Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache...

_reesechong's profile picture

Reese Chong

1,321 subscribers

53,026 Aufrufe • vor 4 Monaten •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Kommentare

Keine Kommentare verfügbar

Kommentare vom Original-Post werden hier angezeigt

Ähnliche Videos

Hot off the press 🔥 AEON vLLM Ultimate → v0.24.0. Link below ⤵️ The ultimate container for AI on NVIDIA DGX Spark — vLLM compiled from source for GB10/sm_121a. 🔓 UNLOCKED (impossible until this release): ▸ DFlash + FP8 KV cache — 2× KV capacity, draft acceptance intact, measured on-device ▸ FP8 KV on vision models — the KV-bandwidth lever is back ▸ DFlash on FlashInfer & Triton backends (was FlashAttention-only) 🛠 Carried ahead of upstream (still-open PRs, merged + validated): ▸ Triton NVFP4 KV cache → up to 3× KV capacity ▸ DFlash sliding-window drafters + prefix-cache stability for DFlash ▸ High-concurrency DFlash — clean through c=64 🧰 Fixes stock v0.24.0 doesn't have: ▸ Gemma-4 tied embeddings (ModelOpt) — stock crashes at load, ours doesn't ▸ Tool calling out of the box (stock --no-deps builds 500) ▸ Unified-memory hardened — no silent KV overcommit on the 121 GB pool 📊 One image, the whole fleet: Gemma-4-26B-A4B · Qwen3.6-27B · Qwen3.6-35B-A3B (hybrid GDN) 536 tok/s @ c16 · 472 tok/s @ c12 · 262K context · image+video · tools + reasoning parsers · full CUDA-graph capture 🐳 docker pull
0:44

Hot off the press 🔥 AEON vLLM Ultimate → v0.24.0. Link below ⤵️ The ultimate container for AI on NVIDIA DGX Spark — vLLM compiled from source for GB10/sm_121a. 🔓 UNLOCKED (impossible until this release): ▸ DFlash + FP8 KV cache — 2× KV capacity, draft acceptance intact, measured on-device ▸ FP8 KV on vision models — the KV-bandwidth lever is back ▸ DFlash on FlashInfer & Triton backends (was FlashAttention-only) 🛠 Carried ahead of upstream (still-open PRs, merged + validated): ▸ Triton NVFP4 KV cache → up to 3× KV capacity ▸ DFlash sliding-window drafters + prefix-cache stability for DFlash ▸ High-concurrency DFlash — clean through c=64 🧰 Fixes stock v0.24.0 doesn't have: ▸ Gemma-4 tied embeddings (ModelOpt) — stock crashes at load, ours doesn't ▸ Tool calling out of the box (stock --no-deps builds 500) ▸ Unified-memory hardened — no silent KV overcommit on the 121 GB pool 📊 One image, the whole fleet: Gemma-4-26B-A4B · Qwen3.6-27B · Qwen3.6-35B-A3B (hybrid GDN) 536 tok/s @ c16 · 472 tok/s @ c12 · 262K context · image+video · tools + reasoning parsers · full CUDA-graph capture 🐳 docker pull

ÆON FORGE ✨

24,800 Aufrufe • vor 1 Monat

Interesting approach to long context/continual learning here from Baseten at Cursor Compact long trajectories by compressing a prefix of the KV cache using an MLP/autoencoder. You can train this "compactor" MLP by learning to reconstruct activations that the original KV cache would produce on subsequent tokens. This maximally reconstructs information from long context that's useful for subsequent outputs. If you run an agent over extremely long context and run this compaction recursively, the activations of this compressed KV cache become like trained weights. Similar to task-specific LoRA or "cartridges". Would not be surprised if OpenAI is running a similar algorithm for their blackbox compaction. Clear benefits here if you can avoid busting the cache, as e.g. compacting via writing to a text file would. Seeing many emerging approaches to fixed size latent states for LLMs that seem promising. If building a task-specific KV cache compression ends up being more sample efficient than running backprop, getting this to "work" feels like one of the 2-3 remaining breakthroughs on the path to true AGI
13:13

Interesting approach to long context/continual learning here from Baseten at Cursor Compact long trajectories by compressing a prefix of the KV cache using an MLP/autoencoder. You can train this "compactor" MLP by learning to reconstruct activations that the original KV cache would produce on subsequent tokens. This maximally reconstructs information from long context that's useful for subsequent outputs. If you run an agent over extremely long context and run this compaction recursively, the activations of this compressed KV cache become like trained weights. Similar to task-specific LoRA or "cartridges". Would not be surprised if OpenAI is running a similar algorithm for their blackbox compaction. Clear benefits here if you can avoid busting the cache, as e.g. compacting via writing to a text file would. Seeing many emerging approaches to fixed size latent states for LLMs that seem promising. If building a task-specific KV cache compression ends up being more sample efficient than running backprop, getting this to "work" feels like one of the 2-3 remaining breakthroughs on the path to true AGI

Jay Hack

29,792 Aufrufe • vor 1 Monat

Google's Gemma 4 26B A4B QAT hits 25+ tokens/sec and 320+ tokens/sec prefill on 8 GB VRAM (RTX 4060) + 16 GB RAM using TurboQuant Prefill just went from 200 → 320+ tok/s on the same 8GB card. 1.6x, no new hardware, no new quant, just a KV cache trick stacked on top of the Gemma 4 26B MoE setup from a few days ago. A few days ago I posted Gemma 4 26B A4B hitting 28 tok/s decode on 8GB VRAM using native MTP. prefill was stuck around 200 tok/s. fair callout by the community. So today I tested something I'd already been meaning to try: TheTom/llama-cpp-turboquant, the TurboQuant KV cache fork by Tom Turney (Tom Turney). (github link in the comments) thanks to him, the fork just got resynced to mainline, so MTP + TurboQuant now run together cleanly (I didnt see any meaningful gains by using MTP with this setup though but you can try). The flags (No MTP): -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -cnv -c 64000 --cache-type-k q8_0 --cache-type-v turbo3 Results on the same RTX 4060 8GB, tested with a 27k token prompt at 64k context loaded: Prefill: 200 tok/s → 320+ tok/s Decode: stayed above 25 tok/s (without MTP) Why it works: TurboQuant uses walsh hadamard rotation + polar quantization on the KV cache. keys are sensitive to compression, values aren't much, so it splits the difference: K stays at q8_0, V drops to turbo3 (~3 bits). bonus from the memory savings: same 8GB card can now stretch to 100-120k context with minimal decode penalty. It should now be snappier with any agent harness such as hermes agent without compromise on intelligence. If you're already running Gemma 4 on a small card, this stacks on top for free. Try --cache-type-k q8_0 --cache-type-v turbo3 on your setup and report back what your prefill/decode split looks like. unsloth model gguf and llama.cpp turboquant fork links in the comments. what's your prefill number before vs after?
0:31

Google's Gemma 4 26B A4B QAT hits 25+ tokens/sec and 320+ tokens/sec prefill on 8 GB VRAM (RTX 4060) + 16 GB RAM using TurboQuant Prefill just went from 200 → 320+ tok/s on the same 8GB card. 1.6x, no new hardware, no new quant, just a KV cache trick stacked on top of the Gemma 4 26B MoE setup from a few days ago. A few days ago I posted Gemma 4 26B A4B hitting 28 tok/s decode on 8GB VRAM using native MTP. prefill was stuck around 200 tok/s. fair callout by the community. So today I tested something I'd already been meaning to try: TheTom/llama-cpp-turboquant, the TurboQuant KV cache fork by Tom Turney (Tom Turney). (github link in the comments) thanks to him, the fork just got resynced to mainline, so MTP + TurboQuant now run together cleanly (I didnt see any meaningful gains by using MTP with this setup though but you can try). The flags (No MTP): -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -cnv -c 64000 --cache-type-k q8_0 --cache-type-v turbo3 Results on the same RTX 4060 8GB, tested with a 27k token prompt at 64k context loaded: Prefill: 200 tok/s → 320+ tok/s Decode: stayed above 25 tok/s (without MTP) Why it works: TurboQuant uses walsh hadamard rotation + polar quantization on the KV cache. keys are sensitive to compression, values aren't much, so it splits the difference: K stays at q8_0, V drops to turbo3 (~3 bits). bonus from the memory savings: same 8GB card can now stretch to 100-120k context with minimal decode penalty. It should now be snappier with any agent harness such as hermes agent without compromise on intelligence. If you're already running Gemma 4 on a small card, this stacks on top for free. Try --cache-type-k q8_0 --cache-type-v turbo3 on your setup and report back what your prefill/decode split looks like. unsloth model gguf and llama.cpp turboquant fork links in the comments. what's your prefill number before vs after?

Alok

119,821 Aufrufe • vor 2 Monaten

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

271,839 Aufrufe • vor 1 Monat

I just crammed the updated Gemma 4 26B A4B QAT (MoE) with 180k context into an 8GB RTX 4060 (8 GB VRAM + 16 GB RAM only!!) and optimized the batch size. 23 tokens/sec decode, 300 tokens/sec prefill Yesterday I showed you a Gemma 4 31B dense model running flawlessly on an RTX 4090. Today, we're breaking the VRAM bank on a budget card using Unsloth’s new Gemma 4 26B (A4B) QAT quants. Following Google’s chat template update that boosted agentic benchmarks by +10%, I pushed this model to its absolute limits. Here is how you squeeze 250k context out of 8GB of VRAM. # The Setup & The Optimization - Hardware: Nvidia RTX 4060 (8GB VRAM) + 16GB System RAM - Environment: CUDA 13.0 build of llama.cpp - Model: gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf - Prompt: 28,000 tokens of prompt for each run If you read my L2 cache breakdown (attached in replies), you know the 4060’s 24MB cache maxes out at `-b 1024 -ub 1024`. Push past that, and prefill crashes. I locked those flags in for every test below to ensure maximum GEMM throughput. # 1. The Raw Context Push (Unquantized KV Cache) First, I wanted to see how far pure 8GB VRAM + 16GB RAM could stretch without touching the KV cache: - 80k Context: Prefill 385 t/s | Decode 25.5 t/s - 120k Context: Prefill 270 t/s | Decode 24 t/s llama.cpp flags: .\llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 120000 --port 8080 -ub 1024 -b 1024 Without KV quantization, 120k is your hard ceiling. push past that prefill throughput drops off a cliff, making the model practically unusable for large agentic workloads. # 2. The Q8 KV Cache Lifeline To survive 250k context on a budget card, you have to quantize the KV cache. I enabled 8 bit KV cache (`-ctk q8_0 -ctv q8_0`) and re ran: - 180k Context: Prefill 280 t/s | Decode 22.8 t/s - 250k Context: Prefill 115 t/s | Decode 20 t/s llama.cpp flags: .\llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 180000 --port 8080 -b 1024 -ub 1024 -ctk q8_0 -ctv q8_0 Result: Q8 KV cache brings 250k context back from the dead. Decode speed stabilizes at a highly usable 20 t/s. You are trading a very small bit amount of reasoning precision for an extra 130,000 tokens of context window. if you own a single rtx 3050, 3060, 3070, 4050, 4060, 5050 or 5060, you must try this model and optimize your batch size for higher prefill. Hugging Face links to the updated Unsloth's QAT quants and performance graph are in the replies below. What model are you running on your 6GB, 8GB or 12GB cards right now? Let's see your setups.
0:20

I just crammed the updated Gemma 4 26B A4B QAT (MoE) with 180k context into an 8GB RTX 4060 (8 GB VRAM + 16 GB RAM only!!) and optimized the batch size. 23 tokens/sec decode, 300 tokens/sec prefill Yesterday I showed you a Gemma 4 31B dense model running flawlessly on an RTX 4090. Today, we're breaking the VRAM bank on a budget card using Unsloth’s new Gemma 4 26B (A4B) QAT quants. Following Google’s chat template update that boosted agentic benchmarks by +10%, I pushed this model to its absolute limits. Here is how you squeeze 250k context out of 8GB of VRAM. # The Setup & The Optimization - Hardware: Nvidia RTX 4060 (8GB VRAM) + 16GB System RAM - Environment: CUDA 13.0 build of llama.cpp - Model: gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf - Prompt: 28,000 tokens of prompt for each run If you read my L2 cache breakdown (attached in replies), you know the 4060’s 24MB cache maxes out at `-b 1024 -ub 1024`. Push past that, and prefill crashes. I locked those flags in for every test below to ensure maximum GEMM throughput. # 1. The Raw Context Push (Unquantized KV Cache) First, I wanted to see how far pure 8GB VRAM + 16GB RAM could stretch without touching the KV cache: - 80k Context: Prefill 385 t/s | Decode 25.5 t/s - 120k Context: Prefill 270 t/s | Decode 24 t/s llama.cpp flags: .\llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 120000 --port 8080 -ub 1024 -b 1024 Without KV quantization, 120k is your hard ceiling. push past that prefill throughput drops off a cliff, making the model practically unusable for large agentic workloads. # 2. The Q8 KV Cache Lifeline To survive 250k context on a budget card, you have to quantize the KV cache. I enabled 8 bit KV cache (`-ctk q8_0 -ctv q8_0`) and re ran: - 180k Context: Prefill 280 t/s | Decode 22.8 t/s - 250k Context: Prefill 115 t/s | Decode 20 t/s llama.cpp flags: .\llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 180000 --port 8080 -b 1024 -ub 1024 -ctk q8_0 -ctv q8_0 Result: Q8 KV cache brings 250k context back from the dead. Decode speed stabilizes at a highly usable 20 t/s. You are trading a very small bit amount of reasoning precision for an extra 130,000 tokens of context window. if you own a single rtx 3050, 3060, 3070, 4050, 4060, 5050 or 5060, you must try this model and optimize your batch size for higher prefill. Hugging Face links to the updated Unsloth's QAT quants and performance graph are in the replies below. What model are you running on your 6GB, 8GB or 12GB cards right now? Let's see your setups.

Alok

36,617 Aufrufe • vor 1 Monat

A single RTX 4090 (24 GB VRAM) can run the updated gemma 4 31B (dense) model with a 190,000 context window at 33 tokens/second. The VRAM barrier is dying. Google quietly updated Gemma 4, and Unsloth immediately compiled the new quants. I built llama.cpp from source on Ubuntu 22 to benchmark it. Google's stealth update 2 days ago enabled uniform Flash Attention 4 on Hopper to boost prefill and patched the chat template to improve tool calling. The agentic reasoning gains on the benchmark charts are massive: TB2 (Agents): +4.5% (to 25.8%) Tau2 (Telecom): +10.1% (to 62.7%) Running on Ubuntu 22, CUDA 13.0 with a single NVIDIA GeForce RTX 4090. Here is the exact step by step benchmarking process with a massive 28k tokens prompt and the commands I used to squeeze out maximum context without killing my throughput: # 1. The Baseline (Unquantized KV Cache) I started with full GPU offload (-ngl 99) and pushed the context to 40k. llama.cpp flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -ngl 99 -c 40000 -fa on --port 8080 -v VRAM: 23.8 GB (maxed out on card) Throughput: Prefill: 2198.81 t/s | Decode: 35.77 t/s (with 28k tokens prompt) # 2. The CPU Split Trap I tried stretching to 80k context by offloading layers to the CPU (-ngl 52). llama.cpp flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 80000 -ngl 52 -fa on --port 8080 -v Throughput: Prefill: 1212.73 t/s | Decode: 5 t/s (with 28k tokens prompt) # 3. The KV Quantization Breakthrough Instead of spilling layers to the CPU, I kept the model fully on card (-ngl 99) but enabled 8-bit KV cache quantization to free up VRAM. flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 100000 --cache-type-k q8_0 --cache-type-v q8_0 -ngl 99 --port 8080 -v VRAM: 23.9 GB Throughput: Prefill: 2139.68 t/s | Decode: 32 t/s (with 28k tokens prompt) Result: 100k tokens of context on a single GPU with practically zero speed loss (and minimal intelligence loss). # 4. The Limit Test (Q4 KV Cache) To find the absolute breaking point, I dropped the KV cache to 4 bit (q4_0) and set -c 190000. flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 190000 --cache-type-k q4_0 --cache-type-v q4_0 -ngl 99 --port 8080 -v VRAM: 23.8 GB Throughput: Prefill: 2206.66 t/s | Decode: 33 t/s (with 28k tokens prompt) (Note: Pushing it to 220k required dropping to -ngl 58 again, which immediately penalized decode down to 17 t/s). # The Tradeoff: For Max Reasoning: Keep your KV cache unquantized (f16). You get pristine reasoning but hit a strict 40k context ceiling. For Massive Document Retrieval: If you need to feed the model giant codebases, use --cache-type-k q4_0. Getting 190k context at 33 tokens/second on a consumer desktop with a 31b dense model is a cheat code. If you’re rocking a single 3090 or 4090 and slept on Gemma 4 earlier, this update is your cue to dust off the terminal. Hugging Face links to the Unsloth QAT quants are in the replies below.
0:20

A single RTX 4090 (24 GB VRAM) can run the updated gemma 4 31B (dense) model with a 190,000 context window at 33 tokens/second. The VRAM barrier is dying. Google quietly updated Gemma 4, and Unsloth immediately compiled the new quants. I built llama.cpp from source on Ubuntu 22 to benchmark it. Google's stealth update 2 days ago enabled uniform Flash Attention 4 on Hopper to boost prefill and patched the chat template to improve tool calling. The agentic reasoning gains on the benchmark charts are massive: TB2 (Agents): +4.5% (to 25.8%) Tau2 (Telecom): +10.1% (to 62.7%) Running on Ubuntu 22, CUDA 13.0 with a single NVIDIA GeForce RTX 4090. Here is the exact step by step benchmarking process with a massive 28k tokens prompt and the commands I used to squeeze out maximum context without killing my throughput: # 1. The Baseline (Unquantized KV Cache) I started with full GPU offload (-ngl 99) and pushed the context to 40k. llama.cpp flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -ngl 99 -c 40000 -fa on --port 8080 -v VRAM: 23.8 GB (maxed out on card) Throughput: Prefill: 2198.81 t/s | Decode: 35.77 t/s (with 28k tokens prompt) # 2. The CPU Split Trap I tried stretching to 80k context by offloading layers to the CPU (-ngl 52). llama.cpp flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 80000 -ngl 52 -fa on --port 8080 -v Throughput: Prefill: 1212.73 t/s | Decode: 5 t/s (with 28k tokens prompt) # 3. The KV Quantization Breakthrough Instead of spilling layers to the CPU, I kept the model fully on card (-ngl 99) but enabled 8-bit KV cache quantization to free up VRAM. flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 100000 --cache-type-k q8_0 --cache-type-v q8_0 -ngl 99 --port 8080 -v VRAM: 23.9 GB Throughput: Prefill: 2139.68 t/s | Decode: 32 t/s (with 28k tokens prompt) Result: 100k tokens of context on a single GPU with practically zero speed loss (and minimal intelligence loss). # 4. The Limit Test (Q4 KV Cache) To find the absolute breaking point, I dropped the KV cache to 4 bit (q4_0) and set -c 190000. flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 190000 --cache-type-k q4_0 --cache-type-v q4_0 -ngl 99 --port 8080 -v VRAM: 23.8 GB Throughput: Prefill: 2206.66 t/s | Decode: 33 t/s (with 28k tokens prompt) (Note: Pushing it to 220k required dropping to -ngl 58 again, which immediately penalized decode down to 17 t/s). # The Tradeoff: For Max Reasoning: Keep your KV cache unquantized (f16). You get pristine reasoning but hit a strict 40k context ceiling. For Massive Document Retrieval: If you need to feed the model giant codebases, use --cache-type-k q4_0. Getting 190k context at 33 tokens/second on a consumer desktop with a 31b dense model is a cheat code. If you’re rocking a single 3090 or 4090 and slept on Gemma 4 earlier, this update is your cue to dust off the terminal. Hugging Face links to the Unsloth QAT quants are in the replies below.

Alok

76,069 Aufrufe • vor 1 Monat

nvidia's 3B mamba destroyed alibaba's 3B deltanet on the same RTX 3090. only 24 days between releases. same active parameters, same VRAM tier, completely different architectures. nemotron cascade 2: 187 tok/s. flat from 4K to 625K context. zero speed loss. flags: -ngl 99 -np 1. that's it. no context flags, no KV cache tricks. auto-allocates 625K. qwen 3.5 35B-A3B: 112 tok/s. flat from 4K to 262K context. zero speed loss. flags: -ngl 99 -np 1 -c 262144 --cache-type-k q8_0 --cache-type-v q8_0. needed KV cache quantization to fit 262K. both models held a flat line across every context level. both architectures are context-independent. but nvidia's mamba2 is 67% faster at generating tokens on the exact same hardware and needs fewer flags to get there. same node, same GPU, same everything. the only variable is the model. gold medal math olympiad winner running at 187 tokens per second on single RTX 3090 a card from 6 years ago. nvidia cooked.
0:36

nvidia's 3B mamba destroyed alibaba's 3B deltanet on the same RTX 3090. only 24 days between releases. same active parameters, same VRAM tier, completely different architectures. nemotron cascade 2: 187 tok/s. flat from 4K to 625K context. zero speed loss. flags: -ngl 99 -np 1. that's it. no context flags, no KV cache tricks. auto-allocates 625K. qwen 3.5 35B-A3B: 112 tok/s. flat from 4K to 262K context. zero speed loss. flags: -ngl 99 -np 1 -c 262144 --cache-type-k q8_0 --cache-type-v q8_0. needed KV cache quantization to fit 262K. both models held a flat line across every context level. both architectures are context-independent. but nvidia's mamba2 is 67% faster at generating tokens on the exact same hardware and needs fewer flags to get there. same node, same GPU, same everything. the only variable is the model. gold medal math olympiad winner running at 187 tokens per second on single RTX 3090 a card from 6 years ago. nvidia cooked.

Sudo su

187,008 Aufrufe • vor 5 Monaten

testing Qwen3.5-35B-A3B latest optimized version by UnslothAI on a single RTX 3090. one detailed prompt. zero handholding. watch a 3B model scaffold an entire multifile game project autonomously. the setup: > model: Qwen3.5-35B-A3B (80B total, only 3B active per token) > quant: UD-Q4_K_XL by Unsloth (MXFP4 layers removed in latest update) > speed: 112 tok/s generation, ~130 tok/s prefill > context: 262K tokens > flags: -ngl 99 -c 262144 -np 1 --cache-type-k q8_0 --cache-type-v q8_0 > engine: llama.cpp > agent: Claude Code talk to localhost:8080 (llama.cpp now has native Anthropic API endpoint. no LiteLLM needed) q8_0 KV cache cuts VRAM usage in half vs f16 at 262K. -np 1 is default but worth noting. parallel slots multiply KV cache and at 262K that's an instant OOM. the prompt was more detailed than this but you get the idea: build a space shooter with parallax backgrounds, particle systems, procedural audio, 4 enemy types, boss fights, power-up system, and ship upgrades. 8 JavaScript modules. no libraries. game's called Octopus Invaders. gameplay footage dropping next.
0:50

testing Qwen3.5-35B-A3B latest optimized version by UnslothAI on a single RTX 3090. one detailed prompt. zero handholding. watch a 3B model scaffold an entire multifile game project autonomously. the setup: > model: Qwen3.5-35B-A3B (80B total, only 3B active per token) > quant: UD-Q4_K_XL by Unsloth (MXFP4 layers removed in latest update) > speed: 112 tok/s generation, ~130 tok/s prefill > context: 262K tokens > flags: -ngl 99 -c 262144 -np 1 --cache-type-k q8_0 --cache-type-v q8_0 > engine: llama.cpp > agent: Claude Code talk to localhost:8080 (llama.cpp now has native Anthropic API endpoint. no LiteLLM needed) q8_0 KV cache cuts VRAM usage in half vs f16 at 262K. -np 1 is default but worth noting. parallel slots multiply KV cache and at 262K that's an instant OOM. the prompt was more detailed than this but you get the idea: build a space shooter with parallax backgrounds, particle systems, procedural audio, 4 enemy types, boss fights, power-up system, and ship upgrades. 8 JavaScript modules. no libraries. game's called Octopus Invaders. gameplay footage dropping next.

Sudo su

167,244 Aufrufe • vor 5 Monaten

I implemented Google Research's TurboQuant as a CUDA-native compression engine on Blackwell B200. 5x KV cache compression on Qwen 2.5-1.5B, near-loseless attention scores, generating live from compressed memory. 5 custom cuTile CUDA kernels ft: - fused attention (with QJL corrections) - online softmax -on-chip cache decompression - pipelined TMA loads Try it out: s/o Bryce, the CUDA Colonel and the cuTile team at NVIDIA for lending me Blackwell GPU access :) cc sunny madra Gavin
4:46

I implemented Google Research's TurboQuant as a CUDA-native compression engine on Blackwell B200. 5x KV cache compression on Qwen 2.5-1.5B, near-loseless attention scores, generating live from compressed memory. 5 custom cuTile CUDA kernels ft: - fused attention (with QJL corrections) - online softmax -on-chip cache decompression - pipelined TMA loads Try it out: s/o Bryce, the CUDA Colonel and the cuTile team at NVIDIA for lending me Blackwell GPU access :) cc sunny madra Gavin

ani

810,322 Aufrufe • vor 4 Monaten

The Cost of Intelligence is Heading to Zero | Hyperspace P2P Distributed Cache We present to you our breakthrough cross-domain work across AI, distributed systems, cryptography, game theory to solve the primary structural inefficiency at the heart of AI infrastructure: most inference is redundant. Google has reported that only 15% of daily searches are truly novel. The rest are repeats or close variants. LLM inference inherits this same power-law distribution. Enterprise chatbots see 70-80% of queries fall into a handful of intent categories. System prompts are identical across 100% of requests within an application. The KV attention state for "You are a helpful assistant" has been computed billions of times, on millions of GPUs, identically. And yet every AI lab, every startup, every self-hosted deployment - computes and caches these results independently. There is no shared layer. No global memory. Every provider pays the full compute cost for every query, even when the answer already exists somewhere in the network. This is the problem Hyperspace solves where distributed cache operates at three levels, each catching a different class of redundancy: 1. Response cache Same prompt, same model, same parameters - instant cached response from any node in the network. SHA-256 hash lookup via DHT, with cryptographic cache proofs linking every response to its original inference execution. No trust required. Fetchers re-announce as providers, so popular responses replicate naturally across more nodes. 2. KV prefix cache Same system prompt tokens - skip the most expensive part of inference entirely. Prefill (computing Key-Value attention states) is deterministic: same model plus same tokens always produces identical KV state. The network caches these states using erasure coding and distributes them via the routing network. New questions that share a common prefix resume generation from cached state instead of recomputing from scratch. 3. Routing to cached nodes Instead of transferring KV state across the network for every request, Hyperspace routes the request to the node that already has the state loaded in VRAM. The request goes to the cache, not the cache to the request. Together, these three layers mean that 70-90% of inference requests at network scale never require full GPU computation. This work doesn't exist in isolation. It builds on research from across the industry: SGLang's RadixAttention demonstrated that automatic prefix sharing can yield up to 5x speedup on structured LLM workloads. Moonshot AI's Mooncake built an entire KV-cache-centric disaggregated architecture for production serving at Kimi. Anthropic, OpenAI, and Google all launched prompt caching products in 2024 - priced at 50-90% discounts - because system prompt reuse is so pervasive that it changes the economics of inference. What all of these systems share is a common limitation: they operate within a single organization's infrastructure. SGLang caches prefixes within one server. Mooncake disaggregates KV cache within one datacenter. Anthropic's prompt caching works within one API provider's fleet. None of them can share cached state across organizational boundaries. Hyperspace removes this boundary. The cache is global. A response computed by a node in Tokyo is immediately available to a node in Berlin. A KV prefix state generated for Qwen-32B on one machine is verifiable and reusable by any other machine running the same model. The routing network provides the delivery guarantees, the erasure coding provides the redundancy, and the cache proofs provide the trust. What this means for the cost of intelligence Big AI labs scale linearly: twice the users means twice the GPU spend. Every query is a cost center. Their internal caching helps, but it's siloed - Lab A's cache can't serve Lab B's users, and neither can serve a self-hosted Llama deployment. Hyperspace scales sub-linearly. Every new node that joins the network adds to the global cache. Every inference result enriches the cache for all future requests. The cache hit rate rises with network size because query distributions follow a power law - the most common questions are asked exponentially more often than rare ones. The implication is simple: as the network grows, the effective cost per inference drops. Not linearly. Logarithmically. At 10 million nodes, we estimate 75-90% of all inference requests can be served from cache, eliminating 400,000+ MWh of energy consumption per year and avoiding over 200,000 tons of CO2 emissions. The first person to ask a question pays the compute cost. Everyone after them gets the answer for free, with cryptographic proof that it's authentic. Training is competitive. Inference is shared Open-weight models are converging on quality with closed models. Labs will continue to differentiate on training - data curation, architecture innovation, RLHF tuning. That's where the real intellectual property lives. But inference is a commodity. Two copies of Qwen-32B running the same prompt produce the same KV state and the same response, byte for byte, regardless of whose GPU runs the matrix multiplication. There is no moat in multiplying matrices. The moat is in training the weights. A global distributed cache makes this separation explicit. It doesn't matter who trained the model. Once the weights are open, the inference cost approaches zero at scale - because the network remembers every answer and can prove it's correct. No lab, no matter how well-funded, can match this. They cannot share caches across competitors. They scale linearly. The network scales logarithmically. The marginal cost of intelligence approaches zero. That's the endgame.
0:23

The Cost of Intelligence is Heading to Zero | Hyperspace P2P Distributed Cache We present to you our breakthrough cross-domain work across AI, distributed systems, cryptography, game theory to solve the primary structural inefficiency at the heart of AI infrastructure: most inference is redundant. Google has reported that only 15% of daily searches are truly novel. The rest are repeats or close variants. LLM inference inherits this same power-law distribution. Enterprise chatbots see 70-80% of queries fall into a handful of intent categories. System prompts are identical across 100% of requests within an application. The KV attention state for "You are a helpful assistant" has been computed billions of times, on millions of GPUs, identically. And yet every AI lab, every startup, every self-hosted deployment - computes and caches these results independently. There is no shared layer. No global memory. Every provider pays the full compute cost for every query, even when the answer already exists somewhere in the network. This is the problem Hyperspace solves where distributed cache operates at three levels, each catching a different class of redundancy: 1. Response cache Same prompt, same model, same parameters - instant cached response from any node in the network. SHA-256 hash lookup via DHT, with cryptographic cache proofs linking every response to its original inference execution. No trust required. Fetchers re-announce as providers, so popular responses replicate naturally across more nodes. 2. KV prefix cache Same system prompt tokens - skip the most expensive part of inference entirely. Prefill (computing Key-Value attention states) is deterministic: same model plus same tokens always produces identical KV state. The network caches these states using erasure coding and distributes them via the routing network. New questions that share a common prefix resume generation from cached state instead of recomputing from scratch. 3. Routing to cached nodes Instead of transferring KV state across the network for every request, Hyperspace routes the request to the node that already has the state loaded in VRAM. The request goes to the cache, not the cache to the request. Together, these three layers mean that 70-90% of inference requests at network scale never require full GPU computation. This work doesn't exist in isolation. It builds on research from across the industry: SGLang's RadixAttention demonstrated that automatic prefix sharing can yield up to 5x speedup on structured LLM workloads. Moonshot AI's Mooncake built an entire KV-cache-centric disaggregated architecture for production serving at Kimi. Anthropic, OpenAI, and Google all launched prompt caching products in 2024 - priced at 50-90% discounts - because system prompt reuse is so pervasive that it changes the economics of inference. What all of these systems share is a common limitation: they operate within a single organization's infrastructure. SGLang caches prefixes within one server. Mooncake disaggregates KV cache within one datacenter. Anthropic's prompt caching works within one API provider's fleet. None of them can share cached state across organizational boundaries. Hyperspace removes this boundary. The cache is global. A response computed by a node in Tokyo is immediately available to a node in Berlin. A KV prefix state generated for Qwen-32B on one machine is verifiable and reusable by any other machine running the same model. The routing network provides the delivery guarantees, the erasure coding provides the redundancy, and the cache proofs provide the trust. What this means for the cost of intelligence Big AI labs scale linearly: twice the users means twice the GPU spend. Every query is a cost center. Their internal caching helps, but it's siloed - Lab A's cache can't serve Lab B's users, and neither can serve a self-hosted Llama deployment. Hyperspace scales sub-linearly. Every new node that joins the network adds to the global cache. Every inference result enriches the cache for all future requests. The cache hit rate rises with network size because query distributions follow a power law - the most common questions are asked exponentially more often than rare ones. The implication is simple: as the network grows, the effective cost per inference drops. Not linearly. Logarithmically. At 10 million nodes, we estimate 75-90% of all inference requests can be served from cache, eliminating 400,000+ MWh of energy consumption per year and avoiding over 200,000 tons of CO2 emissions. The first person to ask a question pays the compute cost. Everyone after them gets the answer for free, with cryptographic proof that it's authentic. Training is competitive. Inference is shared Open-weight models are converging on quality with closed models. Labs will continue to differentiate on training - data curation, architecture innovation, RLHF tuning. That's where the real intellectual property lives. But inference is a commodity. Two copies of Qwen-32B running the same prompt produce the same KV state and the same response, byte for byte, regardless of whose GPU runs the matrix multiplication. There is no moat in multiplying matrices. The moat is in training the weights. A global distributed cache makes this separation explicit. It doesn't matter who trained the model. Once the weights are open, the inference cost approaches zero at scale - because the network remembers every answer and can prove it's correct. No lab, no matter how well-funded, can match this. They cannot share caches across competitors. They scale linearly. The network scales logarithmically. The marginal cost of intelligence approaches zero. That's the endgame.

