fix(pt): optimize graph memory usage (#4006)

- Remove atomic virial graph.
- Remove force graph during inference.

After this, the lammps memory saves **50% for dpa1** (attn_layer=0) and
**80% for dpa2** (layer=12).

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

- **New Features**
- Introduced a new `inference` parameter to key model functions,
enhancing flexibility for inference scenarios during model execution.
- Added functionality to output a mapping array to a CSV file, improving
data handling capabilities.

- **Bug Fixes**
- Improved the behavior of the model during inference versus training,
potentially impacting downstream processing based on the output.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

deepmd/pt/entrypoints/main.py

@@ -335,7 +335,9 @@ def train(FLAGS):
 
 
 def freeze(FLAGS):
-    model = torch.jit.script(inference.Tester(FLAGS.model, head=FLAGS.head).model)
+    model = inference.Tester(FLAGS.model, head=FLAGS.head).model
+    model.eval()
+    model = torch.jit.script(model)
     extra_files = {}
     torch.jit.save(
         model,

deepmd/pt/model/model/make_model.py

@@ -267,6 +267,7 @@ def forward_common_lower(
                 self.atomic_output_def(),
                 cc_ext,
                 do_atomic_virial=do_atomic_virial,
+                create_graph=self.training,
             )
             model_predict = self.output_type_cast(model_predict, input_prec)
             return model_predict

deepmd/pt/model/model/transform_output.py

@@ -33,15 +33,27 @@ def atomic_virial_corr(
     faked_grad = torch.ones_like(sumce0)
     lst = torch.jit.annotate(List[Optional[torch.Tensor]], [faked_grad])
     extended_virial_corr0 = torch.autograd.grad(
-        [sumce0], [extended_coord], grad_outputs=lst, create_graph=True
+        [sumce0],
+        [extended_coord],
+        grad_outputs=lst,
+        create_graph=False,
+        retain_graph=True,
     )[0]
     assert extended_virial_corr0 is not None
     extended_virial_corr1 = torch.autograd.grad(
-        [sumce1], [extended_coord], grad_outputs=lst, create_graph=True
+        [sumce1],
+        [extended_coord],
+        grad_outputs=lst,
+        create_graph=False,
+        retain_graph=True,
     )[0]
     assert extended_virial_corr1 is not None
     extended_virial_corr2 = torch.autograd.grad(
-        [sumce2], [extended_coord], grad_outputs=lst, create_graph=True
+        [sumce2],
+        [extended_coord],
+        grad_outputs=lst,
+        create_graph=False,
+        retain_graph=True,
     )[0]
     assert extended_virial_corr2 is not None
     extended_virial_corr = torch.concat(
@@ -61,11 +73,16 @@ def task_deriv_one(
     extended_coord: torch.Tensor,
     do_virial: bool = True,
     do_atomic_virial: bool = False,
+    create_graph: bool = True,
 ):
     faked_grad = torch.ones_like(energy)
     lst = torch.jit.annotate(List[Optional[torch.Tensor]], [faked_grad])
     extended_force = torch.autograd.grad(
-        [energy], [extended_coord], grad_outputs=lst, create_graph=True
+        [energy],
+        [extended_coord],
+        grad_outputs=lst,
+        create_graph=create_graph,
+        retain_graph=True,
     )[0]
     assert extended_force is not None
     extended_force = -extended_force
@@ -106,6 +123,7 @@ def take_deriv(
     coord_ext: torch.Tensor,
     do_virial: bool = False,
     do_atomic_virial: bool = False,
+    create_graph: bool = True,
 ):
     size = 1
     for ii in vdef.shape:
@@ -123,6 +141,7 @@ def take_deriv(
             coord_ext,
             do_virial=do_virial,
             do_atomic_virial=do_atomic_virial,
+            create_graph=create_graph,
         )
         # nf x nloc x 1 x 3, nf x nloc x 1 x 9
         ffi = ffi.unsqueeze(-2)
@@ -146,6 +165,7 @@ def fit_output_to_model_output(
     fit_output_def: FittingOutputDef,
     coord_ext: torch.Tensor,
     do_atomic_virial: bool = False,
+    create_graph: bool = True,
 ) -> Dict[str, torch.Tensor]:
     """Transform the output of the fitting network to
     the model output.
@@ -169,6 +189,7 @@ def fit_output_to_model_output(
                     coord_ext,
                     do_virial=vdef.c_differentiable,
                     do_atomic_virial=do_atomic_virial,
+                    create_graph=create_graph,
                 )
                 model_ret[kk_derv_r] = dr
                 if vdef.c_differentiable: