README.md

---
license: apache-2.0
base_model: google/gemma-4-31B
tags:
  - rotorquant
  - kv-cache-quantization
  - gemma
  - gemma4
  - quantized
library_name: transformers
pipeline_tag: image-text-to-text
---

# gemma-4-31B-RotorQuant

**RotorQuant KV cache compression** for [google/gemma-4-31B](https://huggingface.co/google/gemma-4-31B).

This is a **documentation repository** that explains how to combine gemma-4-31B's weights with RotorQuant inference-time KV cache compression. No weights are stored here — use the base model directly and apply RotorQuant via the Python package or llama.cpp fork.

## What is this?

KV cache compression reduces the memory used by the attention cache during inference. Unlike weight quantization (which is baked into the GGUF/MLX file), KV cache compression is applied at runtime — so the same base weights can be used with or without compression.

| Technique | Where it's applied | Savings |
|-----------|-------------------|---------|
| Weight quantization (GGUF/MLX/AWQ) | Baked into model file | Reduces disk + weight memory |
| **RotorQuant KV cache** | At inference time | Reduces attention memory (critical for long context) |

Both can be combined for maximum efficiency.

## Quickstart

### Option A — Python / transformers

Install the `rotorquant` package:

```bash
pip install rotorquant
```

Then use it with the base model:

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from rotorquant import IsoQuantCache

tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-31B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-4-31B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

# Apply RotorQuant to the KV cache
cache = IsoQuantCache(bits=4)  # or bits=2 for more aggressive compression

inputs = tokenizer("Hello, how are you?", return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    past_key_values=cache,
    use_cache=True,
)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
```


### Option B — llama.cpp / LM Studio / Ollama (with fork)

RotorQuant KV cache types (`iso3`) are **not** in upstream llama.cpp. They require:
- [llama-cpp-turboquant fork](https://github.com/johndpope/llama-cpp-turboquant/tree/feature/planarquant-kv-cache)

Once built:

```bash
llama-cli -m gemma-4-31B.gguf \
  --cache-type-k iso3 --cache-type-v iso3 \
  -ngl 99 -fa \
  -p "Hello"
```

For standard runtimes (LM Studio, Ollama, upstream llama.cpp), use conventional KV cache types (`q8_0`, `q4_0`). You lose the RotorQuant-specific benefits but keep GGUF weight quantization.

## Model Specifications

| Property | Value |
|----------|-------|
| Base Model | [google/gemma-4-31B](https://huggingface.co/google/gemma-4-31B) |
| Architecture | Dense transformer |
| Parameters | 31B (dense) |
| Context Length | 128K |
| BF16 Size | ~62 GB |
| Modalities | Text + Image |
| License | apache-2.0 |

## What is RotorQuant?

[RotorQuant](https://github.com/scrya-com/rotorquant) is a KV cache compression method based on Clifford algebra (Cl(3,0)) rotors — a faster, more parameter-efficient alternative to Google's TurboQuant. Uses lightweight block-diagonal rotations (independent 2D/4D rotations per pair/quartet) achieving O(d) complexity instead of O(d log d), fully parallelisable with no inter-element dependencies.

**Benchmarks** (from the RotorQuant repository, Llama 3.1 8B on RTX 5090 — results vary by model and hardware):

- Prefill: 3,822 tok/s (vs TurboQuant 722 tok/s)
- Decode: 119 tok/s (vs TurboQuant 93 tok/s)
- Perplexity: 6.91 (vs TurboQuant 7.07)
- Parameters: 4 per rotor (vs TurboQuant 16,384)

> Benchmarks are from the RotorQuant repository using Llama 3.1 8B. Performance on gemma-4-31B will differ. Please open a discussion if you have independent results.

## Current Ecosystem Support

| Runtime | RotorQuant Support | Notes |
|---------|----------------------|-------|
| Python transformers + `rotorquant` | ✅ Full | Drop-in cache class |
| llama.cpp upstream | ❌ Not merged | Use fork below |
| llama-cpp-turboquant fork | ✅ `planar3`, `iso3` | [GitHub](https://github.com/johndpope/llama-cpp-turboquant/tree/feature/planarquant-kv-cache) |
| LM Studio | ❌ [Requested](https://github.com/lmstudio-ai/lmstudio-bug-tracker/issues/1719) | Use `q8_0` as alternative |
| Ollama | ❌ Not supported | Use `OLLAMA_KV_CACHE_TYPE=q8_0` |
| vLLM | ❌ Not supported | — |
| koboldcpp | ❌ Not supported | — |

## Pre-quantized weight variants

If you want combined weight + KV cache compression, majentik hosts pre-quantized versions:

- [MLX (Apple Silicon)](https://huggingface.co/majentik?search=gemma-4-31B+MLX)
- [GGUF (llama.cpp / Ollama / LM Studio)](https://huggingface.co/majentik?search=gemma-4-31B+GGUF)

## See Also

- [RotorQuant GitHub](https://github.com/scrya-com/rotorquant)
- [TurboQuant paper (arXiv 2504.19874)](https://arxiv.org/abs/2504.19874)
- [llama-cpp-turboquant fork](https://github.com/johndpope/llama-cpp-turboquant/tree/feature/planarquant-kv-cache)
- [Base model: google/gemma-4-31B](https://huggingface.co/google/gemma-4-31B)
- [gemma-4-31B announcement](https://blog.google/technology/developers/gemma-4/)