Varun

37,555 Aufrufe • vor 5 Monaten

first impressions of qwen 3.5 27B dense on a single RTX 3090. 35 tok/s. from 4K all the way to 300K+ context. no speed drop. hermes 4.3 started at 35 and degraded to 15 as context filled. qwen dense holds. MoE held 112 flat. 3x faster but only 3B of 35B active per token. architecture tradeoff. Q4_K_M on 16.7GB. native context 262K. pushed past training limit to 376K before VRAM ceiling on 24GB. tried q8 KV cache at 262K, speed collapsed to 11 tok/s. q4_0 KV is the sweet spot. flash attention mandatory. built in reasoning mode. the model thinks step by step before it answers. full chain of thought surviving Q4 quant. 1,799+ token thinking chains with self correction loops. on a single consumer GPU. gave it one prompt: "build a realtime particle galaxy simulation in one HTML file." 3,340 tokens. 95 seconds. one shot. ran on first load. full reasoning and coding in the video below. optimal config if you want to skip the hours of testing: llama-server -ngl 99 -c 262144 -fa on --cache-type-k q4_0 --cache-type-v q4_0 this is just the warmup. octopus invaders is next: 10 files, 3,400+ lines, zero steering. the prompt hermes quit at 22%. already more impressed than expected. full results coming soon.
2:02

first impressions of qwen 3.5 27B dense on a single RTX 3090. 35 tok/s. from 4K all the way to 300K+ context. no speed drop. hermes 4.3 started at 35 and degraded to 15 as context filled. qwen dense holds. MoE held 112 flat. 3x faster but only 3B of 35B active per token. architecture tradeoff. Q4_K_M on 16.7GB. native context 262K. pushed past training limit to 376K before VRAM ceiling on 24GB. tried q8 KV cache at 262K, speed collapsed to 11 tok/s. q4_0 KV is the sweet spot. flash attention mandatory. built in reasoning mode. the model thinks step by step before it answers. full chain of thought surviving Q4 quant. 1,799+ token thinking chains with self correction loops. on a single consumer GPU. gave it one prompt: "build a realtime particle galaxy simulation in one HTML file." 3,340 tokens. 95 seconds. one shot. ran on first load. full reasoning and coding in the video below. optimal config if you want to skip the hours of testing: llama-server -ngl 99 -c 262144 -fa on --cache-type-k q4_0 --cache-type-v q4_0 this is just the warmup. octopus invaders is next: 10 files, 3,400+ lines, zero steering. the prompt hermes quit at 22%. already more impressed than expected. full results coming soon.

Sudo su

122,051 Aufrufe • vor 5 Monaten

after turboquant and qwen3.5-35b-a3b, i got curious: how realistic is it to use kv cache as a document store today? to have vectorless, RAG-less search. so i prefilled 258K out of 262K context window on L4 (a budget GPU popular in prod). ~99% of the slot is pre-computed and stored, users load it on the fly in ~1s. system prompt + query append to the end, generation takes ~3K tokens, enough for search. at 99% fill rate, decoding runs ~20 tps on L4. i prepared some ego datasets (jina papers, which i know best), plus popular novels in chinese and english. the results are actually pretty good. some hallucination, but most answers are solid and well-grounded. what's more interesting is the cost: ~$0.26/h on L4 spot. single LLM. no vector database, no embedding model, no workflow/pipeline engineering. using kv cache as document store is nothing new, like the old CAG paper. but with quantized kv cache and modern attention (hybrid SSM-attention, GQA, MQA, MLA), the economics are changing fast. if we solve cold-prefill speed and decoding speed, and budget GPU costs keep dropping, the future of search could be vectorless. radical, but possible.
1:01

after turboquant and qwen3.5-35b-a3b, i got curious: how realistic is it to use kv cache as a document store today? to have vectorless, RAG-less search. so i prefilled 258K out of 262K context window on L4 (a budget GPU popular in prod). ~99% of the slot is pre-computed and stored, users load it on the fly in ~1s. system prompt + query append to the end, generation takes ~3K tokens, enough for search. at 99% fill rate, decoding runs ~20 tps on L4. i prepared some ego datasets (jina papers, which i know best), plus popular novels in chinese and english. the results are actually pretty good. some hallucination, but most answers are solid and well-grounded. what's more interesting is the cost: ~$0.26/h on L4 spot. single LLM. no vector database, no embedding model, no workflow/pipeline engineering. using kv cache as document store is nothing new, like the old CAG paper. but with quantized kv cache and modern attention (hybrid SSM-attention, GQA, MQA, MLA), the economics are changing fast. if we solve cold-prefill speed and decoding speed, and budget GPU costs keep dropping, the future of search could be vectorless. radical, but possible.

Han Xiao

42,442 Aufrufe • vor 4 Monaten

SITUATION EXPLAINED: Cerebras raised $5.55 billion in their IPO and closing their first day of trading valued at $66 billion, making it the biggest US tech IPO since Snowflake in 2020. Cerebras makes Wafer-Scale Engine chips built for AI inference. We asked Sarah Fong the main difference between wafer-scale chips and traditional GPUs: - GPUs are great at parallel work (graphics, training) - AI inference is sequential, AKA one token at a time This causes the "memory wall" problem: - Every GPU core needs model weights, KV cache, and activations to do its math - On a GPU, that data lives in off-chip memory (HBM) - Cores constantly load and offload from off-chip memory, which is a huge bottleneck; hardware accounts for ~70% of inference latency Cerebras' chips: -Dinner-plate sized (vs. GPUs which are palm-sized) with tens of thousands of cores -Memory sits directly on top of the cores as distributed SRAM -Weights and KV cache can be accessed at on-chip speeds in the PB/s range, compared with off-chip speeds in the TB/s range achieved by GPUs with HBM.
3:14

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

MTS

44,057 Aufrufe • vor 3 Monaten

🚀 Excited to release LongLive 2.0! 🎬 An end-to-end infrastructure for long video generation, with FP4 and parallelism at the core of both training and inference. ⚡45.7 FPS generation speed on 5B model⚡ ✨ LongLive 2.0 supports real-video training, few-step distillation, multi-shot training/inference, sequence-parallel acceleration, NVFP4 KV cache, and async VAE decoding deployment. 🧩 To our knowledge, this is the first open-source 4-bit long video generation infra that covers both training and inference. 🙌 Welcome to check it out, try it, and share feedback! 🔗 Code: 📰 Paper: 🎥 Demo: #LongVideoGeneration #VideoGeneration #Realtime #AIInfra #EfficientAI #FP4 #Parallel #NVIDIA
1:31

🚀 Excited to release LongLive 2.0! 🎬 An end-to-end infrastructure for long video generation, with FP4 and parallelism at the core of both training and inference. ⚡45.7 FPS generation speed on 5B model⚡ ✨ LongLive 2.0 supports real-video training, few-step distillation, multi-shot training/inference, sequence-parallel acceleration, NVFP4 KV cache, and async VAE decoding deployment. 🧩 To our knowledge, this is the first open-source 4-bit long video generation infra that covers both training and inference. 🙌 Welcome to check it out, try it, and share feedback! 🔗 Code: 📰 Paper: 🎥 Demo: #LongVideoGeneration #VideoGeneration #Realtime #AIInfra #EfficientAI #FP4 #Parallel #NVIDIA

Yukang Chen

59,280 Aufrufe • vor 3 Monaten

New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys LMSYS Org and RadixArk RadixArk, and taught by Richard Chen Richard Chen, a Member of Technical Staff at RadixArk. Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests. When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests. Skills you'll gain: - Implement a KV cache from scratch to eliminate redundant computation within a single request - Scale caching across users and requests with RadixAttention, so shared context is only processed once - Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism Join and learn to make LLM inference faster and more cost-efficient at scale!
1:51

New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys LMSYS Org and RadixArk RadixArk, and taught by Richard Chen Richard Chen, a Member of Technical Staff at RadixArk. Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests. When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests. Skills you'll gain: - Implement a KV cache from scratch to eliminate redundant computation within a single request - Scale caching across users and requests with RadixAttention, so shared context is only processed once - Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism Join and learn to make LLM inference faster and more cost-efficient at scale!

