Add minimal vLLM 0.16.1 build repo for BI-V150

vllm/v1/attention/backend.py

@@ -0,0 +1,885 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, replace
+from enum import Enum
+from typing import TYPE_CHECKING, Any, ClassVar, Generic, Protocol, TypeVar, get_args
+
+import numpy as np
+import torch
+from typing_extensions import deprecated
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+    from vllm.config.cache import CacheDType
+    from vllm.model_executor.layers.linear import ColumnParallelLinear
+    from vllm.model_executor.layers.quantization.utils.quant_utils import QuantKey
+    from vllm.platforms.interface import DeviceCapability
+    from vllm.v1.attention.backends.utils import KVCacheLayoutType
+    from vllm.v1.kv_cache_interface import AttentionSpec
+
+
+class AttentionType(str, Enum):
+    """
+    Attention type.
+    Use string to be compatible with `torch.compile`.
+    """
+
+    DECODER = "decoder"
+    """Decoder attention between previous layer Q/K/V."""
+    ENCODER = "encoder"
+    """Encoder attention between previous layer Q/K/V for encoder-decoder."""
+    ENCODER_ONLY = "encoder_only"
+    """Encoder attention between previous layer Q/K/V."""
+    ENCODER_DECODER = "encoder_decoder"
+    """Attention between dec. Q and enc. K/V for encoder-decoder."""
+
+
+class MultipleOf:
+    base: int
+
+    def __init__(self, base: int):
+        self.base = base
+
+
+class AttentionBackend(ABC):
+    """Abstract class for attention backends."""
+
+    # For some attention backends, we allocate an output tensor before
+    # calling the custom op. When piecewise cudagraph is enabled, this
+    # makes sure the output tensor is allocated inside the cudagraph.
+    accept_output_buffer: bool = False
+    supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+    supported_kv_cache_dtypes: ClassVar[list["CacheDType"]] = ["auto", "bfloat16"]
+
+    # Does attention's forward() include kv cache update?
+    forward_includes_kv_cache_update: bool = True
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [MultipleOf(1)]
+
+    @staticmethod
+    @abstractmethod
+    def get_name() -> str:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_impl_cls() -> type["AttentionImplBase"]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_builder_cls():  # -> Type["AttentionMetadataBuilder"]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        raise NotImplementedError
+
+    @staticmethod
+    def get_kv_cache_stride_order(
+        include_num_layers_dimension: bool = False,
+    ) -> tuple[int, ...]:
+        """
+        Get the physical (memory layout) ordering of the kv cache dimensions.
+        e.g. if the KV cache shape is
+        [2, num_blocks, block_size, num_heads, head_size],
+        and get_kv_cache_stride_order returns (1, 3, 0, 2, 4) then the physical
+        ordering of dimensions is
+        [num_blocks, num_heads, 2, block_size, head_size].
+
+        If this function is unimplemented / raises NotImplementedError,
+        the physical layout of the KV cache will match the logical shape.
+
+        Args:
+            include_num_layers_dimension: if True, includes an additional
+                num_layers dimension, which is assumed to be prepended
+                to the logical KV cache shape.
+                With the above example, a return value (2, 4, 0, 1, 3, 5)
+                corresponds to
+                [num_blocks, num_heads, num_layers, 2, block_size, head_size].
+
+                If an additional dimension is NOT included in the returned
+                tuple, the physical layout will not include a layers dimension.
+
+        Returns:
+            A tuple of ints which is a permutation of range(len(shape)).
+        """
+        raise NotImplementedError
+
+    @classmethod
+    def full_cls_name(cls) -> tuple[str, str]:
+        return (cls.__module__, cls.__qualname__)
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return []
+
+    @classmethod
+    def supports_head_size(cls, head_size: int) -> bool:
+        supported_head_sizes = cls.get_supported_head_sizes()
+        return (not supported_head_sizes) or head_size in supported_head_sizes
+
+    @classmethod
+    def supports_dtype(cls, dtype: torch.dtype) -> bool:
+        return dtype in cls.supported_dtypes
+
+    @classmethod
+    def supports_kv_cache_dtype(cls, kv_cache_dtype: "CacheDType | None") -> bool:
+        if kv_cache_dtype is None:
+            return True
+        return (not cls.supported_kv_cache_dtypes) or (
+            kv_cache_dtype in cls.supported_kv_cache_dtypes
+        )
+
+    @classmethod
+    def supports_block_size(cls, block_size: int | None) -> bool:
+        from vllm.config.cache import BlockSize
+
+        if block_size is None:
+            return True
+
+        valid_sizes = get_args(BlockSize)
+        if block_size not in valid_sizes:
+            return False
+
+        supported_kernel_block_sizes = cls.get_supported_kernel_block_sizes()
+        if not supported_kernel_block_sizes:
+            return True
+
+        for supported_size in supported_kernel_block_sizes:
+            if isinstance(supported_size, MultipleOf):
+                supported_size = supported_size.base
+            # With hybrid_blocks feature, the framework-level block size
+            # only needs to be a multiple of the kernel's requirement,
+            # even if the kernel requires a fixed block_size.
+            if block_size % supported_size == 0:
+                return True
+        return False
+
+    @classmethod
+    def is_mla(cls) -> bool:
+        return False
+
+    @classmethod
+    def supports_sink(cls) -> bool:
+        return False
+
+    @classmethod
+    def supports_alibi_sqrt(cls) -> bool:
+        return False
+
+    @classmethod
+    def supports_mm_prefix(cls) -> bool:
+        return False
+
+    @classmethod
+    def is_sparse(cls) -> bool:
+        return False
+
+    @classmethod
+    def supports_per_head_quant_scales(cls) -> bool:
+        return False
+
+    @classmethod
+    def supports_attn_type(cls, attn_type: str) -> bool:
+        """Check if backend supports a given attention type.
+
+        By default, only supports decoder attention.
+        Backends should override this to support other attention types.
+        """
+        return attn_type == AttentionType.DECODER
+
+    @classmethod
+    def supports_compute_capability(cls, capability: "DeviceCapability") -> bool:
+        return True
+
+    @classmethod
+    def supports_combination(
+        cls,
+        head_size: int,
+        dtype: torch.dtype,
+        kv_cache_dtype: "CacheDType | None",
+        block_size: int,
+        use_mla: bool,
+        has_sink: bool,
+        use_sparse: bool,
+        device_capability: "DeviceCapability",
+    ) -> str | None:
+        return None
+
+    @classmethod
+    def validate_configuration(
+        cls,
+        head_size: int,
+        dtype: torch.dtype,
+        kv_cache_dtype: "CacheDType | None",
+        block_size: int,
+        use_mla: bool,
+        has_sink: bool,
+        use_sparse: bool,
+        use_mm_prefix: bool,
+        use_per_head_quant_scales: bool,
+        device_capability: "DeviceCapability",
+        attn_type: str,
+    ) -> list[str]:
+        invalid_reasons = []
+        if not cls.supports_head_size(head_size):
+            invalid_reasons.append("head_size not supported")
+        if not cls.supports_dtype(dtype):
+            invalid_reasons.append("dtype not supported")
+        if not cls.supports_kv_cache_dtype(kv_cache_dtype):
+            invalid_reasons.append("kv_cache_dtype not supported")
+        if not cls.supports_block_size(block_size):
+            invalid_reasons.append("block_size not supported")
+        if use_mm_prefix and not cls.supports_mm_prefix():
+            invalid_reasons.append(
+                "partial multimodal token full attention not supported"
+            )
+        if use_mla != cls.is_mla():
+            if use_mla:
+                invalid_reasons.append("MLA not supported")
+            else:
+                invalid_reasons.append("non-MLA not supported")
+        if has_sink and not cls.supports_sink():
+            invalid_reasons.append("sink setting not supported")
+        if use_sparse != cls.is_sparse():
+            if use_sparse:
+                invalid_reasons.append("sparse not supported")
+            else:
+                invalid_reasons.append("non-sparse not supported")
+        if use_per_head_quant_scales and not cls.supports_per_head_quant_scales():
+            invalid_reasons.append("per-head quant scales not supported")
+        if not cls.supports_compute_capability(device_capability):
+            invalid_reasons.append("compute capability not supported")
+        if not cls.supports_attn_type(attn_type):
+            invalid_reasons.append(f"attention type {attn_type} not supported")
+        combination_reason = cls.supports_combination(
+            head_size,
+            dtype,
+            kv_cache_dtype,
+            block_size,
+            use_mla,
+            has_sink,
+            use_sparse,
+            device_capability,
+        )
+        if combination_reason is not None:
+            invalid_reasons.append(combination_reason)
+        return invalid_reasons
+
+    @classmethod
+    def get_required_kv_cache_layout(cls) -> "KVCacheLayoutType | None":
+        return None
+
+
+class AttentionMetadata:
+    pass
+
+
+T = TypeVar("T", bound=AttentionMetadata)
+
+
+@dataclass
+class CommonAttentionMetadata:
+    """
+    Per-batch attention metadata, shared across layers and backends.
+    AttentionMetadataBuilder instances use it to construct per-layer metadata.
+
+    For many of the tensors we keep both GPU and CPU versions.
+    """
+
+    query_start_loc: torch.Tensor
+    query_start_loc_cpu: torch.Tensor
+    """(batch_size + 1,), the start location of each request in query Tensor"""
+
+    seq_lens: torch.Tensor
+    """(batch_size,), the number of computed tokens for each request"""
+
+    num_reqs: int
+    """Number of requests"""
+    # TODO(lucas): rename to num_tokens since it may be padded and this is misleading
+    num_actual_tokens: int
+    """Total number of tokens in batch"""
+    max_query_len: int
+    """Longest query in batch"""
+    max_seq_len: int
+    """Longest context length (may be an upper bound)"""
+
+    block_table_tensor: torch.Tensor
+    slot_mapping: torch.Tensor
+
+    causal: bool = True
+
+    # Needed by FastPrefillAttentionBuilder
+    logits_indices_padded: torch.Tensor | None = None
+    num_logits_indices: int | None = None
+
+    # Needed by CrossAttentionBuilder
+    encoder_seq_lens: torch.Tensor | None = None
+    encoder_seq_lens_cpu: np.ndarray | None = None
+
+    dcp_local_seq_lens: torch.Tensor | None = None
+    dcp_local_seq_lens_cpu: torch.Tensor | None = None
+    """Sequence lengths of the local rank in decode context parallelism world"""
+
+    # WARNING: Deprecated fields. Will be removed in a future release (v0.15.0)
+    _seq_lens_cpu: torch.Tensor | None = None
+    _num_computed_tokens_cpu: torch.Tensor | None = None
+
+    _num_computed_tokens_cache: torch.Tensor | None = None
+
+    def batch_size(self) -> int:
+        return self.seq_lens.shape[0]
+
+    def naive_query_lens(self) -> torch.Tensor:
+        """Naive because it assumes that query ends where the next query starts."""
+        return self.query_start_loc[1:] - self.query_start_loc[:-1]
+
+    def replace(self, **kwargs) -> "CommonAttentionMetadata":
+        return replace(self, **kwargs)
+
+    @property
+    @deprecated(
+        """
+    Prefer using device seq_lens directly to avoid implicit H<>D sync.
+    If a CPU copy is needed, use `seq_lens.cpu()` instead.
+    Will be removed in a future release, please migrate as soon as possible.
+    """
+    )
+    def seq_lens_cpu(self) -> torch.Tensor:
+        if self._seq_lens_cpu is None:
+            self._seq_lens_cpu = self.seq_lens.to("cpu")
+        return self._seq_lens_cpu
+
+    @property
+    @deprecated(
+        """
+    Prefer using device seq_lens directly to avoid implicit H<>D sync which breaks full
+    async scheduling. If a CPU copy is needed, it can be derived from 
+    query_start_loc_cpu and seq_lens.
+    Will be removed in a future release, please migrate as soon as possible.
+    """
+    )
+    def num_computed_tokens_cpu(self) -> torch.Tensor:
+        if self._num_computed_tokens_cpu is None:
+            query_seq_lens = (
+                self.query_start_loc_cpu[1:] - self.query_start_loc_cpu[:-1]
+            )
+            self._num_computed_tokens_cpu = self.seq_lens_cpu - query_seq_lens
+        return self._num_computed_tokens_cpu
+
+    def compute_num_computed_tokens(self) -> torch.Tensor:
+        """Compute num_computed_tokens on device (seq_lens - query_lens)."""
+        if self._num_computed_tokens_cache is None:
+            query_lens = self.query_start_loc[1:] - self.query_start_loc[:-1]
+            self._num_computed_tokens_cache = self.seq_lens - query_lens
+        return self._num_computed_tokens_cache
+
+    # TODO(lucas): remove once we have FULL-CG spec-decode support
+    def unpadded(
+        self, num_actual_tokens: int, num_actual_reqs: int
+    ) -> "CommonAttentionMetadata":
+        maybe_slice_reqs = lambda x: x[:num_actual_reqs] if x is not None else None
+        return CommonAttentionMetadata(
+            query_start_loc=self.query_start_loc[: num_actual_reqs + 1],
+            query_start_loc_cpu=self.query_start_loc_cpu[: num_actual_reqs + 1],
+            seq_lens=self.seq_lens[:num_actual_reqs],
+            _seq_lens_cpu=self._seq_lens_cpu[:num_actual_reqs]
+            if self._seq_lens_cpu is not None
+            else None,
+            _num_computed_tokens_cpu=self._num_computed_tokens_cpu[:num_actual_reqs]
+            if self._num_computed_tokens_cpu is not None
+            else None,
+            num_reqs=num_actual_reqs,
+            num_actual_tokens=num_actual_tokens,
+            max_query_len=self.max_query_len,
+            max_seq_len=self.max_seq_len,
+            block_table_tensor=self.block_table_tensor[:num_actual_reqs],
+            slot_mapping=self.slot_mapping[:num_actual_tokens],
+            causal=self.causal,
+            logits_indices_padded=self.logits_indices_padded,
+            num_logits_indices=self.num_logits_indices,
+            encoder_seq_lens=maybe_slice_reqs(self.encoder_seq_lens),
+            encoder_seq_lens_cpu=maybe_slice_reqs(self.encoder_seq_lens_cpu),
+            dcp_local_seq_lens=maybe_slice_reqs(self.dcp_local_seq_lens),
+            dcp_local_seq_lens_cpu=maybe_slice_reqs(self.dcp_local_seq_lens_cpu),
+        )
+
+
+M = TypeVar("M")
+
+
+class AttentionCGSupport(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."""
+
+    ALWAYS = 3
+    """Cudagraph always supported; supports mixed-prefill-decode"""
+    UNIFORM_BATCH = 2
+    """Cudagraph supported for batches the only contain query lengths that are
+    the same, this can be used for spec-decode
+        i.e. "decodes" are 1 + num_speculative_tokens"""
+    UNIFORM_SINGLE_TOKEN_DECODE = 1
+    """Cudagraph supported for batches the only contain query_len==1 decodes"""
+    NEVER = 0
+    """NO cudagraph support"""
+
+
+class AttentionMetadataBuilder(ABC, Generic[M]):
+    # Does this backend/builder support CUDA Graphs for attention (default: no).
+    # Do not access directly. Call get_cudagraph_support() instead.
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.NEVER
+    # Does this backend/builder reorder the batch?
+    # If not, set this to None. Otherwise set it to the query
+    # length that will be pulled into the front of the batch.
+    reorder_batch_threshold: int | None = None
+    # Does this backend/builder support updating the block table in existing
+    # metadata
+    supports_update_block_table: bool = False
+
+    @abstractmethod
+    def __init__(
+        self,
+        kv_cache_spec: "AttentionSpec",
+        layer_names: list[str],
+        vllm_config: "VllmConfig",
+        device: torch.device,
+    ):
+        self.kv_cache_spec = kv_cache_spec
+        self.layer_names = layer_names
+        self.vllm_config = vllm_config
+        self.device = device
+
+    @classmethod
+    def get_cudagraph_support(
+        cls: type["AttentionMetadataBuilder"],
+        vllm_config: "VllmConfig",
+        kv_cache_spec: "AttentionSpec",
+    ) -> AttentionCGSupport:
+        """Get the cudagraph support level of this builder class."""
+        return cls._cudagraph_support
+
+    def _init_reorder_batch_threshold(
+        self,
+        reorder_batch_threshold: int | None = 1,
+        supports_spec_as_decode: bool = False,
+        supports_dcp_with_varlen: bool = False,
+    ) -> None:
+        self.reorder_batch_threshold = reorder_batch_threshold
+        if self.reorder_batch_threshold is not None and supports_spec_as_decode:
+            # If the backend supports spec-as-decode kernels, then we can set
+            # the reorder_batch_threshold based on the number of speculative
+            # tokens from the config.
+            speculative_config = self.vllm_config.speculative_config
+            if (
+                speculative_config is not None
+                and speculative_config.num_speculative_tokens is not None
+            ):
+                max_num_queries_for_spec = (
+                    1
+                    + (2 if speculative_config.parallel_drafting else 1)
+                    * speculative_config.num_speculative_tokens
+                )
+                self.reorder_batch_threshold = max(
+                    self.reorder_batch_threshold,
+                    max_num_queries_for_spec,
+                )
+
+        if (
+            self.vllm_config.parallel_config.decode_context_parallel_size > 1
+            and not supports_dcp_with_varlen
+        ):
+            self.reorder_batch_threshold = 1
+
+    @abstractmethod
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> M:
+        """
+        Central method that builds attention metadata.
+        Some builders (MLA) require reorder_batch to be called prior to build.
+
+        Args:
+            common_prefix_len: The length of the common prefix of the batch.
+            common_attn_metadata: The common attention metadata.
+            fast_build: The meta-data will prioritize speed of building over
+                then speed at execution. Can be used for spec-decode where the
+                result of a build call may only be used for few layers/iters.
+        """
+        raise NotImplementedError
+
+    def update_block_table(
+        self,
+        metadata: M,
+        blk_table: torch.Tensor,
+        slot_mapping: torch.Tensor,
+    ) -> M:
+        """
+        Update the block table for the attention metadata.
+        Faster when theres multiple kv-cache groups that create virtually the
+        same metadata but just with different block tables.
+
+        Only needs to be implemented if supports_update_block_table is True.
+        """
+        raise NotImplementedError
+
+    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 build_for_drafting(
+        self,
+        common_attn_metadata: CommonAttentionMetadata,
+        draft_index: int,
+    ) -> M:
+        """
+        Build attention metadata for draft model. Uses build by default.
+
+        Args:
+            common_attn_metadata: The common attention metadata.
+            draft_index: The index of the current draft operation.
+                When speculating a chain of tokens, this index refers to the
+                draft attempt for the i-th token.
+                For tree-based attention, this index instead refers to the
+                draft attempt for the i-th level in the tree of tokens.
+        """
+        return self.build(
+            common_prefix_len=0,
+            common_attn_metadata=common_attn_metadata,
+            fast_build=True,
+        )
+
+    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,
+        use_local_attention: bool,
+        num_sms: int,
+        dcp_world_size: int,
+    ) -> bool:
+        return False
+
+
+class AttentionLayer(Protocol):
+    _q_scale: torch.Tensor
+    _k_scale: torch.Tensor
+    _v_scale: torch.Tensor
+    _q_scale_float: float
+    _k_scale_float: float
+    _v_scale_float: float
+    _prob_scale: torch.Tensor
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor: ...
+
+
+class AttentionImplBase(ABC, Generic[T]):
+    """Base class for attention implementations.
+
+    Contains common attributes and initialization logic shared by both
+    standard AttentionImpl and MLAAttentionImpl. Does not define a forward
+    method - subclasses define their own forward interfaces.
+    """
+
+    # Required attributes that all impls should have
+    num_heads: int
+    head_size: int
+    scale: float
+
+    # Whether the attention impl can return the softmax lse for decode.
+    # Some features like decode context parallelism require the softmax lse.
+    can_return_lse_for_decode: bool = False
+
+    # Whether the attention impl supports Prefill Context Parallelism.
+    supports_pcp: bool = False
+    # Whether the attention impl(or ops) supports MTP
+    # when cp_kv_cache_interleave_size > 1
+    supports_mtp_with_cp_non_trivial_interleave_size: bool = False
+
+    # some attention backends might not always want to return lse
+    # even if they can return lse (for efficiency reasons)
+    need_to_return_lse_for_decode: bool = False
+
+    # Whether this attention implementation supports pre-quantized query input.
+    # When True, the attention layer will quantize queries before passing them
+    # to this backend, allowing torch.compile to fuse the quantization with
+    # previous operations. This is typically supported when using FP8 KV cache
+    # with compatible attention kernels (e.g., TRT-LLM).
+    # Subclasses should set this in __init__.
+    # TODO add support to more backends:
+    # https://github.com/vllm-project/vllm/issues/25584
+    supports_quant_query_input: bool = False
+
+    dcp_world_size: int
+    dcp_rank: int
+
+    pcp_world_size: int
+    pcp_rank: int
+
+    total_cp_world_size: int
+    total_cp_rank: int
+
+    def __new__(cls, *args, **kwargs):
+        # use __new__ so that all subclasses will call this
+        self = super().__new__(cls)
+        try:
+            from vllm.distributed.parallel_state import get_dcp_group
+
+            self.dcp_world_size = get_dcp_group().world_size
+            self.dcp_rank = get_dcp_group().rank_in_group
+        except AssertionError:
+            # DCP might not be initialized in testing
+            self.dcp_world_size = 1
+            self.dcp_rank = 0
+        try:
+            from vllm.distributed.parallel_state import get_pcp_group
+
+            self.pcp_world_size = get_pcp_group().world_size
+            self.pcp_rank = get_pcp_group().rank_in_group
+        except AssertionError:
+            self.pcp_world_size = 1
+            self.pcp_rank = 0
+        self.total_cp_world_size = self.pcp_world_size * self.dcp_world_size
+        self.total_cp_rank = self.pcp_rank * self.dcp_world_size + self.dcp_rank
+
+        self.need_to_return_lse_for_decode = (
+            self.dcp_world_size > 1 and self.can_return_lse_for_decode
+        )
+        return self
+
+    def process_weights_after_loading(self, act_dtype: torch.dtype):
+        pass
+
+
+class AttentionImpl(AttentionImplBase[T], Generic[T]):
+    """Standard attention implementation with forward method."""
+
+    @abstractmethod
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int | None = None,
+        alibi_slopes: list[float] | None = None,
+        sliding_window: int | None = None,
+        kv_cache_dtype: str = "auto",
+        logits_soft_cap: float | None = None,
+        attn_type: str = AttentionType.DECODER,
+        kv_sharing_target_layer_name: str | None = None,
+    ) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: T,
+        output: torch.Tensor | None = None,
+        output_scale: torch.Tensor | None = None,
+        output_block_scale: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        raise NotImplementedError
+
+    def fused_output_quant_supported(self, quant_key: "QuantKey"):
+        """
+        Does this attention implementation support fused output quantization.
+        This is used by the AttnFusionPass to only fuse output quantization
+        onto implementations that support it.
+
+        :param quant_key: QuantKey object that describes the quantization op
+        :return: is fusion supported for this type of quantization
+        """
+        return False
+
+    def fused_rope_kvcache_supported(self):
+        """
+        Does this attention implementation support RoPE+KVCache fusion.
+        This is used by the RopeKVCacheFusionPass to only fuse the RoPE ops
+        with the KV cache update for implementations that support it.
+        """
+        return False
+
+    def do_rope_and_kv_cache_update(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        positions: torch.Tensor,
+        cos_sin_cache: torch.Tensor,
+        is_neox: bool,
+        kv_cache: torch.Tensor,
+        layer_slot_mapping: torch.Tensor,
+    ):
+        """
+        If `fused_rope_kvcache_supported` returns True, this method will be called
+        by torch.ops.vllm.fused_rope_and_unified_kv_cache_update
+        to perform the inplace RoPE and KV cache update.
+        """
+        raise NotImplementedError
+
+
+class MLAAttentionImpl(AttentionImplBase[T], Generic[T]):
+    """MLA attention implementation with forward_mqa and forward_mha methods."""
+
+    @abstractmethod
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        q_lora_rank: int | None,
+        kv_lora_rank: int,
+        qk_nope_head_dim: int,
+        qk_rope_head_dim: int,
+        qk_head_dim: int,
+        v_head_dim: int,
+        kv_b_proj: "ColumnParallelLinear",
+        indexer: object | None = None,
+        q_pad_num_heads: int | None = None,
+    ) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def forward_mha(
+        self,
+        q: torch.Tensor,
+        kv_c_normed: torch.Tensor,
+        k_pe: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: T,
+        k_scale: torch.Tensor,
+        output: torch.Tensor,
+    ) -> None:
+        """MHA-style prefill forward pass."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: T,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        """MQA-style decode forward pass."""
+        raise NotImplementedError
+
+
+class SparseMLAAttentionImpl(AttentionImplBase[T], Generic[T]):
+    """Sparse MLA attention implementation with only forward_mqa method.
+
+    Sparse MLA implementations only support decode (MQA-style) attention.
+    They do not support prefill (MHA-style) attention.
+    """
+
+    @abstractmethod
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        q_lora_rank: int | None,
+        kv_lora_rank: int,
+        qk_nope_head_dim: int,
+        qk_rope_head_dim: int,
+        qk_head_dim: int,
+        v_head_dim: int,
+        kv_b_proj: "ColumnParallelLinear",
+        indexer: object | None = None,
+        q_pad_num_heads: int | None = None,
+    ) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: T,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        """MQA-style decode forward pass."""
+        raise NotImplementedError
+
+
+def is_quantized_kv_cache(kv_cache_dtype: str) -> bool:
+    return kv_cache_dtype.startswith("fp8")
+
+
+def subclass_attention_backend(
+    name_prefix: str,
+    attention_backend_cls: type[AttentionBackend],
+    builder_cls: type[AttentionMetadataBuilder[M]],
+) -> type[AttentionBackend]:
+    """
+    Return a new subclass where `get_builder_cls` returns `builder_cls`.
+    """
+    name: str = name_prefix + attention_backend_cls.__name__  # type: ignore
+
+    return type(
+        name, (attention_backend_cls,), {"get_builder_cls": lambda: builder_cls}
+    )
+
+
+def subclass_attention_backend_with_overrides(
+    name_prefix: str,
+    attention_backend_cls: type[AttentionBackend],
+    overrides: dict[str, Any],
+) -> type[AttentionBackend]:
+    name: str = name_prefix + attention_backend_cls.__name__  # type: ignore
+    return type(name, (attention_backend_cls,), overrides)

vllm/v1/attention/backends/cpu_attn.py

@@ -0,0 +1,503 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.platforms import CpuArchEnum, current_platform
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionLayer,
+    AttentionMetadataBuilder,
+    AttentionType,
+    CommonAttentionMetadata,
+    is_quantized_kv_cache,
+)
+from vllm.v1.attention.backends.utils import (
+    split_decodes_and_prefills,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec, CrossAttentionSpec
+
+logger = init_logger(__name__)
+
+_CPU_ARCH_PREFER_MIXED_BATCH = (CpuArchEnum.X86, CpuArchEnum.ARM, CpuArchEnum.S390X)
+
+
+class CPUAttentionBackend(AttentionBackend):
+    accept_output_buffer: bool = True
+    supported_dtypes: ClassVar[list[torch.dtype]] = [
+        torch.float16,
+        torch.bfloat16,
+        torch.float32,
+    ]
+
+    @classmethod
+    def get_supported_dtypes(cls) -> list[torch.dtype]:
+        return [torch.float16, torch.bfloat16, torch.float32]
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return [32, 64, 80, 96, 112, 128, 160, 192, 224, 256]
+
+    @staticmethod
+    def get_name() -> str:
+        return "CPU_ATTN"
+
+    @classmethod
+    def supports_attn_type(cls, attn_type: str) -> bool:
+        """CPU attention supports decoder,
+        encoder-only and encoder-decoder attention."""
+        return attn_type in (
+            AttentionType.DECODER,
+            AttentionType.ENCODER,
+            AttentionType.ENCODER_ONLY,
+            AttentionType.ENCODER_DECODER,
+        )
+
+    @staticmethod
+    def get_impl_cls() -> type["CPUAttentionBackendImpl"]:
+        return CPUAttentionBackendImpl
+
+    @staticmethod
+    def get_builder_cls() -> type["CPUAttentionMetadataBuilder"]:
+        return CPUAttentionMetadataBuilder
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        return 2, num_blocks, num_kv_heads, block_size, head_size
+
+    @staticmethod
+    def use_cascade_attention(*args, **kwargs) -> bool:
+        return False
+
+
+@dataclass
+class CPUAttentionMetadata:
+    isa: str
+    num_actual_tokens: int  # Number of tokens excluding padding.
+    max_query_len: int
+    query_start_loc: torch.Tensor
+    max_seq_len: int
+    seq_lens: torch.Tensor
+    block_table: torch.Tensor
+    slot_mapping: torch.Tensor
+    scheduler_metadata: torch.Tensor | None
+    causal: bool = True
+
+    # can be removed after deprecate sdpa
+    use_sdpa_prefill: bool = False
+    num_decode_tokens: int = 0
+    sdpa_attn_masks: list[torch.Tensor | None] | None = None
+    sdpa_start_loc: torch.Tensor | None = None
+
+
+class CPUAttentionMetadataBuilder(AttentionMetadataBuilder[CPUAttentionMetadata]):
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ) -> None:
+        super().__init__(kv_cache_spec, layer_names, vllm_config, device)
+
+        self.use_sdpa_prefill = False
+        reorder_batch_threshold = None
+        if current_platform.get_cpu_architecture() not in _CPU_ARCH_PREFER_MIXED_BATCH:
+            # in this case, decode seqs are reordered to the front of prefill seqs
+            # to split decode and prefill. Then use SDPA for prefill and
+            # cpu_attention_with_kv_cache for decode
+            reorder_batch_threshold = 1
+            self.use_sdpa_prefill = True
+
+        self._init_reorder_batch_threshold(reorder_batch_threshold, False)
+
+        self.kv_cache_spec = kv_cache_spec
+        self.vllm_config = vllm_config
+
+        parallel_config = vllm_config.parallel_config
+        self.num_kv_heads = vllm_config.model_config.get_num_kv_heads(parallel_config)
+        self.num_heads = vllm_config.model_config.get_num_attention_heads(
+            parallel_config
+        )
+        self.head_dim = kv_cache_spec.head_size
+        self.dtype = vllm_config.model_config.dtype
+        self.window_size = getattr(kv_cache_spec, "sliding_window", -1)
+        if self.window_size is None:
+            self.window_size = -1
+        self.block_size = vllm_config.cache_config.block_size
+        self.isa = _get_attn_isa(self.dtype, self.block_size, self.head_dim)
+        self.is_cross_attention = isinstance(kv_cache_spec, CrossAttentionSpec)
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> CPUAttentionMetadata:
+        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 = common_attn_metadata.max_seq_len
+        query_start_loc = common_attn_metadata.query_start_loc
+        seq_lens = common_attn_metadata.seq_lens
+        block_table_tensor = common_attn_metadata.block_table_tensor
+        slot_mapping = common_attn_metadata.slot_mapping
+        causal = False if self.is_cross_attention else common_attn_metadata.causal
+
+        sdpa_start_loc = query_start_loc
+        num_decode_tokens = 0
+        if self.use_sdpa_prefill and causal:
+            # Decoder, need reorder and truncate
+            assert self.reorder_batch_threshold
+            (num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens) = (
+                split_decodes_and_prefills(
+                    common_attn_metadata,
+                    decode_threshold=self.reorder_batch_threshold,
+                    require_uniform=True,
+                )
+            )
+            num_reqs = num_decodes
+            sdpa_start_loc = sdpa_start_loc[num_decodes:] - num_decode_tokens
+            seq_lens = seq_lens[:num_decodes]
+            query_start_loc = query_start_loc[: num_decodes + 1]
+            block_table_tensor = block_table_tensor[:num_decodes]
+
+        sheduler_metadata = ops.cpu_attn_get_scheduler_metadata(
+            num_reqs=num_reqs,
+            num_heads=self.num_heads,
+            num_kv_heads=self.num_kv_heads,
+            head_dim=self.head_dim,
+            seq_lens=seq_lens,
+            dtype=self.dtype,
+            query_start_loc=query_start_loc,
+            causal=causal,
+            sliding_window_size=self.window_size,
+            isa=self.isa,
+            enable_kv_split=True,
+        )
+
+        attn_metadata = CPUAttentionMetadata(
+            isa=self.isa,
+            num_actual_tokens=num_actual_tokens,
+            max_query_len=max_query_len,
+            query_start_loc=query_start_loc,
+            max_seq_len=max_seq_len,
+            seq_lens=seq_lens,
+            block_table=block_table_tensor,
+            slot_mapping=slot_mapping,
+            scheduler_metadata=sheduler_metadata,
+            causal=causal,
+            use_sdpa_prefill=self.use_sdpa_prefill,
+            num_decode_tokens=num_decode_tokens,
+            sdpa_start_loc=sdpa_start_loc,
+        )
+
+        return attn_metadata
+
+
+class CPUAttentionBackendImpl(AttentionImpl):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None = None,
+        attn_type: str = AttentionType.DECODER,
+        kv_sharing_target_layer_name: str | None = None,
+        sinks: torch.Tensor | None = None,
+    ) -> None:
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        if logits_soft_cap is not None and attn_type in (
+            AttentionType.ENCODER,
+            AttentionType.ENCODER_ONLY,
+        ):
+            logger.warning_once(
+                "CPU_ATTN does not support logits softcap for"
+                " ENCODER and ENCODER_ONLY, outputs may be slightly off"
+            )
+        if logits_soft_cap is None:
+            logits_soft_cap = 0
+        self.logits_soft_cap = logits_soft_cap
+
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        if sliding_window is None:
+            self.sliding_window = (-1, -1)
+        elif attn_type == AttentionType.ENCODER_ONLY:
+            self.sliding_window = (sliding_window - 1, sliding_window - 1)
+        else:
+            self.sliding_window = (sliding_window - 1, 0)
+        self.kv_cache_dtype = kv_cache_dtype
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        if is_quantized_kv_cache(kv_cache_dtype):
+            raise NotImplementedError("FP8 KV cache is unsupported in CPU_ATTN")
+        self.attn_type = attn_type
+
+        self.sinks = sinks
+        if self.sinks is not None:
+            assert self.sinks.shape[0] == num_heads, (
+                "Sinks must have the same number of heads as the number of "
+                "heads in the layer"
+            )
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: CPUAttentionMetadata | None,
+        output: torch.Tensor | None = None,
+        output_scale: torch.Tensor | None = None,
+        output_block_scale: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        """Forward pass for CPU attention backend.
+
+        Args:
+            query: shape = [num_tokens, num_heads, head_size]
+            key: shape = [num_tokens, num_kv_heads, head_size]
+            value: shape = [num_tokens, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, num_kv_heads, block_size, head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        assert output is not None, "Output tensor must be provided."
+        if output_scale is not None or output_block_scale is not None:
+            raise NotImplementedError(
+                "fused output quantization is not yet supported"
+                " for CPUAttentionBackendImpl"
+            )
+
+        # For warming-up
+        if attn_metadata is None:
+            return output
+
+        num_actual_tokens = attn_metadata.num_actual_tokens
+
+        # Handle encoder attention differently - no KV cache needed
+        if self.attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER):
+            # For encoder attention,
+            return self._run_sdpa_forward(
+                query[:num_actual_tokens],
+                key[:num_actual_tokens],
+                value[:num_actual_tokens],
+                output[:num_actual_tokens],
+                attn_metadata,
+                self.attn_type,
+            )
+
+        # For decoder and cross-attention, use KV cache, size are
+        # [num_blocks, num_kv_heads, block_size, head_size]
+        key_cache, value_cache = kv_cache.unbind(0)
+
+        # key and value may be None in the case of cross attention. They are
+        # calculated once based on the output from the encoder and then cached
+        # in KV cache.
+        if (
+            self.kv_sharing_target_layer_name is None
+            and key is not None
+            and value is not None
+        ):
+            ops.cpu_attn_reshape_and_cache(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping,
+                attn_metadata.isa,
+            )
+
+        if attn_metadata.use_sdpa_prefill:
+            assert self.sinks is None, "Attention sink is unsupported in SDPA prefill"
+            num_decode_tokens = attn_metadata.num_decode_tokens
+            self._run_sdpa_forward(
+                query[num_decode_tokens:num_actual_tokens],
+                key[num_decode_tokens:num_actual_tokens],
+                value[num_decode_tokens:num_actual_tokens],
+                output[num_decode_tokens:num_actual_tokens],
+                attn_metadata,
+                self.attn_type,
+            )
+            num_actual_tokens = num_decode_tokens
+
+        if num_actual_tokens > 0:
+            ops.cpu_attention_with_kv_cache(
+                query=query[:num_actual_tokens],
+                key_cache=key_cache,
+                value_cache=value_cache,
+                output=output[:num_actual_tokens],  # type: ignore
+                query_start_loc=attn_metadata.query_start_loc,
+                seq_lens=attn_metadata.seq_lens,
+                scale=self.scale,
+                causal=attn_metadata.causal,
+                alibi_slopes=self.alibi_slopes,  # type: ignore
+                sliding_window=self.sliding_window,
+                block_table=attn_metadata.block_table,
+                softcap=self.logits_soft_cap,
+                scheduler_metadata=attn_metadata.scheduler_metadata,
+                s_aux=self.sinks,
+            )
+
+        return output
+
+    def _run_sdpa_forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        output: torch.Tensor,
+        attn_metadata: CPUAttentionMetadata,
+        attn_type: str,
+    ) -> torch.Tensor:
+        attn_masks = attn_metadata.sdpa_attn_masks
+        if attn_masks is None:
+            if self.alibi_slopes is not None:
+                attn_masks = _make_alibi_bias(
+                    self.alibi_slopes,
+                    query.dtype,
+                    attn_metadata.sdpa_start_loc,
+                )
+            elif self.sliding_window[0] != -1 or self.sliding_window[1] != -1:
+                assert attn_metadata.seq_lens is not None
+                attn_masks = _make_sliding_window_bias(
+                    attn_metadata.sdpa_start_loc,
+                    self.sliding_window[0],
+                    self.sliding_window[1],
+                    query.dtype,
+                )
+            else:
+                attn_masks = [None] * (attn_metadata.sdpa_start_loc.size(0) - 1)  # type: ignore
+            attn_metadata.sdpa_attn_masks = attn_masks
+
+        query = query.movedim(0, query.dim() - 2)
+        key = key.movedim(0, key.dim() - 2)
+        value = value.movedim(0, value.dim() - 2)
+
+        causal_attn = attn_type == AttentionType.DECODER
+
+        sdpa_start_loc = attn_metadata.sdpa_start_loc.numpy()  # type: ignore
+        for i in range(len(attn_masks)):
+            mask = attn_masks[i]
+            start_q = sdpa_start_loc[i]
+            end_q = sdpa_start_loc[i + 1]
+            sub_out = (
+                torch.nn.functional.scaled_dot_product_attention(
+                    query[None, :, start_q:end_q, :],
+                    key[None, :, start_q:end_q, :],
+                    value[None, :, start_q:end_q, :],
+                    attn_mask=mask,
+                    dropout_p=0.0,
+                    is_causal=causal_attn and mask is None,
+                    scale=self.scale,
+                    enable_gqa=self.num_heads > self.num_kv_heads,
+                )
+                .squeeze(0)
+                .movedim(query.dim() - 2, 0)
+            )
+            output[start_q:end_q, :, :] = sub_out
+        return output
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    dtype: torch.dtype,
+    sdpa_start_loc: torch.Tensor,
+) -> list[torch.Tensor]:
+    attn_biases: list[torch.Tensor] = []
+    seq_num = sdpa_start_loc.size(0) - 1
+    sdpa_start_loc = sdpa_start_loc.numpy()  # type: ignore
+    for i in range(seq_num):
+        seq_len = sdpa_start_loc[i + 1] - sdpa_start_loc[i]
+        bias = torch.arange(seq_len, dtype=dtype)  # type: ignore
+        # NOTE(zhuohan): HF uses
+        #     `bias = bias[None, :].repeat(seq_len, 1)`
+        # here. We find that both biases give the same results, but
+        # the bias below more accurately follows the original ALiBi
+        # paper.
+        bias = bias[None, :] - bias[:, None]
+
+        num_heads = alibi_slopes.shape[0]
+        bias = bias[None, :].repeat((num_heads, 1, 1))
+        bias.mul_(alibi_slopes[:, None, None]).unsqueeze_(0)
+        inf_mask = (
+            torch.empty((1, seq_len, seq_len), dtype=bias.dtype)  # type: ignore
+            .fill_(-torch.inf)
+            .triu_(diagonal=1)
+        )
+        attn_biases.append((bias + inf_mask).to(dtype))
+
+    return attn_biases
+
+
+def _make_sliding_window_bias(
+    sdpa_start_loc: torch.Tensor,
+    left_window_size: int,
+    right_window_size: int,
+    dtype: torch.dtype,
+) -> list[torch.Tensor]:
+    attn_biases: list[torch.Tensor] = []
+    seq_num = sdpa_start_loc.size(0) - 1
+    sdpa_start_loc = sdpa_start_loc.numpy()  # type: ignore
+    for i in range(seq_num):
+        seq_len = sdpa_start_loc[i + 1] - sdpa_start_loc[i]
+        mask = torch.full(  # type: ignore
+            (1, seq_len, seq_len),  # type: ignore
+            fill_value=1,
+            dtype=dtype,
+        )
+
+        if right_window_size != -1:
+            mask = torch.tril(mask, diagonal=right_window_size)
+        if left_window_size != -1:
+            mask = torch.triu(mask, diagonal=-left_window_size)
+        mask = torch.log(mask)
+        attn_biases.append(mask)
+
+    return attn_biases
+
+
+def _get_attn_isa(
+    dtype: torch.dtype, block_size: int, head_size: int | None = None
+) -> str:
+    if head_size is not None and head_size % 32 != 0 and head_size % 16 == 0:
+        return "vec16"
+    supports_amx = torch._C._cpu._is_amx_tile_supported()
+    supports_arm = current_platform.get_cpu_architecture() == CpuArchEnum.ARM
+    supports_vxe = current_platform.get_cpu_architecture() == CpuArchEnum.S390X
+    if supports_amx and dtype in (torch.bfloat16,) and block_size % 32 == 0:
+        return "amx"
+    elif block_size % 32 == 0:
+        if supports_arm:
+            # support ARM NEON FMLA and BFMMLA (bf16) for block size 32
+            return "neon"
+        elif supports_vxe:
+            return "vxe"
+        else:
+            return "vec"
+    else:
+        return "vec16"

