[!IMPORTANT] Naming notice (2026-04-10). The "HLWQ" technique used in this model is being rebranded to HLWQ (Hadamard-Lloyd Weight Quantization). The change is only the name; the algorithm and the weights in this repository are unchanged.

The rebrand resolves a name collision with an unrelated, earlier KV cache quantization method also named HLWQ (Han et al., arXiv:2502.02617, 2025). HLWQ addresses weight quantization with a deterministic Walsh-Hadamard rotation and Lloyd-Max scalar codebook; Han et al.'s HLWQ addresses KV cache quantization with a random polar rotation. The two methods are technically distinct.

Existing loaders that load this repository by ID continue to work without changes. Future model uploads will use the HLWQ name.

Reference paper for this technique: arXiv:2603.29078 (v2 in preparation; v1 still uses the old name).

🧊 HLWQ Unified: Gemma-4-31B-it-HLWQ-Q5

Run Google's Gemma 4 31B-it on consumer GPUs with HLWQ full-stack compression.

Component	Method	Result
Weights	HLWQ Q5 + torchao INT4	62.5 GB → 21.5 GB
KV Cache	HLWQ Q3 (Hadamard + Lloyd-Max)	5.3x compression

🎯 Key Results

Metric	Value
VRAM	21.5 GB (streaming loader)
Speed	24.9 tok/s
Dequant time	37s (per-module GPU-accelerated)
Compression	BF16 62.5 GB → 21.5 GB (2.9x)

Note: WikiText-2 PPL is not reported — Gemma 4 is a multimodal instruct model where raw-text PPL is not meaningful (BF16 baseline PPL = 1002). Generation quality is excellent (see examples below).

🚀 Quick Start — One-Click Colab

Notebook	Description	GPU
▶️ Inference Chat	Gradio chat UI, streaming loader	L4 / RTX 4090 (24 GB)
Quantization	How this model was quantized	A100 80 GB

Streaming Loader (fits on 24 GB GPUs!)

The inference notebook uses a per-module streaming loader that never loads the full BF16 model on GPU:

Load BF16 on CPU (~62 GB RAM)
For each layer: move to GPU → PQ5 dequant → INT4 quantize → keep on GPU
Peak VRAM: ~21.5 GB (accumulated INT4 only)

🏆 GPU Compatibility

GPU	VRAM	Fits?
RTX 4090	24 GB	✅ Yes (2.5 GB headroom)
RTX 5090	32 GB	✅ Comfortable
L4	24 GB	✅ Yes
A6000	48 GB	✅ Plenty
A100 40/80 GB	40-80 GB	✅ Full headroom
T4	16 GB	❌ Too small for 31B

📊 KV Cache Compression

Method	Bits	Compression	tok/s
FP16 (baseline)	16	1.0x	24.9
HLWQ Q4	4	4.0x	24.8
HLWQ Q3	3	5.3x	24.8
HLWQ Q2	2	8.0x	24.8

💬 Generation Quality

The model produces high-quality, coherent responses:

Q: Explain quantum computing in simple terms.

To understand quantum computing, you first have to understand how a regular computer works... A classical bit is like a coin lying on a table — it's either Heads (1) or Tails (0). A qubit is like a coin spinning on the table — it is effectively both at the same time. This state of being in multiple states at once is called Superposition...

Q: Write a Python function for binary search.

Correct implementation with proper docstring, O(log n) complexity, edge case handling.

Q: What causes aurora borealis?

The aurora borealis is caused by charged particles from the sun colliding with Earth's magnetic field...

🔧 Technical Details

Architecture: Gemma 4 (60 layers, 32 attn heads, 16 KV heads, head_dim=256)
Hybrid attention: Sliding window (1024) + global attention
Weight quantization: Hadamard rotation (128×128) + Lloyd-Max Q5 + torchao INT4
KV cache: Hadamard rotation (256×256) + Lloyd-Max Q3 + real bit-packing
Streaming loader: Per-module INT4 via nn.Sequential wrapper (vLLM pattern)
Quantized layers: 602 (text model + vision encoder projections)
Base model: google/gemma-4-31B-it (Apache 2.0)

📖 Citation

bibtex

@article{polarquant2025,
  title={HLWQ: Hadamard-Rotated Lloyd-Max Quantization for LLM Compression},
  author={Vicentino, Caio},
  journal={arXiv preprint arXiv:2603.29078},
  year={2025},
  url={https://arxiv.org/abs/2603.29078}
}

🔗 Resources

🙏 Acknowledgements

Built on Google's Gemma 4 (Apache 2.0). Quantization by HLWQ with torchao.

🚀 Quick Start

Install

pip install git+https://github.com/caiovicentino/polarengine-vllm.git

Load & Generate (1 line!)

python

from polarengine_vllm import HLWQModel

model = HLWQModel.from_pretrained("caiovicentino1/Gemma-4-31B-it-HLWQ-Q5")
print(model.generate("Hello, how are you?", max_new_tokens=100))

With KV Cache Compression (5.3x more context)

python

model = HLWQModel.from_pretrained("caiovicentino1/Gemma-4-31B-it-HLWQ-Q5", kv_cache_nbits=3)
# KV cache now uses 5.3x less memory — fit longer conversations!
print(model.generate("Explain quantum computing in detail.", max_new_tokens=500))

Benchmark

polarquant bench caiovicentino1/Gemma-4-31B-it-HLWQ-Q5 --ppl --chart

Gradio Demo

polarquant demo caiovicentino1/Gemma-4-31B-it-HLWQ-Q5 --share

📦 Method: HLWQ

Hadamard Rotation + Lloyd-Max Optimal Centroids

Unlike GGUF (uniform quantization), HLWQ places quantization levels where weight density is highest — mathematically proven optimal for Gaussian-distributed neural network weights.

HLWQ Q5 (cos_sim > 0.996) > GGUF Q5_K_M (~0.99) at same size

caiovicentino1/Gemma-4-31B-it-HLWQ-Q5