[Bugfix] vLLM should check Inductor config for compile cache enablement status (#27637)

Signed-off-by: Yanan Cao <gmagogsfm@gmail.com>

vllm/compilation/compiler_interface.py

@@ -163,6 +163,23 @@ def get_inductor_factors() -> list[Any]:
     return factors
 
 
+def is_compile_cache_enabled(
+    vllm_additional_inductor_config: dict[str, Any],
+) -> bool:
+    vllm_inductor_config_disable_cache = vllm_additional_inductor_config.get(
+        "force_disable_caches", False
+    )
+
+    # TODO(gmagogsfm): Replace torch._inductor.config.force_disable_caches
+    # with torch.compiler.config.force_disable_caches when minimum PyTorch
+    # version reaches 2.10
+    return (
+        not envs.VLLM_DISABLE_COMPILE_CACHE
+        and not torch._inductor.config.force_disable_caches
+        and not vllm_inductor_config_disable_cache
+    )
+
+
 class InductorStandaloneAdaptor(CompilerInterface):
     """
     The adaptor for the Inductor compiler.
@@ -222,7 +239,8 @@ class InductorStandaloneAdaptor(CompilerInterface):
         # Save the compiled artifact to disk in the specified path
         assert key is not None
         path = os.path.join(self.cache_dir, key)
-        if not envs.VLLM_DISABLE_COMPILE_CACHE:
+
+        if is_compile_cache_enabled(compiler_config):
             compiled_graph.save(path=path, format=self.save_format)
             compilation_counter.num_compiled_artifacts_saved += 1
         return compiled_graph, (key, path)
@@ -472,10 +490,8 @@ class InductorAdaptor(CompilerInterface):
                 config_patches=current_config,
             )
 
-        # We treat VLLM_DISABLE_COMPILE_CACHE as the overall switch for torch
-        # compilation cache. So turn off the checks if we disable the
-        # compilation cache.
-        if not envs.VLLM_DISABLE_COMPILE_CACHE:
+        # Turn off the checks if we disable the compilation cache.
+        if is_compile_cache_enabled(compiler_config):
             if hash_str is None:
                 raise RuntimeError(
                     "vLLM failed to compile the model. The most "