vllm/v1/attention/backends/fa_utils.py

@@ -0,0 +1,177 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Any
+
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+
+logger = init_logger(__name__)
+
+# Track whether upstream flash-attn is available on ROCm.
+# Set during module initialization and never modified afterwards.
+# This module-level flag avoids repeated import attempts and ensures
+# consistent behavior (similar to IS_AITER_FOUND in _aiter_ops.py).
+_ROCM_FLASH_ATTN_AVAILABLE = False
+
+if current_platform.is_cuda():
+    from vllm._custom_ops import reshape_and_cache_flash
+    # from vllm.vllm_flash_attn import (  # type: ignore[attr-defined]
+    #     flash_attn_varlen_func,
+    #     get_scheduler_metadata,
+    # )
+    from ixformer.contrib.vllm_flash_attn import flash_attn_varlen_func, flash_attn_with_kvcache
+
+elif current_platform.is_xpu():
+    from vllm import _custom_ops as ops
+    from vllm._xpu_ops import xpu_ops
+
+    reshape_and_cache_flash = ops.reshape_and_cache_flash
+    flash_attn_varlen_func = xpu_ops.flash_attn_varlen_func  # type: ignore[assignment]
+    get_scheduler_metadata = xpu_ops.get_scheduler_metadata  # type: ignore[assignment]
+elif current_platform.is_rocm():
+    try:
+        from flash_attn import flash_attn_varlen_func  # type: ignore[no-redef]
+
+        # Mark that upstream flash-attn is available on ROCm
+        _ROCM_FLASH_ATTN_AVAILABLE = True
+    except ImportError:
+
+        def flash_attn_varlen_func(*args: Any, **kwargs: Any) -> Any:  # type: ignore[no-redef,misc]
+            raise ImportError(
+                "ROCm platform requires upstream flash-attn "
+                "to be installed. Please install flash-attn first."
+            )
+
+    # ROCm doesn't use scheduler metadata (FA3 feature), provide stub
+    def get_scheduler_metadata(*args: Any, **kwargs: Any) -> None:  # type: ignore[misc]
+        return None
+
+    # ROCm uses the C++ custom op for reshape_and_cache
+    from vllm import _custom_ops as ops
+
+    reshape_and_cache_flash = ops.reshape_and_cache_flash
+
+
+def get_flash_attn_version(requires_alibi: bool = False) -> int | None:
+    # import here to avoid circular dependencies
+    from vllm.platforms import current_platform
+
+    return 3
+
+    if current_platform.is_xpu():
+        return 2
+    if current_platform.is_rocm():
+        # ROCm doesn't use vllm_flash_attn; return None to skip fa_version arg
+        return None
+    try:
+        from vllm.vllm_flash_attn.flash_attn_interface import (
+            fa_version_unsupported_reason,
+            is_fa_version_supported,
+        )
+
+        device_capability = current_platform.get_device_capability()
+
+        assert device_capability is not None
+
+        # 1. default version depending on platform
+        fa_version = (
+            3 if (device_capability.major == 9 and is_fa_version_supported(3)) else 2
+        )
+
+        # 2. override if passed by environment or config
+        from vllm.config import get_current_vllm_config_or_none
+
+        vllm_config = get_current_vllm_config_or_none()
+        if (
+            vllm_config is not None
+            and vllm_config.attention_config.flash_attn_version is not None
+        ):
+            fa_version = vllm_config.attention_config.flash_attn_version
+
+        # 3. fallback for unsupported combinations
+        if device_capability.major == 10 and fa_version == 3:
+            logger.warning_once(
+                "Cannot use FA version 3 on Blackwell platform, "
+                "defaulting to FA version 2."
+            )
+            fa_version = 2
+
+        if requires_alibi and fa_version == 3:
+            logger.warning_once(
+                "Cannot use FA version 3 with ALiBi, defaulting to FA version 2."
+            )
+            fa_version = 2
+
+        if not is_fa_version_supported(fa_version):
+            logger.error(
+                "Cannot use FA version %d is not supported due to %s",
+                fa_version,
+                fa_version_unsupported_reason(fa_version),
+            )
+
+        assert is_fa_version_supported(fa_version)
+        return fa_version
+    except (ImportError, AssertionError):
+        return None
+
+
+def flash_attn_supports_fp8() -> bool:
+    return (
+        get_flash_attn_version() == 3
+        and current_platform.is_device_capability_family(90)
+    )
+
+
+def flash_attn_supports_sinks() -> bool:
+    if current_platform.is_xpu():
+        return True
+    else:
+        return get_flash_attn_version() == 3
+
+
+def flash_attn_supports_mla():
+    from vllm.platforms import current_platform
+
+    if current_platform.is_cuda():
+        try:
+            from vllm.vllm_flash_attn.flash_attn_interface import (
+                is_fa_version_supported,
+            )
+
+            return is_fa_version_supported(
+                3
+            ) and current_platform.is_device_capability_family(90)
+        except (ImportError, AssertionError):
+            pass
+    return False
+
+
+def is_flash_attn_varlen_func_available() -> bool:
+    """Check if flash_attn_varlen_func is available.
+
+    This function determines whether the flash_attn_varlen_func imported at module
+    level is a working implementation or a stub.
+
+    Platform-specific sources:
+    - CUDA: vllm.vllm_flash_attn.flash_attn_varlen_func
+    - XPU: xpu_ops.flash_attn_varlen_func
+    - ROCm: upstream flash_attn.flash_attn_varlen_func (if available)
+
+    Note: This is separate from the AITER flash attention backend (rocm_aiter_fa.py)
+    which uses rocm_aiter_ops.flash_attn_varlen_func. The condition to use AITER is
+    handled separately via _aiter_ops.is_aiter_found_and_supported().
+
+    Returns:
+        bool: True if a working flash_attn_varlen_func implementation is available.
+    """
+    if current_platform.is_cuda() or current_platform.is_xpu():
+        # CUDA and XPU always have flash_attn_varlen_func available
+        return True
+
+    if current_platform.is_rocm():
+        # Use the flag set during module import to check if
+        # upstream flash-attn was successfully imported
+        return _ROCM_FLASH_ATTN_AVAILABLE
+
+    return False

vllm/v1/attention/backends/flash_attn_diffkv.py

@@ -0,0 +1,277 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Attention layer with FlashAttention."""
+
+import torch
+
+from vllm.v1.attention.backend import AttentionType
+from vllm.v1.attention.backends.fa_utils import is_flash_attn_varlen_func_available
+from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
+    triton_reshape_and_cache_flash_diffkv,
+)
+
+if is_flash_attn_varlen_func_available():
+    from vllm.v1.attention.backends.fa_utils import flash_attn_varlen_func
+from vllm.logger import init_logger
+from vllm.v1.attention.backends.utils import get_kv_cache_layout
+
+from .flash_attn import (
+    FlashAttentionBackend,
+    FlashAttentionImpl,
+    FlashAttentionMetadata,
+    cascade_attention,
+)
+
+logger = init_logger(__name__)
+
+
+class FlashAttentionDiffKVBackend(FlashAttentionBackend):
+    # Default to 128 for this backend
+    head_size_v: int = 128
+
+    @classmethod
+    def set_head_size_v(cls, head_size_v: int) -> None:
+        cls.head_size_v = head_size_v
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASH_ATTN_DIFFKV"
+
+    @staticmethod
+    def get_impl_cls() -> type["FlashAttentionImpl"]:
+        return FlashAttentionDiffKVImpl
+
+    # Do not modify the interface of get_kv_cache_shape,
+    # but consider head_size_v when returning result.
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        if block_size % 16 != 0:
+            raise ValueError("Block size must be a multiple of 16.")
+        return (
+            num_blocks,
+            block_size,
+            num_kv_heads,
+            head_size + FlashAttentionDiffKVBackend.head_size_v,
+        )
+
+    @staticmethod
+    def get_kv_cache_stride_order(
+        include_num_layers_dimension: bool = False,
+    ) -> tuple[int, ...]:
+        # `stride_order` indicates the permutation that gets
+        # us from `get_kv_cache_shape` to the actual memory layout we want.
+        cache_layout = get_kv_cache_layout()
+        if cache_layout == "NHD" and include_num_layers_dimension:
+            # (num_blocks, num_layers, block_size,
+            # num_kv_heads, head_size + head_size_v)
+            return (1, 0, 2, 3, 4)
+        elif cache_layout == "NHD":
+            stride_order = (0, 1, 2, 3)
+        elif cache_layout == "HND" and include_num_layers_dimension:
+            # (num_blocks, num_kv_heads, num_layers,
+            # block_size, head_size + head_size_v)
+            return (1, 3, 0, 2, 4)
+        elif cache_layout == "HND":
+            stride_order = (0, 2, 1, 3)
+        else:
+            raise ValueError(f"Unknown cache layout format {cache_layout}.")
+        return stride_order
+
+
+class FlashAttentionDiffKVImpl(FlashAttentionImpl):
+    def forward(
+        self,
+        layer: torch.nn.Module,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: FlashAttentionMetadata,
+        output: torch.Tensor | None = None,
+        output_scale: torch.Tensor | None = None,
+        output_block_scale: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads, head_size]
+            key: shape = [num_tokens, num_kv_heads, head_size]
+            value: shape = [num_tokens, num_kv_heads, head_size_v]
+            kv_cache: shape =
+                [num_blocks, block_size, num_kv_heads, head_size + head_size_v]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size_v]
+        NOTE: FP8 quantization, flash-attn expect the size of
+              {q,k,v}_descale to be (num_sequences, num_kv_heads).
+              We use torch's .expand() to avoid duplicating values
+        """
+        assert output is not None, "Output tensor must be provided."
+        assert self.vllm_flash_attn_version is not None, (
+            "FlashAttention version not detected."
+        )
+
+        if output_scale is not None or output_block_scale is not None:
+            raise NotImplementedError(
+                "fused output quantization is not yet supported for FlashAttentionImpl"
+            )
+
+        if attn_metadata is None:
+            # Profiling run.
+            return output.fill_(0)
+
+        attn_type = self.attn_type
+
+        # IMPORTANT!
+        # NOTE(woosuk): With piece-wise CUDA graphs, this method is executed in
+        # eager-mode PyTorch. Thus, we need to be careful about any CPU overhead
+        # in this method. For example, `view` and `slice` (or `[:n]`) operations
+        # are surprisingly slow even in the case they do not invoke any GPU ops.
+        # Minimize the PyTorch ops in this method as much as possible.
+        # Whenever making a change in this method, please benchmark the
+        # performance to make sure it does not introduce any overhead.
+
+        num_actual_tokens = attn_metadata.num_actual_tokens
+
+        # Handle encoder attention differently - no KV cache needed
+        if attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER):
+            # For encoder attention,
+            # we use direct Q, K, V tensors without caching
+            return self._forward_encoder_attention(
+                query[:num_actual_tokens],
+                key[:num_actual_tokens],
+                value[:num_actual_tokens],
+                output[:num_actual_tokens],
+                attn_metadata,
+                layer,
+            )
+
+        # For decoder and cross-attention, use KV cache as before
+        # Different head_size for K and V
+        key_cache = kv_cache[..., : self.head_size]
+        value_cache = kv_cache[..., self.head_size :]
+
+        # key and value may be None in the case of cross attention. They are
+        # calculated once based on the output from the encoder and then cached
+        # in KV cache.
+        if (
+            self.kv_sharing_target_layer_name is None
+            and key is not None
+            and value is not None
+        ):
+            # Reshape the input keys and values and store them in the cache.
+            # Skip this if sharing KV cache with an earlier attention layer.
+            # NOTE(woosuk): Here, key and value are padded while slot_mapping is
+            # not padded. However, we don't need to do key[:num_actual_tokens]
+            # and value[:num_actual_tokens] because the reshape_and_cache_flash
+            # op uses the slot_mapping's shape to determine the number of
+            # actual tokens.
+
+            # kv_cache update for different head_size K and V
+            triton_reshape_and_cache_flash_diffkv(
+                key,
+                value,
+                kv_cache,
+                attn_metadata.slot_mapping,
+                self.kv_cache_dtype,
+                layer._k_scale,
+                layer._v_scale,
+            )
+
+        if self.kv_cache_dtype.startswith("fp8"):
+            # queries are quantized in the attention layer
+            dtype = FlashAttentionBackend.get_fp8_dtype_for_flashattn(
+                self.kv_cache_dtype
+            )
+            key_cache = key_cache.view(dtype)
+            value_cache = value_cache.view(dtype)
+
+        if not attn_metadata.use_cascade:
+            cu_seqlens_q = attn_metadata.query_start_loc
+            seqused_k = attn_metadata.seq_lens
+            max_seqlen_q = attn_metadata.max_query_len
+            max_seqlen_k = attn_metadata.max_seq_len
+            block_table = attn_metadata.block_table
+            scheduler_metadata = attn_metadata.scheduler_metadata
+
+            descale_shape = (cu_seqlens_q.shape[0] - 1, self.num_kv_heads)
+
+            if self.dcp_world_size > 1:
+                self._forward_with_dcp(
+                    query[:num_actual_tokens],
+                    key[:num_actual_tokens],
+                    value[:num_actual_tokens],
+                    key_cache,
+                    value_cache,
+                    output[:num_actual_tokens],
+                    attn_metadata,
+                    q_descale=layer._q_scale.expand(descale_shape),
+                    k_descale=layer._k_scale.expand(descale_shape),
+                    v_descale=layer._v_scale.expand(descale_shape),
+                )
+                return output
+            else:
+                sliding_window_size = (
+                    list(self.sliding_window)
+                    if self.sliding_window is not None
+                    else None
+                )
+                flash_attn_varlen_func(
+                    q=query[:num_actual_tokens],
+                    k=key_cache,
+                    v=value_cache,
+                    out=output[:num_actual_tokens],
+                    cu_seqlens_q=cu_seqlens_q,
+                    max_seqlen_q=max_seqlen_q,
+                    seqused_k=seqused_k,
+                    max_seqlen_k=max_seqlen_k,
+                    softmax_scale=self.scale,
+                    causal=attn_metadata.causal,
+                    alibi_slopes=self.alibi_slopes,
+                    window_size=sliding_window_size,
+                    block_table=block_table,
+                    softcap=self.logits_soft_cap,
+                    scheduler_metadata=scheduler_metadata,
+                    fa_version=self.vllm_flash_attn_version,
+                    q_descale=layer._q_scale.expand(descale_shape),
+                    k_descale=layer._k_scale.expand(descale_shape),
+                    v_descale=layer._v_scale.expand(descale_shape),
+                    num_splits=attn_metadata.max_num_splits,
+                    s_aux=self.sinks,
+                )
+                return output
+
+        # Cascade attention (rare case).
+        cascade_attention(
+            output[:num_actual_tokens],
+            query[:num_actual_tokens],
+            key_cache,
+            value_cache,
+            cu_query_lens=attn_metadata.query_start_loc,
+            max_query_len=attn_metadata.max_query_len,
+            cu_prefix_query_lens=attn_metadata.cu_prefix_query_lens,
+            prefix_kv_lens=attn_metadata.prefix_kv_lens,
+            suffix_kv_lens=attn_metadata.suffix_kv_lens,
+            max_kv_len=attn_metadata.max_seq_len,
+            softmax_scale=self.scale,
+            alibi_slopes=self.alibi_slopes,
+            sliding_window=self.sliding_window,
+            logits_soft_cap=self.logits_soft_cap,
+            block_table=attn_metadata.block_table,
+            common_prefix_len=attn_metadata.common_prefix_len,
+            max_num_splits=attn_metadata.max_num_splits,
+            fa_version=self.vllm_flash_attn_version,
+            prefix_scheduler_metadata=attn_metadata.prefix_scheduler_metadata,
+            suffix_scheduler_metadata=attn_metadata.scheduler_metadata,
+            q_descale=layer._q_scale,
+            k_descale=layer._k_scale,
+            v_descale=layer._v_scale,
+            s_aux=self.sinks,
+        )
+        return output

vllm/v1/attention/backends/gdn_attn.py

@@ -0,0 +1,430 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Backend for GatedDeltaNet attention."""
+
+from dataclasses import dataclass
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionMetadataBuilder,
+    CommonAttentionMetadata,
+)
+from vllm.v1.attention.backends.utils import (
+    PAD_SLOT_ID,
+    compute_causal_conv1d_metadata,
+    mamba_get_block_table_tensor,
+    split_decodes_and_prefills,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec, MambaSpec
+
+
+class GDNAttentionBackend(AttentionBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "GDN_ATTN"
+
+    @staticmethod
+    def get_builder_cls() -> type["GDNAttentionMetadataBuilder"]:
+        return GDNAttentionMetadataBuilder
+
+
+@dataclass
+class GDNAttentionMetadata:
+    num_prefills: int
+    num_prefill_tokens: int
+    num_decodes: int
+    num_decode_tokens: int
+    num_spec_decodes: int
+    num_spec_decode_tokens: int
+    num_actual_tokens: int
+
+    has_initial_state: torch.Tensor | None = None
+
+    spec_query_start_loc: torch.Tensor | None = None  # shape: [num_spec_decodes + 1,]
+    non_spec_query_start_loc: torch.Tensor | None = (
+        None  # shape: [batch - num_spec_decodes + 1,]
+    )
+
+    spec_state_indices_tensor: torch.Tensor | None = None  # shape: [batch, num_spec]
+    non_spec_state_indices_tensor: torch.Tensor | None = (
+        None  # shape: [batch - num_spec_decodes,]
+    )
+    spec_sequence_masks: torch.Tensor | None = None  # shape: [batch,]
+    spec_token_indx: torch.Tensor | None = None
+    non_spec_token_indx: torch.Tensor | None = None
+
+    num_accepted_tokens: torch.Tensor | None = None  # shape: [batch,]
+
+    # The following attributes are for triton implementation of causal_conv1d
+    nums_dict: dict | None = None
+    batch_ptr: torch.Tensor | None = None
+    token_chunk_offset_ptr: torch.Tensor | None = None
+
+
+class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]):
+    _cudagraph_support = AttentionCGSupport.UNIFORM_BATCH
+
+    reorder_batch_threshold: int = 1
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        assert isinstance(kv_cache_spec, MambaSpec)
+        self.vllm_config = vllm_config
+        self.compilation_config = vllm_config.compilation_config
+        self.speculative_config = vllm_config.speculative_config
+        self.kv_cache_spec = kv_cache_spec
+
+        if self.speculative_config:
+            assert self.speculative_config.num_speculative_tokens is not None
+            self.num_spec: int = self.speculative_config.num_speculative_tokens
+        else:
+            self.num_spec = 0
+        self.use_spec_decode = self.num_spec > 0
+        self._init_reorder_batch_threshold(1, self.use_spec_decode)
+
+        self.use_full_cuda_graph = (
+            self.compilation_config.cudagraph_mode.has_full_cudagraphs()
+        )
+
+        self.decode_cudagraph_max_bs = (
+            self.vllm_config.scheduler_config.max_num_seqs * (self.num_spec + 1)
+        )
+        if self.compilation_config.max_cudagraph_capture_size is not None:
+            self.decode_cudagraph_max_bs = min(
+                self.decode_cudagraph_max_bs,
+                self.compilation_config.max_cudagraph_capture_size,
+            )
+
+        self.spec_state_indices_tensor = torch.empty(
+            (self.decode_cudagraph_max_bs, self.num_spec + 1),
+            dtype=torch.int32,
+            device=device,
+        )
+        self.non_spec_state_indices_tensor = torch.empty(
+            (self.decode_cudagraph_max_bs,),
+            dtype=torch.int32,
+            device=device,
+        )
+        self.spec_sequence_masks = torch.empty(
+            (self.decode_cudagraph_max_bs,),
+            dtype=torch.bool,
+            device=device,
+        )
+        self.spec_token_indx = torch.empty(
+            (self.decode_cudagraph_max_bs * (self.num_spec + 1),),
+            dtype=torch.int32,
+            device=device,
+        )
+        self.non_spec_token_indx = torch.empty(
+            (self.decode_cudagraph_max_bs * (self.num_spec + 1),),
+            dtype=torch.int32,
+            device=device,
+        )
+        self.spec_query_start_loc = torch.empty(
+            (self.decode_cudagraph_max_bs + 1,),
+            dtype=torch.int32,
+            device=device,
+        )
+        self.non_spec_query_start_loc = torch.empty(
+            (self.decode_cudagraph_max_bs + 1,),
+            dtype=torch.int32,
+            device=device,
+        )
+        self.num_accepted_tokens = torch.empty(
+            (self.decode_cudagraph_max_bs,),
+            dtype=torch.int32,
+            device=device,
+        )
+
+    def build(  # type: ignore[override]
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        num_accepted_tokens: torch.Tensor | None = None,
+        num_decode_draft_tokens_cpu: torch.Tensor | None = None,
+        fast_build: bool = False,
+    ) -> GDNAttentionMetadata:
+        m = common_attn_metadata
+
+        query_start_loc = m.query_start_loc
+        query_start_loc_cpu = m.query_start_loc_cpu
+        context_lens_tensor = m.compute_num_computed_tokens()
+        nums_dict, batch_ptr, token_chunk_offset_ptr = None, None, None
+        block_table_tensor = mamba_get_block_table_tensor(
+            m.block_table_tensor,
+            m.seq_lens,
+            self.kv_cache_spec,
+            self.vllm_config.cache_config.mamba_cache_mode,
+        )
+
+        spec_sequence_masks_cpu: torch.Tensor | None = None
+        if (
+            not self.use_spec_decode
+            or num_decode_draft_tokens_cpu is None
+            or num_decode_draft_tokens_cpu[num_decode_draft_tokens_cpu >= 0]
+            .sum()
+            .item()
+            == 0
+        ):
+            spec_sequence_masks = None
+            num_spec_decodes = 0
+        else:
+            spec_sequence_masks_cpu = num_decode_draft_tokens_cpu >= 0
+            num_spec_decodes = spec_sequence_masks_cpu.sum().item()
+            if num_spec_decodes == 0:
+                spec_sequence_masks = None
+                spec_sequence_masks_cpu = None
+            else:
+                spec_sequence_masks = spec_sequence_masks_cpu.to(
+                    query_start_loc.device, non_blocking=True
+                )
+
+        if spec_sequence_masks is None:
+            num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = (
+                split_decodes_and_prefills(m, decode_threshold=1)
+            )
+            num_spec_decode_tokens = 0
+            spec_token_indx = None
+            non_spec_token_indx = None
+            spec_state_indices_tensor = None
+            non_spec_state_indices_tensor = block_table_tensor[:, 0]
+            spec_query_start_loc = None
+            non_spec_query_start_loc = query_start_loc
+            non_spec_query_start_loc_cpu = query_start_loc_cpu
+            num_accepted_tokens = None
+        else:
+            query_lens = query_start_loc[1:] - query_start_loc[:-1]
+            assert spec_sequence_masks_cpu is not None
+            query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
+
+            # Use CPU tensors to avoid CPU-GPU sync
+            non_spec_query_lens_cpu = query_lens_cpu[~spec_sequence_masks_cpu]
+            num_decodes = (non_spec_query_lens_cpu == 1).sum().item()
+            # Exclude zero-length padded sequences from prefill count.
+            num_zero_len = (non_spec_query_lens_cpu == 0).sum().item()
+            num_prefills = non_spec_query_lens_cpu.size(0) - num_decodes - num_zero_len
+            num_decode_tokens = num_decodes
+            num_prefill_tokens = (
+                non_spec_query_lens_cpu.sum().item() - num_decode_tokens
+            )
+            num_spec_decode_tokens = (
+                query_lens_cpu.sum().item() - num_prefill_tokens - num_decode_tokens
+            )
+
+            if num_prefills == 0 and num_decodes == 0:
+                spec_token_size = min(
+                    num_spec_decodes * (self.num_spec + 1),
+                    query_start_loc_cpu[-1].item(),
+                )
+                spec_token_indx = torch.arange(
+                    spec_token_size,
+                    dtype=torch.int32,
+                    device=query_start_loc.device,
+                )
+                non_spec_token_indx = torch.empty(
+                    0, dtype=torch.int32, device=query_start_loc.device
+                )
+                # Filter by spec_sequence_masks to exclude padded sequences
+                spec_state_indices_tensor = block_table_tensor[
+                    spec_sequence_masks, : self.num_spec + 1
+                ]
+                non_spec_state_indices_tensor = None
+                # Padded sequences are always at the back, so the first
+                # num_spec_decodes + 1 entries of query_start_loc already
+                # contain the correct cumulative token counts.
+                spec_query_start_loc = query_start_loc[: num_spec_decodes + 1]
+                non_spec_query_start_loc = None
+                non_spec_query_start_loc_cpu = None
+            else:
+                spec_token_masks = torch.repeat_interleave(
+                    spec_sequence_masks, query_lens
+                )
+                index = torch.argsort(spec_token_masks, stable=True)
+                num_non_spec_tokens = num_prefill_tokens + num_decode_tokens
+                non_spec_token_indx = index[:num_non_spec_tokens]
+                spec_token_indx = index[num_non_spec_tokens:]
+
+                spec_state_indices_tensor = block_table_tensor[
+                    spec_sequence_masks, : self.num_spec + 1
+                ]
+                non_spec_state_indices_tensor = block_table_tensor[
+                    ~spec_sequence_masks, 0
+                ]
+
+                spec_query_start_loc = torch.zeros(
+                    num_spec_decodes + 1,
+                    dtype=torch.int32,
+                    device=query_start_loc.device,
+                )
+                torch.cumsum(
+                    query_lens[spec_sequence_masks], dim=0, out=spec_query_start_loc[1:]
+                )
+                non_spec_query_start_loc = torch.zeros(
+                    query_lens.size(0) - num_spec_decodes + 1,
+                    dtype=torch.int32,
+                    device=query_start_loc.device,
+                )
+                torch.cumsum(
+                    query_lens[~spec_sequence_masks],
+                    dim=0,
+                    out=non_spec_query_start_loc[1:],
+                )
+                non_spec_query_start_loc_cpu = torch.zeros(
+                    query_lens_cpu.size(0) - num_spec_decodes + 1,
+                    dtype=torch.int32,
+                )
+                torch.cumsum(
+                    query_lens_cpu[~spec_sequence_masks_cpu],
+                    dim=0,
+                    out=non_spec_query_start_loc_cpu[1:],
+                )
+
+            assert num_accepted_tokens is not None
+            num_accepted_tokens = num_accepted_tokens[spec_sequence_masks]
+
+        if num_prefills > 0:
+            has_initial_state = context_lens_tensor > 0
+            if spec_sequence_masks is not None:
+                has_initial_state = has_initial_state[~spec_sequence_masks]
+                assert non_spec_query_start_loc_cpu is not None
+            nums_dict, batch_ptr, token_chunk_offset_ptr = (
+                compute_causal_conv1d_metadata(
+                    non_spec_query_start_loc_cpu,
+                    device=query_start_loc.device,
+                )
+            )
+        else:
+            has_initial_state = None
+
+        # Function code counted on either presency non-spec decode or spec decode,
+        # but not both.
+        assert not (num_decodes > 0 and num_spec_decodes > 0), (
+            f"num_decodes: {num_decodes}, num_spec_decodes: {num_spec_decodes}"
+        )
+
+        # Prepare tensors for cudagraph
+        # Note: m.num_actual_tokens is already padded by the model runner for CUDAGraph
+        batch_size = m.num_actual_tokens
+
+        if (
+            self.use_full_cuda_graph
+            and num_prefills == 0
+            and num_decodes == 0
+            and num_spec_decodes <= self.decode_cudagraph_max_bs
+            and num_spec_decode_tokens <= self.decode_cudagraph_max_bs
+        ):
+            self.spec_state_indices_tensor[:num_spec_decodes].copy_(
+                spec_state_indices_tensor, non_blocking=True
+            )
+            spec_state_indices_tensor = self.spec_state_indices_tensor[:batch_size]
+            spec_state_indices_tensor[num_spec_decodes:].fill_(PAD_SLOT_ID)
+
+            self.spec_sequence_masks[:num_spec_decodes].copy_(
+                spec_sequence_masks[:num_spec_decodes], non_blocking=True
+            )
+            spec_sequence_masks = self.spec_sequence_masks[:batch_size]
+            spec_sequence_masks[num_spec_decodes:].fill_(False)
+
+            assert non_spec_token_indx is not None and spec_token_indx is not None
+            self.non_spec_token_indx[: non_spec_token_indx.size(0)].copy_(
+                non_spec_token_indx, non_blocking=True
+            )
+            non_spec_token_indx = self.non_spec_token_indx[
+                : non_spec_token_indx.size(0)
+            ]
+
+            self.spec_token_indx[: spec_token_indx.size(0)].copy_(
+                spec_token_indx, non_blocking=True
+            )
+            spec_token_indx = self.spec_token_indx[: spec_token_indx.size(0)]
+
+            self.spec_query_start_loc[: num_spec_decodes + 1].copy_(
+                spec_query_start_loc, non_blocking=True
+            )
+            spec_num_query_tokens = spec_query_start_loc[-1]  # type: ignore[index]
+            spec_query_start_loc = self.spec_query_start_loc[: batch_size + 1]
+            spec_query_start_loc[num_spec_decodes + 1 :].fill_(spec_num_query_tokens)
+
+            self.num_accepted_tokens[:num_spec_decodes].copy_(
+                num_accepted_tokens, non_blocking=True
+            )
+            num_accepted_tokens = self.num_accepted_tokens[:batch_size]
+            num_accepted_tokens[num_spec_decodes:].fill_(1)
+
+        if (
+            self.use_full_cuda_graph
+            and num_prefills == 0
+            and num_spec_decodes == 0
+            and num_decodes <= self.decode_cudagraph_max_bs
+        ):
+            self.non_spec_state_indices_tensor[:num_decodes].copy_(
+                non_spec_state_indices_tensor, non_blocking=True
+            )
+            non_spec_state_indices_tensor = self.non_spec_state_indices_tensor[
+                :batch_size
+            ]
+            non_spec_state_indices_tensor[num_decodes:].fill_(PAD_SLOT_ID)
+
+            self.non_spec_query_start_loc[: num_decodes + 1].copy_(
+                non_spec_query_start_loc, non_blocking=True
+            )
+            non_spec_num_query_tokens = non_spec_query_start_loc[-1]  # type: ignore[index]
+            non_spec_query_start_loc = self.non_spec_query_start_loc[: batch_size + 1]
+            non_spec_query_start_loc[num_decodes + 1 :].fill_(non_spec_num_query_tokens)
+
+        attn_metadata = GDNAttentionMetadata(
+            num_prefills=num_prefills,
+            num_prefill_tokens=num_prefill_tokens,
+            num_decodes=num_decodes,
+            num_decode_tokens=num_decode_tokens,
+            num_spec_decodes=num_spec_decodes,
+            num_spec_decode_tokens=num_spec_decode_tokens,
+            num_actual_tokens=m.num_actual_tokens,
+            has_initial_state=has_initial_state,
+            spec_query_start_loc=spec_query_start_loc,
+            non_spec_query_start_loc=non_spec_query_start_loc,
+            spec_state_indices_tensor=spec_state_indices_tensor,
+            non_spec_state_indices_tensor=non_spec_state_indices_tensor,
+            spec_sequence_masks=spec_sequence_masks,
+            spec_token_indx=spec_token_indx,
+            non_spec_token_indx=non_spec_token_indx,
+            num_accepted_tokens=num_accepted_tokens,
+            nums_dict=nums_dict,
+            batch_ptr=batch_ptr,
+            token_chunk_offset_ptr=token_chunk_offset_ptr,
+        )
+        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 Mamba.
+        """
+        m = common_attn_metadata
+
+        assert (
+            m.num_reqs <= self.decode_cudagraph_max_bs
+            and m.num_actual_tokens <= self.decode_cudagraph_max_bs
+        ), (
+            f"GDN only supports decode-only full CUDAGraph capture. "
+            f"Make sure batch size ({m.num_reqs}) <= "
+            f"cudagraph capture sizes ({self.decode_cudagraph_max_bs}), "
+            f"and number of tokens ({m.num_actual_tokens}) <= "
+            f"cudagraph capture sizes ({self.decode_cudagraph_max_bs})."
+        )
+
+        num_accepted_tokens = torch.diff(m.query_start_loc)
+        num_decode_draft_tokens_cpu = (num_accepted_tokens - 1).cpu()
+
+        return self.build(0, m, num_accepted_tokens, num_decode_draft_tokens_cpu)

vllm/v1/attention/backends/linear_attn.py

@@ -0,0 +1,89 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionMetadataBuilder,
+    CommonAttentionMetadata,
+)
+from vllm.v1.attention.backends.utils import (
+    mamba_get_block_table_tensor,
+    split_decodes_and_prefills,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec, MambaSpec
+
+
+class LinearAttentionBackend(AttentionBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "LINEAR_ATTN"
+
+    @staticmethod
+    def get_builder_cls() -> type["LinearAttentionMetadataBuilder"]:
+        return LinearAttentionMetadataBuilder
+
+
+@dataclass
+class LinearAttentionMetadata:
+    num_prefills: int
+    num_prefill_tokens: int
+    num_decodes: int
+    num_decode_tokens: int
+    query_start_loc: torch.Tensor
+    seq_lens: torch.Tensor
+
+    state_indices_tensor: torch.Tensor  # shape: [batch,]
+
+
+class LinearAttentionMetadataBuilder(AttentionMetadataBuilder[LinearAttentionMetadata]):
+    reorder_batch_threshold: int = 1
+
+    _cudagraph_support = AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(kv_cache_spec, layer_names, vllm_config, device)
+        assert isinstance(kv_cache_spec, MambaSpec)
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> LinearAttentionMetadata:
+        query_start_loc = common_attn_metadata.query_start_loc
+        seq_lens = common_attn_metadata.seq_lens
+
+        state_indices_tensor = mamba_get_block_table_tensor(
+            common_attn_metadata.block_table_tensor,
+            common_attn_metadata.seq_lens,
+            self.kv_cache_spec,
+            self.vllm_config.cache_config.mamba_cache_mode,
+        )[:, 0]
+
+        num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = (
+            split_decodes_and_prefills(
+                common_attn_metadata, decode_threshold=self.reorder_batch_threshold
+            )
+        )
+
+        attn_metadata = LinearAttentionMetadata(
+            num_prefills=num_prefills,
+            num_prefill_tokens=num_prefill_tokens,
+            num_decodes=num_decodes,
+            num_decode_tokens=num_decode_tokens,
+            query_start_loc=query_start_loc,
+            seq_lens=seq_lens,
+            state_indices_tensor=state_indices_tensor,
+        )
+        return attn_metadata

vllm/v1/attention/backends/mamba1_attn.py

@@ -0,0 +1,31 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+
+from vllm.v1.attention.backend import AttentionBackend
+from vllm.v1.attention.backends.mamba_attn import (
+    BaseMambaAttentionMetadata,
+    BaseMambaAttentionMetadataBuilder,
+)
+
+
+class Mamba1AttentionBackend(AttentionBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "MAMBA1_ATTN"
+
+    @staticmethod
+    def get_builder_cls() -> type["Mamba1AttentionMetadataBuilder"]:
+        return Mamba1AttentionMetadataBuilder
+
+
+@dataclass
+class Mamba1AttentionMetadata(BaseMambaAttentionMetadata):
+    pass
+
+
+class Mamba1AttentionMetadataBuilder(
+    BaseMambaAttentionMetadataBuilder[Mamba1AttentionMetadata]
+):
+    metadata_cls = Mamba1AttentionMetadata

vllm/v1/attention/backends/mamba2_attn.py

@@ -0,0 +1,267 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import itertools
+from dataclasses import dataclass, replace
+from typing import Any
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.utils.math_utils import cdiv
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    CommonAttentionMetadata,
+)
+from vllm.v1.attention.backends.mamba_attn import (
+    BaseMambaAttentionMetadata,
+    BaseMambaAttentionMetadataBuilder,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+
+def compute_varlen_chunk_metadata(
+    query_start_loc: torch.Tensor,
+    chunk_size: int,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """
+    Build chunk-aligned, variable-length metadata used by Mamba2 SSD kernels.
+
+    Given per-sequence cumulative token starts `query_start_loc` of shape [B+1]
+    and a physical `chunk_size`, returns three tensors on the same device:
+      - cu_chunk_seqlens:  (nchunks+1,) int32   exclusive prefix-sum of
+        logical-chunk lengths (each logical chunk never crosses a sequence or
+        physical-chunk boundary).
+      - last_chunk_indices: (B,)       int32   index of the last logical chunk
+        for each sequence (=-1 for empty sequences).
+      - seq_idx_chunks:     (nchunks,) int32   sequence index for each logical
+        chunk in order.
+
+    This is intentionally lightweight and CPU-side; it mirrors the metadata
+    produced by the V1 Mamba2 meta-data builder and is exported so tests
+    (and other callers) can avoid duplicating the logic.
+    """
+    assert query_start_loc.ndim == 1, "query_start_loc must be 1-D [B+1]"
+    assert int(query_start_loc[0].item()) == 0, "query_start_loc[0] must be 0"
+    device = query_start_loc.device
+
+    qsl64 = query_start_loc.to(torch.int64)
+    starts = qsl64[:-1].tolist()
+    ends = qsl64[1:].tolist()
+    total = int(qsl64[-1].item())
+
+    chunk_lens: list[int] = []
+    seq_idx_chunks: list[int] = []
+    last_chunk_indices: list[int] = [-1] * len(starts)
+
+    for b, (s, e) in enumerate(zip(starts, ends)):
+        if e <= s:
+            # empty sequence
+            continue
+        pos = s
+        while pos < e:
+            # split at both sequence boundaries and physical chunk boundaries
+            room = chunk_size - (pos % chunk_size)
+            take = min(room, e - pos)
+            chunk_lens.append(int(take))
+            seq_idx_chunks.append(b)
+            last_chunk_indices[b] = len(chunk_lens) - 1
+            pos += take
+
+    # Exclusive prefix sum over logical-chunk lengths
+    if chunk_lens:
+        cu_chunk_seqlens = torch.tensor(
+            [0] + list(itertools.accumulate(chunk_lens)),
+            device=device,
+            dtype=torch.int32,
+        )
+        # Final boundary must equal total tokens
+        assert int(cu_chunk_seqlens[-1].item()) == total
+    else:
+        cu_chunk_seqlens = torch.tensor([0], device=device, dtype=torch.int32)
+
+    last_chunk_indices_t = (
+        torch.tensor(last_chunk_indices, device=device, dtype=torch.int32)
+        if len(starts) > 0
+        else torch.empty((0,), device=device, dtype=torch.int32)
+    )
+    seq_idx_chunks_t = torch.tensor(seq_idx_chunks, device=device, dtype=torch.int32)
+    return cu_chunk_seqlens, last_chunk_indices_t, seq_idx_chunks_t
+
+
+class Mamba2AttentionBackend(AttentionBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "MAMBA2_ATTN"
+
+    @staticmethod
+    def get_builder_cls() -> type["Mamba2AttentionMetadataBuilder"]:
+        return Mamba2AttentionMetadataBuilder
+
+
+@dataclass
+class Mamba2AttentionMetadata(BaseMambaAttentionMetadata):
+    prep_initial_states: bool = False
+    chunk_size: int = 0
+
+    # Chunk-related metadata (only for prefill)
+    seq_idx_p: torch.Tensor | None = None
+    # cu_chunk_seqlen_p is a tensor of shape (nchunks+1,) that contains, for
+    # each chunk, its offsets into the varlen sequence dimension. It is defined
+    # such that the i-th chunk contains tokens from cu_chunk_seqlen_p[i] to
+    # cu_chunk_seqlen_p[i+1].
+    cu_chunk_seqlen_p: torch.Tensor | None = None
+    # last_chunk_indices_p is a tensor of shape (batch,) that contains the
+    # index of the last chunk for every sequence in the (prefill) batch.
+    last_chunk_indices_p: torch.Tensor | None = None
+
+
+class Mamba2AttentionMetadataBuilder(
+    BaseMambaAttentionMetadataBuilder[Mamba2AttentionMetadata]
+):
+    metadata_cls = Mamba2AttentionMetadata
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(kv_cache_spec, layer_names, vllm_config, device)
+        chunk_size = vllm_config.model_config.get_mamba_chunk_size()
+        assert chunk_size is not None, (
+            "chunk_size needs to be set in the model config for Mamba2 models"
+        )
+        self.chunk_size: int = chunk_size
+
+    def _compute_chunk_metadata(
+        self,
+        num_prefills: int,
+        num_computed_tokens_p_cpu: torch.Tensor,
+        query_start_loc_p_cpu: torch.Tensor,
+    ) -> tuple[list[int], list[int], list[int]]:
+        """
+        Compute chunk-specific metadata for Mamba2.
+
+        The code below carefully constructs the chunks such that:
+        1. Chunks contain tokens from a *single* sequence only.
+        2. For every sequence, we are guaranteed that we can
+           retrieve the mamba state *every* chunk_size tokens.
+        Constraint (1) dramatically simplifies the mamba2 kernels.
+        Constraint (2) dramatically simplifies the implementation
+        of prefix caching for mamba2 (wip). We need to take care
+        of the interaction with chunked prefill in order to
+        satisfy constraint (2).
+        """
+        # TODO (tdoublep): This code could probably be optimized.
+        cu_chunk_seqlen = []
+        seq_idx = []
+        last_chunk_indices = []
+        seqlen_pos = 0
+
+        for req_idx in range(num_prefills):
+            this_num_computed = num_computed_tokens_p_cpu[req_idx].item()
+            this_new_tokens = (
+                query_start_loc_p_cpu[req_idx + 1].item()
+                - query_start_loc_p_cpu[req_idx].item()
+            )
+
+            # if computed tokens are not chunk-aligned, use the first
+            # chunk to finish it off
+            if this_num_computed % self.chunk_size != 0:
+                seq_idx.append(req_idx)
+                cu_chunk_seqlen.append(seqlen_pos)
+                # how many tokens to finish the chunk?
+                chunk_len = (
+                    cdiv(this_num_computed, self.chunk_size) * self.chunk_size
+                    - this_num_computed
+                )
+                # we can only use at most this_new_tokens
+                chunk_len = min(chunk_len, this_new_tokens)
+                seqlen_pos += chunk_len
+                this_new_tokens -= chunk_len
+
+            n_chunks = cdiv(this_new_tokens, self.chunk_size)
+            for chunk in range(n_chunks):
+                seq_idx.append(req_idx)
+                cu_chunk_seqlen.append(seqlen_pos)
+                chunk_len = min(self.chunk_size, this_new_tokens)
+                seqlen_pos += chunk_len
+                this_new_tokens -= chunk_len
+
+            assert this_new_tokens == 0
+            last_chunk_indices.append(len(cu_chunk_seqlen) - 1)
+
+        cu_chunk_seqlen.append(seqlen_pos)
+
+        return cu_chunk_seqlen, seq_idx, last_chunk_indices
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+        **kwargs: Any,
+    ) -> Mamba2AttentionMetadata:
+        common = self._compute_common_metadata(
+            common_attn_metadata, num_accepted_tokens=kwargs.get("num_accepted_tokens")
+        )
+
+        seq_idx_p = None
+        cu_chunk_seqlen_p = None
+        last_chunk_indices_p = None
+        prep_initial_states = False
+
+        # Compute seq_idx for prefill only
+        if common.num_prefills > 0:
+            prep_initial_states = (
+                torch.any(common.has_initial_states_p).item()
+                if common.has_initial_states_p is not None
+                else False
+            )
+
+            num_reqs = common.num_reqs
+            num_prefills = common.num_prefills
+            num_decode_tokens = common.num_decode_tokens
+
+            num_computed_tokens_cpu = (
+                common_attn_metadata.compute_num_computed_tokens().cpu()
+            )
+            num_computed_tokens_p_cpu = num_computed_tokens_cpu[
+                num_reqs - num_prefills : num_reqs
+            ]
+            query_start_loc_p_cpu = (
+                common_attn_metadata.query_start_loc_cpu[-num_prefills - 1 :]
+                - num_decode_tokens
+            )
+
+            cu_chunk_seqlen, seq_idx, last_chunk_indices = self._compute_chunk_metadata(
+                num_prefills,
+                num_computed_tokens_p_cpu,
+                query_start_loc_p_cpu,
+            )
+
+            seq_idx_p = torch.as_tensor(
+                seq_idx,
+                device=common_attn_metadata.query_start_loc.device,
+                dtype=torch.int32,
+            )
+            cu_chunk_seqlen_p = torch.as_tensor(
+                cu_chunk_seqlen,
+                device=common_attn_metadata.query_start_loc.device,
+                dtype=torch.int32,
+            )
+            last_chunk_indices_p = torch.as_tensor(
+                last_chunk_indices,
+                device=common_attn_metadata.query_start_loc.device,
+                dtype=torch.int32,
+            )
+
+        return replace(
+            common,
+            prep_initial_states=prep_initial_states,
+            chunk_size=self.chunk_size,
+            seq_idx_p=seq_idx_p,
+            cu_chunk_seqlen_p=cu_chunk_seqlen_p,
+            last_chunk_indices_p=last_chunk_indices_p,
+        )

vllm/v1/attention/backends/mamba_attn.py

@@ -0,0 +1,464 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import abc
+from dataclasses import dataclass, replace
+from typing import Any, ClassVar, TypeVar
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.utils.math_utils import cdiv
+from vllm.v1.attention.backend import (
+    AttentionCGSupport,
+    AttentionMetadataBuilder,
+    CommonAttentionMetadata,
+)
+from vllm.v1.attention.backends.utils import (
+    PAD_SLOT_ID,
+    compute_causal_conv1d_metadata,
+    mamba_get_block_table_tensor,
+    split_decodes_and_prefills,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec, MambaSpec
+
+M = TypeVar("M", bound="BaseMambaAttentionMetadata")
+
+
+@dataclass
+class BaseMambaAttentionMetadata:
+    num_prefills: int
+    num_prefill_tokens: int
+    num_decodes: int
+    num_decode_tokens: int
+    num_reqs: int
+
+    # The following tensors only contain prefill requests and will be None if
+    # the batch has no prefill requests.
+    has_initial_states_p: torch.Tensor | None
+    query_start_loc_p: torch.Tensor | None
+    num_computed_tokens_p: torch.Tensor | None
+    state_indices_tensor_p: torch.Tensor | None
+
+    # The following tensors are used for decode requests and
+    # speculative decoding compatibility, and will be None if the batch
+    # has no decode requests.
+    state_indices_tensor_d: torch.Tensor | None
+    query_start_loc_d: torch.Tensor | None  # shape: [num_decodes + 1,]
+
+    # Number of accepted tokens for each spec sequence (for loading correct checkpoint)
+    # Includes the bonus token (so minimum is 1)
+    num_accepted_tokens: torch.Tensor | None  # shape: [batch,]
+
+    # The following tensors are only used for prefix caching in all mode and
+    # are None if disabled
+    block_idx_last_scheduled_token: torch.Tensor | None
+    block_idx_first_scheduled_token_p: torch.Tensor | None
+    block_idx_last_computed_token: torch.Tensor | None
+
+    # The following tensor is only used for prefix caching in align mode
+    seq_lens: torch.Tensor
+
+    # The following attributes are for triton implementation of causal_conv1d
+    nums_dict: dict | None = None
+    batch_ptr: torch.Tensor | None = None
+    token_chunk_offset_ptr: torch.Tensor | None = None
+
+
+class BaseMambaAttentionMetadataBuilder(AttentionMetadataBuilder[M], abc.ABC):
+    metadata_cls: type[M]
+    reorder_batch_threshold: int = 1
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH
+
+    # Will be disabled if speculative decoding is used
+    supports_update_block_table: bool = True
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(kv_cache_spec, layer_names, vllm_config, device)
+
+        # Enable speculative decoding support
+        self.speculative_config = vllm_config.speculative_config
+        self.compilation_config = vllm_config.compilation_config
+        self.num_spec_tokens: int = vllm_config.num_speculative_tokens
+        self.use_spec_decode = self.num_spec_tokens > 0
+
+        assert isinstance(kv_cache_spec, MambaSpec)
+        self.compilation_config = vllm_config.compilation_config
+        self.decode_cudagraph_max_bs = self.vllm_config.scheduler_config.max_num_seqs
+        if self.compilation_config.max_cudagraph_capture_size is not None:
+            self.decode_cudagraph_max_bs = min(
+                self.decode_cudagraph_max_bs,
+                self.compilation_config.max_cudagraph_capture_size,
+            )
+
+        if self.vllm_config.cache_config.mamba_cache_mode == "all":
+            max_num_blocks = cdiv(
+                self.vllm_config.model_config.max_model_len,
+                self.kv_cache_spec.block_size,
+            )
+            # Speculative decoding not supported with prefix caching,
+            # so keep shape consistent with prefill buffer
+            # TODO: reduce this size as needed for decode-only cudagraph capture
+            self.state_indices_tensor_d = torch.empty(
+                (
+                    self.decode_cudagraph_max_bs,
+                    max_num_blocks,
+                ),
+                dtype=torch.int32,
+                device=device,
+            )
+            self.block_idx_last_scheduled_token = torch.empty(
+                (self.decode_cudagraph_max_bs,),
+                dtype=torch.int32,
+                device=device,
+            )
+            self.block_idx_last_computed_token = torch.empty(
+                (self.decode_cudagraph_max_bs,),
+                dtype=torch.int32,
+                device=device,
+            )
+        else:
+            self.state_indices_tensor_d = torch.empty(
+                (self.decode_cudagraph_max_bs, 1 + self.num_spec_tokens),
+                dtype=torch.int32,
+                device=device,
+            )
+
+        # For speculative decoding, we need to store the following buffers
+        # for CUDA graph capture during decode
+        if self.num_spec_tokens > 0:
+            self.decode_num_accepted_tokens = torch.empty(
+                (self.decode_cudagraph_max_bs,),
+                dtype=torch.int32,
+                device=device,
+            )
+
+        self._init_reorder_batch_threshold(1, self.use_spec_decode)
+        if self.use_spec_decode:
+            self.supports_update_block_table = False
+
+    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 Mamba.
+        """
+        m = common_attn_metadata
+
+        assert (
+            m.max_query_len <= 1 + self.num_spec_tokens
+            and m.num_reqs <= self.decode_cudagraph_max_bs
+        ), (
+            "Mamba only supports decode-only full CUDAGraph capture. "
+            "Make sure all cudagraph capture sizes <= max_num_seq."
+        )
+
+        assert m.max_query_len == 1 + self.num_spec_tokens  # decode-only
+
+        num_accepted_tokens = None
+        if self.num_spec_tokens > 0:
+            num_accepted_tokens = torch.diff(m.query_start_loc)
+
+        return self.build(0, m, num_accepted_tokens=num_accepted_tokens)
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+        *,
+        num_accepted_tokens: torch.Tensor | None = None,
+        **kwargs: Any,
+    ) -> M:
+        """
+        Default build implementation for Mamba-like attention backends.
+        Subclasses (e.g., Mamba2) can override to add additional metadata.
+        """
+        return self._compute_common_metadata(
+            common_attn_metadata, num_accepted_tokens=num_accepted_tokens
+        )
+
+    def _compute_prefix_caching_block_indices(
+        self,
+        common_attn_metadata: CommonAttentionMetadata,
+        mamba_block_size: int,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        num_computed_tokens = common_attn_metadata.compute_num_computed_tokens()
+        # Block index of the last computed token
+        block_idx_last_computed_token = cdiv(num_computed_tokens, mamba_block_size) - 1
+        # which is <= block index for the first scheduled token
+        block_idx_first_scheduled_token = (
+            cdiv(num_computed_tokens + 1, mamba_block_size) - 1
+        )
+        # which is <= block index of the last scheduled token
+        block_idx_last_scheduled_token = (
+            cdiv(common_attn_metadata.seq_lens, mamba_block_size) - 1
+        )
+        # -1 in case it's non-computed and causes later issues with indexing
+        block_idx_last_computed_token = torch.clamp(
+            block_idx_last_computed_token, min=0
+        )
+        # -1 in the case we have a padded request (0 seq-len)
+        block_idx_last_scheduled_token = torch.clamp(
+            block_idx_last_scheduled_token, min=0
+        )
+
+        return (
+            block_idx_last_computed_token,
+            block_idx_first_scheduled_token,
+            block_idx_last_scheduled_token,
+        )
+
+    def _compute_common_metadata(
+        self,
+        common_attn_metadata: CommonAttentionMetadata,
+        *,
+        num_accepted_tokens: torch.Tensor | None = None,
+    ) -> M:
+        """
+        Compute metadata common to both Mamba1 and Mamba2.
+        """
+        num_reqs = common_attn_metadata.num_reqs
+
+        # Treat multi-token queries as decode requests when
+        # speculative decoding is enabled. Otherwise, use the
+        # default decode threshold to prevent misclassification
+        # of prefill queries as decode requests.
+        decode_threshold = (
+            self.reorder_batch_threshold if num_accepted_tokens is not None else 1
+        )
+
+        num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = (
+            split_decodes_and_prefills(
+                common_attn_metadata, decode_threshold=decode_threshold
+            )
+        )
+
+        # Need flags to indicate if there are initial states
+        has_initial_states_p = None
+        query_start_loc_p = None
+        query_start_loc_d = None
+        num_computed_tokens = None
+        num_computed_tokens_p = None
+
+        # for prefix caching
+        block_idx_first_scheduled_token = None
+        block_idx_first_scheduled_token_p = None
+        block_idx_last_computed_token = None
+        block_idx_last_scheduled_token = None
+
+        # for causal_conv1d
+        nums_dict, batch_ptr, token_chunk_offset_ptr = None, None, None
+
+        if self.vllm_config.cache_config.mamba_cache_mode == "all":
+            num_computed_tokens = common_attn_metadata.compute_num_computed_tokens()
+
+            # Return a tensor of shape (#requests, #max blocks)
+            state_indices_tensor = common_attn_metadata.block_table_tensor
+            # Additional cache-related varaiables:
+            mamba_block_size = self.kv_cache_spec.block_size
+            (
+                block_idx_last_computed_token,
+                block_idx_first_scheduled_token,
+                block_idx_last_scheduled_token,
+            ) = self._compute_prefix_caching_block_indices(
+                common_attn_metadata, mamba_block_size
+            )
+        else:
+            state_indices_tensor = mamba_get_block_table_tensor(
+                common_attn_metadata.block_table_tensor,
+                common_attn_metadata.seq_lens,
+                self.kv_cache_spec,
+                self.vllm_config.cache_config.mamba_cache_mode,
+            )
+
+        if state_indices_tensor.dim() == 1:
+            state_indices_tensor = state_indices_tensor.unsqueeze(-1)
+
+        state_indices_tensor_d, state_indices_tensor_p = torch.split(
+            state_indices_tensor,
+            [num_decodes, num_prefills],
+            dim=0,
+        )
+        if self.vllm_config.cache_config.mamba_cache_mode != "all":
+            state_indices_tensor_d = state_indices_tensor_d[
+                :, : 1 + self.num_spec_tokens
+            ]
+            state_indices_tensor_p = state_indices_tensor_p[:, 0]
+
+        if num_decodes > 0 and self.use_spec_decode:
+            assert num_accepted_tokens is not None
+            query_start_loc_d = common_attn_metadata.query_start_loc[: num_decodes + 1]
+            num_accepted_tokens = num_accepted_tokens[:num_decodes]
+
+        if num_prefills > 0:
+            if num_computed_tokens is None:
+                num_computed_tokens = common_attn_metadata.compute_num_computed_tokens()
+
+            query_start_loc_p_cpu = (
+                common_attn_metadata.query_start_loc_cpu[-num_prefills - 1 :]
+                - num_decode_tokens
+            )
+            query_start_loc_p = (
+                common_attn_metadata.query_start_loc[-num_prefills - 1 :]
+                - num_decode_tokens
+            )
+            has_initial_states_p = (
+                num_computed_tokens[num_reqs - num_prefills : num_reqs] > 0
+            )
+
+            nums_dict, batch_ptr, token_chunk_offset_ptr = (
+                compute_causal_conv1d_metadata(
+                    query_start_loc_p_cpu,
+                    device=common_attn_metadata.query_start_loc.device,
+                )
+            )
+
+            if self.vllm_config.cache_config.mamba_cache_mode == "all":
+                assert num_computed_tokens is not None
+                num_computed_tokens_p = num_computed_tokens[
+                    num_reqs - num_prefills : num_reqs
+                ]
+                assert block_idx_first_scheduled_token is not None
+                block_idx_first_scheduled_token_p = block_idx_first_scheduled_token[
+                    num_reqs - num_prefills : num_reqs
+                ]
+
+        metadata = self.metadata_cls(
+            num_prefills=num_prefills,
+            num_prefill_tokens=num_prefill_tokens,
+            num_decodes=num_decodes,
+            num_decode_tokens=num_decode_tokens,
+            query_start_loc_p=query_start_loc_p,
+            has_initial_states_p=has_initial_states_p,
+            state_indices_tensor_p=state_indices_tensor_p,
+            state_indices_tensor_d=state_indices_tensor_d,
+            num_accepted_tokens=num_accepted_tokens,
+            query_start_loc_d=query_start_loc_d,
+            block_idx_last_scheduled_token=block_idx_last_scheduled_token,
+            block_idx_first_scheduled_token_p=block_idx_first_scheduled_token_p,
+            block_idx_last_computed_token=block_idx_last_computed_token,
+            num_computed_tokens_p=num_computed_tokens_p,
+            num_reqs=num_reqs,
+            seq_lens=common_attn_metadata.seq_lens,
+            nums_dict=nums_dict,
+            batch_ptr=batch_ptr,
+            token_chunk_offset_ptr=token_chunk_offset_ptr,
+        )
+
+        return self._update_metadata_for_cudagraph_capture(metadata)
+
+    def _update_metadata_for_cudagraph_capture(
+        self,
+        metadata: M,
+    ) -> M:
+        """
+        Update the metadata for cudagraph capture.
+        Currently, only decode is supported for full cudagraphs with Mamba.
+        """
+        state_indices_tensor_d = metadata.state_indices_tensor_d
+        query_start_loc_d = metadata.query_start_loc_d
+        num_accepted_tokens = metadata.num_accepted_tokens
+        block_idx_last_scheduled_token = metadata.block_idx_last_scheduled_token
+        block_idx_last_computed_token = metadata.block_idx_last_computed_token
+        if (
+            metadata.num_prefills == 0
+            and metadata.num_decodes <= self.decode_cudagraph_max_bs
+            and self.compilation_config.cudagraph_mode.has_full_cudagraphs()
+        ):
+            padded_bs = metadata.num_reqs
+            self.state_indices_tensor_d[: metadata.num_decodes].copy_(
+                state_indices_tensor_d, non_blocking=True
+            )
+            state_indices_tensor_d = self.state_indices_tensor_d[:padded_bs]
+            state_indices_tensor_d[metadata.num_decodes :] = PAD_SLOT_ID
+
+            if self.use_spec_decode:
+                assert query_start_loc_d is not None
+                assert num_accepted_tokens is not None
+                query_start_loc_d = query_start_loc_d[: padded_bs + 1]
+                self.decode_num_accepted_tokens[: metadata.num_decodes].copy_(
+                    num_accepted_tokens, non_blocking=True
+                )
+                num_accepted_tokens = self.decode_num_accepted_tokens[:padded_bs]
+                num_accepted_tokens[metadata.num_decodes :] = (
+                    1  # pad with 1st slot index
+                )
+
+            if self.vllm_config.cache_config.mamba_cache_mode == "all":
+                assert block_idx_last_scheduled_token is not None
+                assert block_idx_last_computed_token is not None
+                self.block_idx_last_scheduled_token[: metadata.num_decodes].copy_(
+                    block_idx_last_scheduled_token[: metadata.num_decodes],
+                    non_blocking=True,
+                )
+                block_idx_last_scheduled_token = self.block_idx_last_scheduled_token[
+                    : metadata.num_decode_tokens
+                ]
+
+                self.block_idx_last_computed_token[: metadata.num_decodes].copy_(
+                    block_idx_last_computed_token[: metadata.num_decodes],
+                    non_blocking=True,
+                )
+                block_idx_last_computed_token = self.block_idx_last_computed_token[
+                    : metadata.num_decode_tokens
+                ]
+
+        return replace(
+            metadata,
+            state_indices_tensor_d=state_indices_tensor_d,
+            query_start_loc_d=query_start_loc_d,
+            num_accepted_tokens=num_accepted_tokens,
+            block_idx_last_scheduled_token=block_idx_last_scheduled_token,
+            block_idx_last_computed_token=block_idx_last_computed_token,
+        )
+
+    def update_block_table(
+        self,
+        metadata: M,
+        blk_table: torch.Tensor,
+        slot_mapping: torch.Tensor,
+    ) -> M:
+        state_indices_tensor = mamba_get_block_table_tensor(
+            blk_table,
+            metadata.seq_lens,
+            self.kv_cache_spec,
+            self.vllm_config.cache_config.mamba_cache_mode,
+        )
+        if state_indices_tensor.dim() == 1:
+            state_indices_tensor = state_indices_tensor.unsqueeze(-1)
+
+        assert (
+            metadata.num_prefills + metadata.num_decodes
+            == state_indices_tensor.shape[0]
+        ), (
+            "Mismatch in number of requests when updating block table."
+            f" Expected {metadata.num_prefills + metadata.num_decodes}, "
+            f"got {state_indices_tensor.shape[0]}."
+        )
+
+        state_indices_tensor_d, state_indices_tensor_p = torch.split(
+            state_indices_tensor,
+            [metadata.num_decodes, metadata.num_prefills],
+            dim=0,
+        )
+        if self.vllm_config.cache_config.mamba_cache_mode != "all":
+            state_indices_tensor_d = state_indices_tensor_d[
+                :, : 1 + self.num_spec_tokens
+            ]
+            state_indices_tensor_p = state_indices_tensor_p[:, 0]
+
+        new_metadata = replace(
+            metadata,
+            state_indices_tensor_d=state_indices_tensor_d,
+            state_indices_tensor_p=state_indices_tensor_p,
+        )
+
+        return self._update_metadata_for_cudagraph_capture(new_metadata)

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

@@ -0,0 +1,66 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from vllm.v1.attention.backends.mla.rocm_aiter_mla import AiterMLABackend, AiterMLAImpl
+
+
+class AiterTritonMLABackend(AiterMLABackend):
+    @staticmethod
+    def get_name() -> str:
+        return "AITER_TRITON_MLA"
+
+    @staticmethod
+    def get_impl_cls() -> type["AiterTritonMLAImpl"]:
+        return AiterTritonMLAImpl
+
+
+class AiterTritonMLAImpl(AiterMLAImpl):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        **mla_args,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            **mla_args,
+        )
+        from aiter.ops.triton.mha import flash_attn_varlen_func
+
+        self.flash_attn_varlen_func = flash_attn_varlen_func
+
+    def _flash_attn_varlen_diff_headdims(
+        self, q, k, v, return_softmax_lse=False, softmax_scale=None, **kwargs
+    ):
+        result = self.flash_attn_varlen_func(  # type: ignore[call-arg]
+            q,
+            k,
+            v,
+            softmax_scale=softmax_scale,
+            return_lse=return_softmax_lse,
+            **kwargs,
+        )
+        # Transpose the LSE if Triton MHA is used:
+        # (q.shape[0], num_q_heads) to (num_q_heads, q.shape[0])
+        if type(result) is tuple and return_softmax_lse:
+            output, lse = result
+            lse = lse.T.contiguous()
+            return (output, lse)
+        return result

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

@@ -0,0 +1,279 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import os
+from typing import ClassVar
+
+import torch
+
+import vllm._custom_ops as ops
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention.mla_attention import (
+    MLACommonBackend,
+    MLACommonImpl,
+    MLACommonMetadata,
+    MLACommonMetadataBuilder,
+)
+from vllm.platforms.interface import DeviceCapability
+from vllm.utils.platform_utils import num_compute_units
+from vllm.v1.attention.backend import (
+    AttentionCGSupport,
+    AttentionLayer,
+    AttentionType,
+    MultipleOf,
+    is_quantized_kv_cache,
+)
+
+logger = init_logger(__name__)
+
+
+class CutlassMLAMetadataBuilder(MLACommonMetadataBuilder[MLACommonMetadata]):
+    # enable full CUDA Graph support for decode-only capture
+    _cudagraph_support: ClassVar[AttentionCGSupport] = (
+        AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE
+    )
+
+
+class CutlassMLABackend(MLACommonBackend):
+    supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+    supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
+        "auto",
+        "bfloat16",
+        "fp8",
+        "fp8_e4m3",
+    ]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [128]
+
+    @staticmethod
+    def get_name() -> str:
+        return "CUTLASS_MLA"
+
+    @staticmethod
+    def get_impl_cls() -> type["CutlassMLAImpl"]:
+        return CutlassMLAImpl
+
+    @staticmethod
+    def get_builder_cls() -> type["CutlassMLAMetadataBuilder"]:
+        return CutlassMLAMetadataBuilder
+
+    @classmethod
+    def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
+        return capability.major == 10
+
+
+class SM100Workspace:
+    def __init__(self, initial_workspace_size):
+        self._workspace_buf = torch.empty(
+            initial_workspace_size, device="cuda", dtype=torch.uint8
+        )
+
+        self._block_size = 128  # Forced to 128
+
+        # Pre-compute sm_count to avoid recomputing it. Use device 0 as a proxy
+        # (assumes all devices are similar)
+        self._sm_count = num_compute_units(0)
+
+    def get_buf(self):
+        return self._workspace_buf
+
+    def ensure_size(self, attn_metadata: MLACommonMetadata, num_kv_splits: int):
+        batch_size = attn_metadata.num_reqs
+        max_seq_len = attn_metadata.max_query_len
+
+        workspace_size = ops.sm100_cutlass_mla_get_workspace_size(
+            max_seq_len * self._block_size,
+            batch_size,
+            self._sm_count,
+            num_kv_splits=num_kv_splits,
+        )
+
+        if self._workspace_buf.shape[0] < workspace_size:
+            self._workspace_buf.resize_(workspace_size)
+
+
+g_sm100_workspace = SM100Workspace(128 * 1024 * 1024)  # 128MB
+
+MAX_HEADS = 128
+
+
+class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
+    can_return_lse_for_decode: bool = True
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        **mla_args,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            q_pad_num_heads=MAX_HEADS,
+            **mla_args,
+        )
+
+        unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "CutlassMLAImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, logits_soft_cap"
+            )
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError(
+                "Encoder self-attention and "
+                "encoder/decoder cross-attention "
+                "are not implemented for "
+                "CutlassMLAImpl"
+            )
+
+        # TODO: Currently, num_kv_splits is limited to 16 to avoid hanging
+        #       issues. In case the code hangs, use:
+        #       FORCE_NUM_KV_SPLITS=1
+        force_num_kv_splits = os.environ.get("FORCE_NUM_KV_SPLITS", None)
+        if force_num_kv_splits:
+            logger.debug_once("Forcing num_kv_splits to %d", int(force_num_kv_splits))
+            self._num_kv_splits = int(force_num_kv_splits)
+        else:
+            self._num_kv_splits = -1  # => Auto-detect
+
+        # Share workspace buffer across all executions
+        self._workspace = g_sm100_workspace
+
+    def _sm100_cutlass_mla_decode(
+        self,
+        q_nope: torch.Tensor,
+        q_pe: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        seq_lens: torch.Tensor,
+        page_table: torch.Tensor,
+        workspace: torch.Tensor,
+        sm_scale: float,
+        num_kv_splits: int,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        assert q_nope.ndim == 3, f"q_nope must be a 3D tensor, but got {q_nope.ndim}"
+        assert q_pe.ndim == 3, f"q_pe must be a 3D tensor, but got {q_pe.ndim}"
+        assert kv_c_and_k_pe_cache.ndim == 3, (
+            "kv_c_and_k_pe_cache must be a 3D tensor, but got {}".format(
+                kv_c_and_k_pe_cache.ndim
+            )
+        )
+
+        B_q, H, D_q_nope = q_nope.shape
+        B_q_2, H_2, D_q_pe = q_pe.shape
+        assert (B_q == B_q_2) and (H == H_2)
+
+        _, PAGE_SIZE, D_ckv = kv_c_and_k_pe_cache.shape
+
+        D_latent = 512
+        D_rope = 64
+        assert D_q_nope == D_latent
+        assert D_q_pe == D_rope
+        assert D_ckv == D_latent + D_rope
+
+        MAX_HEADS = 128
+        assert H <= MAX_HEADS, f"H must be <= {MAX_HEADS}, but got {H}"
+
+        assert len(page_table.shape) == 2
+        B_block_table, block_num = page_table.shape
+        assert B_block_table == B_q
+        assert block_num > 0, f"block num must be greater than 0, got {block_num}"
+        assert block_num % (128 / PAGE_SIZE) == 0
+
+        assert q_nope.dtype in (torch.float16, torch.bfloat16, torch.float8_e4m3fn), (
+            f"q_nope.dtype needs to be fp16 or bf16 or e4m3 but got {q_nope.dtype}."
+        )
+        assert q_nope.dtype == q_pe.dtype == kv_c_and_k_pe_cache.dtype
+        assert seq_lens.dtype == torch.int32, (
+            f"seq_lens.dtype needs to be int32 but got {seq_lens.dtype}."
+        )
+        assert page_table.dtype == torch.int32, (
+            f"page_table.dtype needs to be int32 but got {page_table.dtype}."
+        )
+
+        dtype = (
+            torch.bfloat16
+            if is_quantized_kv_cache(self.kv_cache_dtype)
+            else q_nope.dtype
+        )
+        out = q_nope.new_empty((B_q, MAX_HEADS, D_latent), dtype=dtype)
+        lse = (
+            torch.empty((B_q, MAX_HEADS), dtype=torch.float32, device=q_nope.device)
+            if self.need_to_return_lse_for_decode
+            else torch.Tensor()
+        )
+
+        ops.sm100_cutlass_mla_decode(
+            out,
+            lse,
+            q_nope,
+            q_pe,
+            kv_c_and_k_pe_cache,
+            seq_lens,
+            page_table,
+            workspace,
+            sm_scale,
+            num_kv_splits,
+        )
+
+        if H < MAX_HEADS:
+            # Extract the subsets of the outputs
+            lse = lse[:, :H] if self.need_to_return_lse_for_decode else lse
+            out = out[:, :H]
+
+        return out, lse
+
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: MLACommonMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        assert kv_c_and_k_pe_cache.numel() > 0
+        assert attn_metadata.decode is not None
+
+        if type(q) is tuple:
+            q_nope, q_pe = q
+        else:
+            q_nope, q_pe = torch.split(
+                q, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+            )
+
+        # Adjust workspace size (if necessary)
+        self._workspace.ensure_size(attn_metadata, self._num_kv_splits)
+
+        # Run MLA
+        o, lse = self._sm100_cutlass_mla_decode(
+            q_nope,
+            q_pe,
+            kv_c_and_k_pe_cache,
+            attn_metadata.decode.seq_lens,
+            attn_metadata.decode.block_table,
+            self._workspace.get_buf(),
+            self.scale,
+            self._num_kv_splits,
+        )
+
+        return o, (lse if self.need_to_return_lse_for_decode else None)

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

@@ -0,0 +1,361 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention.mla_attention import (
+    MLACommonBackend,
+    MLACommonDecodeMetadata,
+    MLACommonImpl,
+    MLACommonMetadata,
+    MLACommonMetadataBuilder,
+    QueryLenSupport,
+)
+from vllm.model_executor.layers.batch_invariant import (
+    vllm_is_batch_invariant,
+)
+from vllm.platforms.interface import DeviceCapability
+from vllm.utils.math_utils import round_up
+from vllm.v1.attention.backend import (
+    AttentionCGSupport,
+    AttentionLayer,
+    AttentionType,
+    MultipleOf,
+    is_quantized_kv_cache,
+)
+from vllm.v1.attention.backends.fa_utils import (
+    flash_attn_supports_mla,
+    get_flash_attn_version,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec
+from vllm.vllm_flash_attn import (  # type: ignore[attr-defined]
+    flash_attn_varlen_func,
+    get_scheduler_metadata,
+)
+
+logger = init_logger(__name__)
+
+
+class FlashAttnMLABackend(MLACommonBackend):
+    supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+    supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
+        "auto",
+        "bfloat16",
+    ]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [MultipleOf(16)]
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASH_ATTN_MLA"
+
+    @staticmethod
+    def get_builder_cls() -> type["FlashAttnMLAMetadataBuilder"]:
+        return FlashAttnMLAMetadataBuilder
+
+    @staticmethod
+    def get_impl_cls() -> type["FlashAttnMLAImpl"]:
+        return FlashAttnMLAImpl
+
+    @classmethod
+    def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
+        return capability.major == 9
+
+    @classmethod
+    def supports_combination(
+        cls,
+        head_size: int,
+        dtype: torch.dtype,
+        kv_cache_dtype: CacheDType | None,
+        block_size: int,
+        use_mla: bool,
+        has_sink: bool,
+        use_sparse: bool,
+        device_capability: DeviceCapability,
+    ) -> str | None:
+        if not flash_attn_supports_mla():
+            return "FlashAttention MLA not supported on this device"
+        return None
+
+
+@dataclass
+class FlashAttnMLADecodeMetadata(MLACommonDecodeMetadata):
+    query_start_loc: torch.Tensor
+    max_query_len: int
+    max_seq_len: int
+    scheduler_metadata: torch.Tensor | None = None
+    max_num_splits: int = 0
+
+
+@dataclass
+class FlashAttnMLAMetadata(MLACommonMetadata[FlashAttnMLADecodeMetadata]):
+    pass
+
+
+class FlashAttnMLAMetadataBuilder(MLACommonMetadataBuilder[FlashAttnMLAMetadata]):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH
+    query_len_support: ClassVar[QueryLenSupport] = QueryLenSupport.VARLEN
+    reorder_batch_threshold: int = 512  # process small prefills with decode pathway
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        interleave_size = vllm_config.parallel_config.cp_kv_cache_interleave_size
+        super().__init__(
+            kv_cache_spec,
+            layer_names,
+            vllm_config,
+            device,
+            FlashAttnMLAMetadata,
+            supports_dcp_with_varlen=(interleave_size == 1),
+        )
+        self.max_num_splits = 0  # No upper bound on the number of splits.
+        self.fa_aot_schedule = get_flash_attn_version() == 3
+
+        self.use_full_cuda_graph = (
+            self.compilation_config.cudagraph_mode.has_full_cudagraphs()
+        )
+        self.max_cudagraph_size = self.compilation_config.max_cudagraph_capture_size
+
+        if self.use_full_cuda_graph and self.fa_aot_schedule:
+            # FA3 scheduler_metadata size: 1 + round_up(batch_size, 4) * 4
+            # The +1 is for the tile_count_semaphore (synchronization).
+            # The 4 slots per batch element (num_prepare_batch_vectors) are:
+            #   prepare_varlen + dynamic_split + sort_batches + head_swizzle
+            # See: https://github.com/vllm-project/flash-attention/blob/5824e6e/hopper/flash_api.cpp#L664-L671  # noqa: E501
+            max_batch_size = max(
+                vllm_config.scheduler_config.max_num_seqs,
+                self.max_cudagraph_size or 0,
+            )
+            self.scheduler_metadata = torch.zeros(
+                1 + round_up(max_batch_size, 4) * 4,
+                dtype=torch.int32,
+                device=self.device,
+            )
+            # When using cuda graph, we need to set the upper bound of the
+            # number of splits so that large enough intermediate buffers are
+            # pre-allocated during capture.
+            self.max_num_splits = (
+                vllm_config.attention_config.flash_attn_max_num_splits_for_cuda_graph
+            )
+
+        if vllm_is_batch_invariant():
+            self.max_num_splits = 1
+
+    def _schedule_decode(
+        self,
+        num_reqs,
+        cu_query_lens,
+        max_query_len,
+        seqlens,
+        max_seq_len,
+        causal,
+        max_num_splits,
+    ):
+        if self.fa_aot_schedule:
+            return get_scheduler_metadata(
+                batch_size=num_reqs,
+                max_seqlen_q=max_query_len,
+                max_seqlen_k=max_seq_len,
+                num_heads_q=self.num_heads * self.dcp_world_size,
+                num_heads_kv=1,
+                headdim=self.mla_dims.qk_rope_head_dim,
+                cache_seqlens=seqlens,
+                qkv_dtype=self.kv_cache_spec.dtype,
+                headdim_v=self.mla_dims.kv_lora_rank,
+                page_size=self.page_size,
+                cu_seqlens_q=cu_query_lens,
+                causal=causal,
+                num_splits=max_num_splits,
+            )
+        return None
+
+    def _build_decode(
+        self,
+        block_table_tensor: torch.Tensor,
+        seq_lens_device: torch.Tensor,
+        max_seq_len: int,
+        query_start_loc_cpu: torch.Tensor,
+        query_start_loc_device: torch.Tensor,
+        num_decode_tokens: int,
+        dcp_tot_seq_lens_device: torch.Tensor | None,
+    ) -> FlashAttnMLADecodeMetadata:
+        query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
+        max_query_len = query_lens_cpu.max().item()
+
+        # For Flash Attention MLA + full cudagraph
+        max_num_splits = 0
+        if (
+            self.use_full_cuda_graph
+            and self.max_cudagraph_size is not None
+            and num_decode_tokens <= self.max_cudagraph_size
+        ):
+            # NOTE(woosuk): Setting num_splits > 1 may increase the memory
+            # usage, because the intermediate buffers of size [num_splits,
+            # num_heads, num_tokens, head_size] are allocated. Therefore,
+            # we only set num_splits when using cuda graphs.
+            max_num_splits = self.max_num_splits
+
+        if vllm_is_batch_invariant():
+            max_num_splits = 1
+
+        scheduler_metadata = self._schedule_decode(
+            num_reqs=seq_lens_device.shape[0],
+            cu_query_lens=query_start_loc_device,
+            max_query_len=max_query_len,
+            seqlens=seq_lens_device,
+            max_seq_len=max_seq_len,
+            causal=True,
+            max_num_splits=max_num_splits,
+        )
+
+        if self.use_full_cuda_graph and scheduler_metadata is not None:
+            n = scheduler_metadata.shape[0]
+            # Ensure the persistent buffer is large enough
+            assert n <= self.scheduler_metadata.shape[0], (
+                f"Scheduler metadata size {n} exceeds buffer size "
+                f"{self.scheduler_metadata.shape[0]}"
+            )
+            self.scheduler_metadata[:n] = scheduler_metadata
+            # NOTE(woosuk): We should zero out the rest of the scheduler
+            # metadata to guarantee the correctness. Otherwise, some thread
+            # blocks may use the invalid scheduler metadata and overwrite the
+            # output buffer.
+            self.scheduler_metadata[n:] = 0
+            scheduler_metadata = self.scheduler_metadata[:n]
+
+        metadata = FlashAttnMLADecodeMetadata(
+            block_table=block_table_tensor,
+            seq_lens=seq_lens_device,
+            query_start_loc=query_start_loc_device,
+            max_query_len=max_query_len,
+            max_seq_len=max_seq_len,
+            scheduler_metadata=scheduler_metadata,
+            max_num_splits=max_num_splits,
+            dcp_tot_seq_lens=dcp_tot_seq_lens_device,
+        )
+        return metadata
+
+
+class FlashAttnMLAImpl(MLACommonImpl[FlashAttnMLAMetadata]):
+    can_return_lse_for_decode: bool = True
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        **mla_args,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            **mla_args,
+        )
+
+        assert flash_attn_supports_mla(), "FlashAttnMLA is not supported on this device"
+
+        unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "FlashAttnMLAImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, logits_soft_cap"
+            )
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError(
+                "Encoder self-attention and "
+                "encoder/decoder cross-attention "
+                "are not implemented for "
+                "FlashAttnMLAImpl"
+            )
+
+        if is_quantized_kv_cache(self.kv_cache_dtype):
+            raise NotImplementedError(
+                "FlashAttnMLA V1 with FP8 KV cache not yet supported"
+            )
+
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: FlashAttnMLAMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        assert kv_c_and_k_pe_cache.numel() > 0
+        assert attn_metadata.decode is not None
+
+        if type(q) is tuple:
+            q_nope, q_pe = q
+        else:
+            q_nope, q_pe = torch.split(
+                q, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+            )
+
+        if self.kv_cache_dtype.startswith("fp8"):
+            raise NotImplementedError("FP8 FlashAttention MLA not yet supported")
+
+        kv_c_cache = kv_c_and_k_pe_cache[..., : self.kv_lora_rank]
+        k_pe_cache = kv_c_and_k_pe_cache[..., self.kv_lora_rank :]
+
+        # NOTE(matt): During CUDA graph capture, max_query_len can be 0, but the
+        # kernel uses this to calculate grid dimensions. Ensure it's at least 1
+        # to prevent invalid grid configuration during graph capture.
+        max_seqlen_q = max(attn_metadata.decode.max_query_len, 1)
+
+        attn_out = flash_attn_varlen_func(
+            q=q_pe,
+            k=k_pe_cache.unsqueeze(-2),  # Add head dim of 1
+            v=kv_c_cache.unsqueeze(-2),  # Add head dim of 1
+            q_v=q_nope,
+            max_seqlen_q=max_seqlen_q,
+            cu_seqlens_q=attn_metadata.decode.query_start_loc,
+            max_seqlen_k=attn_metadata.decode.max_seq_len,
+            seqused_k=attn_metadata.decode.seq_lens,
+            block_table=attn_metadata.decode.block_table,
+            softmax_scale=self.scale,
+            causal=True,
+            return_softmax_lse=self.need_to_return_lse_for_decode,
+            fa_version=3,  # only version 3 is supported
+            scheduler_metadata=attn_metadata.decode.scheduler_metadata,
+            num_splits=attn_metadata.decode.max_num_splits,
+            cp_world_size=self.dcp_world_size,
+            cp_rank=self.dcp_rank,
+            cp_tot_seqused_k=attn_metadata.decode.dcp_tot_seq_lens,
+        )
+
+        if self.need_to_return_lse_for_decode:
+            o, lse = attn_out
+            # FA returns LSE in shape [ H, B ] but DCP wants [ B, H ]
+            return o, lse.transpose(0, 1)  # [ H, B ] -> [ B, H ]
+        else:
+            o = attn_out
+            return o, None

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

@@ -0,0 +1,202 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import ClassVar
+
+import torch
+from flashinfer.decode import trtllm_batch_decode_with_kv_cache_mla
+
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention.mla_attention import (
+    MLACommonBackend,
+    MLACommonImpl,
+    MLACommonMetadata,
+    MLACommonMetadataBuilder,
+    QueryLenSupport,
+)
+from vllm.platforms.interface import DeviceCapability
+from vllm.v1.attention.backend import (
+    AttentionCGSupport,
+    AttentionLayer,
+    AttentionType,
+    MultipleOf,
+)
+from vllm.v1.attention.backends.utils import KVCacheLayoutType
+
+logger = init_logger(__name__)
+
+FLASHINFER_MLA_WORKSPACE_BUFFER_SIZE = 128 * 1024 * 1024
+
+
+class FlashInferMLAMetadataBuilder(MLACommonMetadataBuilder[MLACommonMetadata]):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH
+    query_len_support: ClassVar[QueryLenSupport] = QueryLenSupport.UNIFORM
+
+
+class FlashInferMLABackend(MLACommonBackend):
+    supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+    supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
+        "auto",
+        "bfloat16",
+        "fp8",
+        "fp8_e4m3",
+    ]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [32, 64]
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASHINFER_MLA"
+
+    @staticmethod
+    def get_impl_cls() -> type["FlashInferMLAImpl"]:
+        return FlashInferMLAImpl
+
+    @staticmethod
+    def get_builder_cls() -> type["FlashInferMLAMetadataBuilder"]:
+        return FlashInferMLAMetadataBuilder
+
+    @classmethod
+    def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
+        return capability.major == 10
+
+    @classmethod
+    def supports_combination(
+        cls,
+        head_size: int,
+        dtype: torch.dtype,
+        kv_cache_dtype: CacheDType | None,
+        block_size: int,
+        use_mla: bool,
+        has_sink: bool,
+        use_sparse: bool,
+        device_capability: DeviceCapability,
+    ) -> str | None:
+        # FlashInfer MLA kernel requires qk_nope_head_dim == 128
+        from vllm.config import get_current_vllm_config
+
+        vllm_config = get_current_vllm_config()
+        if vllm_config.model_config is not None:
+            hf_text_config = vllm_config.model_config.hf_text_config
+            qk_nope_head_dim = getattr(hf_text_config, "qk_nope_head_dim", 1)
+            if qk_nope_head_dim != 128:
+                return (
+                    f"FlashInfer MLA kernel requires qk_nope_head_dim == 128, "
+                    f"but got {qk_nope_head_dim}"
+                )
+        return None
+
+    @classmethod
+    def get_required_kv_cache_layout(cls) -> "KVCacheLayoutType | None":
+        return "HND"
+
+
+g_fi_workspace = torch.zeros(
+    FLASHINFER_MLA_WORKSPACE_BUFFER_SIZE,
+    dtype=torch.uint8,
+    device="cuda",
+)
+
+
+class FlashInferMLAImpl(MLACommonImpl[MLACommonMetadata]):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        **mla_args,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            **mla_args,
+        )
+
+        unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "FlashInferMLAImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, logits_soft_cap"
+            )
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError(
+                "Encoder self-attention and "
+                "encoder/decoder cross-attention "
+                "are not implemented for "
+                "FlashInferMLAImpl"
+            )
+
+        self._workspace_buffer = g_fi_workspace
+        self.bmm1_scale: float | None = None
+        self.bmm2_scale: float | None = None
+
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: MLACommonMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        assert kv_c_and_k_pe_cache.numel() > 0
+        assert attn_metadata.decode is not None
+
+        if isinstance(q, tuple):
+            q_nope, q_pe = q
+            q = torch.cat([q_nope, q_pe], dim=-1)
+
+        # trtllm API requires extra dimension q_len_per_request for MTP
+        if attn_metadata.num_decode_tokens % attn_metadata.num_decodes != 0:
+            logger.warning_once(
+                """FlashInferMLAImpl got a query of uneven length.
+                This usually indicates an issue in batch reordering
+                or incorrect setup in dummy_run."""
+            )
+            q = q.unsqueeze(1)
+        else:
+            q = q.view(attn_metadata.num_decodes, -1, q.shape[-2], q.shape[-1])
+
+        if self.bmm1_scale is None:
+            self.bmm1_scale = layer._q_scale_float * layer._k_scale_float * self.scale
+        if self.bmm2_scale is None:
+            self.bmm2_scale = layer._v_scale_float
+
+        o = trtllm_batch_decode_with_kv_cache_mla(
+            query=q,
+            kv_cache=kv_c_and_k_pe_cache.unsqueeze(1),
+            workspace_buffer=self._workspace_buffer,
+            qk_nope_head_dim=self.qk_nope_head_dim,
+            kv_lora_rank=self.kv_lora_rank,
+            qk_rope_head_dim=self.qk_rope_head_dim,
+            block_tables=attn_metadata.decode.block_table,
+            seq_lens=attn_metadata.decode.seq_lens,
+            max_seq_len=attn_metadata.max_seq_len,
+            bmm1_scale=self.bmm1_scale,
+            bmm2_scale=self.bmm2_scale,
+        )
+
+        # Flatten the output for consistent shape
+        o = o.view(-1, o.shape[-2], o.shape[-1])
+
+        # TODO: Return LSE pending support from Flashinfer API:
+        # https://github.com/flashinfer-ai/flashinfer/pull/1566
+        return o, None

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

@@ -0,0 +1,353 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""FlashInfer MLA Sparse Attention Backend.
+
+This backend uses the FlashInfer TRT-LLM MLA kernel with sparse_mla_top_k
+for models like DeepSeek-V3.2 that use index-based sparse attention.
+
+For sparse MLA:
+- block_tables shape changes from [batch_size, max_num_blocks] (dense)
+  to [batch_size, q_len_per_request, sparse_mla_top_k] (sparse)
+- The sparse indices represent physical cache slot positions to attend to
+- sparse_mla_top_k parameter must be set to the topk value
+"""
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, ClassVar
+
+import numpy as np
+import torch
+from flashinfer.decode import trtllm_batch_decode_with_kv_cache_mla
+
+from vllm.config import VllmConfig
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention.mla_attention import (
+    get_mla_dims,
+)
+from vllm.platforms.interface import DeviceCapability
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionLayer,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+    AttentionType,
+    CommonAttentionMetadata,
+    MultipleOf,
+    SparseMLAAttentionImpl,
+)
+from vllm.v1.attention.backends.mla.sparse_utils import (
+    triton_convert_req_index_to_global_index,
+)
+from vllm.v1.attention.backends.utils import KVCacheLayoutType
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+if TYPE_CHECKING:
+    from vllm.model_executor.models.deepseek_v2 import Indexer
+
+logger = init_logger(__name__)
+
+FLASHINFER_MLA_SPARSE_WORKSPACE_BUFFER_SIZE = 128 * 1024 * 1024
+
+
+class FlashInferMLASparseBackend(AttentionBackend):
+    """FlashInfer MLA backend with sparse attention support.
+
+    This backend uses the FlashInfer TRT-LLM MLA kernel with sparse_mla_top_k
+    for models like DeepSeek-V3.2 that use index-based sparse attention.
+    """
+
+    accept_output_buffer: bool = True
+    supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+    supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
+        "auto",
+        "bfloat16",
+    ]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [32, 64]
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASHINFER_MLA_SPARSE"
+
+    @staticmethod
+    def get_impl_cls() -> type["FlashInferMLASparseImpl"]:
+        return FlashInferMLASparseImpl
+
+    @staticmethod
+    def get_builder_cls() -> type["FlashInferMLASparseMetadataBuilder"]:
+        return FlashInferMLASparseMetadataBuilder
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return [576]
+
+    @classmethod
+    def is_mla(cls) -> bool:
+        return True
+
+    @classmethod
+    def is_sparse(cls) -> bool:
+        return True
+
+    @classmethod
+    def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
+        # FlashInfer sparse MLA targets Blackwell (SM 10.x)
+        return capability.major == 10
+
+    @classmethod
+    def supports_combination(
+        cls,
+        head_size: int,
+        dtype: torch.dtype,
+        kv_cache_dtype: CacheDType | None,
+        block_size: int,
+        use_mla: bool,
+        has_sink: bool,
+        use_sparse: bool,
+        device_capability: DeviceCapability,
+    ) -> str | None:
+        # FlashInfer MLA sparse kernel requires qk_nope_head_dim == 128
+        from vllm.config import get_current_vllm_config
+
+        vllm_config = get_current_vllm_config()
+        if vllm_config.model_config is not None:
+            hf_text_config = vllm_config.model_config.hf_text_config
+            qk_nope_head_dim = getattr(hf_text_config, "qk_nope_head_dim", 1)
+            if qk_nope_head_dim != 128:
+                return (
+                    f"FlashInfer MLA Sparse kernel requires qk_nope_head_dim == 128, "
+                    f"but got {qk_nope_head_dim}"
+                )
+            # Check for index_topk which indicates sparse model
+            if not hasattr(hf_text_config, "index_topk"):
+                return "FlashInfer MLA Sparse requires model with index_topk config"
+        return None
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,  # assumed to be 1 for MLA
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        return (num_blocks, block_size, head_size)
+
+    @classmethod
+    def get_required_kv_cache_layout(cls) -> "KVCacheLayoutType | None":
+        return "HND"
+
+
+@dataclass
+class FlashInferMLASparseMetadata(AttentionMetadata):
+    """Attention metadata for FlashInfer MLA Sparse backend."""
+
+    num_reqs: int
+    max_query_len: int
+    max_seq_len: int
+    num_actual_tokens: int
+
+    # Query start locations
+    query_start_loc: torch.Tensor
+    slot_mapping: torch.Tensor
+    block_table: torch.Tensor
+    req_id_per_token: torch.Tensor
+
+    # Sequence lengths for all requests (context + query)
+    seq_lens: torch.Tensor
+
+    # Sparse-specific
+    block_size: int = 64
+    topk_tokens: int = 2048
+
+
+class FlashInferMLASparseMetadataBuilder(
+    AttentionMetadataBuilder[FlashInferMLASparseMetadata]
+):
+    """Builder for FlashInfer MLA Sparse attention metadata."""
+
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ) -> None:
+        self.vllm_config = vllm_config
+        self.layer_names = layer_names
+        self.kv_cache_spec = kv_cache_spec
+        self.model_config = vllm_config.model_config
+        self.device = device
+
+        self.mla_dims = get_mla_dims(self.model_config)
+        self.topk_tokens = vllm_config.model_config.hf_config.index_topk
+
+        self.req_id_per_token_buffer = torch.empty(
+            (vllm_config.scheduler_config.max_num_batched_tokens,),
+            dtype=torch.int32,
+            device=device,
+        )
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> FlashInferMLASparseMetadata:
+        cm = common_attn_metadata
+        num_tokens = cm.num_actual_tokens
+
+        # Build req_id_per_token mapping
+        starts = np.asarray(cm.query_start_loc_cpu, dtype=np.int32)
+        seg_lengths = np.diff(starts)
+        req_id_per_token = np.repeat(
+            np.arange(seg_lengths.shape[0], dtype=np.int32), seg_lengths
+        )
+
+        # Zero-fill for cudagraphs
+        self.req_id_per_token_buffer.fill_(0)
+        self.req_id_per_token_buffer[: req_id_per_token.shape[0]].copy_(
+            torch.from_numpy(req_id_per_token), non_blocking=True
+        )
+        req_id_per_token_tensor = self.req_id_per_token_buffer[:num_tokens]
+
+        return FlashInferMLASparseMetadata(
+            num_reqs=cm.num_reqs,
+            max_query_len=cm.max_query_len,
+            max_seq_len=cm.max_seq_len,
+            num_actual_tokens=cm.num_actual_tokens,
+            query_start_loc=cm.query_start_loc,
+            slot_mapping=cm.slot_mapping,
+            block_table=cm.block_table_tensor,
+            req_id_per_token=req_id_per_token_tensor,
+            seq_lens=cm.seq_lens,
+            block_size=self.kv_cache_spec.block_size,
+            topk_tokens=self.topk_tokens,
+        )
+
+
+# Global workspace buffer (lazily initialized)
+_fi_sparse_workspace: torch.Tensor | None = None
+
+
+def _get_workspace_buffer(device: torch.device) -> torch.Tensor:
+    global _fi_sparse_workspace
+    if _fi_sparse_workspace is None:
+        _fi_sparse_workspace = torch.zeros(
+            FLASHINFER_MLA_SPARSE_WORKSPACE_BUFFER_SIZE,
+            dtype=torch.uint8,
+            device=device,
+        )
+    return _fi_sparse_workspace
+
+
+class FlashInferMLASparseImpl(SparseMLAAttentionImpl[FlashInferMLASparseMetadata]):
+    """FlashInfer MLA Sparse implementation.
+
+    Uses the TRT-LLM MLA kernel with sparse_mla_top_k parameter for
+    sparse attention computation.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        topk_indice_buffer: torch.Tensor | None = None,
+        indexer: "Indexer | None" = None,
+        **mla_args,
+    ) -> None:
+        unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "FlashInferMLASparseImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, logits_soft_cap"
+            )
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError(
+                "Encoder self-attention and "
+                "encoder/decoder cross-attention "
+                "are not implemented for "
+                "FlashInferMLASparseImpl"
+            )
+
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        self.kv_cache_dtype = kv_cache_dtype
+
+        # MLA-specific dimensions
+        self.kv_lora_rank: int = mla_args["kv_lora_rank"]
+        self.qk_nope_head_dim: int = mla_args["qk_nope_head_dim"]
+        self.qk_rope_head_dim: int = mla_args["qk_rope_head_dim"]
+
+        assert indexer is not None, "Indexer required for sparse MLA"
+        self.topk_indices_buffer: torch.Tensor | None = indexer.topk_indices_buffer
+
+        self._workspace_buffer: torch.Tensor | None = None
+        self.bmm1_scale: float | None = None
+        self.bmm2_scale: float | None = None
+
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: FlashInferMLASparseMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        if isinstance(q, tuple):
+            q = torch.cat(q, dim=-1)
+
+        num_actual_toks = q.shape[0]
+
+        assert self.topk_indices_buffer is not None
+        topk_indices = self.topk_indices_buffer[:num_actual_toks]
+
+        topk_indices_physical, seq_lens = triton_convert_req_index_to_global_index(
+            attn_metadata.req_id_per_token[:num_actual_toks],
+            attn_metadata.block_table,
+            topk_indices,
+            BLOCK_SIZE=attn_metadata.block_size,
+            NUM_TOPK_TOKENS=topk_indices.shape[1],
+            return_valid_counts=True,
+        )
+
+        if self._workspace_buffer is None:
+            self._workspace_buffer = _get_workspace_buffer(q.device)
+
+        if self.bmm1_scale is None:
+            self.bmm1_scale = layer._q_scale_float * layer._k_scale_float * self.scale
+        if self.bmm2_scale is None:
+            self.bmm2_scale = layer._v_scale_float
+
+        o = trtllm_batch_decode_with_kv_cache_mla(
+            query=q.unsqueeze(1),
+            kv_cache=kv_c_and_k_pe_cache.unsqueeze(1),
+            workspace_buffer=self._workspace_buffer,
+            qk_nope_head_dim=self.qk_nope_head_dim,
+            kv_lora_rank=self.kv_lora_rank,
+            qk_rope_head_dim=self.qk_rope_head_dim,
+            block_tables=topk_indices_physical.unsqueeze(1),
+            seq_lens=seq_lens,
+            max_seq_len=attn_metadata.topk_tokens,
+            bmm1_scale=self.bmm1_scale,
+            bmm2_scale=self.bmm2_scale,
+            sparse_mla_top_k=attn_metadata.topk_tokens,
+        )
+        return o.view(-1, o.shape[-2], o.shape[-1]), None

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

@@ -0,0 +1,317 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention.mla_attention import (
+    MLACommonBackend,
+    MLACommonDecodeMetadata,
+    MLACommonImpl,
+    MLACommonMetadata,
+    MLACommonMetadataBuilder,
+    QueryLenSupport,
+)
+from vllm.model_executor.layers.batch_invariant import (
+    vllm_is_batch_invariant,
+)
+from vllm.platforms.interface import DeviceCapability
+from vllm.utils.platform_utils import num_compute_units
+from vllm.v1.attention.backend import (
+    AttentionCGSupport,
+    AttentionLayer,
+    AttentionType,
+    MultipleOf,
+)
+from vllm.v1.attention.backends.utils import (
+    reshape_attn_output_for_spec_decode,
+    reshape_query_for_spec_decode,
+)
+from vllm.v1.attention.ops.flashmla import (
+    FlashMLASchedMeta,
+    flash_mla_with_kvcache,
+    flash_mla_with_kvcache_fp8,
+    get_mla_metadata,
+    get_mla_metadata_dense_fp8,
+    is_flashmla_dense_supported,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+logger = init_logger(__name__)
+
+
+class FlashMLABackend(MLACommonBackend):
+    supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+    supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
+        "auto",
+        "bfloat16",
+        "fp8",
+        "fp8_e4m3",
+    ]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [64]
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASHMLA"
+
+    @staticmethod
+    def get_builder_cls() -> type["FlashMLAMetadataBuilder"]:
+        return FlashMLAMetadataBuilder
+
+    @staticmethod
+    def get_impl_cls() -> type["FlashMLAImpl"]:
+        return FlashMLAImpl
+
+    @classmethod
+    def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
+        return capability.major in [9, 10]
+
+    @classmethod
+    def supports_combination(
+        cls,
+        head_size: int,
+        dtype: torch.dtype,
+        kv_cache_dtype: CacheDType | None,
+        block_size: int,
+        use_mla: bool,
+        has_sink: bool,
+        use_sparse: bool,
+        device_capability: DeviceCapability,
+    ) -> str | None:
+        if use_sparse:
+            from vllm.v1.attention.ops.flashmla import is_flashmla_sparse_supported
+
+            return is_flashmla_sparse_supported()[1]
+        else:
+            from vllm.v1.attention.ops.flashmla import is_flashmla_dense_supported
+
+            return is_flashmla_dense_supported()[1]
+
+
+@dataclass
+class FlashMLADecodeMetadata(MLACommonDecodeMetadata):
+    scheduler_metadata: FlashMLASchedMeta
+
+
+@dataclass
+class FlashMLAMetadata(MLACommonMetadata[FlashMLADecodeMetadata]):
+    pass
+
+
+class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH
+    query_len_support: ClassVar[QueryLenSupport] = QueryLenSupport.UNIFORM
+    reorder_batch_threshold: int = 128  # process small prefills with decode pathway
+    # ^ TODO(matt): tune this
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(
+            kv_cache_spec, layer_names, vllm_config, device, FlashMLAMetadata
+        )
+
+        self.num_q_heads = vllm_config.model_config.get_num_attention_heads(
+            vllm_config.parallel_config
+        )
+
+        self.cg_buf_tile_scheduler_metadata = None
+        self.cg_buf_num_splits = None
+        self.is_fp8_kvcache = vllm_config.cache_config.cache_dtype.startswith("fp8")
+
+        num_sms = num_compute_units(self.device.index)
+
+        if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
+            self.cg_buf_tile_scheduler_metadata = torch.zeros(
+                # Upper bound on size (<= #SMs, TileSchedulerMetaDataSize)
+                # TileSchedulerMetaDataSize = 8
+                (num_sms, 8),
+                device=self.device,
+                dtype=torch.int32,
+            )
+            self.cg_buf_num_splits = torch.empty(
+                (vllm_config.scheduler_config.max_num_seqs + 1),
+                device=self.device,
+                dtype=torch.int32,
+            )
+
+    def _build_decode(
+        self,
+        block_table_tensor: torch.Tensor,
+        seq_lens_device: torch.Tensor,
+        max_seq_len: int,
+        query_start_loc_cpu: torch.Tensor,
+        query_start_loc_device: torch.Tensor,
+        num_decode_tokens: int,
+        dcp_tot_seq_lens_device: torch.Tensor | None,
+    ) -> FlashMLADecodeMetadata:
+        query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
+        # we use the max but all should be the same due to uniform length requirement
+        max_query_len = query_lens_cpu.max().item()
+        num_q_tokens_per_head_k = max_query_len * self.num_q_heads // 1
+        scheduler_metadata, _ = get_mla_metadata(
+            seq_lens_device,
+            num_q_tokens_per_head_k,
+            1,  # MQA for the decode path
+            is_fp8_kvcache=self.is_fp8_kvcache,
+        )
+        if self.is_fp8_kvcache:
+            tile_scheduler_metadata, num_splits = get_mla_metadata_dense_fp8(
+                seq_lens_device,
+                num_q_tokens_per_head_k,
+                1,  # MQA for the decode path
+            )
+            scheduler_metadata.tile_scheduler_metadata = tile_scheduler_metadata
+            scheduler_metadata.num_splits = num_splits
+
+        return FlashMLADecodeMetadata(
+            block_table=block_table_tensor,
+            seq_lens=seq_lens_device,
+            scheduler_metadata=scheduler_metadata,
+            dcp_tot_seq_lens=dcp_tot_seq_lens_device,
+        )
+
+
+class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
+    can_return_lse_for_decode: bool = True
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        **mla_args,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            **mla_args,
+        )
+
+        is_supported, reason = is_flashmla_dense_supported()
+        assert is_supported, reason
+
+        unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "FlashMLAImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, 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_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: FlashMLAMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        # TODO: (zyongye) decode function for mla here
+        assert kv_c_and_k_pe_cache.numel() > 0
+        assert attn_metadata.decode is not None
+
+        if type(q) is tuple:
+            q = torch.cat(q, dim=-1)
+
+        # mypy assertion: q is now always a tensor
+        assert isinstance(q, torch.Tensor)
+
+        num_decodes = attn_metadata.num_decodes
+        q = reshape_query_for_spec_decode(q, num_decodes)
+
+        scheduler_metadata = attn_metadata.decode.scheduler_metadata
+        if vllm_is_batch_invariant() and not self.kv_cache_dtype.startswith("fp8"):
+            device = q.device
+            dtype = torch.int32
+
+            B = q.shape[0]
+            # block_table shape: [batch_size, max_num_blocks_per_seq]
+            # The number of blocks per sequence is in the second dimension
+            topk = attn_metadata.decode.block_table.shape[-1]
+            B_TOPK = 64
+            assert topk % B_TOPK == 0, f"topk ({topk}) must be divisible by {B_TOPK}"
+            end_block_idx = topk // B_TOPK
+
+            # Single partition => num_sm_parts = 1
+            # TileSchedulerMetaDataSize = 8, layout:
+            # [begin_idx, begin_block_idx, end_idx, end_block_idx,
+            #  begin_n_split_idx, _, _, _]
+            tile_scheduler_metadata = torch.zeros((1, 8), dtype=dtype, device=device)
+            tile_scheduler_metadata[0, 0] = 0  # begin_idx
+            tile_scheduler_metadata[0, 1] = 0  # sched_begin_block_idx
+            tile_scheduler_metadata[0, 2] = B - 1  # end_idx
+            tile_scheduler_metadata[0, 3] = end_block_idx
+            tile_scheduler_metadata[0, 4] = 0  # begin_n_split_idx
+            # fields [5..7] stay 0
+
+            # Non-split path ignores num_splits, but the API requires it:
+            # zeros of length B+1
+            num_splits = torch.zeros((B + 1,), dtype=dtype, device=device)
+            scheduler_metadata.tile_scheduler_metadata = tile_scheduler_metadata
+            scheduler_metadata.num_splits = num_splits
+
+        if self.kv_cache_dtype.startswith("fp8"):
+            o, lse = flash_mla_with_kvcache_fp8(
+                q=q,
+                k_cache=kv_c_and_k_pe_cache.unsqueeze(-2),  # Add head dim of 1
+                block_table=attn_metadata.decode.block_table,
+                cache_seqlens=attn_metadata.decode.seq_lens,
+                head_dim_v=self.kv_lora_rank,
+                tile_scheduler_metadata=scheduler_metadata.tile_scheduler_metadata,
+                num_splits=scheduler_metadata.num_splits,
+                softmax_scale=self.scale,
+                causal=True,
+                descale_q=layer._q_scale.reshape(1),
+                descale_k=layer._k_scale.reshape(1),
+            )
+        else:
+            o, lse = flash_mla_with_kvcache(
+                q=q,
+                k_cache=kv_c_and_k_pe_cache.unsqueeze(-2),  # Add head dim of 1
+                block_table=attn_metadata.decode.block_table,
+                cache_seqlens=attn_metadata.decode.seq_lens,
+                head_dim_v=self.kv_lora_rank,
+                tile_scheduler_metadata=scheduler_metadata,
+                softmax_scale=self.scale,
+                causal=True,
+                is_fp8_kvcache=False,
+            )
+
+        o = reshape_attn_output_for_spec_decode(o)
+
+        return o, lse

