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

* 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

View File

@@ -56,7 +56,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
row_layer_for_check = ['wte', 'h[0].mlp.c_proj']
# check grad
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-4, 1e-3
else:
@@ -120,6 +120,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
'use_lazy_init': True,
'enable_sequence_parallelism': True,
'precision': 'fp32',
}, {
'tp_size': 2,
'pp_size': 1,
'enable_all_optimization': True,
'use_lazy_init': True,
'zero_stage': 2,
'precision': 'fp16',
'initial_scale': 1
}])
@clear_cache_before_run()
def run_gpt2_test(test_config):