fix bugs in CPU adam (#633)

* add cpu adam counter for all cpu adam

* fixed updating error in adam kernel

colossalai/kernel/cuda_native/csrc/cpu_adam.cpp

@@ -493,7 +493,7 @@ int adam_step(int optimizer_id,
                 grads_ptr,
                 exp_avg_ptr,
                 exp_avg_sq_ptr,
-                params_c.size(0),
+                params_c.numel(),
                 (params.options().dtype() == at::kHalf),
                 (grads.options().dtype() == at::kHalf),
                 loss_scale);

colossalai/nn/optimizer/__init__.py

@@ -1,3 +1,4 @@
+from .utils import CPU_ADAM_CNT
 from .colossalai_optimizer import ColossalaiOptimizer
 from .fused_adam import FusedAdam
 from .fused_lamb import FusedLAMB
@@ -7,4 +8,5 @@ from .lars import Lars
 from .cpu_adam import CPUAdam
 from .hybrid_adam import HybridAdam
 
-__all__ = ['ColossalaiOptimizer', 'FusedLAMB', 'FusedAdam', 'FusedSGD', 'Lamb', 'Lars', 'CPUAdam', 'HybridAdam']
+__all__ = ['ColossalaiOptimizer', 'FusedLAMB', 'FusedAdam', 'FusedSGD',
+           'Lamb', 'Lars', 'CPUAdam', 'HybridAdam', 'CPU_ADAM_CNT']

colossalai/nn/optimizer/cpu_adam.py

@@ -2,6 +2,7 @@ import math
 import torch
 
 from colossalai.registry import OPTIMIZERS
+from colossalai.nn.optimizer import CPU_ADAM_CNT
 
 
 @OPTIMIZERS.register_module
@@ -51,7 +52,6 @@ class CPUAdam(torch.optim.Optimizer):
         https://openreview.net/forum?id=ryQu7f-RZ
     """
 
-    optimizer_id = 0
     # Number of fp32 shards for per parameter
     # Param weight, grad, momentum and variance
     num_fp32_shards_per_param = 4
@@ -68,8 +68,7 @@ class CPUAdam(torch.optim.Optimizer):
 
         default_args = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, bias_correction=bias_correction)
         super(CPUAdam, self).__init__(model_params, default_args)
-        self.opt_id = CPUAdam.optimizer_id
-        CPUAdam.optimizer_id = CPUAdam.optimizer_id + 1
+        self.opt_id = CPU_ADAM_CNT()
         self.adamw_mode = adamw_mode
         try:
             import cpu_adam

colossalai/nn/optimizer/hybrid_adam.py

@@ -2,6 +2,7 @@ import torch
 
 from colossalai.utils import multi_tensor_applier
 from colossalai.registry import OPTIMIZERS
+from colossalai.nn.optimizer import CPU_ADAM_CNT
 
 
 @OPTIMIZERS.register_module
@@ -50,7 +51,6 @@ class HybridAdam(torch.optim.Optimizer):
         https://openreview.net/forum?id=ryQu7f-RZ
     """
 
-    optimizer_id = 0
     # Number of fp32 shards for per parameter
     # Param weight, grad, momentum and variance
     num_fp32_shards_per_param = 4
@@ -67,8 +67,7 @@ class HybridAdam(torch.optim.Optimizer):
 
         default_args = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, bias_correction=bias_correction)
         super(HybridAdam, self).__init__(model_params, default_args)
-        self.opt_id = HybridAdam.optimizer_id
-        HybridAdam.optimizer_id = HybridAdam.optimizer_id + 1
+        self.opt_id = CPU_ADAM_CNT()
         self.adamw_mode = adamw_mode
         try:
             import cpu_adam

colossalai/nn/optimizer/utils.py

@@ -0,0 +1,14 @@
+class CpuAdamCounter(object):
+    """Used to record the total number of CPU Adam.
+    We must use it to avoid hybrid cpu adam and cpu adam using the same id.
+    """
+
+    def __init__(self):
+        self.number = 0
+
+    def __call__(self):
+        self.number += 1
+        return self.number - 1
+
+
+CPU_ADAM_CNT = CpuAdamCounter()

tests/test_moe/test_moe_zero_optim.py

@@ -1,7 +1,6 @@
 from functools import partial
 
 import colossalai
-from colossalai.utils.cuda import get_current_device
 import pytest
 import torch
 import torch.multiprocessing as mp
@@ -51,11 +50,10 @@ def _run_step(model, optimizer, data, label, criterion, grad_handler):
 @parameterize("use_cpuadam", [True, False])
 @parameterize("shard_strategy_class", [TensorShardStrategy, BucketTensorShardStrategy])
 def _run_test_sharded_optim_v2(cpu_offload, shard_strategy_class, use_cpuadam, gpu_margin_mem_ratio=0.0):
-    MOE_CONTEXT.reset_loss()
     shard_strategy = shard_strategy_class()
     if use_cpuadam and cpu_offload is False:
         return
-
+    MOE_CONTEXT.reset_loss()
     get_components_func = non_distributed_component_funcs.get_callable('no_leaf_module')
     _, train_dataloader, _, optimizer_class, criterion = get_components_func()