github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-13 13:11:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer]update t5 tests for using all optimizations. (#4407)

Browse Source 
* [shardformer] gpt2 tests fix

[shardformer] test all optimizations (#4399)

[shardformer] test all optimizations

[shardformer] test all optimizations

[shardformer] test all optimizations

[shardformer] gpt2 tests fix

* [shardformer]update t5 to use all optimizations
This commit is contained in:
flybird11111
2023-08-14 15:49:13 +08:00
committed by Hongxin Liu
parent 1edc9b5fb3
commit 108e54a0b4
3 changed files with 33 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
tests/test_shardformer/test_model/test_shard_t5.py
View File

@@ -37,11 +37,15 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
# check last hidden state & loss
if stage_manager is None or stage_manager.is_last_stage():
if test_config['precision'] == 'fp32':
atol, rtol = 1e-5, 1e-3
else:
atol, rtol = 5e-3, 5e-3
if org_model.__class__.__name__ != 'T5ForConditionalGeneration':
check_output_hidden_state(org_output, sharded_output, stage_manager, atol=1e-5, rtol=1e-3)
check_output_hidden_state(org_output, sharded_output, stage_manager, atol=atol, rtol=rtol)
check_loss(org_loss, sharded_loss, atol=1e-5, rtol=1e-3)
check_loss(org_loss, sharded_loss, atol=atol, rtol=rtol)
# unwrap model
t5 = org_model
@@ -50,14 +54,22 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
row_layer_for_check = ['shared', 'encoder.block[0].layer[0].SelfAttention.q']
# check weights and gradients
if test_config['precision'] == 'fp32':
atol, rtol = 1e-5, 1e-3
else:
atol, rtol = 5e-3, 5e-3
if stage_manager is None or stage_manager.is_first_stage():
check_grad(t5, sharded_t5, row_layer_for_check, tp_group, atol=1e-5, rtol=1e-3, dim=0)
check_grad(t5, sharded_t5, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0)
# check weights after optimizer.step()
org_optimizer.step()
sharded_optimizer.step()
if test_config['precision'] == 'fp32':
atol, rtol = 1e-4, 1e-3
else:
atol, rtol = 5e-3, 5e-3
if stage_manager is None or stage_manager.is_first_stage():
check_weight(t5, sharded_t5, row_layer_for_check, tp_group, atol=1e-4, rtol=1e-3, dim=0, verbose=False)
check_weight(t5, sharded_t5, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False)
torch.cuda.empty_cache()
@@ -66,23 +78,29 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
'tp_size': 2,
'pp_size': 2,
'num_microbatches': 2,
'enable_fused_normalization': True,
'use_lazy_init': True
'enable_all_optimization': True,
'use_lazy_init': True,
'precision': 'fp16',
'initial_scale': 1,
}, {
'tp_size': 1,
'pp_size': 2,
'num_microbatches': 4,
'use_lazy_init': False
'use_lazy_init': False,
'precision': 'fp16',
'initial_scale': 1,
}, {
'tp_size': 4,
'pp_size': 1,
'enable_fused_normalization': True,
'use_lazy_init': False
'enable_all_optimization': True,
'use_lazy_init': False,
'precision': 'fp32',
}, {
'tp_size': 1,
'pp_size': 4,
'num_microbatches': 4,
'use_lazy_init': False
'use_lazy_init': False,
'precision': 'fp32',
}])
@clear_cache_before_run()
def run_t5_test(test_config):
@@ -93,7 +111,6 @@ def run_t5_test(test_config):
# TODO: add test_config for flash attention & jit operator after supporting
sub_model_zoo = model_zoo.get_sub_registry('transformers_t5')
test_config['precision'] = 'float' # Do not use fp16/bf16 in testing
for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():