[CUDA] Enable full cudagraph for FlashMLA (#18581)

Signed-off-by: luka <luka@neuralmagic.com>

tests/compile/piecewise/test_full_cudagraph.py

@@ -2,15 +2,16 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import contextlib
 import os
+import weakref
+from contextlib import ExitStack
 
 import pytest
 
+from tests.utils import wait_for_gpu_memory_to_clear
 from vllm import LLM, SamplingParams
 from vllm.config import CompilationConfig
 from vllm.platforms import current_platform
 
-MODEL = "Qwen/Qwen2-1.5B-Instruct"
-
 
 @contextlib.contextmanager
 def temporary_environ(env_vars):
@@ -31,64 +32,119 @@ def temporary_environ(env_vars):
                 os.environ[k] = v
 
 
-@pytest.fixture(scope="module")
-def full_cudagraph_llm():
-    with temporary_environ({
-            "VLLM_USE_V1": "1",
-            "VLLM_FLASH_ATTN_VERSION": "3"
-    }):
-        return LLM(model=MODEL,
-                   gpu_memory_utilization=0.3,
-                   compilation_config=CompilationConfig(full_cuda_graph=True))
-
+@pytest.fixture(scope="class")
+def llm_pair(request):
+    model = request.param
 
-@pytest.fixture(scope="module")
-def piecewise_llm():
     with temporary_environ({
             "VLLM_USE_V1": "1",
             "VLLM_FLASH_ATTN_VERSION": "3"
     }):
-        return LLM(model=MODEL,
-                   gpu_memory_utilization=0.6,
-                   compilation_config=CompilationConfig())
-
-
-def generate_text(llm: LLM, batch_size: int, max_tokens: int):
-    prompts = ["Hi my name is"] * batch_size
-    sampling_params = SamplingParams(temperature=0.0,
-                                     max_tokens=max_tokens,
-                                     top_p=0.95)
-
-    return llm.generate(prompts, sampling_params)
-
-
+        full = LLM(
+            model=model,
+            gpu_memory_utilization=0.45,
+            trust_remote_code=True,
+            max_model_len=1024,
+            compilation_config=CompilationConfig(full_cuda_graph=True),
+        )
+        piecewise = LLM(
+            model=model,
+            gpu_memory_utilization=0.45,
+            trust_remote_code=True,
+            max_model_len=1024,
+            compilation_config=CompilationConfig(),
+        )
+
+    # PyTest caches the fixture values so we use weakref.proxy to enable GC
+    yield weakref.proxy(full), weakref.proxy(piecewise)
+    del full
+    del piecewise
+
+    wait_for_gpu_memory_to_clear(
+        devices=[0],
+        threshold_ratio=0.1,
+    )
+
+
+@pytest.mark.parametrize(
+    "llm_pair",
+    [
+        # Model names for the llm_pair fixture
+        "deepseek-ai/DeepSeek-V2-Lite",
+        "Qwen/Qwen2-1.5B-Instruct"
+    ],
+    indirect=True)
 @pytest.mark.skipif(current_platform.get_device_capability() != (9, 0),
-                    reason="Only Hopper GPUs support FlashAttention 3")
-@pytest.mark.parametrize(("batch_size", "max_tokens"), [(1, 10), (7, 10),
-                                                        (16, 10), (25, 10),
-                                                        (32, 10), (45, 10),
-                                                        (64, 10), (8, 5),
-                                                        (8, 20), (8, 200)])
-def test_full_cudagraph(batch_size, max_tokens, full_cudagraph_llm,
-                        piecewise_llm):
+                    reason="Only Hopper GPUs support FA3 and FlashMLA")
+class TestFullCUDAGraph:
     """
-    Load full cudagraph model and piecewise model once, and at the same time to
-    reuse them across various test cases.
+    Use a class such that an llm pair is constructed once for all
+    batch_size/max_tokens combinations and released immediately after.
 
-    Test various batch sizes and max_tokens to ensure that the full cudagraph
-    compilation works for padded cases too.
+    Module-scope fixtures would stick around the whole time,
+    meaning there would be multiple LLM instances hogging memory simultaneously.
     """
-    piecewise_responses = generate_text(piecewise_llm,
-                                        batch_size=batch_size,
-                                        max_tokens=max_tokens)
-    full_cudagraph_responses = generate_text(full_cudagraph_llm,
-                                             batch_size=batch_size,
-                                             max_tokens=max_tokens)
 
-    # Check that all responses are the same
-    for i in range(len(piecewise_responses)):
-        assert piecewise_responses[i].outputs[
-            0].text == full_cudagraph_responses[i].outputs[0].text
+    @pytest.mark.parametrize(("batch_size", "max_tokens"), [
+        (1, 10),
+        (7, 10),
+        (16, 10),
+        (25, 10),
+        (32, 10),
+        (45, 10),
+        (64, 10),
+        (123, 10),
+        (8, 5),
+        (8, 30),
+    ])
+    def test_full_cudagraph(self, batch_size, max_tokens,
+                            llm_pair: tuple[LLM, LLM]):
+        """
+        Test various batch sizes and max_tokens to ensure that the
+        full cudagraph compilation works for padded cases too.
+        """
+
+        piecewise_llm, full_cudagraph_llm = llm_pair
+
+        prompts = ["Hello, my name is"] * batch_size
+        sampling_params = SamplingParams(temperature=0.0,
+                                         max_tokens=max_tokens,
+                                         top_p=0.95)
+
+        piecewise_responses = piecewise_llm.generate(prompts, sampling_params)
+        full_responses = full_cudagraph_llm.generate(prompts, sampling_params)
+
+        # Check that all responses are the same
+        for piecewise_res, full_res in zip(piecewise_responses,
+                                           full_responses):
+            assert piecewise_res.outputs[0].text == full_res.outputs[0].text
+
+
+@pytest.mark.parametrize(
+    "model, supported",
+    [
+        ("Qwen/Qwen2-1.5B-Instruct", True),
+        # MLA does not support capturing CUDA Graphs with size > max_num_seqs
+        ("deepseek-ai/DeepSeek-V2-Lite", False),
+    ])
+@pytest.mark.skipif(current_platform.get_device_capability() != (9, 0),
+                    reason="Only Hopper GPUs support FA3 and FlashMLA")
+def test_lower_max_num_seqs(model, supported):
+    with temporary_environ({
+            "VLLM_USE_V1": "1",
+            "VLLM_FLASH_ATTN_VERSION": "3"
+    }), ExitStack() as stack:
+        if not supported:
+            stack.enter_context(pytest.raises(RuntimeError))
+
+        llm = LLM(model=model,
+                  max_num_seqs=256,
+                  trust_remote_code=True,
+                  max_model_len=1024,
+                  compilation_config=CompilationConfig(
+                      full_cuda_graph=True,
+                      cudagraph_capture_sizes=[64, 256, 512]))
+        llm.generate(["Hello, my name is"] * 10)
 
