github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-12 12:47:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer] made tensor parallelism configurable (#4144)

Browse Source 
* [shardformer] made tensor parallelism configurable

* polish code
This commit is contained in:
Frank Lee
2023-07-04 09:57:03 +08:00
parent 74257cb446
commit 1fb0d95df0
15 changed files with 819 additions and 673 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
tests/test_shardformer/test_model/test_shard_bert.py
View File

@@ -3,7 +3,14 @@ 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 colossalai.tensor.d_tensor.api import is_customized_distributed_tensor, is_distributed_tensor
from colossalai.testing import (
assert_hf_output_close,
clear_cache_before_run,
parameterize,
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
@@ -33,34 +40,48 @@ def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transfo
# compare self attention grad
org_grad = bert.encoder.layer[0].attention.self.query.weight.grad
shard_grad = sharded_bert.encoder.layer[0].attention.self.query.weight.grad
shard_weight = sharded_bert.encoder.layer[0].attention.self.query.weight
shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(2)]
shard_grad = torch.distributed.all_gather(shard_grad_list, shard_grad)
all_shard_grad = torch.cat(shard_grad_list, dim=0)
if is_distributed_tensor(shard_weight) or is_customized_distributed_tensor(shard_weight):
shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(2)]
shard_grad = torch.distributed.all_gather(shard_grad_list, shard_grad)
all_shard_grad = torch.cat(shard_grad_list, dim=0)
else:
all_shard_grad = shard_grad
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{all_shard_grad}"
# compare embedding grad
org_grad = bert.embeddings.word_embeddings.weight.grad
shard_grad = sharded_bert.embeddings.word_embeddings.weight.grad
shard_weight = sharded_bert.embeddings.word_embeddings.weight
if is_distributed_tensor(shard_weight) or is_customized_distributed_tensor(shard_weight):
shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(2)]
shard_grad = torch.distributed.all_gather(shard_grad_list, shard_grad)
all_shard_grad = torch.cat(shard_grad_list, dim=0)
else:
all_shard_grad = shard_grad
shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(2)]
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_grad, all_shard_grad,
atol=1e-5), f"shard model grad is not equal to orgin model grad\n{org_grad}\n{all_shard_grad}"
@parameterize('enable_fused_normalization', [True, False])
@parameterize('enable_tensor_parallelism', [True, False])
def run_bert_test(enable_fused_normalization, enable_tensor_parallelism):
sub_model_zoo = model_zoo.get_sub_registry('transformers_bert')
for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
org_model, sharded_model = build_model(model_fn, enable_fused_normalization, enable_tensor_parallelism)
check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn)
torch.cuda.empty_cache()
def check_bert(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_bert')
for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
org_model, sharded_model = build_model(model_fn)
check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn)
torch.cuda.empty_cache()
run_bert_test()
@pytest.mark.dist