ColossalAI/tests/test_shardformer/test_model/test_shard_llama.py

68 lines
2.5 KiB
Python
Raw Normal View History

import os
import pytest
import torch
import colossalai
from colossalai.logging import disable_existing_loggers
from colossalai.testing import assert_hf_output_close, clear_cache_before_run, rerun_if_address_is_in_use, spawn
from tests.kit.model_zoo import model_zoo
from tests.test_shardformer.test_model._utils import build_model, run_forward
os.environ['TRANSFORMERS_NO_ADVISORY_WARNINGS'] = 'true'
def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn):
org_output, org_loss, shard_output, shard_loss = run_forward(org_model, sharded_model, data_gen_fn,
output_transform_fn, loss_fn)
# forward check
assert_hf_output_close(org_output, shard_output, ignore_keys=['past_key_values'], rtol=1e-4)
# run backward
org_loss.backward()
shard_loss.backward()
# check grad
if hasattr(org_model, 'model'):
llama_model = org_model.model
shard_llama_model = sharded_model.model
else:
llama_model = org_model
shard_llama_model = sharded_model
org_grad = llama_model.layers[0].self_attn.q_proj.weight.grad
shard_grad = shard_llama_model.layers[0].self_attn.q_proj.weight.grad
shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(4)]
shard_grad = torch.distributed.all_gather(shard_grad_list, shard_grad)
all_shard_grad = torch.cat(shard_grad_list, dim=0)
assert torch.allclose(org_loss, shard_loss,
atol=1e-5), f"shard model loss is not equal to orgin model loss\n{org_loss}\n{shard_loss}"
assert torch.allclose(org_grad, all_shard_grad,
atol=1e-5), f"shard model grad is not equal to orgin model grad\n{org_grad}\n{shard_grad}"
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')
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():
org_model, sharded_model = build_model(world_size, model_fn)
check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn)
torch.cuda.empty_cache()
@pytest.mark.dist
@rerun_if_address_is_in_use()
@clear_cache_before_run()
def test_llama():
spawn(check_llama, 4)
if __name__ == "__main__":
test_llama()