Performance Comparison

Short-sequence training

NVIDIA A100 * 1

Method	Bits	TGS	VRAM	Speed
HF	16	2,392	18GB	100%
HF+FA2	16	2,954	17GB	123%
Unsloth+FA2	16	4,007	16GB	168%
HF	4	2,415	9GB	101%
Unsloth+FA2	4	3,726	7GB	160%

NVIDIA A100 * 2

Method	Bits	TGS	VRAM	Speed
HF	16	2,155	29GB	100%
HF+FA2	16	2,556	28GB	119%
Unsloth+FA2	16	3,400	27GB	158%

• TGS: tokens per GPU per second
• Model: LLaMA2-7B
• Batch size: 4
• Gradient accumulation: 2
• LoRA rank: 8
• LoRA modules: all
• Max length: 1024

Long-sequence training

VRAM	1,024	2,048	4,096	8,192	16,384	32,768	65,536	100,000
FlashAttention2	6GB	7GB	9GB	12GB	19GB	32GB	OOM	OOM
Unsloth+FA2	5GB	6GB	7GB	8GB	10GB	16GB	25GB	37GB

TGS	1,024	2,048	4,096	8,192	16,384	32,768	65,536	100,000
FlashAttention2	2,295	2,741	2,926	3,128	3,542	2,216	OOM	OOM
Unsloth+FA2	2,556	3,178	3,413	3,632	4,050	2,456	1,820	1,202
Improvement	111%	116%	117%	116%	114%	111%

• TGS: tokens per GPU per second
• GPU: NVIDIA A100 40GB * 1
• Model: LLaMA2-7B
• Batch size: 1
• Gradient accumulation: 4
• LoRA rank: 8
• LoRA modules: all
• Quantization bit: 4