[!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).

🧊 Gemma-4-26B-A4B-it-HLWQ-Q5

25.2B MoE (3.8B active) + Vision — PQ5 quantized weights including ALL MoE experts.

Download: 26.9 GB (vs 51.6 GB BF16 original)

📊 Charts

🚀 Quick Start

Expert Offloading (8.6 GB GPU — best for consumer GPUs)

# vLLM MoE expert cache is in PR #37190 (still open). Use the fork until it lands:
pip install git+https://github.com/caiovicentino/vllm-expert-offload.git

from vllm import LLM, SamplingParams
llm = LLM('google/gemma-4-26B-A4B-it', dtype='bfloat16',
          moe_expert_cache_size=8, enforce_eager=True,
          kernel_config={'moe_backend': 'triton'})

Streaming Loader (PQ5 dequant + INT4)

See POLARQUANT_GEMMA4_26B_A4B_VISION.ipynb

🏆 GPU Support

📓 Notebooks

🔧 Technical Details

• MoE experts: 3D nn.Parameter (128, out, in) — each expert quantized independently
• gate_up_proj: (128, 1408, 2816) per layer
• down_proj: (128, 2816, 704) per layer
• 128 experts × 2 params × 30 layers = 7,680 expert quantizations
• Quantization time: 50 seconds on A100
• PQ5 codes: int8 + fp16 norms + Hadamard rotation + Lloyd-Max centroids

📖 Citation

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

📄 Paper · 💻 GitHub · 📦 pip install polarquant

🚀 Quick Start

Install

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

Load & Generate (1 line!)

from polarengine_vllm import HLWQModel

model = HLWQModel.from_pretrained("caiovicentino1/Gemma-4-26B-A4B-it-HLWQ-Q5")
print(model.generate("Hello, how are you?", max_new_tokens=100))

With KV Cache Compression (5.3x more context)

model = HLWQModel.from_pretrained("caiovicentino1/Gemma-4-26B-A4B-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-4-26B-A4B-it-HLWQ-Q5 --ppl --chart

Gradio Demo

polarquant demo caiovicentino1/Gemma-4-26B-A4B-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

🔗 Links

• 📄 Paper — arXiv:2603.29078
• 💻 GitHub — HLWQ-Engine
• 📦 PyPI — pip install polarquant