Spaces:
Runtime error
Runtime error
File size: 4,548 Bytes
0558aa4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 |
# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import os
import pytest
import torch
from megatron.core.optimizer import OptimizerConfig
from nemo import lightning as nl
from nemo.collections import llm
from nemo.collections.llm.gpt.data import PreTrainingDataModule
from nemo.collections.nlp.modules.common.tokenizer_utils import get_nmt_tokenizer
from nemo.lightning.pytorch.callbacks import DdpParityChecker
def make_parser():
parser = argparse.ArgumentParser(description='Train a small GPT model using NeMo 2.0')
parser.add_argument('--data-path', type=str, help="Path to data file")
parser.add_argument('--vocab-path', type=str, help="Path to vocab file")
parser.add_argument('--merges-path', type=str, help="Path to merges file")
return parser
def wrap_config(config, trainer):
class ConfigWrapper(type(config)):
def configure_model(self, tokenizer, vp_stage=None) -> "MCoreGPTModel":
return make_byzantine_model_wrapper(super().configure_model(tokenizer), trainer)
config.__class__ = ConfigWrapper
return config
def make_byzantine_model_wrapper(model, trainer):
class ByzantineModel(type(model)):
def forward(self, *ans, **kwargs):
ans = super().forward(*ans, **kwargs)
with torch.no_grad():
import random
rank = int(os.environ['LOCAL_RANK'])
if rank != 1:
return ans
for opt in trainer.strategy.model.optim._optimizers:
for g in opt.param_groups:
for param in g['params']:
param.fill_(random.uniform(0, 1))
return ans
model.__class__ = ByzantineModel
return model
@pytest.mark.skip(reason="tested with GH")
def test_failing(trainer, ddp_parity, optim, data, tokenizer):
config = llm.Llama2Config7B(num_layers=2)
config = wrap_config(config, trainer)
model = llm.LlamaModel(config, tokenizer=tokenizer, optim=optim)
trainer.fit(model, data)
@pytest.mark.skip(reason="tested with GH")
def test_working(trainer, ddp_parity, optim, data, tokenizer):
config = llm.Llama2Config7B(num_layers=2)
model = llm.LlamaModel(config, tokenizer=tokenizer, optim=optim)
trainer.fit(model, data)
def make_trainer_optim(args):
ddp_parity = DdpParityChecker(1)
trainer = nl.Trainer(
devices=2,
max_steps=4,
accelerator="gpu",
strategy=nl.MegatronStrategy(
ckpt_load_optimizer=False,
ckpt_save_optimizer=False,
),
plugins=nl.MegatronMixedPrecision(precision="bf16-mixed"),
limit_val_batches=1,
num_sanity_val_steps=0,
log_every_n_steps=1,
logger=None,
callbacks=[ddp_parity],
)
optim = nl.MegatronOptimizerModule(
config=OptimizerConfig(
optimizer="adam",
lr=1e-5,
use_distributed_optimizer=False,
fp16=False,
bf16=True,
params_dtype=torch.float32,
),
)
tokenizer = get_nmt_tokenizer(
"megatron",
"GPT2BPETokenizer",
vocab_file=args.vocab_path,
merges_file=args.merges_path,
)
data = PreTrainingDataModule(
paths=args.data_path,
seq_length=2048,
global_batch_size=32,
seed=1234,
tokenizer=tokenizer,
)
return trainer, ddp_parity, optim, data, tokenizer
@pytest.mark.skip(reason="tested with GH")
def main():
args = make_parser().parse_args()
trainer, ddp_parity, optim, data, tokenizer = make_trainer_optim(args)
test_failing(trainer, ddp_parity, optim, data, tokenizer)
if trainer.should_stop != True:
raise ValueError("DDP parity checking failed.")
try:
test_working(*make_trainer_optim(args))
print("DDP parity checking worked as expected")
except:
raise
if __name__ == "__main__":
main()
|