github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-15 22:19:38 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer] fix opt test hanging (#4521)

Browse Source 
* [shardformer] fix opt test hanging

* fix

* test

* test

* test

* fix test

* fix test

* remove print

* add fix
This commit is contained in:
flybird11111
2023-08-30 14:50:34 +08:00
committed by GitHub
parent e241b74f24
commit d367b88785
13 changed files with 472 additions and 249 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
tests/test_shardformer/test_model/test_shard_gpt2.py
View File

@@ -9,10 +9,11 @@ from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_ad
from tests.kit.model_zoo import model_zoo
from tests.test_shardformer.test_model._utils import (
build_model_from_hybrid_plugin,
check_grad,
check_all_grad_tensors,
check_loss,
check_output_hidden_state,
check_weight,
get_grad_tensors_for_check,
run_forward_backward_with_hybrid_plugin,
unwrap_model,
)
@@ -36,6 +37,43 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
stage_manager = booster.plugin.stage_manager
tp_group = booster.plugin.tp_group
# unwrap model
gpt2 = unwrap_model(org_model, 'GPT2Model', 'transformer')
sharded_gpt2 = unwrap_model(sharded_model, 'GPT2Model', 'transformer')
col_layer_for_check = ['h[0].mlp.c_fc']
row_layer_for_check = ['wte', 'h[0].mlp.c_proj']
# Save gradient tensors for comparison between the original model and the sharded model.
grads_to_check = {}
if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
if test_config['precision'] == 'fp32':
atol, rtol = 1e-4, 1e-3
else:
atol, rtol = 5e-3, 5e-3
col_layer_grads = get_grad_tensors_for_check(gpt2,
sharded_gpt2,
col_layer_for_check,
tp_group,
atol=atol,
rtol=rtol,
dim=1,
verbose=False)
row_layer_grads = get_grad_tensors_for_check(gpt2,
sharded_gpt2,
row_layer_for_check,
tp_group,
atol=atol,
rtol=rtol,
dim=0,
verbose=False)
grads_to_check.update(col_layer_grads)
grads_to_check.update(row_layer_grads)
# optimizer executes step
org_optimizer.step()
sharded_optimizer.step()
# check last hidden state & loss
if stage_manager is None or stage_manager.is_last_stage():
if test_config['precision'] == 'fp32':
@@ -48,25 +86,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
check_loss(org_loss, sharded_loss, atol=atol, rtol=rtol)
# unwrap model
gpt2 = unwrap_model(org_model, 'GPT2Model', 'transformer')
sharded_gpt2 = unwrap_model(sharded_model, 'GPT2Model', 'transformer')
col_layer_for_check = ['h[0].mlp.c_fc']
row_layer_for_check = ['wte', 'h[0].mlp.c_proj']
# check grad
if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
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()
# check weights
if stage_manager is None or stage_manager.is_first_stage():
if test_config['precision'] == 'fp32':
atol, rtol = 5e-3, 1e-3
@@ -74,6 +94,9 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
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)
# check grads
check_all_grad_tensors(grads_to_check)
torch.cuda.empty_cache()