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

@@ -0,0 +1,847 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, ClassVar
+
+import numpy as np
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.config import VllmConfig, get_current_vllm_config
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention.mla_attention import (
+    get_mla_dims,
+)
+from vllm.platforms import current_platform
+from vllm.platforms.interface import DeviceCapability
+from vllm.utils.platform_utils import num_compute_units
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionLayer,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+    CommonAttentionMetadata,
+    MultipleOf,
+    SparseMLAAttentionImpl,
+)
+from vllm.v1.attention.backends.mla.sparse_utils import (
+    triton_convert_req_index_to_global_index,
+)
+from vllm.v1.attention.backends.utils import (
+    reshape_attn_output_for_spec_decode,
+    reshape_query_for_spec_decode,
+    split_decodes_and_prefills,
+    split_prefill_chunks,
+)
+from vllm.v1.attention.ops.flashmla import (
+    FlashMLASchedMeta,
+    flash_mla_sparse_fwd,
+    flash_mla_with_kvcache,
+    get_mla_metadata,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec
+from vllm.v1.worker.workspace import current_workspace_manager
+
+if TYPE_CHECKING:
+    from vllm.model_executor.models.deepseek_v2 import Indexer
+
+logger = init_logger(__name__)
+
+# For FP8 sparse attention we have two impelementations:
+# 1. Mixed batch mode: use the FP8 decode kernel for both prefill and decode this is
+#    done by treating all tokens as single batch.
+# 2. Separate prefill and decode mode: use the BF16 prefill kernel for prefill
+#    (upconverting the FP8 cache to BF16 then calling the prefill kernel) and using
+#    the FP8 decode kernel for decode.
+# Currently we use #1 when the number of heads per rank is low (i.e. TP) since the BF16
+# prefill kernel requires padding the numer of heads to 128 while the decode does not
+# so when the per ranke head count is below MIN_HEADS_FOR_BF16_PREFILL we use the mixed
+# batch mode (#2).
+MIN_HEADS_FOR_BF16_PREFILL = 32
+
+"""
+NOTE: FlashMLA Sparse uses an fp8 cache with the following format
+
+In the "FP8 with scale" format, each token's KV cache is 656 Bytes,
+structured as:
+-   **First 512 bytes:** The "quantized NoPE" part, containing 512
+    `float8_e4m3` values.
+-   **Next 16 bytes:** Scale factors, containing 4 `float32` values.
+    The first `float32` is the scale for the first 128 `float8_e4m3` values,
+    the second for the next 128, and so on.
+-   **Last 128 bytes:** The "RoPE" part, containing 64 `bfloat16` values. This
+    part is not quantized for accuracy.
+"""
+
+
+class FlashMLASparseBackend(AttentionBackend):
+    accept_output_buffer: bool = True
+    supported_dtypes: ClassVar[list[torch.dtype]] = [torch.bfloat16]
+    supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
+        "auto",
+        "bfloat16",
+        "fp8_ds_mla",
+    ]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [64]
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASHMLA_SPARSE"
+
+    @staticmethod
+    def get_builder_cls() -> type["FlashMLASparseMetadataBuilder"]:
+        return FlashMLASparseMetadataBuilder
+
+    @staticmethod
+    def get_impl_cls() -> type["FlashMLASparseImpl"]:
+        return FlashMLASparseImpl
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return [576]
+
+    @classmethod
+    def is_mla(cls) -> bool:
+        return True
+
+    @classmethod
+    def is_sparse(cls) -> bool:
+        return True
+
+    @classmethod
+    def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
+        return capability.major in [9, 10]
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,  # assumed to be 1 for MLA
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        if cache_dtype_str == "fp8_ds_mla":
+            # custom storage fromat is 656 bytes
+            #  see FlashMLA readme.md for details
+            return (num_blocks, block_size, 656)
+        else:
+            return (num_blocks, block_size, head_size)
+
+
+@dataclass
+class FlashMLASparseMetadata(AttentionMetadata):
+    num_reqs: int
+    max_query_len: int
+    max_seq_len: int
+
+    num_actual_tokens: int  # Number of tokens excluding padding.
+    query_start_loc: torch.Tensor
+    slot_mapping: torch.Tensor
+
+    block_table: torch.Tensor
+    req_id_per_token: torch.Tensor
+    block_size: int = 64
+    topk_tokens: int = 2048
+
+    @dataclass
+    class FP8KernelMetadata:
+        scheduler_metadata: FlashMLASchedMeta
+        dummy_block_table: torch.Tensor
+        cache_lens: torch.Tensor
+
+    @dataclass
+    class FP8SeparatePrefillDecode:
+        @dataclass
+        class Decode:
+            kernel_metadata: "FlashMLASparseMetadata.FP8KernelMetadata"
+            decode_query_len: int  # needed for reshape in spec decode
+
+        @dataclass
+        class Prefill:
+            # Sequence lengths (context + query) for prefill requests
+            # Shape: [num_prefill_reqs]
+            seq_lens: torch.Tensor
+
+            # Request ID for each token: -1 for decode tokens, request index
+            # (0, 1, 2, ...) for prefill tokens.
+            # Shape: [num_actual_tokens]
+            request_ids: torch.Tensor
+
+            # Workspace start offsets for all prefill requests
+            # Shape: [num_prefill_reqs], adjusted in-place per chunk to be
+            # 0-indexed within each chunk. Used to map prefill tokens to workspace
+            # offsets in convert_logical_index_to_physical_index
+            workspace_starts: torch.Tensor
+
+            @dataclass
+            class Chunk:
+                """Metadata for a chunk of prefill requests.
+
+                Prefill requests may be chunked to fit within the fixed workspace size.
+                """
+
+                seq_lens: torch.Tensor
+                tokens_slice: slice
+                block_table: torch.Tensor
+                req_start_idx: int
+                workspace_starts: torch.Tensor
+                chunk_tot_seqlen: int
+
+            chunks: list[Chunk]
+
+        num_prefills: int = 0
+        num_decodes: int = 0
+        num_prefill_tokens: int = 0
+        num_decode_tokens: int = 0
+
+        decode: Decode | None = None
+        prefill: Prefill | None = None
+
+    fp8_extra_metadata: FP8SeparatePrefillDecode | FP8KernelMetadata | None = None
+    fp8_use_mixed_batch: bool = False
+
+
+def get_prefill_workspace_size(max_model_len: int):
+    # NOTE(Lucas): 5 is a magic number for controlling the prefill buffer size.
+    # May be tuned later.
+    # Memory usage: 5 * max_model_len * 576 * 2 bytes
+    #   Example: DeepSeek-V3.2 with max_model_len=163840 ->
+    #            5 * 163840 * 576 * 2 = ~900 MB
+    # This fits nicely below the typical MoE workspace size of >2GB so this is "free"
+    return max_model_len * 5
+
+
+class FlashMLASparseMetadataBuilder(AttentionMetadataBuilder[FlashMLASparseMetadata]):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ) -> None:
+        self.vllm_config = vllm_config
+        self.layer_names = layer_names
+        cache_config = vllm_config.cache_config
+        self.kv_cache_spec = kv_cache_spec
+        self.model_config = vllm_config.model_config
+        parallel_config = vllm_config.parallel_config
+        self.device = device
+
+        # Treat requests with query length <= 1 as decodes to match the
+        # DeepGEMM indexer constraint (fp8_paged_mqa_logits only supports next_n <= 2)
+        self._init_reorder_batch_threshold(1, supports_spec_as_decode=True)
+
+        sm_count = num_compute_units(device.index)
+
+        self.num_heads = self.model_config.get_num_attention_heads(parallel_config)
+        self.mla_dims = get_mla_dims(self.model_config)
+        # FP8 decode kernel only supports h_q = 64 or 128, so we need to pad
+        self.fp8_decode_padded_heads = (
+            FlashMLASparseImpl._compute_fp8_decode_padded_heads(self.num_heads)
+        )
+
+        self.topk_tokens = vllm_config.model_config.hf_config.index_topk
+        self.use_fp8_kv_cache = cache_config.cache_dtype == "fp8_ds_mla"
+        max_num_seqs = vllm_config.scheduler_config.max_num_seqs
+        # Shape: [max_num_seqs], all elements = topk_tokens (constant for full-CG)
+        self.topk_tokens_tensor = torch.full(
+            (max_num_seqs,), self.topk_tokens, device=device, dtype=torch.int32
+        )
+        # Shape: [max_num_seqs], all elements = max_model_len
+        self.max_model_len_tensor = torch.full(
+            (max_num_seqs,),
+            self.model_config.max_model_len,
+            device=device,
+            dtype=torch.int32,
+        )
+        # this is ignored by `flash_mla_with_kvcache` if indices not None
+        self.dummy_block_table = torch.empty(
+            (max_num_seqs, 1), dtype=torch.int32, device=self.device
+        )
+
+        # Equation taken from FlashMLA/csrc/api/sparse_decode.h
+        # For sparse FP8 decode, the formula depends on architecture:
+        # - SM90 (Hopper): num_sm_parts = num_sms / s_q / (h_q/64)
+        # - SM100 (Blackwell head64/head64x2): num_sm_parts = num_sms / s_q
+        # - SM100 (Blackwell head128): num_sm_parts = num_sms / s_q / 2
+        # For max buffer size, use s_q = 1 (the case that produces largest output)
+        # Use padded head count since that's what will be passed to the kernel
+        h_q = self.fp8_decode_padded_heads
+        if current_platform.is_device_capability_family(100):
+            # SM100 head64 or head64x2 uses full SM count
+            max_num_sm_parts = sm_count
+        else:
+            # SM90 uses h_q/64 divisor
+            max_num_sm_parts = sm_count // max(1, h_q // 64)
+        self.tile_scheduler_metadata_buffer = torch.empty(
+            # TileSchedulerMetaDataSize = 8
+            # see: FlashMLA/csrc/params.h
+            (max_num_sm_parts, 8),
+            dtype=torch.int32,
+            device=device,
+        )
+        # Sized for per-request batching (num_decodes + 1)
+        self.num_splits_buffer = torch.empty(
+            (max_num_seqs + 1,),
+            dtype=torch.int32,
+            device=device,
+        )
+        self.req_id_per_token_buffer = torch.empty(
+            (vllm_config.scheduler_config.max_num_batched_tokens,),
+            dtype=torch.int32,
+            device=device,
+        )
+
+    def _build_fp8_mixed_decode_prefill(
+        self,
+        common_attn_metadata: CommonAttentionMetadata,
+    ) -> "FlashMLASparseMetadata.FP8KernelMetadata":
+        """Build FP8 metadata treating all tokens as one mixed batch.
+
+        This matches main branch's approach and avoids the BF16 prefill kernel
+        which has head padding overhead when num_heads is small (high TP case).
+        """
+        num_tokens = common_attn_metadata.num_actual_tokens
+
+        # Use padded head count since that's what the kernel will see
+        padded_heads = self.fp8_decode_padded_heads
+
+        # Build metadata for all tokens as a single batch
+        scheduler_metadata, _ = get_mla_metadata(
+            cache_seqlens=self.topk_tokens_tensor[:1],  # Single batch
+            num_q_tokens_per_head_k=num_tokens * padded_heads,
+            topk=self.topk_tokens,
+            num_heads_q=padded_heads,
+            num_heads_k=1,
+            is_fp8_kvcache=True,
+        )
+
+        fp8_metadata = FlashMLASparseMetadata.FP8KernelMetadata(
+            scheduler_metadata=scheduler_metadata,
+            cache_lens=self.max_model_len_tensor[:1],
+            dummy_block_table=self.dummy_block_table[:1],
+        )
+
+        return fp8_metadata
+
+    def _build_fp8_separate_prefill_decode(
+        self,
+        common_attn_metadata: CommonAttentionMetadata,
+    ) -> "FlashMLASparseMetadata.FP8SeparatePrefillDecode":
+        num_tokens = common_attn_metadata.num_actual_tokens
+
+        (num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens) = (
+            split_decodes_and_prefills(
+                common_attn_metadata,
+                decode_threshold=self.reorder_batch_threshold or 1,
+                require_uniform=True,
+            )
+        )
+
+        FP8Meta = FlashMLASparseMetadata.FP8SeparatePrefillDecode
+        fp8_metadata = FP8Meta(
+            num_decodes=num_decodes,
+            num_prefills=num_prefills,
+            num_decode_tokens=num_decode_tokens,
+            num_prefill_tokens=num_prefill_tokens,
+        )
+
+        # Extract prefill sequence lengths (context + query, not just query)
+        # Decode requests come first in the batch, prefill requests follow
+        prefill_seq_lens = None
+        prefill_request_id = None
+        prefill_workspace_starts = None
+        prefill_chunks = None
+
+        # For pure decode batches, prefill_request_id will be None
+        # For mixed batches, it will have -1 for decode and request_id for prefill
+        if num_prefills > 0:
+            seq_lens_cpu = common_attn_metadata.seq_lens.cpu()
+            seq_lens = common_attn_metadata.seq_lens
+            query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
+
+            prefill_seq_lens_cpu = seq_lens_cpu[num_decodes:]
+            prefill_seq_lens = seq_lens[num_decodes:]
+
+            # Build prefill_request_id: -1 for decode, request index for
+            # prefill. This enables a single
+            # convert_logical_index_to_physical_index call for all tokens
+            prefill_request_id = torch.full(
+                (num_tokens,), -1, dtype=torch.int32, device=self.device
+            )
+            # Map prefill tokens to their request IDs (0, 1, 2, ...)
+            for req_idx in range(num_prefills):
+                # Get query token range for this prefill request
+                global_req_idx = num_decodes + req_idx
+                req_query_start = query_start_loc_cpu[global_req_idx]
+                req_query_end = query_start_loc_cpu[global_req_idx + 1]
+                prefill_request_id[req_query_start:req_query_end] = req_idx
+
+            # will be adjusted by chunk loop
+            prefill_workspace_starts_cpu = torch.zeros(
+                num_prefills, dtype=torch.int32, pin_memory=True
+            )
+            prefill_workspace_starts_cpu[1:] = torch.cumsum(
+                prefill_seq_lens_cpu[:-1], dim=0
+            )
+            # populated by non-blocking copy after prefill_workspace_starts_cpu is
+            # updated by each chunk
+            prefill_workspace_starts = torch.empty(
+                num_prefills, dtype=torch.int32, device=self.device
+            )
+
+            # Chunk prefill requests to fit within workspace size
+            max_prefill_buffer_size = get_prefill_workspace_size(
+                self.vllm_config.model_config.max_model_len
+            )
+            chunk_bounds = split_prefill_chunks(
+                prefill_seq_lens_cpu, max_prefill_buffer_size
+            )
+
+            prefill_chunks = []
+            for chunk_start, chunk_end in chunk_bounds:
+                # Adjust workspace_starts in-place per chunk to be
+                # 0-indexed within each chunk
+                # Example: seq_lens=[10,15,20,5], chunks=[[0,2],[2,4]]
+                #   Initial: workspace_starts=[0,10,25,45]
+                #   After:   workspace_starts=[0,10,0,20]
+                #           (chunk 0 starts at 0, chunk 1 starts at 0)
+                offset = prefill_workspace_starts_cpu[chunk_start].item()
+                prefill_workspace_starts_cpu[chunk_start:chunk_end] -= offset
+
+                chunk_seq_lens = prefill_seq_lens[chunk_start:chunk_end]
+                chunk_tot_seqlen = prefill_seq_lens_cpu[chunk_start:chunk_end].sum()
+                token_start = query_start_loc_cpu[num_decodes + chunk_start].item()
+                token_end = query_start_loc_cpu[num_decodes + chunk_end].item()
+                tokens_slice = slice(token_start, token_end)
+
+                # Create chunk view of gpu tensor
+                chunk_workspace_starts = prefill_workspace_starts[chunk_start:chunk_end]
+                chunk_block_table = common_attn_metadata.block_table_tensor[
+                    num_decodes + chunk_start : num_decodes + chunk_end
+                ]
+
+                prefill_chunks.append(
+                    FP8Meta.Prefill.Chunk(
+                        seq_lens=chunk_seq_lens,
+                        tokens_slice=tokens_slice,
+                        block_table=chunk_block_table,
+                        req_start_idx=chunk_start,
+                        workspace_starts=chunk_workspace_starts,
+                        chunk_tot_seqlen=chunk_tot_seqlen,
+                    )
+                )
+
+            prefill_workspace_starts.copy_(
+                prefill_workspace_starts_cpu, non_blocking=True
+            )
+
+            fp8_metadata.prefill = FP8Meta.Prefill(
+                seq_lens=prefill_seq_lens,
+                request_ids=prefill_request_id,
+                workspace_starts=prefill_workspace_starts,
+                chunks=prefill_chunks,
+            )
+
+        if num_decodes > 0:
+            # Compute decode_query_len for spec decode (uniform due to require_uniform)
+            query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
+            decode_query_len = (query_start_loc_cpu[1] - query_start_loc_cpu[0]).item()
+
+            # Use padded head count since that's what the kernel will see
+            padded_heads = self.fp8_decode_padded_heads
+            scheduler_metadata, _ = get_mla_metadata(
+                cache_seqlens=self.topk_tokens_tensor[:num_decodes],
+                num_q_tokens_per_head_k=decode_query_len * padded_heads,
+                topk=self.topk_tokens,
+                num_heads_q=padded_heads,
+                num_heads_k=1,
+                is_fp8_kvcache=True,
+            )
+
+            kernel_meta = FlashMLASparseMetadata.FP8KernelMetadata(
+                scheduler_metadata=scheduler_metadata,
+                dummy_block_table=self.dummy_block_table[:num_decodes],
+                cache_lens=self.max_model_len_tensor[:num_decodes],
+            )
+            fp8_metadata.decode = FP8Meta.Decode(
+                kernel_metadata=kernel_meta,
+                decode_query_len=decode_query_len,
+            )
+
+        return fp8_metadata
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> FlashMLASparseMetadata:
+        cm = common_attn_metadata
+        num_tokens = cm.num_actual_tokens
+        starts = np.asarray(cm.query_start_loc_cpu, dtype=np.int32)
+        seg_lengths = np.diff(starts)
+        req_id_per_token = np.repeat(
+            np.arange(seg_lengths.shape[0], dtype=np.int32), seg_lengths
+        )
+        # Zero-fill for cudagraphs
+        self.req_id_per_token_buffer.fill_(0)
+        self.req_id_per_token_buffer[: req_id_per_token.shape[0]].copy_(
+            torch.from_numpy(req_id_per_token), non_blocking=True
+        )
+        req_id_per_token = self.req_id_per_token_buffer[:num_tokens]
+
+        fp8_extra_metadata: (
+            FlashMLASparseMetadata.FP8SeparatePrefillDecode
+            | FlashMLASparseMetadata.FP8KernelMetadata
+            | None
+        ) = None
+        fp8_use_mixed_batch = self.num_heads < MIN_HEADS_FOR_BF16_PREFILL
+        if self.use_fp8_kv_cache:
+            if fp8_use_mixed_batch:
+                fp8_extra_metadata = self._build_fp8_mixed_decode_prefill(cm)
+            else:
+                fp8_extra_metadata = self._build_fp8_separate_prefill_decode(cm)
+
+        metadata = FlashMLASparseMetadata(
+            num_reqs=cm.num_reqs,
+            max_query_len=cm.max_query_len,
+            max_seq_len=cm.max_seq_len,
+            num_actual_tokens=cm.num_actual_tokens,
+            query_start_loc=cm.query_start_loc,
+            slot_mapping=cm.slot_mapping,
+            block_table=cm.block_table_tensor,
+            req_id_per_token=req_id_per_token,
+            block_size=self.kv_cache_spec.block_size,
+            topk_tokens=self.topk_tokens,
+            fp8_extra_metadata=fp8_extra_metadata,
+            fp8_use_mixed_batch=fp8_use_mixed_batch,
+        )
+
+        return metadata
+
+
+class FlashMLASparseImpl(SparseMLAAttentionImpl[FlashMLASparseMetadata]):
+    @staticmethod
+    def _compute_fp8_decode_padded_heads(num_heads: int) -> int:
+        # FP8 decode kernel only supports h_q = 64 or 128
+        # Compute padded head count for decode
+        return 64 if num_heads <= 64 else 128
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        topk_indice_buffer: torch.Tensor | None = None,
+        indexer: "Indexer | None" = None,
+        **mla_args,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        self.kv_cache_dtype = kv_cache_dtype
+        self.kv_lora_rank: int = mla_args["kv_lora_rank"]
+        self.softmax_scale = scale
+        assert indexer is not None
+        self.topk_indices_buffer: torch.Tensor | None = indexer.topk_indices_buffer
+        # Prefill BF16 kernel requires 64 on Hopper, 128 on Blackwell
+        self.prefill_padding = (
+            128 if current_platform.is_device_capability_family(100) else 64
+        )
+        self.fp8_decode_padded_heads = self._compute_fp8_decode_padded_heads(num_heads)
+
+        if kv_cache_dtype == "fp8_ds_mla":
+            # Reserve workspace during initialization
+            vllm_config = get_current_vllm_config()
+            assert vllm_config is not None and vllm_config.model_config is not None
+            prefill_workspace_size = get_prefill_workspace_size(
+                vllm_config.model_config.max_model_len
+            )
+            self.prefill_workspace_shape = (prefill_workspace_size, head_size)
+            (self.prefill_bf16_workspace,) = (
+                current_workspace_manager().get_simultaneous(
+                    (self.prefill_workspace_shape, torch.bfloat16)
+                )
+            )
+
+    def _forward_bf16_kv(
+        self,
+        q: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        topk_indices: torch.Tensor,
+        attn_metadata: FlashMLASparseMetadata,
+    ) -> torch.Tensor:
+        # Convert per-request indices to global slots (decode) or workspace
+        # offsets (prefill).
+        topk_indices = triton_convert_req_index_to_global_index(
+            attn_metadata.req_id_per_token,
+            attn_metadata.block_table,
+            topk_indices,
+            BLOCK_SIZE=attn_metadata.block_size,
+            NUM_TOPK_TOKENS=topk_indices.shape[1],
+        )
+
+        return self._bf16_flash_mla_kernel(q, kv_c_and_k_pe_cache, topk_indices)
+
+    def _forward_fp8_kv_separate_prefill_decode(
+        self,
+        q: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        topk_indices: torch.Tensor,
+        attn_metadata: FlashMLASparseMetadata,
+    ) -> torch.Tensor:
+        fp8_metadata = attn_metadata.fp8_extra_metadata
+        assert isinstance(fp8_metadata, FlashMLASparseMetadata.FP8SeparatePrefillDecode)
+        num_decodes = fp8_metadata.num_decodes
+
+        prefill_request_ids = None
+        prefill_workspace_starts = None
+        has_prefill_workspace = False
+        if fp8_metadata.prefill is not None:
+            prefill_request_ids = fp8_metadata.prefill.request_ids
+            prefill_workspace_starts = fp8_metadata.prefill.workspace_starts
+            has_prefill_workspace = True
+
+        # Convert per-request indices to global slots (decode) or workspace
+        # offsets (prefill).
+        # For FP8 cache: prefill uses workspace mapping (upconverted to BF16)
+        # For BF16 cache: always use global cache slots (no workspace)
+        # prefill_workspace_starts has been adjusted in-place per chunk so
+        # prefill indices automatically come out chunk-local
+        topk_indices = triton_convert_req_index_to_global_index(
+            attn_metadata.req_id_per_token,
+            attn_metadata.block_table,
+            topk_indices,
+            BLOCK_SIZE=attn_metadata.block_size,
+            NUM_TOPK_TOKENS=topk_indices.shape[1],
+            HAS_PREFILL_WORKSPACE=has_prefill_workspace,
+            prefill_workspace_request_ids=prefill_request_ids,
+            prefill_workspace_starts=prefill_workspace_starts,
+        )
+
+        fp8_metadata = attn_metadata.fp8_extra_metadata
+        assert isinstance(fp8_metadata, FlashMLASparseMetadata.FP8SeparatePrefillDecode)
+
+        def _fp8_decode(q: torch.Tensor, topk_indices: torch.Tensor) -> torch.Tensor:
+            # Reshape q: (num_decode_tokens, num_heads, head_dim)
+            #         -> (num_decodes, seq_len, num_heads, head_dim)
+            q = reshape_query_for_spec_decode(q, num_decodes)
+            seq_len = q.shape[1]
+            # Reshape topk_indices: (num_decode_tokens, topk)
+            #                    -> (num_decodes, seq_len, topk)
+            topk_indices = topk_indices.view(num_decodes, seq_len, -1)
+            assert fp8_metadata.decode is not None
+            attn_out, _ = self._fp8_flash_mla_kernel(
+                q=q,
+                kv_c_and_k_pe_cache=kv_c_and_k_pe_cache,
+                topk_indices=topk_indices,
+                kernel_metadata=fp8_metadata.decode.kernel_metadata,
+            )
+            # Reshape output: (num_decodes, seq_len, num_heads, head_dim_v)
+            #              -> (num_decode_tokens, num_heads, head_dim_v)
+            return reshape_attn_output_for_spec_decode(attn_out)
+
+        num_decode_tokens = fp8_metadata.num_decode_tokens
+        num_prefill_tokens = fp8_metadata.num_prefill_tokens
+
+        # Pure decode: direct call without allocation
+        if num_decode_tokens > 0 and num_prefill_tokens == 0:
+            assert fp8_metadata.decode is not None
+            attn_out = _fp8_decode(q, topk_indices)
+        else:
+            # Mixed or pure prefill: allocate output tensor
+            attn_out = q.new_empty(
+                (attn_metadata.num_actual_tokens, self.num_heads, self.kv_lora_rank),
+                dtype=q.dtype,
+                device=q.device,
+            )
+
+            if num_decode_tokens > 0:
+                attn_out[:num_decode_tokens] = _fp8_decode(
+                    q[:num_decode_tokens], topk_indices[:num_decode_tokens]
+                )
+
+            assert fp8_metadata.prefill is not None
+            for chunk in fp8_metadata.prefill.chunks:
+                chunk_workspace = self.prefill_bf16_workspace[: chunk.chunk_tot_seqlen]
+                ops.cp_gather_and_upconvert_fp8_kv_cache(
+                    kv_c_and_k_pe_cache,
+                    chunk_workspace,
+                    chunk.block_table,
+                    chunk.seq_lens,
+                    chunk.workspace_starts,
+                    len(chunk.block_table),
+                )
+
+                chunk_q = q[chunk.tokens_slice]
+                chunk_topk_indices_workspace = topk_indices[chunk.tokens_slice]
+
+                attn_out[chunk.tokens_slice] = self._bf16_flash_mla_kernel(
+                    chunk_q,
+                    chunk_workspace,
+                    chunk_topk_indices_workspace,
+                )
+
+        return attn_out
+
+    def _forward_fp8_kv_mixed_batch(
+        self,
+        q: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        topk_indices: torch.Tensor,
+        attn_metadata: FlashMLASparseMetadata,
+    ) -> torch.Tensor:
+        """Mixed batch FP8 forward path that treats all tokens as one batch.
+
+        This is equivalent to main branch's approach and avoids the BF16
+        prefill kernel which has head padding overhead when num_heads is small.
+        Used when use_mixed_batch is True.
+        """
+        # Convert per-request indices to global slots (decode) or workspace
+        # offsets (prefill).
+        topk_indices = triton_convert_req_index_to_global_index(
+            attn_metadata.req_id_per_token,
+            attn_metadata.block_table,
+            topk_indices,
+            BLOCK_SIZE=attn_metadata.block_size,
+            NUM_TOPK_TOKENS=topk_indices.shape[1],
+        )
+
+        assert attn_metadata.fp8_extra_metadata is not None
+        assert isinstance(
+            attn_metadata.fp8_extra_metadata, FlashMLASparseMetadata.FP8KernelMetadata
+        )
+        fp8_metadata = attn_metadata.fp8_extra_metadata
+
+        _attn_out, _ = self._fp8_flash_mla_kernel(
+            q=q.unsqueeze(0),  # unsqueeze to add batch_dim: (T, H, D) -> (1, T, H, D)
+            kv_c_and_k_pe_cache=kv_c_and_k_pe_cache,
+            topk_indices=topk_indices.unsqueeze(0),  # (T, topk) -> (1, T, topk)
+            kernel_metadata=fp8_metadata,
+        )
+
+        # Output is (1, T, H, D_v), squeeze back to (T, H, D_v)
+        return _attn_out.squeeze(0)
+
+    def _fp8_flash_mla_kernel(
+        self,
+        q: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        topk_indices: torch.Tensor,
+        kernel_metadata: FlashMLASparseMetadata.FP8KernelMetadata,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        # q shape: (batch, seq_len, num_heads, head_dim)
+        actual_num_heads = q.size(2)
+        padded_num_heads = self.fp8_decode_padded_heads
+
+        # Pad query if needed (kernel only supports h_q = 64 or 128)
+        if actual_num_heads < padded_num_heads:
+            logger.warning_once(
+                f"Padding num_heads from {actual_num_heads} to "
+                f"{padded_num_heads} for FP8 sparse decode kernel"
+            )
+            q_padded = q.new_zeros((q.size(0), q.size(1), padded_num_heads, q.size(3)))
+            q_padded[:, :, :actual_num_heads, :] = q
+            q = q_padded
+
+        out, lse = flash_mla_with_kvcache(
+            q=q,
+            k_cache=kv_c_and_k_pe_cache.view(torch.uint8).unsqueeze(-2),
+            block_table=kernel_metadata.dummy_block_table,
+            head_dim_v=512,
+            cache_seqlens=kernel_metadata.cache_lens,
+            tile_scheduler_metadata=kernel_metadata.scheduler_metadata,
+            is_fp8_kvcache=True,
+            indices=topk_indices,
+            softmax_scale=self.softmax_scale,
+        )
+
+        # Slice output back to actual head count if we padded
+        if actual_num_heads < padded_num_heads:
+            out = out[:, :, :actual_num_heads, :]
+
+        return out, lse
+
+    def _bf16_flash_mla_kernel(
+        self,
+        q: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        topk_indices: torch.Tensor,
+    ) -> torch.Tensor:
+        num_tokens = q.shape[0]
+        kv_c_and_k_pe_cache = kv_c_and_k_pe_cache.view(
+            -1, 1, kv_c_and_k_pe_cache.shape[-1]
+        )
+
+        # NOTE(Chen): kernel requires num_local_head to be a multiple of
+        # 64 on hopper and 128 on blackwell
+        if self.num_heads % self.prefill_padding != 0:
+            assert self.prefill_padding % self.num_heads == 0
+            logger.warning_once(
+                f"Padding num_heads from {self.num_heads} to "
+                f"{self.prefill_padding} for BF16 sparse prefill kernel"
+            )
+            q_padded = q.new_empty((q.shape[0], self.prefill_padding, q.shape[2]))
+            q_padded[:, : self.num_heads, :] = q
+            q = q_padded
+
+        topk_indices = topk_indices.view(num_tokens, 1, -1)
+        output = flash_mla_sparse_fwd(
+            q, kv_c_and_k_pe_cache, topk_indices, self.softmax_scale
+        )[0]
+        output = output[:, : self.num_heads, :]
+        return output
+
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: FlashMLASparseMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        # NOTE(lucas): for the sparse FlashMLA kernels the kernels want to use
+        # MQA 576/512 approach for both prefill and decode
+
+        # Concatenate q if it's a tuple (ql_nope, q_pe)
+        if isinstance(q, tuple):
+            q = torch.cat(q, dim=-1)
+
+        num_actual_toks = q.shape[0]
+
+        # Get topk indices
+        assert self.topk_indices_buffer is not None
+        topk_indices = self.topk_indices_buffer[:num_actual_toks]
+
+        use_fp8_cache = self.kv_cache_dtype == "fp8_ds_mla"
+
+        if not use_fp8_cache:
+            attn_out = self._forward_bf16_kv(
+                q, kv_c_and_k_pe_cache, topk_indices, attn_metadata
+            )
+        elif attn_metadata.fp8_use_mixed_batch:
+            attn_out = self._forward_fp8_kv_mixed_batch(
+                q, kv_c_and_k_pe_cache, topk_indices, attn_metadata
+            )
+        else:
+            attn_out = self._forward_fp8_kv_separate_prefill_decode(
+                q, kv_c_and_k_pe_cache, topk_indices, attn_metadata
+            )
+
+        return attn_out, None

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

@@ -0,0 +1,386 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+from vllm.utils.deep_gemm import get_paged_mqa_logits_metadata, has_deep_gemm
+from vllm.utils.platform_utils import num_compute_units
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionMetadataBuilder,
+    CommonAttentionMetadata,
+    MultipleOf,
+)
+from vllm.v1.attention.backends.utils import (
+    split_decodes_and_prefills,
+    split_prefill_chunks,
+)
+
+logger = init_logger(__name__)
+
+
+class DeepseekV32IndexerBackend(AttentionBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "DEEPSEEK_V32_INDEXER"
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [1 if current_platform.is_rocm() else 64]
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return [32, 64, 128]
+
+    @staticmethod
+    def get_builder_cls() -> type["DeepseekV32IndexerMetadataBuilder"]:
+        return DeepseekV32IndexerMetadataBuilder
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        assert num_kv_heads == 1
+        return (num_blocks, block_size, head_size)
+
+    @staticmethod
+    def get_kv_cache_stride_order(
+        include_num_layers_dimension: bool = False,
+    ) -> tuple[int, ...]:
+        if include_num_layers_dimension:
+            return (0, 1, 2, 3)
+        return (0, 1, 2)
+
+
+@dataclass
+class DeepseekV32IndexerPrefillChunkMetadata:
+    block_table: torch.Tensor
+    cu_seqlen_ks: torch.Tensor
+    cu_seqlen_ke: torch.Tensor
+    cu_seq_lens: torch.Tensor
+    token_to_seq: torch.Tensor
+    total_seq_lens: int
+    token_start: int
+    token_end: int
+    num_reqs: int
+
+
+@dataclass
+class DeepseekV32IndexerPrefillMetadata:
+    chunks: list[DeepseekV32IndexerPrefillChunkMetadata]
+
+
+@dataclass
+class DeepSeekV32IndexerDecodeMetadata:
+    block_table: torch.Tensor
+    seq_lens: torch.Tensor
+    decode_lens: torch.Tensor
+    requires_padding: bool
+    schedule_metadata: torch.Tensor
+    use_large_context_topk: bool
+    offsets: torch.Tensor | None  # Precomputed offsets for speculative decoding
+
+
+@dataclass
+class DeepseekV32IndexerMetadata:
+    # FIXME (zyongye)
+    # hacky way to access the data now, need to be in chunked meta
+    seq_lens: torch.Tensor
+
+    num_reqs: int
+    max_query_len: int
+    max_seq_len: int
+
+    num_actual_tokens: int  # Number of tokens excluding padding.
+    query_start_loc: torch.Tensor
+    slot_mapping: torch.Tensor
+    # The dimension of the attention heads
+    head_dim: int
+
+    # New for MLA (compared to FlashAttention)
+    # For handling prefill decode split
+    num_decodes: int
+    num_decode_tokens: int
+    num_prefills: int
+    num_prefill_tokens: int
+
+    decode: DeepSeekV32IndexerDecodeMetadata | None = None
+    prefill: DeepseekV32IndexerPrefillMetadata | None = None
+
+
+# TODO (zyongye) optimize this, this is now vibe coded
+def kv_spans_from_batches(
+    start_seq_loc: torch.Tensor, seq_len_per_batch: torch.Tensor, device: torch.device
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Args:
+      start_seq_loc: 1D long tensor [B+1], cumulative counts of
+                     selected tokens per batch.
+            Example: [0, 2, 4, 7] ->
+                     batch sizes (selected) [2, 2, 3], N=7 tokens total.
+      seq_len_per_batch: 1D long tensor [B],
+                         full sequence length (KV length) of each batch.
+                         Example: [5, 9, 4].
+
+    Returns:
+      start_tensor: 1D long tensor [N], start offset in the
+                    concatenated KV cache for each token's batch.
+      end_location: 1D long tensor [N],
+                    **exclusive** end = start + token's local position.
+                    (So the attended KV slice is kv[start:end].)
+
+    Assumes each batch contributes its full `seq_len_per_batch[i]`
+    keys to the KV cache, andthe selected tokens within a batch
+    are the **last** `counts[i]` positions of that sequence.
+    """
+    q = start_seq_loc.to(dtype=torch.long)
+    L = seq_len_per_batch.to(dtype=torch.long)
+    assert q.dim() == 1 and L.dim() == 1
+    assert q.numel() == L.numel() + 1, "start_seq_loc must have length B+1"
+
+    # Selected tokens per batch and totals
+    counts = q[1:] - q[:-1]  # [B]
+    N = int(q[-1].item())  # total selected tokens
+    B = L.numel()
+
+    if N == 0:
+        return (
+            torch.empty(0, dtype=torch.long, device=device),
+            torch.empty(0, dtype=torch.long, device=device),
+        )
+
+    # KV start offsets per batch in the concatenated KV cache
+    kv_starts_per_batch = torch.cumsum(L, dim=0) - L  # [B]
+
+    # For each selected token, which batch does it belong to?
+    batch_id = torch.repeat_interleave(torch.arange(B), counts)  # [N]
+
+    # Map batch KV start to each token
+    start_tensor = kv_starts_per_batch[batch_id]  # [N]
+
+    # End-align local positions inside each batch:
+    # local_pos = L[b] - counts[b] + (1..counts[b])  for each batch b
+    L_expand = torch.repeat_interleave(L, counts)  # [N]
+    m_expand = torch.repeat_interleave(counts, counts)  # [N]
+    # position within the selected block: 1..counts[b]
+    pos_within = (
+        torch.arange(N, dtype=torch.long) - torch.repeat_interleave(q[:-1], counts) + 1
+    )
+
+    local_pos = L_expand - m_expand + pos_within  # [N], 1-based
+    end_location = start_tensor + local_pos  # exclusive end
+
+    return start_tensor.int().to(device), end_location.int().to(device)
+
+
+def get_max_prefill_buffer_size(vllm_config: VllmConfig):
+    max_model_len = vllm_config.model_config.max_model_len
+    # NOTE(Chen): 40 is a magic number for controlling the prefill buffer size.
+    # Each entry is 128 fp8 bytes and 4 scale bytes for a total of 132 bytes.
+    # The flashmla_sparse backend uses a workspace size of 5 * max_model_len.
+    # The memory usage of the workspace there is 576 * 2 bytes; so we size this as
+    # (576 * 2 // 132) * 5 = 40 to maximize this workspace size while still fitting
+    # within the flashmla_sparse workspace.
+    # For DeepSeek-V3.2, the max_model_len is 163840.
+    #   40 * 163840 * 132 = 865075200 bytes = 825 MB
+    return max_model_len * 40
+
+
+class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH
+
+    reorder_batch_threshold: int = 1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        scheduler_config = self.vllm_config.scheduler_config
+        # NOTE(Chen):an estimated max size of flattened_kv. Need to double check.
+        self.max_prefill_buffer_size = get_max_prefill_buffer_size(self.vllm_config)
+        self.num_speculative_tokens = (
+            self.vllm_config.speculative_config.num_speculative_tokens
+            if self.vllm_config.speculative_config
+            else 0
+        )
+        if self.num_speculative_tokens > 1:
+            raise ValueError(
+                "Sparse MLA only supports "
+                "num_speculative_tokens <= 1 because the DeepGEMM "
+                "fp8_paged_mqa_logits kernel does not support next_n > 2. "
+                f"Got num_speculative_tokens={self.num_speculative_tokens}."
+            )
+        self.reorder_batch_threshold += self.num_speculative_tokens
+
+        sm_count = num_compute_units(self.device.index)
+        self.num_sms = sm_count
+
+        self.decode_lens_buffer = torch.empty(
+            (scheduler_config.max_num_seqs,), dtype=torch.int32, device=self.device
+        )
+
+        # See: DeepGMM/csrc/apis/attention.hpp
+        self.scheduler_metadata_buffer = torch.empty(
+            (self.num_sms + 1, 2), dtype=torch.int32, device=self.device
+        )
+
+    def build_one_prefill_chunk(
+        self, reqs_start, reqs_end, query_start_loc_cpu, seq_lens_cpu, block_table
+    ):
+        prefill_query_start_loc = (
+            query_start_loc_cpu[reqs_start : reqs_end + 1]
+            - query_start_loc_cpu[reqs_start]
+        )
+        cu_seqlen_ks, cu_seqlen_ke = kv_spans_from_batches(
+            prefill_query_start_loc, seq_lens_cpu[reqs_start:reqs_end], self.device
+        )
+        token_start = query_start_loc_cpu[reqs_start].item()
+        token_end = query_start_loc_cpu[reqs_end].item()
+        total_seq_lens = seq_lens_cpu[reqs_start:reqs_end].sum()
+        seq_idx = torch.arange(0, reqs_end - reqs_start, dtype=torch.int32)
+        token_to_seq = torch.repeat_interleave(
+            seq_idx, seq_lens_cpu[reqs_start:reqs_end]
+        ).to(self.device)
+        assert total_seq_lens <= self.max_prefill_buffer_size
+        cu_seq_lens = (
+            torch.cat(
+                [
+                    torch.zeros(1, dtype=torch.int32),
+                    seq_lens_cpu[reqs_start:reqs_end].cumsum(dim=0),
+                ]
+            )
+            .to(torch.int32)
+            .to(self.device)
+        )
+        return DeepseekV32IndexerPrefillChunkMetadata(
+            cu_seqlen_ks=cu_seqlen_ks,
+            cu_seqlen_ke=cu_seqlen_ke,
+            cu_seq_lens=cu_seq_lens,
+            token_to_seq=token_to_seq,
+            total_seq_lens=total_seq_lens,
+            block_table=block_table[reqs_start:reqs_end],
+            token_start=token_start,
+            token_end=token_end,
+            num_reqs=reqs_end - reqs_start,
+        )
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> DeepseekV32IndexerMetadata:
+        num_reqs = common_attn_metadata.num_reqs
+        num_tokens = common_attn_metadata.num_actual_tokens
+
+        query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
+        num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = (
+            split_decodes_and_prefills(
+                common_attn_metadata, decode_threshold=self.reorder_batch_threshold
+            )
+        )
+
+        assert num_decodes + num_prefills == num_reqs
+        assert num_decode_tokens + num_prefill_tokens == num_tokens
+
+        prefill_metadata = None
+        if num_prefills > 0:
+            chunk_seq_ids = split_prefill_chunks(
+                common_attn_metadata.seq_lens_cpu[num_decodes:],
+                self.max_prefill_buffer_size,
+                request_offset=num_decodes,
+            )
+            chunks = [
+                self.build_one_prefill_chunk(
+                    reqs_start,
+                    reqs_end,
+                    query_start_loc_cpu,
+                    common_attn_metadata.seq_lens_cpu,
+                    common_attn_metadata.block_table_tensor,
+                )
+                for reqs_start, reqs_end in chunk_seq_ids
+            ]
+            prefill_metadata = DeepseekV32IndexerPrefillMetadata(
+                chunks=chunks,
+            )
+
+        decode_metadata = None
+        if num_decodes > 0:
+            torch.diff(
+                common_attn_metadata.query_start_loc[: num_decodes + 1],
+                out=self.decode_lens_buffer[:num_decodes],
+            )
+            decode_lens = self.decode_lens_buffer[:num_decodes]
+            decode_lens_cpu = torch.diff(
+                common_attn_metadata.query_start_loc_cpu[: num_decodes + 1]
+            )
+
+            # Use CPU to avoid GPU sync; breaking async scheduling
+            requires_padding = (decode_lens_cpu.max() > decode_lens_cpu.min()).item()
+
+            # Decide which top-k kernel to use based on batch size and sequence length
+            batch_size = num_decodes
+            _is_large_context = common_attn_metadata.max_seq_len > 8192
+
+            # Decision logic based on micro-benchmark results:
+            # - large_context_topk wins for batch <= 128 and seq_len > 8K
+            # - top_k_per_row_decode wins for batch > 128 or seq_len <= 8K
+            use_large_context_topk = batch_size <= 128 and _is_large_context
+
+            next_n = 1 + self.num_speculative_tokens
+            if next_n > 1:
+                offsets = torch.arange(next_n, device=self.device, dtype=torch.int32)
+            else:
+                offsets = None
+
+            seq_lens = common_attn_metadata.seq_lens[:num_decodes]
+
+            # DeepGEMM is required for the paged MQA logits on CUDA devices
+            if current_platform.is_cuda() and has_deep_gemm():
+                self.scheduler_metadata_buffer[:] = get_paged_mqa_logits_metadata(
+                    seq_lens, self.kv_cache_spec.block_size, self.num_sms
+                )
+            block_table = common_attn_metadata.block_table_tensor[:num_decodes, ...]
+            # Padded CUDA graph requests have block_table entries of -1.
+            # Clamp to 0 to prevent OOB access in the DeepGEMM kernel.
+            # This is safe because padded requests have seq_lens=0, so the
+            # kernel produces no meaningful output for those rows.
+            block_table.clamp_(min=0)
+            decode_metadata = DeepSeekV32IndexerDecodeMetadata(
+                block_table=block_table,
+                seq_lens=common_attn_metadata.seq_lens[:num_decodes],
+                decode_lens=decode_lens,
+                requires_padding=requires_padding,
+                schedule_metadata=self.scheduler_metadata_buffer,
+                use_large_context_topk=use_large_context_topk,
+                offsets=offsets,
+            )
+
+        attn_metadata = DeepseekV32IndexerMetadata(
+            seq_lens=common_attn_metadata.seq_lens,
+            num_reqs=common_attn_metadata.num_reqs,
+            max_query_len=common_attn_metadata.max_query_len,
+            max_seq_len=common_attn_metadata.max_seq_len,
+            num_actual_tokens=common_attn_metadata.num_actual_tokens,
+            query_start_loc=common_attn_metadata.query_start_loc,
+            slot_mapping=common_attn_metadata.slot_mapping,
+            head_dim=128,
+            num_decodes=num_decodes,
+            num_decode_tokens=num_decode_tokens,
+            num_prefills=num_prefills,
+            num_prefill_tokens=num_prefill_tokens,
+            prefill=prefill_metadata,
+            decode=decode_metadata,
+        )
+
+        # if get_tensor_model_parallel_rank() == 0:
+        #     logger.info(f"attn_metadata: {attn_metadata}")
+        return attn_metadata

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

@@ -0,0 +1,284 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm._aiter_ops import rocm_aiter_ops
+from vllm.config import VllmConfig
+from vllm.model_executor.layers.attention.mla_attention import (
+    MLACommonBackend,
+    MLACommonDecodeMetadata,
+    MLACommonImpl,
+    MLACommonMetadata,
+    MLACommonMetadataBuilder,
+    QueryLenSupport,
+)
+from vllm.v1.attention.backend import AttentionCGSupport, AttentionLayer, MultipleOf
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+
+class AiterMLABackend(MLACommonBackend):
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [1]
+
+    @staticmethod
+    def get_name() -> str:
+        return "ROCM_AITER_MLA"
+
+    @staticmethod
+    def get_impl_cls() -> type["AiterMLAImpl"]:
+        return AiterMLAImpl
+
+    @staticmethod
+    def get_builder_cls() -> type["AiterMLAMetadataBuilder"]:
+        return AiterMLAMetadataBuilder
+
+
+@dataclass
+class AiterMLADecodeMetadata(MLACommonDecodeMetadata):
+    # The indptr of the paged kv cache, shape: [batch_size + 1]
+    paged_kv_indptr: torch.Tensor | None = None
+    # The page indices of the paged kv cache
+    paged_kv_indices: torch.Tensor | None = None
+    # The number of entries in the last page of each request in
+    # the paged kv cache, shape: [batch_size]
+    paged_kv_last_page_len: torch.Tensor | None = None
+    # The query indptr, shape : [num_decode + 1]
+    qo_indptr: torch.Tensor | None = None
+    # The dtype of MLA out tensor
+    attn_out_dtype: torch.dtype = torch.bfloat16
+    # The max query output length: int
+    max_qo_len: int | None = None
+
+
+class AiterMLAMetadata(MLACommonMetadata[AiterMLADecodeMetadata]):
+    pass
+
+
+class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
+    # TODO(luka, lucas): audit this as part of:
+    #  https://github.com/vllm-project/vllm/issues/22945
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH
+    query_len_support: ClassVar[QueryLenSupport] = QueryLenSupport.UNIFORM
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(
+            kv_cache_spec, layer_names, vllm_config, device, AiterMLAMetadata
+        )
+
+        self.compilation_config = vllm_config.compilation_config
+        self.decode_attn_out_dtype = vllm_config.model_config.dtype
+        # kernel block size is always 1.
+        max_num_pages_per_req = vllm_config.model_config.max_model_len
+        max_num_reqs = vllm_config.scheduler_config.max_num_seqs
+        max_num_pages = max_num_reqs * max_num_pages_per_req
+
+        # Preparing persistent buffers
+        # TODO: we can disambiguate between decode and mixed-prefill decode here
+        # so we can only use the persistent buffer if a cudagraph is actually
+        # being used.
+
+        # paged_kv_last_page_len is always 1s (kernel block size is always 1),
+        # so we create it once and reuse slices in both eager and cudagraph modes.
+        self.paged_kv_last_page_len = torch.ones(
+            max_num_reqs, dtype=torch.int32, device=device
+        )
+
+        if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
+            self.paged_kv_indptr = torch.zeros(
+                max_num_reqs + 1, dtype=torch.int32, device=device
+            )
+            self.paged_kv_indices = torch.zeros(
+                max_num_pages, dtype=torch.int32, device=device
+            )
+
+            self.qo_indptr = torch.zeros(
+                max_num_reqs + 1, dtype=torch.int32, device=device
+            )
+
+    def _build_decode(
+        self,
+        block_table_tensor: torch.Tensor,
+        seq_lens_device: torch.Tensor,
+        max_seq_len: int,
+        query_start_loc_cpu: torch.Tensor,
+        query_start_loc_device: torch.Tensor,
+        num_decode_tokens: int,
+        dcp_tot_seq_lens_device: torch.Tensor | None,
+    ) -> AiterMLADecodeMetadata:
+        # kernel block size is always 1, although the kv block size is not 1.
+        device = self.device
+        num_reqs = seq_lens_device.size(0)
+
+        mask = torch.arange(
+            block_table_tensor.size(1), dtype=block_table_tensor.dtype, device=device
+        ).unsqueeze(0) < seq_lens_device.unsqueeze(1)
+        paged_kv_indices = block_table_tensor[mask]
+
+        # kernel block size is always 1, so each page has exactly 1 token.
+        # last_page_len is always 1 - just slice the pre-initialized buffer.
+        paged_kv_last_page_len = self.paged_kv_last_page_len[:num_reqs]
+
+        paged_kv_indptr = torch.cat(
+            [
+                torch.zeros(1, dtype=seq_lens_device.dtype, device=device),
+                seq_lens_device.cumsum(dim=0, dtype=torch.int32),
+            ]
+        )
+        qo_len = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
+        max_qo_len = qo_len.max().item()
+
+        if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
+            num_actual_pages = paged_kv_indices.size(0)
+
+            self.paged_kv_indices[:num_actual_pages].copy_(
+                paged_kv_indices, non_blocking=True
+            )
+            self.paged_kv_indices[num_actual_pages:].fill_(-1)
+            paged_kv_indices = self.paged_kv_indices[:num_actual_pages]
+
+            self.paged_kv_indptr[: 1 + num_reqs].copy_(
+                paged_kv_indptr, non_blocking=True
+            )
+            self.paged_kv_indptr[1 + num_reqs :].fill_(paged_kv_indptr[-1])
+            paged_kv_indptr = self.paged_kv_indptr[: 1 + num_reqs]
+
+            # paged_kv_last_page_len already uses the pre-initialized buffer slice
+            # (set above), so no copy needed - buffer is always 1s.
+
+            self.qo_indptr[: 1 + num_reqs].copy_(
+                query_start_loc_device, non_blocking=True
+            )
+            self.qo_indptr[1 + num_reqs :] = query_start_loc_device[-1]
+            qo_indptr = self.qo_indptr[: 1 + num_reqs]
+
+        else:
+            qo_indptr = torch.arange(
+                0, num_reqs + 1, step=1, dtype=torch.int32, device=device
+            )
+
+        attn_metadata = AiterMLADecodeMetadata(
+            block_table=block_table_tensor,
+            seq_lens=seq_lens_device,
+            paged_kv_indptr=paged_kv_indptr,
+            paged_kv_indices=paged_kv_indices,
+            paged_kv_last_page_len=paged_kv_last_page_len,
+            qo_indptr=qo_indptr,
+            dcp_tot_seq_lens=dcp_tot_seq_lens_device,
+            max_qo_len=max_qo_len,
+            attn_out_dtype=self.decode_attn_out_dtype,
+        )
+
+        return attn_metadata
+
+
+class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        **mla_args,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            **mla_args,
+        )
+        assert num_heads == 16 or num_heads == 128, (
+            f"Aiter MLA only supports 16 or 128 number of heads.\n"
+            f"Provided {num_heads} number of heads.\n"
+            "Try adjusting tensor_parallel_size value."
+        )
+        unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "Aiter MLA does not support one of the following: "
+                "alibi_slopes, sliding_window, logits_soft_cap"
+            )
+
+        from aiter import flash_attn_varlen_func
+
+        self.flash_attn_varlen_func = flash_attn_varlen_func
+
+    def _flash_attn_varlen_diff_headdims(
+        self, q, k, v, return_softmax_lse=False, softmax_scale=None, **kwargs
+    ):
+        output = self.flash_attn_varlen_func(  # type: ignore[call-arg]
+            q=q,
+            k=k,
+            v=v,
+            softmax_scale=softmax_scale,
+            return_lse=return_softmax_lse,
+            **kwargs,
+        )
+
+        return output
+
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: AiterMLAMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        assert kv_c_and_k_pe_cache.numel() > 0
+        assert attn_metadata.decode is not None
+        assert attn_metadata.decode.max_qo_len is not None
+
+        if type(q) is tuple:
+            q = torch.cat(q, dim=-1)
+
+        assert isinstance(q, torch.Tensor)
+        B = q.shape[0]
+        o = torch.zeros(
+            B,
+            self.num_heads,
+            self.kv_lora_rank,
+            dtype=attn_metadata.decode.attn_out_dtype,
+            device=q.device,
+        )
+
+        kv_buffer = kv_c_and_k_pe_cache.unsqueeze(2)
+
+        rocm_aiter_ops.mla_decode_fwd(
+            q,
+            kv_buffer,
+            o,
+            self.scale,
+            attn_metadata.decode.qo_indptr,
+            attn_metadata.decode.max_qo_len,
+            attn_metadata.decode.paged_kv_indptr,
+            attn_metadata.decode.paged_kv_indices,
+            attn_metadata.decode.paged_kv_last_page_len,
+            q_scale=layer._q_scale,
+            kv_scale=layer._k_scale,
+        )
+
+        return o, None

