Based on the recent TRiP source code by Carlo Valenti.

Ported to Zig and headless Vulkan Compute shaders.

TurboQuant added an optional inference path.

Achieves 120 tok/s on RTX 3090 for Gemma.

Notes regarding TurboQuant:

Right now Algorithm 1 only, RHT pre-conditioner + Lloyd-Max scalar quantization to a global 4-bit codebook + a small norm-correction γ.

We deliberately drop QJL (Algorithm 2)

Five independent practitioner reproductions converged on this decision.

The sign-bit residual eliminates bias but explodes attention-score variance, which softmax tolerates much worse than bias.

Randomized Hadamard Transform, not random orthogonal.

At 4 bits, plain random rotation this gives PPL 604 vs RHT's 10.12 on Qwen3-1.7B per arclabs001's benchmarks.

Norm-correction γ (TheTom / spiritbuun)

We store original L2 / ‖reconstruction‖ instead of raw L2.

This provides free PPL, and guarantees the dequantized block has the original L2 norm exactly.

Asymmetric K= fp / V=TQ4 by default (the dense-model recommendation from llama.cpp practitioner data).

The TQ4 pack kernel produces 256/256 indices bit-exact versus both the CPU oracle and Python reference on a deterministic input ramp (regeneration script in scripts/cross_validate_turboquant.py).

Memory savings on Gemma 2B at max_pos = 2048

V cache shrinks from 36 MiB to 4.6 MiB across 18 layers (\~5.5×), plus a 2 MiB shared dequant scratchpad.

Hardware Requirements

Any Vulkan 1.3 GPU (AMD / Intel / NVIDIA / Apple via MoltenVK / Android).

One SPIR-V binary per shader, across any vendor.

https://github.com/Foundation42/valkyr