Skip to content

[shardformer] update bloom/llama/vit/chatglm tests for checking all optimizations. #4420

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
flybird11111 merged 1 commit into from
Aug 14, 2023
Merged

[shardformer] update bloom/llama/vit/chatglm tests for checking all optimizations. #4420

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 27 additions & 16 deletions tests/test_shardformer/test_model/test_shard_bloom.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,11 +36,14 @@ 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__ == 'BloomModel':
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__ == 'BloomModel':
Expand All @@ -54,14 +57,22 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
row_layer_for_check = ['h[0].self_attention.query_key_value', 'word_embeddings']
col_layer_for_check = ['h[0].self_attention.dense']
if stage_manager is None or stage_manager.is_first_stage():
check_grad(bloom, sharded_bloom, row_layer_for_check, tp_group, atol=1e-6, rtol=1e-5, dim=0, verbose=False)
check_grad(bloom, sharded_bloom, col_layer_for_check, tp_group, atol=1e-6, rtol=1e-5, dim=1, verbose=False)
if test_config['precision'] == 'fp32':
atol, rtol = 1e-6, 1e-5
else:
atol, rtol = 5e-3, 5e-3
check_grad(bloom, sharded_bloom, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False)
check_grad(bloom, sharded_bloom, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False)

# check weights after optimizer.step()
org_optimizer.step()
sharded_optimizer.step()
if stage_manager is None or stage_manager.is_first_stage():
check_weight(bloom, sharded_bloom, col_layer_for_check, tp_group, atol=1e-4, rtol=1e-3, dim=1, verbose=False)
if test_config['precision'] == 'fp32':
atol, rtol = 1e-4, 1e-3
else:
atol, rtol = 5e-3, 5e-3
check_weight(bloom, sharded_bloom, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False)

torch.cuda.empty_cache()

Expand All @@ -70,29 +81,29 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
'tp_size': 2,
'pp_size': 2,
'num_microbatches': 4,
'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,
'enable_fused_normalization': False,
'use_lazy_init': False
'enable_all_optimization': 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',
}])
def run_bloom_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_bloom')
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)
Expand Down
48 changes: 30 additions & 18 deletions tests/test_shardformer/test_model/test_shard_chatglm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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__ == 'ChatGLMModel':
check_output_hidden_state(org_output, sharded_output, stage_manager, atol=1e-5, rtol=1e-3, dim=1)
check_output_hidden_state(org_output, sharded_output, stage_manager, atol=atol, rtol=rtol, dim=1)

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__ == 'ChatGLMModel':
Expand All @@ -55,34 +59,42 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
row_layer_for_check = ['encoder.layers[0].self_attention.query_key_value', 'embedding.word_embeddings']
col_layer_for_check = ['encoder.layers[0].self_attention.dense']
if stage_manager is None or stage_manager.is_first_stage():
if test_config['precision'] == 'fp32':
atol, rtol = 1e-6, 1e-3
else:
atol, rtol = 5e-3, 5e-3
check_grad(chatglm_model,
shard_chatglm_model,
row_layer_for_check,
tp_group,
atol=1e-6,
rtol=1e-3,
atol=atol,
rtol=rtol,
dim=0,
verbose=False)

check_grad(chatglm_model,
shard_chatglm_model,
col_layer_for_check,
tp_group,
atol=1e-6,
rtol=1e-3,
atol=atol,
rtol=rtol,
dim=1,
verbose=False)

# check weights after optimizer.step()
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(chatglm_model,
shard_chatglm_model,
col_layer_for_check,
tp_group,
atol=1e-4,
rtol=1e-3,
atol=atol,
rtol=rtol,
dim=1,
verbose=False)

Expand All @@ -93,29 +105,29 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
'tp_size': 2,
'pp_size': 2,
'num_microbatches': 4,
'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,
'enable_fused_normalization': False,
'use_lazy_init': False
'enable_all_optimization': 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',
}])
def run_chatglm_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_chatglm')
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)
Expand Down
16 changes: 8 additions & 8 deletions tests/test_shardformer/test_model/test_shard_gpt2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,22 +63,22 @@ def unwrap(module):
row_layer_for_check = ['wte', 'h[0].mlp.c_proj']

# check grad
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():
if test_config['precision'] == 'fp32':
atol, rtol = 1e-4, 1e-3
else:
atol, rtol = 5e-3, 5e-3
check_grad(gpt2, sharded_gpt2, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False)
check_grad(gpt2, sharded_gpt2, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False)

# check weights after optimizer.step()
org_optimizer.step()
sharded_optimizer.step()
if test_config['precision'] == 'fp32':
atol, rtol = 5e-3, 1e-3
else:
atol, rtol = 5e-3, 5e-3
if stage_manager is None or stage_manager.is_first_stage():
if test_config['precision'] == 'fp32':
atol, rtol = 5e-3, 1e-3
else:
atol, rtol = 5e-3, 5e-3
check_weight(gpt2, sharded_gpt2, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False)

torch.cuda.empty_cache()
Expand Down
49 changes: 31 additions & 18 deletions tests/test_shardformer/test_model/test_shard_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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':
Expand All @@ -59,33 +63,41 @@ 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)

# check weights after optimizer.step()
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)

Expand All @@ -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)
Expand Down
Loading