Andrew Ng

101,713 Aufrufe • vor 4 Monaten

QVAC SDK 0.10.0 is now live, bringing advanced local compute capabilities and specialized hardware optimization directly to your device Key Features and Updates: - Image-to-Image Diffusion: Transform and edit images using simple prompts with 100% local compute—no cloud uploads or external servers required - Dynamic Tooling & KV Cache Management:Your local LLM now receives a tailored toolbox for every interaction, with automatic KV cache clearing to maintain high-speed inference - Doctor CLI: A new diagnostic tool that analyzes your hardware and memory, providing specific instructions on how to optimize your GPU for local AI - Suspend & Resume API: Specifically designed for mobile environments, this allows apps to pause P2P swarms and RAG workspaces to meet background rules without losing model state - GPT-OSS Compatibility: Added support for the latest GPT-OSS models loaded externally, expanding the range of open-source intelligence available on the platform Build the future of private, unstoppable AI:
1:11

QVAC SDK 0.10.0 is now live, bringing advanced local compute capabilities and specialized hardware optimization directly to your device Key Features and Updates: - Image-to-Image Diffusion: Transform and edit images using simple prompts with 100% local compute—no cloud uploads or external servers required - Dynamic Tooling & KV Cache Management:Your local LLM now receives a tailored toolbox for every interaction, with automatic KV cache clearing to maintain high-speed inference - Doctor CLI: A new diagnostic tool that analyzes your hardware and memory, providing specific instructions on how to optimize your GPU for local AI - Suspend & Resume API: Specifically designed for mobile environments, this allows apps to pause P2P swarms and RAG workspaces to meet background rules without losing model state - GPT-OSS Compatibility: Added support for the latest GPT-OSS models loaded externally, expanding the range of open-source intelligence available on the platform Build the future of private, unstoppable AI:

QVAC

34,508 Aufrufe • vor 3 Monaten

$NVDA $MU $SNDK $LITE PAPER OVERVIEW AND CORE CLAIMS The paper “KV Cache Transform Coding for Compact Storage in LLM Inference” introduces kvtc, a transform-coding pipeline that compresses transformer key-value (KV) caches primarily for storage and transfer in LLM serving, rather than for accelerating the per-token attention kernel during active decoding. The method combines 3 stages: (1) feature decorrelation via a PCA basis computed from a calibration dataset and reused across requests; (2) adaptive, variable-precision quantization with bit allocation solved via dynamic programming (DP), including groupwise scaling/shift overhead; and (3) lossless entropy coding (DEFLATE via nvCOMP in the reference implementation) to exploit residual redundancy after quantization. The central empirical claim is that KV tensors contain large, exploitable redundancy across heads and layers, enabling approximately 20× compression versus a 16-bit baseline with negligible degradation across a broad set of accuracy and long-context benchmarks, with materially higher compression (≥40×) available at modest quality cost in some regimes. The system claim is that such compression materially improves the economics of multi-turn, prefix-reuse serving by extending effective KV cache capacity in GPU HBM and host tiers (DRAM/NVMe) and by reducing inter-node and GPU↔host bandwidth demands, thereby improving cache hit rates and reducing time-to-first-token (TTFT) relative to recomputation when caches would otherwise be evicted. KV CACHE AS THE DOMINANT STATE VARIABLE IN INFERENCE ECONOMICS KV cache growth is linear in context length and is multiplicative in layers and attention heads, making it an increasingly dominant constraint as (a) context lengths expand, (b) models add layers and maintain large hidden dimensions, and (c) production workloads shift toward iterative and tool-augmented interactions that repeatedly reuse long prefixes. The paper uses the canonical 16-bit KV cache size formula (4·l·h·d_head·t) bytes and reports 16-bit KV cache sizes per 1K tokens of context that are already operationally large: 128MiB for Llama 3.1 8B, 160MiB for Mistral NeMo 12B, and 320MiB for Llama 3.3 70B Instruct. In binary units, these figures imply per-token KV footprints of 128KiB/token (Llama 3.1 8B), 160KiB/token (Mistral NeMo 12B), and 320KiB/token (Llama 3.3 70B Instruct) at 16-bit. For a 10K-token prompt (10×1K in the paper’s binary convention), the 16-bit KV cache sizes scale to approximately 1.25GiB (Llama 3.1 8B), 1.56GiB (Mistral NeMo 12B), and 3.13GiB (Llama 3.3 70B Instruct). These magnitudes explain why stale caches create a throughput–latency dilemma: retaining them in HBM maximizes responsiveness on future turns but crowds out concurrent sessions; evicting them forces quadratic-cost prefill recomputation and increases TTFT; offloading them to host or storage introduces large transfer overhead and consumes DRAM/NVMe capacity. A key operational nuance emphasized is that modern serving stacks increasingly treat KV caches as a database, leveraging block paging and shared-prefix reuse. In the common disaggregated serving design (separate prefill and decode nodes), KV cache transfer becomes a dominant category of cross-node traffic. Under that design, any reduction in KV cache size directly increases effective fabric capacity and reduces tail latency attributable to congestion, while also enabling longer cache lifetimes in “hot” (HBM) and “warm” (CPU DRAM) tiers that raise cache hit rates and reduce recomputation frequency. The paper’s quantitative example illustrates the economic stakes: a 1,000-line code file tokenized at ~10 tokens/line yields ~10K tokens; for Llama 3.3 70B, an 8-bit KV cache for that context is ~1.6GiB. Reuse across subsequent turns or parallel chats around the same file is valuable, but HBM scarcity makes retaining many such caches infeasible without compression. TECHNICAL MECHANISM: WHY KV CACHES ARE COMPRESSIBLE AND HOW KVTC EXPLOITS IT The technical rationale begins with an empirical observation: keys (and, to a lesser extent, values) across different attention heads can be aligned into a shared latent space using orthogonal transformations (Procrustes alignment). This supports the hypothesis that head-specific projections introduce rotations of a common subspace rather than completely distinct information, implying that concatenating across heads and layers should reveal low-rank structure suitable for linear decorrelation and dimensionality reduction. The method operationalizes this using a PCA/SVD basis learned from calibration data rather than recomputing a decomposition per prompt. This design choice targets production viability: per-prompt SVD is computationally expensive and scales poorly with long prompts and frequent cache updates. kvtc is explicitly structured as an offline-calibrated, online-applied codec: Calibration (performed 1 time per model and compression setting for DP allocation) A calibration dataset is forwarded through the model to collect KV caches. Token positions are pooled, and a subset of positions is sampled. Keys and values are processed separately. Several implementation choices are highlighted as decisive for stability: Rotary positional embeddings are effectively removed prior to compression (“undo positional rotations”), because positional rotations degrade the apparent low-rank structure of keys. “Attention sink” tokens (the earliest tokens in the sequence) and a sliding window of most recent tokens are excluded from compression because they disproportionately affect attention patterns and are empirically more sensitive to reconstruction error. Cross-layer concatenation is used: keys (or values) from multiple layers and heads at the same token position are concatenated along the feature axis to form a higher-dimensional feature vector. PCA is computed over these concatenated vectors, improving robustness relative to per-layer or per-head PCA. The PCA basis is computed via SVD of centered calibration data, using randomized SVD for scalability with a target rank cutoff. The paper reports calibration regimes of 160K tokens for several models with a 10K PCA dimension cutoff (8K for Qwen variants with fewer KV heads), selected to fit within a single 80GB H100 memory envelope and complete within minutes. A critical economic detail is that the same PCA basis can be reused across multiple compression ratios; only the DP-derived precision assignment changes per compression target. Compression (applied between inference phases) Compression operates on stored KV cache tensors, not on weights, and does not modify attention computation. The KV cache is projected into the PCA basis, quantized, packed, and then entropy-coded. Compression is positioned as a background or between-phase operation (after decoding, or between prefill and decode), executed on GPU or CPU depending on where the cache currently resides. The design intent is that compression should not sit on the critical per-token decoding path; it is a storage and transport optimization. Decompression (performed prior to reuse) Decompression reverses the entropy coding and quantization and applies the inverse PCA projection. A practical latency optimization is proposed: inverse projection can be performed layer-by-layer using submatrices of the PCA basis, allowing generation to begin before the full cache is reconstructed, reducing TTFT. Quantization and bit allocation are the core differentiators versus simpler PCA truncation. PCA provides ordered components by variance; kvtc uses DP to allocate a global bit budget across PCA coordinates (and across groups of coordinates) to minimize reconstruction error in the decorrelated domain. Groups of subsequent PCA coordinates share 16-bit shift and scale factors (a microscaling-inspired design), and the DP algorithm jointly selects group size and precision type under a bit budget, including the overhead of per-group metadata. DP commonly assigns 0 bits to many trailing PCA components, which both increases compression and provides a mechanism to trim the PCA basis to the subset of components that actually carry payload, reducing compute and storage overhead of the projection matrices in deployment. Lossless entropy coding then exploits the structure induced by quantization. DEFLATE is used in the reference implementation, and the paper emphasizes that the incremental gain from the lossless stage is content-dependent but meaningful, with an average uplift of ~1.23× on top of quantization in the reported regime. An ablation in the appendices indicates that GPU-friendly variants (GDeflate) can achieve nearly identical compression ratios (≤0.1 difference in measured cases), implying that throughput-optimized lossless codecs can likely be substituted without sacrificing meaningful compression. EMPIRICAL RESULTS: ACCURACY, COMPRESSION, AND LATENCY General-purpose 8B–12B dense models The paper evaluates Llama 3.1 8B, MN-Minitron 8B, and Mistral NeMo 12B across math/knowledge (GSM8K, MMLU) and long-context tasks (Qasper, Lost in the Middle, RULER Variable Tracking) under a simulated multi-turn regime where compression/decompression is applied periodically, with a sliding window of recent tokens excluded. A consistent pattern appears: kvtc maintains near-vanilla performance through 16× compression settings, and remains competitive at 32×, with degradation becoming task- and model-dependent at 64×, particularly on long-context retrieval metrics when compression is pushed aggressively. Selected quantitative anchor points from the paper’s standard-error table (all values are reported with the paper’s evaluation setup and token-window exclusions): Llama 3.1 8B Vanilla: GSM8K 56.8, MMLU 60.5, Qasper 40.4, LITM 99.4, RULER-VT 99.8 kvtc16×: GSM8K 56.9, MMLU 60.1, Qasper 40.7, LITM 99.3, RULER-VT 99.1 kvtc32×: GSM8K 57.8, MMLU 60.6, Qasper 39.4, LITM 99.1, RULER-VT 98.9 kvtc64×: GSM8K 57.2, MMLU 60.7, Qasper 37.8, LITM 90.2, RULER-VT 95.9 These results indicate that, for this model, long-context sensitivity emerges at 64× with meaningful drops in LITM and RULER-VT, while math/knowledge scores remain stable, implying a differential sensitivity consistent with key-vector precision being more critical for retrieval-style behavior. Mistral NeMo 12B Vanilla: GSM8K 61.9, MMLU 64.5, Qasper 38.4, LITM 99.5, RULER-VT 99.8 kvtc16×: GSM8K 62.0, MMLU 64.4, Qasper 37.6, LITM 99.8, RULER-VT 99.5 kvtc32×: GSM8K 62.2, MMLU 63.8, Qasper 37.5, LITM 99.6, RULER-VT 98.7 kvtc64×: GSM8K 61.9, MMLU 61.4, Qasper 38.0, LITM 95.3, RULER-VT 98.0 Here, degradation at 64× is visible but materially smaller than the Llama 3.1 8B LITM drop, suggesting model-architecture or training-data differences can change the tolerance envelope for aggressive KV cache distortion. MN-Minitron 8B Vanilla: GSM8K 59.1, MMLU 64.3, Qasper 38.2, LITM 99.8, RULER-VT 99.4 kvtc16×: GSM8K 60.3, MMLU 64.1, Qasper 38.6, LITM 99.3, RULER-VT 98.8 kvtc32×: GSM8K 59.1, MMLU 63.7, Qasper 37.7, LITM 86.9, RULER-VT 96.0 kvtc64×: GSM8K 57.8, MMLU 62.1, Qasper 38.1, LITM 59.5, RULER-VT 93.4 This model shows markedly higher sensitivity on LITM at 32× and 64×, despite stable short-context metrics, reinforcing that “compression safety” is not monotonic in parameter count and that pruning/distillation choices can alter KV cache redundancy or robustness. Comparisons to baselines The paper compares kvtc to quantization baselines (KIVI, GEAR, FP8) and eviction baselines (H2O, TOVA), plus an SVD-based prefill-optimization method (xKV). Across the reported tasks: Low-bit quantization methods at modest compression (2-bit KV schemes) show earlier degradation in long-context behavior than kvtc at substantially higher compression settings. Eviction methods perform poorly as generic compressors for long-context tasks, consistent with their objective function (selective pruning) being misaligned with “lossless-ish storage for reuse.” xKV shows competitive results on some tasks but a consistent underperformance on Qasper relative to kvtc and vanilla in the provided tables, consistent with method-specific distortions introduced by its decomposition regime. Reasoning models and high-variance tasks For DeepSeek-R1-distilled Qwen 2.5 reasoning models, the paper evaluates AIME 2024/2025 and LiveCodeBench coding. Results are averaged over 8 runs with large variance, but a key inference is that kvtc at ~9×–21× compression achieves broadly similar AIME scores within variance bands, while coding performance remains stable at ~9× and degrades more visibly at ~18×–21× on the 7B model. An important nuance is that smaller reasoning models already have smaller KV footprints (reported ~29KiB/token for Qwen R1 1.5B versus 131KiB/token for Llama 3.1 8B), so the economic value of aggressive KV cache compression is proportionally higher for large models and long contexts than for small models with short contexts, unless the serving system’s bottleneck is dominated by cache transfer rather than HBM capacity. Multi-GPU inference and pipeline parallel For Llama 3.3 70B Instruct run pipeline-parallel across 4 GPUs (20 layers per GPU), the paper compresses KV cache chunks independently per GPU. On MATH-500, the reported accuracy declines from 75.6 (vanilla) to 74.4 at 10× and 72.6 at 20×, with standard errors near ~1.9. NIAH and LITM remain at 100.0 for all tested ratios in that table. The paper notes that joint compression across chunks could improve accuracy for some offload scenarios but is not required for feasibility, highlighting an engineering trade-off between deployment simplicity in distributed settings and optimal global compression. Latency and TTFT economics A critical system result is the measured compression/decompression latency on an H100 for a non-fused implementation. For Mistral NeMo 12B in bfloat16: BS=8, CTX=8K: compression 379ms, decompression 267ms; vanilla recompute TTFT 3098ms; kvtc decompression TTFT 380ms BS=2, CTX=16K: compression 194ms, decompression 143ms; vanilla recompute TTFT 1780ms; kvtc decompression TTFT 208ms These measurements imply that, when a cache would otherwise be recomputed, decompressing a stored compressed cache can reduce TTFT by ~8×–9× in these scenarios, even without kernel fusion. The decomposition of runtime shows PCA projection and entropy coding as the largest contributors, implying that GPU-optimized kernels and faster GPU-native lossless codecs could reduce overhead further. The fundamental economic conclusion is that, in multi-turn settings with long prefixes, compression-induced overhead is likely dominated by the avoided prefill compute and avoided transfer overhead for uncompressed caches. KEY DEPLOYMENT-SENSITIVE DESIGN CHOICES AND FAILURE MODES Several design choices appear to be “hard requirements” rather than optional optimizations: Sink tokens and sliding window exclusions The paper’s ablations show that compressing early “sink” tokens can catastrophically degrade accuracy at high compression ratios (example: Llama 3.1 8B at 64× collapses on multiple tasks when sink tokens are compressed). Similarly, compressing the most recent tokens hurts performance, motivating a sliding window (default 128 tokens) that remains uncompressed. This introduces a predictable engineering constraint: kvtc is not a uniform compression of the full cache; it is a policy-driven, token-position-dependent codec. Production integration therefore requires correct handling of token positions, attention sinks, and window management, and these policies must be aligned with attention-kernel behavior and model-specific sink dynamics. RoPE handling Removing positional rotations prior to compression is described as important for preserving low-rank structure. In deployment, this implies that the codec must be position-aware and must invert and reapply RoPE correctly. This is an additional source of complexity relative to pure per-token quantization and is sensitive to model variants and RoPE parameterizations. Calibration set representativeness The method’s quality hinges on the PCA basis generalizing from calibration data to production data. The paper demonstrates relative stability with 160K–200K calibration tokens and explores domain shifts (general web text vs math traces vs code). Results suggest that moderate domain mismatch is tolerated at 16×–64×, while extreme compression (e.g., 256× in ablations) becomes materially more sensitive to calibration choice. In production, this implies that operators targeting the “negligible degradation” regime should be able to calibrate with broadly representative corpora, while operators targeting ultra-high compression for specialized workloads should expect tighter coupling between calibration domain and achieved quality. PCA matrix storage overhead and operational footprint A non-trivial hidden cost is the need to store PCA projection matrices per model. The paper reports that, prior to DP trimming, PCA matrices stored at 16-bit can amount to a meaningful fraction of model parameter count (examples reported: ~2.4% for Llama 3.3 70B, ~8.7% for Llama 3.1 8B). This overhead is amortized across all cached sessions for a model but competes with HBM/DRAM budgets in multi-model serving. DP-driven trimming can reduce this overhead at higher compression ratios by removing zero-bit components, but the directionality is not guaranteed at low compression ratios if many components remain active. In distributed inference (pipeline parallel), per-chunk PCA can reduce matrix sizes, but may reduce cross-layer decorrelation benefits if fewer layers are concatenated. SYSTEM-LEVEL IMPLICATIONS FOR GENERATIVE AI INFRASTRUCTURE GPU AND HBM The principal infrastructure implication is that KV cache compression at storage time targets the dominant memory allocator stressor in stateful serving: the accumulation of idle or warm conversation state. For workloads with long reusable prefixes (code assistants, enterprise agents with large system prompts, repeated RAG scaffolds, document chat), the limiting resource frequently becomes HBM reserved for KV caches rather than compute. By compressing stale caches by ~20× (or more), the same HBM budget can retain a materially larger working set of cached prefixes, increasing cache hit rates and reducing recomputation. This effect is multiplicative with cache-aware routing and prefix sharing: more prefixes can remain resident (hot or warm) and can be routed to nodes that already hold them, improving both throughput and tail latency. However, kvtc as described does not reduce the active KV cache footprint during the actual attention computation for a currently decoding sequence, because the model operates on decompressed KV caches during decoding. Therefore, the method does not directly reduce HBM bandwidth consumed by attention kernels during steady-state decode, and does not directly address the “memory traffic per generated token” bottleneck that motivates online KV quantization and eviction strategies. The primary HBM benefit is increased effective capacity for caches between turns and reduced HBM pressure from storing many idle sessions, not reduced per-token decode bandwidth. Compression and decompression themselves consume GPU compute and memory bandwidth. The measured decompression TTFT of ~208ms–380ms in the provided benchmarks indicates that the overhead is real but can be materially smaller than recomputation of long prefixes. In an HBM-constrained serving environment, this overhead can be interpreted as a trade between (a) maintaining more caches warm and paying decompression on reuse versus (b) evicting caches and paying full prefill recomputation. The decision boundary will depend on distribution of inter-turn idle times, probability of reuse, and SLA sensitivity to TTFT. kvtc expands the feasible region where keeping caches is economically rational, especially for long prompts. CPU AND DRAM The method implies a stronger role for CPU DRAM as a warm KV cache tier. A ~20× compression ratio changes the practical scale of “warm state” that can be stored per server. Using the paper’s reported KV cache sizes, a 10K-token 16-bit KV cache for Llama 3.3 70B is ~3.13GiB; compressing by ~20× would reduce this to ~160MiB. At that size, storing hundreds to thousands of warm conversation states in DRAM becomes materially more feasible, increasing cache hit rates and reducing NVMe dependence. This can shift system design from “HBM-only hot caches with aggressive eviction” toward “HBM hot + DRAM warm with long retention,” which is structurally analogous to CPU page cache hierarchies in classical systems design. CPU compute implications depend on where compression is executed. The paper explicitly allows compression on CPU if the cache is already in storage, but the strongest bandwidth savings are achieved when compression happens before moving KV caches off the GPU. If an operator chooses GPU-side compression prior to PCIe/NVLink transfer, CPU compute overhead is modest (orchestrating and DP calibration offline). If an operator instead transfers uncompressed caches to CPU for compression, bandwidth savings are forfeited and CPU memory bandwidth becomes a bottleneck. Therefore, the most economically coherent deployment path is GPU-native compression/decompression with CPU DRAM used as the warm storage reservoir.
12:09

$NVDA $MU $SNDK $LITE PAPER OVERVIEW AND CORE CLAIMS The paper “KV Cache Transform Coding for Compact Storage in LLM Inference” introduces kvtc, a transform-coding pipeline that compresses transformer key-value (KV) caches primarily for storage and transfer in LLM serving, rather than for accelerating the per-token attention kernel during active decoding. The method combines 3 stages: (1) feature decorrelation via a PCA basis computed from a calibration dataset and reused across requests; (2) adaptive, variable-precision quantization with bit allocation solved via dynamic programming (DP), including groupwise scaling/shift overhead; and (3) lossless entropy coding (DEFLATE via nvCOMP in the reference implementation) to exploit residual redundancy after quantization. The central empirical claim is that KV tensors contain large, exploitable redundancy across heads and layers, enabling approximately 20× compression versus a 16-bit baseline with negligible degradation across a broad set of accuracy and long-context benchmarks, with materially higher compression (≥40×) available at modest quality cost in some regimes. The system claim is that such compression materially improves the economics of multi-turn, prefix-reuse serving by extending effective KV cache capacity in GPU HBM and host tiers (DRAM/NVMe) and by reducing inter-node and GPU↔host bandwidth demands, thereby improving cache hit rates and reducing time-to-first-token (TTFT) relative to recomputation when caches would otherwise be evicted. KV CACHE AS THE DOMINANT STATE VARIABLE IN INFERENCE ECONOMICS KV cache growth is linear in context length and is multiplicative in layers and attention heads, making it an increasingly dominant constraint as (a) context lengths expand, (b) models add layers and maintain large hidden dimensions, and (c) production workloads shift toward iterative and tool-augmented interactions that repeatedly reuse long prefixes. The paper uses the canonical 16-bit KV cache size formula (4·l·h·d_head·t) bytes and reports 16-bit KV cache sizes per 1K tokens of context that are already operationally large: 128MiB for Llama 3.1 8B, 160MiB for Mistral NeMo 12B, and 320MiB for Llama 3.3 70B Instruct. In binary units, these figures imply per-token KV footprints of 128KiB/token (Llama 3.1 8B), 160KiB/token (Mistral NeMo 12B), and 320KiB/token (Llama 3.3 70B Instruct) at 16-bit. For a 10K-token prompt (10×1K in the paper’s binary convention), the 16-bit KV cache sizes scale to approximately 1.25GiB (Llama 3.1 8B), 1.56GiB (Mistral NeMo 12B), and 3.13GiB (Llama 3.3 70B Instruct). These magnitudes explain why stale caches create a throughput–latency dilemma: retaining them in HBM maximizes responsiveness on future turns but crowds out concurrent sessions; evicting them forces quadratic-cost prefill recomputation and increases TTFT; offloading them to host or storage introduces large transfer overhead and consumes DRAM/NVMe capacity. A key operational nuance emphasized is that modern serving stacks increasingly treat KV caches as a database, leveraging block paging and shared-prefix reuse. In the common disaggregated serving design (separate prefill and decode nodes), KV cache transfer becomes a dominant category of cross-node traffic. Under that design, any reduction in KV cache size directly increases effective fabric capacity and reduces tail latency attributable to congestion, while also enabling longer cache lifetimes in “hot” (HBM) and “warm” (CPU DRAM) tiers that raise cache hit rates and reduce recomputation frequency. The paper’s quantitative example illustrates the economic stakes: a 1,000-line code file tokenized at ~10 tokens/line yields ~10K tokens; for Llama 3.3 70B, an 8-bit KV cache for that context is ~1.6GiB. Reuse across subsequent turns or parallel chats around the same file is valuable, but HBM scarcity makes retaining many such caches infeasible without compression. TECHNICAL MECHANISM: WHY KV CACHES ARE COMPRESSIBLE AND HOW KVTC EXPLOITS IT The technical rationale begins with an empirical observation: keys (and, to a lesser extent, values) across different attention heads can be aligned into a shared latent space using orthogonal transformations (Procrustes alignment). This supports the hypothesis that head-specific projections introduce rotations of a common subspace rather than completely distinct information, implying that concatenating across heads and layers should reveal low-rank structure suitable for linear decorrelation and dimensionality reduction. The method operationalizes this using a PCA/SVD basis learned from calibration data rather than recomputing a decomposition per prompt. This design choice targets production viability: per-prompt SVD is computationally expensive and scales poorly with long prompts and frequent cache updates. kvtc is explicitly structured as an offline-calibrated, online-applied codec: Calibration (performed 1 time per model and compression setting for DP allocation) A calibration dataset is forwarded through the model to collect KV caches. Token positions are pooled, and a subset of positions is sampled. Keys and values are processed separately. Several implementation choices are highlighted as decisive for stability: Rotary positional embeddings are effectively removed prior to compression (“undo positional rotations”), because positional rotations degrade the apparent low-rank structure of keys. “Attention sink” tokens (the earliest tokens in the sequence) and a sliding window of most recent tokens are excluded from compression because they disproportionately affect attention patterns and are empirically more sensitive to reconstruction error. Cross-layer concatenation is used: keys (or values) from multiple layers and heads at the same token position are concatenated along the feature axis to form a higher-dimensional feature vector. PCA is computed over these concatenated vectors, improving robustness relative to per-layer or per-head PCA. The PCA basis is computed via SVD of centered calibration data, using randomized SVD for scalability with a target rank cutoff. The paper reports calibration regimes of 160K tokens for several models with a 10K PCA dimension cutoff (8K for Qwen variants with fewer KV heads), selected to fit within a single 80GB H100 memory envelope and complete within minutes. A critical economic detail is that the same PCA basis can be reused across multiple compression ratios; only the DP-derived precision assignment changes per compression target. Compression (applied between inference phases) Compression operates on stored KV cache tensors, not on weights, and does not modify attention computation. The KV cache is projected into the PCA basis, quantized, packed, and then entropy-coded. Compression is positioned as a background or between-phase operation (after decoding, or between prefill and decode), executed on GPU or CPU depending on where the cache currently resides. The design intent is that compression should not sit on the critical per-token decoding path; it is a storage and transport optimization. Decompression (performed prior to reuse) Decompression reverses the entropy coding and quantization and applies the inverse PCA projection. A practical latency optimization is proposed: inverse projection can be performed layer-by-layer using submatrices of the PCA basis, allowing generation to begin before the full cache is reconstructed, reducing TTFT. Quantization and bit allocation are the core differentiators versus simpler PCA truncation. PCA provides ordered components by variance; kvtc uses DP to allocate a global bit budget across PCA coordinates (and across groups of coordinates) to minimize reconstruction error in the decorrelated domain. Groups of subsequent PCA coordinates share 16-bit shift and scale factors (a microscaling-inspired design), and the DP algorithm jointly selects group size and precision type under a bit budget, including the overhead of per-group metadata. DP commonly assigns 0 bits to many trailing PCA components, which both increases compression and provides a mechanism to trim the PCA basis to the subset of components that actually carry payload, reducing compute and storage overhead of the projection matrices in deployment. Lossless entropy coding then exploits the structure induced by quantization. DEFLATE is used in the reference implementation, and the paper emphasizes that the incremental gain from the lossless stage is content-dependent but meaningful, with an average uplift of ~1.23× on top of quantization in the reported regime. An ablation in the appendices indicates that GPU-friendly variants (GDeflate) can achieve nearly identical compression ratios (≤0.1 difference in measured cases), implying that throughput-optimized lossless codecs can likely be substituted without sacrificing meaningful compression. EMPIRICAL RESULTS: ACCURACY, COMPRESSION, AND LATENCY General-purpose 8B–12B dense models The paper evaluates Llama 3.1 8B, MN-Minitron 8B, and Mistral NeMo 12B across math/knowledge (GSM8K, MMLU) and long-context tasks (Qasper, Lost in the Middle, RULER Variable Tracking) under a simulated multi-turn regime where compression/decompression is applied periodically, with a sliding window of recent tokens excluded. A consistent pattern appears: kvtc maintains near-vanilla performance through 16× compression settings, and remains competitive at 32×, with degradation becoming task- and model-dependent at 64×, particularly on long-context retrieval metrics when compression is pushed aggressively. Selected quantitative anchor points from the paper’s standard-error table (all values are reported with the paper’s evaluation setup and token-window exclusions): Llama 3.1 8B Vanilla: GSM8K 56.8, MMLU 60.5, Qasper 40.4, LITM 99.4, RULER-VT 99.8 kvtc16×: GSM8K 56.9, MMLU 60.1, Qasper 40.7, LITM 99.3, RULER-VT 99.1 kvtc32×: GSM8K 57.8, MMLU 60.6, Qasper 39.4, LITM 99.1, RULER-VT 98.9 kvtc64×: GSM8K 57.2, MMLU 60.7, Qasper 37.8, LITM 90.2, RULER-VT 95.9 These results indicate that, for this model, long-context sensitivity emerges at 64× with meaningful drops in LITM and RULER-VT, while math/knowledge scores remain stable, implying a differential sensitivity consistent with key-vector precision being more critical for retrieval-style behavior. Mistral NeMo 12B Vanilla: GSM8K 61.9, MMLU 64.5, Qasper 38.4, LITM 99.5, RULER-VT 99.8 kvtc16×: GSM8K 62.0, MMLU 64.4, Qasper 37.6, LITM 99.8, RULER-VT 99.5 kvtc32×: GSM8K 62.2, MMLU 63.8, Qasper 37.5, LITM 99.6, RULER-VT 98.7 kvtc64×: GSM8K 61.9, MMLU 61.4, Qasper 38.0, LITM 95.3, RULER-VT 98.0 Here, degradation at 64× is visible but materially smaller than the Llama 3.1 8B LITM drop, suggesting model-architecture or training-data differences can change the tolerance envelope for aggressive KV cache distortion. MN-Minitron 8B Vanilla: GSM8K 59.1, MMLU 64.3, Qasper 38.2, LITM 99.8, RULER-VT 99.4 kvtc16×: GSM8K 60.3, MMLU 64.1, Qasper 38.6, LITM 99.3, RULER-VT 98.8 kvtc32×: GSM8K 59.1, MMLU 63.7, Qasper 37.7, LITM 86.9, RULER-VT 96.0 kvtc64×: GSM8K 57.8, MMLU 62.1, Qasper 38.1, LITM 59.5, RULER-VT 93.4 This model shows markedly higher sensitivity on LITM at 32× and 64×, despite stable short-context metrics, reinforcing that “compression safety” is not monotonic in parameter count and that pruning/distillation choices can alter KV cache redundancy or robustness. Comparisons to baselines The paper compares kvtc to quantization baselines (KIVI, GEAR, FP8) and eviction baselines (H2O, TOVA), plus an SVD-based prefill-optimization method (xKV). Across the reported tasks: Low-bit quantization methods at modest compression (2-bit KV schemes) show earlier degradation in long-context behavior than kvtc at substantially higher compression settings. Eviction methods perform poorly as generic compressors for long-context tasks, consistent with their objective function (selective pruning) being misaligned with “lossless-ish storage for reuse.” xKV shows competitive results on some tasks but a consistent underperformance on Qasper relative to kvtc and vanilla in the provided tables, consistent with method-specific distortions introduced by its decomposition regime. Reasoning models and high-variance tasks For DeepSeek-R1-distilled Qwen 2.5 reasoning models, the paper evaluates AIME 2024/2025 and LiveCodeBench coding. Results are averaged over 8 runs with large variance, but a key inference is that kvtc at ~9×–21× compression achieves broadly similar AIME scores within variance bands, while coding performance remains stable at ~9× and degrades more visibly at ~18×–21× on the 7B model. An important nuance is that smaller reasoning models already have smaller KV footprints (reported ~29KiB/token for Qwen R1 1.5B versus 131KiB/token for Llama 3.1 8B), so the economic value of aggressive KV cache compression is proportionally higher for large models and long contexts than for small models with short contexts, unless the serving system’s bottleneck is dominated by cache transfer rather than HBM capacity. Multi-GPU inference and pipeline parallel For Llama 3.3 70B Instruct run pipeline-parallel across 4 GPUs (20 layers per GPU), the paper compresses KV cache chunks independently per GPU. On MATH-500, the reported accuracy declines from 75.6 (vanilla) to 74.4 at 10× and 72.6 at 20×, with standard errors near ~1.9. NIAH and LITM remain at 100.0 for all tested ratios in that table. The paper notes that joint compression across chunks could improve accuracy for some offload scenarios but is not required for feasibility, highlighting an engineering trade-off between deployment simplicity in distributed settings and optimal global compression. Latency and TTFT economics A critical system result is the measured compression/decompression latency on an H100 for a non-fused implementation. For Mistral NeMo 12B in bfloat16: BS=8, CTX=8K: compression 379ms, decompression 267ms; vanilla recompute TTFT 3098ms; kvtc decompression TTFT 380ms BS=2, CTX=16K: compression 194ms, decompression 143ms; vanilla recompute TTFT 1780ms; kvtc decompression TTFT 208ms These measurements imply that, when a cache would otherwise be recomputed, decompressing a stored compressed cache can reduce TTFT by ~8×–9× in these scenarios, even without kernel fusion. The decomposition of runtime shows PCA projection and entropy coding as the largest contributors, implying that GPU-optimized kernels and faster GPU-native lossless codecs could reduce overhead further. The fundamental economic conclusion is that, in multi-turn settings with long prefixes, compression-induced overhead is likely dominated by the avoided prefill compute and avoided transfer overhead for uncompressed caches. KEY DEPLOYMENT-SENSITIVE DESIGN CHOICES AND FAILURE MODES Several design choices appear to be “hard requirements” rather than optional optimizations: Sink tokens and sliding window exclusions The paper’s ablations show that compressing early “sink” tokens can catastrophically degrade accuracy at high compression ratios (example: Llama 3.1 8B at 64× collapses on multiple tasks when sink tokens are compressed). Similarly, compressing the most recent tokens hurts performance, motivating a sliding window (default 128 tokens) that remains uncompressed. This introduces a predictable engineering constraint: kvtc is not a uniform compression of the full cache; it is a policy-driven, token-position-dependent codec. Production integration therefore requires correct handling of token positions, attention sinks, and window management, and these policies must be aligned with attention-kernel behavior and model-specific sink dynamics. RoPE handling Removing positional rotations prior to compression is described as important for preserving low-rank structure. In deployment, this implies that the codec must be position-aware and must invert and reapply RoPE correctly. This is an additional source of complexity relative to pure per-token quantization and is sensitive to model variants and RoPE parameterizations. Calibration set representativeness The method’s quality hinges on the PCA basis generalizing from calibration data to production data. The paper demonstrates relative stability with 160K–200K calibration tokens and explores domain shifts (general web text vs math traces vs code). Results suggest that moderate domain mismatch is tolerated at 16×–64×, while extreme compression (e.g., 256× in ablations) becomes materially more sensitive to calibration choice. In production, this implies that operators targeting the “negligible degradation” regime should be able to calibrate with broadly representative corpora, while operators targeting ultra-high compression for specialized workloads should expect tighter coupling between calibration domain and achieved quality. PCA matrix storage overhead and operational footprint A non-trivial hidden cost is the need to store PCA projection matrices per model. The paper reports that, prior to DP trimming, PCA matrices stored at 16-bit can amount to a meaningful fraction of model parameter count (examples reported: ~2.4% for Llama 3.3 70B, ~8.7% for Llama 3.1 8B). This overhead is amortized across all cached sessions for a model but competes with HBM/DRAM budgets in multi-model serving. DP-driven trimming can reduce this overhead at higher compression ratios by removing zero-bit components, but the directionality is not guaranteed at low compression ratios if many components remain active. In distributed inference (pipeline parallel), per-chunk PCA can reduce matrix sizes, but may reduce cross-layer decorrelation benefits if fewer layers are concatenated. SYSTEM-LEVEL IMPLICATIONS FOR GENERATIVE AI INFRASTRUCTURE GPU AND HBM The principal infrastructure implication is that KV cache compression at storage time targets the dominant memory allocator stressor in stateful serving: the accumulation of idle or warm conversation state. For workloads with long reusable prefixes (code assistants, enterprise agents with large system prompts, repeated RAG scaffolds, document chat), the limiting resource frequently becomes HBM reserved for KV caches rather than compute. By compressing stale caches by ~20× (or more), the same HBM budget can retain a materially larger working set of cached prefixes, increasing cache hit rates and reducing recomputation. This effect is multiplicative with cache-aware routing and prefix sharing: more prefixes can remain resident (hot or warm) and can be routed to nodes that already hold them, improving both throughput and tail latency. However, kvtc as described does not reduce the active KV cache footprint during the actual attention computation for a currently decoding sequence, because the model operates on decompressed KV caches during decoding. Therefore, the method does not directly reduce HBM bandwidth consumed by attention kernels during steady-state decode, and does not directly address the “memory traffic per generated token” bottleneck that motivates online KV quantization and eviction strategies. The primary HBM benefit is increased effective capacity for caches between turns and reduced HBM pressure from storing many idle sessions, not reduced per-token decode bandwidth. Compression and decompression themselves consume GPU compute and memory bandwidth. The measured decompression TTFT of ~208ms–380ms in the provided benchmarks indicates that the overhead is real but can be materially smaller than recomputation of long prefixes. In an HBM-constrained serving environment, this overhead can be interpreted as a trade between (a) maintaining more caches warm and paying decompression on reuse versus (b) evicting caches and paying full prefill recomputation. The decision boundary will depend on distribution of inter-turn idle times, probability of reuse, and SLA sensitivity to TTFT. kvtc expands the feasible region where keeping caches is economically rational, especially for long prompts. CPU AND DRAM The method implies a stronger role for CPU DRAM as a warm KV cache tier. A ~20× compression ratio changes the practical scale of “warm state” that can be stored per server. Using the paper’s reported KV cache sizes, a 10K-token 16-bit KV cache for Llama 3.3 70B is ~3.13GiB; compressing by ~20× would reduce this to ~160MiB. At that size, storing hundreds to thousands of warm conversation states in DRAM becomes materially more feasible, increasing cache hit rates and reducing NVMe dependence. This can shift system design from “HBM-only hot caches with aggressive eviction” toward “HBM hot + DRAM warm with long retention,” which is structurally analogous to CPU page cache hierarchies in classical systems design. CPU compute implications depend on where compression is executed. The paper explicitly allows compression on CPU if the cache is already in storage, but the strongest bandwidth savings are achieved when compression happens before moving KV caches off the GPU. If an operator chooses GPU-side compression prior to PCIe/NVLink transfer, CPU compute overhead is modest (orchestrating and DP calibration offline). If an operator instead transfers uncompressed caches to CPU for compression, bandwidth savings are forfeited and CPU memory bandwidth becomes a bottleneck. Therefore, the most economically coherent deployment path is GPU-native compression/decompression with CPU DRAM used as the warm storage reservoir.

