[torch.compile] Document the workaround to standalone_compile failing (#33571)

Signed-off-by: Richard Zou <zou3519@gmail.com>

docs/design/debug_vllm_compile.md

@@ -282,6 +282,15 @@ If vLLM's compile cache is wrong, this usually means that a factor is missing.
 Please see [this example](https://github.com/vllm-project/vllm/blob/18b39828d90413d05d770dfd2e2f48304f4ca0eb/vllm/config/model.py#L310)
 of how vLLM computes part of the cache key.
 
+vLLM's compilation cache requires that the code being compiled ends up being serializable.
+If this is not the case, then it will error out on save. Usually the fixes are to either:
+
+- rewrite the non-serializable pieces (perhaps difficult because it's difficult to
+  tell right now what is serializable and what isn't)
+- file a bug report
+- ignore the error by setting `VLLM_DISABLE_COMPILE_CACHE=1` (note that this will
+  make warm server starts a lot slower).
+
 ## Debugging CUDAGraphs
 
 CUDAGraphs is a feature that allows one to:

vllm/compilation/compiler_interface.py

@@ -273,7 +273,26 @@ class InductorStandaloneAdaptor(CompilerInterface):
         assert key is not None
         path = os.path.join(self.cache_dir, key)
 
+        def is_saveable_2_10(compiled_artifact):
+            # can just use compiled_artifact.is_saveable in 2.11
+            if compiled_artifact._artifacts is None:
+                return False
+            _, cache_info = compiled_artifact._artifacts
+            return len(cache_info.aot_autograd_artifacts) == 1
+
         if is_compile_cache_enabled(compiler_config):
+            if not is_saveable_2_10(compiled_graph):
+                raise RuntimeError(
+                    "The compiled artifact is not serializable. This usually means "
+                    "that the model code has something that is not serializable "
+                    "by torch.compile in it. You can fix this by either "
+                    "figuring out what is not serializable and rewriting it, "
+                    "filing a bug report, "
+                    "or suppressing this error by "
+                    "disabling vLLM's compilation cache via "
+                    "VLLM_DISABLE_COMPILE_CACHE=1 "
+                    "(this will greatly increase vLLM server warm start times)."
+                )
             compiled_graph.save(path=path, format=self.save_format)
             compilation_counter.num_compiled_artifacts_saved += 1
         return compiled_graph, (key, path)