2023-06-13 06:44:40 +00:00
|
|
|
import os
|
|
|
|
|
|
|
|
import pytest
|
|
|
|
import torch
|
2023-08-11 07:43:23 +00:00
|
|
|
from torch import distributed as dist
|
2023-06-13 06:44:40 +00:00
|
|
|
|
|
|
|
import colossalai
|
|
|
|
from colossalai.logging import disable_existing_loggers
|
2023-08-18 13:29:25 +00:00
|
|
|
from colossalai.shardformer.layer.utils import Randomizer
|
2023-08-11 07:43:23 +00:00
|
|
|
from colossalai.tensor.d_tensor.api import clear_layout_converter
|
|
|
|
from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
|
2023-06-21 01:32:46 +00:00
|
|
|
from tests.kit.model_zoo import model_zoo
|
2023-08-11 07:43:23 +00:00
|
|
|
from tests.test_shardformer.test_model._utils import (
|
|
|
|
build_model_from_hybrid_plugin,
|
|
|
|
check_grad,
|
|
|
|
check_loss,
|
|
|
|
check_output_hidden_state,
|
|
|
|
check_weight,
|
|
|
|
run_forward_backward_with_hybrid_plugin,
|
2023-08-16 08:11:57 +00:00
|
|
|
unwrap_model,
|
2023-08-11 07:43:23 +00:00
|
|
|
)
|
2023-06-13 06:44:40 +00:00
|
|
|
|
|
|
|
os.environ['TRANSFORMERS_NO_ADVISORY_WARNINGS'] = 'true'
|
2023-06-19 09:57:37 +00:00
|
|
|
|
2023-06-21 01:32:46 +00:00
|
|
|
|
2023-08-11 07:43:23 +00:00
|
|
|
def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config):
|
|
|
|
|
|
|
|
org_model, org_optimizer, sharded_model, sharded_optimizer, criterion, booster = \
|
|
|
|
build_model_from_hybrid_plugin(model_fn, loss_fn, test_config)
|
|
|
|
|
|
|
|
org_loss, org_output, sharded_loss, sharded_output = \
|
|
|
|
run_forward_backward_with_hybrid_plugin(
|
|
|
|
org_model,
|
|
|
|
sharded_model,
|
|
|
|
sharded_optimizer,
|
|
|
|
data_gen_fn,
|
|
|
|
output_transform_fn,
|
|
|
|
criterion,
|
|
|
|
booster)
|
2023-06-21 01:32:46 +00:00
|
|
|
|
2023-08-11 07:43:23 +00:00
|
|
|
stage_manager = booster.plugin.stage_manager
|
|
|
|
tp_group = booster.plugin.tp_group
|
2023-06-19 09:57:37 +00:00
|
|
|
|
2023-08-11 07:43:23 +00:00
|
|
|
# check last hidden state & loss
|
|
|
|
if stage_manager is None or stage_manager.is_last_stage():
|
2023-08-14 07:51:13 +00:00
|
|
|
if test_config['precision'] == 'fp32':
|
|
|
|
atol, rtol = 1e-5, 1e-3
|
|
|
|
else:
|
|
|
|
atol, rtol = 5e-3, 5e-3
|
2023-06-19 09:57:37 +00:00
|
|
|
|
2023-08-11 07:43:23 +00:00
|
|
|
if org_model.__class__.__name__ == 'LlamaModel':
|
2023-08-14 07:51:13 +00:00
|
|
|
check_output_hidden_state(org_output, sharded_output, stage_manager, atol=atol, rtol=rtol)
|
2023-08-11 07:43:23 +00:00
|
|
|
|
2023-08-14 07:51:13 +00:00
|
|
|
check_loss(org_loss, sharded_loss, atol=atol, rtol=rtol)
|
2023-06-30 08:16:44 +00:00
|
|
|
|
|
|
|
# unwrap model
|
2023-08-16 08:11:57 +00:00
|
|
|
llama_model = unwrap_model(org_model, 'LlamaModel', 'model')
|
|
|
|
shard_llama_model = unwrap_model(sharded_model, 'LlamaModel', 'model')
|
2023-08-03 06:51:36 +00:00
|
|
|
# check grad
|
