[ez] Delete more torch version checks <= 2.8 (#33288)

Signed-off-by: angelayi <yiangela7@gmail.com>

vllm/compilation/compiler_interface.py

@@ -375,14 +375,13 @@ class InductorAdaptor(CompilerInterface):
         # it to get the hash of the compiled graph directly.
 
         hash_str, file_path = None, None
-        from torch._inductor.codecache import FxGraphCache, compiled_fx_graph_hash
+        from torch._inductor.codecache import compiled_fx_graph_hash
 
-        if torch.__version__.startswith("2.5"):
-            original_load = FxGraphCache.load
-            original_load_name = "torch._inductor.codecache.FxGraphCache.load"
-
-            def hijack_load(*args: Any, **kwargs: Any) -> Any:
-                inductor_compiled_graph = original_load(*args, **kwargs)
+        def hijacked_compile_fx_inner(*args: Any, **kwargs: Any) -> Any:
+            output = torch._inductor.compile_fx.compile_fx_inner(*args, **kwargs)
+            nonlocal hash_str
+            inductor_compiled_graph = output
+            if inductor_compiled_graph is not None:
                 nonlocal file_path
                 compiled_fn = inductor_compiled_graph.current_callable
                 file_path = compiled_fn.__code__.co_filename  # noqa
@@ -395,44 +394,14 @@ class InductorAdaptor(CompilerInterface):
                     for cell in compiled_fn.__closure__:
                         if not callable(cell.cell_contents):
                             continue
-                        if cell.cell_contents.__code__.co_filename.startswith(
-                            self.base_cache_dir
-                        ):
-                            # this is the real file path compiled from Inductor
-                            file_path = cell.cell_contents.__code__.co_filename
+                        code = cell.cell_contents.__code__
+                        if code.co_filename.startswith(self.base_cache_dir):
+                            # this is the real file path
+                            # compiled from Inductor
+                            file_path = code.co_filename
                             break
-                return inductor_compiled_graph
-
-            hijacked_compile_fx_inner = torch._inductor.compile_fx.compile_fx_inner  # noqa
-        elif torch.__version__ >= "2.6":
-            # function renamed in 2.6
-            original_load_name = None
-
-            def hijacked_compile_fx_inner(*args: Any, **kwargs: Any) -> Any:
-                output = torch._inductor.compile_fx.compile_fx_inner(*args, **kwargs)
-                nonlocal hash_str
-                inductor_compiled_graph = output
-                if inductor_compiled_graph is not None:
-                    nonlocal file_path
-                    compiled_fn = inductor_compiled_graph.current_callable
-                    file_path = compiled_fn.__code__.co_filename  # noqa
-                    if (
-                        not file_path.startswith(self.base_cache_dir)
-                        and compiled_fn.__closure__ is not None
-                    ):
-                        # hooked in the align_inputs_from_check_idxs function
-                        # in torch/_inductor/utils.py
-                        for cell in compiled_fn.__closure__:
-                            if not callable(cell.cell_contents):
-                                continue
-                            code = cell.cell_contents.__code__
-                            if code.co_filename.startswith(self.base_cache_dir):
-                                # this is the real file path
-                                # compiled from Inductor
-                                file_path = code.co_filename
-                                break
-                    hash_str = inductor_compiled_graph._fx_graph_cache_key
-                return output
+                hash_str = inductor_compiled_graph._fx_graph_cache_key
+            return output
 
         def hijack_compiled_fx_graph_hash(*args: Any, **kwargs: Any) -> Any:
             out = compiled_fx_graph_hash(*args, **kwargs)
@@ -453,10 +422,6 @@ class InductorAdaptor(CompilerInterface):
             return AlwaysHitShapeEnv()
 
         with ExitStack() as stack:
-            # hijack to get the compiled graph itself
-            if original_load_name is not None:
-                stack.enter_context(patch(original_load_name, hijack_load))
-
             # for hijacking the hash of the compiled graph
             stack.enter_context(
                 patch(
@@ -573,25 +538,16 @@ class InductorAdaptor(CompilerInterface):
             # Dynamo metrics context, see method for more details.
             exit_stack.enter_context(self.metrics_context())
 
-            if torch.__version__.startswith("2.5"):
-                inductor_compiled_graph = FxGraphCache._lookup_graph(
-                    hash_str, example_inputs, True, False
-                )
-                assert inductor_compiled_graph is not None, (
-                    "Inductor cache lookup failed. Please remove "
-                    f"the cache directory and try again."  # noqa
-                )
-            elif torch.__version__ >= "2.6":
-                from torch._inductor.output_code import CompiledFxGraphConstantsWithGm
+            from torch._inductor.output_code import CompiledFxGraphConstantsWithGm
 
-                constants = CompiledFxGraphConstantsWithGm(graph)
-                inductor_compiled_graph, _ = FxGraphCache._lookup_graph(
-                    hash_str, example_inputs, True, None, constants
-                )
-                assert inductor_compiled_graph is not None, (
-                    "Inductor cache lookup failed. Please remove "
-                    f"the cache directory and try again."  # noqa
-                )
+            constants = CompiledFxGraphConstantsWithGm(graph)
+            inductor_compiled_graph, _ = FxGraphCache._lookup_graph(
+                hash_str, example_inputs, True, None, constants
+            )
+            assert inductor_compiled_graph is not None, (
+                "Inductor cache lookup failed. Please remove "
+                f"the cache directory and try again."  # noqa
+            )
 
         # Inductor calling convention (function signature):
         # f(list) -> tuple

vllm/model_executor/layers/quantization/kernels/scaled_mm/pytorch.py

@@ -3,7 +3,6 @@
 
 
 import torch
-from packaging import version
 
 from vllm.config import CompilationMode, get_current_vllm_config
 from vllm.platforms import current_platform
@@ -98,9 +97,6 @@ class RowWiseTorchFP8ScaledMMLinearKernel(TorchFP8ScaledMMLinearKernel):
         if compute_capability is not None and compute_capability < 94:
             return False, "requires compute capability 94 and above."
 
-        if not version.parse(torch.__version__) >= version.parse("2.7"):
-            return False, "requires pytorch version >=2.7."
-
         return True, None
 
     @classmethod