the model is ternary bonsai 2 with the heretic uncensoring pass, the weights are ternary, just -1, 0 and +1, so the whole 27b is a file small enough that a gaming card from 2020 holds it with room to spare, and it's a reasoning model. it thinks before it answers, and it still keeps the pace the speed is the head i grafted on, a draft model that predicts the next token before the big model checks it. speculative decoding. the card fires a guess, the 27b confirms, and you get 50 tok/s instead of 39 the whole thing, weights, head and a 128k context window, sits in 10.7 of the 12 gb. kv cache at q4. nothing exotic, one card, one power plug why uncensored. because it's mine. no api keys, no terms of service, no model deciding what an adult is allowed to ask. the refusals left in it are ones it chose, and there are almost none that's the whole point of running local. the policy lives in your house, not someone else's

Sudo su

59,906 görüntüleme • 18 gün önce

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

13,282 görüntüleme • 5 ay önce

Just a little bit faster than everyone else on OpenAI GPT-OSS-120B. > 10 times faster than the leading NVIDIA GPU provider
0:23

Just a little bit faster than everyone else on OpenAI GPT-OSS-120B. > 10 times faster than the leading NVIDIA GPU provider

Andrew Feldman

124,175 görüntüleme • 1 yıl önce

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 SGLang. Across batch sizes 1 to 256, DSpark gives the best throughput/latency tradeoff on DeepSeek-V4-Flash, ahead of both MTP and non-spec. At high concurrency, dynamic scheduling provides up to ~20% higher throughput compared to a fixed budget, while maintaining high verification quality across workloads. With fused kernels and zero-overhead scheduling, DeepSeek-V4-Pro reaches 383.7 tok/s at B=1 on B300. DSpark is now available in SGLang with support for Qwen3 and DeepSeek-V4. Thanks DeepSeek for open-sourcing! Blog with full technical details and commands to run below 👇
0:33

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 SGLang. Across batch sizes 1 to 256, DSpark gives the best throughput/latency tradeoff on DeepSeek-V4-Flash, ahead of both MTP and non-spec. At high concurrency, dynamic scheduling provides up to ~20% higher throughput compared to a fixed budget, while maintaining high verification quality across workloads. With fused kernels and zero-overhead scheduling, DeepSeek-V4-Pro reaches 383.7 tok/s at B=1 on B300. DSpark is now available in SGLang with support for Qwen3 and DeepSeek-V4. Thanks DeepSeek for open-sourcing! Blog with full technical details and commands to run below 👇

LMSYS Org

170,752 görüntüleme • 3 ay önce