2023-08-11 07:43:23 +00:00
|
|
|
row_layer_for_check = ['layers[0].self_attn.q_proj', 'embed_tokens']
|
|
|
|
col_layer_for_check = ['layers[0].self_attn.o_proj']
|
2023-08-21 04:04:52 +00:00
|
|
|
if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
|
2023-08-14 07:51:13 +00:00
|
|
|
if test_config['precision'] == 'fp32':
|
|
|
|
atol, rtol = 1e-6, 1e-4
|
|
|
|
else:
|
|
|
|
atol, rtol = 5e-3, 5e-3
|
2023-08-11 07:43:23 +00:00
|
|
|
check_grad(llama_model,
|
|
|
|
shard_llama_model,
|
|
|
|
row_layer_for_check,
|
|
|
|
tp_group,
|
2023-08-14 07:51:13 +00:00
|
|
|
atol=atol,
|
|
|
|
rtol=rtol,
|
2023-08-11 07:43:23 +00:00
|
|
|
dim=0,
|
|
|
|
verbose=False)
|
|
|
|
check_grad(llama_model,
|
|
|
|
shard_llama_model,
|
|
|
|
col_layer_for_check,
|
|
|
|
tp_group,
|
2023-08-14 07:51:13 +00:00
|
|
|
atol=atol,
|
|
|
|
rtol=rtol,
|
2023-08-11 07:43:23 +00:00
|
|
|
dim=1,
|
|
|
|
verbose=False)
|
|
|
|
|
|
|
|
# check weights after optimizer.step()
|
|
|
|
org_optimizer.step()
|
|
|
|
sharded_optimizer.step()
|
|
|
|
if stage_manager is None or stage_manager.is_first_stage():
|
2023-08-14 07:51:13 +00:00
|
|
|
if test_config['precision'] == 'fp32':
|
|
|
|
atol, rtol = 1e-4, 1e-3
|
|
|
|
else:
|
|
|
|
atol, rtol = 5e-3, 5e-3
|
2023-08-11 07:43:23 +00:00
|
|
|
check_weight(llama_model,
|
|
|
|
shard_llama_model,
|
|
|
|
col_layer_for_check,
|
|
|
|
tp_group,
|
2023-08-14 07:51:13 +00:00
|
|
|
atol=atol,
|
|
|
|
rtol=rtol,
|
2023-08-11 07:43:23 +00:00
|
|
|
dim=1,
|
|
|
|
verbose=False)
|
|
|
|
|
|
|
|
torch.cuda.empty_cache()
|
2023-06-13 06:44:40 +00:00
|
|
|
|
|
|
|
|
2023-08-11 07:43:23 +00:00
|
|
|
@parameterize('test_config', [{
|
|
|
|
'tp_size': 2,
|
|
|
|
'pp_size': 2,
|
|
|
|
'num_microbatches': 2,
|
2023-08-14 07:51:13 +00:00
|
|
|
'enable_all_optimization': True,
|
|
|
|
'use_lazy_init': True,
|
|
|
|
'precision': 'fp16',
|
|
|
|
'initial_scale': 1,
|
2023-08-11 07:43:23 +00:00
|
|
|
}, {
|
|
|
|
'tp_size': 1,
|
|
|
|
'pp_size': 2,
|
|
|
|
'num_microbatches': 4,
|
2023-08-14 07:51:13 +00:00
|
|
|
'use_lazy_init': False,
|
|
|
|
'precision': 'fp32',
|
2023-08-11 07:43:23 +00:00
|
|
|
}, {
|
|
|
|
'tp_size': 4,
|
|
|
|
'pp_size': 1,
|
2023-08-14 07:51:13 +00:00
|
|
|
'enable_all_optimization': True,
|
|
|
|
'use_lazy_init': False,
|
|
|
|
'precision': 'fp32',
|
2023-08-11 07:43:23 +00:00
|
|
|
}, {
|
|
|
|
'tp_size': 1,
|
|
|
|
'pp_size': 4,
|
|
|
|
'num_microbatches': 4,
|
2023-08-16 08:11:57 +00:00
|
|
|
'enable_all_optimization': False,
|
2023-08-14 07:51:13 +00:00
|
|
|
'use_lazy_init': False,
|
2023-08-16 08:11:57 +00:00
|
|
|
'precision': 'fp32'
|
|
|
|
}, {
|
|
|
|
'tp_size': 2,
|
|
|
|
'pp_size': 1,
|
|
|
|
'enable_all_optimization': True,
|
|
|
|
'use_lazy_init': False,
|
|
|
|
'precision': 'fp32'
|
2023-08-21 04:04:52 +00:00
|
|
|
}, {
|
|
|
|
'tp_size': 2,
|
|
|
|