 
 def test_full_cudagraph_with_invalid_backend():
@@ -97,5 +153,5 @@ def test_full_cudagraph_with_invalid_backend():
             "VLLM_FLASH_ATTN_VERSION":
             "2"  #FA2 not supported with full_cuda_graph
     }), pytest.raises(RuntimeError):
-        LLM(model=MODEL,
+        LLM(model="Qwen/Qwen2-1.5B-Instruct",
             compilation_config=CompilationConfig(full_cuda_graph=True))

tests/compile/piecewise/test_simple.py

@@ -4,7 +4,7 @@
 Test the piecewise compilation with a simple model so that we
 can exactly calculate the expected output and side effects.
 """
-
+import pytest
 import torch
 from torch import nn
 from torch.library import Library
@@ -14,6 +14,7 @@ from vllm.compilation.decorators import support_torch_compile
 from vllm.config import (CompilationConfig, CompilationLevel, VllmConfig,
                          set_current_vllm_config)
 from vllm.envs import VLLM_USE_V1
+from vllm.forward_context import set_forward_context
 from vllm.utils import direct_register_custom_op
 
 global_counter = 0
@@ -76,7 +77,8 @@ class SillyModel(nn.Module):
         return x
 
 
-def _test_simple_piecewise_compile(*, use_inductor):
+@pytest.mark.parametrize("use_inductor", [True, False])
+def test_simple_piecewise_compile(use_inductor):
     assert VLLM_USE_V1
 
     vllm_config = VllmConfig(compilation_config=CompilationConfig(
@@ -99,7 +101,7 @@ def _test_simple_piecewise_compile(*, use_inductor):
             num_backend_compilations=3,  # num_piecewise_capturable_graphs_seen
             num_cudagraph_captured=
             6,  # num_cudagraph_sizes * num_piecewise_capturable_graphs_seen
-    ):
+    ), set_forward_context({}, vllm_config=vllm_config):
 
         model(inputs)
 
@@ -112,11 +114,3 @@ def _test_simple_piecewise_compile(*, use_inductor):
         output = model(input)
         assert global_counter == 2
         assert torch.allclose(output.cpu(), torch.tensor([3., 1.]))
-
-
-def test_simple_piecewise_compile_inductor():
-    _test_simple_piecewise_compile(use_inductor=True)
-
-
-def test_simple_piecewise_compile_no_inductor():
-    _test_simple_piecewise_compile(use_inductor=False)

tests/compile/piecewise/test_toy_llama.py

@@ -11,6 +11,7 @@ initialized randomly with a fixed seed.
 from dataclasses import dataclass
 from typing import Any, Optional
 
+import pytest
 import torch
 from torch import nn
 from torch.library import Library
@@ -19,6 +20,7 @@ from vllm.compilation.counter import compilation_counter
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import (CompilationConfig, CompilationLevel, VllmConfig,
                          set_current_vllm_config)
+from vllm.forward_context import set_forward_context
 from vllm.utils import direct_register_custom_op
 
 # create a library to hold the custom op
@@ -285,29 +287,32 @@ def run_model(llama_config,
                            vllm_config=vllm_config,
                            prefix="").eval().cuda()
 
-    B = 16  # max batch size
-    input_ids = torch.randint(0, llama_config.vocab_size, (B, )).cuda()
-    positions = torch.arange(B).cuda()
+    with set_forward_context({}, vllm_config=vllm_config):
+        B = 16  # max batch size
+        input_ids = torch.randint(0, llama_config.vocab_size, (B, )).cuda()
+        positions = torch.arange(B).cuda()
 
-    model(input_ids, positions)
-    model(input_ids[:2], positions[:2])
-    model(input_ids[:1], positions[:1])
+        model(input_ids, positions)
+        model(input_ids[:2], positions[:2])
+        model(input_ids[:1], positions[:1])
 
-    input_ids[:2].zero_()
-    output = model(input_ids[:2], positions[:2])
+        input_ids[:2].zero_()
+        output = model(input_ids[:2], positions[:2])
 
-    output = output.cpu()
+        output = output.cpu()
 
-    if llama_config.tractable_init:
-        expected_output = tractable_computation(input_ids[:2], positions[:2],
-                                                llama_config).cpu()
+        if llama_config.tractable_init:
+            expected_output = tractable_computation(input_ids[:2],
+                                                    positions[:2],
+                                                    llama_config).cpu()
 
-        assert torch.allclose(output, expected_output)
-    else:
-        return output.cpu()
+            assert torch.allclose(output, expected_output)
+        else:
+            return output.cpu()
 
 
-def _test_toy_llama(*, use_inductor):
+@pytest.mark.parametrize("use_inductor", [True, False])
+def test_toy_llama(use_inductor: bool):
     # compare output with and without piecewise compilation
 
     llama_config = LlamaConfig(hidden_size=128,
@@ -379,14 +384,6 @@ def _test_toy_llama(*, use_inductor):
         assert torch.allclose(outputs[0], outputs[i])
 
 
-def test_toy_llama_inductor():
-    _test_toy_llama(use_inductor=True)
-
-
-def test_toy_no_inductor():
-    _test_toy_llama(use_inductor=False)
-
-
 @torch.inference_mode
 def benchmark():
     from triton.testing import do_bench

tests/utils.py

@@ -667,42 +667,54 @@ def get_physical_device_indices(devices):
 
