[DOC] Benchmark LLM Studio on a10g GPUs

[ozahavi]

We should have a benchmark for running LLM Studio on g5.12xlarge and g5.48xlarge machines.

[pascal-pfeiffer]

hardware setup	llm backbone	quantization	train	validation
8xA10G	h2oai/h2ogpt-4096-llama2-7b	bfloat16	11:35	3:32
4xA10G	h2oai/h2ogpt-4096-llama2-7b	bfloat16	21:13	06:35
2xA10G	h2oai/h2ogpt-4096-llama2-7b	bfloat16	37:04	12:21
1xA10G	h2oai/h2ogpt-4096-llama2-7b	bfloat16	1:25:29	15:50
8xA10G	h2oai/h2ogpt-4096-llama2-7b	nf4	14:26	06:13
4xA10G	h2oai/h2ogpt-4096-llama2-7b	nf4	26:55	11:59
2xA10G	h2oai/h2ogpt-4096-llama2-7b	nf4	48:24	23:37
1xA10G	h2oai/h2ogpt-4096-llama2-7b	nf4	1:26:59	42:17
8xA10G	h2oai/h2ogpt-4096-llama2-13b	bfloat16	OOM	OOM
4xA10G	h2oai/h2ogpt-4096-llama2-13b	bfloat16	OOM	OOM
2xA10G	h2oai/h2ogpt-4096-llama2-13b	bfloat16	OOM	OOM
1xA10G	h2oai/h2ogpt-4096-llama2-13b	bfloat16	OOM	OOM
8xA10G	h2oai/h2ogpt-4096-llama2-13b	nf4	25:07	10:58
4xA10G	h2oai/h2ogpt-4096-llama2-13b	nf4	48:43	21:25
2xA10G	h2oai/h2ogpt-4096-llama2-13b	nf4	1:30:45	42:06
1xA10G	h2oai/h2ogpt-4096-llama2-13b	nf4	2:44:36	1:14:20
8xA10G	h2oai/h2ogpt-4096-llama2-70b	nf4	OOM	OOM
4xA10G	h2oai/h2ogpt-4096-llama2-70b	nf4	OOM	OOM
2xA10G	h2oai/h2ogpt-4096-llama2-70b	nf4	OOM	OOM
1xA10G	h2oai/h2ogpt-4096-llama2-70b	nf4	OOM	OOM
---	---	---	---	---
4xA100 80GB	h2oai/h2ogpt-4096-llama2-7b	bfloat16	7:04	3:55
2xA100 80GB	h2oai/h2ogpt-4096-llama2-7b	bfloat16	13:14	7:23
1xA100 80GB	h2oai/h2ogpt-4096-llama2-7b	bfloat16	23:36	13:25
4xA100 80GB	h2oai/h2ogpt-4096-llama2-7b	nf4	9:44	6:30
2xA100 80GB	h2oai/h2ogpt-4096-llama2-7b	nf4	18:34	12:16
1xA100 80GB	h2oai/h2ogpt-4096-llama2-7b	nf4	34:06	21:51
4xA100 80GB	h2oai/h2ogpt-4096-llama2-13b	bfloat16	11:46	5:56
2xA100 80GB	h2oai/h2ogpt-4096-llama2-13b	bfloat16	21:54	11:17
1xA100 80GB	h2oai/h2ogpt-4096-llama2-13b	bfloat16	39:10	18:55
4xA100 80GB	h2oai/h2ogpt-4096-llama2-13b	nf4	16:51	10:35
2xA100 80GB	h2oai/h2ogpt-4096-llama2-13b	nf4	32:05	21:00
1xA100 80GB	h2oai/h2ogpt-4096-llama2-13b	nf4	59:11	36:53
4xA100 80GB	h2oai/h2ogpt-4096-llama2-70b	nf4	1:13:33	46:02
2xA100 80GB	h2oai/h2ogpt-4096-llama2-70b	nf4	2:20:44	1:33:42
1xA100 80GB	h2oai/h2ogpt-4096-llama2-70b	nf4	4:23:57	2:44:51

Runtimes were gathered with default parameters:

architecture:
    backbone_dtype: int4
    force_embedding_gradients: false
    gradient_checkpointing: true
    intermediate_dropout: 0.0
    pretrained: true
    pretrained_weights: ''
augmentation:
    random_parent_probability: 0.0
    skip_parent_probability: 0.0
    token_mask_probability: 0.0
dataset:
    add_eos_token_to_answer: true
    add_eos_token_to_prompt: true
    add_eos_token_to_system: true
    answer_column: output
    chatbot_author: H2O.ai
    chatbot_name: h2oGPT
    data_sample: 1.0
    data_sample_choice:
    - Train
    - Validation
    limit_chained_samples: false
    mask_prompt_labels: true
    parent_id_column: None
    personalize: false
    prompt_column:
    - instruction
    system_column: None
    text_answer_separator: <|answer|>
    text_prompt_start: <|prompt|>
    text_system_start: <|system|>
    train_dataframe: /data/user/oasst/train_full.pq
    validation_dataframe: None
    validation_size: 0.01
    validation_strategy: automatic
environment:
    compile_model: false
    find_unused_parameters: false
    gpus:
    - '0'
    - '1'
    - '2'
    - '3'
    - '4'
    - '5'
    - '6'
    - '7'
    huggingface_branch: main
    mixed_precision: true
    number_of_workers: 8
    seed: -1
    trust_remote_code: true
    use_fsdp: false
experiment_name: default-8-a10g
llm_backbone: h2oai/h2ogpt-4096-llama2-7b
logging:
    logger: None
    neptune_project: ''
output_directory: /output/...
prediction:
    batch_size_inference: 0
    do_sample: false
    max_length_inference: 256
    metric: BLEU
    metric_gpt_model: gpt-3.5-turbo-0301
    min_length_inference: 2
    num_beams: 1
    num_history: 4
    repetition_penalty: 1.2
    stop_tokens: ''
    temperature: 0.3
    top_k: 0
    top_p: 1.0
problem_type: text_causal_language_modeling
tokenizer:
    add_prefix_space: false
    add_prompt_answer_tokens: false
    max_length: 512
    max_length_answer: 256
    max_length_prompt: 256
    padding_quantile: 1.0
    use_fast: true
training:
    batch_size: 2
    differential_learning_rate: 1.0e-05
    differential_learning_rate_layers: []
    drop_last_batch: true
    epochs: 1
    evaluate_before_training: false
    evaluation_epochs: 1.0
    grad_accumulation: 1
    gradient_clip: 0.0
    learning_rate: 0.0001
    lora: true
    lora_alpha: 16
    lora_dropout: 0.05
    lora_r: 4
    lora_target_modules: ''
    loss_function: TokenAveragedCrossEntropy
    optimizer: AdamW
    save_best_checkpoint: false
    schedule: Cosine
    train_validation_data: false
    warmup_epochs: 0.0
    weight_decay: 0.0

[psinger]

@pascal-pfeiffer shall we add these to the docs / README?

[pascal-pfeiffer]

Certainly

Docs would be a good place, I guess. And maybe a subset for the README and then link to the docs from there.