[v1][torch.compile] support managing cudagraph buffer (#10203)

Signed-off-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

tests/compile/piecewise/piecewise_compilation_config.json

@@ -1,4 +1,5 @@
 {
     "use_cudagraph": true,
-    "non_cudagraph_ops": ["silly.attention"]
+    "non_cudagraph_ops": ["silly.attention"],
+    "cudagraph_copy_inputs": true
 }

tests/compile/piecewise/test_simple.py

@@ -80,7 +80,7 @@ def test_simple_piecewise_compile():
     config = os.path.join(directory, "piecewise_compilation_config.json")
     os.environ["VLLM_TORCH_COMPILE_CONFIG"] = config
 
-    input_buffer = torch.randn(100).cuda()
+    inputs = torch.randn(100).cuda()
 
     with compilation_counter.expect(
             num_graphs_seen=1,  # one graph for the model
@@ -92,15 +92,15 @@ def test_simple_piecewise_compile():
     ):
 
         with set_compile_context([1, 2]):
-            model(input_buffer)
+            model(inputs)
 
-            model(input_buffer[:2])
+            model(torch.randn(2).cuda())
-            model(input_buffer[:1])
+            model(torch.randn(1).cuda())
 
-        input_buffer[:2].zero_()
+        input = torch.zeros(2).cuda()
         global global_counter
         global_counter = 0
-        output = model(input_buffer[:2])
+        output = model(input)
         assert global_counter == 2
         assert torch.allclose(output.cpu(), torch.tensor([3., 1.]))
 

vllm/compilation/backends.py

@@ -389,6 +389,8 @@ class VllmBackend:
     returned_callable: Callable
     # Inductor passes to run on the graph pre-defunctionalization
     post_grad_passes: Sequence[Callable]
+    sym_tensor_indices: List[int]
+    input_buffers: List[torch.Tensor]
 
     def __init__(self, post_grad_passes: Sequence[Callable] = ()):
         global global_graph_pool
@@ -401,6 +403,9 @@ class VllmBackend:
         self.graph_pool = global_graph_pool
         self.post_grad_passes = post_grad_passes
 
+        self.sym_tensor_indices = []
+        self.input_buffers = []
+
         # `torch.compile` is JIT compiled, so we don't need to
         # do anything here
 
@@ -461,7 +466,46 @@ class VllmBackend:
 
         self._called = True
 
-        return self.split_gm
+        if not self.compilation_configs.use_cudagraph or \
+            not self.compilation_configs.cudagraph_copy_inputs:
+            return self.split_gm
+
+        # if we need to copy input buffers for cudagraph
+        from torch._guards import detect_fake_mode
+        fake_mode = detect_fake_mode()
+        fake_args = [
+            fake_mode.from_tensor(t) if isinstance(t, torch.Tensor) else t
+            for t in example_inputs
+        ]
+
+        # index of tensors that have symbolic shapes (batch size)
+        self.sym_tensor_indices = [
+            i for i, x in enumerate(fake_args)
+            if isinstance(x, torch._subclasses.fake_tensor.FakeTensor)
+        ]
+
+        # compiler managed cudagraph input buffers
+        # we assume the first run with symbolic shapes
+        # has the maximum size among all the tensors
+        self.input_buffers = [
+            example_inputs[x].clone() for x in self.sym_tensor_indices
+        ]
+
+        def copy_and_call(*args):
+            list_args = list(args)
+            for i, index in enumerate(self.sym_tensor_indices):
+                runtime_tensor = list_args[index]
+                runtime_shape = runtime_tensor.shape[0]
+                static_tensor = self.input_buffers[i][:runtime_shape]
+
+                # copy the tensor to the static buffer
+                static_tensor.copy_(runtime_tensor)
+
+                # replace the tensor in the list_args to the static buffer
+                list_args[index] = static_tensor
+            return self.split_gm(*list_args)
+
+        return copy_and_call
 
 
 @dataclasses.dataclass

vllm/compilation/config.py

@@ -32,6 +32,11 @@ class CompilationConfig(BaseModel):
             It means the first several runs will be treated as warmup runs.
             Only after that, the execution will be recorded, and the recorded
             cudagraph will be used for subsequent runs.
+        - cudagraph_copy_inputs: whether to copy input tensors for
+            cudagraph. If the caller can guarantee that the same input buffers
+            are always used, it can set this to False. Otherwise, it should
+            set this to True, and the compiler will copy the input to an
+            internally managed buffer. Default is False.
     - Inductor compilation:
         - use_inductor: whether to use inductor compilation.
             - False: inductor compilation is not used. graph runs in eager.
@@ -78,6 +83,7 @@ class CompilationConfig(BaseModel):
     non_cudagraph_ops: List[str] = Field(default_factory=list)
     cudagraph_num_of_warmups: int = 0
     cudagraph_capture_sizes: Optional[List[int]] = None
+    cudagraph_copy_inputs: bool = False
 
     dump_graph_stages: List[str] = Field(default_factory=list)
     dump_graph_dir: Path = Field(default=Path("."))