 
 @_nvml()
-def wait_for_gpu_memory_to_clear(devices: list[int],
-                                 threshold_bytes: int,
+def wait_for_gpu_memory_to_clear(*,
+                                 devices: list[int],
+                                 threshold_bytes: Optional[int] = None,
+                                 threshold_ratio: Optional[float] = None,
                                  timeout_s: float = 120) -> None:
+    assert threshold_bytes is not None or threshold_ratio is not None
     # Use nvml instead of pytorch to reduce measurement error from torch cuda
     # context.
     devices = get_physical_device_indices(devices)
     start_time = time.time()
     while True:
         output: dict[int, str] = {}
-        output_raw: dict[int, float] = {}
+        output_raw: dict[int, tuple[float, float]] = {}
         for device in devices:
             if current_platform.is_rocm():
                 dev_handle = amdsmi_get_processor_handles()[device]
                 mem_info = amdsmi_get_gpu_vram_usage(dev_handle)
                 gb_used = mem_info["vram_used"] / 2**10
+                gb_total = mem_info["vram_total"] / 2**10
             else:
                 dev_handle = nvmlDeviceGetHandleByIndex(device)
                 mem_info = nvmlDeviceGetMemoryInfo(dev_handle)
                 gb_used = mem_info.used / 2**30
-            output_raw[device] = gb_used
-            output[device] = f'{gb_used:.02f}'
+                gb_total = mem_info.total / 2**30
+            output_raw[device] = (gb_used, gb_total)
+            output[device] = f'{gb_used:.02f}/{gb_total:.02f}'
 
-        print('gpu memory used (GB): ', end='')
+        print('gpu memory used/total (GiB): ', end='')
         for k, v in output.items():
             print(f'{k}={v}; ', end='')
         print('')
 
+        if threshold_bytes is not None:
+            is_free = lambda used, total: used <= threshold_bytes / 2**30
+            threshold = f"{threshold_bytes/2**30} GiB"
+        else:
+            is_free = lambda used, total: used / total <= threshold_ratio
+            threshold = f"{threshold_ratio:.2f}"
+
         dur_s = time.time() - start_time
-        if all(v <= (threshold_bytes / 2**30) for v in output_raw.values()):
+        if all(is_free(used, total) for used, total in output_raw.values()):
             print(f'Done waiting for free GPU memory on devices {devices=} '
-                  f'({threshold_bytes/2**30=}) {dur_s=:.02f}')
+                  f'({threshold=}) {dur_s=:.02f}')
             break
 
         if dur_s >= timeout_s:
             raise ValueError(f'Memory of devices {devices=} not free after '
-                             f'{dur_s=:.02f} ({threshold_bytes/2**30=})')
+                             f'{dur_s=:.02f} ({threshold=})')
 
         time.sleep(5)
 

vllm/compilation/cuda_piecewise_backend.py

@@ -14,6 +14,7 @@ from vllm.compilation.backends import VllmBackend
 from vllm.compilation.counter import compilation_counter
 from vllm.compilation.monitor import end_monitoring_torch_compile
 from vllm.config import VllmConfig
+from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
 from vllm.utils import weak_ref_tensors
 
@@ -137,7 +138,10 @@ class CUDAPiecewiseBackend:
             if self.is_last_graph and not self.to_be_compiled_sizes:
                 self.check_for_ending_compilation()
 
-        if not entry.use_cudagraph:
+        # Skip CUDA graphs if this entry doesn't use them OR
+        # if we're supposed to skip them globally
+        skip_cuda_graphs = get_forward_context().skip_cuda_graphs
+        if not entry.use_cudagraph or skip_cuda_graphs:
             return entry.runnable(*args)
 
         if entry.cudagraph is None:

vllm/entrypoints/llm.py

@@ -179,7 +179,8 @@ class LLM:
         hf_overrides: Optional[HfOverrides] = None,
         mm_processor_kwargs: Optional[dict[str, Any]] = None,
         override_pooler_config: Optional[PoolerConfig] = None,
-        compilation_config: Optional[Union[int, dict[str, Any]]] = None,
+        compilation_config: Optional[Union[int, dict[str, Any],
+                                           CompilationConfig]] = None,
         **kwargs,
     ) -> None:
         """LLM constructor."""

vllm/forward_context.py

@@ -94,6 +94,7 @@ class ForwardContext:
     virtual_engine: int  # set dynamically for each forward pass
     # set dynamically for each forward pass
     dp_metadata: Optional[DPMetadata] = None
+    skip_cuda_graphs: bool = False
 
 
 _forward_context: Optional[ForwardContext] = None
@@ -108,11 +109,14 @@ def get_forward_context() -> ForwardContext:
 
 
 @contextmanager
-def set_forward_context(attn_metadata: Any,
-                        vllm_config: VllmConfig,
-                        virtual_engine: int = 0,
-                        num_tokens: Optional[int] = None,
-                        num_tokens_across_dp: Optional[torch.Tensor] = None):
+def set_forward_context(
+    attn_metadata: Any,
+    vllm_config: VllmConfig,
+    virtual_engine: int = 0,
+    num_tokens: Optional[int] = None,
+    num_tokens_across_dp: Optional[torch.Tensor] = None,
+    skip_cuda_graphs: bool = False,
+):
     """A context manager that stores the current forward context,
     can be attention metadata, etc.
     Here we can inject common logic for every model forward pass.
@@ -135,7 +139,9 @@ def set_forward_context(attn_metadata: Any,
         static_forward_context,
         virtual_engine=virtual_engine,
         attn_metadata=attn_metadata,
-        dp_metadata=dp_metadata)
+        dp_metadata=dp_metadata,
+        skip_cuda_graphs=skip_cuda_graphs,
+    )
 
     try:
         yield

vllm/v1/attention/backends/cpu_attn.py

@@ -7,7 +7,8 @@ from vllm.attention.backends.torch_sdpa import (TorchSDPABackendImpl,
                                                 TorchSDPAMetadata)
 from vllm.attention.backends.utils import CommonAttentionState
 from vllm.attention.ops.ipex_attn import PagedAttention
-from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+                                              CommonAttentionMetadata)
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.kv_cache_interface import AttentionSpec
 from vllm.v1.worker.block_table import BlockTable
@@ -53,7 +54,7 @@ class TorchSDPABackend:
         return False
 
