[torch.compile] integration with compilation control (#9058)

.buildkite/test-pipeline.yaml

@@ -121,7 +121,9 @@ steps:
   - vllm/core/
   - tests/distributed
   - tests/spec_decode/e2e/test_integration_dist_tp4
+  - tests/compile
   commands:
+  - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
 
@@ -231,14 +233,16 @@ steps:
   - vllm/
   - tests/compile
   commands:
-  - pytest -v -s compile/test_full_graph_smoke.py
+  - pytest -v -s compile/test_basic_correctness.py
 
-- label: "PyTorch Fullgraph Test" # 18min
+# TODO: re-write in comparison tests, and fix symbolic shape
-  source_file_dependencies:
+# for quantization ops.
-  - vllm/
+# - label: "PyTorch Fullgraph Test" # 18min
-  - tests/compile
+#   source_file_dependencies:
-  commands:
+#   - vllm/
-  - pytest -v -s compile/test_full_graph.py
+#   - tests/compile
+#   commands:
+#   - pytest -v -s compile/test_full_graph.py
 
 - label: Kernels Test %N # 1h each
   mirror_hardwares: [amd]
@@ -394,7 +398,7 @@ steps:
   - tests/distributed/
   - vllm/compilation
   commands:
-  - pytest -v -s ./compile/test_full_graph_multi_gpu.py
+  - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep -q 'Same node test passed'
   - TARGET_TEST_SUITE=L4 VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1  pytest basic_correctness/ -v -s -m distributed_2_gpus

tests/compile/test_basic_correctness.py

@@ -0,0 +1,48 @@
+from typing import Dict, List, Optional
+
+import pytest
+
+from vllm.compilation.levels import CompilationLevel
+from vllm.utils import cuda_device_count_stateless
+
+from ..utils import compare_all_settings
+
+
+# we cannot afford testing the full Catesian product
+# of all models and all levels
+@pytest.mark.parametrize(
+    "model, model_args, pp_size, tp_size, attn_backend, method, fullgraph",
+    [
+        ("meta-llama/Meta-Llama-3-8B", [], 2, 2, "FLASH_ATTN", "generate",
+         True),
+        ("nm-testing/Meta-Llama-3-8B-Instruct-W8A8-Dyn-Per-Token-2048-Samples",
+         ["--quantization", "compressed-tensors"
+          ], 1, 1, "FLASH_ATTN", "generate", True),
+        ("google/gemma-2-2b-it", [], 1, 2, "FLASHINFER", "generate", True),
+        # TODO: add multi-modality test for llava
+        ("llava-hf/llava-1.5-7b-hf", [], 2, 1, "FLASHINFER", "generate", False)
+    ])
+def test_compile_correctness(model, model_args, pp_size, tp_size, attn_backend,
+                             method, fullgraph):
+    # this test is run under multiple suits, with different GPUs.
+    # make sure we only run the test with correct CUDA devices.
+    # don't use "<", as it will duplicate the tests.
+    if cuda_device_count_stateless() != pp_size * tp_size:
+        pytest.skip("Not correct CUDA devices for the test.")
+    import os
+    os.environ["VLLM_ATTENTION_BACKEND"] = attn_backend
+    if not fullgraph:
+        os.environ["VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE"] = "0"
+    all_args = [["--enforce-eager"] + model_args + ["--max_model_len", "1024"]
+                + ["-pp", str(pp_size)] + ["-tp", str(tp_size)]] * 3
+    # don't test VLLM_TORCH_COMPILE_LEVEL == 3 case
+    # inductor will change the output, so we cannot compare them.
+    all_envs: List[Optional[Dict[str, str]]] = [{
+        "VLLM_TORCH_COMPILE_LEVEL":
+        str(level)
+    } for level in [
+        CompilationLevel.NO_COMPILATION,
+        CompilationLevel.DYNAMO_AS_IS,
+        CompilationLevel.DYNAMO_ONCE,
+    ]]
+    compare_all_settings(model, all_args, all_envs, method=method)

tests/compile/test_full_graph.py

@@ -1,13 +1,20 @@
 import pytest
 
-from vllm.compilation.backends import vllm_backend
+from vllm.compilation.levels import CompilationLevel
 
