[Kernel] FlashMLA integration (#13747)

Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

CMakeLists.txt

@@ -575,77 +575,8 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     WITH_SOABI)
 endif()
 
-# vllm-flash-attn currently only supported on CUDA
+# For CUDA we also build and ship some external projects.
-if (NOT VLLM_GPU_LANG STREQUAL "CUDA")
+if (VLLM_GPU_LANG STREQUAL "CUDA")
-  return()
+    include(cmake/external_projects/flashmla.cmake)
+    include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()
-
-# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target
-# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the
-# arches in the CUDA case (and instead set the gencodes on a per file basis)
-# we need to manually set VLLM_GPU_ARCHES here.
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  foreach(_ARCH ${CUDA_ARCHS})
-    string(REPLACE "." "" _ARCH "${_ARCH}")
-    list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
-  endforeach()
-endif()
-
-#
-# Build vLLM flash attention from source
-#
-# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
-# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
-# They should be identical but if they aren't, this is a massive footgun.
-#
-# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
-# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
-# If no component is specified, vllm-flash-attn is still installed.
-
-# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
-# This is to enable local development of vllm-flash-attn within vLLM.
-# It can be set as an environment variable or passed as a cmake argument.
-# The environment variable takes precedence.
-if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
-  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
-endif()
-
-if(VLLM_FLASH_ATTN_SRC_DIR)
-  FetchContent_Declare(
-          vllm-flash-attn SOURCE_DIR
-          ${VLLM_FLASH_ATTN_SRC_DIR}
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-else()
-  FetchContent_Declare(
-          vllm-flash-attn
-          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 720c94869cf2e0ff5a706e9c7f1dce0939686ade
-          GIT_PROGRESS TRUE
-          # Don't share the vllm-flash-attn build between build types
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-endif()
-
-
-# Fetch the vllm-flash-attn library
-FetchContent_MakeAvailable(vllm-flash-attn)
-message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
-
-# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
-# case only one is built, in the case both are built redundant work is done)
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
-  COMPONENT _vllm_fa2_C
-  FILES_MATCHING PATTERN "*.py"
-)
-
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
-  COMPONENT _vllm_fa3_C
-  FILES_MATCHING PATTERN "*.py"
-)
-
-# Nothing after vllm-flash-attn, see comment about macros above

cmake/external_projects/flashmla.cmake

@@ -0,0 +1,66 @@
+include(FetchContent)
+
+# If FLASH_MLA_SRC_DIR is set, flash-mla is installed from that directory 
+# instead of downloading.
+# It can be set as an environment variable or passed as a cmake argument.
+# The environment variable takes precedence.
+if (DEFINED ENV{FLASH_MLA_SRC_DIR})
+  set(FLASH_MLA_SRC_DIR $ENV{FLASH_MLA_SRC_DIR})
+endif()
+
+if(FLASH_MLA_SRC_DIR)
+  FetchContent_Declare(
+        flashmla 
+        SOURCE_DIR ${FLASH_MLA_SRC_DIR}
+        CONFIGURE_COMMAND ""
+        BUILD_COMMAND ""
+  )
+else()
+  FetchContent_Declare(
+        flashmla
+        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA.git
+        GIT_TAG 575f7724b9762f265bbee5889df9c7d630801845
+        GIT_PROGRESS TRUE
+        CONFIGURE_COMMAND ""
+        BUILD_COMMAND ""
+  )
+endif()
+
+
+FetchContent_MakeAvailable(flashmla)
+message(STATUS "FlashMLA is available at ${flashmla_SOURCE_DIR}")
+
+# The FlashMLA kernels only work on hopper and require CUDA 12.3 or later.
+# Only build FlashMLA kernels if we are building for something compatible with 
+# sm90a
+cuda_archs_loose_intersection(FLASH_MLA_ARCHS "9.0a" "${CUDA_ARCHS}")
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3 AND FLASH_MLA_ARCHS)
+    set(FlashMLA_SOURCES
+        ${flashmla_SOURCE_DIR}/csrc/flash_api.cpp
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_bf16_sm90.cu
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_fp16_sm90.cu
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_metadata.cu)
+
+    set(FlashMLA_INCLUDES
+        ${flashmla_SOURCE_DIR}/csrc/cutlass/include
+        ${flashmla_SOURCE_DIR}/csrc/include)
+
+    set_gencode_flags_for_srcs(
+        SRCS "${FlashMLA_SOURCES}"
+        CUDA_ARCHS "${FLASH_MLA_ARCHS}")
+
+    define_gpu_extension_target(
+        _flashmla_C
+        DESTINATION vllm
+        LANGUAGE ${VLLM_GPU_LANG}
+        SOURCES ${FlashMLA_SOURCES}
+        COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+        ARCHITECTURES ${VLLM_GPU_ARCHES}
+        INCLUDE_DIRECTORIES ${FlashMLA_INCLUDES}
+        USE_SABI 3
+        WITH_SOABI)
+else()
+    # Create an empty target for setup.py when not targeting sm90a systems
+    add_custom_target(_flashmla_C)
+endif()
+