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

@@ -0,0 +1,368 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, ClassVar
+
+import numpy as np
+import torch
+
+from vllm._aiter_ops import rocm_aiter_ops
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention.mla_attention import (
+    get_mla_dims,
+)
+from vllm.triton_utils import tl, triton
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionLayer,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+    CommonAttentionMetadata,
+    SparseMLAAttentionImpl,
+)
+from vllm.v1.attention.backends.mla.flashmla_sparse import (
+    triton_convert_req_index_to_global_index,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+if TYPE_CHECKING:
+    from vllm.model_executor.models.deepseek_v2 import Indexer
+logger = init_logger(__name__)
+
+
+@triton.jit
+def fetch_id_to_ragged_kernel(
+    in_tensor_ptr,  # [num_seq, topk]
+    cumsum_ptr,  # [num_seq + 1]
+    out_tensor_ptr,  # [max_num_seq * topk]
+    in_tensor_ptr_stride,
+    TOPK: tl.constexpr,
+    TOKEN_NUM: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    seq_id = tl.program_id(0)
+    block_id = tl.program_id(1)
+    offset = tl.arange(0, BLOCK_SIZE)
+    token_start = tl.load(cumsum_ptr + seq_id)
+    token_end = tl.load(cumsum_ptr + seq_id + 1)
+    token_num = token_end - token_start
+    row_offset = block_id * BLOCK_SIZE
+    if row_offset >= token_num:
+        return
+    in_tensor_offset = seq_id * in_tensor_ptr_stride + row_offset + offset
+    in_tensor_mask = (row_offset + offset) < TOPK
+    in_tensor_val = tl.load(in_tensor_ptr + in_tensor_offset, mask=in_tensor_mask)
+    out_tensor_offset = token_start + row_offset + offset
+    out_tensor_mask = (out_tensor_offset < token_end) & in_tensor_mask
+    tl.store(out_tensor_ptr + out_tensor_offset, in_tensor_val, mask=out_tensor_mask)
+
+
+def fetch_id_to_ragged_triton(
+    in_tensor: torch.Tensor, cumsum: torch.Tensor, out_tensor: torch.Tensor, topk
+):
+    num_tokens = in_tensor.size(0)
+    block_size = 64
+    num_block_per_row = triton.cdiv(topk, block_size)
+    grid = (
+        num_tokens,
+        num_block_per_row,
+    )
+    fetch_id_to_ragged_kernel[grid](
+        in_tensor, cumsum, out_tensor, in_tensor.stride(0), topk, num_tokens, block_size
+    )
+
+
+class ROCMAiterMLASparseBackend(AttentionBackend):
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_name() -> str:
+        return "ROCM_AITER_MLA_SPARSE"
+
+    @staticmethod
+    def get_metadata_cls() -> type["ROCMAiterMLASparseMetadata"]:
+        return ROCMAiterMLASparseMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["ROCMAiterMLASparseMetadataBuilder"]:
+        return ROCMAiterMLASparseMetadataBuilder
+
+    @staticmethod
+    def get_impl_cls() -> type["ROCMAiterMLASparseImpl"]:
+        return ROCMAiterMLASparseImpl
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,  # assumed to be 1 for MLA
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        return (num_blocks, block_size, head_size)
+
+    @classmethod
+    def get_supported_dtypes(cls) -> list[torch.dtype]:
+        return [torch.bfloat16]
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return [576]
+
+
+@dataclass
+class ROCMAiterMLASparseMetadata(AttentionMetadata):
+    num_reqs: int
+    max_query_len: int
+    max_seq_len: int
+
+    num_actual_tokens: int  # Number of tokens excluding padding.
+    query_start_loc: torch.Tensor
+    slot_mapping: torch.Tensor
+
+    block_table: torch.Tensor
+    req_id_per_token: torch.Tensor
+
+    qo_indptr: torch.Tensor
+    paged_kv_last_page_len: torch.Tensor
+    paged_kv_indices: torch.Tensor
+    paged_kv_indptr: torch.Tensor
+    paged_kv_indptr_rest: torch.Tensor
+
+    block_size: int = 1
+    topk_tokens: int = 2048
+
+
+@dataclass
+class ROCMAiterMLASparseMetadataBuilder(
+    AttentionMetadataBuilder[ROCMAiterMLASparseMetadata]
+):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.NEVER
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        self.kv_cache_spec = kv_cache_spec
+        self.model_config = vllm_config.model_config
+        parallel_config = vllm_config.parallel_config
+        self.device = device
+        max_num_batched_tokens = vllm_config.scheduler_config.max_num_batched_tokens
+
+        self.num_heads = self.model_config.get_num_attention_heads(parallel_config)
+        self.mla_dims = get_mla_dims(self.model_config)
+        self.topk_tokens = vllm_config.model_config.hf_config.index_topk
+        self.topk_tokens_tensor = torch.tensor(
+            [self.topk_tokens], device=device, dtype=torch.int32
+        )
+        self.max_model_len_tensor = torch.tensor(
+            [self.model_config.max_model_len], device=device, dtype=torch.int32
+        )
+        # this is ignored by `flash_mla_with_kvcache` if indices not None
+        self.dummy_block_table = torch.empty(
+            (1, 1), dtype=torch.int32, device=self.device
+        )
+
+        self.req_id_per_token_buffer = torch.empty(
+            (vllm_config.scheduler_config.max_num_batched_tokens,),
+            dtype=torch.int32,
+            device=device,
+        )
+        self.qo_indptr = torch.arange(
+            0, max_num_batched_tokens + 1, dtype=torch.int32, device=device
+        )
+        self.paged_kv_last_page_len = torch.ones(
+            max_num_batched_tokens, dtype=torch.int32, device=device
+        )
+
+        # These two needs to be calculated in runtime,
+        # but we still needs to prepare the buffer
+        self.paged_kv_indices = torch.zeros(
+            [max_num_batched_tokens * self.topk_tokens],
+            dtype=torch.int32,
+            device=device,
+        )
+        self.paged_kv_indptr = torch.zeros(
+            [max_num_batched_tokens + 1], dtype=torch.int32, device=device
+        )
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> ROCMAiterMLASparseMetadata:
+        num_tokens = common_attn_metadata.num_actual_tokens
+        starts = np.asarray(common_attn_metadata.query_start_loc_cpu, dtype=np.int32)
+        seg_lengths = np.diff(starts)
+        req_id_per_token = np.repeat(
+            np.arange(seg_lengths.shape[0], dtype=np.int32), seg_lengths
+        )
+        # Zero-fill for cudagraphs
+        self.req_id_per_token_buffer.fill_(0)
+        self.req_id_per_token_buffer[: req_id_per_token.shape[0]].copy_(
+            torch.from_numpy(req_id_per_token), non_blocking=True
+        )
+        self.paged_kv_indices.fill_(0)
+        self.paged_kv_indptr.fill_(0)
+
+        req_id_per_token = self.req_id_per_token_buffer[:num_tokens]
+        qo_indptr = self.qo_indptr[: num_tokens + 1]
+        paged_kv_last_page_len = self.paged_kv_last_page_len[:num_tokens]
+        paged_kv_indices = self.paged_kv_indices[: num_tokens * self.topk_tokens]
+        paged_kv_indptr = self.paged_kv_indptr[: num_tokens + 1]
+        paged_kv_indptr_rest = self.paged_kv_indptr[num_tokens + 1 :]
+
+        metadata = ROCMAiterMLASparseMetadata(
+            num_reqs=common_attn_metadata.num_reqs,
+            max_query_len=common_attn_metadata.max_query_len,
+            max_seq_len=common_attn_metadata.max_seq_len,
+            num_actual_tokens=common_attn_metadata.num_actual_tokens,
+            query_start_loc=common_attn_metadata.query_start_loc,
+            slot_mapping=common_attn_metadata.slot_mapping,
+            block_table=common_attn_metadata.block_table_tensor,
+            req_id_per_token=req_id_per_token,
+            block_size=self.kv_cache_spec.block_size,
+            topk_tokens=self.topk_tokens,
+            qo_indptr=qo_indptr,
+            paged_kv_last_page_len=paged_kv_last_page_len,
+            paged_kv_indices=paged_kv_indices,
+            paged_kv_indptr=paged_kv_indptr,
+            paged_kv_indptr_rest=paged_kv_indptr_rest,
+        )
+        return metadata
+
+
+# Take from
+# https://github.com/deepseek-ai/FlashMLA/blob/main/tests/test_flash_mla_prefill.py#L72
+def reference_mla_sparse_prefill(
+    q: torch.Tensor, kv: torch.Tensor, indices: torch.Tensor, sm_scale: float, d_v: int
+) -> tuple[torch.Tensor, torch.Tensor]:
+    import math
+
+    def log2sumexp2(a: torch.Tensor, dim: int) -> torch.Tensor:
+        return torch.logsumexp(a * math.log(2), dim=dim) * math.log2(math.e)
+
+    skv = kv.shape[0]
+    sq = q.shape[0]
+    topk = indices.shape[-1]
+    dqk = q.shape[-1]
+    indices = indices[:, 0, :]  # [s_q, topk]
+    invalid_indices_mask = (indices < 0) | (indices >= skv)
+    indices[invalid_indices_mask] = 0
+    qs = q  # [s_q, h_q, d_qk]
+    kvs = kv[:, 0, :][indices].view(sq, topk, dqk)  # [s_q, topk, d_qk]
+
+    attn_score = (qs @ kvs.transpose(1, 2)).float()  # [s_q, h_q, topk]
+    attn_score.masked_fill_(invalid_indices_mask.unsqueeze(1), float("-inf"))
+    attn_score *= sm_scale * math.log2(math.e)
+    lse = log2sumexp2(attn_score, dim=-1)  # [s_q, h_q]
+    attn_score = torch.exp2(attn_score - lse.unsqueeze(-1))  # [s_q, h_q, topk]
+    result = attn_score.to(q.dtype) @ kvs[:, :, :d_v]
+    return (result, lse)
+
+
+class ROCMAiterMLASparseImpl(SparseMLAAttentionImpl[ROCMAiterMLASparseMetadata]):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        topk_indice_buffer: torch.Tensor | None = None,
+        indexer: "Indexer | None" = None,
+        **mla_args,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        self.kv_cache_dtype = kv_cache_dtype
+        self.kv_lora_rank: int = mla_args["kv_lora_rank"]
+        self.softmax_scale = scale
+        assert indexer is not None
+        self.topk_indices_buffer: torch.Tensor | None = indexer.topk_indices_buffer
+
+    def _forward_bf16_kv(
+        self,
+        q: torch.Tensor,  # [sq, heads, d_qk]
+        kv_c_and_k_pe_cache: torch.Tensor,  # [blocks, heads, d_qk]
+        topk_indices: torch.Tensor,  # [sq, topk]
+        attn_metadata: ROCMAiterMLASparseMetadata,
+    ) -> torch.Tensor:
+        num_tokens = q.shape[0]
+        output = torch.empty(
+            [num_tokens, self.num_heads, self.kv_lora_rank],
+            dtype=q.dtype,
+            device=q.device,
+        )
+        seq_len = (topk_indices != -1).sum(dim=-1)
+        torch.cumsum(seq_len, dim=0, out=attn_metadata.paged_kv_indptr[1:])
+        attn_metadata.paged_kv_indptr_rest.fill_(attn_metadata.paged_kv_indptr[-1])
+        fetch_id_to_ragged_triton(
+            topk_indices,
+            attn_metadata.paged_kv_indptr,
+            attn_metadata.paged_kv_indices,
+            attn_metadata.topk_tokens,
+        )
+
+        rocm_aiter_ops.mla_decode_fwd(
+            q,
+            kv_c_and_k_pe_cache,
+            output,
+            self.scale,
+            attn_metadata.qo_indptr,
+            1,
+            attn_metadata.paged_kv_indptr,
+            attn_metadata.paged_kv_indices,
+            attn_metadata.paged_kv_last_page_len,
+        )
+
+        return output[:, : self.num_heads, :]
+
+    def forward_mqa(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: ROCMAiterMLASparseMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        # NOTE(lucas): for the sparse FlashMLA kernels the kernels want to use
+        # MQA 576/512 approach for both prefill and decode
+
+        # Concatenate q if it's a tuple (ql_nope, q_pe)
+        if isinstance(q, tuple):
+            q = torch.cat(q, dim=-1)
+
+        num_actual_toks = q.shape[0]
+
+        # Get topk indices
+        assert self.topk_indices_buffer is not None
+        topk_indices = self.topk_indices_buffer[:num_actual_toks]
+
+        topk_indices_global = triton_convert_req_index_to_global_index(
+            attn_metadata.req_id_per_token,
+            attn_metadata.block_table,
+            topk_indices,
+            BLOCK_SIZE=attn_metadata.block_size,
+            NUM_TOPK_TOKENS=attn_metadata.topk_tokens,
+        )
+
+        attn_out = self._forward_bf16_kv(
+            q, kv_c_and_k_pe_cache, topk_indices_global, attn_metadata
+        )
+
+        return attn_out, None

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

@@ -0,0 +1,191 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Utility functions for sparse MLA backends."""
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+
+# Kernel with prefill workspace support and valid count tracking
+@triton.jit
+def _convert_req_index_to_global_index_kernel(
+    req_id_ptr,  # int32 [num_tokens]
+    block_table_ptr,  # int32 [num_requests, max_num_blocks_per_req]
+    token_indices_ptr,  # int32 [num_tokens, NUM_TOPK_TOKENS]
+    out_ptr,  # int32 [num_tokens, NUM_TOPK_TOKENS]
+    valid_count_ptr,  # int32 [num_tokens] - output valid count per row
+    prefill_request_id_ptr,  # int32 [num_tokens], -1 for decode, >=0 for prefill
+    workspace_starts_ptr,  # int32 [num_prefill_reqs+1] or nullptr
+    # shapes (compile-time where possible)
+    max_num_blocks_per_req: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    BLOCK_N: tl.constexpr,  # tile width along columns
+    HAS_PREFILL: tl.constexpr,
+    COUNT_VALID: tl.constexpr,  # whether to count valid indices
+    # strides (in elements)
+    bt_stride0,
+    bt_stride1,
+    ti_stride0,
+    ti_stride1,
+    out_stride0,
+    out_stride1,
+):
+    # program_id(0) -> token_id (row)
+    # program_id(1) -> tile index along columns
+    token_id = tl.program_id(0)
+    tile_id = tl.program_id(1)
+
+    # Each program covers BLOCK_N consecutive columns
+    indice_id = tile_id * BLOCK_N + tl.arange(0, BLOCK_N)
+
+    # Load request id for this token (no mask: grid is exact)
+    req = tl.load(req_id_ptr + token_id)
+
+    # Load token indices for this tile
+    ti_ptr = token_indices_ptr + token_id * ti_stride0 + indice_id * ti_stride1
+    tok = tl.load(ti_ptr)  # int32
+
+    # Only token == -1 should propagate as -1
+    is_invalid_tok = tok < 0
+    is_prefill = False
+    if HAS_PREFILL:
+        prefill_req_id = tl.load(prefill_request_id_ptr + token_id)
+        is_prefill = prefill_req_id >= 0
+    # Compute block id and in-block offset
+    block_id = tok // BLOCK_SIZE
+    inblock_off = tok % BLOCK_SIZE
+
+    # Guard block_table access
+    valid_block = (block_id < max_num_blocks_per_req) & (block_id >= 0)
+    bt_ptr = block_table_ptr + req * bt_stride0 + block_id * bt_stride1
+    is_invalid_tok |= ~valid_block
+    base = tl.load(bt_ptr, mask=valid_block & ~is_prefill, other=0)
+    out_val = base * BLOCK_SIZE + inblock_off
+
+    # Override with prefill output if prefill is enabled
+    if HAS_PREFILL:
+        workspace_start = tl.load(
+            workspace_starts_ptr + prefill_req_id, mask=is_prefill, other=0
+        )
+        prefill_out = workspace_start + tok
+        out_val = tl.where(is_prefill, prefill_out, out_val)
+    out_val = tl.where(is_invalid_tok, -1, out_val)
+
+    # Store results
+    out_ptr_ij = out_ptr + token_id * out_stride0 + indice_id * out_stride1
+    tl.store(out_ptr_ij, out_val)
+
+    # Count valid indices in this tile and atomically add to row total
+    if COUNT_VALID:
+        tile_valid_count = tl.sum((~is_invalid_tok).to(tl.int32))
+        tl.atomic_add(valid_count_ptr + token_id, tile_valid_count)
+
+
+def triton_convert_req_index_to_global_index(
+    req_id: torch.Tensor,  # int32 [num_tokens]
+    block_table: torch.Tensor,  # int32 [num_requests, max_num_blocks_per_req]
+    token_indices: torch.Tensor,  # int32 [num_tokens, NUM_TOPK_TOKENS]
+    BLOCK_SIZE: int = 64,
+    NUM_TOPK_TOKENS: int = 2048,
+    BLOCK_N: int = 128,  # tile width along columns
+    HAS_PREFILL_WORKSPACE: bool = False,
+    prefill_workspace_request_ids: torch.Tensor | None = None,
+    prefill_workspace_starts: torch.Tensor | None = None,
+    return_valid_counts: bool = False,
+) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+    """
+    out[token_id, indice_id] =
+        block_table[req_id[token_id],
+            token_indices[token_id, indice_id] // BLOCK_SIZE] * BLOCK_SIZE
+        + token_indices[token_id, indice_id] % BLOCK_SIZE
+
+    Only when token_indices[token_id, indice_id] == -1 do we output -1.
+    For safety, we also output -1 if the derived block_id would be
+        out-of-bounds.
+
+    When HAS_PREFILL_WORKSPACE is True, prefill tokens are mapped to workspace offsets
+    instead of global cache slots. prefill_workspace_request_ids and
+    prefill_workspace_starts must be provided.
+
+    prefill_workspace_request_ids: int32 [num_tokens], -1 for decode else
+        prefill request index (maps to prefill_workspace_starts)
+    prefill_workspace_starts: int32 [num_prefills], 0-indexed workspace
+        starts for each prefill request
+
+    When return_valid_counts is True, also returns the count of valid (non -1)
+    indices per row, computed during the same kernel pass (no extra overhead).
+    """
+    assert req_id.dtype == torch.int32
+    assert block_table.dtype == torch.int32
+    assert token_indices.dtype == torch.int32
+    assert token_indices.shape[1] == NUM_TOPK_TOKENS
+    assert NUM_TOPK_TOKENS % BLOCK_N == 0, (
+        f"NUM_TOPK_TOKENS ({NUM_TOPK_TOKENS}) must be divisible by BLOCK_N ({BLOCK_N})"
+    )
+
+    if HAS_PREFILL_WORKSPACE:
+        assert prefill_workspace_request_ids is not None
+        assert prefill_workspace_starts is not None
+        assert prefill_workspace_request_ids.dtype == torch.int32
+        assert prefill_workspace_starts.dtype == torch.int32
+
+    num_tokens = req_id.shape[0]
+    max_num_blocks_per_req = block_table.shape[1]
+    tiles_per_row = NUM_TOPK_TOKENS // BLOCK_N
+
+    # Ensure contiguous tensors on the same device
+    req_id_c = req_id.contiguous()
+    block_table_c = block_table.contiguous()
+    token_indices_c = token_indices.contiguous()
+    out = torch.empty_like(token_indices_c)
+
+    # Allocate valid count buffer if needed (must be zero-initialized for atomics)
+    valid_counts: torch.Tensor | None = None
+    if return_valid_counts:
+        valid_counts = torch.zeros(
+            num_tokens, dtype=torch.int32, device=token_indices.device
+        )
+
+    # Strides in elements
+    bt_stride0, bt_stride1 = block_table_c.stride()
+    ti_stride0, ti_stride1 = token_indices_c.stride()
+    out_stride0, out_stride1 = out.stride()
+
+    # Prepare prefill pointers
+    if HAS_PREFILL_WORKSPACE:
+        assert prefill_workspace_request_ids is not None  # for mypy
+        assert prefill_workspace_starts is not None  # for mypy
+        assert prefill_workspace_request_ids.is_contiguous()
+        assert prefill_workspace_starts.is_contiguous()
+
+    # Exact 2D grid: tokens × column tiles
+    grid = (num_tokens, tiles_per_row)
+
+    _convert_req_index_to_global_index_kernel[grid](
+        req_id_c,
+        block_table_c,
+        token_indices_c,
+        out,
+        valid_counts,
+        prefill_workspace_request_ids,
+        prefill_workspace_starts,
+        # shapes / constexprs
+        max_num_blocks_per_req,
+        BLOCK_SIZE,
+        BLOCK_N,
+        HAS_PREFILL_WORKSPACE,
+        return_valid_counts,
+        # strides
+        bt_stride0,
+        bt_stride1,
+        ti_stride0,
+        ti_stride1,
+        out_stride0,
+        out_stride1,
+    )
+
+    if return_valid_counts:
+        assert valid_counts is not None
+        return out, valid_counts
+    return out

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

@@ -0,0 +1,210 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import ClassVar
+
+import torch
+
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention.mla_attention import (
+    MLACommonBackend,
+    MLACommonImpl,
+    MLACommonMetadata,
+)
+from vllm.model_executor.layers.batch_invariant import (
+    vllm_is_batch_invariant,
+)
+from vllm.platforms.interface import DeviceCapability
+from vllm.v1.attention.backend import (
+    AttentionLayer,
+    AttentionType,
+    is_quantized_kv_cache,
+)
+from vllm.v1.attention.ops.triton_decode_attention import decode_attention_fwd
+
+import ixformer.inference.functions as ixf_ops
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.distributed.parallel_state import get_dcp_group
+logger = init_logger(__name__)
+
+
+class TritonMLABackend(MLACommonBackend):
+    # supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+    # supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
+    #     "auto",
+    #     "bfloat16",
+    # ]
+
+    @staticmethod
+    def get_name() -> str:
+        return "TRITON_MLA"
+
+    @staticmethod
+    def get_impl_cls() -> type["TritonMLAImpl"]:
+        return TritonMLAImpl
+
+    @classmethod
+    def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
+        return True
+
+
+class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
+    can_return_lse_for_decode: bool = True
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        **mla_args,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            **mla_args,
+        )
+
+        unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "TritonMLAImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, logits_soft_cap"
+            )
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError(
+                "Encoder self-attention and "
+                "encoder/decoder cross-attention "
+                "are not implemented for "
+                "TritonMLAImpl"
+            )
+
+        if is_quantized_kv_cache(self.kv_cache_dtype):
+            raise NotImplementedError(
+                "TritonMLA V1 with FP8 KV cache not yet supported"
+            )
+
+    def _flash_attn_varlen_diff_headdims(
+        self, q, k, v, return_softmax_lse=False, softmax_scale=None, **kwargs
+    ):
+        return super()._flash_attn_varlen_diff_headdims(
+            q,
+            k,
+            v,
+            return_softmax_lse=return_softmax_lse,
+            softmax_scale=softmax_scale,
+            **kwargs,
+        )
+
+    def forward_mqa(
+        self,
+        q_nope: torch.Tensor,
+        q_pe: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: MLACommonMetadata,
+        # layer: AttentionLayer,
+        k_c_normed: torch.Tensor |None = None,
+        k_pe: torch.Tensor |None = None,
+        kv_c_and_k_pe_cache_scale: torch.Tensor |None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        assert kv_c_and_k_pe_cache.numel() > 0
+        assert attn_metadata.decode is not None
+
+        if self.kv_cache_dtype.startswith("fp8"):
+            raise NotImplementedError("FP8 Triton MLA not yet supported")
+
+        decode_meta = attn_metadata.decode
+        q_nope = self._k_up_proj(q_nope)
+        q_nope = q_nope.view(-1, self.num_heads, self.kv_lora_rank)
+
+        B = q_nope.shape[0]
+
+        if self.dcp_world_size > 1:
+            q = torch.cat([q_nope, q_pe], dim=-1)
+            q = get_dcp_group().all_gather(q, dim=1)
+            o = torch.empty(B,
+                        q.shape[1],
+                        self.kv_lora_rank,
+                        dtype=q_nope.dtype,
+                        device=q_nope.device)
+            if envs.VLLM_USE_INT8_MLA:
+                q_int8, q_scale = ops.quant_kv(q)
+                attn_out, softmax_lse = ixf_ops.ref_vllm_paged_attention_mla_int8(
+                o,
+                q_int8,
+                q_scale,
+                kv_c_and_k_pe_cache,
+                kv_c_and_k_pe_cache_scale, 
+                self.scale,
+                attn_metadata.decode.block_table,
+                attn_metadata.decode.seq_lens,
+                attn_metadata.decode.max_decode_seq_len,
+                return_softmax_lse=True   
+        )
+            else:
+                attn_out, softmax_lse = ixf_ops.ref_vllm_paged_attention_mla(
+                    output=o,
+                    query=q, 
+                    kv_cache=kv_c_and_k_pe_cache, 
+                    scale=self.scale, 
+                    block_tables=attn_metadata.decode.block_table,
+                    context_lens=attn_metadata.decode.seq_lens,
+                    max_context_len=decode_meta.max_decode_seq_len,
+                    return_softmax_lse=True)
+            return attn_out, softmax_lse
+
+        o = torch.empty(B,
+                        self.num_heads,
+                        self.kv_lora_rank,
+                        dtype=q_nope.dtype,
+                        device=q_nope.device)
+
+        if envs.VLLM_USE_INT8_MLA:
+            q = torch.cat([q_nope, q_pe], dim=-1)
+            q_int8, q_scale = ops.quant_kv(q)
+            ixf_ops.vllm_paged_attention_mla_int8(
+                o,
+                q_int8,
+                q_scale,
+                kv_c_and_k_pe_cache,
+                kv_c_and_k_pe_cache_scale, 
+                self.scale,
+                attn_metadata.decode.block_table,
+                attn_metadata.decode.seq_lens,
+                attn_metadata.decode.max_decode_seq_len,
+                attn_metadata.decode.use_cuda_graph
+        )
+        else:
+        # fused q concat & cache write
+            ixf_ops.vllm_paged_attention_mla_fused(
+                output=o,
+                q_nope=q_nope,
+                q_pe=q_pe.contiguous(),
+                kv_cache=kv_c_and_k_pe_cache,
+                scale=self.scale,
+                block_tables=attn_metadata.decode.block_table,
+                context_lens=attn_metadata.decode.seq_lens,
+                max_context_len=decode_meta.max_decode_seq_len,
+                k_c_normed=k_c_normed,
+                k_pe=k_pe,
+                use_cuda_graph=decode_meta.use_cuda_graph
+            )
+        return self._v_up_proj(o), None

vllm/v1/attention/backends/registry.py

@@ -0,0 +1,261 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Attention backend registry"""
+
+from collections.abc import Callable
+from enum import Enum, EnumMeta
+from typing import TYPE_CHECKING, cast
+
+from vllm.logger import init_logger
+from vllm.utils.import_utils import resolve_obj_by_qualname
+
+if TYPE_CHECKING:
+    from vllm.v1.attention.backend import AttentionBackend
+
+logger = init_logger(__name__)
+
+
+class _AttentionBackendEnumMeta(EnumMeta):
+    """Metaclass for AttentionBackendEnum to provide better error messages."""
+
+    def __getitem__(cls, name: str):
+        """Get backend by name with helpful error messages."""
+        try:
+            return super().__getitem__(name)
+        except KeyError:
+            members = cast("dict[str, Enum]", cls.__members__).keys()
+            valid_backends = ", ".join(members)
+            raise ValueError(
+                f"Unknown attention backend: '{name}'. "
+                f"Valid options are: {valid_backends}"
+            ) from None
+
+
+class AttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta):
+    """Enumeration of all supported attention backends.
+
+    The enum value is the default class path, but this can be overridden
+    at runtime using register_backend().
+
+    To get the actual backend class (respecting overrides), use:
+        backend.get_class()
+    """
+
+    FLASH_ATTN = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"
+    FLASH_ATTN_DIFFKV = (
+        "vllm.v1.attention.backends.flash_attn_diffkv.FlashAttentionDiffKVBackend"
+    )
+    TRITON_ATTN = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend"
+    ROCM_ATTN = "vllm.v1.attention.backends.rocm_attn.RocmAttentionBackend"
+    ROCM_AITER_MLA = "vllm.v1.attention.backends.mla.rocm_aiter_mla.AiterMLABackend"
+    ROCM_AITER_TRITON_MLA = (
+        "vllm.v1.attention.backends.mla.aiter_triton_mla.AiterTritonMLABackend"
+    )
+    ROCM_AITER_FA = (
+        "vllm.v1.attention.backends.rocm_aiter_fa.AiterFlashAttentionBackend"
+    )
+    ROCM_AITER_MLA_SPARSE = (
+        "vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse.ROCMAiterMLASparseBackend"
+    )
+    TORCH_SDPA = ""  # this tag is only used for ViT
+    FLASHINFER = "vllm.v1.attention.backends.flashinfer.FlashInferBackend"
+    FLASHINFER_MLA = (
+        "vllm.v1.attention.backends.mla.flashinfer_mla.FlashInferMLABackend"
+    )
+    FLASHINFER_MLA_SPARSE = (
+        "vllm.v1.attention.backends.mla.flashinfer_mla_sparse."
+        "FlashInferMLASparseBackend"
+    )
+    TRITON_MLA = "vllm.v1.attention.backends.mla.triton_mla.TritonMLABackend"
+    CUTLASS_MLA = "vllm.v1.attention.backends.mla.cutlass_mla.CutlassMLABackend"
+    FLASHMLA = "vllm.v1.attention.backends.mla.flashmla.FlashMLABackend"
+    FLASHMLA_SPARSE = (
+        "vllm.v1.attention.backends.mla.flashmla_sparse.FlashMLASparseBackend"
+    )
+    FLASH_ATTN_MLA = "vllm.v1.attention.backends.mla.flashattn_mla.FlashAttnMLABackend"
+    NO_ATTENTION = "vllm.v1.attention.backends.no_attention.NoAttentionBackend"
+    FLEX_ATTENTION = "vllm.v1.attention.backends.flex_attention.FlexAttentionBackend"
+    TREE_ATTN = "vllm.v1.attention.backends.tree_attn.TreeAttentionBackend"
+    ROCM_AITER_UNIFIED_ATTN = (
+        "vllm.v1.attention.backends.rocm_aiter_unified_attn."
+        "RocmAiterUnifiedAttentionBackend"
+    )
+    CPU_ATTN = "vllm.v1.attention.backends.cpu_attn.CPUAttentionBackend"
+    # Placeholder for third-party/custom backends - must be registered before use
+    # set to None to avoid alias with other backend, whose value is an empty string
+    CUSTOM = None
+
+    def get_path(self, include_classname: bool = True) -> str:
+        """Get the class path for this backend (respects overrides).
+
+        Returns:
+            The fully qualified class path string
+
+        Raises:
+            ValueError: If Backend.CUSTOM is used without being registered
+        """
+        path = _ATTN_OVERRIDES.get(self, self.value)
+        if not path:
+            raise ValueError(
+                f"Backend {self.name} must be registered before use. "
+                f"Use register_backend(Backend.{self.name}, 'your.module.YourClass')"
+            )
+        if not include_classname:
+            path = path.rsplit(".", 1)[0]
+        return path
+
+    def get_class(self) -> "type[AttentionBackend]":
+        """Get the backend class (respects overrides).
+
+        Returns:
+            The backend class
+
+        Raises:
+            ImportError: If the backend class cannot be imported
+            ValueError: If Backend.CUSTOM is used without being registered
+        """
+        return resolve_obj_by_qualname(self.get_path())
+
+    def is_overridden(self) -> bool:
+        """Check if this backend has been overridden.
+
+        Returns:
+            True if the backend has a registered override
+        """
+        return self in _ATTN_OVERRIDES
+
+    def clear_override(self) -> None:
+        """Clear any override for this backend, reverting to the default."""
+        _ATTN_OVERRIDES.pop(self, None)
+
+
+class MambaAttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta):
+    """Enumeration of all supported mamba attention backends.
+
+    The enum value is the default class path, but this can be overridden
+    at runtime using register_backend().
+
+    To get the actual backend class (respecting overrides), use:
+        backend.get_class()
+    """
+
+    MAMBA1 = "vllm.v1.attention.backends.mamba1_attn.Mamba1AttentionBackend"
+    MAMBA2 = "vllm.v1.attention.backends.mamba2_attn.Mamba2AttentionBackend"
+    SHORT_CONV = "vllm.v1.attention.backends.short_conv_attn.ShortConvAttentionBackend"
+    LINEAR = "vllm.v1.attention.backends.linear_attn.LinearAttentionBackend"
+    GDN_ATTN = "vllm.v1.attention.backends.gdn_attn.GDNAttentionBackend"
+    # Placeholder for third-party/custom backends - must be registered before use
+    # set to None to avoid alias with other backend, whose value is an empty string
+    CUSTOM = None
+
+    def get_path(self, include_classname: bool = True) -> str:
+        """Get the class path for this backend (respects overrides).
+
+        Returns:
+            The fully qualified class path string
+
+        Raises:
+            ValueError: If Backend.CUSTOM is used without being registered
+        """
+        path = _MAMBA_ATTN_OVERRIDES.get(self, self.value)
+        if not path:
+            raise ValueError(
+                f"Backend {self.name} must be registered before use. "
+                f"Use register_backend(Backend.{self.name}, 'your.module.YourClass')"
+            )
+        if not include_classname:
+            path = path.rsplit(".", 1)[0]
+        return path
+
+    def get_class(self) -> "type[AttentionBackend]":
+        """Get the backend class (respects overrides).
+
+        Returns:
+            The backend class
+
+        Raises:
+            ImportError: If the backend class cannot be imported
+            ValueError: If Backend.CUSTOM is used without being registered
+        """
+        return resolve_obj_by_qualname(self.get_path())
+
+    def is_overridden(self) -> bool:
+        """Check if this backend has been overridden.
+
+        Returns:
+            True if the backend has a registered override
+        """
+        return self in _MAMBA_ATTN_OVERRIDES
+
+    def clear_override(self) -> None:
+        """Clear any override for this backend, reverting to the default."""
+        _MAMBA_ATTN_OVERRIDES.pop(self, None)
+
+
+MAMBA_TYPE_TO_BACKEND_MAP = {
+    "mamba1": MambaAttentionBackendEnum.MAMBA1.name,
+    "mamba2": MambaAttentionBackendEnum.MAMBA2.name,
+    "short_conv": MambaAttentionBackendEnum.SHORT_CONV.name,
+    "linear_attention": MambaAttentionBackendEnum.LINEAR.name,
+    "gdn_attention": MambaAttentionBackendEnum.GDN_ATTN.name,
+    "custom": MambaAttentionBackendEnum.CUSTOM.name,
+}
+
+
+_ATTN_OVERRIDES: dict[AttentionBackendEnum, str] = {}
+_MAMBA_ATTN_OVERRIDES: dict[MambaAttentionBackendEnum, str] = {}
+
+
+def register_backend(
+    backend: AttentionBackendEnum | MambaAttentionBackendEnum,
+    class_path: str | None = None,
+    is_mamba: bool = False,
+) -> Callable[[type], type]:
+    """Register or override a backend implementation.
+
+    Args:
+        backend: The AttentionBackendEnum member to register
+        class_path: Optional class path. If not provided and used as
+            decorator, will be auto-generated from the class.
+
+    Returns:
+        Decorator function if class_path is None, otherwise a no-op
+
+    Examples:
+        # Override an existing attention backend
+        @register_backend(AttentionBackendEnum.FLASH_ATTN)
+        class MyCustomFlashAttn:
+            ...
+
+        # Override an existing mamba attention backend
+        @register_backend(MambaAttentionBackendEnum.LINEAR, is_mamba=True)
+        class MyCustomMambaAttn:
+            ...
+
+        # Register a custom third-party attention backend
+        @register_backend(AttentionBackendEnum.CUSTOM)
+        class MyCustomBackend:
+            ...
+
+        # Direct registration
+        register_backend(
+            AttentionBackendEnum.CUSTOM,
+            "my.module.MyCustomBackend"
+        )
+    """
+
+    def decorator(cls: type) -> type:
+        if is_mamba:
+            _MAMBA_ATTN_OVERRIDES[backend] = f"{cls.__module__}.{cls.__qualname__}"  # type: ignore[index]
+        else:
+            _ATTN_OVERRIDES[backend] = f"{cls.__module__}.{cls.__qualname__}"  # type: ignore[index]
+        return cls
+
+    if class_path is not None:
+        if is_mamba:
+            _MAMBA_ATTN_OVERRIDES[backend] = class_path  # type: ignore[index]
+        else:
+            _ATTN_OVERRIDES[backend] = class_path  # type: ignore[index]
+        return lambda x: x
+
+    return decorator

vllm/v1/attention/backends/rocm_aiter_unified_attn.py

@@ -0,0 +1,249 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Attention layer with PagedAttention and Triton prefix prefill."""
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm._aiter_ops import rocm_aiter_ops
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    QuantKey,
+    kFp8StaticTensorSym,
+)
+from vllm.v1.attention.backend import AttentionLayer, AttentionType
+from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
+from vllm.v1.attention.backends.rocm_attn import (
+    RocmAttentionBackend,
+    RocmAttentionImpl,
+    RocmAttentionMetadataBuilder,
+)
+
+logger = init_logger(__name__)
+
+
+class RocmAiterUnifiedAttentionBackend(RocmAttentionBackend):
+    accept_output_buffer: bool = True
+
+    forward_includes_kv_cache_update: bool = False
+
+    @staticmethod
+    def get_name() -> str:
+        return "ROCM_AITER_UNIFIED_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["RocmAiterUnifiedAttentionImpl"]:
+        return RocmAiterUnifiedAttentionImpl
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        if block_size % 16 != 0:
+            raise ValueError("Block size must be a multiple of 16.")
+        return (2, num_blocks, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def use_cascade_attention(*args, **kwargs) -> bool:
+        return False
+
+    @staticmethod
+    def get_builder_cls() -> type["RocmAttentionMetadataBuilder"]:
+        return RocmAttentionMetadataBuilder
+
+
+class RocmAiterUnifiedAttentionImpl(RocmAttentionImpl):
+    def fused_output_quant_supported(self, quant_key: QuantKey):
+        return quant_key == kFp8StaticTensorSym
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None = None,
+        attn_type: AttentionType = AttentionType.DECODER,
+        kv_sharing_target_layer_name: int | None = None,
+        sinks: torch.Tensor | None = None,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            sinks,
+        )
+        logger.info_once(
+            "Using aiter unified attention for RocmAiterUnifiedAttentionImpl"
+        )
+        from aiter.ops.triton.unified_attention import unified_attention
+
+        self.unified_attention = unified_attention
+
+    def forward(
+        self,
+        layer: torch.nn.Module,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: FlashAttentionMetadata,
+        output: torch.Tensor | None = None,
+        output_scale: torch.Tensor | None = None,
+        output_block_scale: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads, head_size]
+            key: shape = [num_tokens, num_kv_heads, head_size]
+            value: shape = [num_tokens, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        assert output is not None, "Output tensor must be provided."
+
+        if output_block_scale is not None:
+            raise NotImplementedError(
+                "fused block_scale output quantization is not yet supported"
+                " for RocmAttentionImpl"
+            )
+
+        if attn_metadata is None:
+            # Profiling run.
+            return output.fill_(0)
+
+        assert attn_metadata.use_cascade is False
+
+        # IMPORTANT!
+        # NOTE(woosuk): With piece-wise CUDA graphs, this method is executed in
+        # eager-mode PyTorch. Thus, we need to be careful about any CPU overhead
+        # in this method. For example, `view` and `slice` (or `[:n]`) operations
+        # are surprisingly slow even in the case they do not invoke any GPU ops.
+        # Minimize the PyTorch ops in this method as much as possible.
+        # Whenever making a change in this method, please benchmark the
+        # performance to make sure it does not introduce any overhead.
+
+        num_actual_tokens = attn_metadata.num_actual_tokens
+
+        key_cache, value_cache = kv_cache.unbind(0)
+
+        if self.kv_cache_dtype.startswith("fp8"):
+            key_cache = key_cache.view(self.fp8_dtype)
+            value_cache = value_cache.view(self.fp8_dtype)
+            assert layer._q_scale_float == 1.0, (
+                "A non 1.0 q_scale is not currently supported."
+            )
+
+        cu_seqlens_q = attn_metadata.query_start_loc
+        seqused_k = attn_metadata.seq_lens
+        max_seqlen_q = attn_metadata.max_query_len
+        max_seqlen_k = attn_metadata.max_seq_len
+        block_table = attn_metadata.block_table
+
+        descale_shape = (
+            cu_seqlens_q.shape[0] - 1,
+            key.shape[1] if key is not None else self.num_kv_heads,
+        )
+
+        self.unified_attention(
+            q=query[:num_actual_tokens],
+            k=key_cache,
+            v=value_cache,
+            out=output[:num_actual_tokens],
+            cu_seqlens_q=cu_seqlens_q,
+            max_seqlen_q=max_seqlen_q,
+            seqused_k=seqused_k,
+            max_seqlen_k=max_seqlen_k,
+            softmax_scale=self.scale,
+            causal=True,
+            alibi_slopes=self.alibi_slopes,
+            window_size=self.sliding_window,
+            block_table=block_table,
+            softcap=self.logits_soft_cap,
+            q_descale=None,  # Not supported
+            k_descale=layer._k_scale.expand(descale_shape),
+            v_descale=layer._v_scale.expand(descale_shape),
+            sinks=self.sinks,
+            output_scale=output_scale,
+        )
+
+        return output
+
+    def do_kv_cache_update(
+        self,
+        layer: AttentionLayer,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+    ):
+        key_cache, value_cache = kv_cache.unbind(0)
+
+        # Reshape the input keys and values and store them in the cache.
+        ops.reshape_and_cache_flash(
+            key,
+            value,
+            key_cache,
+            value_cache,
+            slot_mapping,
+            self.kv_cache_dtype,
+            layer._k_scale,
+            layer._v_scale,
+        )
+
+    def fused_rope_kvcache_supported(self):
+        return rocm_aiter_ops.is_enabled()
+
+    def do_rope_and_kv_cache_update(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        positions: torch.Tensor,
+        cos_sin_cache: torch.Tensor,
+        is_neox: bool,
+        kv_cache: torch.Tensor,
+        layer_slot_mapping: torch.Tensor,
+    ):
+        key_cache, value_cache = kv_cache.unbind(0)
+        flash_layout = True
+
+        is_fp8_kv_cache = self.kv_cache_dtype.startswith("fp8")
+        if is_fp8_kv_cache:
+            key_cache = key_cache.view(self.fp8_dtype)
+            value_cache = value_cache.view(self.fp8_dtype)
+
+        rocm_aiter_ops.triton_rope_and_cache(
+            query,
+            key,
+            value,
+            positions,
+            cos_sin_cache,
+            is_neox,
+            key_cache,
+            value_cache,
+            layer_slot_mapping,
+            layer._k_scale,
+            layer._v_scale,
+            flash_layout,
+            is_fp8_kv_cache,
+        )

vllm/v1/attention/backends/rocm_attn.py

@@ -0,0 +1,461 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Attention layer with PagedAttention and Triton prefix prefill."""
+
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm._aiter_ops import rocm_aiter_ops
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    QuantKey,
+    kFp8StaticTensorSym,
+)
+from vllm.platforms import current_platform
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionImpl,
+    AttentionLayer,
+    AttentionMetadataBuilder,
+    AttentionType,
+    CommonAttentionMetadata,
+    MultipleOf,
+)
+from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
+from vllm.v1.attention.ops.chunked_prefill_paged_decode import (
+    chunked_prefill_paged_decode,
+)
+from vllm.v1.attention.ops.paged_attn import PagedAttention
+from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
+    triton_reshape_and_cache_flash,
+)
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+logger = init_logger(__name__)
+
+
+@dataclass
+class RocmAttentionMetadata:
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ---------------------|
+    #                                   |-- query_len ---|
+
+    num_actual_tokens: int  # Number of tokens excluding padding.
+    max_query_len: int
+    query_start_loc: torch.Tensor
+    max_seq_len: int
+    seq_lens: torch.Tensor
+    block_table: torch.Tensor
+    slot_mapping: torch.Tensor
+
+    # For cascade attention.
+    use_cascade: bool
+    common_prefix_len: int
+    cu_prefix_query_lens: torch.Tensor | None
+    prefix_kv_lens: torch.Tensor | None
+    suffix_kv_lens: torch.Tensor | None
+
+    # Optional aot scheduling
+    scheduler_metadata: torch.Tensor | None = None
+    prefix_scheduler_metadata: torch.Tensor | None = None
+
+
+class RocmAttentionMetadataBuilder(AttentionMetadataBuilder[RocmAttentionMetadata]):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.ALWAYS
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(kv_cache_spec, layer_names, vllm_config, device)
+
+        self.block_size = kv_cache_spec.block_size
+
+        model_config = vllm_config.model_config
+        self.num_heads_q = model_config.get_num_attention_heads(
+            vllm_config.parallel_config
+        )
+        self.num_heads_kv = model_config.get_num_kv_heads(vllm_config.parallel_config)
+        self.headdim = model_config.get_head_size()
+
+    def build_for_cudagraph_capture(
+        self, common_attn_metadata: CommonAttentionMetadata
+    ) -> RocmAttentionMetadata:
+        attn_metadata = self.build(0, common_attn_metadata)
+        # When doing full graph capture, setting seq_lens to
+        # max_model_len will cause graph capture to be extremely
+        # slow, so here we set it to 1.
+        attn_metadata.seq_lens.fill_(1)
+
+        # Here we set the query start locs to 0. This is to
+        # cover up an invalid memory access in the prefix_prefil kernel
+        # that we run into during graph capture (#25985)
+        common_attn_metadata.query_start_loc.zero_()
+        common_attn_metadata.query_start_loc_cpu.zero_()
+
+        return attn_metadata
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> RocmAttentionMetadata:
+        num_actual_tokens = common_attn_metadata.num_actual_tokens
+        max_query_len = common_attn_metadata.max_query_len
+
+        max_seq_len = common_attn_metadata.max_seq_len
+        query_start_loc = common_attn_metadata.query_start_loc
+        seq_lens = common_attn_metadata.seq_lens
+        block_table_tensor = common_attn_metadata.block_table_tensor
+        slot_mapping = common_attn_metadata.slot_mapping
+
+        use_cascade = common_prefix_len > 0
+
+        if use_cascade:
+            cu_prefix_query_lens = torch.tensor(
+                [0, num_actual_tokens], dtype=torch.int32, device=self.device
+            )
+            prefix_kv_lens = torch.tensor(
+                [common_prefix_len], dtype=torch.int32, device=self.device
+            )
+            suffix_kv_lens = common_attn_metadata.seq_lens.cpu() - common_prefix_len
+            suffix_kv_lens = suffix_kv_lens.to(self.device)
+        else:
+            cu_prefix_query_lens = None
+            prefix_kv_lens = None
+            suffix_kv_lens = None
+            prefix_scheduler_metadata = None
+
+        attn_metadata = RocmAttentionMetadata(
+            num_actual_tokens=num_actual_tokens,
+            max_query_len=max_query_len,
+            query_start_loc=query_start_loc,
+            max_seq_len=max_seq_len,
+            seq_lens=seq_lens,
+            block_table=block_table_tensor,
+            slot_mapping=slot_mapping,
+            use_cascade=use_cascade,
+            common_prefix_len=common_prefix_len,
+            cu_prefix_query_lens=cu_prefix_query_lens,
+            prefix_kv_lens=prefix_kv_lens,
+            suffix_kv_lens=suffix_kv_lens,
+            prefix_scheduler_metadata=prefix_scheduler_metadata,
+        )
+        return attn_metadata
+
+
+class RocmAttentionBackend(AttentionBackend):
+    accept_output_buffer: bool = True
+    supported_dtypes: ClassVar[list[torch.dtype]] = [
+        torch.float16,
+        torch.bfloat16,
+        torch.float32,
+    ]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        # ROCM paged attention kernel only supports block sizes 16 and 32
+        # due to shared memory (LDS) constraints on AMD GPUs.
+        # See csrc/rocm/attention.cu CALL_CUSTOM_LAUNCHER_BLK macro.
+
+        # However, The limitations in [16, 32] are reasonable for a native C++ kernel,
+        # but vLLM should allow support for non-standard sizes via the Triton path,
+        # as addressed in this PR: https://github.com/vllm-project/vllm/pull/31380,
+        # where the Triton kernel under rocm_atten does not support inference
+        # for a non-standard qwen3-next model with a block_size of 544.
+        # We have fixed the Triton kernel so that the standard model uses the original
+        # bit-addressing logic, while the non-standard model
+        # uses our optimized kernel logic.
+        return [16, 32, 544]
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return [32, 64, 96, 128, 160, 192, 224, 256]
+
+    @classmethod
+    def validate_head_size(cls, head_size: int) -> None:
+        if not cls.supports_head_size(head_size):
+            attn_type = cls.__name__.removesuffix("Backend")
+            raise ValueError(
+                f"Head size {head_size} is not supported by {attn_type}. "
+                f"Supported head sizes are: {cls.get_supported_head_sizes()}. "
+                "Set --attention-backend=FLEX_ATTENTION to use "
+                "FlexAttention backend which supports all head sizes."
+            )
+
+    forward_includes_kv_cache_update: bool = False
+
+    @staticmethod
+    def get_name() -> str:
+        return "ROCM_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["RocmAttentionImpl"]:
+        return RocmAttentionImpl
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        if block_size % 16 != 0:
+            raise ValueError("Block size must be a multiple of 16.")
+        return (2, num_blocks, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def use_cascade_attention(*args, **kwargs) -> bool:
+        return False
+
+    @staticmethod
+    def get_builder_cls() -> type["RocmAttentionMetadataBuilder"]:
+        return RocmAttentionMetadataBuilder
+
+
+class RocmAttentionImpl(AttentionImpl):
+    def fused_output_quant_supported(self, quant_key: QuantKey):
+        return quant_key == kFp8StaticTensorSym
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None = None,
+        attn_type: AttentionType = AttentionType.DECODER,
+        kv_sharing_target_layer_name: int | None = None,
+        sinks: torch.Tensor | None = None,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        if sliding_window is None:
+            self.sliding_window = (-1, -1)
+        else:
+            self.sliding_window = (sliding_window - 1, 0)
+        self.kv_cache_dtype = kv_cache_dtype
+        if logits_soft_cap is None:
+            # In flash-attn, setting logits_soft_cap as 0 means no soft cap.
+            logits_soft_cap = 0
+        self.logits_soft_cap = logits_soft_cap
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
+
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        RocmAttentionBackend.validate_head_size(head_size)
+
+        if attn_type not in [AttentionType.DECODER, AttentionType.ENCODER_DECODER]:
+            raise NotImplementedError(
+                "Encoder self-attention is not implemented for RocmAttentionImpl"
+            )
+
+        self.fp8_dtype = current_platform.fp8_dtype()
+
+        self.sinks = sinks
+        if sinks is not None:
+            assert sinks.shape[0] == num_heads, (
+                "Sinks must have the same number of heads as the number of "
+                f"heads in the layer. Sinks shape: {sinks.shape}, "
+                f"num_heads: {num_heads}."
+            )
+
+    def forward(
+        self,
+        layer: torch.nn.Module,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: FlashAttentionMetadata,
+        output: torch.Tensor | None = None,
+        output_scale: torch.Tensor | None = None,
+        output_block_scale: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads, head_size]
+            key: shape = [num_tokens, num_kv_heads, head_size]
+            value: shape = [num_tokens, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        assert output is not None, "Output tensor must be provided."
+
+        if output_block_scale is not None:
+            raise NotImplementedError(
+                "fused block_scale output quantization is not yet supported"
+                " for RocmAttentionImpl"
+            )
+
+        if attn_metadata is None:
+            # Profiling run.
+            return output.fill_(0)
+
+        assert attn_metadata.use_cascade is False
+
+        # IMPORTANT!
+        # NOTE(woosuk): With piece-wise CUDA graphs, this method is executed in
+        # eager-mode PyTorch. Thus, we need to be careful about any CPU overhead
+        # in this method. For example, `view` and `slice` (or `[:n]`) operations
+        # are surprisingly slow even in the case they do not invoke any GPU ops.
+        # Minimize the PyTorch ops in this method as much as possible.
+        # Whenever making a change in this method, please benchmark the
+        # performance to make sure it does not introduce any overhead.
+
+        num_actual_tokens = attn_metadata.num_actual_tokens
+
+        key_cache, value_cache = PagedAttention.split_kv_cache(
+            kv_cache, self.num_kv_heads, self.head_size
+        )
+
+        if self.kv_cache_dtype.startswith("fp8"):
+            key_cache = key_cache.view(self.fp8_dtype)
+            value_cache = value_cache.view(self.fp8_dtype)
+            assert layer._q_scale_float == 1.0, (
+                "A non 1.0 q_scale is not currently supported."
+            )
+
+        cu_seqlens_q = attn_metadata.query_start_loc
+        seqused_k = attn_metadata.seq_lens
+        max_seqlen_q = attn_metadata.max_query_len
+        max_seqlen_k = attn_metadata.max_seq_len
+        block_table = attn_metadata.block_table
+
+        # Compute attention and update output up to `num_actual_tokens`.
+        chunked_prefill_paged_decode(
+            query=query[:num_actual_tokens],
+            key=key[:num_actual_tokens] if key is not None else None,
+            value=value[:num_actual_tokens] if value is not None else None,
+            output=output[:num_actual_tokens],
+            kv_cache_dtype=self.kv_cache_dtype,
+            key_cache=key_cache,
+            value_cache=value_cache,
+            block_table=block_table,
+            query_start_loc=cu_seqlens_q,
+            seq_lens=seqused_k,
+            max_seq_len=max_seqlen_k,
+            max_query_len=max_seqlen_q,
+            k_scale=layer._k_scale,
+            v_scale=layer._v_scale,
+            alibi_slopes=self.alibi_slopes,
+            sliding_window=self.sliding_window[0],
+            sm_scale=self.scale,
+            output_scale=output_scale,
+            sinks=self.sinks,
+        )
+
+        return output
+
+    def do_kv_cache_update(
+        self,
+        layer: AttentionLayer,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+    ):
+        key_cache, value_cache = PagedAttention.split_kv_cache(
+            kv_cache, self.num_kv_heads, self.head_size
+        )
+
+        # Reshape the input keys and values and store them in the cache.
+        # Get the actual block_size from value_cache
+        # value_cache shape: [num_blocks, num_heads, head_size, block_size]
+        block_size = value_cache.shape[3]
+        # Determine if it is a power of 2
+        is_pow2 = block_size > 0 and (block_size & (block_size - 1) == 0)
+
+        if is_pow2:
+            # Normal 16, 32, 64, etc., use vLLM native HIP C++ logic
+            PagedAttention.write_to_paged_cache(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                slot_mapping,
+                self.kv_cache_dtype,
+                layer._k_scale,
+                layer._v_scale,
+            )
+        else:
+            # Case B: Non-standard blocks (e.g., 544 in Qwen3),
+            # force using our modified Triton logic
+            triton_reshape_and_cache_flash(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                slot_mapping,
+                self.kv_cache_dtype,
+                layer._k_scale,
+                layer._v_scale,
+            )
+
+    def fused_rope_kvcache_supported(self):
+        return rocm_aiter_ops.is_enabled()
+
+    def do_rope_and_kv_cache_update(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        positions: torch.Tensor,
+        cos_sin_cache: torch.Tensor,
+        is_neox: bool,
+        kv_cache: torch.Tensor,
+        layer_slot_mapping: torch.Tensor,
+    ):
+        key_cache, value_cache = PagedAttention.split_kv_cache(
+            kv_cache,
+            layer.num_kv_heads,  # type: ignore[attr-defined]
+            layer.head_size,  # type: ignore[attr-defined]
+        )
+        flash_layout = False
+
+        is_fp8_kv_cache = self.kv_cache_dtype.startswith("fp8")
+        if is_fp8_kv_cache:
+            key_cache = key_cache.view(self.fp8_dtype)
+            value_cache = value_cache.view(self.fp8_dtype)
+
+        rocm_aiter_ops.triton_rope_and_cache(
+            query,
+            key,
+            value,
+            positions,
+            cos_sin_cache,
+            is_neox,
+            key_cache,
+            value_cache,
+            layer_slot_mapping,
+            layer._k_scale,
+            layer._v_scale,
+            flash_layout,
+            is_fp8_kv_cache,
+        )

vllm/v1/attention/backends/short_conv_attn.py

@@ -0,0 +1,30 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+
+from vllm.v1.attention.backend import AttentionBackend
+from vllm.v1.attention.backends.mamba_attn import (
+    BaseMambaAttentionMetadata,
+    BaseMambaAttentionMetadataBuilder,
+)
+
+
+class ShortConvAttentionBackend(AttentionBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "SHORT_CONV_ATTN"
+
+    @staticmethod
+    def get_builder_cls() -> type["ShortConvAttentionMetadataBuilder"]:
+        return ShortConvAttentionMetadataBuilder
+
+
+@dataclass
+class ShortConvAttentionMetadata(BaseMambaAttentionMetadata):
+    pass
+
+
+class ShortConvAttentionMetadataBuilder(
+    BaseMambaAttentionMetadataBuilder[ShortConvAttentionMetadata]
+):
+    metadata_cls = ShortConvAttentionMetadata

vllm/v1/attention/backends/tree_attn.py

@@ -0,0 +1,430 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Attention layer with TreeAttention."""
+
+import ast
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadataBuilder,
+    AttentionType,
+    CommonAttentionMetadata,
+    MultipleOf,
+)
+from vllm.v1.attention.backends.utils import (
+    split_decodes_and_prefills,
+)
+from vllm.v1.attention.ops.triton_unified_attention import unified_attention
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+logger = init_logger(__name__)
+
+
+class TreeAttentionBackend(AttentionBackend):
+    accept_output_buffer: bool = True
+    supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [MultipleOf(16)]
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return [32, 64, 96, 128, 160, 192, 224, 256]
+
+    @staticmethod
+    def get_name() -> str:
+        return "TREE_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["TreeAttentionImpl"]:
+        return TreeAttentionImpl
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        if block_size % 16 != 0:
+            raise ValueError("Block size must be a multiple of 16.")
+        return (2, num_blocks, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def get_builder_cls() -> type["TreeAttentionMetadataBuilder"]:
+        return TreeAttentionMetadataBuilder
+
+    @staticmethod
+    def use_cascade_attention(*args, **kwargs) -> bool:
+        return False
+
+
+@dataclass
+class TreeAttentionMetadata:
+    num_actual_tokens: int  # Number of tokens excluding padding.
+    max_query_len: int
+    query_start_loc: torch.Tensor
+    max_seq_len: int
+    seq_lens: torch.Tensor
+    block_table: torch.Tensor
+    slot_mapping: torch.Tensor
+
+    num_prefill_tokens: int = 0
+    num_decode_tokens: int = 0
+    num_prefills: int = 0
+    num_decodes: int = 0
+
+    tree_attn_bias: torch.Tensor | None = None
+
+    # Cached Prefill/decode metadata.
+    _cached_prefill_metadata: "TreeAttentionMetadata | None" = None
+    _cached_decode_metadata: "TreeAttentionMetadata | None" = None
+
+    @property
+    def prefill_metadata(self) -> "TreeAttentionMetadata | None":
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            # Recover cached prefill-phase attention
+            # metadata structure
+            return self._cached_prefill_metadata
+
+        q_start_loc = self.query_start_loc[self.num_decodes :]
+        q_seqlens = torch.diff(q_start_loc)
+        kv_seqlens = self.seq_lens[self.num_decodes :]
+        # Construct & cache prefill-phase attention metadata structure
+        self._cached_prefill_metadata = TreeAttentionMetadata(
+            num_actual_tokens=self.num_prefill_tokens,
+            max_query_len=int(q_seqlens.max().item()),
+            query_start_loc=q_start_loc - q_start_loc[0],
+            max_seq_len=int(kv_seqlens.max().item()),
+            seq_lens=kv_seqlens,
+            block_table=self.block_table[self.num_decodes :],
+            slot_mapping=self.slot_mapping[self.num_decode_tokens :],
+        )
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> "TreeAttentionMetadata | None":
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            # Recover cached decode-phase attention
+            # metadata structure
+            return self._cached_decode_metadata
+
+        q_start_loc = self.query_start_loc[: self.num_decodes + 1]
+        q_seqlens = torch.diff(q_start_loc)
+        kv_seqlens = self.seq_lens[: self.num_decodes]
+        # Construct & cache decode-phase attention metadata structure
+        self._cached_decode_metadata = TreeAttentionMetadata(
+            num_actual_tokens=self.num_decode_tokens,
+            max_query_len=int(q_seqlens.max().item()),
+            query_start_loc=q_start_loc,
+            max_seq_len=int(kv_seqlens.max().item()),
+            seq_lens=kv_seqlens,
+            block_table=self.block_table[: self.num_decodes],
+            slot_mapping=self.slot_mapping[: self.num_decode_tokens],
+            tree_attn_bias=self.tree_attn_bias,
+        )
+        return self._cached_decode_metadata
+
+
+class TreeAttentionMetadataBuilder(AttentionMetadataBuilder[TreeAttentionMetadata]):
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(kv_cache_spec, layer_names, vllm_config, device)
+
+        self.block_size = kv_cache_spec.block_size
+
+        spec_config = vllm_config.speculative_config
+        spec_token_tree: str | None = None
+        if spec := spec_config:
+            spec_token_tree = spec.speculative_token_tree
+        tree_choices: list[tuple[int, ...]] = (
+            ast.literal_eval(spec_token_tree) if spec_token_tree is not None else [(0,)]
+        )
+        # Construct the tree attention bias.
+        depth_counts = _get_depth_counts(tree_choices)
+        self.tree_attn_bias = _prepare_tree_attn_bias(
+            tree_choices,
+            depth_counts,
+            dtype=torch.float32,
+            device=device,
+        )
+
+        self.reorder_batch_threshold = self.tree_attn_bias.shape[0]
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> TreeAttentionMetadata:
+        decode_threshold = self.tree_attn_bias.shape[0]
+        num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = (
+            split_decodes_and_prefills(
+                common_attn_metadata, decode_threshold=decode_threshold
+            )
+        )
+
+        num_actual_tokens = common_attn_metadata.num_actual_tokens
+        q_start_loc = common_attn_metadata.query_start_loc
+        max_query_len = common_attn_metadata.max_query_len
+        kv_seqlens = common_attn_metadata.seq_lens
+        max_seq_len = common_attn_metadata.max_seq_len
+        block_table = common_attn_metadata.block_table_tensor
+        slot_mapping = common_attn_metadata.slot_mapping
+
+        return TreeAttentionMetadata(
+            num_actual_tokens=num_actual_tokens,
+            num_prefill_tokens=num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            num_prefills=num_prefills,
+            num_decodes=num_decodes,
+            max_query_len=max_query_len,
+            query_start_loc=q_start_loc,
+            max_seq_len=max_seq_len,
+            seq_lens=kv_seqlens,
+            block_table=block_table,
+            slot_mapping=slot_mapping,
+            tree_attn_bias=self.tree_attn_bias,
+        )
+
+    def build_for_drafting(
+        self,
+        common_attn_metadata: CommonAttentionMetadata,
+        draft_index: int,
+    ) -> TreeAttentionMetadata:
+        # Cache the original tree attention bias.
+        orig_tree_attn_bias = self.tree_attn_bias
+
+        if draft_index == 0:
+            # Use prefill for drafting at the root level.
+            self.tree_attn_bias = torch.empty(0)
+        else:
+            # Slice the tree attention bias for drafting. Exclude
+            # the root level.
+            start, end = 1, 1 + common_attn_metadata.max_query_len
+            self.tree_attn_bias = self.tree_attn_bias[start:end, start:end].contiguous()
+
+        # Build attention bias.
+        attn_metadata = self.build(0, common_attn_metadata, fast_build=True)
+
+        # Reset the tree attention bias to the original value.
+        self.tree_attn_bias = orig_tree_attn_bias
+        return attn_metadata
+
+
+def _get_depth_counts(sorted_tree_choices: list[tuple[int, ...]]) -> list[int]:
+    # Count the number of choices at each depth of the tree.
+    depth_counts = []
+    prev_depth = 0
+    for path in sorted_tree_choices:
+        depth = len(path)
+        if depth != prev_depth:
+            depth_counts.append(0)
+        depth_counts[depth - 1] += 1
+        prev_depth = depth
+    return depth_counts
+
+
+def _prepare_tree_attn_bias(
+    sorted_tree_choices: list[tuple[int, ...]],
+    depth_counts: list[int],
+    dtype: torch.dtype | None,
+    device: torch.device | None,
+) -> torch.Tensor:
+    # +1 comes from the additional root node.
+    tree_len = len(sorted_tree_choices) + 1
+    tree_attn_mask = torch.full(
+        (tree_len, tree_len), -torch.inf, device=device, dtype=dtype
+    )
+
+    # Set diagonal to all zeros. Each token should
+    # attend to itself.
+    mask_val = 0
+    for i in range(tree_len):
+        tree_attn_mask[i, i] = mask_val
+
+    # Set root to all zeros. All tokens attend to it.
+    tree_attn_mask[:, 0] = mask_val
+
+    # Set all ancestors to zeros.
+    start = 0
+    for i in range(len(depth_counts)):
+        for j in range(depth_counts[i]):
+            cur_tree_choice = sorted_tree_choices[start + j]
+            # Retrieve ancestor position.
+            if len(cur_tree_choice) == 1:
+                continue
+            ancestor_idx = []
+            for c in range(len(cur_tree_choice) - 1):
+                ancestor_idx.append(
+                    sorted_tree_choices.index(cur_tree_choice[: c + 1]) + 1
+                )
+            tree_attn_mask[j + start + 1, ancestor_idx] = mask_val
+        start += depth_counts[i]
+    return tree_attn_mask
+
+
+class TreeAttentionImpl(AttentionImpl):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None = None,
+        attn_type: AttentionType = AttentionType.DECODER,
+        kv_sharing_target_layer_name: str | None = None,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        self.kv_cache_dtype = kv_cache_dtype
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        if logits_soft_cap is None:
+            # Setting logits_soft_cap to 0 means no soft cap.
+            logits_soft_cap = 0
+        self.logits_soft_cap = logits_soft_cap
+        if sliding_window is None:
+            self.sliding_window = (-1, -1)
+        else:
+            self.sliding_window = (sliding_window - 1, 0)
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError(
+                "Encoder self-attention and "
+                "encoder/decoder cross-attention "
+                "are not implemented for "
+                "TreeAttentionImpl."
+            )
+
+    def forward(
+        self,
+        layer: torch.nn.Module,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: TreeAttentionMetadata,
+        output: torch.Tensor | None = None,
+        output_scale: torch.Tensor | None = None,
+        output_block_scale: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        """Forward pass with TreeAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads, head_size]
+            key: shape = [num_tokens, num_kv_heads, head_size]
+            value: shape = [num_tokens, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        assert output is not None, "Output tensor must be provided."
+
+        if output_scale is not None or output_block_scale is not None:
+            raise NotImplementedError(
+                "fused output quantization is not yet supported for TreeAttentionImpl"
+            )
+
+        if attn_metadata is None:
+            # Profiling run.
+            return output.fill_(0)
+
+        # Cache the input KVs.
+        key_cache, value_cache = kv_cache.unbind(0)
+        if self.kv_sharing_target_layer_name is None:
+            # Reshape the input keys and values and store them in the cache.
+            # Skip this if sharing KV cache with an earlier attention layer.
+            # NOTE(woosuk): Here, key and value are padded while slot_mapping is
+            # not padded. However, we don't need to do key[:num_actual_tokens]
+            # and value[:num_actual_tokens] because the reshape_and_cache_flash
+            # op uses the slot_mapping's shape to determine the number of
+            # actual tokens.
+            ops.reshape_and_cache_flash(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping,
+                self.kv_cache_dtype,
+                layer._k_scale,
+                layer._v_scale,
+            )
+
+        num_actual_tokens = attn_metadata.num_actual_tokens
+        num_decode_tokens = attn_metadata.num_decode_tokens
+        descale_shape = (attn_metadata.query_start_loc.shape[0] - 1, key.shape[1])
+        if prefill_meta := attn_metadata.prefill_metadata:
+            unified_attention(
+                q=query[num_decode_tokens:num_actual_tokens],
+                k=key_cache,
+                v=value_cache,
+                out=output[num_decode_tokens:num_actual_tokens],
+                cu_seqlens_q=prefill_meta.query_start_loc,
+                max_seqlen_q=prefill_meta.max_query_len,
+                seqused_k=prefill_meta.seq_lens,
+                max_seqlen_k=prefill_meta.max_seq_len,
+                softmax_scale=self.scale,
+                causal=True,
+                alibi_slopes=self.alibi_slopes,
+                window_size=self.sliding_window,
+                block_table=prefill_meta.block_table,
+                softcap=self.logits_soft_cap,
+                q_descale=None,  # Not supported
+                k_descale=layer._k_scale.expand(descale_shape),
+                v_descale=layer._v_scale.expand(descale_shape),
+            )
+
+        if decode_meta := attn_metadata.decode_metadata:
+            unified_attention(
+                q=query[:num_decode_tokens],
+                k=key_cache,
+                v=value_cache,
+                out=output[:num_decode_tokens],
+                cu_seqlens_q=decode_meta.query_start_loc,
+                max_seqlen_q=decode_meta.max_query_len,
+                seqused_k=decode_meta.seq_lens,
+                max_seqlen_k=decode_meta.max_seq_len,
+                softmax_scale=self.scale,
+                causal=True,
+                alibi_slopes=self.alibi_slopes,
+                qq_bias=decode_meta.tree_attn_bias,
+                window_size=self.sliding_window,
+                block_table=decode_meta.block_table,
+                softcap=self.logits_soft_cap,
+                q_descale=None,  # Not supported
+                k_descale=layer._k_scale.expand(descale_shape),
+                v_descale=layer._v_scale.expand(descale_shape),
+            )
+        return output

vllm/v1/attention/backends/triton_attn.py

@@ -0,0 +1,638 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""High-Performance Triton-only Attention layer."""
+
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm._aiter_ops import rocm_aiter_ops
+from vllm.config import CUDAGraphMode, VllmConfig
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    QuantKey,
+    kFp8StaticTensorSym,
+)
+from vllm.platforms import current_platform
+from vllm.platforms.interface import DeviceCapability
+from vllm.utils.math_utils import next_power_of_2
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionImpl,
+    AttentionLayer,
+    AttentionMetadataBuilder,
+    AttentionType,
+    CommonAttentionMetadata,
+    MultipleOf,
+)
+from vllm.v1.attention.ops.triton_prefill_attention import context_attention_fwd
+from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
+    triton_reshape_and_cache_flash,
+)
+from vllm.v1.attention.ops.triton_unified_attention import unified_attention
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+logger = init_logger(__name__)
+
+
+# constants
+MIN_LAUNCH_GRID_SIZE_2D = 128  # Minimum launch grid size of 2D kernel
+NUM_PAR_SOFTMAX_SEGMENTS = 16  # Number of parallel tiled softmax segments
+
+
+@dataclass
+class TritonAttentionMetadata:
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ---------------------|
+    #                                   |-- query_len ---|
+
+    num_actual_tokens: int  # Number of tokens excluding padding.
+    max_query_len: int
+    query_start_loc: torch.Tensor
+    max_seq_len: int
+    seq_lens: torch.Tensor
+    block_table: torch.Tensor
+    slot_mapping: torch.Tensor
+
+    seq_threshold_3D: int
+    num_par_softmax_segments: int
+    softmax_segm_output: torch.Tensor
+    softmax_segm_max: torch.Tensor
+    softmax_segm_expsum: torch.Tensor
+
+    # For cascade attention.
+    use_cascade: bool
+    common_prefix_len: int
+    cu_prefix_query_lens: torch.Tensor | None
+    prefix_kv_lens: torch.Tensor | None
+    suffix_kv_lens: torch.Tensor | None
+
+    # Optional aot scheduling
+    scheduler_metadata: torch.Tensor | None = None
+    prefix_scheduler_metadata: torch.Tensor | None = None
+    mm_prefix_range: dict[int, list[tuple[int, int]]] | None = None
+
+    @property
+    def mm_prefix_range_tensor(self) -> torch.Tensor | None:
+        """Convert mm_prefix_range dict to padded tensor for Triton kernel.
+
+        Returns shape: (num_seqs, max_ranges, 2) with 0-padding for empty ranges.
+        Empty ranges have start==end==0, which kernel skips via is_valid check.
+        """
+        # TODO(Isotr0py): Move to model runner's attention metadata
+        # preparation to avoid duplicate computation.
+        if self.mm_prefix_range is None:
+            return None
+
+        num_seqs = self.seq_lens.shape[0]
+        device = self.seq_lens.device
+
+        # Collect ranges, using [(0,0)] for empty sequences to ensure uniform dims
+        range_lists = [
+            self.mm_prefix_range.get(i, [(0, 0)]) or [(0, 0)] for i in range(num_seqs)
+        ]
+
+        # Return None if all ranges are trivial (only (0,0) placeholders)
+        if all(r == [(0, 0)] for r in range_lists):
+            return None
+
+        # Create 2D tensors with shape (num_ranges, 2) for each sequence
+        range_tensors = [
+            torch.tensor(r, dtype=torch.int32, device=device).view(-1, 2)
+            for r in range_lists
+        ]
+
+        return torch.nested.nested_tensor(
+            range_tensors, layout=torch.jagged
+        ).to_padded_tensor(0)
+
+
+class TritonAttentionMetadataBuilder(AttentionMetadataBuilder[TritonAttentionMetadata]):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.ALWAYS
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(kv_cache_spec, layer_names, vllm_config, device)
+
+        self.block_size = kv_cache_spec.block_size
+
+        model_config = vllm_config.model_config
+        self.num_heads_q = model_config.get_num_attention_heads(
+            vllm_config.parallel_config
+        )
+        self.num_heads_kv = model_config.get_num_kv_heads(vllm_config.parallel_config)
+        self.headdim = model_config.get_head_size()
+
+        # Check if CUDA Graphs are enabled for decode
+        self.decode_cudagraph_enabled = (
+            self.vllm_config.compilation_config.cudagraph_mode
+            in (
+                CUDAGraphMode.FULL_AND_PIECEWISE,
+                CUDAGraphMode.FULL_DECODE_ONLY,
+                CUDAGraphMode.FULL,
+            )
+        )
+
+        # The launch grid for the 2D kernel is defined as (num_q_blocks, num_heads_kv).
+        # A lower bound for num_q_blocks is the number of sequences.
+        # To ensure the minimum launch grid size is achieved, the number of sequences
+        # must be at least equal to the threshold below.
+        # If this threshold is not reached (i.e., the batch size is not large enough),
+        # the 3D kernel will be selected instead.
+        self.seq_threshold_3D = MIN_LAUNCH_GRID_SIZE_2D // self.num_heads_kv
+
+        # Modify the threshold if needed.
+        if self.decode_cudagraph_enabled:
+            capture_sizes = self.vllm_config.compilation_config.cudagraph_capture_sizes
+            assert capture_sizes, "CUDA Graphs enabled but no capture sizes specified."
+
+            # Select the CUDA Graph capture size closest to self.seq_threshold_3D
+            # as threshold. This ensures that each captured graph covers the
+            # correct execution path.
+            self.seq_threshold_3D = min(
+                capture_sizes,
+                key=lambda x: abs(x - self.seq_threshold_3D),
+            )
+
+        self.num_par_softmax_segments = NUM_PAR_SOFTMAX_SEGMENTS
+        headdim_padded = next_power_of_2(self.headdim)
+        self.softmax_segm_output = torch.empty(
+            (
+                self.seq_threshold_3D,
+                self.num_heads_q,
+                self.num_par_softmax_segments,
+                headdim_padded,
+            ),
+            dtype=torch.float32,
+            device=device,
+        )
+        self.softmax_segm_max = torch.empty(
+            (self.seq_threshold_3D, self.num_heads_q, self.num_par_softmax_segments),
+            dtype=torch.float32,
+            device=device,
+        )
+        self.softmax_segm_expsum = torch.empty(
+            (self.seq_threshold_3D, self.num_heads_q, self.num_par_softmax_segments),
+            dtype=torch.float32,
+            device=device,
+        )
+
+    def build_for_cudagraph_capture(
+        self, common_attn_metadata: CommonAttentionMetadata
+    ) -> TritonAttentionMetadata:
+        attn_metadata = self.build(0, common_attn_metadata)
+        # When doing full graph capture, setting seq_lens to
+        # max_model_len will cause graph capture to be extremely
+        # slow, so here we set it to 1.
+        attn_metadata.seq_lens.fill_(1)
+        return attn_metadata
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> TritonAttentionMetadata:
+        num_actual_tokens = common_attn_metadata.num_actual_tokens
+        max_query_len = common_attn_metadata.max_query_len
+
+        max_seq_len = common_attn_metadata.max_seq_len
+        query_start_loc = common_attn_metadata.query_start_loc
+        seq_lens = common_attn_metadata.seq_lens
+        block_table_tensor = common_attn_metadata.block_table_tensor
+        slot_mapping = common_attn_metadata.slot_mapping
+
+        use_cascade = common_prefix_len > 0
+
+        if use_cascade:
+            cu_prefix_query_lens = torch.tensor(
+                [0, num_actual_tokens], dtype=torch.int32, device=self.device
+            )
+            prefix_kv_lens = torch.tensor(
+                [common_prefix_len], dtype=torch.int32, device=self.device
+            )
+            suffix_kv_lens = common_attn_metadata.seq_lens.cpu() - common_prefix_len
+            suffix_kv_lens = suffix_kv_lens.to(self.device)
+        else:
+            cu_prefix_query_lens = None
+            prefix_kv_lens = None
+            suffix_kv_lens = None
+            prefix_scheduler_metadata = None
+
+        attn_metadata = TritonAttentionMetadata(
+            num_actual_tokens=num_actual_tokens,
+            max_query_len=max_query_len,
+            query_start_loc=query_start_loc,
+            max_seq_len=max_seq_len,
+            seq_lens=seq_lens,
+            block_table=block_table_tensor,
+            slot_mapping=slot_mapping,
+            use_cascade=use_cascade,
+            common_prefix_len=common_prefix_len,
+            cu_prefix_query_lens=cu_prefix_query_lens,
+            prefix_kv_lens=prefix_kv_lens,
+            suffix_kv_lens=suffix_kv_lens,
+            prefix_scheduler_metadata=prefix_scheduler_metadata,
+            seq_threshold_3D=self.seq_threshold_3D,
+            num_par_softmax_segments=self.num_par_softmax_segments,
+            softmax_segm_output=self.softmax_segm_output,
+            softmax_segm_max=self.softmax_segm_max,
+            softmax_segm_expsum=self.softmax_segm_expsum,
+        )
+        return attn_metadata
+
+
+class TritonAttentionBackend(AttentionBackend):
+    accept_output_buffer: bool = True
+    supported_dtypes: ClassVar[list[torch.dtype]] = [
+        torch.float16,
+        torch.bfloat16,
+        torch.float32,
+    ]
+    supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
+        "auto",
+        "bfloat16",
+        "fp8",
+        "fp8_e4m3",
+        "fp8_e5m2",
+    ]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [MultipleOf(16)]
+
+    forward_includes_kv_cache_update: bool = False
+
+    @staticmethod
+    def get_name() -> str:
+        return "TRITON_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["TritonAttentionImpl"]:
+        return TritonAttentionImpl
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        if block_size % 16 != 0:
+            raise ValueError("Block size must be a multiple of 16.")
+        return (num_blocks, 2, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def get_kv_cache_stride_order(
+        include_num_layers_dimension: bool = False,
+    ) -> tuple[int, ...]:
+        # `stride_order` indicates the permutation that gets
+        # us from `get_kv_cache_shape` to the actual memory layout we want.
+        if include_num_layers_dimension:
+            # (num_blocks, num_layers, 2, block_size, num_kv_heads, head_size)
+            return (1, 0, 2, 3, 4, 5)
+
+        # (num_blocks, 2, block_size, num_kv_heads, head_size)
+        return (0, 1, 2, 3, 4)
+
+    @staticmethod
+    def use_cascade_attention(*args, **kwargs) -> bool:
+        return False
+
+    @staticmethod
+    def get_builder_cls() -> type["TritonAttentionMetadataBuilder"]:
+        return TritonAttentionMetadataBuilder
+
+    @classmethod
+    def supports_head_size(cls, head_size: int) -> bool:
+        return head_size >= 32
+
+    @classmethod
+    def supports_mm_prefix(cls) -> bool:
+        return True
+
+    @classmethod
+    def supports_sink(cls) -> bool:
+        return True
+
+    @classmethod
+    def supports_attn_type(cls, attn_type: str) -> bool:
+        """TritonAttention supports all attention types."""
+        return attn_type in (
+            AttentionType.DECODER,
+            AttentionType.ENCODER,
+            AttentionType.ENCODER_ONLY,
+            AttentionType.ENCODER_DECODER,
+        )
+
+    @classmethod
+    def supports_alibi_sqrt(cls) -> bool:
+        return True
+
+    @classmethod
+    def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
+        return True
+
+
+class TritonAttentionImpl(AttentionImpl):
+    def fused_output_quant_supported(self, quant_key: QuantKey):
+        return quant_key == kFp8StaticTensorSym
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None = None,
+        attn_type: AttentionType = AttentionType.DECODER,
+        kv_sharing_target_layer_name: int | None = None,
+        sinks: torch.Tensor | None = None,
+        use_alibi_sqrt: bool = False,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        if sliding_window is None:
+            self.sliding_window = (-1, -1)
+        elif attn_type in (AttentionType.ENCODER, AttentionType.ENCODER_ONLY):
+            self.sliding_window = (sliding_window - 1, sliding_window - 1)
+        else:
+            self.sliding_window = (sliding_window - 1, 0)
+        self.kv_cache_dtype = kv_cache_dtype
+        if logits_soft_cap is None:
+            # In flash-attn, setting logits_soft_cap as 0 means no soft cap.
+            logits_soft_cap = 0
+        self.logits_soft_cap = logits_soft_cap
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
+
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        self.attn_type = attn_type
+        self.fp8_dtype = current_platform.fp8_dtype()
+
+        self.sinks = sinks
+        if sinks is not None:
+            assert sinks.shape[0] == num_heads, (
+                "Sinks must have the same number of heads as the number of "
+                f"heads in the layer. Sinks shape: {sinks.shape}, "
+                f"num_heads: {num_heads}."
+            )
+        self.use_alibi_sqrt = use_alibi_sqrt
+        self.supports_quant_query_input = current_platform.is_cuda()
+
+    def forward(
+        self,
+        layer: torch.nn.Module,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: TritonAttentionMetadata,
+        output: torch.Tensor | None = None,
+        output_scale: torch.Tensor | None = None,
+        output_block_scale: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        """Forward pass with Paged Attention impl. in Triton.
+
+        Args:
+            query: shape = [num_tokens, num_heads, head_size]
+            key: shape = [num_tokens, num_kv_heads, head_size]
+            value: shape = [num_tokens, num_kv_heads, head_size]
+            kv_cache: shape =
+                [num_blocks, 2, block_size, num_kv_heads, head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        assert output is not None, "Output tensor must be provided."
+
+        if output_block_scale is not None:
+            raise NotImplementedError(
+                "fused block_scale output quantization is not yet supported"
+                " for TritonAttentionImpl"
+            )
+
+        if attn_metadata is None:
+            # Profiling run.
+            return output.fill_(0)
+
+        assert attn_metadata.use_cascade is False
+
+        # IMPORTANT!
+        # NOTE(woosuk): With piece-wise CUDA graphs, this method is executed in
+        # eager-mode PyTorch. Thus, we need to be careful about any CPU overhead
+        # in this method. For example, `view` and `slice` (or `[:n]`) operations
+        # are surprisingly slow even in the case they do not invoke any GPU ops.
+        # Minimize the PyTorch ops in this method as much as possible.
+        # Whenever making a change in this method, please benchmark the
+        # performance to make sure it does not introduce any overhead.
+
+        num_actual_tokens = attn_metadata.num_actual_tokens
+
+        # Handle encoder attention differently - no KV cache needed
+        if self.attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER):
+            # For encoder attention,
+            # we use direct Q, K, V tensors without caching
+            return self._forward_encoder_attention(
+                query[:num_actual_tokens],
+                key[:num_actual_tokens],
+                value[:num_actual_tokens],
+                output[:num_actual_tokens],
+                attn_metadata,
+                layer,
+            )
+
+        # For decoder and cross-attention, use KV cache as before
+        key_cache, value_cache = kv_cache.unbind(1)
+        if self.kv_cache_dtype.startswith("fp8"):
+            if key_cache.dtype != self.fp8_dtype:
+                key_cache = key_cache.view(self.fp8_dtype)
+                value_cache = value_cache.view(self.fp8_dtype)
+            assert layer._q_scale_float == 1.0, (
+                "A non 1.0 q_scale is not currently supported."
+            )
+
+        cu_seqlens_q = attn_metadata.query_start_loc
+        seqused_k = attn_metadata.seq_lens
+        max_seqlen_q = attn_metadata.max_query_len
+        max_seqlen_k = attn_metadata.max_seq_len
+        block_table = attn_metadata.block_table
+
+        seq_threshold_3D = attn_metadata.seq_threshold_3D
+        num_par_softmax_segments = attn_metadata.num_par_softmax_segments
+        softmax_segm_output = attn_metadata.softmax_segm_output
+        softmax_segm_max = attn_metadata.softmax_segm_max
+        softmax_segm_expsum = attn_metadata.softmax_segm_expsum
+
+        descale_shape = (cu_seqlens_q.shape[0] - 1, key_cache.shape[2])
+        mm_prefix_range_tensor = attn_metadata.mm_prefix_range_tensor
+
+        unified_attention(
+            q=query[:num_actual_tokens],
+            k=key_cache,
+            v=value_cache,
+            out=output[:num_actual_tokens],
+            cu_seqlens_q=cu_seqlens_q,
+            max_seqlen_q=max_seqlen_q,
+            seqused_k=seqused_k,
+            max_seqlen_k=max_seqlen_k,
+            softmax_scale=self.scale,
+            causal=True,
+            alibi_slopes=self.alibi_slopes,
+            use_alibi_sqrt=self.use_alibi_sqrt,
+            window_size=self.sliding_window,
+            block_table=block_table,
+            softcap=self.logits_soft_cap,
+            q_descale=None,  # Not supported
+            k_descale=layer._k_scale.expand(descale_shape),
+            v_descale=layer._v_scale.expand(descale_shape),
+            seq_threshold_3D=seq_threshold_3D,
+            num_par_softmax_segments=num_par_softmax_segments,
+            softmax_segm_output=softmax_segm_output,
+            softmax_segm_max=softmax_segm_max,
+            softmax_segm_expsum=softmax_segm_expsum,
+            sinks=self.sinks,
+            output_scale=output_scale,
+            mm_prefix_range=mm_prefix_range_tensor,
+        )
+
+        return output
+
+    def _forward_encoder_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        output: torch.Tensor,
+        attn_metadata: TritonAttentionMetadata,
+        layer: torch.nn.Module,
+    ) -> torch.Tensor:
+        """Forward pass for encoder attention without KV cache.
+
+        Args:
+            query: shape = [num_encoder_tokens, num_heads, head_size]
+            key: shape = [num_encoder_tokens, num_kv_heads, head_size]
+            value: shape = [num_encoder_tokens, num_kv_heads, head_size]
+            output: shape = [num_encoder_tokens, num_heads, head_size]
+            attn_metadata: Encoder attention metadata
+            layer: The attention layer
+        """
+        # For encoder attention, process FP8 quantization if needed
+        if self.kv_cache_dtype.startswith("fp8"):
+            raise NotImplementedError(
+                "quantization is not supported for encoder attention"
+            )
+
+        # Use encoder-specific metadata for sequence information
+        query_start_loc = attn_metadata.query_start_loc
+        seq_lens = attn_metadata.seq_lens
+        max_query_len = attn_metadata.max_query_len
+
+        # Call flash attention directly on Q, K, V tensors
+        context_attention_fwd(
+            q=query,
+            k=key,
+            v=value,
+            o=output,
+            b_start_loc=query_start_loc,
+            b_seq_len=seq_lens,
+            max_input_len=max_query_len,
+            is_causal=False,  # Encoder attention is bidirectional
+            softmax_scale=self.scale,
+            sliding_window_q=self.sliding_window[0],
+            sliding_window_k=self.sliding_window[1],
+        )
+        return output
+
+    def do_kv_cache_update(
+        self,
+        layer: AttentionLayer,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+    ):
+        if self.attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER):
+            # For encoder attention,
+            # we use direct Q, K, V tensors without caching
+            return
+        # For decoder and cross-attention, use KV cache as before
+        key_cache, value_cache = kv_cache.unbind(1)
+
+        # Reshape the input keys and values and store them in the cache.
+        if self.kv_cache_dtype.startswith("fp8"):
+            key_cache = key_cache.view(self.fp8_dtype)
+            value_cache = value_cache.view(self.fp8_dtype)
+            # triton kernel does not support uint8 kv_cache
+            #  (because some explicit casts (e.g. float8_e4m3fnuz)
+            #   are not supported)
+        triton_reshape_and_cache_flash(
+            key,
+            value,
+            key_cache,
+            value_cache,
+            slot_mapping,
+            self.kv_cache_dtype,
+            layer._k_scale,
+            layer._v_scale,
+        )
+
+    def fused_rope_kvcache_supported(self):
+        return rocm_aiter_ops.is_enabled()
+
+    def do_rope_and_kv_cache_update(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        positions: torch.Tensor,
+        cos_sin_cache: torch.Tensor,
+        is_neox: bool,
+        kv_cache: torch.Tensor,
+        layer_slot_mapping: torch.Tensor,
+    ):
+        key_cache, value_cache = kv_cache.unbind(1)
+        flash_layout = True
+
+        is_fp8_kv_cache = self.kv_cache_dtype.startswith("fp8")
+        if is_fp8_kv_cache:
+            key_cache = key_cache.view(self.fp8_dtype)
+            value_cache = value_cache.view(self.fp8_dtype)
+
+        rocm_aiter_ops.triton_rope_and_cache(
+            query,
+            key,
+            value,
+            positions,
+            cos_sin_cache,
+            is_neox,
+            key_cache,
+            value_cache,
+            layer_slot_mapping,
+            layer._k_scale,
+            layer._v_scale,
+            flash_layout,
+            is_fp8_kv_cache,
+        )