 
-class TorchSDPAMetadataBuilderV1:
+class TorchSDPAMetadataBuilderV1(AttentionMetadataBuilder[TorchSDPAMetadata]):
 
     def __init__(self, runner: CPUModelRunner, kv_cache_spec: AttentionSpec,
                  block_table: BlockTable) -> None:
@@ -118,9 +119,12 @@ class TorchSDPAMetadataBuilderV1:
 
         return True
 
-    def build(self, num_reqs: int, num_actual_tokens: int, max_query_len: int,
-              common_prefix_len: int,
+    def build(self, common_prefix_len: int,
               common_attn_metadata: CommonAttentionMetadata):
+        num_reqs = common_attn_metadata.num_reqs
+        num_actual_tokens = common_attn_metadata.num_actual_tokens
+        max_query_len = common_attn_metadata.max_query_len
+
         runner = self.runner
         block_table = self.block_table
         seq_lens_np = runner.seq_lens_np[:num_reqs]

vllm/v1/attention/backends/flash_attn.py

@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Attention layer with FlashAttention."""
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Optional
+from typing import TYPE_CHECKING, Any, ClassVar, Optional
 
 import numpy as np
 import torch
@@ -21,13 +21,12 @@ from vllm.distributed.kv_transfer.kv_connector.utils import (
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils import cdiv
-from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+                                              CommonAttentionMetadata)
 from vllm.v1.kv_cache_interface import AttentionSpec
 from vllm.v1.worker.block_table import BlockTable
 
 if TYPE_CHECKING:
-    from vllm.v1.core.sched.output import SchedulerOutput
-    from vllm.v1.worker.gpu_input_batch import InputBatch
     from vllm.v1.worker.gpu_model_runner import GPUModelRunner
 
 if current_platform.is_cuda():
@@ -306,7 +305,9 @@ def _get_sliding_window_configs(
     return sliding_window_configs
 
 
-class FlashAttentionMetadataBuilder:
+class FlashAttentionMetadataBuilder(
+        AttentionMetadataBuilder[FlashAttentionMetadata]):
+    full_cudagraph_supported: ClassVar[bool] = get_flash_attn_version() == 3
 
     def __init__(self, runner: "GPUModelRunner", kv_cache_spec: AttentionSpec,
                  block_table: BlockTable):
@@ -336,13 +337,14 @@ class FlashAttentionMetadataBuilder:
         # populated on first build() call.
         self.aot_sliding_window: Optional[tuple[int, int]] = None
 
-    def reorder_batch(self, input_batch: "InputBatch",
-                      scheduler_output: "SchedulerOutput") -> bool:
-        return False
+    def build(
+        self, common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata
+    ) -> FlashAttentionMetadata:
+        num_reqs = common_attn_metadata.num_reqs
+        num_actual_tokens = common_attn_metadata.num_actual_tokens
+        max_query_len = common_attn_metadata.max_query_len
 
-    def build(self, num_reqs: int, num_actual_tokens: int, max_query_len: int,
-              common_prefix_len: int,
-              common_attn_metadata: CommonAttentionMetadata):
         max_seq_len = int(self.runner.seq_lens_np[:num_reqs].max())
         query_start_loc = common_attn_metadata.query_start_loc
         seq_lens = common_attn_metadata.seq_lens
@@ -496,6 +498,11 @@ class FlashAttentionMetadataBuilder:
         )
         return attn_metadata
 
+    def can_run_in_cudagraph(
+            self, common_attn_metadata: CommonAttentionMetadata) -> bool:
+        # Full CUDA Graph always supported (FA2 support checked separately)
+        return True
+
     def use_cascade_attention(self, *args, **kwargs) -> bool:
         return use_cascade_attention(*args, **kwargs)
 

vllm/v1/attention/backends/flashinfer.py

@@ -18,7 +18,8 @@ from vllm.attention.layer import Attention
 from vllm.config import VllmConfig, get_layers_from_vllm_config
 from vllm.logger import init_logger
 from vllm.v1.attention.backends.flash_attn import use_cascade_attention
-from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+                                              CommonAttentionMetadata)
 from vllm.v1.kv_cache_interface import AttentionSpec
 from vllm.v1.worker.block_table import BlockTable
 
@@ -202,7 +203,7 @@ class FlashInferMetadata:
                 f" received {self.head_dim}.")
 
 
-class FlashInferMetadataBuilder:
+class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
 
     def __init__(self, runner: GPUModelRunner, kv_cache_spec: AttentionSpec,
                  block_table: BlockTable):
@@ -399,9 +400,11 @@ class FlashInferMetadataBuilder:
                     kv_data_type=attn_metadata.data_type,
                 )
 
-    def build(self, num_reqs: int, num_actual_tokens: int, max_query_len: int,
-              common_prefix_len: int,
+    def build(self, common_prefix_len: int,
               common_attn_metadata: CommonAttentionMetadata):
+        num_reqs = common_attn_metadata.num_reqs
+        num_actual_tokens = common_attn_metadata.num_actual_tokens
+
         assert self._num_decodes + self._num_prefills == num_reqs
         assert (self._num_decode_tokens +
                 self._num_prefill_tokens == num_actual_tokens)

vllm/v1/attention/backends/flex_attention.py

@@ -15,7 +15,8 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               is_quantized_kv_cache)
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+                                              CommonAttentionMetadata)
 from vllm.v1.kv_cache_interface import AttentionSpec
 from vllm.v1.worker.block_table import BlockTable
 
@@ -25,8 +26,6 @@ if current_platform.is_cuda():
 logger = init_logger(__name__)
 
 if TYPE_CHECKING:
-    from vllm.v1.core.sched.output import SchedulerOutput
-    from vllm.v1.worker.gpu_input_batch import InputBatch
     from vllm.v1.worker.gpu_model_runner import GPUModelRunner
 
 create_block_mask_compiled = torch.compile(create_block_mask,
@@ -256,7 +255,8 @@ class FlexAttentionMetadata:
         self.block_mask = self.build_block_mask()
 
 
-class FlexAttentionMetadataBuilder:
+class FlexAttentionMetadataBuilder(
+        AttentionMetadataBuilder[FlexAttentionMetadata]):
 
     def __init__(self, runner: "GPUModelRunner", kv_cache_spec: AttentionSpec,
                  block_table: BlockTable):
@@ -272,13 +272,12 @@ class FlexAttentionMetadataBuilder:
         self.kv_cache_spec = kv_cache_spec
         self.block_table = block_table
 
-    def reorder_batch(self, input_batch: "InputBatch",
-                      scheduler_output: "SchedulerOutput") -> bool:
-        return False
-
-    def build(self, num_reqs: int, num_actual_tokens: int, max_query_len: int,
-              common_prefix_len: int,
+    def build(self, common_prefix_len: int,
               common_attn_metadata: CommonAttentionMetadata):
+        num_reqs = common_attn_metadata.num_reqs
+        num_actual_tokens = common_attn_metadata.num_actual_tokens
+        max_query_len = common_attn_metadata.max_query_len
+
         max_seq_len = self.runner.seq_lens_np[:num_reqs].max()
         query_start_loc = common_attn_metadata.query_start_loc
         seq_lens = common_attn_metadata.seq_lens
@@ -332,9 +331,6 @@ class FlexAttentionMetadataBuilder:
         )
         return out
 
-    def use_cascade_attention(self, *args, **kwargs) -> bool:
-        return False
-
 
 class FlexAttentionImpl(AttentionImpl):
     sliding_window: Optional[tuple[int, int]]

vllm/v1/attention/backends/mla/common.py

@@ -207,7 +207,8 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                UnquantizedLinearMethod)
 from vllm.platforms import current_platform
 from vllm.utils import cdiv, round_down
-from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+                                              CommonAttentionMetadata)
 from vllm.v1.kv_cache_interface import AttentionSpec
 from vllm.v1.worker.block_table import BlockTable
 
@@ -329,7 +330,7 @@ class MLACommonMetadata(Generic[D]):
 M = TypeVar("M", bound=MLACommonMetadata)
 
 
-class MLACommonMetadataBuilder(Generic[M]):
+class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
     """
     NOTE: Please read the comment at the top of the file before trying to
     understand this class
@@ -450,9 +451,32 @@ class MLACommonMetadataBuilder(Generic[M]):
             seq_lens=seq_lens,
         )
 
-    def build(self, num_reqs: int, num_actual_tokens: int, max_query_len: int,
-              common_prefix_len: int,
+    def build_for_cudagraph_capture(
+            self, common_attn_metadata: CommonAttentionMetadata) -> M:
+        """
+        This method builds the metadata for full cudagraph capture.
+        Currently, only decode is supported for full cudagraphs with MLA.
+        """
+        m = common_attn_metadata
+        assert m.num_reqs == m.num_actual_tokens, \
+            "MLA only supports decode-only full CUDAGraph capture. " \
+            "Make sure all cudagraph capture sizes <= max_num_seq."
+
+        m.max_query_len = 1  # decode-only
+
+        # Update state usually set in reorder_batch.
+        self._num_decodes = m.num_reqs
+        self._num_decode_tokens = m.num_actual_tokens
+        self._num_prefills = 0
+        self._num_prefill_tokens = 0
+        return self.build(0, m)
+
+    def build(self, common_prefix_len: int,
               common_attn_metadata: CommonAttentionMetadata) -> M:
+        num_reqs = common_attn_metadata.num_reqs
+        num_actual_tokens = common_attn_metadata.num_actual_tokens
+        max_query_len = common_attn_metadata.max_query_len
+
         assert self._num_decodes + self._num_prefills == num_reqs
 
         # Note(simon): be careful about the CPU <> GPU memory movement in this
@@ -461,8 +485,11 @@ class MLACommonMetadataBuilder(Generic[M]):
         device = self.runner.device
         block_table = self.block_table
         block_table_tensor = block_table.get_device_tensor()[:num_reqs]
-        slot_mapping = block_table.slot_mapping_cpu[:num_actual_tokens].to(
-            device, non_blocking=True).long()
+        block_table.slot_mapping[:num_actual_tokens].copy_(
+            block_table.slot_mapping_cpu[:num_actual_tokens],
+            non_blocking=True)
+        block_table.slot_mapping[num_actual_tokens:].fill_(-1)
+        slot_mapping = block_table.slot_mapping[:num_actual_tokens]
 
         query_start_loc = common_attn_metadata.query_start_loc
         seq_lens = common_attn_metadata.seq_lens
@@ -564,8 +591,9 @@ class MLACommonMetadataBuilder(Generic[M]):
             decode=decode_metadata,
         )
 
-    def use_cascade_attention(self, *args, **kwargs) -> bool:
-        return False
+    def can_run_in_cudagraph(
+            self, common_attn_metadata: CommonAttentionMetadata) -> bool:
+        return common_attn_metadata.max_query_len == 1
 
 
 class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):

vllm/v1/attention/backends/mla/flashmla.py

@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from dataclasses import dataclass
-from typing import Any, Optional
+from typing import Any, ClassVar, Optional
 
 import torch
 
@@ -44,7 +44,7 @@ class FlashMLABackend(MLACommonBackend):
 
 @dataclass
 class FlashMLADecodeMetadata(MLACommonDecodeMetadata):
-    tile_scheduler_metadata: tuple[torch.Tensor, torch.Tensor]
+    tile_scheduler_metadata: torch.Tensor
     num_splits: torch.Tensor
 
 
@@ -54,14 +54,18 @@ class FlashMLAMetadata(MLACommonMetadata[FlashMLADecodeMetadata]):
 
 
 class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
+    full_cudagraph_supported: ClassVar[bool] = True  # Decode-only
 
     def __init__(self, runner, kv_cache_spec: AttentionSpec,
                  block_table: BlockTable):
-        super().__init__(runner, kv_cache_spec, block_table)
+        super().__init__(runner, kv_cache_spec, block_table, FlashMLAMetadata)
 
         self.num_q_heads = self.runner.model_config.get_num_attention_heads(
             self.runner.parallel_config)
 
+        self.cg_buf_tile_scheduler_metadata = None
+        self.cg_buf_num_splits = None
+
     def _build_decode(self, block_table_tensor: torch.Tensor,
                       seq_lens: torch.Tensor) -> FlashMLADecodeMetadata:
         tile_scheduler_metadata, num_splits = \
@@ -71,6 +75,30 @@ class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
             1, # MQA for the decode path
         )
 
