generate.py not utilizing GPU in full #476
Description
I tried to run text generation with prompts using generate.py. I provided a large list of prompts, approximately 20K, and tried to run the generation on 10 RTX 8000 GPUs. However, the GPU utilization by nvidia-smi shows that the GPU utilization during generation is averaging at about 50-60%, which is not ideal. Thank you!
My configuration is:
{
# Text gen type: `input-file`, `unconditional` or `interactive`
"text-gen-type": "input-file", #"input-file",
# Params for all
"maximum_tokens": 256,
"temperature": 0.2,
"top_p": 0.95,
"top_k": 0,
"recompute": false,
# `unconditional`/`input-file`: samples
"num-samples": 100,
# input/output file
"sample-input-file": "0",
"data-path": "data/code/code_text_document",
# or for weighted datasets:
# "train-data-paths": ["data/enron/enron_text_document", "data/enron/enron_text_document"],
# "test-data-paths": ["data/enron/enron_text_document", "data/enron/enron_text_document"],
# "valid-data-paths": ["data/enron/enron_text_document", "data/enron/enron_text_document"],
# "train-data-weights": [1., 2.],
# "test-data-weights": [2., 1.],
# "valid-data-weights": [0.5, 0.4],
# If weight_by_num_documents is True, Builds dataset weights from a multinomial distribution over groups of data according to the number of documents in each group.
# WARNING: setting this to True will override any user provided weights
# "weight_by_num_documents": false,
# "weighted_sampler_alpha": 0.3,
"vocab-file": "data/code-vocab.json",
"merge-file": "data/code-merges.txt",
"save": "checkpoints",
"load": "checkpoints",
"checkpoint_validation_with_forward_pass": False,
"tensorboard-dir": "tensorboard",
"log-dir": "logs",
"use_wandb": True,
"wandb_host": "https://api.wandb.ai",
"wandb_project": "neox",
}
And the model config:
# GPT-2 pretraining setup
{
# parallelism settings ( you will want to change these based on your cluster setup, ideally scheduling pipeline stages
# across the node boundaries )
"pipe-parallel-size": 1,
"model-parallel-size": 1,
# model settings
"num-layers": 32,
"hidden-size": 2560,
"num-attention-heads": 32,
"seq-length": 2048,
"max-position-embeddings": 2048,
"norm": "layernorm",
"pos-emb": "rotary",
"no-weight-tying": true,
# these should provide some speedup but takes a while to build, set to true if desired
"scaled-upper-triang-masked-softmax-fusion": true,
"bias-gelu-fusion": true,
# optimizer settings
"zero_allow_untested_optimizer": true,
"optimizer": {
"type": "adam",
"params": {
"lr": 0.00016,
"betas": [0.9, 0.999],
"eps": 1.0e-8,
}
},
"zero_optimization": {
"stage": 1,
"allgather_partitions": True,
"allgather_bucket_size": 500000000,
"overlap_comm": True,
"reduce_scatter": True,
"reduce_bucket_size": 500000000,
"contiguous_gradients": True,
"cpu_offload": False
},
# batch / data settings
"train_micro_batch_size_per_gpu": 16,
"gradient_accumulation_steps": 1,
"data-impl": "mmap",
"split": "989,10,1",
# activation checkpointing
"checkpoint-activations": true,
"checkpoint-num-layers": 1,
"partition-activations": true,
"synchronize-each-layer": true,
# regularization
"gradient_clipping": 1.0,
"weight-decay": 0,
"hidden-dropout": 0,
"attention-dropout": 0,
# precision settings
"fp16": {
"fp16": true,
"enabled": true,
"loss_scale": 0,
"initial_scale_power": 16,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
# misc. training settings
"train-iters": 160000,
"lr-decay-iters": 160000,
"distributed-backend": "nccl",
"lr-decay-style": "cosine",
"warmup": 0.01,
"save-interval": 1000,
"eval-interval": 1000,
"eval-iters": 10,
# logging
"log-interval": 100,
"steps_per_print": 10,
"keep-last-n-checkpoints": 1,
"wall_clock_breakdown": true,
}