github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-15 06:00:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

[misc] update pre-commit and run all files (#4752)

Browse Source 
* [misc] update pre-commit

* [misc] run pre-commit

* [misc] remove useless configuration files

* [misc] ignore cuda for clang-format
This commit is contained in:
Hongxin Liu
2023-09-19 14:20:26 +08:00
committed by GitHub
parent 3c6b831c26
commit 079bf3cb26
1268 changed files with 50037 additions and 38444 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
tests/test_shardformer/test_model/test_shard_blip2.py
View File

@@ -16,16 +16,18 @@ from tests.test_shardformer.test_model._utils import build_model, check_grad, ru
def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn):
# check forward
org_output, org_loss, shard_output, shard_loss = run_forward(org_model, sharded_model, data_gen_fn,
output_transform_fn, loss_fn)
assert_hf_output_close(org_output, shard_output, ignore_keys=['past_key_values'])
org_output, org_loss, shard_output, shard_loss = run_forward(
org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn
)
assert_hf_output_close(org_output, shard_output, ignore_keys=["past_key_values"])
# do backward
org_loss.backward()
shard_loss.backward()
assert torch.allclose(org_loss, shard_loss,
atol=1e-5), f"shard model loss is not equal to orgin model loss\n{org_loss}\n{shard_loss}"
assert torch.allclose(
org_loss, shard_loss, atol=1e-5
), f"shard model loss is not equal to orgin model loss\n{org_loss}\n{shard_loss}"
# check grad
@@ -34,26 +36,29 @@ def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transfo
# check grad
col_layer_for_check = [
'vision_model.encoder.layers[0].self_attn.qkv', 'qformer.encoder.layer[0].attention.attention.query',
'language_model.model.decoder.layers[0].self_attn.k_proj'
"vision_model.encoder.layers[0].self_attn.qkv",
"qformer.encoder.layer[0].attention.attention.query",
"language_model.model.decoder.layers[0].self_attn.k_proj",
]
row_layer_for_check = [
'vision_model.encoder.layers[0].self_attn.projection', 'qformer.encoder.layer[0].attention.output.dense',
'language_model.model.decoder.layers[0].self_attn.out_proj'
"vision_model.encoder.layers[0].self_attn.projection",
"qformer.encoder.layer[0].attention.output.dense",
"language_model.model.decoder.layers[0].self_attn.out_proj",
]
check_grad(blip2, sharded_blip2, col_layer_for_check, atol=1e-6, rtol=1e-5, dim=0, verbose=False)
check_grad(blip2, sharded_blip2, row_layer_for_check, atol=1e-6, rtol=1e-5, dim=1, verbose=False)
@parameterize('enable_fused_normalization', [True, False])
@parameterize('enable_tensor_parallelism', [True, False])
@parameterize('enable_flash_attention', [True, False])
@parameterize('enable_jit_fused', [True, False])
@parameterize("enable_fused_normalization", [True, False])
@parameterize("enable_tensor_parallelism", [True, False])
@parameterize("enable_flash_attention", [True, False])
@parameterize("enable_jit_fused", [True, False])
def run_blip2_test(enable_fused_normalization, enable_tensor_parallelism, enable_flash_attention, enable_jit_fused):
sub_model_zoo = model_zoo.get_sub_registry('transformers_blip2')
sub_model_zoo = model_zoo.get_sub_registry("transformers_blip2")
for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
org_model, sharded_model = build_model(model_fn, enable_fused_normalization, enable_tensor_parallelism,
enable_flash_attention, enable_jit_fused)
org_model, sharded_model = build_model(
model_fn, enable_fused_normalization, enable_tensor_parallelism, enable_flash_attention, enable_jit_fused
)
check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn)
torch.cuda.empty_cache()
@@ -61,7 +66,7 @@ def run_blip2_test(enable_fused_normalization, enable_tensor_parallelism, enable
def check_blip2(rank, world_size, port):
disable_existing_loggers()
colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
run_blip2_test()