[Bugfix] VLLM_V1 supports passing other compilation levels (#19340)

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

tests/compile/test_config.py

@@ -26,6 +26,8 @@ def test_use_cudagraphs_dynamic(monkeypatch):
     assert not vllm_config.compilation_config.use_cudagraph
 
 
+# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
+@pytest.mark.forked
 # 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.
@@ -33,8 +35,8 @@ def test_use_cudagraphs_dynamic(monkeypatch):
 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.
+    # Disable multiprocessing so that the counter is in the same process
-    monkeypatch.setenv('VLLM_WORKER_MULTIPROC_METHOD', "spawn")
+    monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
     monkeypatch.setenv('VLLM_DISABLE_COMPILE_CACHE', val)
 
     compilation_config = {
@@ -50,6 +52,8 @@ def test_VLLM_DISABLE_COMPILE_CACHE(vllm_runner, monkeypatch, val):
         pass
 
 
+# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
+@pytest.mark.forked
 @pytest.mark.parametrize("enabled", [True, False])
 def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
     assert vllm.envs.VLLM_USE_V1
@@ -72,3 +76,50 @@ def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
                         compilation_config=compilation_config,
                         gpu_memory_utilization=0.4) as _):
         pass
+
+
+# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
+@pytest.mark.forked
+def test_dynamo_as_is(vllm_runner, monkeypatch):
+    # Disable multiprocessing so that the counter is in the same process
+    monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
+
+    with (
+            compilation_counter.expect(dynamo_as_is_count=1),
+            # loading the model causes compilation (if enabled) to happen
+            vllm_runner('facebook/opt-125m',
+                        compilation_config={"level": 1},
+                        gpu_memory_utilization=0.4) as _):
+        pass
+
+
+# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
+@pytest.mark.forked
+def test_no_compilation(vllm_runner, monkeypatch):
+    # Disable multiprocessing so that the counter is in the same process
+    monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
+
+    with (
+            compilation_counter.expect(num_graphs_seen=0,
+                                       dynamo_as_is_count=0),
+            # loading the model causes compilation (if enabled) to happen
+            vllm_runner('facebook/opt-125m',
+                        compilation_config={"level": 0},
+                        gpu_memory_utilization=0.4) as _):
+        pass
+
+
+# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
+@pytest.mark.forked
+def test_enforce_eager(vllm_runner, monkeypatch):
+    # Disable multiprocessing so that the counter is in the same process
+    monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
+
+    with (
+            compilation_counter.expect(num_graphs_seen=0,
+                                       dynamo_as_is_count=0),
+            # loading the model causes compilation (if enabled) to happen
+            vllm_runner('facebook/opt-125m',
+                        enforce_eager=True,
+                        gpu_memory_utilization=0.4) as _):
+        pass

vllm/compilation/counter.py

@@ -27,6 +27,8 @@ class CompilationCounter:
     num_cache_entries_updated: int = 0
     # The number of standalone_compile compiled artifacts saved
     num_compiled_artifacts_saved: int = 0
+    # Number of times a model was loaded with CompilationLevel.DYNAMO_AS_IS
+    dynamo_as_is_count: int = 0
 
     def clone(self) -> "CompilationCounter":
         return copy.deepcopy(self)

vllm/config.py

@@ -4106,9 +4106,11 @@ class CompilationConfig:
         certain small batchsizes, where inductor is good at optimizing.
     """
     # Top-level Compilation control
-    level: int = 0
+    level: Optional[int] = None
     """The level of compilation:
 
+    - None: If None, we will select the default compilation level.
+      For V1 engine this is 3, for V0 engine this is 0.
     - 0: no compilation.
     - 1: dynamo as is.
     - 2: dynamo once.
@@ -4664,6 +4666,22 @@ class VllmConfig:
                 "To workaround this limitation, vLLM will set 'ieee' input "
                 "precision for chunked prefill triton kernels.")
 
+        # If the user does not explicitly set a compilation level, then
+        # we use the default level. The default level depends on other
+        # settings (see the below code).
+        if self.compilation_config.level is None:
+            if envs.VLLM_USE_V1:
+                if (self.model_config is not None
+                        and not self.model_config.enforce_eager):
+                    self.compilation_config.level = CompilationLevel.PIECEWISE
+                else:
+                    self.compilation_config.level = \
+                            CompilationLevel.NO_COMPILATION
+            else:
+                # NB: Passing both --enforce-eager and a compilation level
+                # in V0 means the compilation level wins out.
+                self.compilation_config.level = CompilationLevel.NO_COMPILATION
+
         # async tp is built on top of sequence parallelism
         # and requires it to be enabled.
         if self.compilation_config.pass_config.enable_async_tp:
@@ -4676,7 +4694,6 @@ class VllmConfig:
             # By default, V1 uses piecewise CUDA graphs. If full_cuda_graph
             # is set to True, full CUDA graphs will be used.
             self.compilation_config.cudagraph_num_of_warmups = 1
-            self.compilation_config.level = CompilationLevel.PIECEWISE
             self.compilation_config.set_splitting_ops_for_v1()
 
         self._set_cudagraph_sizes()

vllm/v1/worker/gpu_model_runner.py

@@ -43,7 +43,7 @@ from vllm.sequence import IntermediateTensors, PoolerOutput
 from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
                         GiB_bytes, LazyLoader, check_use_alibi, get_dtype_size,
-                        is_pin_memory_available, round_up)
+                        is_pin_memory_available, round_up, supports_dynamo)
 from vllm.v1.attention.backends.mamba_selectors import get_mamba_attn_backend
 from vllm.v1.attention.backends.utils import (
     AttentionMetadataBuilder, CommonAttentionMetadata,
@@ -1930,6 +1930,17 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                 rank_mapping,
             )
 
+        if (
+            self.vllm_config.compilation_config.level == \
+                CompilationLevel.DYNAMO_AS_IS and supports_dynamo()
+        ):
+            backend = self.vllm_config.compilation_config.init_backend(
+                self.vllm_config)
+            compilation_counter.dynamo_as_is_count += 1
+            self.model.compile(
+                fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
+                backend=backend)
+
     def reload_weights(self) -> None:
         assert getattr(self, "model", None) is not None, \
             "Cannot reload weights before model is loaded."

vllm/worker/model_runner.py

@@ -22,6 +22,7 @@ import vllm.envs as envs
 from vllm.attention import AttentionMetadata, get_attn_backend
 from vllm.attention.backends.abstract import AttentionState
 from vllm.attention.backends.utils import CommonAttentionState
+from vllm.compilation.counter import compilation_counter
 from vllm.config import CompilationLevel, VllmConfig
 from vllm.core.scheduler import SchedulerOutputs
 from vllm.distributed import broadcast_tensor_dict, get_pp_group
@@ -1121,6 +1122,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
             CompilationLevel.DYNAMO_AS_IS and supports_dynamo():
             backend = self.vllm_config.compilation_config.init_backend(
                 self.vllm_config)
+            compilation_counter.dynamo_as_is_count += 1
             self.model = torch.compile(
                 self.model,
                 fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,