+from ..utils import fork_new_process_for_each_test
 from .utils import TEST_MODELS, check_full_graph_support
 
 
 @pytest.mark.parametrize("model_info", TEST_MODELS)
-@pytest.mark.parametrize("backend", ["eager", vllm_backend])
+@pytest.mark.parametrize(
-def test_full_graph(model_info, backend):
+    "optimization_level",
+    [CompilationLevel.DYNAMO_ONCE, CompilationLevel.INDUCTOR])
+@fork_new_process_for_each_test
+def test_full_graph(model_info, optimization_level):
     model = model_info[0]
     model_kwargs = model_info[1]
-    check_full_graph_support(model, model_kwargs, backend, tp_size=1)
+    check_full_graph_support(model,
+                             model_kwargs,
+                             optimization_level,
+                             tp_size=1)

tests/compile/test_full_graph_multi_gpu.py

@@ -1,22 +0,0 @@
-import pytest
-
-from vllm.compilation.backends import vllm_backend
-from vllm.utils import cuda_device_count_stateless
-
-from ..utils import fork_new_process_for_each_test
-from .utils import TEST_MODELS_SMOKE, check_full_graph_support
-
-
-@pytest.mark.parametrize("model_info", TEST_MODELS_SMOKE)
-@pytest.mark.parametrize("tp_size", [2])
-@pytest.mark.parametrize("backend", ["eager", vllm_backend])
-@fork_new_process_for_each_test
-def test_full_graph_multi_gpu(model_info, tp_size, backend):
-    model = model_info[0]
-    model_kwargs = model_info[1]
-
-    # Skip the test if there are not enough CUDA devices.
-    if cuda_device_count_stateless() < tp_size:
-        pytest.skip("Not enough CUDA devices for the test.")
-
-    check_full_graph_support(model, model_kwargs, backend, tp_size=tp_size)

tests/compile/test_full_graph_smoke.py

@@ -1,13 +0,0 @@
-import pytest
-
-from vllm.compilation.backends import vllm_backend
-
-from .utils import TEST_MODELS_SMOKE, check_full_graph_support
-
-
-@pytest.mark.parametrize("model_info", TEST_MODELS_SMOKE)
-@pytest.mark.parametrize("backend", ["eager", vllm_backend])
-def test_full_graph(model_info, backend):
-    model = model_info[0]
-    model_kwargs = model_info[1]
-    check_full_graph_support(model, model_kwargs, backend, tp_size=1)

tests/compile/utils.py

@@ -4,16 +4,9 @@ import torch
 
 from tests.quantization.utils import is_quant_method_supported
 from vllm import LLM, SamplingParams
-from vllm.plugins import set_torch_compile_backend
+from vllm.compilation.levels import CompilationLevel
 from vllm.utils import is_hip
 
-TEST_MODELS_SMOKE = [
-    ("nm-testing/Meta-Llama-3-8B-Instruct-W8A8-Dyn-Per-Token-2048-Samples", {
-        "quantization": "compressed-tensors"
-    }),
-    ("meta-llama/Meta-Llama-3-8B", {}),
-]
-
 TEST_MODELS = [
     ("facebook/opt-125m", {}),
     ("nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change", {
@@ -68,20 +61,21 @@ if not is_hip() and is_quant_method_supported("awq"):
     }))
 
 
-def check_full_graph_support(model, model_kwargs, backend, tp_size=1):
+def check_full_graph_support(model,
+                             model_kwargs,
+                             optimization_level,
+                             tp_size=1):
     # make sure these models can be captured in full graph mode
-    if "VLLM_TEST_DYNAMO_GRAPH_CAPTURE" not in os.environ:
+    os.environ["VLLM_TORCH_COMPILE_LEVEL"] = str(optimization_level)
-        os.environ["VLLM_TEST_DYNAMO_GRAPH_CAPTURE"] = "1"
     os.environ["VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE"] = "1"
 
     # Inductor doesn't support fp8/gptq_marlin_24 yet.
     quantization = model_kwargs.get("quantization")
     if (quantization == "fp8" or quantization == "gptq_marlin"
-            or quantization == "gptq_marlin_24") and backend != "eager":
+            or quantization == "gptq_marlin_24"
+        ) and optimization_level >= CompilationLevel.INDUCTOR:
         return
 
-    set_torch_compile_backend(backend)
-
     prompts = [
         "Hello, my name is",
         "The president of the United States is",

tests/tpu/test_compilation.py

@@ -5,9 +5,11 @@ import tempfile
 
 import depyf
 
+from vllm.compilation.levels import CompilationLevel
+
 # disable custom dispatcher, let Dynamo takes over
 # all the control
-os.environ['VLLM_DYNAMO_USE_CUSTOM_DISPATCHER'] = "0"
+os.environ['VLLM_TORCH_COMPILE_LEVEL'] = str(CompilationLevel.DYNAMO_AS_IS)
 
 temp_dir = tempfile.mkdtemp()
 with depyf.prepare_debug(temp_dir):

tests/tpu/test_custom_dispatcher.py

@@ -1,5 +1,7 @@
 import os
 
+from vllm.compilation.levels import CompilationLevel
+
 from ..utils import compare_two_settings
 
 # --enforce-eager on TPU causes graph compilation
@@ -9,8 +11,9 @@ os.environ["VLLM_RPC_TIMEOUT"] = "30000"
 
 
 def test_custom_dispatcher():
-    compare_two_settings("google/gemma-2b",
+    compare_two_settings(
+        "google/gemma-2b",
         arg1=["--enforce-eager"],
         arg2=["--enforce-eager"],
-                         env1={"VLLM_DYNAMO_USE_CUSTOM_DISPATCHER": "0"},
+        env1={"VLLM_TORCH_COMPILE_LEVEL": str(CompilationLevel.DYNAMO_ONCE)},
-                         env2={})
+        env2={"VLLM_TORCH_COMPILE_LEVEL": str(CompilationLevel.DYNAMO_AS_IS)})

vllm/compilation/backends.py

@@ -1,8 +1,17 @@
+import copy
 import operator
+from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import torch
 import torch.fx as fx
 
+from vllm.logger import init_logger
+
+from .compile_context import get_compile_context
+from .levels import CompilationLevel
+
+logger = init_logger(__name__)
+
 
 def fix_functionalization(graph: fx.Graph):
     """
@@ -148,9 +157,113 @@ def fix_functionalization(graph: fx.Graph):
     #     print(graph.python_code(root_module="self", verbose=True).src, file=f)
 
 
-def vllm_backend(graph, example_inputs):
+def wrap_inductor(graph, example_inputs, additional_inductor_config):
     from torch._inductor import config
     current_config = config.shallow_copy_dict()
     from torch._inductor.compile_fx import compile_fx
+
+    if additional_inductor_config is not None:
+        current_config.update(additional_inductor_config)
+    if current_config['post_grad_custom_post_pass'] is not None:
+        logger.warning(
+            "post_grad_custom_post_pass is already set in the config. "
+            "Overwriting it with the fix_functionalization")
     current_config['post_grad_custom_post_pass'] = fix_functionalization
     return compile_fx(graph, example_inputs, config_patches=current_config)
+
+
+def vllm_backend(
+        graph,
+        example_inputs,
+        additional_inductor_config: Optional[Dict] = None) -> Callable:
+
+    context = get_compile_context()
+    context = copy.deepcopy(context) if context is not None else []
+    sizes_to_specialize: List[int] = context
+
+    # flags for all the seen shapes, whether we need to specialize
+    runtime_shapes_to_compile_flags: Dict[Tuple[int, ...], bool] = {}
+
+    # if we need to specialize, the compiled graph for that shape
+    runtime_shapes_to_compiled_graph: Dict[Tuple[int, ...], Callable] = {}
+
+    # this is the first compilation, we will compile a graph with
+    # dynamic shape, as the caller will mark first dimension as dynamic
+    logger.info("Compiling a graph for general shapes")
+    graph_for_symbolic_shape = wrap_inductor(graph, example_inputs,
+                                             additional_inductor_config)
+
+    # TODO: Dynamo does not pass all dynamic shapes.
+    # Need to investigate why. It works now because all the dynamic
+    # shapes have the same value, and either of them can be used.
+    sym_shape_indices = [
+        i for i, x in enumerate(example_inputs) if isinstance(x, torch.SymInt)
+    ]
+
+    first_run = True
+
+    # this is the function we return to Dynamo to run finally
+    def compiled_graph_wrapper(*args):
+
+        runtime_shapes: Tuple[int,
+                              ...] = tuple(args[i] for i in sym_shape_indices)
+
+        nonlocal first_run
+        nonlocal runtime_shapes_to_compile_flags
+        nonlocal runtime_shapes_to_compiled_graph
+
+        if first_run:
+            # the first compilation is for profiling, we directly run it
+            first_run = False
+            return graph_for_symbolic_shape(*args)
+
+        if runtime_shapes not in runtime_shapes_to_compile_flags:
+            # we haven't seen this shape before
+            # query if we need to specialize for this shape
+            # we only specialize for the first dimension.
+            # TODO: investigate if any model needs to specialize
+            # beyond the first dimension
+            runtime_shapes_to_compile_flags[runtime_shapes] = runtime_shapes[
+                0] in sizes_to_specialize
+
+        if not runtime_shapes_to_compile_flags[runtime_shapes]:
+            # we don't need to specialize for this shape
+            return graph_for_symbolic_shape(*args)
+
+        if runtime_shapes not in runtime_shapes_to_compiled_graph:
+            # we need to specialize for this shape, and we haven't compiled
+            # compile the graph for this shape
+            logger.info("Compiling a graph for shapes %s", runtime_shapes)
+            runtime_shapes_to_compiled_graph[runtime_shapes] = wrap_inductor(
+                graph, args, additional_inductor_config)
+
+        return runtime_shapes_to_compiled_graph[runtime_shapes](*args)
+
+    return compiled_graph_wrapper
+
+
+def select_default_backend(level: int) -> Union[str, Callable]:
+    if level in [CompilationLevel.DYNAMO_AS_IS, CompilationLevel.DYNAMO_ONCE]:
+        backend = "eager"
+        return backend
+    assert level in [
+        CompilationLevel.INDUCTOR, CompilationLevel.INDUCTOR_MAX_AUTOTUNE
+    ], f"Invalid level {level}"
+
+    from vllm.compilation.backends import vllm_backend
+    from vllm.plugins import get_inductor_additional_configs
+    additional_configs = get_inductor_additional_configs()
+
+    if level == CompilationLevel.INDUCTOR_MAX_AUTOTUNE:
+        if "max_autotune" in additional_configs and not additional_configs[
+                "max_autotune"]:
+            logger.warning(
+                "max_autotune is disabled, but is overridden by level %s",
+                CompilationLevel.INDUCTOR_MAX_AUTOTUNE)
+        additional_configs['max_autotune'] = True
+
+    from functools import partial
+    backend = partial(vllm_backend,
+                      additional_inductor_config=additional_configs)
+
+    return backend

vllm/compilation/compile_context.py

@@ -0,0 +1,23 @@
+from contextlib import contextmanager
+from typing import Any
+
+_compile_context: Any = None
+
+
+def get_compile_context() -> Any:
+    """Get the current compile context."""
+    return _compile_context
+
+
+@contextmanager
+def set_compile_context(context: Any):
+    """A context manager that stores the current compile context,
+    usually it is a list of sizes to specialize.
+    """
+    global _compile_context
+    prev_context = _compile_context
+    _compile_context = context
+    try:
+        yield
+    finally:
+        _compile_context = prev_context

