Falcon

Released in April 2023, TII’s Falcon is an Apache 2.0 license model based on the transformer decoder framework with key adjustments such as using multi-group attention, RoPE, parallel attention and MLP blocks, and removal of bias from linear layers. More information is available in the companion paper: “The Falcon Series of Open Language Models”. Falcon is offered in sizes 7B, 40B, and 180B.

• Falcon
• Data Preparation
• Training with Predefined Configurations
• Model Evaluation
• Parameter Efficient Fine-Tuning (PEFT)

Feature	Status
Data parallelism	✓
Tensor parallelism	✓
Pipeline parallelism	✓
Interleaved Pipeline Parallelism Sched	N/A
Sequence parallelism	✓
Selective activation checkpointing	✓
Gradient checkpointing	✓
Partial gradient checkpointing	✓
FP32/TF32	✓
AMP/FP16	✗
BF16	✓
TransformerEngine/FP8	✗
Multi-GPU	✓
Multi-Node	✓
Inference	N/A
Slurm	✓
Base Command Manager	✓
Base Command Platform	✓
Distributed data preprcessing	✓
NVfuser	✗
P-Tuning and Prompt Tuning	✓
IA3 and Adapter learning	✓
Distributed Optimizer	✓
Distributed Checkpoint	✓
Fully Shared Data Parallel	N/A