github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-07 20:10:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer] update bloom/llama/vit/chatglm tests (#4420)

Browse Source 
[shardformer] update bloom/llama/vit/chatglm tests

[shardformer] update opt tests

[shardformer] update opt tests

[shardformer] update bloom/llama/vit/chatglm tests

[shardformer] update bloom/llama/vit/chatglm tests

[shardformer] update bloom/llama/vit/chatglm tests
This commit is contained in:
flybird11111
2023-08-14 15:51:13 +08:00
committed by Hongxin Liu
parent 108e54a0b4
commit 328a791d10
6 changed files with 157 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -41,11 +41,15 @@ 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__ == 'LlamaModel':
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-6, rtol=1e-3)
check_loss(org_loss, sharded_loss, atol=atol, rtol=rtol)
# unwrap model
if org_model.__class__.__name__ == 'LlamaModel':
@@ -59,20 +63,24 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
row_layer_for_check = ['layers[0].self_attn.q_proj', 'embed_tokens']
col_layer_for_check = ['layers[0].self_attn.o_proj']
if stage_manager is None or stage_manager.is_first_stage():
if test_config['precision'] == 'fp32':
atol, rtol = 1e-6, 1e-4
else:
atol, rtol = 5e-3, 5e-3
check_grad(llama_model,
shard_llama_model,
row_layer_for_check,
tp_group,
atol=1e-6,
rtol=1e-4,
atol=atol,
rtol=rtol,
dim=0,
verbose=False)
check_grad(llama_model,
shard_llama_model,
col_layer_for_check,
tp_group,
atol=1e-6,
rtol=1e-4,
atol=atol,
rtol=rtol,
dim=1,
verbose=False)
@@ -80,12 +88,16 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
org_optimizer.step()
sharded_optimizer.step()
if stage_manager is None or stage_manager.is_first_stage():
if test_config['precision'] == 'fp32':
atol, rtol = 1e-4, 1e-3
else:
atol, rtol = 5e-3, 5e-3
check_weight(llama_model,
shard_llama_model,
col_layer_for_check,
tp_group,
atol=1e-4,
rtol=1e-3,
atol=atol,
rtol=rtol,
dim=1,
verbose=False)
@@ -96,33 +108,34 @@ 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': 'fp32',
}, {
'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',
}])
def run_llama_test(test_config):
# TODO: add test_config for TP+DP after supporting & debugging it
# {'tp_size': 2, 'pp_size': 1, 'enable_fused_normalization': True}
# TODO: add test_config for flash attention & jit operator after supporting
sub_model_zoo = model_zoo.get_sub_registry('transformers_llama')
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():
check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config)