cmake/external_projects/vllm_flash_attn.cmake

@@ -0,0 +1,67 @@
+# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target
+# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the
+# arches in the CUDA case (and instead set the gencodes on a per file basis)
+# we need to manually set VLLM_GPU_ARCHES here.
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  foreach(_ARCH ${CUDA_ARCHS})
+    string(REPLACE "." "" _ARCH "${_ARCH}")
+    list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
+  endforeach()
+endif()
+
+#
+# Build vLLM flash attention from source
+#
+# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
+# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
+# They should be identical but if they aren't, this is a massive footgun.
+#
+# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
+# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
+# If no component is specified, vllm-flash-attn is still installed.
+
+# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
+# This is to enable local development of vllm-flash-attn within vLLM.
+# It can be set as an environment variable or passed as a cmake argument.
+# The environment variable takes precedence.
+if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
+  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
+endif()
+
+if(VLLM_FLASH_ATTN_SRC_DIR)
+  FetchContent_Declare(
+          vllm-flash-attn SOURCE_DIR 
+          ${VLLM_FLASH_ATTN_SRC_DIR}
+          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
+  )
+else()
+  FetchContent_Declare(
+          vllm-flash-attn
+          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
+          GIT_TAG 720c94869cf2e0ff5a706e9c7f1dce0939686ade
+          GIT_PROGRESS TRUE
+          # Don't share the vllm-flash-attn build between build types
+          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
+  )
+endif()
+
+
+# Fetch the vllm-flash-attn library
+FetchContent_MakeAvailable(vllm-flash-attn)
+message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
+
+# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
+# case only one is built, in the case both are built redundant work is done)
+install(
+  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+  DESTINATION vllm_flash_attn
+  COMPONENT _vllm_fa2_C
+  FILES_MATCHING PATTERN "*.py"
+)
+
+install(
+  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+  DESTINATION vllm_flash_attn
+  COMPONENT _vllm_fa3_C
+  FILES_MATCHING PATTERN "*.py"
+)

setup.py

@@ -328,6 +328,7 @@ class repackage_wheel(build_ext):
             files_to_copy = [
                 "vllm/_C.abi3.so",
                 "vllm/_moe_C.abi3.so",
+                "vllm/_flashmla_C.abi3.so",
                 "vllm/vllm_flash_attn/_vllm_fa2_C.abi3.so",
                 "vllm/vllm_flash_attn/_vllm_fa3_C.abi3.so",
                 "vllm/vllm_flash_attn/flash_attn_interface.py",
@@ -612,6 +613,11 @@ if _is_cuda():
         # FA3 requires CUDA 12.0 or later
         ext_modules.append(
             CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa3_C"))
+    if envs.VLLM_USE_PRECOMPILED or get_nvcc_cuda_version() >= Version("12.3"):
+        # Optional since this doesn't get built (produce an .so file) when
+        # not targeting a hopper system
+        ext_modules.append(
+            CMakeExtension(name="vllm._flashmla_C", optional=True))
     ext_modules.append(CMakeExtension(name="vllm.cumem_allocator"))
 
 if _build_custom_ops():

tests/kernels/test_flashmla.py

@@ -0,0 +1,132 @@
+# Adapted from: https://github.com/deepseek-ai/FlashMLA/blob/main/tests/test_flash_mla.py
+# SPDX-License-Identifier: Apache-2.0
+import math
+import random
+
+import pytest
+import torch
+import triton
+
+from vllm.attention.ops.flashmla import (flash_mla_with_kvcache,
+                                         get_mla_metadata,
+                                         is_flashmla_supported)
+
+
+def cal_diff(x: torch.Tensor, y: torch.Tensor, name: str) -> None:
+    x, y = x.double(), y.double()
+    cos_diff = 1 - 2 * (x * y).sum().item() / max(
+        (x * x + y * y).sum().item(), 1e-12)
+    assert cos_diff < 1e-5
+
+FLASH_MLA_UNSUPPORTED_REASON = is_flashmla_supported()[1] \
+    if not is_flashmla_supported()[0] else "FlashMLA is supported"
+
+
+@pytest.mark.skipif(not is_flashmla_supported()[0],
+                    reason=FLASH_MLA_UNSUPPORTED_REASON)
+@pytest.mark.parametrize("b", [128])
+@pytest.mark.parametrize("s_q", [1, 2])
+@pytest.mark.parametrize("mean_sk", [4096, 8192])
+@pytest.mark.parametrize("h_q", [16, 32, 64, 128])
+@pytest.mark.parametrize("h_kv", [1])
+@pytest.mark.parametrize("d", [576])
+@pytest.mark.parametrize("dv", [512])
+@pytest.mark.parametrize("block_size", [64])
+@pytest.mark.parametrize("causal", [True])
+@pytest.mark.parametrize("varlen", [False, True])
+@torch.inference_mode()
+def test_flash_mla(b, s_q, mean_sk, h_q, h_kv, d, dv, block_size, causal,
+                   varlen):
+    # TODO: parametrize using pytest
+    dtype = torch.bfloat16
+    device = torch.device("cuda:0")
+    torch.set_default_dtype(dtype)
+    torch.set_default_device(device)
+    torch.cuda.set_device(device)
+    torch.manual_seed(0)
+    random.seed(0)
+
+    print(f"{b=}, {s_q=}, {mean_sk=}, {h_q=}, {h_kv=}, "
+          f"{d=}, {dv=}, {causal=}, {varlen=}")
+
+    cache_seqlens = torch.full((b, ), mean_sk, dtype=torch.int32)
+    if varlen:
+        for i in range(b):
+            cache_seqlens[i] = max(random.normalvariate(mean_sk, mean_sk / 2),
+                                   s_q)
+    total_seqlens = cache_seqlens.sum().item()
+    max_seqlen = cache_seqlens.max().item()
+    max_seqlen_pad = triton.cdiv(max_seqlen, 256) * 256
+
+    q = torch.randn(b, s_q, h_q, d)
+    block_table = torch.arange(b * max_seqlen_pad // block_size,
+                               dtype=torch.int32).view(
+                                   b, max_seqlen_pad // block_size)
+    blocked_k = torch.randn(block_table.numel(), block_size, h_kv, d)
+    for i in range(b):
+        blocked_k.view(b, max_seqlen_pad, h_kv,
+                       d)[i, cache_seqlens[i].item():] = float("nan")
+    blocked_v = blocked_k[..., :dv]
+
+    tile_scheduler_metadata, num_splits = get_mla_metadata(
+        cache_seqlens, s_q * h_q // h_kv, h_kv)
+
+    def flash_mla():
+        return flash_mla_with_kvcache(
+            q,
+            blocked_k,
+            block_table,
+            cache_seqlens,
+            dv,
+            tile_scheduler_metadata,
+            num_splits,
+            causal=causal,
+        )
+
+    def scaled_dot_product_attention(query, key, value, is_causal=False):
+        query = query.float()
+        key = key.float()
+        value = value.float()
+        key = key.repeat_interleave(h_q // h_kv, dim=0)
+        value = value.repeat_interleave(h_q // h_kv, dim=0)
+        attn_weight = query @ key.transpose(-2, -1) / math.sqrt(query.size(-1))
+        if is_causal:
+            s_q = query.shape[-2]
+            s_k = key.shape[-2]
+            attn_bias = torch.zeros(s_q, s_k, dtype=query.dtype)
+            temp_mask = torch.ones(s_q, s_k,
+                                   dtype=torch.bool).tril(diagonal=s_k - s_q)
+            attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
+            attn_bias.to(query.dtype)
+            attn_weight += attn_bias
+        lse = attn_weight.logsumexp(dim=-1)
+        attn_weight = torch.softmax(attn_weight, dim=-1, dtype=torch.float32)
+        return attn_weight @ value, lse
+
+    def ref_mla():
+        out = torch.empty(b, s_q, h_q, dv, dtype=torch.float32)
+        lse = torch.empty(b, h_q, s_q, dtype=torch.float32)
+        for i in range(b):
+            begin = i * max_seqlen_pad
+            end = begin + cache_seqlens[i]
+            ref_O, LSE = scaled_dot_product_attention(
+                q[i].transpose(0, 1),
+                blocked_k.view(-1, h_kv, d)[begin:end].transpose(0, 1),
+                blocked_v.view(-1, h_kv, dv)[begin:end].transpose(0, 1),
+                is_causal=causal,
+            )
+            out[i] = ref_O.transpose(0, 1)
+            lse[i] = LSE
+        return out, lse
+
+    out_flash, lse_flash = flash_mla()
+    out_torch, lse_torch = ref_mla()
+    cal_diff(out_flash, out_torch, "out")
+    cal_diff(lse_flash, lse_torch, "lse")
+
+    t = triton.testing.do_bench(flash_mla, fast_flush=False)
+    FLOPS = s_q * total_seqlens * h_q * (d + dv) * 2
+    bytes = (total_seqlens * h_kv * d + b * s_q * h_q * d +
+             b * s_q * h_q * dv) * (torch.finfo(dtype).bits // 8)
+    print(f"{t:.3f} ms, {FLOPS / 10 ** 9 / t:.0f} "
+          f"TFLOPS, {bytes / 10 ** 6 / t:.0f} GB/s")

vllm/_custom_ops.py

@@ -1163,3 +1163,67 @@ def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
 def register_graph_buffers(fa: int, handles: List[List[int]],
                            offsets: List[List[int]]) -> None:
     torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
+
+
+def get_flash_mla_metadata(
+    cache_seqlens: torch.Tensor,
+    num_heads_per_head_k: int,
+    num_heads_k: int,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Arguments:
+        cache_seqlens: (batch_size), dtype torch.int32.
+        num_heads_per_head_k: Equals to seq_len_q * num_heads_q // num_heads_k.
+        num_heads_k: num_heads_k.
+
+    Return:
+        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), dtype torch.int32.
+        num_splits: (batch_size + 1), dtype torch.int32.
+    """
+    return torch.ops._C.get_flash_mla_metadata(cache_seqlens,
+                                               num_heads_per_head_k,
+                                               num_heads_k)
+
+
+def flash_mla_with_kvcache(
+    q: torch.Tensor,
+    k_cache: torch.Tensor,
+    block_table: torch.Tensor,
+    cache_seqlens: torch.Tensor,
+    head_dim_v: int,
+    tile_scheduler_metadata: torch.Tensor,
+    num_splits: torch.Tensor,
+    softmax_scale: Optional[float] = None,
+    causal: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Arguments:
+        q: (batch_size, seq_len_q, num_heads_q, head_dim).
+        k_cache: (num_blocks, page_block_size, num_heads_k, head_dim).
+        block_table: (batch_size, max_num_blocks_per_seq), torch.int32.
+        cache_seqlens: (batch_size), torch.int32.
+        head_dim_v: Head_dim of v.
+        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), torch.int32, return by get_mla_metadata.
+        num_splits: (batch_size + 1), torch.int32, return by get_mla_metadata.
+        softmax_scale: float. The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim).
+        causal: bool. Whether to apply causal attention mask.
+
+    Return:
+        out: (batch_size, seq_len_q, num_heads_q, head_dim_v).
+        softmax_lse: (batch_size, num_heads_q, seq_len_q), torch.float32.
+    """
+    if softmax_scale is None:
+        softmax_scale = q.shape[-1]**(-0.5)
+    out, softmax_lse = torch.ops._C.flash_mla_fwd_kvcache(
+        q,
+        k_cache,
+        None,
+        head_dim_v,
+        cache_seqlens,
+        block_table,
+        softmax_scale,
+        causal,
+        tile_scheduler_metadata,
+        num_splits,
+    )
+    return out, softmax_lse

vllm/attention/backends/flashmla.py

@@ -0,0 +1,239 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from contextlib import contextmanager
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.backends.mla.common import (MLACommonBackend,
+                                                MLACommonImpl,
+                                                MLACommonMetadata,
+                                                MLACommonMetadataBuilder,
+                                                MLACommonState)
+from vllm.attention.ops.flashmla import (flash_mla_with_kvcache,
+                                         get_mla_metadata,
+                                         is_flashmla_supported)
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
+
+
+class FlashMLABackend(MLACommonBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASHMLA"
+
+    @staticmethod
+    def get_impl_cls() -> Type["FlashMLAImpl"]:
+        return FlashMLAImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["FlashMLAMetadata"]:
+        return FlashMLAMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["FlashMLAMetadataBuilder"]:
+        return FlashMLAMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["FlashMLAState"]:
+        return FlashMLAState
+
+
+@dataclass
+class FlashMLAMetadata(MLACommonMetadata):
+    decode_tile_scheduler_metadata: Optional[Tuple[torch.Tensor,
+                                                   torch.Tensor]] = None
+    decode_num_splits: Optional[torch.Tensor] = None
+
+    @property
+    def decode_metadata(self):
+        decode_metadata = super().decode_metadata
+        # TODO: cache assignment?
+        if decode_metadata is not None:
+            decode_metadata.decode_tile_scheduler_metadata=\
+                self.decode_tile_scheduler_metadata
+            decode_metadata.decode_num_splits=\
+                self.decode_num_splits
+        return decode_metadata
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        raise NotImplementedError(
+            "advance_step is not implemented for FlashMLA")
+
+
+class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        self.num_q_heads = self.runner.model_config.get_num_attention_heads(
+            self.runner.parallel_config)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        m = super().build(seq_lens, query_lens, cuda_graph_pad_size,
+                          batch_size)
+
+        if m.num_decode_tokens > 0:
+            m.decode_tile_scheduler_metadata, m.decode_num_splits = \
+                get_mla_metadata(
+                m.seq_lens_tensor[m.num_prefills:],
+                self.num_q_heads,
+                1, # MQA for the decode path
+            )
+
+        return m
+
+
+class FlashMLAState(MLACommonState[FlashMLAMetadata]):
+
+    def __init__(self, *args, **kwds):
+        super().__init__(*args, **kwds)
+
+        self.num_q_heads = self.runner.model_config.get_num_attention_heads(
+            self.runner.parallel_config)
+
+    @contextmanager
+    def graph_capture(self, max_batch_size: int):
+        # Run a dummy `get_mla_metadata` so we can get the right shapes
+        self._graph_decoder_tile_scheduler_metadata, \
+            self._graph_decode_num_splits = get_mla_metadata(
+            torch.ones(
+                max_batch_size, dtype=torch.int32, device=self.runner.device),
+            self.num_q_heads,
+            1, # MQA for the decode path
+        )
+
+        with super().graph_capture(max_batch_size):
+            yield
+
+        del self._graph_decoder_tile_scheduler_metadata
+        del self._graph_decode_num_splits
+
+    def graph_capture_get_metadata_for_batch(
+            self, batch_size: int, is_encoder_decoder_model: bool = False):
+        metadata = super().graph_capture_get_metadata_for_batch(
+            batch_size, is_encoder_decoder_model)
+        assert metadata.num_decode_tokens > 0
+
+        decoder_tile_scheduler_metadata, decode_num_splits = get_mla_metadata(
+            self._graph_seq_lens[:batch_size],
+            self.num_q_heads,
+            1,  # MQA for the decode path
+        )
+
+        self._graph_decoder_tile_scheduler_metadata.copy_(
+            decoder_tile_scheduler_metadata)
+        self._graph_decode_num_splits[:batch_size + 1].copy_(decode_num_splits)
+
+        metadata.decode_tile_scheduler_metadata=\
+            self._graph_decoder_tile_scheduler_metadata
+        metadata.decode_num_splits=\
+            self._graph_decode_num_splits[:batch_size + 1]
+
+        return metadata
+
+    def get_graph_input_buffers(self,
+                                attn_metadata,
+                                is_encoder_decoder_model: bool = False):
+        input_buffers = super().get_graph_input_buffers(
+            attn_metadata, is_encoder_decoder_model)
+        input_buffers["decode_tile_scheduler_metadata"] = \
+                attn_metadata.decode_metadata.decode_tile_scheduler_metadata
+        input_buffers["decode_num_splits"] = \
+                attn_metadata.decode_metadata.decode_num_splits
+
+        return input_buffers
+
+    def prepare_graph_input_buffers(self,
+                                    input_buffers,
+                                    attn_metadata,
+                                    is_encoder_decoder_model: bool = False):
+        super().prepare_graph_input_buffers(input_buffers, attn_metadata,
+                                            is_encoder_decoder_model)
+
+        input_buffers["decode_tile_scheduler_metadata"].copy_(
+            attn_metadata.decode_metadata.decode_tile_scheduler_metadata)
+        input_buffers["decode_num_splits"].copy_(
+            attn_metadata.decode_metadata.decode_num_splits)
+
+
+class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
+
+    def __init__(
+            self,
+            num_heads: int,
+            head_size: int,
+            scale: float,
+            num_kv_heads: int,
+            alibi_slopes: Optional[List[float]],
+            sliding_window: Optional[int],
+            kv_cache_dtype: str,
+            blocksparse_params: Optional[Dict[str, Any]],
+            logits_soft_cap: Optional[float],
+            attn_type: str,
+            # MLA Specific Arguments
+            **mla_args) -> None:
+        super().__init__(num_heads, head_size, scale, num_kv_heads,
+                         alibi_slopes, sliding_window, kv_cache_dtype,
+                         blocksparse_params, logits_soft_cap, attn_type,
+                         **mla_args)
+
+        assert is_flashmla_supported(), \
+            "FlashMLA is not supported on this device"
+
+        unsupported_features = [
+            alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap
+        ]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "FlashMLAImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, blocksparse_params, "
+                "logits_soft_cap")
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "FlashMLAImpl")
+
+    def _forward_decode(
+        self,
+        q_nope: torch.Tensor,
+        q_pe: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: FlashMLAMetadata,
+    ) -> torch.Tensor:
+        assert kv_c_and_k_pe_cache.numel() > 0
+        if self.kv_cache_dtype.startswith("fp8"):
+            raise NotImplementedError("FP8 FlashMLA not yet supported")
+
+        decode_meta = attn_metadata.decode_metadata
+        assert decode_meta is not None
+
+        q = torch.cat([q_nope, q_pe], dim=-1)\
+            .unsqueeze(1) # Add seqlen dim of 1 (decode)
+
+        o, _ = flash_mla_with_kvcache(
+            q=q,
+            k_cache=kv_c_and_k_pe_cache.unsqueeze(-2),  # Add head dim of 1
+            block_table=decode_meta.block_tables,
+            cache_seqlens=decode_meta.seq_lens_tensor,
+            head_dim_v=self.kv_lora_rank,
+            tile_scheduler_metadata=decode_meta.decode_tile_scheduler_metadata,
+            num_splits=decode_meta.decode_num_splits,
+            softmax_scale=self.scale,
+            causal=True,
+        )
+
+        return self._v_up_proj_and_o_proj(o)

vllm/attention/backends/mla/common.py

@@ -293,7 +293,10 @@ class MLACommonBackend(AttentionBackend):
         return [576]
 
 
-class MLACommonState(AttentionState):
+T = TypeVar("T", bound="MLACommonMetadata")
+
+
+class MLACommonState(AttentionState, Generic[T]):
 
     def __init__(self, runner):
         self.runner = runner
@@ -355,7 +358,9 @@ class MLACommonState(AttentionState):
         return self.__class__(self.runner)
 
     def graph_capture_get_metadata_for_batch(
-            self, batch_size: int, is_encoder_decoder_model: bool = False):
+            self,
+            batch_size: int,
+            is_encoder_decoder_model: bool = False) -> T:
         assert self._is_graph_capturing
 
         attn_metadata = self.runner.attn_backend.make_metadata(
@@ -507,8 +512,8 @@ class MLACommonMetadata(AttentionMetadata):
     # [4, 6], it is [0, 4, 10].
     seq_start_loc: Optional[torch.Tensor] = None
 
-    _cached_prefill_metadata: Optional["MLACommonMetadata"] = None
+    _cached_prefill_metadata: Optional[Any] = None
-    _cached_decode_metadata: Optional["MLACommonMetadata"] = None
+    _cached_decode_metadata: Optional[Any] = None
 
     num_prefill_tokens: int
 
@@ -537,7 +542,7 @@ class MLACommonMetadata(AttentionMetadata):
                 f" received {self.head_dim}.")
 
     @property
-    def prefill_metadata(self) -> Optional["MLACommonMetadata"]:
+    def prefill_metadata(self):
         if self.num_prefills == 0:
             return None
 
@@ -565,7 +570,7 @@ class MLACommonMetadata(AttentionMetadata):
         input_positions = (None if self.input_positions is None else
                            self.input_positions[:self.num_prefill_tokens])
 
-        self._cached_prefill_metadata = MLACommonMetadata(
+        self._cached_prefill_metadata = self.__class__(
             # Required by ModelRunner
             use_cuda_graph=False,  # Not Attention Related
             # Required by Attention Metadata
@@ -599,7 +604,7 @@ class MLACommonMetadata(AttentionMetadata):
         return self._cached_prefill_metadata
 
     @property
-    def decode_metadata(self) -> Optional["MLACommonMetadata"]:
+    def decode_metadata(self):
         if self.num_decode_tokens == 0:
             return None
 
@@ -617,7 +622,7 @@ class MLACommonMetadata(AttentionMetadata):
         input_positions = (None if self.input_positions is None else
                            self.input_positions[self.num_prefill_tokens:])
 
-        self._cached_decode_metadata = MLACommonMetadata(
+        self._cached_decode_metadata = self.__class__(
             # Required by ModelRunner
             use_cuda_graph=self.use_cuda_graph,  # Not Attention Related
             # Required by Attention Metadata
@@ -723,10 +728,7 @@ class MLACommonMetadata(AttentionMetadata):
                                    block_tables=self.block_tables)
 
 
-T = TypeVar("T", bound=MLACommonMetadata)
+class MLACommonMetadataBuilder(AttentionMetadataBuilder[T], Generic[T]):
-
-
-class MLACommonMetadataBuilder(AttentionMetadataBuilder[MLACommonMetadata]):
     """
     NOTE: Please read the comment at the top of the file before trying to 
     understand this class
@@ -959,7 +961,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[MLACommonMetadata]):
             assert max(context_chunk_seq_tot) <= \
                 self.chunked_prefill_workspace_size
 
-        return MLACommonMetadata(
+        return self.runner.attn_backend.make_metadata(
             # Required by ModelRunner
             use_cuda_graph=use_captured_graph,  # Not Attention Related
             # Required by Attention Metadata

vllm/attention/ops/flashmla.py

@@ -0,0 +1,115 @@
+# SPDX-License-Identifier: Apache-2.0
+# adapted from: https://github.com/deepseek-ai/FlashMLA/blob/main/flash_mla/flash_mla_interface.py
+from typing import Optional, Tuple
+
+import torch
+
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+
+logger = init_logger(__name__)
+
+if current_platform.is_cuda():
+    try:
+        import vllm._flashmla_C  # noqa: F401
+        _flashmla_C_AVAILABLE = True
+    except ImportError:
+        _flashmla_C_AVAILABLE = False
+else:
+    _flashmla_C_AVAILABLE = False
+
+
+def is_flashmla_supported() -> Tuple[bool, Optional[str]]:
+    """
+    Return: is_supported_flag, unsupported_reason (optional).
+    """
+    if not current_platform.is_cuda():
+        return False, "FlashMLA is only supported on CUDA devices."
+    if current_platform.get_device_capability()[0] != 9:
+        return False, "FlashMLA is only supported on Hopper devices."
+    if not _flashmla_C_AVAILABLE:
+        return False, "vllm._flashmla_C is not available, likely was not "\
+            "compiled due to insufficient nvcc version or a supported arch "\
+            "(only sm90a currently) was not in the list of target arches to "\
+            "compile for."
+    return True, None
+
+
+def get_mla_metadata(
+    cache_seqlens: torch.Tensor,
+    num_heads_per_head_k: int,
+    num_heads_k: int,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Arguments:
+        cache_seqlens: (batch_size), dtype torch.int32.
+        num_heads_per_head_k: Equals to seq_len_q * num_heads_q // num_heads_k.
+        num_heads_k: num_heads_k.
+
+    Return:
+        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), 
+                                 dtype torch.int32.
+        num_splits: (batch_size + 1), dtype torch.int32.
+    """
+    return torch.ops._flashmla_C.get_mla_metadata(cache_seqlens,
+                                                  num_heads_per_head_k,
+                                                  num_heads_k)
+
+
+def flash_mla_with_kvcache(
+    q: torch.Tensor,
+    k_cache: torch.Tensor,
+    block_table: torch.Tensor,
+    cache_seqlens: torch.Tensor,
+    head_dim_v: int,
+    tile_scheduler_metadata: torch.Tensor,
+    num_splits: torch.Tensor,
+    softmax_scale: Optional[float] = None,
+    causal: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Arguments:
+        q: (batch_size, seq_len_q, num_heads_q, head_dim).
+        k_cache: (num_blocks, page_block_size, num_heads_k, head_dim).
+        block_table: (batch_size, max_num_blocks_per_seq), torch.int32.
+        cache_seqlens: (batch_size), torch.int32.
+        head_dim_v: Head_dim of v.
+        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), 
+                                 torch.int32, return by get_mla_metadata.
+        num_splits: (batch_size + 1), torch.int32, return by get_mla_metadata.
+        softmax_scale: float. The scaling of QK^T before applying softmax. 
+                       Default to 1 / sqrt(head_dim).
+        causal: bool. Whether to apply causal attention mask.
+
+    Return:
+        out: (batch_size, seq_len_q, num_heads_q, head_dim_v).
+        softmax_lse: (batch_size, num_heads_q, seq_len_q), torch.float32.
+    """
+    if softmax_scale is None:
+        softmax_scale = q.shape[-1]**(-0.5)
+    out, softmax_lse = torch.ops._flashmla_C.fwd_kvcache_mla(
+        q,
+        k_cache,
+        None,
+        head_dim_v,
+        cache_seqlens,
+        block_table,
+        softmax_scale,
+        causal,
+        tile_scheduler_metadata,
+        num_splits,
+    )
+    return out, softmax_lse
+
+
+#
+# TODO: Add fake functions
+#
+# @register_fake("_flashmla_C::get_mla_metadata")
+# def _get_mla_metadata_fake(....) -> Tuple[torch.Tensor, torch.Tensor]:
+#     return ....
+#
+# @register_fake("_flashmla_C::fwd_kvcache_mla")
+# def _fwd_kvcache_mla_fake(....) -> Tuple[torch.Tensor, torch.Tensor]:
+#     return ....
+#

vllm/platforms/cuda.py

@@ -141,6 +141,14 @@ class CudaPlatformBase(Platform):
         cache_config = vllm_config.cache_config
         if cache_config and cache_config.block_size is None:
             cache_config.block_size = 16
+        # TODO(lucas): handle this more gracefully
+        if envs.VLLM_ATTENTION_BACKEND is not None \
+           and envs.VLLM_ATTENTION_BACKEND == "FLASHMLA" \
+           and cache_config.block_size != 64:
+            cache_config.block_size = 64
+            logger.info(
+                "FlashMLA: Forcing kv cache block size to 64 since this"
+                " is currently the only block size supported by the kernel.")
 
     @classmethod
     def get_current_memory_usage(cls,
@@ -157,6 +165,22 @@ class CudaPlatformBase(Platform):
             logger.info("Using Flash Attention backend on V1 engine.")
             return "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"
         if use_mla:
+            if selected_backend == _Backend.FLASHMLA:
+                from vllm.attention.backends.flashmla import (
+                    is_flashmla_supported)
+                if not is_flashmla_supported()[0]:
+                    logger.warning(
+                        "FlashMLA backend is not supported due to %s",
+                        is_flashmla_supported()[1])
+                elif block_size != 64:
+                    logger.warning(
+                        "FlashMLA backend is not supported for block size %d"
+                        " (currently only supports block size 64).",
+                        block_size)
+                else:
+                    logger.info("Using FlashMLA backend.")
+                    return "vllm.attention.backends.flashmla.FlashMLABackend"
+
             logger.info("Using Triton MLA backend.")
             return "vllm.attention.backends.triton_mla.TritonMLABackend"
         if selected_backend == _Backend.FLASHINFER:

vllm/platforms/interface.py

@@ -35,6 +35,7 @@ class _Backend(enum.Enum):
     OPENVINO = enum.auto()
     FLASHINFER = enum.auto()
     TRITON_MLA = enum.auto()
+    FLASHMLA = enum.auto()
     HPU_ATTN = enum.auto()
     PALLAS = enum.auto()
     PALLAS_VLLM_V1 = enum.auto()