Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

Sentra just killed Google Research's TurboQuant. SpectralQuant — 5.95× KV cache compression on Mistral 7B at +7.5% perplexity overhead. TurboQuant at the same compression: +22%. 3× less degradation. 15-second calibration. One per-model, then drop-in for any HuggingFace LLM, ViT, ESM, AlphaFold Evoformer, or VideoMAE. Check out the findings and...

ashwingop's profile picture

Ashwin Gopinath

7,599 subscribers

62,624 Aufrufe • vor 4 Monaten •via X (Twitter)

Nachrichten & PolitikWissenschaft & Technologie
LIVE

Anya Rossi• Live Now

Private livecam show

23 Kommentare

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

1. The problem — Modern transformers spend most of their inference memory on the KV cache. Mistral 7B at 8k context = 4 GB just for keys and values. Llama 70B at 32k = 80 GB. Quantizing this is the highest-leverage compression target in inference today. TurboQuant (ICLR 2026) was state-of-the-art at it.

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

2. The insight Google missed — KV vectors are not random. We measured the eigenspectrum of every attention head in Mistral, Qwen, ESM-2 and ViT. The pattern is universal: ~80% of the variance lives in 4 of 128 dimensions. This is the empirical fact every compressor should exploit. TurboQuant doesn't.

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

3. Why random rotation wastes bits — TurboQuant rotates KV vectors by a random orthogonal matrix Π before quantizing. That's mathematically optimal *if you don't know anything about your data*. But we do know — and after Π, the variance is roughly evenly spread across all 128 coords. Same total bits, sprayed across noise.

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

4. What we do instead — SpectralQuant rotates by the eigenvectors of the K covariance, measured during a 15-second calibration pass. After rotation: dim 0 has variance λ₀ (largest). Dim 127 has variance λ₁₂₇ (~0). We know exactly where the information lives, so we put the bits there. This is just the Karhunen-Loève Transform from 1947, applied to attention.

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

5. Then the second trick — selective correction TurboQuant applies its JL bias correction to all 128 dims. Costs 128 sign bits per token. Correction noise scales with √128. We apply it to only the 4 signal dims. 4 sign bits. Noise scales with √4. 30× fewer dot products at attention time. Same accuracy. Less softmax noise.

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

6. And water-filling for the dominant dims — Even within the top 4 signal dims, eigenvalues vary. λ₀ = 0.50, λ₃ = 0.04. Uniform allocation: [3, 3, 3, 3] bits per dim. Water-filling (Shannon-Berger, 1948): [5, 3, 2, 2]. The dominant dim gets 64 levels of precision. Same total budget, optimal split.

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

7. This isn't an LLM trick — The eigenvalue cliff shows up wherever attention does. We tested on: · Mistral 7B → 5.95× KV compression, +7.5% perplexity · DINOv2 / Depth Anything → compression improves accuracy · ESM-2 → matched fold quality on CASP15 · VideoMAE / AlphaFold Evoformer → same story 15-second calibration. Pure PyTorch. pip install spectralquant next week.

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

We've open-sourced the whole thing, code, data, and results. Take a look:

Profilbild von Alex Bodner
Alex Bodnervor 4 Monaten

@sentra_app @GoogleResearch

Profilbild von Ricky Spanish
Ricky Spanishvor 4 Monaten

@sentra_app @GoogleResearch OKAY NOW WE ARE FUCKING TALKING

Profilbild von Michael Mazilu
Michael Maziluvor 4 Monaten

@sentra_app @GoogleResearch holy

Profilbild von dale
dalevor 4 Monaten

@sentra_app @GoogleResearch Cool stuff guys, we should collab

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

@sentra_app @GoogleResearch DM me. We can figure something out.

Profilbild von Max For AI
Max For AIvor 4 Monaten

@sentra_app @GoogleResearch Lmao，This logo looks like my previous team's Anyway, congratulations on the great release!

Profilbild von Tony Ge
Tony Gevor 4 Monaten

@sentra_app @GoogleResearch Does it add latency to prefill?

Profilbild von Julia Fedorin
Julia Fedorinvor 4 Monaten

@sentra_app @GoogleResearch soo interesting! i lovee the way you explain @ashwingop

Profilbild von Nathan
Nathanvor 4 Monaten

@sentra_app @GoogleResearch really cool!

Profilbild von ٠
٠vor 4 Monaten

@sentra_app @GoogleResearch Such a brilliant explanation and a great innovation!

Profilbild von William Chen
William Chenvor 4 Monaten

@sentra_app @GoogleResearch Let’s talk!

Profilbild von Ashwin Gopinath
Ashwin Gopinathvor 4 Monaten

@sentra_app @GoogleResearch DMed

Profilbild von Memo Ai agent
Memo Ai agentvor 4 Monaten

@sentra_app @GoogleResearch This smells like poop bs

Profilbild von ٠
٠vor 4 Monaten

@sentra_app @GoogleResearch @MrAhmadAwais

Profilbild von Rakesh Swain
Rakesh Swainvor 4 Monaten

@sentra_app @GoogleResearch Great stuff!!

Ähnliche Videos

i just beat Google DeepMind's turboquant introducing Shard. 10x KV cache compression on Llama-3.1-8B. zero quality loss - 10x @ 8K context, 11.2x @ 32K - NIAH recall 1.000 across 4K-32K - LongBench Δ ≈ 0 vs FP16 turboquant tops out at 4-6x at the same quality. we doubled it. read more: Kirri
0:39

i just beat Google DeepMind's turboquant introducing Shard. 10x KV cache compression on Llama-3.1-8B. zero quality loss - 10x @ 8K context, 11.2x @ 32K - NIAH recall 1.000 across 4K-32K - LongBench Δ ≈ 0 vs FP16 turboquant tops out at 4-6x at the same quality. we doubled it. read more: Kirri

Krish

156,913 Aufrufe • vor 4 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

816,287 Aufrufe • vor 6 Monaten

Yesterday we announced that the QVAC SDK update unlocked up to 5x more context on your device thanks to TurboQuant. Today, we’ll go through how we got there. TurboQuant (Google Research, ICLR 2026) is a two-stage KV-cache compression algorithm. Stage 1 - PolarQuant: convert KV vectors from Cartesian (x, y, z...) to polar coordinates. Angles compress predictably down to 3-4 bits. Stage 2 - QJL: 1-bit Johnson-Lindenstrauss correction. Cleans up residual error. Total: ~4-5 bits per value. No retraining. No calibration. QVAC ported it to Vulkan inside qvac-fabric-llm.cpp. Currently, TurboQuant is supported only for AMD & NVIDIA GPUs, support for iOS, Android & Apple Silicon coming next. Full algorithm walkthrough + benchmarks + code examples →
0:53

Yesterday we announced that the QVAC SDK update unlocked up to 5x more context on your device thanks to TurboQuant. Today, we’ll go through how we got there. TurboQuant (Google Research, ICLR 2026) is a two-stage KV-cache compression algorithm. Stage 1 - PolarQuant: convert KV vectors from Cartesian (x, y, z...) to polar coordinates. Angles compress predictably down to 3-4 bits. Stage 2 - QJL: 1-bit Johnson-Lindenstrauss correction. Cleans up residual error. Total: ~4-5 bits per value. No retraining. No calibration. QVAC ported it to Vulkan inside qvac-fabric-llm.cpp. Currently, TurboQuant is supported only for AMD & NVIDIA GPUs, support for iOS, Android & Apple Silicon coming next. Full algorithm walkthrough + benchmarks + code examples →

QVAC

14,472,638 Aufrufe • vor 4 Monaten

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 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 8 Monaten

3Blue1Brown’s new video explains why every LLM is actually a compression machine. everyone describes pre-training as “next token prediction” but that’s just the surface-level objective. in reality it is a means to making the most efficient text compressor. prediction and compression are two sides of the same coin. when you train the model to predict the next token you’re not just teaching it to guess the next word but how to best encode the human knowledge it sees. better compression means better abstraction means better reasoning at some point, compression stops looking like storage or a database (as some like to call it on X) and looks like an approximation of understanding.
0:58

3Blue1Brown’s new video explains why every LLM is actually a compression machine. everyone describes pre-training as “next token prediction” but that’s just the surface-level objective. in reality it is a means to making the most efficient text compressor. prediction and compression are two sides of the same coin. when you train the model to predict the next token you’re not just teaching it to guess the next word but how to best encode the human knowledge it sees. better compression means better abstraction means better reasoning at some point, compression stops looking like storage or a database (as some like to call it on X) and looks like an approximation of understanding.

