Fix DeepCompile for PyTorch v2.8 (#7496)

This PR updates the kernel generation function arguments in Inductor to
ensure DeepCompile is compatible with PyTorch v2.8.
It also fixes the logging output of DeepCompile.

Author: tohtana

deepspeed/compile/backend.py

@@ -311,7 +311,7 @@ def make_bw_graph(gm, sample_inputs):
             graph_index = get_index_by_graph_id(graph_order, graph_id)
             log_rank0(
                 f"Bwd start {graph_index} graph_id={graph_id} alloc_mem={get_accelerator().memory_allocated()} graph={gm.graph}",
-                enable=True)
+                enable=debug_log)
 
             bwd_inputs_stack = get_backward_inputs()
 

deepspeed/compile/inductor.py

@@ -17,6 +17,7 @@
 except ImportError:
     pass
 
+from deepspeed.utils.torch import required_torch_version
 from .util import get_input_nodes
 from .graph_param import DSGraphParamManager
 
@@ -172,7 +173,11 @@ def codegen(self, wrapper):
                     self.codegen_comment(wrapper)
                     args = [*self.codegen_args(), *self.codegen_kwargs()]
 
-                    V.graph.wrapper_code.generate_fallback_kernel(self, args)
+                    if required_torch_version(min_version=2.8):
+                        V.graph.wrapper_code.generate_fallback_kernel(self)
+                    else:
+                        V.graph.wrapper_code.generate_fallback_kernel(self, args)
+
                     if isinstance(self.layout, Layout):
                         self.codegen_size_asserts(wrapper)