[V1][CUDA] Full cudagraph support for FlashInfer (#21367)

vllm/v1/attention/backends/flash_attn.py

@@ -25,7 +25,8 @@ if is_flash_attn_varlen_func_available():
 from vllm.config import VllmConfig, get_layers_from_vllm_config
 from vllm.logger import init_logger
 from vllm.utils import cdiv
-from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+from vllm.v1.attention.backends.utils import (AttentionCGSupport,
+                                              AttentionMetadataBuilder,
                                               CommonAttentionMetadata,
                                               get_kv_cache_layout)
 from vllm.v1.kv_cache_interface import AttentionSpec
@@ -153,7 +154,9 @@ def _get_sliding_window_configs(
 
 class FlashAttentionMetadataBuilder(
         AttentionMetadataBuilder[FlashAttentionMetadata]):
-    full_cudagraph_supported: ClassVar[bool] = get_flash_attn_version() == 3
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.NEVER if get_flash_attn_version() == 2 \
+        else AttentionCGSupport.ALWAYS
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):

vllm/v1/attention/backends/flashinfer.py

@@ -4,26 +4,28 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING, ClassVar, Optional, Union
 
 import torch
 from flashinfer import (BatchDecodeWithPagedKVCacheWrapper,
                         BatchPrefillWithPagedKVCacheWrapper,
                         MultiLevelCascadeAttentionWrapper)
-from flashinfer.decode import trtllm_batch_decode_with_kv_cache
+from flashinfer.decode import (_get_range_buf, get_seq_lens,
+                               trtllm_batch_decode_with_kv_cache)
 
 import vllm.envs as envs
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionType)
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
-from vllm.utils import cdiv
+from vllm.utils import cdiv, is_pin_memory_available
 from vllm.utils.flashinfer import use_trtllm_decode_attention
 from vllm.v1.attention.backends.flash_attn import use_cascade_attention
 from vllm.v1.attention.backends.utils import (
-    AttentionMetadataBuilder, CommonAttentionMetadata, get_kv_cache_layout,
-    get_per_layer_parameters, infer_global_hyperparameters,
-    reorder_batch_to_split_decodes_and_prefills, split_decodes_and_prefills)
+    AttentionCGSupport, AttentionMetadataBuilder, CommonAttentionMetadata,
+    get_kv_cache_layout, get_per_layer_parameters,
+    infer_global_hyperparameters, reorder_batch_to_split_decodes_and_prefills,
+    split_decodes_and_prefills)
 from vllm.v1.kv_cache_interface import AttentionSpec
 
 if TYPE_CHECKING:
@@ -174,26 +176,66 @@ class FlashInferMetadata:
 
 
 class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.PURE_DECODE_ONLY
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):
         self.device = device
+        self.vllm_config = vllm_config
+        self.cache_config = vllm_config.cache_config
+        self.kv_cache_spec = kv_cache_spec
         self._workspace_buffer = None
         self._prefill_wrapper = None  # Wrapper for prefill/append
-        self._decode_wrapper = None  # Wrapper for decode
+        self._decode_wrapper = None  # Wrapper for decode (general shape)
+
+        self.compilation_config = vllm_config.compilation_config
+        max_num_pages_per_req = cdiv(vllm_config.model_config.max_model_len,
+                                     self.kv_cache_spec.block_size)
+        max_num_reqs = vllm_config.scheduler_config.max_num_seqs
+        max_num_pages = max_num_reqs * max_num_pages_per_req
+        self.enable_cuda_graph = self.compilation_config.full_cuda_graph
+        if self.enable_cuda_graph:
+            # For full cudagraph capture, one `decode_wrapper` for each batch
+            # size is needed for FlashInfer.
+            self._decode_wrappers_cudagraph: dict[
+                int, BatchDecodeWithPagedKVCacheWrapper] = {}
+            self._decode_cudagraph_max_bs = min(
+                max_num_reqs, self.compilation_config.max_capture_size)
+
         self._cascade_wrapper = None  # Wrapper for cascade attention
 
         # Global hyperparameters shared by all attention layers
         self.global_hyperparameters = infer_global_hyperparameters(
             get_per_layer_parameters(vllm_config, layer_names, FlashInferImpl))
 