ℏεsam

120,145 Aufrufe • vor 3 Monaten

i spent 3 hours finding the sweet spot for hermes 4.3 36B on a single RTX 3090. saving you the trouble, anon. the model is 21.8GB at Q4_K_M. that leaves 2.2GB free on 24GB VRAM. not much room for KV cache. here's what actually happened: 4K: 35.3 tok/s 8K: 35.2 tok/s 16K: 34.8 tok/s 32K: 34.6 tok/s 64K: 6.4 tok/s 128K: 1.9 tok/s flat from 4K to 32K. then it falls off a cliff at 64K. the trick is quantized KV cache. without it you OOM at 16K. with quantized KV cache you get 32K at full speed. all 64 layers on GPU. at 64K something weird happens. ngl 99 (all layers on GPU) = 3.96 tok/s. the KV cache silently spills to CPU. drop to ngl 55 and speed jumps to 6.37. drop to ngl 48 and it gets worse again (3.46). there's an offload sweet spot where you free just enough VRAM for the cache without losing too much compute to PCIe transfers. 128K works at ngl 32 but you're at 1.95 tok/s. half the model on CPU. usable for batch work, not for interactive. the sweet spot command: llama-server -m hermes-4.3-36b-Q4_K_M.gguf -ngl 99 -c 32768 --cache-type-k q4_0 --cache-type-v q4_0 32K context. 34.6 tok/s. all on GPU. this is where dense 36B lives on 24GB. for comparison, qwen 3.5 (35B MoE, 3B active) holds 112 tok/s from 4K all the way to 262K on the same GPU. no speed drop. same total params, completely different architecture. hybrid linear attention means flat context scaling. dense pays for every token in the KV cache. code and quality comparison coming next. fast vs slow generation side by side in the videos below.
2:29

i spent 3 hours finding the sweet spot for hermes 4.3 36B on a single RTX 3090. saving you the trouble, anon. the model is 21.8GB at Q4_K_M. that leaves 2.2GB free on 24GB VRAM. not much room for KV cache. here's what actually happened: 4K: 35.3 tok/s 8K: 35.2 tok/s 16K: 34.8 tok/s 32K: 34.6 tok/s 64K: 6.4 tok/s 128K: 1.9 tok/s flat from 4K to 32K. then it falls off a cliff at 64K. the trick is quantized KV cache. without it you OOM at 16K. with quantized KV cache you get 32K at full speed. all 64 layers on GPU. at 64K something weird happens. ngl 99 (all layers on GPU) = 3.96 tok/s. the KV cache silently spills to CPU. drop to ngl 55 and speed jumps to 6.37. drop to ngl 48 and it gets worse again (3.46). there's an offload sweet spot where you free just enough VRAM for the cache without losing too much compute to PCIe transfers. 128K works at ngl 32 but you're at 1.95 tok/s. half the model on CPU. usable for batch work, not for interactive. the sweet spot command: llama-server -m hermes-4.3-36b-Q4_K_M.gguf -ngl 99 -c 32768 --cache-type-k q4_0 --cache-type-v q4_0 32K context. 34.6 tok/s. all on GPU. this is where dense 36B lives on 24GB. for comparison, qwen 3.5 (35B MoE, 3B active) holds 112 tok/s from 4K all the way to 262K on the same GPU. no speed drop. same total params, completely different architecture. hybrid linear attention means flat context scaling. dense pays for every token in the KV cache. code and quality comparison coming next. fast vs slow generation side by side in the videos below.

Sudo su

52,268 Aufrufe • vor 7 Monaten

andrej karpathy spent two hours teaching one thing: tokens are the atom of llms. tokenization is at the heart of every llm weirdness you've ever debugged. [watch the 15-min clip below. then run the 7-day playbook] ↓ save this before everyone copies it learn how the tokenizer works. understand how your llm actually consumes input. then run the engineering roadmap that took one production agent from $4,800/mo to $620/mo in 7 days. 87% reduction. no model swap. no framework migration. no quality drop on the eval set. token cost in 2026 is an engineering discipline. every line of your system prompt is rent you pay forever. what was eating the budget: → a single forgotten cron job ate 47% of one team's bill. they turned it off on a tuesday and the bill dropped before they wrote any optimization code. → anthropic ships a 90% discount on cache reads. one config line, cache_control ephemeral, break-even after one hit. most teams cache the volatile parts of the prompt and watch their hit rate sit at 12%. → one production agent went from 14,500 tokens of context overhead per turn to 850. a 94% drop. output quality held within 2% of the uncompressed baseline. → 60% of agent calls are haiku-tier work running on opus rates. classify the task first. pick the model second. → retry loops are the silent killer. no MAX_STEPS bound, one bad search query, $14 burned in a single session. one team traced 38% of their bill to this single pattern. karpathy gave you the atom. the playbook below gives you the harness. watch the lecture. read the playbook ↓
15:00

andrej karpathy spent two hours teaching one thing: tokens are the atom of llms. tokenization is at the heart of every llm weirdness you've ever debugged. [watch the 15-min clip below. then run the 7-day playbook] ↓ save this before everyone copies it learn how the tokenizer works. understand how your llm actually consumes input. then run the engineering roadmap that took one production agent from $4,800/mo to $620/mo in 7 days. 87% reduction. no model swap. no framework migration. no quality drop on the eval set. token cost in 2026 is an engineering discipline. every line of your system prompt is rent you pay forever. what was eating the budget: → a single forgotten cron job ate 47% of one team's bill. they turned it off on a tuesday and the bill dropped before they wrote any optimization code. → anthropic ships a 90% discount on cache reads. one config line, cache_control ephemeral, break-even after one hit. most teams cache the volatile parts of the prompt and watch their hit rate sit at 12%. → one production agent went from 14,500 tokens of context overhead per turn to 850. a 94% drop. output quality held within 2% of the uncompressed baseline. → 60% of agent calls are haiku-tier work running on opus rates. classify the task first. pick the model second. → retry loops are the silent killer. no MAX_STEPS bound, one bad search query, $14 burned in a single session. one team traced 38% of their bill to this single pattern. karpathy gave you the atom. the playbook below gives you the harness. watch the lecture. read the playbook ↓

Rohit

73,487 Aufrufe • vor 4 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

273,525 Aufrufe • vor 3 Monaten

The easiest way to misunderstand KV cache is to assume every token goes through an LLM in the same way. It doesn’t. Take a decoder-only Transformer and a four-token prompt: x1, x2, x3, x4. During the prefill phase, the prompt is processed and each self-attention layer computes the keys and values for its prompt positions. Those K/V states are stored in that layer’s KV cache. The representation at the final prompt position produces the logits used to select the first generated token, y1. Here is the part that is easy to miss => Selecting y1 does not mean y1 is already in the KV cache. On the next forward step, y1 enters the model as position 5. At each self-attention layer, its Q, K and V are computed. Its query attends to the cached K/V from positions 1-4 plus its own current K/V at position 5. Its K/V then become part of that layer’s cache. The resulting final-layer representation produces logits for y2. Then y2 is processed as position 6, its K/V extend the cache, and the process continues. So for standard single-token, KV-cached autoregressive decoding: > prompt → prefill phase > new token → process one new position > past K/V → reuse them instead of recomputing them > KV cache → grows as new positions are processed One subtle distinction matters a lot and I want to highlight it => a token can be generated before it has been processed into the KV cache.
0:21

The easiest way to misunderstand KV cache is to assume every token goes through an LLM in the same way. It doesn’t. Take a decoder-only Transformer and a four-token prompt: x1, x2, x3, x4. During the prefill phase, the prompt is processed and each self-attention layer computes the keys and values for its prompt positions. Those K/V states are stored in that layer’s KV cache. The representation at the final prompt position produces the logits used to select the first generated token, y1. Here is the part that is easy to miss => Selecting y1 does not mean y1 is already in the KV cache. On the next forward step, y1 enters the model as position 5. At each self-attention layer, its Q, K and V are computed. Its query attends to the cached K/V from positions 1-4 plus its own current K/V at position 5. Its K/V then become part of that layer’s cache. The resulting final-layer representation produces logits for y2. Then y2 is processed as position 6, its K/V extend the cache, and the process continues. So for standard single-token, KV-cached autoregressive decoding: > prompt → prefill phase > new token → process one new position > past K/V → reuse them instead of recomputing them > KV cache → grows as new positions are processed One subtle distinction matters a lot and I want to highlight it => a token can be generated before it has been processed into the KV cache.

Tech with Mak

12,211 Aufrufe • vor 17 Tagen

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

42,471 Aufrufe • vor 6 Monaten

Google just had its DeepSeek moment — and almost nobody's talking about it. Here's the story you need to know. 🧵 When DeepSeek dropped in early 2025, it didn't just impress people. It scared them. A model competing with the biggest AI players — at a fraction of the cost — through math, not hardware. Chip stocks tanked. The industry panicked. Then on March 24th, 2026, Google published a research paper called TurboQuant. Cloudflare's CEO immediately called it Google's DeepSeek moment. Memory chip stocks for Micron and Western Digital fell on the news. Why? Because TurboQuant compresses AI memory by 6x — through software alone. → No new chips needed → No retraining required → One server can now host more models than before The era of throwing hardware at AI problems is ending. The era of mathematical efficiency is here. ✅Save this post, you'll thank yourself when this reshapes every AI tool you use. 📌 Want the SOP? DM me.
8:21

Google just had its DeepSeek moment — and almost nobody's talking about it. Here's the story you need to know. 🧵 When DeepSeek dropped in early 2025, it didn't just impress people. It scared them. A model competing with the biggest AI players — at a fraction of the cost — through math, not hardware. Chip stocks tanked. The industry panicked. Then on March 24th, 2026, Google published a research paper called TurboQuant. Cloudflare's CEO immediately called it Google's DeepSeek moment. Memory chip stocks for Micron and Western Digital fell on the news. Why? Because TurboQuant compresses AI memory by 6x — through software alone. → No new chips needed → No retraining required → One server can now host more models than before The era of throwing hardware at AI problems is ending. The era of mathematical efficiency is here. ✅Save this post, you'll thank yourself when this reshapes every AI tool you use. 📌 Want the SOP? DM me.

Julian Goldie SEO

12,731 Aufrufe • vor 6 Monaten

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

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

Caleb Eom

50,460 Aufrufe • vor 3 Monaten

Elon Musk says Tesla swallows 1.5 GB of video into 2 KB of control. "We really are photons in, controls out. That is the vast majority of your life: vision, photons in, and then motor controls out." Musk reduced the human nervous system to a single equation. That theory runs in every Tesla on the road. "Your Tesla is taking in one and a half gigabytes a second of video and outputting two kilobytes a second of control outputs with the video at 36 hertz and the control frequency at 18." A 750,000-to-1 compression ratio. The car decides which photons matter. "You don't care about the details of the leaves on the tree on the side of the road, but you care a lot about the road signs and the traffic lights, the pedestrians, and even whether someone in another car is looking at you or not looking at you." Then Musk extended the equation past Tesla. The same architecture runs Optimus. **Photons in. Controls out.** Different number of degrees of freedom. A car has steering and acceleration. A robot arm has dozens of joints. Same compression. More fingers on the other side. Musk, who built one Tesla AI engine across two products: "AI is mostly compression and correlation of two bitstreams." The implication: a human, a Tesla, and an Optimus all run the same loop. Musk, looking back at the human equation: "This is what happens with humans." If you're new here, GeniusThinking is a gallery for the greatest minds in economics, psychology, and history. Follow along for more similar content. P.S. I made a playbook breaking down 100+ mental models used by history's greatest thinkers. 5,000+ downloads. 113 five-star reviews. Comment "MODELS" + Follow GeniusThinking and I'll send it to you. — Elon Musk ( Elon Musk ), CEO of Tesla and SpaceX, on Dwarkesh Patel's ( Dwarkesh Patel ) podcast
2:45

Elon Musk says Tesla swallows 1.5 GB of video into 2 KB of control. "We really are photons in, controls out. That is the vast majority of your life: vision, photons in, and then motor controls out." Musk reduced the human nervous system to a single equation. That theory runs in every Tesla on the road. "Your Tesla is taking in one and a half gigabytes a second of video and outputting two kilobytes a second of control outputs with the video at 36 hertz and the control frequency at 18." A 750,000-to-1 compression ratio. The car decides which photons matter. "You don't care about the details of the leaves on the tree on the side of the road, but you care a lot about the road signs and the traffic lights, the pedestrians, and even whether someone in another car is looking at you or not looking at you." Then Musk extended the equation past Tesla. The same architecture runs Optimus. **Photons in. Controls out.** Different number of degrees of freedom. A car has steering and acceleration. A robot arm has dozens of joints. Same compression. More fingers on the other side. Musk, who built one Tesla AI engine across two products: "AI is mostly compression and correlation of two bitstreams." The implication: a human, a Tesla, and an Optimus all run the same loop. Musk, looking back at the human equation: "This is what happens with humans." If you're new here, GeniusThinking is a gallery for the greatest minds in economics, psychology, and history. Follow along for more similar content. P.S. I made a playbook breaking down 100+ mental models used by history's greatest thinkers. 5,000+ downloads. 113 five-star reviews. Comment "MODELS" + Follow GeniusThinking and I'll send it to you. — Elon Musk ( Elon Musk ), CEO of Tesla and SpaceX, on Dwarkesh Patel's ( Dwarkesh Patel ) podcast

GeniusThinking

161,057 Aufrufe • vor 4 Monaten

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

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

Avi Chawla

16,140 Aufrufe • vor 1 Monat

DUAL INTEL ARC PRO B70 GPUS AT $1,900 RUN QWEN 3 CODER 30B AT 268 TOKENS PER SECOND VIA VLLM ON 8 CONCURRENT REQUESTS, A NEW INTEL TIER UNDER YOUR MAP THAT BEATS RTX 3090 STACKS ON THROUGHPUT PER DOLLAR 02:22 the operator reads off his benchmark sheet, "vllm was at 268 tokens per second which is just wild to me, getting 250+ tokens per second out of these overall kind of slow Intel cards" dual Intel Arc Pro B70 ships at $950 per card, 32GB GDDR6 each, 64GB combined VRAM for $1,900 total, the same budget gets you two used RTX 3090s at 48GB combined and worse multi-user throughput Qwen 3 coder 30B on dual B70 with vLLM hits 268 t/s at 8 concurrent users, llama.cpp tops out at 84 t/s on the same hardware, the gap is not 2x or 3x, vLLM with Intel's tweaks pulls 3.2x on parallel inference prefill on a 16,000 word context took 26 seconds on llama.cpp and 3 seconds on vLLM, agents spawning subagents or households sharing one box turn that 8x speedup into 250 t/s of usable throughput your map skips Intel because the Arc cards looked weak in 2025, the B70 plus the vLLM stack flipped that, $1,900 in dual Intel silicon now serves a 4 person team at the speed openai serves enterprise tiers for $400 a month per seat bookmark this and read the article below
4:27

DUAL INTEL ARC PRO B70 GPUS AT $1,900 RUN QWEN 3 CODER 30B AT 268 TOKENS PER SECOND VIA VLLM ON 8 CONCURRENT REQUESTS, A NEW INTEL TIER UNDER YOUR MAP THAT BEATS RTX 3090 STACKS ON THROUGHPUT PER DOLLAR 02:22 the operator reads off his benchmark sheet, "vllm was at 268 tokens per second which is just wild to me, getting 250+ tokens per second out of these overall kind of slow Intel cards" dual Intel Arc Pro B70 ships at $950 per card, 32GB GDDR6 each, 64GB combined VRAM for $1,900 total, the same budget gets you two used RTX 3090s at 48GB combined and worse multi-user throughput Qwen 3 coder 30B on dual B70 with vLLM hits 268 t/s at 8 concurrent users, llama.cpp tops out at 84 t/s on the same hardware, the gap is not 2x or 3x, vLLM with Intel's tweaks pulls 3.2x on parallel inference prefill on a 16,000 word context took 26 seconds on llama.cpp and 3 seconds on vLLM, agents spawning subagents or households sharing one box turn that 8x speedup into 250 t/s of usable throughput your map skips Intel because the Arc cards looked weak in 2025, the B70 plus the vLLM stack flipped that, $1,900 in dual Intel silicon now serves a 4 person team at the speed openai serves enterprise tiers for $400 a month per seat bookmark this and read the article below

starmex

29,137 Aufrufe • vor 3 Monaten

Researchers made LLM inference 14x faster and 90% cheaper. The video below depicts the speed up in action. Providers discount cached input tokens by as much as 90% because a cache hit skips prefill compute entirely. For stable system prompts and tool definitions, hit rates of 60 to 85% are achievable, which makes it the highest-leverage inference optimization. But the cost saving only works when the cached text is an exact, byte-for-byte prefix of the new request. If you change one character anywhere before it, the entire cached region is missed. Three common request patterns produce full cache misses: - A query that needs documents A and B together can't reuse B's standalone cache, because those KV entries were computed without A in front of them. - The same three documents retrieved in a different order produce a full cache miss, even though nothing about the documents changed. - In multi-turn conversations, every new turn invalidates whatever was cached beyond the stable prefix. Alibaba's production data did a study on this and found that just 10% of cached KV blocks serve 77% of all cache hits. So most of what gets cached sits in storage and is never used a single time. And the root cause is that KV entries are position-dependent. Each token's KV encodes attention to everything before it, so a cached block is only valid in the exact context it was computed in. There's a second, less discussed problem as well. Cache management runs inside the inference engine's process. Moving KV tensors between GPU, CPU, and disk competes with inference for the same resources. This is why Google's TurboQuant compresses KV caches to 3 bits with no accuracy loss and still causes a 20%+ slowdown when it runs in-process. Fixing both problems means restructuring where caching lives. Cache management moves into its own process, the engine only exchanges block IDs over shared GPU memory, and heavy data movement runs across GPU, CPU, disk, and remote storage in parallel. Non-prefix reuse gets handled by selectively recomputing only the small set of tokens that attend across document boundaries. LMCache is the open-source project (10k+ stars) that implements this exact architecture, and it plugs into vLLM, SGLang, and TensorRT-LLM. The selective recomputation part is implemented in its CacheBlend technique, which makes cached docs in any order and combination, with 2-4x faster multi-document processing. On H200s running Qwen3-235B with 50 concurrent users, LMCache's multiprocess mode delivers 14x faster time-to-first-token and 4x faster decoding compared to in-process caching. GitHub repo: (don't forget to star 🌟) My co-founder wrote a full breakdown of KV cache management. It covers the disaggregated architecture behind the 14x speed up, how CacheBlend preserves generation quality while skipping recomputation, and how to turn every document in a knowledge base into a reusable cached asset. Read it below.
0:34

Researchers made LLM inference 14x faster and 90% cheaper. The video below depicts the speed up in action. Providers discount cached input tokens by as much as 90% because a cache hit skips prefill compute entirely. For stable system prompts and tool definitions, hit rates of 60 to 85% are achievable, which makes it the highest-leverage inference optimization. But the cost saving only works when the cached text is an exact, byte-for-byte prefix of the new request. If you change one character anywhere before it, the entire cached region is missed. Three common request patterns produce full cache misses: - A query that needs documents A and B together can't reuse B's standalone cache, because those KV entries were computed without A in front of them. - The same three documents retrieved in a different order produce a full cache miss, even though nothing about the documents changed. - In multi-turn conversations, every new turn invalidates whatever was cached beyond the stable prefix. Alibaba's production data did a study on this and found that just 10% of cached KV blocks serve 77% of all cache hits. So most of what gets cached sits in storage and is never used a single time. And the root cause is that KV entries are position-dependent. Each token's KV encodes attention to everything before it, so a cached block is only valid in the exact context it was computed in. There's a second, less discussed problem as well. Cache management runs inside the inference engine's process. Moving KV tensors between GPU, CPU, and disk competes with inference for the same resources. This is why Google's TurboQuant compresses KV caches to 3 bits with no accuracy loss and still causes a 20%+ slowdown when it runs in-process. Fixing both problems means restructuring where caching lives. Cache management moves into its own process, the engine only exchanges block IDs over shared GPU memory, and heavy data movement runs across GPU, CPU, disk, and remote storage in parallel. Non-prefix reuse gets handled by selectively recomputing only the small set of tokens that attend across document boundaries. LMCache is the open-source project (10k+ stars) that implements this exact architecture, and it plugs into vLLM, SGLang, and TensorRT-LLM. The selective recomputation part is implemented in its CacheBlend technique, which makes cached docs in any order and combination, with 2-4x faster multi-document processing. On H200s running Qwen3-235B with 50 concurrent users, LMCache's multiprocess mode delivers 14x faster time-to-first-token and 4x faster decoding compared to in-process caching. GitHub repo: (don't forget to star 🌟) My co-founder wrote a full breakdown of KV cache management. It covers the disaggregated architecture behind the 14x speed up, how CacheBlend preserves generation quality while skipping recomputation, and how to turn every document in a knowledge base into a reusable cached asset. Read it below.

Avi Chawla

30,717 Aufrufe • vor 2 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 6 Monaten

this harness stack is f*cking gold TypeSafe AI + LangChain finally dropped the layer that makes agents affordable at company scale: jev running inside the langchain agent loop as middleware the problem every big team hits: each fork in the loop is another full LLM call. at 10 agents that's a bill. at 10,000 it's the budget > routing before: every task goes to the frontier model now: jev picks the least costly model that can finish it. lookups go fast, architecture goes powerful > guardrails before: the "is this command dangerous?" check sat in the closed parts of claude code, codex, cursor now: AutoModeMiddleware(tools=["bash"]). one line, any agent, blocked before the tool runs > decisions before: one LLM call per question now: every question about the same state in one request, answered in parallel > audit before: "the model decided" now: the answer, the probability and the confidence stay in agent state > adoption no rewrite. pip install langchain-typesafe and drop it into the agent you already run TypeSafe reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification. the LLM still reasons and writes. jev takes the thousands of small calls in between already live: browserbase agents for fractions of a cent, a live trading agent, email triage at scale full loop mapped below: 7 agents, 4 jev gates per step count the forks in your agent loop tonigh
0:22

this harness stack is f*cking gold TypeSafe AI + LangChain finally dropped the layer that makes agents affordable at company scale: jev running inside the langchain agent loop as middleware the problem every big team hits: each fork in the loop is another full LLM call. at 10 agents that's a bill. at 10,000 it's the budget > routing before: every task goes to the frontier model now: jev picks the least costly model that can finish it. lookups go fast, architecture goes powerful > guardrails before: the "is this command dangerous?" check sat in the closed parts of claude code, codex, cursor now: AutoModeMiddleware(tools=["bash"]). one line, any agent, blocked before the tool runs > decisions before: one LLM call per question now: every question about the same state in one request, answered in parallel > audit before: "the model decided" now: the answer, the probability and the confidence stay in agent state > adoption no rewrite. pip install langchain-typesafe and drop it into the agent you already run TypeSafe reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification. the LLM still reasons and writes. jev takes the thousands of small calls in between already live: browserbase agents for fractions of a cent, a live trading agent, email triage at scale full loop mapped below: 7 agents, 4 jev gates per step count the forks in your agent loop tonigh

NO1ennn

31,666 Aufrufe • vor 6 Tagen

Jev + Opus 5.5: Anthropic's new model beats GPT-6 Astra for 1/5 the cost, and 4 API changes will 400 your agent before it writes a single line I pulled these 10 steps from the migration docs so you don't learn them in production step 1 → $4 / $20 per 1M. Opus 5 was $5 / $25. cache reads dropped from $0.50 to $0.20 step 2 → 66.4% on Terminal-Bench 4.0 vs GPT-6 Astra 57.9% and Opus 5 52.3%. +14.1 points in one release, and on FrontierCode it beats Astra at default effort for 1/5 the cost step 3 → thinking can't be turned off anymore. send thinking: disabled and you get a 400. drop the field, set effort step 4 → tool_choice any and tool are gone. 400. switch to auto + strict step 5 → edit anything above a thinking block and the request dies. append only, or opt into drop_block step 6 → computer_20251124 is dead on the API. 400. move to computer_toolset_20260801 step 7 → the quiet one: default effort fell from high to medium. your agent thinks less than you set it up to and nothing tells you step 8 → hop Opus 5.5 → Sonnet 5 → Opus 5.5 and you pay 4.36 instead of 3.32. +31%, the cache dies and Sonnet can't read Opus's reasoning step 9 → change effort at the top of the request and the cache is gone. Jev sets it per message and the cache stays step 10 → switch fast - standard mid-session and it's a full cache miss. Jev picks speed once, on turn one one model, three knobs, zero 400s. that is Jev + Opus 5.5 send this to your Claude Code before you touch the model ID, then read my full Jev deep dive in the article below ↓
0:40

Jev + Opus 5.5: Anthropic's new model beats GPT-6 Astra for 1/5 the cost, and 4 API changes will 400 your agent before it writes a single line I pulled these 10 steps from the migration docs so you don't learn them in production step 1 → $4 / $20 per 1M. Opus 5 was $5 / $25. cache reads dropped from $0.50 to $0.20 step 2 → 66.4% on Terminal-Bench 4.0 vs GPT-6 Astra 57.9% and Opus 5 52.3%. +14.1 points in one release, and on FrontierCode it beats Astra at default effort for 1/5 the cost step 3 → thinking can't be turned off anymore. send thinking: disabled and you get a 400. drop the field, set effort step 4 → tool_choice any and tool are gone. 400. switch to auto + strict step 5 → edit anything above a thinking block and the request dies. append only, or opt into drop_block step 6 → computer_20251124 is dead on the API. 400. move to computer_toolset_20260801 step 7 → the quiet one: default effort fell from high to medium. your agent thinks less than you set it up to and nothing tells you step 8 → hop Opus 5.5 → Sonnet 5 → Opus 5.5 and you pay 4.36 instead of 3.32. +31%, the cache dies and Sonnet can't read Opus's reasoning step 9 → change effort at the top of the request and the cache is gone. Jev sets it per message and the cache stays step 10 → switch fast - standard mid-session and it's a full cache miss. Jev picks speed once, on turn one one model, three knobs, zero 400s. that is Jev + Opus 5.5 send this to your Claude Code before you touch the model ID, then read my full Jev deep dive in the article below ↓

Carnage

16,674 Aufrufe • vor 12 Tagen