Anvil: TurboQuant KV Cache & Speculative Decoding Engine

Overview

As frontier language models scale to massive context windows (32k–262k+ tokens), the Key-Value (KV) cache becomes the primary memory bottleneck. Google TurboQuant (TQ) solves this by rotating high-dimensional activation vectors using Fast Walsh-Hadamard Transforms (FWHT), eliminating outlier activation spikes without requiring expensive retraining.

Anvil (Solstice-Labs/anvil) is Solstice Labs’ dedicated inference engine and model management CLI:

1. TurboQuant WHT Rotation: Fast Walsh-Hadamard Transform ($O(d \log d)$) for ~4.3× to 6.4× KV cache compression with backend-native kernels (TurboFlash on Metal, CUDA, Vulkan, HIP).
2. Single-Word Launch: Run any registered checkpoint instantly (anvil <model> or anvil run <model>).
3. Speculative Decoding: Integrated NextN draft reuse and Gemma 4 Multi-Token Prediction (MTP).

Input Tokens ──► FWHT Rotation ──► Lloyd-Max Quantizer ──► turbo3 / turbo2 KV Cache
                      │                                            │
                      ▼                                            ▼
               QJL Error Correction ◄────────────────────── Attention Compute (Flash-Attn)

---

1. How Google TurboQuant Works

The Outlier Bottleneck

Standard INT8/INT4 quantization fails on raw KV tensors because a few outlier channels carry disproportionately high amplitudes. Clamping these outliers degrades reasoning, while expanding quantization bounds wastes bit-depth on near-zero dimensions.

The TurboQuant Solution

1. Decorrelation via Fast Walsh-Hadamard Transform (FWHT): TurboQuant rotates key-value vectors into a coordinate frame where energy is uniformly distributed: $$\mathbf{v}’ = \frac{1}{\sqrt{d}} \mathbf{H}_d \mathbf{v}$$ Computed in $O(d \log d)$ operations via butterfly network kernels.
2. Scalar Quantization: The rotated distribution approaches a zero-mean Gaussian, enabling optimal Lloyd-Max quantization at 2-bit, 3-bit, or 4-bit precision.
3. Quantized Johnson-Lindenstrauss (QJL) Correction: Preserves inner-product similarity for exact attention score calculation.

---

2. Using the Anvil CLI (anvil)

Pulling Checkpoints

# Pull directly from Hugging Face Hub into Anvil registry
anvil pull hf:Solstice-AI/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-Q8_0-GGUF

Running 262k Long-Context Inference

To compress the full 262k context window on Qwen 3.8 and derivative models into ~4 GB of VRAM, pass the TurboQuant cache presets:

# Interactive chat REPL with 262k context and TurboQuant 3-bit KV cache
anvil run huihui-qwen --ctx 262144 --type-k turbo4 --type-v turbo3

# Save flags so future runs launch instantly with a single word
anvil run huihui-qwen --ctx 262144 --type-k turbo4 --type-v turbo3 --save

# Now run instantly with one word
anvil huihui-qwen

# Single-shot prompt generation
anvil run huihui-qwen -p "Analyze this 260,000 token code repository and verify all concurrency safety invariants..."

TurboQuant KV Presets

Preset	Command Flags	Compression vs FP16	Recommended Use Case
Recommended	--type-k turbo4 --type-v turbo3	~4.3×	Sweet spot (262k in ~4.1 GB VRAM, ~1% loss)
Quality+	--type-k q8_0 --type-v turbo3	~3.0×	Near-lossless precision (<1% quality loss)
High Compression	--type-k turbo4 --type-v turbo2	~6.1×	Maximum context density (262k in ~2.8 GB VRAM)

---

3. Serving OpenAI-Compatible API Server

# Launch multi-slot server with cross-request prefix caching
anvil serve --port 8080 --slots 4

---

4. Empirical Performance Benchmarks

Evaluated on Qwen 3.8 27B and Huihui 35B across 262k context lengths on 1x RTX 4090 / M4 Max:

KV Cache Configuration	32k Context	128k Context	262k Context	Degradation vs FP16
Standard FP16	2.3 GB	9.2 GB	18.6 GB (OOM risk)	Baseline (0.00)
Standard INT8	1.2 GB	4.8 GB	9.6 GB	+0.18 PPL
Anvil TurboQuant 4-bit (turbo4)	0.6 GB	2.4 GB	4.9 GB	+0.03 PPL
Anvil TurboQuant 3-bit (turbo3)	0.5 GB	2.0 GB	4.1 GB	+0.05 PPL
Anvil TurboQuant 2-bit (turbo2)	0.35 GB	1.4 GB	2.8 GB	+0.12 PPL