[BugFix] VLLM_DISABLE_COMPILE_CACHE=1 should disable all reads and writes from the cache (#20942)

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

tests/compile/test_config.py

@@ -26,6 +26,30 @@ def test_use_cudagraphs_dynamic(monkeypatch):
     assert not vllm_config.compilation_config.use_cudagraph
 
 
+# NB: We don't test VLLM_DISABLE_COMPILE_CACHE=0 because that depends
+# on the state of the cache directory on the current machine, which
+# may be influenced by other tests.
+@pytest.mark.parametrize("val", ["1"])
+def test_VLLM_DISABLE_COMPILE_CACHE(vllm_runner, monkeypatch, val):
+    assert vllm.envs.VLLM_USE_V1
+
+    # spawn means that the counters are in the same process.
+    monkeypatch.setenv('VLLM_WORKER_MULTIPROC_METHOD', "spawn")
+    monkeypatch.setenv('VLLM_DISABLE_COMPILE_CACHE', val)
+
+    compilation_config = {
+        "use_cudagraph": False,  # speed things up a bit
+    }
+    with (
+            compilation_counter.expect(num_cache_entries_updated=0,
+                                       num_compiled_artifacts_saved=0),
+            # loading the model causes compilation (if enabled) to happen
+            vllm_runner('facebook/opt-125m',
+                        compilation_config=compilation_config,
+                        gpu_memory_utilization=0.4) as _):
+        pass
+
+
 @pytest.mark.parametrize("enabled", [True, False])
 def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
     assert vllm.envs.VLLM_USE_V1

vllm/compilation/backends.py

@@ -183,9 +183,10 @@ class CompilerManager:
         assert compiled_graph is not None, "Failed to compile the graph"
 
         # store the artifact in the cache
-        if handle is not None:
+        if not envs.VLLM_DISABLE_COMPILE_CACHE and handle is not None:
             self.cache[(runtime_shape, graph_index,
                         self.compiler.name)] = handle
+            compilation_counter.num_cache_entries_updated += 1
             self.is_cache_updated = True
             if graph_index == 0:
                 # adds some info logging for the first graph

vllm/compilation/compiler_interface.py

@@ -213,7 +213,9 @@ 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)
-        compiled_graph.save(path=path, format="unpacked")
+        if not envs.VLLM_DISABLE_COMPILE_CACHE:
+            compiled_graph.save(path=path, format="unpacked")
+            compilation_counter.num_compiled_artifacts_saved += 1
         return compiled_graph, (key, path)
 
     def load(self,

vllm/compilation/counter.py

@@ -23,6 +23,10 @@ class CompilationCounter:
     num_inductor_compiles: int = 0
     # EagerAdapter.compile calls
     num_eager_compiles: int = 0
+    # The number of time vLLM's compiler cache entry was updated
+    num_cache_entries_updated: int = 0
+    # The number of standalone_compile compiled artifacts saved
+    num_compiled_artifacts_saved: int = 0
 
     def clone(self) -> "CompilationCounter":
         return copy.deepcopy(self)