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

* support tp+zero/input type cast for hybridplugin

* add tp+zero tests

* fix bucket arguments

tests/test_shardformer/test_model/test_shard_bert.py

@@ -56,9 +56,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
         atol, rtol = 1e-4, 1e-3
     else:
         atol, rtol = 5e-3, 5e-3
-    if stage_manager is None or stage_manager.is_first_stage():
-        #check_weight(bert.embeddings.word_embeddings, sharded_bert.embeddings.word_embeddings, tp_group, atol=1e-5, rtol=1e-3)
-        #check_weight(bert.encoder.layer[0].attention.self.query, sharded_bert.encoder.layer[0].attention.self.query, tp_group, atol=5e-3, rtol=1e-3)
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         check_grad(bert, sharded_bert, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False)
         check_grad(bert, sharded_bert, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False)
 
@@ -101,6 +99,14 @@ 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_bert_test(test_config):
 

tests/test_shardformer/test_model/test_shard_bloom.py

@@ -53,7 +53,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     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():
+    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-5
         else:
@@ -101,6 +101,14 @@ 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_bloom_test(test_config):
 

tests/test_shardformer/test_model/test_shard_chatglm.py

@@ -55,7 +55,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     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 (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-3
         else:
@@ -125,6 +125,14 @@ 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_chatglm_test(test_config):
 

tests/test_shardformer/test_model/test_shard_gpt2.py

@@ -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):

tests/test_shardformer/test_model/test_shard_llama.py

@@ -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):
 

tests/test_shardformer/test_model/test_shard_opt.py

@@ -58,7 +58,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     row_layer_for_check = ['decoder.layers[0].self_attn.q_proj', 'decoder.embed_tokens']    # 'decoder.embed_tokens'
     col_layer_for_check = ['decoder.layers[0].self_attn.out_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-3
         else:
@@ -127,6 +127,14 @@ 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_opt_test(test_config):
 

tests/test_shardformer/test_model/test_shard_t5.py

@@ -55,12 +55,12 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
 
     row_layer_for_check = ['shared', 'encoder.block[0].layer[0].SelfAttention.q']
 
-    # check weights and gradients
+    # check grad
     if test_config['precision'] == 'fp32':
         atol, rtol = 1e-5, 1e-3
     else:
         atol, rtol = 5e-3, 5e-3
-    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:
         check_grad(t5, sharded_t5, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0)
 
     # check weights after optimizer.step()
@@ -110,6 +110,14 @@ 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
 }])
 @clear_cache_before_run()
 def run_t5_test(test_config):

tests/test_shardformer/test_model/test_shard_vit.py

@@ -55,7 +55,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     row_layer_for_check = ['encoder.layer[0].attention.attention.query', 'embeddings.patch_embeddings.projection']
     col_layer_for_check = ['encoder.layer[0].attention.output.dense']
-    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-5, 1e-3
         else:
@@ -124,6 +124,14 @@ 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': False,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 def run_vit_test(test_config):