+        if self.runner.full_cuda_graph:
+            # First time around (CUDAGraph capture), allocate the static buffer
+            if self.cg_buf_tile_scheduler_metadata is None:
+                self.cg_buf_tile_scheduler_metadata = tile_scheduler_metadata
+                self.cg_buf_num_splits = num_splits
+            else:
+                assert self.cg_buf_num_splits is not None
+
+                # Metadata per-SM, fixed size (#SMs, TileMetadataSize)
+                assert (self.cg_buf_tile_scheduler_metadata.size() ==
+                        tile_scheduler_metadata.size())
+                self.cg_buf_tile_scheduler_metadata.\
+                    copy_(tile_scheduler_metadata)
+                tile_scheduler_metadata = self.cg_buf_tile_scheduler_metadata
+
+                # Num splits is per-batch, varying size (batch_size,)
+                n = num_splits.size(0)
+                # make sure static buffer is large enough
+                assert n <= self.cg_buf_num_splits.size(0)
+                num_splits_view = self.cg_buf_num_splits[:n]
+                num_splits_view.copy_(num_splits)
+                self.cg_buf_num_splits[n:].fill_(0)  # fill the rest with 0s
+                num_splits = num_splits_view
+
         return FlashMLADecodeMetadata(
             block_table=block_table_tensor,
             seq_lens=seq_lens,

vllm/v1/attention/backends/mla/rocm_aiter_mla.py

@@ -66,7 +66,7 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
 
     def __init__(self, runner, kv_cache_spec: AttentionSpec,
                  block_table: BlockTable):
-        super().__init__(runner, kv_cache_spec, block_table)
+        super().__init__(runner, kv_cache_spec, block_table, AiterMLAMetadata)
         assert self.kv_cache_spec.block_size == 1, "AITER MLA" \
             "only supports block size 1."
 

vllm/v1/attention/backends/utils.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import abc
+from abc import abstractmethod
 from dataclasses import dataclass
+from typing import TYPE_CHECKING, ClassVar, Generic, TypeVar
 
+import numpy as np
 import torch
 
+if TYPE_CHECKING:
+    from vllm.v1.core.sched.output import SchedulerOutput
+    from vllm.v1.worker.gpu_input_batch import InputBatch
+
 
 @dataclass
 class CommonAttentionMetadata:
     """
-    Attention metadata attributes that can be shared by layers in different KV
-    cache groups and thus having different block table.
+    Per-batch attention metadata, shared across layers and backends.
+    AttentionMetadataBuilder instances use it to construct per-layer metadata.
     """
 
     query_start_loc: torch.Tensor
@@ -18,6 +26,67 @@ class CommonAttentionMetadata:
     """(batch_size,), the length of each request including both computed tokens
     and newly scheduled tokens"""
 
+    num_reqs: int
+    """Number of requests"""
+    num_actual_tokens: int
+    """Total number of tokens in batch"""
+    max_query_len: int
+    """Longest query in batch"""
+
+
+M = TypeVar("M")
+
+
+class AttentionMetadataBuilder(abc.ABC, Generic[M]):
+    # Does this backend/builder support CUDA Graphs for attention.
+    full_cudagraph_supported: ClassVar[bool] = False
+
+    @abstractmethod
+    def build(self, common_prefix_len: int,
+              common_attn_metadata: CommonAttentionMetadata) -> M:
+        """
+        Central method that builds attention metadata.
+        Some builders (MLA) require reorder_batch to be called prior to build.
+        """
+        raise NotImplementedError
+
+    def can_run_in_cudagraph(
+            self, common_attn_metadata: CommonAttentionMetadata) -> bool:
+        """
+        Can this batch (with given metadata) use CUDA Graphs for attention.
+        """
+        return False
+
+    def build_for_cudagraph_capture(
+            self, common_attn_metadata: CommonAttentionMetadata) -> M:
+        """
+        Build attention metadata for CUDA graph capture. Uses build by default.
+        Subclasses that override this method should call self.build or
+        super().build_for_cudagraph_capture.
+        """
+        return self.build(common_prefix_len=0,
+                          common_attn_metadata=common_attn_metadata)
+
+    def use_cascade_attention(
+        self,
+        common_prefix_len: int,
+        query_lens: np.ndarray,
+        num_query_heads: int,
+        num_kv_heads: int,
+        use_alibi: bool,
+        use_sliding_window: bool,
+        num_sms: int,
+    ) -> bool:
+        return False
+
+    def reorder_batch(self, input_batch: "InputBatch",
+                      scheduler_output: "SchedulerOutput") -> bool:
+        """
+        This method can reorder the batch if desired by the backend.
+        :return: Has the batch been reordered (default False).
+        """
+        return False
+
 
 def validate_kv_sharing_target(current_layer_name, target_layer_name,
                                static_forward_context):

vllm/v1/spec_decode/eagle.py

@@ -138,15 +138,17 @@ class EagleProposer:
             max_query_len = query_lens.max().item()
 
             common_attn_metadata = CommonAttentionMetadata(
-                query_start_loc=cu_num_tokens, seq_lens=seq_lens)
+                query_start_loc=cu_num_tokens,
+                seq_lens=seq_lens,
+                num_reqs=batch_size,
+                num_actual_tokens=num_tokens,
+                max_query_len=max_query_len,
+            )
 
             assert self.runner is not None
 
             # FIXME: need to consider multiple kv_cache_groups
             attn_metadata = self.runner.attn_metadata_builder.build(
-                num_reqs=batch_size,
-                num_actual_tokens=num_tokens,
-                max_query_len=max_query_len,
                 common_prefix_len=0,
                 common_attn_metadata=common_attn_metadata,
             )

vllm/v1/worker/gpu_model_runner.py

@@ -16,10 +16,8 @@ from tqdm import tqdm
 
 import vllm.envs as envs
 from vllm.attention import AttentionType, get_attn_backend
-from vllm.attention.backends.abstract import (AttentionBackend,
-                                              AttentionMetadataBuilder)
+from vllm.attention.backends.abstract import AttentionBackend
 from vllm.attention.layer import Attention
-from vllm.attention.utils.fa_utils import get_flash_attn_version
 from vllm.config import (CompilationLevel, VllmConfig,
                          get_layers_from_vllm_config)
 from vllm.distributed.kv_transfer import (get_kv_transfer_group,
@@ -41,7 +39,8 @@ from vllm.sequence import IntermediateTensors
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
                         GiB_bytes, LazyLoader, async_tensor_h2d, cdiv,
                         check_use_alibi, is_pin_memory_available)
-from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+                                              CommonAttentionMetadata)
 from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
 from vllm.v1.kv_cache_interface import (AttentionSpec, FullAttentionSpec,
                                         KVCacheConfig, KVCacheSpec,
@@ -89,6 +88,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         self.vllm_config = vllm_config
         self.model_config = vllm_config.model_config
         self.cache_config = vllm_config.cache_config
+        self.compilation_config = vllm_config.compilation_config
         self.lora_config = vllm_config.lora_config
         self.load_config = vllm_config.load_config
         self.parallel_config = vllm_config.parallel_config
@@ -197,7 +197,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             block_sizes=[self.cache_config.block_size],
         )
 
