github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-09 04:50:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer] support tp+zero for shardformer (#4472)

Browse Source 
* support tp+zero/input type cast for hybridplugin

* add tp+zero tests

* fix bucket arguments
This commit is contained in:
Baizhou Zhang
2023-08-21 12:04:52 +08:00
committed by GitHub
parent 8739aa7fa0
commit 1c7df566e2
9 changed files with 136 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/test_shardformer/test_model/test_shard_llama.py
View File

@@ -60,7 +60,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
# check grad
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 (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
if test_config['precision'] == 'fp32':
atol, rtol = 1e-6, 1e-4
else:
@@ -135,6 +135,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
'enable_all_optimization': True,
'use_lazy_init': False,
'precision': 'fp32'
}, {
'tp_size': 2,
'pp_size': 1,
'enable_all_optimization': True,
'use_lazy_init': True,
'zero_stage': 2,
'precision': 'fp16',
'initial_scale': 1
}])
def run_llama_test(test_config):