Replace torch.jit.script with torch.compile (#7835) (#7840)

Fixes #7835.

On torch==2.10.0, importing DeepSpeed emitted deprecation warnings from
import-time JIT-decorated helpers.
This change updates the compatibility path to align with PyTorch
guidance while keeping import clean.

---------

Signed-off-by: Masahiro Tanaka <mtanaka@anyscale.com>
Co-authored-by: Hongwei Chen <33092912+hwchen2017@users.noreply.github.com>

Author: tohtana

deepspeed/moe/sharded_moe.py

@@ -18,6 +18,7 @@
 from deepspeed.utils.timer import SynchronizedWallClockTimer
 from deepspeed.utils import logger
 from deepspeed.utils.bwc import bwc_tensor_model_parallel_world_size
+from deepspeed.utils.torch import jit_script_compat
 from typing import Callable, Dict, TYPE_CHECKING, Any, Optional, Tuple, Union
 
 import torch
@@ -157,7 +158,7 @@ def einsum(rule, a, b):
 # includes stateful caching logic which is incompatible with ONNX.
 
 
-@torch.jit.script
+@jit_script_compat
 def _capacity(gates: Tensor, capacity_factor: Tensor, min_capacity: Tensor) -> Tensor:
     # gates has shape of SE
     num_tokens = gates.shape[0]
@@ -170,12 +171,12 @@ def _capacity(gates: Tensor, capacity_factor: Tensor, min_capacity: Tensor) -> T
     return capacity
 
 
-@torch.jit.script
+@jit_script_compat
 def _top_idx(source, k):
     return torch.topk(source, k=k, dim=0)[1]
 
 
-@torch.jit.script
+@jit_script_compat
 def _one_hot_to_float(x, num_classes):
     return F.one_hot(x, num_classes=num_classes).float()
 

deepspeed/runtime/zero/mics_utils.py

@@ -11,20 +11,20 @@
 from typing import List
 
 import numpy as np
-import torch
 from torch import Tensor
 
 from deepspeed import comm as dist
 from deepspeed.accelerator import get_accelerator
 from deepspeed.utils import logger
+from deepspeed.utils.torch import jit_script_compat
 
 
 def _log_rank0(msg):
     if dist.get_rank() == 0:
         logger.info(msg)
 
 
-@torch.jit.script
+@jit_script_compat
 def scale_tensors(tensors: List[Tensor], scale: int):
     for t in tensors:
         t.div_(scale)

deepspeed/sequence/fpdt_layer.py

@@ -10,6 +10,7 @@
 from packaging import version
 import deepspeed.comm as dist
 from deepspeed.accelerator import get_accelerator
+from deepspeed.utils.torch import jit_script_compat
 
 try:
     import flash_attn
@@ -1040,12 +1041,12 @@ def forward(self,
         return output, self.qkv_dense_bias if self.reture_bias else None
 
 
-@torch.jit.script
+@jit_script_compat
 def bias_gelu(x):
     return x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))
 
 
-@torch.jit.script
+@jit_script_compat
 def bias_gelu_back(g, x):
     tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
     # sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243

deepspeed/utils/torch.py

@@ -29,3 +29,9 @@ def register_grad_hook(param, hook):
         param_tmp = param.expand_as(param)
         grad_acc = param_tmp.grad_fn.next_functions[0][0]
         return grad_acc.register_hook(hook)
+
+
+def jit_script_compat(fn):
+    if required_torch_version(min_version=2.0) and hasattr(torch, "compile"):
+        return torch.compile(fn)
+    return torch.jit.script(fn)