-        self.use_cuda_graph = (self.vllm_config.compilation_config.level
+        self.use_cuda_graph = (self.compilation_config.level
                                == CompilationLevel.PIECEWISE
                                and not self.model_config.enforce_eager)
         # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
@@ -205,8 +205,9 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         # self.cudagraph_batch_sizes sorts in ascending order.
         # The batch sizes in the config are in descending order.
         self.cudagraph_batch_sizes = list(
-            reversed(
-                self.vllm_config.compilation_config.cudagraph_capture_sizes))
+            reversed(self.compilation_config.cudagraph_capture_sizes))
+
+        self.full_cuda_graph = self.compilation_config.full_cuda_graph
 
         # Cache the device properties.
         self._init_device_properties()
@@ -555,7 +556,15 @@ class GPUModelRunner(LoRAModelRunnerMixin):
     def _prepare_inputs(
         self,
         scheduler_output: "SchedulerOutput",
-    ) -> tuple[dict[str, Any], torch.Tensor, Optional[SpecDecodeMetadata]]:
+    ) -> tuple[dict[str, Any], bool, torch.Tensor,
+               Optional[SpecDecodeMetadata]]:
+        """
+        :return: tuple[
+            attn_metadata: layer-to-attention_metadata mapping,
+            attention_cuda_graphs: whether attention can run in cudagraph
+            logits_indices, spec_decode_metadata
+        ]
+        """
         total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
         assert total_num_scheduled_tokens > 0
         num_reqs = self.input_batch.num_reqs
@@ -669,7 +678,12 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         seq_lens = self.seq_lens[:num_reqs]
 
         common_attn_metadata = CommonAttentionMetadata(
-            query_start_loc=query_start_loc, seq_lens=seq_lens)
+            query_start_loc=query_start_loc,
+            seq_lens=seq_lens,
+            num_reqs=num_reqs,
+            num_actual_tokens=total_num_scheduled_tokens,
+            max_query_len=max_num_scheduled_tokens,
+        )
 
         attn_metadata: dict[str, Any] = {}
         # Prepare the attention metadata for each KV cache group and make layers
@@ -679,25 +693,28 @@ class GPUModelRunner(LoRAModelRunnerMixin):
 
             # Prepare for cascade attention if enabled & beneficial.
             common_prefix_len = 0
+            builder = self.attn_metadata_builders[kv_cache_group_id]
             if self.cascade_attn_enabled:
                 common_prefix_len = self._compute_cascade_attn_prefix_len(
                     num_scheduled_tokens,
                     scheduler_output.
                     num_common_prefix_blocks[kv_cache_group_id],
                     kv_cache_group_spec.kv_cache_spec,
-                    self.attn_metadata_builders[kv_cache_group_id],
+                    builder,
                 )
 
-            attn_metadata_i = (
-                self.attn_metadata_builders[kv_cache_group_id].build(
-                    num_reqs=num_reqs,
-                    num_actual_tokens=total_num_scheduled_tokens,
-                    max_query_len=max_num_scheduled_tokens,
-                    common_prefix_len=common_prefix_len,
-                    common_attn_metadata=common_attn_metadata))
+            attn_metadata_i = (builder.build(
+                common_prefix_len=common_prefix_len,
+                common_attn_metadata=common_attn_metadata,
+            ))
+
             for layer_name in kv_cache_group_spec.layer_names:
                 attn_metadata[layer_name] = attn_metadata_i
 
+        attention_cuda_graphs = all(
+            b.can_run_in_cudagraph(common_attn_metadata)
+            for b in self.attn_metadata_builders)
+
         use_spec_decode = len(
             scheduler_output.scheduled_spec_decode_tokens) > 0
         if not use_spec_decode:
@@ -726,7 +743,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         if self.lora_config:
             self.set_active_loras(self.input_batch, num_scheduled_tokens)
 
-        return attn_metadata, logits_indices, spec_decode_metadata
+        return (attn_metadata, attention_cuda_graphs, logits_indices,
+                spec_decode_metadata)
 
     def _compute_cascade_attn_prefix_len(
         self,
@@ -1121,7 +1139,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         assert self.intermediate_tensors is not None
 
         tp = self.vllm_config.parallel_config.tensor_parallel_size
-        enabled_sp = self.vllm_config.compilation_config.pass_config. \
+        enabled_sp = self.compilation_config.pass_config. \
             enable_sequence_parallelism
         if enabled_sp:
             # When sequence parallelism is enabled, we always pad num_tokens
@@ -1189,8 +1207,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             return self.kv_connector_no_forward(scheduler_output)
 
         # Prepare the decoder inputs.
-        attn_metadata, logits_indices, spec_decode_metadata = (
-            self._prepare_inputs(scheduler_output))
+        (attn_metadata, attention_cuda_graphs, logits_indices,
+         spec_decode_metadata) = (self._prepare_inputs(scheduler_output))
         num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
         if (self.use_cuda_graph
                 and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
@@ -1203,7 +1221,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             # Pad tokens to multiple of tensor_parallel_size when
             # enabled collective fusion for SP
             tp_size = self.vllm_config.parallel_config.tensor_parallel_size
-            if self.vllm_config.compilation_config.pass_config. \
+            if self.compilation_config.pass_config. \
                 enable_sequence_parallelism and tp_size > 1:
                 from vllm.utils import round_up
                 num_input_tokens = round_up(num_scheduled_tokens, tp_size)
@@ -1255,12 +1273,20 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                 num_input_tokens, intermediate_tensors, True)
 
+        # Some attention backends only support CUDA Graphs in pure decode.
+        # If attention doesn't support CUDA Graphs for this batch, but we
+        # compiled with full CUDA graphs, we have to skip them entirely.
+        skip_cuda_graphs = self.full_cuda_graph and not attention_cuda_graphs
+
         # Run the decoder.
         # Use persistent buffers for CUDA graphs.
-        with set_forward_context(attn_metadata,
-                                 self.vllm_config,
-                                 num_tokens=num_input_tokens,
-                                 num_tokens_across_dp=num_tokens_across_dp):
+        with set_forward_context(
+                attn_metadata,
+                self.vllm_config,
+                num_tokens=num_input_tokens,
+                num_tokens_across_dp=num_tokens_across_dp,
+                skip_cuda_graphs=skip_cuda_graphs,
+        ):
             self.maybe_setup_kv_connector(scheduler_output)
 
             model_output = self.model(
@@ -1769,7 +1795,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
     def _dummy_run(
         self,
         num_tokens: int,
-        skip_attn: bool = True,
+        capture_attn_cudagraph: bool = False,
     ) -> torch.Tensor:
 
         # Padding for DP
@@ -1790,9 +1816,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         num_scheduled_tokens = np.array(num_scheduled_tokens_list,
                                         dtype=np.int32)
 
-        if skip_attn:
-            attn_metadata: Optional[dict[str, Any]] = None
-        else:
+        attn_metadata: Optional[dict[str, Any]] = None
+        if capture_attn_cudagraph:
+            attn_metadata = {}
+
             query_start_loc = self.query_start_loc[:num_reqs + 1]
             # Make sure max_model_len is used at the graph capture time.
             self.seq_lens_np[:num_reqs] = self.max_model_len
@@ -1802,19 +1829,19 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             seq_lens = self.seq_lens[:num_reqs]
 
             common_attn_metadata = CommonAttentionMetadata(
-                query_start_loc=query_start_loc, seq_lens=seq_lens)
+                query_start_loc=query_start_loc,
+                seq_lens=seq_lens,
+                num_reqs=num_reqs,
+                num_actual_tokens=num_tokens,
+                max_query_len=num_tokens,
+            )
 
-            attn_metadata = {}
             for kv_cache_group_id, kv_cache_group_spec in enumerate(
                     self.kv_cache_config.kv_cache_groups):
-                attn_metadata_i = (
-                    self.attn_metadata_builders[kv_cache_group_id].build(
-                        num_reqs=num_reqs,
-                        num_actual_tokens=num_tokens,
-                        max_query_len=num_tokens,
-                        common_prefix_len=0,
-                        common_attn_metadata=common_attn_metadata,
-                    ))
+
+                attn_metadata_i = self.attn_metadata_builders[
+                    kv_cache_group_id].build_for_cudagraph_capture(
+                        common_attn_metadata)
                 for layer_name in kv_cache_group_spec.layer_names:
                     attn_metadata[layer_name] = attn_metadata_i
 
@@ -2039,14 +2066,14 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         # Capture the large shapes first so that the smaller shapes
         # can reuse the memory pool allocated for the large shapes.
         with graph_capture(device=self.device):
-            skip_attn = not self.vllm_config.compilation_config.full_cuda_graph
+            full_cg = self.full_cuda_graph
             for num_tokens in tqdm(reversed(self.cudagraph_batch_sizes),
                                    desc="Capturing CUDA graphs",
                                    total=len(self.cudagraph_batch_sizes)):
-                for _ in range(self.vllm_config.compilation_config.
-                               cudagraph_num_of_warmups):
-                    self._dummy_run(num_tokens, skip_attn=skip_attn)
-                self._dummy_run(num_tokens, skip_attn=skip_attn)
+                for _ in range(
+                        self.compilation_config.cudagraph_num_of_warmups):
+                    self._dummy_run(num_tokens, capture_attn_cudagraph=full_cg)
+                self._dummy_run(num_tokens, capture_attn_cudagraph=full_cg)
 
         end_time = time.perf_counter()
         end_free_gpu_memory = torch.cuda.mem_get_info()[0]
@@ -2089,20 +2116,20 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                     "Non-Attention backend is not supported by V1 "
                     "GPUModelRunner.")
 
-            if self.vllm_config.compilation_config.full_cuda_graph:
-                attn_backend_name = attn_backend_i.__name__
-                flash_attn_version = get_flash_attn_version()
-                if attn_backend_name != "FlashAttentionBackend" or \
-                    flash_attn_version != 3:
-                    raise ValueError(
-                        f"full_cuda_graph is only supported with "
-                        f"FA3. Current attention backend is "
-                        f"{attn_backend_name}, FlashAttention version is "
-                        f"{flash_attn_version}.")
-
             block_table_i = self.input_batch.block_table[i]
             attn_metadata_builder_i = attn_backend_i.get_builder_cls()(
-                weakref.proxy(self), kv_cache_spec, block_table_i)
+                weakref.proxy(self),
+                kv_cache_spec,
+                block_table_i,
+            )
+
+            if (self.full_cuda_graph
+                    and not attn_metadata_builder_i.full_cudagraph_supported):
+                raise ValueError(
+                    f"Full CUDAGraph not supported for "
+                    f"{attn_backend_i.__name__}. Turn off CompilationConfig."
+                    f"full_cuda_graph or use a different attention backend.")
+
             self.attn_backends.append(attn_backend_i)
             self.attn_metadata_builders.append(attn_metadata_builder_i)
 
@@ -2142,9 +2169,9 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         to be reshaped to the desired shape before being used by the models.
 
         Args:
-            kv_cache_config: The KV cache config 
+            kv_cache_config: The KV cache config
         Returns:
-            dict[str, torch.Tensor]: A map between layer names to their 
+            dict[str, torch.Tensor]: A map between layer names to their
             corresponding memory buffer for KV cache.
          """
         kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
@@ -2171,11 +2198,11 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         Reshape the KV cache tensors to the desired shape and dtype.
 
         Args:
-            kv_cache_config: The KV cache config 
-            kv_cache_raw_tensors: The KV cache buffer of each layer, with 
+            kv_cache_config: The KV cache config
+            kv_cache_raw_tensors: The KV cache buffer of each layer, with
             correct size but uninitialized shape.
         Returns:
-            Dict[str, torch.Tensor]: A map between layer names to their 
+            Dict[str, torch.Tensor]: A map between layer names to their
             corresponding memory buffer for KV cache.
         """
         kv_caches: dict[str, torch.Tensor] = {}
@@ -2227,7 +2254,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         Args:
             kv_cache_config: The KV cache config
         Returns:
-            Dict[str, torch.Tensor]: A map between layer names to their 
+            Dict[str, torch.Tensor]: A map between layer names to their
             corresponding memory buffer for KV cache.
         """
         # Initialize the memory buffer for KV cache
@@ -2245,10 +2272,9 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 kv_caches,
             )
 
-        bind_kv_cache(
-            kv_caches,
-            self.vllm_config.compilation_config.static_forward_context,
-            self.kv_caches)
+        bind_kv_cache(kv_caches,
+                      self.compilation_config.static_forward_context,
+                      self.kv_caches)
         return kv_caches
 
     def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None: