ApX logoApX logo

Falcon-7B

Parameters

7B

Context Length

2K

Modality

Text

Architecture

Dense

License

Apache 2.0

Release Date

5 Jun 2023

Knowledge Cutoff

-

System Requirements

VRAM requirements for different quantization methods and context sizes

1,024 tokens

16.22 GB VRAM

Consumer

1x RTX 4090

24GB VRAM

Datacenter

1x NVIDIA A100

80GB VRAM

Apple Silicon

1x Apple M3 Max

128GB VRAM

2,048 tokens

16.24 GB VRAM

Consumer

1x RTX 4090

24GB VRAM

Datacenter

1x NVIDIA A100

80GB VRAM

Apple Silicon

1x Apple M3 Max

128GB VRAM

Architecture Diagram

Input TokensToken EmbeddingPosition: RoPEHidden: 4.5k · Context: 2K · Vocab: 65kx 32 layersLayerNormPre-AttentionMulti-Query Attention71Q / 1KV headsHead dim: 64+LayerNormPre-FFNFeed-Forward NetworkActivation+Final LayerNormOutput Logits

Evaluation Benchmarks

No evaluation benchmarks for Falcon-7B available.

Rankings

Overall Rank

-

Coding Rank

-

About Falcon-7B

Falcon-7B is a 7 billion parameter causal decoder-only language model developed by the Technology Innovation Institute (TII). Its primary purpose is to serve as a high-performance, efficient foundation for a wide array of natural language processing tasks, encompassing both language understanding and generation capabilities. The model's design emphasizes utility within research and commercial applications, providing a robust open-source option for developers and practitioners.

Architecturally, Falcon-7B builds upon the transformer framework, incorporating specific modifications to optimize performance and efficiency. A core innovation is the implementation of Multi-Query Attention (MQA), which enhances inference speed and reduces memory overhead by allowing all attention heads to share a single key and value projection. This contrasts with traditional multi-head attention that uses separate projections for each head. Furthermore, the model integrates FlashAttention, a technique that significantly accelerates both training and inference computations through memory-efficient attention mechanisms. Positional encoding is handled via Rotary Positional Embeddings (RoPE), contributing to the model's ability to process sequence information effectively. The decoder blocks feature a parallel arrangement of attention and Multi-Layer Perceptron (MLP) components, unified by a single layer normalization.

Trained on a vast dataset of 1,500 billion tokens, primarily sourced from the RefinedWeb corpus and augmented with curated datasets, Falcon-7B exhibits proficiency in generating coherent and contextually relevant text. Its architectural optimizations are specifically tailored to facilitate efficient inference, making it well-suited for deployment in scenarios where rapid response times are critical. Common use cases include text generation, chatbots, summarization, and question answering. The model is released under the Apache 2.0 license, permitting broad commercial use and fostering its integration into various AI-driven solutions and continued research endeavors.

Technical Specifications

Attention

Attention Structure

Multi-Query Attention

Attention Heads

71

Key-Value Heads

1

Attention Head Dimension

-

Position Embedding

ROPE

RoPE Theta

-

Sliding Window Attention

No

Sliding Window Size

-

Sliding Window Ratio

-

Linear Attention

-

Linear Attention Ratio

-

Normalization

Layer Normalization

Activation Function

-

Dimensions

Hidden Dimension Size

4,544

Number of Layers

32

FFN Intermediate Size (Dense)

-

Multi-Token Prediction Heads

-

Tokenizer

Vocabulary Size

65,024

About Falcon

The TII Falcon model family comprises causal decoder-only language models (7B, 40B). Their architecture, adapted from GPT-3, integrates rotary positional embeddings, Multi-Query Attention for inference efficiency, and FlashAttention for accelerated operations. Models are trained on the RefinedWeb dataset.


Other Falcon Models
Falcon-7B: Specifications and GPU VRAM Requirements