'pp_size': 1,
|
|
|
|
'enable_all_optimization': True,
|
|
|
|
'use_lazy_init': True,
|
|
|
|
'zero_stage': 2,
|
|
|
|
'precision': 'fp16',
|
|
|
|
'initial_scale': 1
|
2023-08-28 02:51:16 +00:00
|
|
|
}, {
|
|
|
|
'tp_size': 1,
|
|
|
|
'pp_size': 2,
|
|
|
|
'num_microbatches': 2,
|
|
|
|
'enable_all_optimization': True,
|
|
|
|
'use_lazy_init': True,
|
|
|
|
'zero_stage': 1,
|
|
|
|
'precision': 'fp16',
|
|
|
|
'initial_scale': 1
|
2023-08-11 07:43:23 +00:00
|
|
|
}])
|
|
|
|
def run_llama_test(test_config):
|
|
|
|
|
2023-06-21 01:32:46 +00:00
|
|
|
sub_model_zoo = model_zoo.get_sub_registry('transformers_llama')
|
2023-08-11 07:43:23 +00:00
|
|
|
|
2023-06-21 01:32:46 +00:00
|
|
|
for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
|
2023-08-11 07:43:23 +00:00
|
|
|
check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config)
|
|
|
|
|
|
|
|
clear_layout_converter()
|
2023-08-18 13:29:25 +00:00
|
|
|
Randomizer.reset_index()
|
2023-06-13 06:44:40 +00:00
|
|
|
torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
2023-08-23 07:05:24 +00:00
|
|
|
@parameterize('test_config', [
|
|
|
|
{
|
|
|
|
'tp_size': 2,
|
|
|
|
'pp_size': 2,
|
|
|
|
'num_microbatches': 4,
|
|
|
|
'enable_all_optimization': False,
|
|
|
|
'use_lazy_init': False,
|
|
|
|
'precision': 'fp32',
|
|
|
|
'initial_scale': 1,
|
|
|
|
},
|
|
|
|
])
|
|
|
|
def run_llama_3d_test(test_config):
|
|
|
|
sub_model_zoo = model_zoo.get_sub_registry('transformers_llama')
|
|
|
|
|
|
|
|
for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
|
|
|
|
check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config)
|
|
|
|
|
|
|
|
clear_layout_converter()
|
|
|
|
Randomizer.reset_index()
|
|
|
|
torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
2023-07-04 01:57:03 +00:00
|
|
|
def check_llama(rank, world_size, port):
|
|
|
|
disable_existing_loggers()
|
|
|
|
colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
|
2023-08-11 07:43:23 +00:00
|
|
|
run_llama_test()
|
2023-07-04 01:57:03 +00:00
|
|
|
|
|
|
|
|
2023-08-23 07:05:24 +00:00
|
|
|
def check_llama_3d(rank, world_size, port):
|
|
|
|
disable_existing_loggers()
|
|
|
|
colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
|
|
|
|
run_llama_3d_test()
|
|
|
|
|
|
|
|
|
2023-06-13 06:44:40 +00:00
|
|
|
@pytest.mark.dist
|
|
|
|
@rerun_if_address_is_in_use()
|
2023-06-19 09:57:37 +00:00
|
|
|
@clear_cache_before_run()
|
2023-06-13 06:44:40 +00:00
|
|
|
def test_llama():
|
|
|
|
spawn(check_llama, 4)
|
|
|
|
|
|
|
|
|
2023-08-23 07:05:24 +00:00
|
|
|
@pytest.mark.largedist
|
|
|
|
@rerun_if_address_is_in_use()
|
|
|
|
@clear_cache_before_run()
|
|
|
|
def test_llama_3d():
|
|
|
|
spawn(check_llama_3d, 8)
|
|
|
|
|
|
|
|
|
2023-06-13 06:44:40 +00:00
|
|
|
if __name__ == "__main__":
|
|
|
|
test_llama()
|
2023-08-23 07:05:24 +00:00
|
|
|
test_llama_3d()
|