github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-06 19:40:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

[test] Hotfix/fix some model test and refactor check util api (#4369)

Browse Source 
* fix llama test

* fix test bug of bert, blip2, bloom, gpt2

* fix llama test

* fix opt test

* fix sam test

* fix sam test

* fix t5 test

* fix vit test

* fix whisper test

* fix whisper test

* polish code

* adjust allclose parameter

* Add mistakenly deleted code

* addjust allclose

* change loss function for some base model
This commit is contained in:
Bin Jia
2023-08-03 14:51:36 +08:00
committed by Hongxin Liu
parent c3ca53cf05
commit 5c6f183192
16 changed files with 135 additions and 336 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
tests/test_shardformer/test_model/test_shard_bert.py
View File

@@ -15,10 +15,18 @@ from colossalai.testing import (
spawn,
)
from tests.kit.model_zoo import model_zoo
from tests.test_shardformer.test_model._utils import build_model, check_state_dict, run_forward
from tests.test_shardformer.test_model._utils import build_model, check_grad, check_state_dict, run_forward
def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn):
# unwarp model
if org_model.__class__.__name__ == 'BertModel':
bert = org_model
sharded_bert = sharded_model
else:
bert = org_model.bert
sharded_bert = sharded_model.bert
# check forward
org_output, org_loss, shard_output, shard_loss = run_forward(org_model, sharded_model, data_gen_fn,
output_transform_fn, loss_fn)
@@ -32,42 +40,10 @@ def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transfo
atol=1e-5), f"shard model loss is not equal to orgin model loss\n{org_loss}\n{shard_loss}"
# check grad
if org_model.__class__.__name__ == 'BertModel':
bert = org_model
sharded_bert = sharded_model
else:
bert = org_model.bert
sharded_bert = sharded_model.bert
# 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
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
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}"
col_layer_for_check = ['encoder.layer[0].attention.self.query', 'embeddings.word_embeddings']
row_layer_for_check = ['encoder.layer[0].attention.output.dense']
check_grad(bert, sharded_bert, col_layer_for_check, atol=1e-7, rtol=1e-3, dim=0, verbose=False)
check_grad(bert, sharded_bert, row_layer_for_check, atol=1e-7, rtol=1e-3, dim=1, verbose=False)
@parameterize('enable_fused_normalization', [False, True])