diff --git a/bmtrain/optim/adam.py b/bmtrain/optim/adam.py
index 03830fdd..06ee195e 100644
--- a/bmtrain/optim/adam.py
+++ b/bmtrain/optim/adam.py
@@ -107,11 +107,16 @@ def step(self, closure=None):
                     # update the steps for each param group update
                     state['step'] += 1
 
+                    if ('maximize' in group) and (group['maximize'] is True):
+                        grad = -p.grad
+                    else:
+                        grad = p.grad
+                        
                     if p.dtype == torch.half:
                         C.f_adam(
                             state["_param_fp32"],    # fp32
                             p,                      # fp16
-                            p.grad,                 # fp16
+                            grad,                 # fp16
                             state['exp_avg'],       # fp16: m
                             state["exp_avg_sq"],    # fp32: v
                             group['betas'][0], group['betas'][1],
@@ -127,7 +132,7 @@ def step(self, closure=None):
                             other_kwargs['maximize'] = False
                         F.adam(
                             [p],
-                            [p.grad / self._scale],
+                            [grad / self._scale],
                             [state['exp_avg']],
                             [state["exp_avg_sq"]],
                             [],
diff --git a/bmtrain/optim/adam_offload.py b/bmtrain/optim/adam_offload.py
index 9d19c042..4f106b07 100644
--- a/bmtrain/optim/adam_offload.py
+++ b/bmtrain/optim/adam_offload.py
@@ -129,10 +129,14 @@ def step(self, closure=None):
 
             # update parameters
             if param.dtype == torch.half:
+                if ('maximize' in group) and (group['maximize'] is True):
+                    grad = -state["_grad_fp16"]
+                else:
+                    grad = state["_grad_fp16"]
                 C.f_adam_cpu(
                     state["_param_fp32"].view(-1),
                     state["_param_fp16"].view(-1),
-                    state["_grad_fp16"].view(-1),
+                    grad.view(-1),
                     state["exp_avg"].view(-1),
                     state["exp_avg_sq"].view(-1),
                     beta1, beta2,
@@ -145,13 +149,16 @@ def step(self, closure=None):
                 param.copy_(state["_param_fp16"], non_blocking=True)
             else:
                 state["_grad_fp32"].mul_(1.0 / self._scale)
-                
+                if ('maximize' in group) and (group['maximize'] is True):
+                    grad = -state["_grad_fp32"]
+                else:
+                    grad = state["_grad_fp32"]
                 other_kwargs = {}
                 if 'maximize' in inspect.signature(F.adam).parameters:
                     other_kwargs['maximize'] = False
                 F.adam(
                     [state["_param_fp32"]],
-                    [state["_grad_fp32"]],
+                    [grad],
                     [state["exp_avg"]],
                     [state["exp_avg_sq"]],
                     [],