Gemma 2 27B
Parameters
27B
Context Length
8K
Modality
Text
Architecture
Dense
License
Gemma License
Release Date
27 Jun 2024
Knowledge Cutoff
-
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
3x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
8,192 tokens
Consumer
3x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#144
| Benchmark | Score | Rank |
|---|---|---|
QA Assistant | 0.804 | 19 |
Summarization | 0.59 | 25 |
General Knowledge | 0.752 | 27 |
General Text | 1289 | 109 |
Rankings
Overall Rank
#144
Coding Rank
-
About Gemma 2 27B
Gemma 2 is a family of advanced, open models developed by Google DeepMind, stemming from the same research that informed the Gemini models. This model family aims to provide robust capabilities for a range of text generation tasks, including but not limited to question answering, summarization, and reasoning. The 27B variant is engineered for efficient inference, facilitating deployment across various hardware environments, from high-performance workstations to more constrained consumer devices.
The architecture of Gemma 2 represents a progression in Transformer design, integrating several key innovations. These include the adoption of Grouped-Query Attention (GQA) and a strategic interleaving of local and global attention layers. This architectural refinement contributes to enhanced performance and improved inference efficiency, particularly when processing extended contexts. Furthermore, the model employs Logit soft-capping for training stability and incorporates Rotary Position Embeddings (RoPE) for effective positional encoding. Notably, the smaller 2B and 9B models within the Gemma 2 family were developed using knowledge distillation from a larger teacher model.
The Gemma 2 27B model is designed to achieve a high level of performance within its parameter class, while prioritizing computational efficiency. This efficiency enables cost-effective deployment, as the model supports full precision inference on a single high-performance GPU or TPU. The model's capabilities are applicable to tasks requiring sophisticated natural language understanding and generation, making it suitable for applications in content creation, conversational AI systems, and fundamental natural language processing research.
Technical Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
32
Key-Value Heads
16
Attention Head Dimension
-
Position Embedding
ROPE
RoPE Theta
-
Sliding Window Attention
-
Sliding Window Size
-
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
GELU
Dimensions
Hidden Dimension Size
4,096
Number of Layers
46
FFN Intermediate Size (Dense)
-
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
-
About Gemma 2
Gemma 2 is Google's family of open large language models, offering 2B, 9B, and 27B parameter sizes. Built upon the Gemma architecture, it incorporates innovations such as interleaved local and global attention, logit soft-capping for training stability, and Grouped Query Attention for inference efficiency. The smaller models leverage knowledge distillation.