[shardformer] shardformer support opt models (#4091)

* [shardformer] shardformer support opt models

* [shardformer] shardformer support opt models, fix

* [shardformer] shardformer support opt models, fix

* [shardformer] shardformer support opt models, fix

tests/test_shardformer/test_model/_utils.py

@@ -25,7 +25,6 @@ def run_forward(original_model, sharded_model, data_gen_fn, output_transform_fn,
     # switch to train mode
     original_model.train()
     sharded_model.train()
-
     # run forward
     org_output = original_model(**data)
     org_output = output_transform_fn(org_output)
@@ -34,5 +33,4 @@ def run_forward(original_model, sharded_model, data_gen_fn, output_transform_fn,
     shard_output = sharded_model(**data)
     shard_output = output_transform_fn(shard_output)
     shard_loss = loss_fn(shard_output)
-
-    return org_output, org_loss, shard_output, shard_loss
+    return org_output, org_loss, shard_output, shard_loss

tests/test_shardformer/test_model/test_shard_opt.py

@@ -0,0 +1,67 @@
+import copy
+import os
+
+import pytest
+import torch
+
+import colossalai
+from colossalai.logging import disable_existing_loggers
+from colossalai.testing import (
+    assert_hf_output_close,
+    check_state_dict_equal,
+    clear_cache_before_run,
+    rerun_if_address_is_in_use,
+    spawn,
+)
+from tests.kit.model_zoo import model_zoo
+from tests.test_shardformer.test_model._utils import build_model, run_forward
+
+os.environ['TRANSFORMERS_NO_ADVISORY_WARNINGS'] = 'true'
+
+
+def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn):
+    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'], rtol=1e-4)
+
+    # run backward
+    org_loss.backward()
+    shard_loss.backward()
+
+    # check grad
+    if hasattr(org_model, 'model'):
+        opt_model = org_model.model
+        shard_opt_model = sharded_model.model
+    else:
+        opt_model = org_model
+        shard_opt_model = sharded_model
+
+    org_grad = opt_model.decoder.layers[0].self_attn.q_proj.weight.grad
+    shard_grad = shard_opt_model.decoder.layers[0].self_attn.q_proj.weight.grad
+
+    shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(4)]
+    shard_grad = torch.distributed.all_gather(shard_grad_list, shard_grad)
+    all_shard_grad = torch.cat(shard_grad_list, dim=0)
+    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_grad, all_shard_grad,
+                          atol=1e-5), f"shard model grad is not equal to orgin model grad\n{org_grad}\n{shard_grad}"
+
+
+def check_OPTModel(rank, world_size, port):
+    disable_existing_loggers()
+    colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
+
+    sub_model_zoo = model_zoo.get_sub_registry('transformers_opt')
+    for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
+        org_model, sharded_model = build_model(world_size, model_fn)
+        check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn)
+
+    torch.cuda.empty_cache()
+
+
+@pytest.mark.dist
+@rerun_if_address_is_in_use()
+@clear_cache_before_run()
+def test_OPTModel():
+    spawn(check_OPTModel, 4)