paddlepaddle--paddlenlp/slm/examples/benchmark/peft

Benchmark Results

配置

• 硬件: A100-80G with NVLink, 具体卡数见表
• Torch 环境: 见 torch/requirements.txt
• FP16配置: torch 使用 cuda amp fp16, paddle 使用 fp16 O2 opt level, intokens 设置为 1024, 并开启了 flash attention

Bloom

• 数据: 10k 条Chinese-Vicuna/guanaco_belle_merge_v1.0

Model	Method	Num GPUs	Batch Size	Paddle Setup	Paddle Effective Tokens/s	Torch Setup	Torch Effective Tokens/s	Speedup
Bloomz-7b1-mt	LoRA	1	4		4097.03		1980.32	107%
Bloomz-7b1-mt	Finetune	4	8	MP 4	4136.69	ZeRO 3	1702.00	143%
Bloomz-7b1-mt	Finetune	4	16	MP 4	4359.72	ZeRO 3	2849.90	53%

多卡分布式实验记录

• Finetuning with 4 GPUs

Model	Setup	Paddle Effective Tokens /s	Torch Effective Tokens /s	Speedup
Bloomz-7b1-mt	bsz 8 MP4	7421.09	N/A	N/A
Bloomz-7b1-mt	bsz 8 ZeRO 3	6063.23	1702.00	256%
Bloomz-7b1-mt	bsz 8 ZeRO 2	5191.47	1891.16	175%
Bloomz-7b1-mt	bsz 16 MP4	8214.55	N/A	N/A
Bloomz-7b1-mt	bsz 16 ZeRO 3	5822.23	2849.90	104
Bloomz-7b1-mt	bsz 16 ZeRO 2	5572.13	2719.92	105%

Llama

• 数据: 使用10k 条tatsu-lab/alpaca

Model	Method	Num GPUs	Batch Size	Paddle Setup	Paddle Effective Tokens/s	Torch Setup	Torch Effective Tokens/s	Speedup
Llama-7b	LoRA	1	4		4406.23		1895.90	132%
Llama-13b	LoRA	1	4		1975.94		1101.85	79%
Llama-13b	LoRA	1	8	recompute	1869.60	gradient ckpt	768.26	143%
Llama-7b	Finetune	4	8	MP4	3275.90	ZeRO 2	1621.52	102%
Llama-7b	Finetune	4	16	sharding 2	6798.72	ZeRO 2	2465.55	176%
Llama-13b	Finetune	4	8	MP4 recompute	1938.19	ZeRO 3	736.19	127%
Llama-65b	LoRA	4	8	MP4 recompute	840.57	gradient ckpt, bits 4, max_memory_MB 50000, qlora	327.75	156%
Llama-65b	LoRA	4	16	MP4 recompute	993.38	gradient ckpt, bits 4, max_memory_MB 50000, qlora	405.90	122%

多卡分布式实验记录

• Finetuning with 4 GPUs

Model	Setup	Paddle Effective Tokens /s	Torch Effective Tokens /s	Speedup
LLaMA-7b	bsz 8 MP4	3841.61	N/A	N/A
LLaMA-7b	bsz 8 ZeRO 3	4189.43	1177.93	256%
LLaMA-7b	bsz 8 ZeRO 2	4611.10	1621.52	184%
LLaMA-7b	bsz 16 (4*4) MP4	4829.47	N/A	N/A
LLaMA-7b	bsz 16 ZeRO 3	4048.61	2268.16	78%
LLaMA-7b	bsz 16 ZeRO 2	3463.45	2465.55	40%
LLaMA-13b	bsz 8 MP4 recompute	2509.50	N/A	N/A
LLaMA-13b	bsz 8 ZeRO 3	1867.99	736.19	154%
LLaMA-13b	bsz 8 ZeRO 2	1201.75	OOM	N/A

ChatGLM

Model	Method	Num GPUs	Batch Size	Paddle Setup	Paddle Effective Tokens/s	Torch Setup	Torch Effective Tokens/s	Speedup
chatglm-6b	LoRA	1	4		4216.76		1866.48	126%
chatglm-6b	Finetune	4	8	MP 4	3799.78	ZeRO 2	2124.17	79%
chatglm-6b	Finetune	4	16	MP 4	5720.21	ZeRO 3	3191.35	79%

多卡分布式实验记录

• Finetuning with 4 GPUs

Model	Setup	Paddle Effective Tokens /s	Torch Effective Tokens /s	Speedup
chatglm-6b	bsz 8 MP4	4564.94	N/A	N/A
chatglm-6b	bsz 8 ZeRO 3	6480.36	1840.99	252%
chatglm-6b	bsz 8 ZeRO 2	4707.74	2124.17	122%
chatglm-6b	bsz 16 MP4	4972.21	N/A	N/A
chatglm-6b	bsz 16 ZeRO 3	5282.28	3184.26	66%
chatglm-6b	bsz 16 ZeRO 2	5751.00	3151.07	83%

GPT 3

Model	Method	Num GPUs	Batch Size	Paddle Setup	Paddle Effective Tokens/s	Torch Setup	Torch Effective Tokens/s	Speedup
gpt3-6.7b	LoRA	1	4		3450.06		1186.74	191%
gpt3-13b	LoRA	1	4		2008.40		969.60	107%
gpt3-6.7b	Finetune	4	8	MP 4	3301.49	ZeRO 2	1441.65	129%
gpt3-13b	Finetune	4	8	MP 4	1890.38	ZeRO 2	783.26	141%
gpt3-6.7b	Finetune	4	16	MP 4	4666.19	ZeRO 3	1756.42	166%