vllm/compilation/decorators.py

@@ -0,0 +1,85 @@
+from typing import List, Optional, Union
+
+import torch
+
+import vllm.envs as envs
+from vllm.attention import AttentionMetadata
+from vllm.compilation.levels import CompilationLevel
+from vllm.compilation.wrapper import TorchCompileWrapperWithCustomDispatcher
+from vllm.sequence import IntermediateTensors
+from vllm.utils import supports_dynamo
+
+
+def support_compile_llama_style(cls: type):
+    """
+    A decorator to add support for compiling the forward method of a class.
+    If a module's **forward signature** is compatible with llama, this 
+    decorator can be used to enable the compilation of the forward method.
+    """
+
+    # for CompilationLevel.DYNAMO_AS_IS , the upper level model runner
+    # will handle the compilation, so we don't need to do anything here.
+    if envs.VLLM_TORCH_COMPILE_LEVEL in [
+            CompilationLevel.NO_COMPILATION, CompilationLevel.DYNAMO_AS_IS
+    ] or not supports_dynamo():
+        return cls
+
+    # take care of method resolution order
+    # make sure super().__init__ is called on the base class
+    #  other than TorchCompileWrapperWithCustomDispatcher
+    cls.__bases__ = cls.__bases__ + (TorchCompileWrapperWithCustomDispatcher, )
+
+    old_init = cls.__init__
+
+    def __init__(self, *args, **kwargs):
+        old_init(self, *args, **kwargs)
+        TorchCompileWrapperWithCustomDispatcher.__init__(self)
+
+    cls.__init__ = __init__
+
+    def __call__(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        # torch.compiler.is_compiling() means we are inside the compilation
+        # e.g. TPU has the compilation logic in model runner, so we don't
+        # need to compile the model inside.
+        if torch.compiler.is_compiling():
+            return self.forward(input_ids, positions, kv_caches, attn_metadata,
+                                intermediate_tensors, inputs_embeds)
+
+        # the first compilation needs to have dynamic shapes marked
+        if len(self.compiled_codes) < 1:
+            if input_ids is not None:
+                torch._dynamo.mark_dynamic(input_ids, 0)
+            torch._dynamo.mark_dynamic(positions, 0)
+            if inputs_embeds is not None:
+                torch._dynamo.mark_dynamic(inputs_embeds, 0)
+            if intermediate_tensors is not None:
+                for tensors in intermediate_tensors.tensors.values():
+                    torch._dynamo.mark_dynamic(tensors, 0)
+
+        # if we don't use custom dispatcher, we can directly call the
+        # compiled function and let torch.compile handle the dispatching,
+        # with the overhead of guard evaluation and recompilation.
+        if len(self.compiled_codes) < 1 or not self.use_custom_dispatcher:
+            return self.compiled_callable(input_ids, positions, kv_caches,
+                                          attn_metadata, intermediate_tensors,
+                                          inputs_embeds)
+
+        # usually, capturing the model once is enough, and then we can
+        # dispatch to the compiled code directly, without going through
+        # the Dynamo guard mechanism.
+        with self.dispatch_to_code(0):
+            model_output = self.forward(input_ids, positions, kv_caches,
+                                        attn_metadata, intermediate_tensors,
+                                        inputs_embeds)
+            return model_output
+
+    cls.__call__ = __call__
+    return cls

vllm/compilation/levels.py

@@ -0,0 +1,9 @@
+# constants for the levels of the compilation process
+
+
+class CompilationLevel:
+    NO_COMPILATION = 0
+    DYNAMO_AS_IS = 1
+    DYNAMO_ONCE = 2
+    INDUCTOR = 3
+    INDUCTOR_MAX_AUTOTUNE = 4

vllm/compilation/wrapper.py

@@ -3,12 +3,14 @@ import sys
 from abc import abstractmethod
 from contextlib import contextmanager
 from types import CodeType
-from typing import Callable, List
+from typing import Callable, List, Optional
 
 import torch
 
 import vllm.envs as envs
 
+from .levels import CompilationLevel
+
 
 class TorchCompileWrapperWithCustomDispatcher:
     """
@@ -23,7 +25,26 @@ class TorchCompileWrapperWithCustomDispatcher:
         `torch.compile` over the forward method.
     """
 
-    def __init__(self, compiled_callable: Callable):
+    def __init__(self, compiled_callable: Optional[Callable] = None):
+
+        if compiled_callable is None:
+            # default compilation settings
+            # compiling the forward method
+
+            # choose the compile backend
+
+            # if the user has set the backend, use it
+            from vllm.plugins import get_torch_compile_backend
+            backend = get_torch_compile_backend()
+            if backend is None:
+                from vllm.compilation.backends import select_default_backend
+                backend = select_default_backend(envs.VLLM_TORCH_COMPILE_LEVEL)
+
+            compiled_callable = torch.compile(
+                self.forward,
+                fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
+                backend=backend)
+
         self.compiled_callable = compiled_callable
         self.original_code_object = self.__class__.forward.__code__
         self.compiled_codes: List[CodeType] = []
@@ -33,7 +54,7 @@ class TorchCompileWrapperWithCustomDispatcher:
         # subclasses can use this to switch between the custom dispatcher
         # and the default Dynamo guard mechanism.
         self.use_custom_dispatcher: bool = \
-            envs.VLLM_DYNAMO_USE_CUSTOM_DISPATCHER
+            envs.VLLM_TORCH_COMPILE_LEVEL >= CompilationLevel.DYNAMO_ONCE
 
     def __call__(self, *args, **kwargs):
         """Implement the dispatch logic here, beyond the torch.compile level.

vllm/envs.py

@@ -65,6 +65,7 @@ if TYPE_CHECKING:
     VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
     VLLM_SKIP_P2P_CHECK: bool = False
     VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1: bool = False
+    VLLM_TORCH_COMPILE_LEVEL: int = 0
 
 
 def get_default_cache_root():
@@ -198,23 +199,12 @@ environment_variables: Dict[str, Callable[[], Any]] = {
     lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "True").lower() in
              ("true", "1")),
 
-    # Internal flag to enable Dynamo graph capture
-    "VLLM_TEST_DYNAMO_GRAPH_CAPTURE":
-    lambda: int(os.environ.get("VLLM_TEST_DYNAMO_GRAPH_CAPTURE", "0")),
-    "VLLM_DYNAMO_USE_CUSTOM_DISPATCHER":
-    lambda:
-    (os.environ.get("VLLM_DYNAMO_USE_CUSTOM_DISPATCHER", "True").lower() in
-     ("true", "1")),
-
-    # Internal flag to control whether we use custom op,
-    # or use the native pytorch implementation
-    "VLLM_TEST_COMPILE_NO_CUSTOM_OPS":
-    lambda: int(os.environ.get("VLLM_TEST_COMPILE_NO_CUSTOM_OPS", "0")),
-
     # Internal flag to enable Dynamo fullgraph capture
     "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE":
     lambda: bool(
         os.environ.get("VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE", "1") != "0"),
+    "VLLM_TORCH_COMPILE_LEVEL":
+    lambda: int(os.environ.get("VLLM_TORCH_COMPILE_LEVEL", "0")),
 
     # local rank of the process in the distributed setting, used to determine
     # the GPU device id

vllm/model_executor/custom_op.py

@@ -1,6 +1,7 @@
 import torch.nn as nn
 
 import vllm.envs as envs
+from vllm.compilation.levels import CompilationLevel
 from vllm.platforms import current_platform
 from vllm.utils import is_cpu, is_hip, is_xpu
 
@@ -55,7 +56,7 @@ class CustomOp(nn.Module):
         # NOTE(woosuk): Here we assume that vLLM was built for only one
         # specific backend. Currently, we do not support dynamic dispatching.
 
-        if envs.VLLM_TEST_COMPILE_NO_CUSTOM_OPS:
+        if envs.VLLM_TORCH_COMPILE_LEVEL >= CompilationLevel.INDUCTOR:
             return self.forward_native
 
         if is_hip():

vllm/model_executor/models/gemma2.py

@@ -21,6 +21,7 @@ from torch import nn
 from transformers import Gemma2Config
 
 from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_compile_llama_style
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
 from vllm.logger import init_logger
@@ -238,6 +239,7 @@ class Gemma2DecoderLayer(nn.Module):
         return hidden_states, residual
 
 
+@support_compile_llama_style
 class Gemma2Model(nn.Module):
 
     def __init__(

vllm/model_executor/models/llama.py

@@ -28,6 +28,7 @@ from torch import nn
 from transformers import LlamaConfig
 
 from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_compile_llama_style
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size)
@@ -265,6 +266,7 @@ class LlamaDecoderLayer(nn.Module):
         return hidden_states, residual
 