TheValueist

16,549 Aufrufe • vor 6 Monaten

During an agentic session, a quick “hi” after a coffee break can cost a full dollar. But it’s not because of the “hi.” It’s the million tokens of context you’re repaying for the moment your cache goes cold. Here's the caching math from Evan: 0:00 Why you're billed for last turn's context, not just this one 0:17 Cache hits: ~10% of full price 0:41 Agentic AI = way more back-and-forth than chat 1:16 Paying full price for context even on a 1-token tool call 1:40 Cache hit vs. cache miss 2:13 The $1 "hi": walk away for an hour, come back to a cold cache 2:59 Cost grows by the square, even with caching 3:28 How context compaction helps 4:03 Why Claude Code/Codex may compact at ~200-300K, not the full 1M window
4:34

During an agentic session, a quick “hi” after a coffee break can cost a full dollar. But it’s not because of the “hi.” It’s the million tokens of context you’re repaying for the moment your cache goes cold. Here's the caching math from Evan: 0:00 Why you're billed for last turn's context, not just this one 0:17 Cache hits: ~10% of full price 0:41 Agentic AI = way more back-and-forth than chat 1:16 Paying full price for context even on a 1-token tool call 1:40 Cache hit vs. cache miss 2:13 The $1 "hi": walk away for an hour, come back to a cold cache 2:59 Cost grows by the square, even with caching 3:28 How context compaction helps 4:03 Why Claude Code/Codex may compact at ~200-300K, not the full 1M window

Arena.ai

51,372 Aufrufe • vor 4 Tagen

Dylan Patel of SemiAnalysis says speed is the moat in AI inference, and getting there runs through new compute (Groq/Cerebras), KV-cache offload, and networking optimization (Aria). "Most of the infrastructure is pretty optimized for high throughput, I would say, but more and more people want to pay for speed. It turns out agents are very sequential, task-oriented, and models running in sequential tasks." "Trying something, verifying it, checking if it's right, and you do that in a loop. It ends up being very slow." "And speed is the moat. The faster you go, speed is the moat." "So a lot of the workloads and high-end use cases that people are developing now are all focused on high interactivity, and the infrastructure there is quite different." "So let's break that down into a few components." "One is you can do new compute: you can do Groq or Cerebras-style compute, which is very much focused on high interactivity." "But even with existing GPUs you can do a few things. One, if you're focused heavily on time to first token, you might offload KV cache to storage, keep it cached, move it back into the compute whenever you need it, much faster than recomputing." "And the third thing you can do is you have to optimize the network."
1:12

Dylan Patel of SemiAnalysis says speed is the moat in AI inference, and getting there runs through new compute (Groq/Cerebras), KV-cache offload, and networking optimization (Aria). "Most of the infrastructure is pretty optimized for high throughput, I would say, but more and more people want to pay for speed. It turns out agents are very sequential, task-oriented, and models running in sequential tasks." "Trying something, verifying it, checking if it's right, and you do that in a loop. It ends up being very slow." "And speed is the moat. The faster you go, speed is the moat." "So a lot of the workloads and high-end use cases that people are developing now are all focused on high interactivity, and the infrastructure there is quite different." "So let's break that down into a few components." "One is you can do new compute: you can do Groq or Cerebras-style compute, which is very much focused on high interactivity." "But even with existing GPUs you can do a few things. One, if you're focused heavily on time to first token, you might offload KV cache to storage, keep it cached, move it back into the compute whenever you need it, much faster than recomputing." "And the third thing you can do is you have to optimize the network."

Fireside Alpha

282,905 Aufrufe • vor 27 Tagen

🎉 Congrats to Fish Audio on launching Fish Audio S2, a frontier TTS model with fine-grained prosody & emotion control via natural-language inline tags. SGLang Day-0 support is now live! 🏆 Best WER on Seed-TTS Eval; 81.88% win rate on EmergentTTS-Eval 🎙️ Voice cloning with 86.4% prefix-cache hit rate via RadixAttention ⚡️ RTF 0.34, 63.3 tok/s on single H200 (single batch) 🌍 Trained on 10M+ hours of audio across ~100 languages, GRPO-aligned 🔧 Dual-AR (Slow + Fast AR) is LLM-isomorphic: continuous batching, paged KV cache & CUDA graphs inherited natively 🗣️ Native multi-speaker: turn-taking, interruptions & cross-speaker emotion in a single pass 👉Cookbook: 👉Blog: 🎬 Curious how to run with SGLang? Check out this voice cloning demo from Chayenne Zhao with Fishaudio-S2-Pro:
1:15

🎉 Congrats to Fish Audio on launching Fish Audio S2, a frontier TTS model with fine-grained prosody & emotion control via natural-language inline tags. SGLang Day-0 support is now live! 🏆 Best WER on Seed-TTS Eval; 81.88% win rate on EmergentTTS-Eval 🎙️ Voice cloning with 86.4% prefix-cache hit rate via RadixAttention ⚡️ RTF 0.34, 63.3 tok/s on single H200 (single batch) 🌍 Trained on 10M+ hours of audio across ~100 languages, GRPO-aligned 🔧 Dual-AR (Slow + Fast AR) is LLM-isomorphic: continuous batching, paged KV cache & CUDA graphs inherited natively 🗣️ Native multi-speaker: turn-taking, interruptions & cross-speaker emotion in a single pass 👉Cookbook: 👉Blog: 🎬 Curious how to run with SGLang? Check out this voice cloning demo from Chayenne Zhao with Fishaudio-S2-Pro:

LMSYS Org

39,939 Aufrufe • vor 5 Monaten