[unit test] refactor test tensor (#1005)

* polish test_gpt

* update op unit tests

* update test model

tests/components_to_test/__init__.py

@@ -1 +1 @@
-from . import repeated_computed_layer, resnet, nested_model, bert, no_leaf_module, simple_net
+from . import repeated_computed_layer, resnet, nested_model, bert, no_leaf_module, simple_net, gpt

tests/components_to_test/gpt.py

@@ -0,0 +1,79 @@
+import torch
+import torch.nn as nn
+from .registry import non_distributed_component_funcs
+from transformers import GPT2Config, GPT2LMHeadModel
+from .utils.dummy_data_generator import DummyDataGenerator
+from colossalai.utils.cuda import get_current_device
+
+
+class DummyDataLoader(DummyDataGenerator):
+    vocab_size = 50304
+    batch_size = 4
+    seq_len = 1024
+
+    def generate(self):
+        input_ids = torch.randint(0,
+                                  DummyDataLoader.vocab_size, (DummyDataLoader.batch_size, DummyDataLoader.seq_len),
+                                  device=get_current_device())
+        attention_mask = torch.ones_like(input_ids)
+        return input_ids, attention_mask
+
+
+class GPTLMModel(nn.Module):
+
+    def __init__(self,
+                 hidden_size=768,
+                 num_layers=12,
+                 num_attention_heads=12,
+                 max_seq_len=1024,
+                 vocab_size=50304,
+                 checkpoint=False):
+        super().__init__()
+        self.checkpoint = checkpoint
+        self.model = GPT2LMHeadModel(
+            GPT2Config(n_embd=hidden_size,
+                       n_layer=num_layers,
+                       n_head=num_attention_heads,
+                       n_positions=max_seq_len,
+                       n_ctx=max_seq_len,
+                       vocab_size=vocab_size,
+                       resid_pdrop=0.0,
+                       embd_pdrop=0.0,
+                       attn_pdrop=0.0))
+        if checkpoint:
+            self.model.gradient_checkpointing_enable()
+
+    def forward(self, input_ids, attention_mask):
+        # Only return lm_logits
+        return self.model(input_ids=input_ids, attention_mask=attention_mask, use_cache=not self.checkpoint)[0]
+
+
+def gpt2_s(checkpoint=True):
+    return GPTLMModel(checkpoint=checkpoint)
+
+
+def gpt2_m(checkpoint=True):
+    return GPTLMModel(hidden_size=1024, num_layers=24, num_attention_heads=16, checkpoint=checkpoint)
+
+
+class GPTLMLoss(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        self.loss_fn = nn.CrossEntropyLoss()
+
+    def forward(self, logits, labels):
+        shift_logits = logits[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+        # Flatten the tokens
+        return self.loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+
+@non_distributed_component_funcs.register(name='gpt2')
+def get_training_components():
+
+    trainloader = DummyDataLoader()
+    testloader = DummyDataLoader()
+
+    criterion = GPTLMLoss()
+    return gpt2_s, trainloader, testloader, torch.optim.Adam, criterion

tests/test_tensor/_utils/_util.py

@@ -1,5 +1,19 @@
+import os
+import random
+import numpy as np
 import torch
 import torch.distributed as dist
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+
+
+def set_seed(seed):
+    random.seed(seed)
+    os.environ['PYTHONHASHSEED'] = str(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.backends.cudnn.deterministic = True
 
 
 def check_equal(A, B):
@@ -25,3 +39,19 @@ def broadcast_tensor_chunk(tensor, chunk_size=1, local_rank=0):
 
 def tensor_equal(A, B):
     return torch.allclose(A, B, rtol=1e-3, atol=1e-1)
+
+
+def tensor_shard_equal(tensor: torch.Tensor, shard: torch.Tensor):
+    assert tensor.ndim == shard.ndim
+    if tensor.shape == shard.shape:
+        return tensor_equal(tensor, shard)
+    else:
+        dims_not_eq = torch.nonzero(torch.tensor(tensor.shape) != torch.tensor(shard.shape))
+        if dims_not_eq.numel() == 1:
+            # 1D shard
+            dim = dims_not_eq.item()
+            world_size = gpc.get_world_size(ParallelMode.PARALLEL_1D)
+            rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
+            return tensor_equal(tensor.chunk(world_size, dim)[rank], shard)
+        else:
+            raise NotImplementedError

tests/test_tensor/test_addmm_tp.py

@@ -11,6 +11,7 @@
 from colossalai.utils import free_port
 from functools import partial
 from colossalai.core import global_context as gpc
+from _utils import tensor_shard_equal, tensor_equal
 
 
 class Conv1D(nn.Module):
@@ -45,13 +46,6 @@ def init_1d_row(weight, bias):
         weight.set_spec(spec)
 