 
+@support_compile_llama_style
 class LlamaModel(nn.Module):
 
     def __init__(

vllm/model_executor/models/llava.py

@@ -365,6 +365,8 @@ class LlavaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
             input_ids = None
             inputs_embeds = None
         else:
+            # always pass the input via `inputs_embeds`
+            # to make sure the computation graph is consistent
             image_input = self._parse_and_validate_image_input(**kwargs)
 
             if image_input is not None:
@@ -375,10 +377,10 @@ class LlavaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
                 inputs_embeds = merge_multimodal_embeddings(
                     input_ids, inputs_embeds, vision_embeddings,
                     self.config.image_token_index)
-
-                input_ids = None
             else:
-                inputs_embeds = None
+                inputs_embeds = self.language_model.model.get_input_embeddings(
+                    input_ids)
+            input_ids = None
 
         hidden_states = self.language_model.model(input_ids,
                                                   positions,

vllm/platforms/tpu.py

@@ -1,7 +1,21 @@
+import os
+
 import torch
 
+import vllm.envs as envs
+from vllm.compilation.levels import CompilationLevel
+from vllm.plugins import set_torch_compile_backend
+
 from .interface import Platform, PlatformEnum
 
+if "VLLM_TORCH_COMPILE_LEVEL" not in os.environ:
+    os.environ["VLLM_TORCH_COMPILE_LEVEL"] = str(CompilationLevel.DYNAMO_ONCE)
+
+assert envs.VLLM_TORCH_COMPILE_LEVEL < CompilationLevel.INDUCTOR,\
+     "TPU does not support Inductor."
+
+set_torch_compile_backend("openxla")
+
 
 class TpuPlatform(Platform):
     _enum = PlatformEnum.TPU

vllm/plugins/__init__.py

@@ -1,5 +1,5 @@
 import logging
-from typing import Callable, Optional, Union
+from typing import Callable, Dict, Optional, Union
 
 import vllm.envs as envs
 
@@ -42,3 +42,15 @@ def set_torch_compile_backend(backend: Union[Callable, str]):
 
 def get_torch_compile_backend() -> Optional[Union[Callable, str]]:
     return _torch_compile_backend
+
+
+_inductor_additional_configs: Dict = {}
+
+
+def set_inductor_additional_configs(configs: Dict):
+    global _inductor_additional_configs
+    _inductor_additional_configs = configs
+
+
+def get_inductor_additional_configs() -> Dict:
+    return _inductor_additional_configs

vllm/sequence.py

@@ -1137,10 +1137,9 @@ class EmbeddingSequenceGroupOutput(
         return self.embeddings == other.embeddings
 
 
-class IntermediateTensors(
+# cannot use msgspec.Struct here because Dynamo does not support it
-        msgspec.Struct,
+@dataclass
-        omit_defaults=True,  # type: ignore[call-arg]
+class IntermediateTensors:
-        array_like=True):  # type: ignore[call-arg]
     """For all pipeline stages except the last, we need to return the hidden
     states and residuals to be sent to the next stage. This data structure
     contains the hidden states and residuals for a request.

vllm/worker/model_runner.py

@@ -18,6 +18,8 @@ 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.compile_context import set_compile_context
+from vllm.compilation.levels import CompilationLevel
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
                          ModelConfig, ObservabilityConfig, ParallelConfig,
                          PromptAdapterConfig, SchedulerConfig)
@@ -1126,10 +1128,10 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                     "provided. Defaulting to scaling factors of 1.0. "
                     "This may lead to less accurate results!")
 
-        if envs.VLLM_TEST_DYNAMO_GRAPH_CAPTURE and supports_dynamo():
+        if envs.VLLM_TORCH_COMPILE_LEVEL == CompilationLevel.DYNAMO_AS_IS \
-            from vllm.compilation.backends import vllm_backend
+            and supports_dynamo():
             from vllm.plugins import get_torch_compile_backend
-            backend = get_torch_compile_backend() or vllm_backend
+            backend = get_torch_compile_backend() or "eager"
             self.model = torch.compile(
                 self.model,
                 fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
@@ -1289,6 +1291,14 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                 batch_size=batch_size,
                 dtype=self.model_config.dtype,
                 device=self.device)
+
+        graph_batch_size = self.max_batchsize_to_capture
+        batch_size_capture_list = [
+            bs for bs in _BATCH_SIZES_TO_CAPTURE if bs <= graph_batch_size
+        ]
+        if self.model_config.enforce_eager:
+            batch_size_capture_list = []
+        with set_compile_context(batch_size_capture_list):
             self.execute_model(model_input, kv_caches, intermediate_tensors)
         torch.cuda.synchronize()
         return