[Shardformer] Support the Qwen2 model (#5699)

* feat: support qwen2 model

* fix: modify model config and add Qwen2RMSNorm

* fix qwen2 model conflicts

* test: add qwen2 shard test

* to: add qwen2 auto policy

* support qwen model

* fix the conflicts

* add try catch

* add transformers version for qwen2

* add the ColoAttention for the qwen2 model

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* add the unit test version check

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix the test input bug

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix the version check

* fix the version check

---------

Co-authored-by: Wenhao Chen <cwher@outlook.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

tests/kit/model_zoo/transformers/__init__.py

@@ -17,3 +17,8 @@ try:
     from .mistral import *
 except ImportError:
     print("This version of transformers doesn't support mistral.")
+
+try:
+    from .qwen2 import *
+except ImportError:
+    print("This version of transformers doesn't support qwen2.")

tests/kit/model_zoo/transformers/qwen2.py

@@ -0,0 +1,89 @@
+import torch
+import transformers
+
+from ..registry import ModelAttribute, model_zoo
+
+try:
+    from transformers import Qwen2Config
+
+    HAS_QWEN2 = True
+except ImportError:
+    HAS_QWEN2 = False
+
+if HAS_QWEN2:
+    # ===============================
+    # Register Qwen2
+    # ===============================
+
+    def data_gen():
+        # the input ids are corresponding to the sentence
+        # 'Hello, my dog is cute'
+        #
+        # the code is give below:
+        # -----------------------------------
+        # from transformers import Qwen2TokenizerFast
+        # tokenizer = Qwen2TokenizerFast.from_pretrained("Qwen/Qwen1.5-7B-Chat")
+        # input = 'Hello, my dog is cute'
+        # tokenized_input = tokenizer(input, return_tensors='pt').to('cuda')
+        # -----------------------------------
+
+        input_ids = torch.Tensor(
+            [[9707, 11, 847, 5562, 374, 13, 123, 18838], [9707, 11, 847, 5562, 374, 17, 89, 18838]]
+        ).long()
+        attention_mask = torch.Tensor([[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]]).long()
+        return dict(input_ids=input_ids, attention_mask=attention_mask)
+
+    # label is needed for casual lm
+    def data_gen_for_casual_lm():
+        data = data_gen()
+        labels = data["input_ids"].clone()
+        data["labels"] = labels
+        return data
+
+    # transform the output to a dict
+    output_transform_fn = lambda x: x
+
+    # function to get the loss
+    loss_fn = lambda output: output["last_hidden_state"].mean()
+    loss_fn_for_casual_lm = lambda output: output["loss"]
+    loss_fn_for_seq_classification = lambda output: output["logits"].mean()
+
+    config = Qwen2Config(
+        hidden_size=128,
+        intermediate_size=256,
+        max_window_layers=4,
+        num_attention_heads=16,
+        num_hidden_layers=4,
+        num_key_value_heads=16,
+    )
+
+    config.pad_token_id = 0
+
+    # register the following models
+    # transformers.Qwen2Model,
+    # transformers.Qwen2ForCausalLM,
+    # transformers.Qwen2ForSequenceClassification,
+    model_zoo.register(
+        name="transformers_qwen2",
+        model_fn=lambda: transformers.Qwen2Model(config),
+        data_gen_fn=data_gen,
+        output_transform_fn=output_transform_fn,
+        loss_fn=loss_fn,
+        model_attribute=ModelAttribute(has_control_flow=True),
+    )
+    model_zoo.register(
+        name="transformers_qwen2_for_casual_lm",
+        model_fn=lambda: transformers.Qwen2ForCausalLM(config),
+        data_gen_fn=data_gen_for_casual_lm,
+        output_transform_fn=output_transform_fn,
+        loss_fn=loss_fn_for_casual_lm,
+        model_attribute=ModelAttribute(has_control_flow=True),
+    )
+    model_zoo.register(
+        name="transformers_qwen2_for_sequence_classification",
+        model_fn=lambda: transformers.Qwen2ForSequenceClassification(config),
+        data_gen_fn=data_gen,
+        output_transform_fn=output_transform_fn,
+        loss_fn=loss_fn_for_seq_classification,
+        model_attribute=ModelAttribute(has_control_flow=True),
+    )

tests/test_shardformer/test_model/test_shard_qwen2.py

@@ -0,0 +1,235 @@
+import os
+
+import pytest
+import torch
+import transformers
+
+import colossalai
+from colossalai.logging import disable_existing_loggers
+from colossalai.shardformer.layer.utils import Randomizer
+from colossalai.tensor.d_tensor.api import clear_layout_converter
+from colossalai.testing import clear_cache_before_run, parameterize, 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_from_hybrid_plugin,
+    check_all_grad_tensors,
+    check_loss,
+    check_output_hidden_state,
+    check_weight,
+    get_grad_tensors_for_check,
+    run_forward_backward_with_hybrid_plugin,
+    unwrap_model,
+)
+
+os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "true"
+
+
+def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config):
+    org_model, org_optimizer, sharded_model, sharded_optimizer, criterion, booster = build_model_from_hybrid_plugin(
+        model_fn, loss_fn, test_config
+    )
+
+    org_loss, org_output, sharded_loss, sharded_output = run_forward_backward_with_hybrid_plugin(
+        org_model, sharded_model, sharded_optimizer, data_gen_fn, output_transform_fn, criterion, booster
+    )
+
+    stage_manager = booster.plugin.stage_manager
+    tp_group = booster.plugin.tp_group
+
+    # unwrap model
+    qwen2_model = unwrap_model(org_model, "Qwen2Model", "model")
+    shard_qwen2_model = unwrap_model(sharded_model, "Qwen2Model", "model")
+
+    row_layer_for_check = ["layers[0].self_attn.q_proj", "embed_tokens"]
+    col_layer_for_check = ["layers[0].self_attn.o_proj"]
+
+    # Save gradient tensors for comparison between the original model and the sharded model before optimizer step.
+    grads_to_check = {}
+    if (stage_manager is None or stage_manager.is_first_stage(ignore_chunk=True)) and booster.plugin.zero_stage == 0:
+        if test_config["precision"] == "fp32":
+            atol, rtol = 1e-6, 1e-4
+        else:
+            atol, rtol = 5e-3, 5e-3
+        row_layer_grads = get_grad_tensors_for_check(
+            qwen2_model, shard_qwen2_model, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False
+        )
+        col_layer_grads = get_grad_tensors_for_check(
+            qwen2_model, shard_qwen2_model, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False
+        )
+        grads_to_check.update(col_layer_grads)
+        grads_to_check.update(row_layer_grads)
+
+    # optimizer executes step
+    org_optimizer.step()
+    sharded_optimizer.step()
+
+    # check last hidden state & loss
+    if stage_manager is None or stage_manager.is_last_stage(ignore_chunk=True):
+        if test_config["precision"] == "fp32":
+            atol, rtol = 1e-5, 1e-3
+        else:
+            atol, rtol = 5e-3, 5e-3
+
+        if org_model.__class__.__name__ == "Qwen2Model":
+            check_output_hidden_state(org_output, sharded_output, stage_manager, atol=atol, rtol=rtol)
+
+        check_loss(org_loss, sharded_loss, atol=atol, rtol=rtol)
+
+    # check weights
+    if stage_manager is None or stage_manager.is_first_stage(ignore_chunk=True):
+        if test_config["precision"] == "fp32":
+            atol, rtol = 1e-4, 1e-3
+        else:
+            atol, rtol = 5e-3, 5e-3
+        check_weight(
+            qwen2_model, shard_qwen2_model, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False
+        )
+
+    # check grads
+    check_all_grad_tensors(grads_to_check)
+
+    torch.cuda.empty_cache()
+
+
+@parameterize(
+    "test_config",
+    [
+        {
+            "tp_size": 2,
+            "pp_size": 2,
+            "num_microbatches": 2,
+            "enable_all_optimization": True,
+            "use_lazy_init": True,
+            "precision": "fp16",
+            "initial_scale": 1,
+        },
+        {
+            "tp_size": 1,
+            "pp_size": 2,
+            "num_microbatches": 4,
+            "use_lazy_init": False,
+            "precision": "fp32",
+        },
+        {
+            "tp_size": 4,
+            "pp_size": 1,
+            "enable_all_optimization": True,
+            "use_lazy_init": False,
+            "precision": "fp32",
+        },
+        {
+            "tp_size": 1,
+            "pp_size": 4,
+            "num_microbatches": 4,
+            "enable_all_optimization": False,
+            "use_lazy_init": False,
+            "precision": "fp32",
+        },
+        {"tp_size": 2, "pp_size": 1, "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,
+        },
+        {
+            "tp_size": 1,
+            "pp_size": 2,
+            "num_microbatches": 2,
+            "enable_all_optimization": True,
+            "use_lazy_init": True,
+            "zero_stage": 1,
+            "precision": "fp16",
+            "initial_scale": 1,
+        },
+    ],
+)
+def run_qwen2_test(test_config):
+    sub_model_zoo = model_zoo.get_sub_registry("transformers_qwen2")
+
+    for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
+        check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config)
+
+    clear_layout_converter()
+    Randomizer.reset_index()
+    torch.cuda.empty_cache()
+
+
+@parameterize(
+    "test_config",
+    [
+        {
+            "tp_size": 2,
+            "pp_size": 2,
+            "num_microbatches": 4,
+            "enable_all_optimization": False,
+            "use_lazy_init": False,
+            "precision": "fp32",
+            "initial_scale": 1,
+        },
+        {
+            "tp_size": 2,
+            "pp_size": 2,
+            "num_microbatches": 4,
+            "enable_all_optimization": False,
+            "use_lazy_init": False,
+            "precision": "fp16",
+            "zero_stage": 1,
+            "initial_scale": 1,
+        },
+        {
+            "tp_size": 2,
+            "pp_size": 2,
+            "pp_style": "interleaved",
+            "num_model_chunks": 2,
+            "num_microbatches": 4,
+            "enable_all_optimization": False,
+            "precision": "fp16",
+            "zero_stage": 1,
+            "initial_scale": 1,
+        },
+    ],
+)
+def run_qwen2_3d_test(test_config):
+    sub_model_zoo = model_zoo.get_sub_registry("transformers_qwen2")
+
+    for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
+        check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config)
+
+    clear_layout_converter()
+    Randomizer.reset_index()
+    torch.cuda.empty_cache()
+
+
+def check_qwen2(rank, world_size, port):
+    disable_existing_loggers()
+    colossalai.launch(rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
+    run_qwen2_test()
+
+
+def check_qwen2_3d(rank, world_size, port):
+    disable_existing_loggers()
+    colossalai.launch(rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
+    run_qwen2_3d_test()
+
+
+@pytest.mark.skipif(transformers.__version__ < "4.39.1", reason="Requires transformers version 4.39.1 or later")
+@rerun_if_address_is_in_use()
+@clear_cache_before_run()
+def test_qwen2():
+    spawn(check_qwen2, 4)
+
+
+@pytest.mark.skipif(transformers.__version__ < "4.39.1", reason="Requires transformers version 4.39.1 or later")
+@rerun_if_address_is_in_use()
+@clear_cache_before_run()
+def test_qwen2_3d():
+    spawn(check_qwen2_3d, 8)
+
+
+if __name__ == "__main__":
+    test_qwen2()
+    test_qwen2_3d()