 
-def check_grad_1d_row(model: torch.nn.Module, weight, bias):
-    rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
-    size = gpc.get_world_size(ParallelMode.PARALLEL_1D)
-    assert torch.allclose(model.weight.grad.chunk(size, 0)[rank], weight.grad)
-    assert torch.allclose(model.bias.grad, bias.grad)
-
-
 def init_1d_col(weight, bias):
     spec = TensorSpec(
         distspec.shard(gpc.get_group(ParallelMode.PARALLEL_1D), [-1], [gpc.get_world_size(ParallelMode.PARALLEL_1D)]),
@@ -61,33 +55,27 @@ def init_1d_col(weight, bias):
         bias.set_spec(spec)
 
 
-def check_grad_1d_col(model: torch.nn.Module, weight, bias):
-    rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
-    size = gpc.get_world_size(ParallelMode.PARALLEL_1D)
-    assert torch.allclose(model.weight.grad.chunk(size, -1)[rank], weight.grad)
-    assert torch.allclose(model.bias.grad.chunk(size, -1)[rank], bias.grad)
-
-
-def run_with_spec(spec_init_func, check_grad_func):
+def run_with_spec(spec_init_func):
     model = Conv1D(4, 16).cuda()
     weight = ColoTensor(torch.nn.Parameter(model.weight.detach()))
     bias = ColoTensor(torch.nn.Parameter(model.bias.detach()))
     spec_init_func(weight, bias)
     x = torch.rand(2, 16).cuda()
     out = model(x)
     colo_out = torch.addmm(bias, x, weight)
-    assert torch.allclose(out, colo_out)
+    assert tensor_equal(out, colo_out)
     grad = torch.rand_like(out)
     out.backward(grad)
     colo_out.backward(grad)
-    check_grad_func(model, weight, bias)
+    tensor_shard_equal(model.weight.grad, weight.grad)
+    tensor_shard_equal(model.bias.grad, bias.grad)
 
 
 def run_dist(rank, world_size, port):
     config = dict(parallel=dict(tensor=dict(mode="1d", size=world_size),))
     colossalai.launch(config=config, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
-    run_with_spec(init_1d_row, check_grad_1d_row)
-    run_with_spec(init_1d_col, check_grad_1d_col)
+    run_with_spec(init_1d_row)
+    run_with_spec(init_1d_col)
 
 
 @pytest.mark.dist

tests/test_tensor/test_embedding_tp.py

@@ -12,6 +12,7 @@
 from colossalai.utils import free_port
 from colossalai.core import global_context as gpc
 from colossalai.tensor import TensorSpec, ComputePattern, ParallelAction, DistSpecManager
+from _utils import tensor_equal, tensor_shard_equal
 
 
 def init_1d_row(weight):
@@ -22,12 +23,6 @@ def init_1d_row(weight):
         weight.set_spec(spec)
 
 
-def check_grad_1d_row(model: torch.nn.Module, weight):
-    rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
-    size = gpc.get_world_size(ParallelMode.PARALLEL_1D)
-    assert torch.allclose(model.weight.grad.chunk(size, 0)[rank], weight.grad)
-
-
 def init_1d_col(weight):
     spec = TensorSpec(
         distspec.shard(gpc.get_group(ParallelMode.PARALLEL_1D), [-1], [gpc.get_world_size(ParallelMode.PARALLEL_1D)]),
@@ -36,31 +31,25 @@ def init_1d_col(weight):
         weight.set_spec(spec)
 
 
-def check_grad_1d_col(model: torch.nn.Module, weight):
-    rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
-    size = gpc.get_world_size(ParallelMode.PARALLEL_1D)
-    assert torch.allclose(model.weight.grad.chunk(size, -1)[rank], weight.grad)
-
-
-def run_with_spec(spec_init_func, check_grad_func):
+def run_with_spec(spec_init_func):
     model = torch.nn.Embedding(12, 32).cuda()
     weight = ColoTensor(torch.nn.Parameter(model.weight.detach()))
     spec_init_func(weight)
     x = torch.tensor((0, 3, 6, 9)).cuda()
     out = model(x)
     colo_out = F.embedding(x, weight)
-    assert torch.allclose(out, colo_out)
+    assert tensor_equal(out, colo_out)
     grad = torch.rand_like(out)
     out.backward(grad)
     colo_out.backward(grad)
-    check_grad_func(model, weight)
+    assert tensor_shard_equal(model.weight.grad, weight.grad)
 
 
 def run_dist(rank, world_size, port):
     config = dict(parallel=dict(tensor=dict(mode="1d", size=world_size),))
     colossalai.launch(config=config, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
-    run_with_spec(init_1d_row, check_grad_1d_row)
-    run_with_spec(init_1d_col, check_grad_1d_col)
+    run_with_spec(init_1d_row)
+    run_with_spec(init_1d_col)
 
 
 @pytest.mark.dist