-        self.vllm_config = vllm_config
-        self.cache_config = vllm_config.cache_config
-        self.kv_cache_spec = kv_cache_spec
-        max_num_blocks_per_request = cdiv(
-            vllm_config.model_config.max_model_len,
-            self.kv_cache_spec.block_size)
-        self.block_table_arange = torch.arange(max_num_blocks_per_request,
+        # Preparing persistent buffers (device-side)
+        self.paged_kv_indptr = torch.zeros(max_num_reqs + 1,
+                                           dtype=torch.int32,
+                                           device=self.device)
+        self.paged_kv_indices = torch.zeros(
+            max_num_pages,  # max num pages possible
+            dtype=torch.int32,
+            device=self.device)
+        self.paged_kv_last_page_len = torch.zeros(max_num_reqs,
+                                                  dtype=torch.int32,
+                                                  device=self.device)
+        # host-side buffer
+        pin_memory = is_pin_memory_available()
+        self.paged_kv_indptr_cpu = torch.zeros(max_num_reqs + 1,
+                                               dtype=torch.int32,
+                                               device="cpu",
+                                               pin_memory=pin_memory)
+        self.paged_kv_indices_cpu = torch.zeros(max_num_pages,
+                                                dtype=torch.int32,
+                                                device="cpu",
+                                                pin_memory=pin_memory)
+        self.paged_kv_last_page_len_cpu = torch.zeros(max_num_reqs,
+                                                      dtype=torch.int32,
+                                                      device="cpu",
+                                                      pin_memory=pin_memory)
+
+        self.block_table_arange = torch.arange(max_num_pages_per_req,
                                                dtype=torch.int32,
                                                device=self.device)
 
@@ -217,8 +259,16 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
                 self._get_workspace_buffer(), get_kv_cache_layout())
         return self._prefill_wrapper
 
-    def _get_decode_wrapper(self):
-        if self._decode_wrapper is None:
+    def _get_decode_wrapper(self,
+                            batch_size: int,
+                            use_cudagraph: bool = False):
+        if use_cudagraph:
+            decode_wrapper = self._decode_wrappers_cudagraph.get(
+                batch_size, None)
+        else:
+            decode_wrapper = self._decode_wrapper
+
+        if decode_wrapper is None:
             num_qo_heads = (
                 self.vllm_config.model_config.get_num_attention_heads(
                     self.vllm_config.parallel_config))
@@ -226,11 +276,32 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
                 self.vllm_config.parallel_config)
             use_tensor_cores = envs.VLLM_FLASHINFER_FORCE_TENSOR_CORES or (
                 num_qo_heads // num_kv_heads > 4)
-            self._decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
+
+            if use_cudagraph:
+                paged_kv_indptr = self.paged_kv_indptr[:batch_size + 1]
+                paged_kv_indices = self.paged_kv_indices
+                paged_kv_last_page_len = self.paged_kv_last_page_len[:
+                                                                     batch_size]
+            else:
+                paged_kv_indptr = None
+                paged_kv_indices = None
+                paged_kv_last_page_len = None
+            decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
                 self._get_workspace_buffer(),
                 get_kv_cache_layout(),
+                use_cuda_graph=use_cudagraph,
+                paged_kv_indptr_buffer=paged_kv_indptr,
+                paged_kv_indices_buffer=paged_kv_indices,
+                paged_kv_last_page_len_buffer=paged_kv_last_page_len,
                 use_tensor_cores=use_tensor_cores)
-        return self._decode_wrapper
+
+            # save the decode wrapper
+            if use_cudagraph:
+                self._decode_wrappers_cudagraph[batch_size] = decode_wrapper
+            else:
+                self._decode_wrapper = decode_wrapper
+
+        return decode_wrapper
 
     def _get_cascade_wrapper(self):
         if self._cascade_wrapper is None:
@@ -308,16 +379,44 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
                 )
 
             if num_decodes > 0:
-                attn_metadata.decode_wrapper = self._get_decode_wrapper()
+                pure_decode = num_prefills == 0
+                # possible required padding for cudagraph replay
+                use_cudagraph = (self.enable_cuda_graph and pure_decode and
+                                 num_decodes <= self._decode_cudagraph_max_bs)
+                if use_cudagraph:
+                    num_input_tokens = (
+                        self.vllm_config.pad_for_cudagraph(num_decodes))
+                    # Carefully fulfill the padding region with reasonable value
+                    # on cpu.
+                    # Make sure paged_kv_indptr_cpu is not decreasing
+                    self.paged_kv_indptr_cpu[1 + num_decodes:1 +
+                                             num_input_tokens].fill_(
+                                                 attn_metadata.
+                                                 paged_kv_indptr_cpu[-1])
+                    # Fill the remaining paged_kv_last_page_len_cpu with 1.
+                    # This is because flashinfer treats 0 as a full page
+                    # instead of empty.
+                    self.paged_kv_last_page_len_cpu[
+                        num_decodes:num_input_tokens].fill_(1)
+
+                else:
+                    num_input_tokens = num_decodes
+
+                attn_metadata.decode_wrapper = self._get_decode_wrapper(
+                    num_input_tokens, use_cudagraph)
                 if not use_trtllm_decode_attention(
                         num_decodes, attn_metadata.max_seq_len,
                         self.cache_config.cache_dtype,
                         attn_metadata.num_qo_heads, attn_metadata.num_kv_heads,
                         attn_metadata.head_dim):
-                    attn_metadata.decode_wrapper.plan(
-                        attn_metadata.paged_kv_indptr_cpu[:num_decodes + 1],
+                    # Use the persistent buffer with padding length,
+                    # instead of the same address but chunked version
+                    # in atten_metadata when using cudagraph.
+                    fast_plan_decode(
+                        attn_metadata.decode_wrapper,
+                        self.paged_kv_indptr_cpu[:num_input_tokens + 1],
                         attn_metadata.paged_kv_indices,
-                        attn_metadata.paged_kv_last_page_len_cpu[:num_decodes],
+                        self.paged_kv_last_page_len_cpu[:num_input_tokens],
                         attn_metadata.num_qo_heads,
                         attn_metadata.num_kv_heads,
                         attn_metadata.head_dim,
@@ -336,6 +435,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
               common_prefix_len: int,
               common_attn_metadata: CommonAttentionMetadata,
               fast_build: bool = False) -> FlashInferMetadata:
+        num_reqs = common_attn_metadata.num_reqs
         num_actual_tokens = common_attn_metadata.num_actual_tokens
         num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens =\
             split_decodes_and_prefills(common_attn_metadata)
@@ -381,18 +481,26 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
                                                        non_blocking=True)
         mask = (self.block_table_arange[:max_num_blocks].unsqueeze(0)
                 < block_table_bounds.unsqueeze(1))
-        paged_kv_indices = block_table_tensor[:, :max_num_blocks][mask]
-
-        paged_kv_indptr_cpu = torch.zeros(len(block_table_bounds_cpu) + 1,
-                                          dtype=torch.int32,
-                                          device='cpu')
-        paged_kv_indptr_cpu[1:] = block_table_bounds_cpu.cumsum(
-            dim=0, dtype=torch.int32)
+        # write self.paged_kv_indices inplace
+        num_actual_pages = torch.sum(mask)
+        paged_kv_indices = self.paged_kv_indices[:num_actual_pages]
+        torch.masked_select(block_table_tensor[:, :max_num_blocks],
+                            mask,
+                            out=paged_kv_indices)
+
+        # write self.paged_kv_indptr_cpu inplace (0-index is always 0)
+        torch.cumsum(block_table_bounds_cpu,
+                     dim=0,
+                     dtype=torch.int32,
+                     out=self.paged_kv_indptr_cpu[1:1 + num_reqs])
 
         paged_kv_last_page_len_cpu = seq_lens_cpu % page_size
-        paged_kv_last_page_len_cpu = torch.where(
-            paged_kv_last_page_len_cpu == 0, page_size,
-            paged_kv_last_page_len_cpu)
+        # write self.paged_kv_last_page_len_cpu inplace
+        torch.where(paged_kv_last_page_len_cpu == 0,
+                    torch.tensor(page_size),
+                    paged_kv_last_page_len_cpu,
+                    out=self.paged_kv_last_page_len_cpu[:num_reqs])
+
         cache_dtype = self.cache_config.cache_dtype
         if cache_dtype.startswith("fp8"):
             kv_cache_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
@@ -402,9 +510,10 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
         attn_metadata = FlashInferMetadata(
             num_actual_tokens=num_actual_tokens,
             qo_indptr_cpu=common_attn_metadata.query_start_loc_cpu,
-            paged_kv_indptr_cpu=paged_kv_indptr_cpu,
+            paged_kv_indptr_cpu=self.paged_kv_indptr_cpu[:1 + num_reqs],
             paged_kv_indices=paged_kv_indices,
-            paged_kv_last_page_len_cpu=paged_kv_last_page_len_cpu,
+            paged_kv_last_page_len_cpu=self.
+            paged_kv_last_page_len_cpu[:num_reqs],
             num_qo_heads=self.vllm_config.model_config.get_num_attention_heads(
                 self.vllm_config.parallel_config),
             num_kv_heads=self.kv_cache_spec.num_kv_heads,
@@ -431,6 +540,26 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
 
         return attn_metadata
 
+    def build_for_cudagraph_capture(
+            self, common_attn_metadata: CommonAttentionMetadata):
+        """
+        This method builds the metadata for full cudagraph capture.
+        Currently, only decode is supported for full cudagraphs with FlashInfer.
+        """
+        m = common_attn_metadata
+
+        assert m.num_reqs == m.num_actual_tokens, \
+            "FlashInfer only supports decode-only full CUDAGraph capture. " \
+            "Make sure all cudagraph capture sizes <= max_num_seq."
+
+        m.max_query_len = 1  # decode-only
+
+        return self.build(0, m)
+
+    def can_run_in_cudagraph(
+            self, common_attn_metadata: CommonAttentionMetadata) -> bool:
+        return common_attn_metadata.max_query_len == 1
+
     def use_cascade_attention(self, *args, **kwargs) -> bool:
         if self.kv_cache_spec.dtype != self.vllm_config.model_config.dtype:
             # TODO: The cascade wrapper currently does not support setting
@@ -638,3 +767,163 @@ class FlashInferImpl(AttentionImpl):
                         out=output[:num_decode_tokens],
                     )
         return output_padded
+
+
+def fast_plan_decode(
+    self,  # decode wrapper
+    indptr_cpu: torch.Tensor,
+    indices: torch.Tensor,
+    last_page_len_cpu: torch.Tensor,
+    num_qo_heads: int,
+    num_kv_heads: int,
+    head_dim: int,
+    page_size: int,
+    pos_encoding_mode: str = "NONE",
+    window_left: int = -1,
+    logits_soft_cap: Optional[float] = None,
+    q_data_type: Optional[Union[str, torch.dtype]] = "float16",
+    kv_data_type: Optional[Union[str, torch.dtype]] = None,
+    data_type: Optional[Union[str, torch.dtype]] = None,
+    sm_scale: Optional[float] = None,
+    rope_scale: Optional[float] = None,
+    rope_theta: Optional[float] = None,
+    non_blocking: bool = True,
+) -> None:
+    """
+    A faster version of BatchDecodeWithPagedKVCacheWrapper::plan used for 
+    cudagraph capture/replay, while the no cudagraph version turns back 
+    to the original plan.
+    using original plan after passing host-side buffers:
+    - only host-to-device copy of indptr and last_page_len buffers
+    Modifications for cudagraph:
+    - only host-to-device copy of indptr and last_page_len buffers.
+    - avoid device-to-device copy of indices buffer.
+
+    Part of the code get inspiration from the original plan from FlashInfer repo
+    and the implementation of fast_decode_plan for FlashInfer in SGlang repo.
+    """
+    # Warm up with the original plan if it is first call, and always run the
+    # original plan if we run for dynamic shape. For fixed shape (cudagraph),
+    # this warm up is to generate the _cached_module for the decode wrapper.
+    if not self.is_cuda_graph_enabled or \
+        getattr(self, "vllm_first_call", True):
+        self.plan(
+            indptr_cpu,
+            indices,
+            last_page_len_cpu,
+            num_qo_heads,
+            num_kv_heads,
+            head_dim,
+            page_size,
+            pos_encoding_mode,
+            window_left,
+            logits_soft_cap,
+            q_data_type,
+            kv_data_type,
+            data_type,
+            sm_scale,
+            rope_scale,
+            rope_theta,
+            non_blocking,
+        )
+        self.vllm_first_call = False
+        return
+
+    assert self.is_cuda_graph_enabled, "Should be cudagraph only here"
+
+    batch_size = len(last_page_len_cpu)
+    if logits_soft_cap is None:
+        logits_soft_cap = 0.0
+
+    # Handle data types consistently
+    if data_type is not None:
+        if q_data_type is None:
+            q_data_type = data_type
+        if kv_data_type is None:
+            kv_data_type = data_type
+    elif q_data_type is None:
+        q_data_type = "float16"
+
+    if kv_data_type is None:
+        kv_data_type = q_data_type
+    q_data_type = getattr(torch, q_data_type) if isinstance(
+        q_data_type, str) else q_data_type
+    kv_data_type = getattr(torch, kv_data_type) if isinstance(
+        kv_data_type, str) else kv_data_type
+
+    if self.use_tensor_cores:
+        qo_indptr_host = _get_range_buf(batch_size + 1, "cpu")
+
+    if batch_size != self._fixed_batch_size:
+        raise ValueError(
+            "The batch size should be fixed in cudagraph mode, the runtime "
+            "batch size {} mismatches the batch size set during "
+            "initialization {}".format(batch_size, self._fixed_batch_size))
+    if len(indices) > len(self._paged_kv_indices_buf):
+        raise ValueError(
+            "The size of indices should be less than or equal to the "
+            "allocated buffer")
+
+    # host-to-device copy for the indptr buffer
+    self._paged_kv_indptr_buf.copy_(indptr_cpu, non_blocking=True)
+    # host-to-device copy for the last_page_len buffer
+    self._paged_kv_last_page_len_buf.copy_(last_page_len_cpu,
+                                           non_blocking=True)
+
+    indptr_host = indptr_cpu
+    last_page_len_host = last_page_len_cpu
+
+    if self.use_tensor_cores:
+        kv_lens_arr_host = get_seq_lens(indptr_host, last_page_len_host,
+                                        page_size)
+
+        try:
+            # Make sure we pass exactly 15 arguments for tensor core version
+            self._plan_info = self._cached_module.plan(
+                self._float_workspace_buffer,
+                self._int_workspace_buffer,
+                self._pin_memory_int_workspace_buffer,
+                qo_indptr_host,
+                indptr_host,
+                kv_lens_arr_host,
+                batch_size,  # total_num_rows
+                batch_size,
+                num_qo_heads,
+                num_kv_heads,
+                page_size,
+                self.is_cuda_graph_enabled,
+                head_dim,
+                head_dim,
+                False,  # causal
+            )
+        except Exception as e:
+            raise RuntimeError(f"Error in tensor core plan: {e}") from e
+    else:
+        try:
+            # Make sure we pass exactly 15 arguments for standard version
+            self._plan_info = self._cached_module.plan(
+                self._float_workspace_buffer,
+                self._int_workspace_buffer,
+                self._pin_memory_int_workspace_buffer,
+                indptr_host,
+                batch_size,
+                num_qo_heads,
+                num_kv_heads,
+                page_size,
+                self.is_cuda_graph_enabled,
+                window_left,
+                logits_soft_cap,
+                head_dim,
+                head_dim,
+                torch.empty(0, dtype=q_data_type),
+                torch.empty(0, dtype=kv_data_type),
+            )
+        except Exception as e:
+            raise RuntimeError(f"Error in standard plan: {e}") from e
+
+    self._pos_encoding_mode = pos_encoding_mode
+    self._window_left = window_left
+    self._logits_soft_cap = logits_soft_cap
+    self._sm_scale = sm_scale
+    self._rope_scale = rope_scale
+    self._rope_theta = rope_theta

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

@@ -18,6 +18,7 @@ from vllm.v1.attention.backends.mla.common import (MLACommonBackend,
                                                    MLACommonImpl,
                                                    MLACommonMetadata,
                                                    MLACommonMetadataBuilder)
+from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec
 
 logger = init_logger(__name__)
@@ -54,7 +55,8 @@ class FlashMLAMetadata(MLACommonMetadata[FlashMLADecodeMetadata]):
 
 
 class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
-    full_cudagraph_supported: ClassVar[bool] = True  # Decode-only
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.PURE_DECODE_ONLY
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):

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

@@ -17,6 +17,7 @@ from vllm.v1.attention.backends.mla.common import (MLACommonBackend,
                                                    MLACommonImpl,
                                                    MLACommonMetadata,
                                                    MLACommonMetadataBuilder)
+from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec
 
 # yapf: enable
@@ -64,7 +65,8 @@ class AiterMLAMetadata(MLACommonMetadata[AiterMLADecodeMetadata]):
 
 
 class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
-    full_cudagraph_supported: ClassVar[bool] = True  # decode only
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.PURE_DECODE_ONLY
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):

vllm/v1/attention/backends/triton_attn.py

@@ -18,7 +18,8 @@ from vllm.config import VllmConfig
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
-from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+from vllm.v1.attention.backends.utils import (AttentionCGSupport,
+                                              AttentionMetadataBuilder,
                                               CommonAttentionMetadata)
 from vllm.v1.kv_cache_interface import AttentionSpec
 
@@ -57,7 +58,8 @@ class TritonAttentionMetadata:
 
 class TritonAttentionMetadataBuilder(
         AttentionMetadataBuilder[TritonAttentionMetadata]):
-    full_cudagraph_supported: ClassVar[bool] = True
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.ALWAYS
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):

vllm/v1/attention/backends/utils.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import abc
+import enum
 import functools
 from abc import abstractmethod
 from dataclasses import dataclass, make_dataclass
@@ -65,9 +66,24 @@ class CommonAttentionMetadata:
 M = TypeVar("M")
 
 
+class AttentionCGSupport(enum.Enum):
+    """ Constants for the cudagraph support of the attention backend
+    Here we do not consider the cascade attention, as currently
+    it is never cudagraph supported."""
+
+    NEVER = 0
+    """NO cudagraph support"""
+    PURE_DECODE_ONLY = 1
+    """Cudagraph supported for pure decode, need to run without
+    cudagraph for mixed prefill-decode batches"""
+    ALWAYS = 2
+    """Cudagraph always supported"""
+
+
 class AttentionMetadataBuilder(abc.ABC, Generic[M]):
     # Does this backend/builder support CUDA Graphs for attention.
-    full_cudagraph_supported: ClassVar[bool] = False
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.NEVER
 
     @abstractmethod
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],

vllm/v1/worker/gpu_model_runner.py

@@ -47,7 +47,7 @@ from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
                         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,
+    AttentionCGSupport, AttentionMetadataBuilder, CommonAttentionMetadata,
     make_kv_sharing_fast_prefill_attention_metadata,
     make_local_attention_virtual_batches)
 from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
@@ -2619,12 +2619,22 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
             self.device,
         )
 
-        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.")
+        if self.full_cuda_graph:
+            if attn_metadata_builder_i.attn_cudagraph_support == \
+                AttentionCGSupport.NEVER:
+                raise ValueError(f"Full CUDAGraph not supported for "
+                                 f"{attn_backend_i.__name__}. Turn off "
+                                 f"CompilationConfig.full_cuda_graph or use a "
+                                 f" different attention backend.")
+            if attn_metadata_builder_i.attn_cudagraph_support == \
+                AttentionCGSupport.PURE_DECODE_ONLY:
+                # Limit the max cudagraph size to the max number of
+                # sequences for pure decode only cudagraph backend,
+                # whose max_query_len is 1.
+                self.cudagraph_batch_sizes = [
+                    size for size in self.cudagraph_batch_sizes
+                    if size <= self.scheduler_config.max_num_seqs
+                ]
         return attn_backend_i, attn_metadata_builder_i
 
     def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:

vllm/v1/worker/gpu_worker.py

@@ -321,11 +321,16 @@ class Worker(WorkerBase):
         if get_pp_group().is_last_rank:
             max_num_reqs = min(self.scheduler_config.max_num_seqs,
                                self.scheduler_config.max_num_batched_tokens)
+            # activate building attn_metadata for this dummy run to avoid
+            # potential illegal memory access for full cudagraph relay.
+            attn_cudagraph = self.compilation_config.full_cuda_graph and\
+                not self.model_config.enforce_eager
 
             # We skip EPLB here since we don't want to record dummy metrics
             hidden_states, last_hidden_states = \
                 self.model_runner._dummy_run(
                     num_tokens=max_num_reqs,
+                    capture_attn_cudagraph=attn_cudagraph,
                     skip_eplb=True,
                 )
             if self.model_runner.is_pooling_model: