majentik/gemma-4-31B-RotorQuant-AWQ-4bit

Gemma 4 31B - RotorQuant AWQ 4-bit

4-bit AWQ-quantized version of google/gemma-4-31B (31B dense) with RotorQuant KV-cache quantization. AWQ (Activation-aware Weight Quantization) is an activation-aware method optimal for GPU inference. RotorQuant delivers 5.3x faster prefill and 28% faster decode compared to TurboQuant -- especially valuable for long-context serving on a single 24GB GPU.

Approximate model size: ~17 GB

Note: RotorQuant KV cache modes (planar3, iso3) require the RotorQuant fork or the llama-cpp-turboquant fork. The AWQ weights themselves load cleanly in stock AutoAWQ / vLLM; RotorQuant KV-cache kernels are opt-in.

Model Specifications

Property	Value
Base Model	google/gemma-4-31B
Parameters	~31 billion
Architecture	Dense transformer
Modality	Multimodal: image + text input, text output
License	Apache 2.0
Weight Quantization	AWQ 4-bit (~17 GB)
Group Size	128
KV-Cache Quantization	RotorQuant (planar3 / iso3)
Framework	transformers + AutoAWQ / vLLM

Quickstart

AutoAWQ

from awq import AutoAWQForCausalLM
from transformers import AutoTokenizer

model = AutoAWQForCausalLM.from_quantized(
    "majentik/gemma-4-31B-RotorQuant-AWQ-4bit",
    device_map="auto",
    fuse_layers=True,
)
tokenizer = AutoTokenizer.from_pretrained("majentik/gemma-4-31B-RotorQuant-AWQ-4bit")

prompt = "The history of artificial intelligence began"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(out[0], skip_special_tokens=True))

vLLM

vllm serve majentik/gemma-4-31B-RotorQuant-AWQ-4bit \
  --quantization awq_marlin \
  --max-model-len 8192

With RotorQuant KV cache (fork)

from rotorquant import RotorQuantCache
cache = RotorQuantCache(model, mode="iso3")  # or "planar3"

What is RotorQuant?

RotorQuant is a high-performance KV-cache quantization method using block-diagonal Clifford-algebra rotors. Combined with AWQ 4-bit weights, this delivers a dual compression strategy with superior KV-cache performance for GPU inference.

Key advantages over TurboQuant:

• 5.3x faster prefill
• 28% faster decode
• Equivalent memory savings
• planar3 / iso3 3-bit KV cache modes

KV-Cache Quantization Comparison

Method	Prefill Speed	Decode Speed	Memory Savings	Reference
TurboQuant	1x (baseline)	1x (baseline)	High	arXiv: 2504.19874
RotorQuant	5.3x faster	28% faster	High	GitHub

AWQ vs GGUF vs MLX

Format	Target Hardware	Runtime	Best For
AWQ	NVIDIA / AMD GPU (CUDA/ROCm)	AutoAWQ, vLLM, TGI	GPU-native inference, production serving
GGUF	CPU + GPU (cross-platform)	llama.cpp, Ollama, LM Studio	Laptops, CPU-only boxes, mixed offload
MLX	Apple Silicon	MLX, mlx-lm, mlx-vlm	Macs with unified memory

This repo ships AWQ. See the "See Also" section for GGUF and MLX siblings.

Memory Estimates (Gemma 4 31B)

Precision	Approximate Size	VRAM Tier
FP16 (original)	~62 GB	80 GB+ (A100/H100)
AWQ 8-bit	~31 GB	48 GB+ (L40S, A100 40GB tight)
AWQ 4-bit	~17 GB	24 GB+ (RTX 4090, A5000, A6000)

Fits on a single RTX 4090 (24GB) with room for moderate contexts; comfortable on A6000 / L40 (48GB).

Hardware Requirements

• NVIDIA GPU with >=24 GB VRAM (RTX 4090, A5000, A6000, L40, A100, H100)
• CUDA 12.x recommended
• For vLLM: compute capability >= 7.5 (Turing or newer) for Marlin kernels
• For RotorQuant KV cache: scrya-com/rotorquant fork

See Also

• google/gemma-4-31B -- Base model
• majentik/gemma-4-31B-RotorQuant -- RotorQuant KV-cache only (transformers)
• majentik/gemma-4-31B-RotorQuant-AWQ-8bit -- AWQ 8-bit variant
• majentik/gemma-4-31B-TurboQuant-AWQ-4bit -- TurboQuant AWQ 4-bit variant
• majentik/gemma-4-31B-RotorQuant-MLX-4bit -- MLX variant (Apple Silicon)
• RotorQuant GitHub
• llama-cpp-turboquant fork
• AutoAWQ
• vLLM

Quant trade-off (AWQ lane)

Bits	Approx size	Use case	Recommendation
4-bit	~13 GB	Activation-aware 4-bit weight quant	GPU inference (vLLM, transformers, AutoAWQ)
8-bit	~24 GB	Activation-aware 8-bit weight quant	Quality-sensitive GPU inference

(Current variant — 4bit — is bolded.)

Variants in this family

(Showing 18 sibling variants under majentik/gemma4-31b-*. The current variant — RotorQuant-AWQ-4bit — is bolded.)

Variant	Runtime	Approx size	Use case
RotorQuant	runtime modifier	n/a	KV-cache root (weight-agnostic)
RotorQuant-AWQ-4bit	transformers	~19 GB	GPU 4-bit (AutoAWQ)
RotorQuant-AWQ-8bit	transformers	~34 GB	GPU 8-bit (AutoAWQ)
RotorQuant-GGUF-IQ4_XS	llama.cpp	~27 GB	Lossy 4-bit, low-RAM CPU/edge
RotorQuant-GGUF-Q2_K	llama.cpp	~19 GB	Lossy, low-RAM CPU/edge
RotorQuant-GGUF-Q3_K_M	llama.cpp	~24 GB	Smaller 3-bit, CPU-friendly
RotorQuant-GGUF-Q4_K_M	llama.cpp	~34 GB	Balanced default
RotorQuant-GGUF-Q5_K_M	llama.cpp	~41 GB	Higher fidelity, more RAM
RotorQuant-GGUF-Q8_0	llama.cpp	~65 GB	Near-lossless reference
RotorQuant-MLX-2bit	mlx-lm	~9.9 GB	Apple Silicon, smallest
RotorQuant-MLX-4bit	mlx-lm	~19 GB	Apple Silicon balanced
RotorQuant-MLX-8bit	mlx-lm	~37 GB	Apple Silicon reference
TurboQuant	runtime modifier	n/a	KV-cache root (weight-agnostic)
TurboQuant-AWQ-4bit	transformers	~19 GB	GPU 4-bit (AutoAWQ)
TurboQuant-AWQ-8bit	transformers	~34 GB	GPU 8-bit (AutoAWQ)
TurboQuant-MLX-2bit	mlx-lm	~9.9 GB	Apple Silicon, smallest
TurboQuant-MLX-4bit	mlx-lm	~19 GB	Apple Silicon balanced
TurboQuant-MLX-8bit	mlx-lm	~37 GB	Apple Silicon reference