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

[!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-4-31B-it-HLWQ-Q5-Vision

Multimodal (Image+Text → Text) on consumer GPUs with HLWQ.

🎯 Key Results

📊 Charts

[image] [image] [image] [image]

🏆 GPU Support

🚀 Quick Start

from transformers import AutoModelForMultimodalLM, AutoProcessor
import torch

MODEL = "google/gemma-4-31B-it"
processor = AutoProcessor.from_pretrained(MODEL)

# Load with streaming loader (see notebook for full code)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL, dtype=torch.bfloat16, device_map="cpu",
    attn_implementation="sdpa",
)
# Apply PQ5 dequant + INT4 per-module (text only, vision stays BF16)
# ... see notebook ...

# Image + Text inference
messages = [{
    "role": "user",
    "content": [
        {"type": "image", "url": "https://example.com/image.jpg"},
        {"type": "text", "text": "Describe this image."},
    ]
}]
inputs = processor.apply_chat_template(messages, tokenize=True, return_dict=True,
                                        return_tensors="pt", add_generation_prompt=True).to("cuda")
output = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

📊 KV Cache Compression

🔧 Technical Details

• —Architecture: Gemma 4 (60 layers, 32 attn heads, 16 KV heads, head_dim=256)
• —Hybrid attention: Sliding window (1024) + global attention
• —Text quantization: Hadamard rotation (128×128) + Lloyd-Max Q5 + torchao INT4
• —Vision encoder: Preserved in BF16 (550M params) for full image quality
• —KV cache: Hadamard rotation (256×256) + Lloyd-Max Q3 + real bit-packing
• —Streaming loader: Per-module INT4 via nn.Sequential wrapper — peak VRAM = final model only
• —Base model: google/gemma-4-31B-it (Apache 2.0)

📖 Citation

@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

• —arXiv Paper
• —GitHub
• —Text-only version
• —HLWQ Collection
• —HLWQ Skills

🙏 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!)

from polarengine_vllm import HLWQModel

model = HLWQModel.from_pretrained("caiovicentino1/Gemma-4-31B-it-HLWQ-Q5-Vision")
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-4-31B-it-HLWQ-Q5-Vision", 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-Vision --ppl --chart

Gradio Demo

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