kastan

Kastan's group workspace

Group: Aug-05__13:54

1-8

of 8

Tags

Aug-05__13:54

BATCH_SIZE32

MICRO_BATCH_SIZE=4

NUM_EPOCHS=20

NUM_MICRO_BATCHES=8

PP=2

SLURM=513916

TP=4

WORLD_SIZE=8

Author

kastan

State

Crashed

Start time

August 5th, 2022 6:55:24 PM

Runtime

2m 15s

Tracked hours

29s

Run path

kastan/LLM-Distributed-Quantization/jtjcwm6i

Linux-4.18.0-305.49.1.el8_4.x86_64-x86_64-with-glibc2.28

Python version

3.9.12

Command

/u/kastanday/LLM-Distributed-Quantization/benchmarks/gpt/v2_train.py --config /u/kastanday/LLM-Distributed-Quantization/benchmarks/gpt/configs/gpt2_PP2_TP4_2d.py --host gpub076 --port 29500 --world_size 8 --rank 3

System Hardware

CPU count	64
GPU count	4
GPU type	NVIDIA A40

W&B CLI Version

0.13.0

Group

Aug-05__13:54

Config parameters are your model's inputs. Learn more

▶
Config parameters:{} 26 keys
- BATCH_SIZE:
  32
- clip_grad_norm:
  1
- conda_env_name:
  "col_ai_quant"
- data_dir:
  "/u/kastanday/LLM-Distributed-Quantization/datasets/small-gpt-dataset.json"
- ▶
  fp16:{} 1 key
  - mode:
    "AMP_TYPE.NAIVE"
- gpt2_medium:
  "titans.model.gpt.gpt.gpt2_medium"
- gradient_accumulation:
  1
- LEARNING_RATE:
  0.00015
- ▶
  loss:{} 1 key
  - type:
    "titans.loss.lm_loss.gpt_lmloss.GPTLMLoss"
- MICRO_BATCH_SIZE:
  4
- ▶
  model:{} 4 keys
  - checkpoint:
    true
  - dtype:
    "torch.float16"
  - max_position_embeddings:
    1,024
  - vocab_size:
    50,304
- NUM_EPOCHS:
  20
- num_gpus_per_node:
  "4"
- NUM_MICRO_BATCHES:
  8
- ▶
  optimizer:{} 2 keys
  - lr:
    0.00015
  - weight_decay:
    0.01
- ▶
  parallel:{} 2 keys
  - pipeline:
    2
  - ▶
    tensor:{} 2 keys
    - mode:
      "2d"
    - size:
      4
- PIPELINE_SIZE:
  2
- ▶
  schedule:{} 4 keys
  - num_microbatches:
    8
  - scatter_gather_tensors:
    true
  - ▶
    tensor_shape:[] 3 items
    - 0:
      4
    - 1:
      1,024
    - 2:
      3,072
  - type:
    "colossalai.engine.schedule._pipeline_schedule.PipelineSchedule"
- SEQ_LENGTH:
  1,024
- TENSOR_PARALLEL_MODE:
  "2d"
- TENSOR_PARALLEL_SIZE:
  4
- TOTAL_BATCH_SIZE:
  32
- total_gpus:
  "8"
- VOCAB_SIZE:
  50,304
- WARMUP_EPOCHS:
  1
- WEIGHT_DECAY:
  0.01

Summary metrics are your model's outputs. Learn more

▶
Summary metrics:{} 3 keys
- embed_layer_norm:
  0
- loss:
  null
- per_step_metrics.lr:
  0.000025

This run produced these artifacts as outputs. Total: 1. Learn more

wandb-history

run-jtjcwm6i-history:v0