Open
Description
I'm using keras_nlp and I get Segmentation fault (core dumped) in the following code:
import os
os.environ["KERAS_BACKEND"] = "torch" # "jax" or "tensorflow" or "torch"
import keras_nlp
import keras_core as keras
import keras_core.backend as K
import torch
import tensorflow as tf
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib as mpl
cmap = mpl.cm.get_cmap('coolwarm')
class CFG:
verbose = 0 # Verbosity
wandb = True # Weights & Biases logging
competition = 'llm-detect-ai-generated-text' # Competition name
_wandb_kernel = 'awsaf49' # WandB kernel
comment = 'DebertaV3-MaxSeq_200-ext_s-torch' # Comment description
preset = "deberta_v3_base_en" # Name of pretrained models
sequence_length = 200 # Input sequence length
device = 'TPU' # Device
seed = 42 # Random seed
num_folds = 5 # Total folds
selected_folds = [0, 1, 2] # Folds to train on
epochs = 3 # Training epochs
batch_size = 3 # Batch size
drop_remainder = True # Drop incomplete batches
cache = True # Caches data after one iteration, use only with `TPU` to avoid OOM
scheduler = 'cosine' # Learning rate scheduler
class_names = ["real", "fake"] # Class names [A, B, C, D, E]
num_classes = len(class_names) # Number of classes
class_labels = list(range(num_classes)) # Class labels [0, 1, 2, 3, 4]
label2name = dict(zip(class_labels, class_names)) # Label to class name mapping
name2label = {v: k for k, v in label2name.items()} # Class name to label mapping
keras.utils.set_random_seed(CFG.seed)
def get_device():
"Detect and intializes GPU/TPU automatically"
try:
# Connect to TPU
tpu = tf.distribute.cluster_resolver.TPUClusterResolver.connect()
# Set TPU strategy
strategy = tf.distribute.TPUStrategy(tpu)
print(f'> Running on TPU', tpu.master(), end=' | ')
print('Num of TPUs: ', strategy.num_replicas_in_sync)
device=CFG.device
except:
# If TPU is not available, detect GPUs
gpus = tf.config.list_logical_devices('GPU')
ngpu = len(gpus)
# Check number of GPUs
if ngpu:
# Set GPU strategy
strategy = tf.distribute.MirroredStrategy(gpus) # single-GPU or multi-GPU
# Print GPU details
print("> Running on GPU", end=' | ')
print("Num of GPUs: ", ngpu)
device='GPU'
else:
# If no GPUs are available, use CPU
print("> Running on CPU")
strategy = tf.distribute.get_strategy()
device='CPU'
return strategy, device
# Initialize GPU/TPU/TPU-VM
strategy, CFG.device = get_device()
CFG.replicas = strategy.num_replicas_in_sync
BASE_PATH = '/some/path/'
print(1)
preprocessor = keras_nlp.models.DebertaV3Preprocessor.from_preset(
preset=CFG.preset, # Name of the model
sequence_length=CFG.sequence_length, # Max sequence length, will be padded if shorter
)
print(2)
The complete log is as follows:
$python test.py
Using PyTorch backend.
/mypath/test.py:22: MatplotlibDeprecationWarning: The get_cmap function was deprecated in Matplotlib 3.7 and will be removed two minor releases later. Use ``matplotlib.colormaps[name]`` or ``matplotlib.colormaps.get_cmap(obj)`` instead.
cmap = mpl.cm.get_cmap('coolwarm')
> Running on GPU | Num of GPUs: 1
1
Segmentation fault (core dumped)
What's wrong with this? I'm using a Nvidia-A100 with 80GB memory, it should be large enough I guess. On CPU It works well, this only happens on GPU.
Thank you all for helping me!!!