vllm/v1/attention/backends/utils.py

@@ -0,0 +1,866 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import functools
+from collections.abc import Callable
+from dataclasses import dataclass, field, fields, make_dataclass
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Literal,
+    Protocol,
+    get_args,
+)
+
+import numpy as np
+import torch
+from typing_extensions import runtime_checkable
+
+from vllm.config import VllmConfig, get_layers_from_vllm_config
+from vllm.utils.math_utils import cdiv
+from vllm.v1.kv_cache_interface import KVCacheSpec, MambaSpec
+
+if TYPE_CHECKING:
+    from vllm.v1.core.sched.output import SchedulerOutput
+    from vllm.v1.worker.gpu_input_batch import InputBatch
+
+import vllm.envs as envs
+from vllm.distributed.kv_transfer.kv_connector.utils import (
+    get_kv_connector_cache_layout,
+)
+from vllm.logger import init_logger
+from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
+from vllm.v1.attention.backend import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+    CommonAttentionMetadata,
+    subclass_attention_backend,
+)
+
+logger = init_logger(__name__)
+KVCacheLayoutType = Literal["NHD", "HND"]
+_KV_CACHE_LAYOUT_OVERRIDE: KVCacheLayoutType | None = None
+
+PAD_SLOT_ID = -1
+
+
+def is_valid_kv_cache_layout(value: str) -> bool:
+    return value in get_args(KVCacheLayoutType)
+
+
+@functools.lru_cache
+def get_kv_cache_layout():
+    # Format specified by the code.
+    global _KV_CACHE_LAYOUT_OVERRIDE
+
+    cache_layout: Literal["NHD", "HND"] | None = None
+    if _KV_CACHE_LAYOUT_OVERRIDE is not None:
+        cache_layout = _KV_CACHE_LAYOUT_OVERRIDE
+        logger.info_once(
+            "`_KV_CACHE_LAYOUT_OVERRIDE` variable detected. "
+            "Setting KV cache layout to %s.",
+            cache_layout,
+        )
+        return cache_layout
+
+    # Format specified by the user.
+    cache_layout = envs.VLLM_KV_CACHE_LAYOUT
+    # When neither the user nor the override specified a layout, get default
+    if cache_layout is None:
+        cache_layout = get_kv_connector_cache_layout()
+    else:
+        assert is_valid_kv_cache_layout(cache_layout)
+        logger.info_once(
+            "`VLLM_KV_CACHE_LAYOUT` environment variable "
+            "detected. Setting KV cache layout to %s.",
+            cache_layout,
+        )
+    return cache_layout
+
+
+def set_kv_cache_layout(cache_layout: KVCacheLayoutType):
+    global _KV_CACHE_LAYOUT_OVERRIDE
+    _KV_CACHE_LAYOUT_OVERRIDE = cache_layout
+
+
+@dataclass
+class PerLayerParameters:
+    """
+    Currently, FlashInfer backend only support models in which all layers share
+    the same values for the following hyperparameters. Should not be used for
+    trtllm-gen backend since it supports different values for the following
+    hyperparameters.
+    """
+
+    window_left: int
+    logits_soft_cap: float | None
+    sm_scale: float
+    has_sinks: bool = False
+    # has same params for all layers
+    has_same_window_lefts: bool | None = field(default=None, compare=False)
+    has_same_all_params: bool | None = field(default=None, compare=False)
+
+
+def get_per_layer_parameters(
+    vllm_config: VllmConfig, layer_names: list[str], cls_: type["AttentionImpl"]
+) -> dict[str, PerLayerParameters]:
+    """
+    Scan layers in `layer_names` and determine some hyperparameters
+    to use during `plan`.
+    """
+
+    layers = get_layers_from_vllm_config(
+        vllm_config,
+        AttentionLayerBase,  # type: ignore[type-abstract]
+        layer_names,
+    )
+    per_layer_params: dict[str, PerLayerParameters] = {}
+
+    for key, layer in layers.items():
+        impl = layer.impl
+        assert isinstance(impl, cls_)
+
+        # Infer hyperparameters from the attention layer
+        window_size = getattr(impl, "sliding_window", None)
+        window_left = window_size[0] if window_size is not None else -1
+        logits_soft_cap = getattr(impl, "logits_soft_cap", None)
+        sm_scale = impl.scale
+        has_sinks = getattr(impl, "sinks", None) is not None
+
+        per_layer_params[key] = PerLayerParameters(
+            window_left, logits_soft_cap, sm_scale, has_sinks
+        )
+
+    return per_layer_params
+
+
+def infer_global_hyperparameters(
+    per_layer_params: dict[str, PerLayerParameters],
+) -> PerLayerParameters:
+    """
+    Currently, FlashInfer backend other than trtllm-gen
+    only support models in which all layers share
+    the same values for the following hyperparameters:
+    - `window_left`
+    - `logits_soft_cap`
+    - `sm_scale`
+
+    So this function asserts that all layers share the same values for these
+    hyperparameters and returns the global values.
+    """
+
+    assert len(per_layer_params) > 0, "No attention layers found in the model."
+
+    param_sets = list(per_layer_params.values())
+    global_params = param_sets[0]
+
+    global_params.has_same_window_lefts = all(
+        params.window_left == global_params.window_left for params in param_sets
+    )
+    global_params.has_same_all_params = all(
+        params == global_params for params in param_sets
+    )
+
+    return global_params
+
+
+#
+# Take in `query_start_loc_np` and `seq_lens_np` and break the sequences into
+# local attention blocks, where each block is passed to the attention kernel
+# as an independent local ("virtual") batch item.
+#
+# For example, if are performing a chunked prefill a batch of 3 sequences:
+#   q_seqlens  = [4, 10, 5]
+#   kv_seqlens = [6, 17, 9]
+# Then normally for regular attention we would compute with an attention mask
+#  for batch idx 0 (q_seqlens = 4, kv_seqlens = 6) like:
+#   batch idx: 0 (q_seqlens = 4, kv_seqlens = 6)
+#        k_toks >   0 1 2 3 4 5
+#        q_toks v  _____________
+#               0 | 1 1 1
+#               1 | 1 1 1 1
+#               2 | 1 1 1 1 1
+#               3 | 1 1 1 1 1 1
+#
+# for local attention (with attn_chunk_size = 4) we would compute with an
+#  attention mask like:
+#   batch idx: 0  (q_seqlens = 4, kv_seqlens = 6, attn_chunk_size = 4)
+#        k_toks >   0 1 2 3 4 5
+#        q_toks v  _____________
+#               0 | 1 1 1
+#               1 | 1 1 1 1
+#               2 |         1
+#               3 |         1 1
+#
+# We can simulate this mask using standard flash-attention by breaking the
+#  sequences into local ("virtual") batches, where each local batch item is a
+#  local attention block, so in this case batch idx 0 would be broken up into:
+#
+#   local-batch idx: 0 (q_seqlens = 2, kv_seqlens = 4)  (batch 0)
+#        k_toks >   0 1 2 3
+#        q_toks v  _____________
+#               0 | 1 1 1
+#               1 | 1 1 1 1
+#   local-batch idx: 1 (q_seqlens = 2, kv_seqlens = 2) (batch 0)
+#        k_toks >   4 5
+#        q_toks v  _____________
+#               2 | 1
+#               3 | 1 1
+#
+# e.g. if we have:
+#   attn_chunk_size = 4
+#   query_start_loc_np = [0, 4, 14, 19] (q_seqlens = [4, 10, 5])
+# Then this function would return:
+#                           __b0__  ______b1______  __b2__ < orig batch indices
+#   q_seqlens_local    = [   2,  2,  1,  4,  4,  1,  4,  1]
+#   cu_seqlens_q_local = [0, 4,  6, 10, 14, 18, 19, 23, 24]
+#   seqlens_k_local    = [   4,  2,  4,  4,  4,  1,  4,  1]
+#   block_table_local  : shape[local_virtual_batches, pages_per_local_batch]
+def make_local_attention_virtual_batches(
+    attn_chunk_size: int,
+    common_attn_metadata: CommonAttentionMetadata,
+    block_size: int = 0,
+) -> tuple[CommonAttentionMetadata, Callable[[torch.Tensor], torch.Tensor]]:
+    query_start_loc_np = common_attn_metadata.query_start_loc_cpu.numpy()
+    seq_lens_np = common_attn_metadata.seq_lens_cpu.numpy()
+    block_table = common_attn_metadata.block_table_tensor
+    device = common_attn_metadata.query_start_loc.device
+
+    q_seqlens = query_start_loc_np[1:] - query_start_loc_np[:-1]
+    actual_batch_size = seq_lens_np.shape[0]
+
+    # Handle if we are starting in the middle of a local attention block,
+    #  we assume q_seqlens > 0 (for all elements), for each batch idx we compute
+    #  the number of tokens that are not in the first local attention block and
+    #  then we can simply use a cdiv for the rest.
+    # For example if we have:
+    #   attn_chunk_size = 4
+    #   q_seqlens = [4, 10, 5]
+    #   k_seqlens = [6, 17, 9]
+    # Then we would get:
+    #   new_tokens_in_first_block = [2, 1, 4]
+    #   local_blocks = [2, 4, 2]
+    q_tokens_in_first_block = np.minimum(
+        attn_chunk_size - ((seq_lens_np - q_seqlens) % attn_chunk_size), q_seqlens
+    ).astype(np.int32)
+    tokens_in_last_block = attn_chunk_size + (seq_lens_np % -attn_chunk_size)
+    local_blocks = 1 + cdiv(q_seqlens - q_tokens_in_first_block, attn_chunk_size)
+
+    # Once we know the number of local blocks we can compute the request spans
+    #  for each batch idx, we can figure out the number of "virtual" requests we
+    #  have to make,
+    # For the above example we would get:
+    #   seqlens_q_local = [2, 2, 1, 4, 4, 1, 4, 1]
+    #
+    # First Get batched arange. (E.g., [2, 4, 2] -> [0, 1, 0, 1, 2, 3, 0, 1])
+    #   (TODO: max a utility to share this code with _prepare_inputs)
+    # arange step 1. [2, 4, 2] -> [2, 6, 8]
+    cu_num_blocks = np.cumsum(local_blocks)
+    virtual_batches = cu_num_blocks[-1]
+    # arange step 2. [2, 6, 8] -> [0, 0, 2, 2, 2, 2, 6, 6]
+    block_offsets = np.repeat(cu_num_blocks - local_blocks, local_blocks)
+    # arange step 3. [0, 1, 0, 1, 2, 3, 0, 1]
+    arange = np.arange(virtual_batches, dtype=np.int32) - block_offsets
+    # also compute reverse arange (i.e. [1, 0, 3, 2, 1, 0, 1, 0])
+    rarange = np.repeat(local_blocks, local_blocks) - arange - 1
+    # Then we can compute the seqlens_q_local, handling the fact that the
+    #  first and last blocks could be partial
+    seqlens_q_local = np.repeat(q_seqlens - q_tokens_in_first_block, local_blocks)
+    # set the first block since this may be a partial block
+    seqlens_q_local[arange == 0] = q_tokens_in_first_block
+    # set the remaining blocks
+    seqlens_q_local[arange > 0] = np.minimum(
+        seqlens_q_local - attn_chunk_size * (arange - 1), attn_chunk_size
+    )[arange > 0]
+
+    # convert from q_seqlens to cu_seqlens_q
+    cu_seqlens_q_local = np.empty(virtual_batches + 1, dtype=np.int32)
+    np.cumsum(seqlens_q_local, out=cu_seqlens_q_local[1:])
+    cu_seqlens_q_local[0] = 0
+
+    # compute the seqlens_k_local,
+    #  basically a full local attention block for all but the last block in each
+    #  batch
+    # For our example this will be:
+    #   seqlens_k_local = [4, 2, 4, 4, 4, 1, 4, 1]
+    seqlens_k_local = np.full(cu_num_blocks[-1], attn_chunk_size, dtype=np.int32)
+    seqlens_k_local[cu_num_blocks - 1] = tokens_in_last_block
+    num_computed_tokens_local = seqlens_k_local - seqlens_q_local
+
+    k_seqstarts_absolute = np.repeat(seq_lens_np, local_blocks) - (
+        rarange * attn_chunk_size + np.repeat(tokens_in_last_block, local_blocks)
+    )
+    # For the example the local attention blocks start at:
+    #                           _b0_  _____b1_____  _b2_
+    #   k_seqstarts_absolute = [0, 4, 4, 8, 12, 16, 4, 8]
+    block_starts = k_seqstarts_absolute // block_size
+    assert attn_chunk_size % block_size == 0, (
+        f"attn_chunk_size {attn_chunk_size} is not divisible by block_size {block_size}"
+    )
+    pages_per_local_batch = attn_chunk_size // block_size
+
+    # Create a block_table for the local attention blocks
+    # For out example if we have a block-table like (assuming block_size=2):
+    #   block_table = [
+    #     [ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9],  < batch 0
+    #     [10, 11, 12, 13, 14, 15, 16, 17, 18, 19],  < batch 1
+    #     [20, 21, 22, 23, 24, 25, 26, 27, 28, 29],  < batch 2
+    #   ]
+    # Then for the local batches we would want a block-table like
+    #   block_table_local = [
+    #     [  0,  1 ], < local-batch 0, (batch 0, starting from k[0])
+    #     [  2,  3 ], < local-batch 1, (batch 0, starting from k[4])
+    #     [ 12, 13 ], < local-batch 2, (batch 1, starting from k[4])
+    #     [ 14, 15 ], < local-batch 3, (batch 1, starting from k[8])
+    #     [ 16, 17 ], < local-batch 4, (batch 1, starting from k[12])
+    #     [ 18, 19 ], < local-batch 5, (batch 1, starting from k[16])
+    #     [ 22, 23 ], < local-batch 6, (batch 2, starting from k[4])
+    #     [ 24, 25 ], < local-batch 7, (batch 2, starting from k[8])
+    #   ]
+    block_indices = block_starts[:, None] + np.arange(
+        pages_per_local_batch, dtype=np.int32
+    )
+    block_indices = block_indices.reshape(-1).clip(max=block_table.shape[1] - 1)
+    batch_indices = np.repeat(
+        np.arange(actual_batch_size, dtype=np.int32),
+        local_blocks * pages_per_local_batch,
+    )
+
+    # NOTE: https://github.com/pytorch/pytorch/pull/160256 causes performance
+    # regression when using numpy arrays (batch and block indices) to index into
+    # torch tensor (block_table). As a workaround, convert numpy arrays to torch
+    # tensor first, which recovers perf.
+    batch_indices_torch = torch.from_numpy(batch_indices)
+    block_indices_torch = torch.from_numpy(block_indices)
+
+    # Save as a lambda so we can return this for update_block_table
+    make_block_table = lambda block_table: block_table[
+        batch_indices_torch, block_indices_torch
+    ].view(virtual_batches, -1)
+    block_table_local = make_block_table(block_table)
+
+    query_start_loc_cpu = torch.from_numpy(cu_seqlens_q_local)
+    seq_lens_cpu = torch.from_numpy(seqlens_k_local)
+    max_seq_len = int(seq_lens_cpu.max())
+
+    return CommonAttentionMetadata(
+        query_start_loc_cpu=query_start_loc_cpu,
+        query_start_loc=query_start_loc_cpu.to(device=device, non_blocking=True),
+        seq_lens=seq_lens_cpu.to(device=device, non_blocking=True),
+        num_reqs=len(seq_lens_cpu),
+        num_actual_tokens=common_attn_metadata.num_actual_tokens,
+        max_query_len=seqlens_q_local.max(),
+        max_seq_len=max_seq_len,
+        block_table_tensor=block_table_local,
+        slot_mapping=common_attn_metadata.slot_mapping,
+        causal=True,
+        _seq_lens_cpu=seq_lens_cpu,
+        _num_computed_tokens_cpu=torch.from_numpy(num_computed_tokens_local),
+    ), make_block_table
+
+
+def make_kv_sharing_fast_prefill_common_attn_metadata(
+    common_attn_metadata: CommonAttentionMetadata,
+) -> CommonAttentionMetadata:
+    if common_attn_metadata.max_query_len == 1:
+        # All requests are decode (assume 1 token for now)
+        # Skip computing fast prefill path
+        return common_attn_metadata
+
+    assert common_attn_metadata.logits_indices_padded is not None
+    assert common_attn_metadata.num_logits_indices is not None
+
+    logits_indices_padded = common_attn_metadata.logits_indices_padded
+    num_logits_indices = common_attn_metadata.num_logits_indices
+    # Get rid of CUDAGraph padding, if any
+    logits_indices = logits_indices_padded[:num_logits_indices]
+    num_reqs = common_attn_metadata.num_reqs
+    query_start_loc = common_attn_metadata.query_start_loc
+    # Example inputs
+    # num_reqs: 3
+    # generation_indices:  [14, 18, 19, 27]
+    # query_start_loc: [0, 15, 20, 28]
+    # seq_lens:        [41, 31, 40]
+
+    # Find how many decode indices belong to each request
+    # request_ids: [0, 1, 1, 2]
+    request_ids = torch.bucketize(logits_indices, query_start_loc[1:], right=True)
+
+    # Figure out how many tokens are in each request
+    # num_decode_tokens: [1, 2, 1]
+    num_decode_tokens = torch.bincount(request_ids, minlength=num_reqs)
+
+    # Calculate new query_start_loc with tokens in generation_indices
+    # decode_query_start_loc: [0, 1, 3, 4]
+    decode_query_start_loc = torch.empty(
+        num_reqs + 1, device=query_start_loc.device, dtype=query_start_loc.dtype
+    )
+
+    decode_query_start_loc[0] = 0
+    decode_query_start_loc[1:] = torch.cumsum(num_decode_tokens, dim=0)
+    decode_max_query_len = int(num_decode_tokens.max().item())
+    total_num_decode_tokens = int(num_decode_tokens.sum().item())
+
+    common_attn_metadata = CommonAttentionMetadata(
+        query_start_loc=decode_query_start_loc,
+        query_start_loc_cpu=decode_query_start_loc.to("cpu", non_blocking=True),
+        seq_lens=common_attn_metadata.seq_lens,
+        num_reqs=num_reqs,
+        num_actual_tokens=total_num_decode_tokens,
+        max_query_len=decode_max_query_len,
+        max_seq_len=common_attn_metadata.max_seq_len,
+        block_table_tensor=common_attn_metadata.block_table_tensor,
+        slot_mapping=common_attn_metadata.slot_mapping,
+        causal=True,
+        _seq_lens_cpu=common_attn_metadata._seq_lens_cpu,
+        _num_computed_tokens_cpu=common_attn_metadata._num_computed_tokens_cpu,
+    )
+    return common_attn_metadata
+
+
+def split_decodes_prefills_and_extends(
+    common_attn_metadata: CommonAttentionMetadata,
+    decode_threshold: int = 1,
+) -> tuple[int, int, int, int, int, int]:
+    """
+    Assuming a reordered batch, finds the boundary between prefill and decode
+    requests.
+
+    Args:
+        common_attn_metadata: CommonAttentionMetadata object containing the
+            batch metadata.
+        decode_threshold: The maximum query length to be considered a decode.
+
+    Returns:
+        num_decodes: The number of decode requests.
+        num_extends: The number of extend requests.
+        num_prefills: The number of prefill requests.
+        num_decode_tokens: The number of tokens in the decode requests.
+        num_extend_tokens: The number of tokens in the extend requests.
+        num_prefill_tokens: The number of tokens in the prefill requests.
+    """
+    max_query_len = common_attn_metadata.max_query_len
+    num_reqs = common_attn_metadata.num_reqs
+    num_tokens = common_attn_metadata.num_actual_tokens
+    query_start_loc = common_attn_metadata.query_start_loc_cpu
+    seq_lens = common_attn_metadata.seq_lens_cpu
+
+    if max_query_len <= decode_threshold:
+        return num_reqs, 0, 0, num_tokens, 0, 0
+
+    query_lens = query_start_loc[1:] - query_start_loc[:-1]
+    is_prefill_or_extend = query_lens > decode_threshold
+    is_prefill = (seq_lens == query_lens) & is_prefill_or_extend
+    first_extend = is_prefill_or_extend.int().argmax(dim=-1).item()
+    first_prefill = is_prefill.int().argmax(dim=-1).item()
+    num_decodes = first_extend
+    num_decode_tokens = query_start_loc[first_extend].item()
+    if not torch.any(is_prefill_or_extend):
+        return (num_decodes, 0, 0, num_decode_tokens, 0, 0)
+
+    num_prefills_or_extends = num_reqs - num_decodes
+    num_prefill_or_extend_tokens = num_tokens - num_decode_tokens
+    if not torch.any(is_prefill):
+        return (
+            num_decodes,
+            num_prefills_or_extends,
+            0,
+            num_decode_tokens,
+            num_prefill_or_extend_tokens,
+            0,
+        )
+
+    num_extends = first_prefill - num_decodes
+    num_prefills = num_reqs - first_prefill
+
+    num_prefill_tokens = num_tokens - query_start_loc[first_prefill]
+    num_extend_tokens = num_prefill_or_extend_tokens - num_prefill_tokens
+    return (
+        num_decodes,
+        num_extends,
+        num_prefills,
+        num_decode_tokens,
+        num_extend_tokens,
+        num_prefill_tokens,
+    )
+
+
+def split_decodes_and_prefills(
+    common_attn_metadata: CommonAttentionMetadata,
+    decode_threshold: int = 1,
+    require_uniform: bool = False,
+) -> tuple[int, int, int, int]:
+    """
+    Assuming a reordered batch, finds the boundary between prefill and decode
+    requests.
+
+    Args:
+        common_attn_metadata: CommonAttentionMetadata object containing the
+            batch metadata.
+        decode_threshold: The maximum query length to be considered a decode.
+        require_uniform: If True, requires that all decode requests have the
+            same query length. When set, some queries may be considered prefills
+            even if they are <= decode_threshold, in order to ensure uniformity.
+
+    Returns:
+        num_decodes: The number of decode requests.
+        num_prefills: The number of prefill requests.
+        num_decode_tokens: The number of tokens in the decode requests.
+        num_prefill_tokens: The number of tokens in the prefill requests.
+    """
+    max_query_len = common_attn_metadata.max_query_len
+    num_reqs = common_attn_metadata.num_reqs
+    num_tokens = common_attn_metadata.num_actual_tokens
+    query_start_loc = common_attn_metadata.query_start_loc_cpu
+
+    if max_query_len <= decode_threshold and (
+        not require_uniform or decode_threshold <= 1
+    ):
+        return num_reqs, 0, num_tokens, 0
+
+    query_lens = query_start_loc[1:] - query_start_loc[:-1]
+    if query_lens[0].item() > decode_threshold:
+        # first request is not decode, so no decode requests
+        return 0, num_reqs, 0, num_tokens
+
+    if require_uniform:
+        # check if we are in a padded uniform batch; this is used for full-CGs, some
+        # requests may have a query length of 0 but since they are padding its fine
+        # to treat them as decodes (ensures num_decodes matches the captured size)
+        if torch.all((query_lens == query_lens[0]) | (query_lens == 0)):
+            assert num_reqs * query_lens[0] == num_tokens, "tokens not padded correctly"
+            return num_reqs, 0, num_tokens, 0  # all decodes
+        is_prefill = query_lens != query_lens[0]
+    else:
+        is_prefill = query_lens > decode_threshold
+
+    if not torch.any(is_prefill):
+        return num_reqs, 0, num_tokens, 0
+
+    first_prefill = is_prefill.int().argmax(dim=-1).item()
+    assert torch.all(query_lens[:first_prefill] <= decode_threshold)
+    num_decodes = first_prefill
+    num_prefills = num_reqs - num_decodes
+    num_decode_tokens = query_start_loc[first_prefill].item()
+    num_prefill_tokens = num_tokens - num_decode_tokens
+    return (num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens)
+
+
+def split_prefill_chunks(
+    seq_lens_cpu: torch.Tensor, workspace_size: int, request_offset: int = 0
+) -> list[tuple[int, int]]:
+    """
+    Split the prefill requests into chunks such that the total sequence length
+    of each chunk is less than or equal to the workspace size.
+
+    Args:
+        seq_lens_cpu: The sequence lengths of the prefill requests on CPU.
+        workspace_size: The maximum workspace size (in tokens) per chunk.
+        request_offset: The offset to add to the request indices.
+    Returns:
+        A list of tuples of (reqs_start, reqs_end) representing chunk boundaries.
+    """
+    chunk_bounds = []
+    i, n = 0, len(seq_lens_cpu)
+    assert torch.all(seq_lens_cpu <= workspace_size).item()
+
+    while i < n:
+        start, chunk_total = i, 0
+        while i < n and (chunk_total + (s := seq_lens_cpu[i].item())) <= workspace_size:
+            chunk_total += s
+            i += 1
+        chunk_bounds.append((start + request_offset, i + request_offset))
+    return chunk_bounds
+
+
+def reorder_batch_to_split_decodes_and_prefills(
+    input_batch: "InputBatch",
+    scheduler_output: "SchedulerOutput",
+    decode_threshold: int = 1,
+) -> bool:
+    """
+    Reorders the batch to split into prefill and decode requests; places all
+    requests with <= decode_threshold tokens at the front of the batch.
+
+    Returns:
+        True if the batch was modified, False otherwise.
+    """
+    # We now want to reorder the batch into decode → extend → prefill order
+    # where:
+    #   decode: request with num_scheduled_tokens <= decode_threshold
+    #   extend: non-decode request with existing context
+    #   prefill: non-decode request with no existing context
+    # NOTE for now we loosely use "decode" to mean requests where attention is
+    #  likely memory-bound and "prefill" to mean requests where attention is
+    #  likely compute-bound,
+    num_reqs = len(input_batch.req_ids)
+    num_scheduled_tokens = [
+        scheduler_output.num_scheduled_tokens[id] for id in input_batch.req_ids
+    ]
+    num_scheduled_tokens_np = np.array(num_scheduled_tokens)
+    num_computed_tokens_np = input_batch.num_computed_tokens_cpu[:num_reqs]
+
+    is_prefill = num_computed_tokens_np == 0
+    is_decode = (num_scheduled_tokens_np <= decode_threshold) & (~is_prefill)
+    is_extend = (num_scheduled_tokens_np > decode_threshold) & (~is_prefill)
+
+    # Desired order: decode → extend → prefill
+    req_regions = np.zeros(is_decode.shape, dtype=np.int32)  # 0 = decode by default
+    req_regions[is_extend] = 1
+    req_regions[is_prefill] = 2
+
+    num_decodes = int(is_decode.sum())
+    num_extends = int(is_extend.sum())
+
+    target_regions = np.zeros(num_reqs, dtype=np.int32)
+    target_regions[num_decodes : num_decodes + num_extends] = 1
+    target_regions[num_decodes + num_extends :] = 2
+
+    needs_swap = req_regions != target_regions
+
+    if not needs_swap.any():
+        return False
+
+    # Extract indices that need swapping and sort by target region
+    orig_indices = np.where(needs_swap)[0]
+    sorted_order = np.argsort(req_regions[needs_swap], kind="stable")
+    src_indices = orig_indices[sorted_order]
+
+    src_dest_map = {int(src): int(dst) for src, dst in zip(src_indices, orig_indices)}
+
+    for src in src_dest_map:
+        dst = src_dest_map[src]
+        while src != dst:
+            input_batch.swap_states(src, dst)
+            # Mark dst as done by updating its destination to itself
+            next_dst = src_dest_map.get(dst, dst)
+            src_dest_map[dst] = dst
+            dst = next_dst
+
+    return True
+
+
+def reshape_query_for_spec_decode(query: torch.Tensor, batch_size: int) -> torch.Tensor:
+    """
+    Reshapes the query tensor for the specified batch size, so that
+    it has shape (batch_size, seq_len, num_heads, head_dim).
+    """
+    assert query.dim() == 3, f"query must be 3D, got {query.dim()}D"
+    total_tokens = query.shape[0]
+    num_heads = query.shape[1]
+    head_dim = query.shape[2]
+    assert total_tokens % batch_size == 0, (
+        f"{total_tokens=} is not divisible by {batch_size=}"
+    )
+    seq_len = total_tokens // batch_size
+    return query.view(batch_size, seq_len, num_heads, head_dim)
+
+
+def reshape_attn_output_for_spec_decode(attn_output: torch.Tensor) -> torch.Tensor:
+    """
+    Reshapes the attention output tensor, so that
+    the batch_size and seq_len dimensions are combined.
+    """
+    if attn_output.dim() == 3:
+        # Already in the correct shape
+        return attn_output
+    assert attn_output.dim() == 4, f"attn_output must be 4D, got {attn_output.dim()}D"
+    total_tokens = attn_output.shape[0] * attn_output.shape[1]
+    return attn_output.view(total_tokens, attn_output.shape[2], attn_output.shape[3])
+
+
+def subclass_attention_metadata(
+    name_prefix: str,
+    metadata_cls: Any,
+    fields: list[tuple[str, Any, Any]],
+) -> Any:
+    """
+    Return a new subclass of `metadata_cls` with additional fields
+    """
+    name: str = name_prefix + metadata_cls.__name__  # type: ignore
+    Wrapped = make_dataclass(name, fields, bases=(metadata_cls,))
+    return Wrapped
+
+
+@runtime_checkable
+class KVSharingFastPrefillMetadata(Protocol):
+    logits_indices_padded: torch.Tensor | None = None
+    num_logits_indices: int | None = None
+
+
+def create_fast_prefill_custom_backend(
+    prefix: str,
+    underlying_attn_backend: type[AttentionBackend],
+) -> type[AttentionBackend]:
+    underlying_builder = underlying_attn_backend.get_builder_cls()
+
+    class FastPrefillAttentionBuilder(underlying_builder):  # type: ignore
+        def build(
+            self,
+            common_prefix_len: int,
+            common_attn_metadata: CommonAttentionMetadata,
+            fast_build: bool = False,
+        ) -> AttentionMetadata:
+            new_common_attn_metadata = (
+                make_kv_sharing_fast_prefill_common_attn_metadata(common_attn_metadata)
+            )
+            metadata = super().build(
+                common_prefix_len, new_common_attn_metadata, fast_build
+            )
+
+            class KVSharingFastPrefillAttentionMetadata(
+                metadata.__class__,  #  type: ignore
+                KVSharingFastPrefillMetadata,
+            ):
+                def __init__(self, metadata, common_attn_metadata):
+                    # Shallow copy all fields in metadata cls
+                    for _field in fields(metadata.__class__):
+                        setattr(self, _field.name, getattr(metadata, _field.name))
+
+                    self.logits_indices_padded = (
+                        common_attn_metadata.logits_indices_padded
+                    )
+                    self.num_logits_indices = common_attn_metadata.num_logits_indices
+
+            return KVSharingFastPrefillAttentionMetadata(metadata, common_attn_metadata)
+
+    attn_backend = subclass_attention_backend(
+        name_prefix=prefix,
+        attention_backend_cls=underlying_attn_backend,
+        builder_cls=FastPrefillAttentionBuilder,
+    )
+
+    return attn_backend
+
+
+def compute_causal_conv1d_metadata(
+    query_start_loc_p_cpu: torch.Tensor,
+    *,
+    device: torch.device,
+):
+    # Needed for causal_conv1d. Use the CPU query_start_loc to avoid DtoH sync.
+    assert query_start_loc_p_cpu.device.type == "cpu"
+    seqlens = query_start_loc_p_cpu.diff()
+    nums_dict = {}  # type: ignore
+    batch_ptr = None
+    token_chunk_offset_ptr = None
+    for BLOCK_M in [8]:  # cover all BLOCK_M values
+        nums = -(-seqlens // BLOCK_M)
+        nums_dict[BLOCK_M] = {}
+        nums_dict[BLOCK_M]["nums"] = nums
+        nums_dict[BLOCK_M]["tot"] = nums.sum().item()
+        mlist = torch.from_numpy(np.repeat(np.arange(len(nums)), nums))
+        nums_dict[BLOCK_M]["mlist"] = mlist
+        mlist_len = len(nums_dict[BLOCK_M]["mlist"])
+        nums_dict[BLOCK_M]["mlist_len"] = mlist_len
+        MAX_NUM_PROGRAMS = max(1024, mlist_len) * 2
+        offsetlist = []  # type: ignore
+        for idx, num in enumerate(nums):
+            offsetlist.extend(range(num))
+        offsetlist = torch.tensor(offsetlist, dtype=torch.int32)
+        nums_dict[BLOCK_M]["offsetlist"] = offsetlist
+
+        if batch_ptr is None:
+            # Update default value after class definition
+            batch_ptr = torch.full(
+                (MAX_NUM_PROGRAMS,), PAD_SLOT_ID, dtype=torch.int32, device=device
+            )
+            token_chunk_offset_ptr = torch.full(
+                (MAX_NUM_PROGRAMS,), PAD_SLOT_ID, dtype=torch.int32, device=device
+            )
+        else:
+            if batch_ptr.nelement() < MAX_NUM_PROGRAMS:
+                batch_ptr.resize_(MAX_NUM_PROGRAMS).fill_(PAD_SLOT_ID)
+                token_chunk_offset_ptr.resize_(  # type: ignore
+                    MAX_NUM_PROGRAMS
+                ).fill_(PAD_SLOT_ID)
+
+        batch_ptr[0:mlist_len].copy_(mlist, non_blocking=True)
+        token_chunk_offset_ptr[  # type: ignore
+            0:mlist_len
+        ].copy_(offsetlist, non_blocking=True)
+        nums_dict[BLOCK_M]["batch_ptr"] = batch_ptr
+        nums_dict[BLOCK_M]["token_chunk_offset_ptr"] = token_chunk_offset_ptr  # type: ignore
+
+    return nums_dict, batch_ptr, token_chunk_offset_ptr
+
+
+def get_dcp_local_seq_lens(
+    seq_lens: torch.Tensor,
+    dcp_size: int = 1,
+    dcp_rank: int | None = None,
+    cp_kv_cache_interleave_size: int = 1,
+) -> torch.Tensor:
+    """While using dcp, kv_cache size stored on each rank may be different,
+    use this function to calculate split decode seq_lens of each dcp rank.
+    Only consider dcp now, we can extend the case of cp based on this.
+    """
+    num_requests = seq_lens.size(0)
+    if dcp_rank is None:
+        rank_offsets = (
+            torch.arange(dcp_size, dtype=torch.int32, device=seq_lens.device)
+            .unsqueeze(0)
+            .repeat(num_requests, 1)
+        )
+    else:
+        rank_offsets = torch.tensor(
+            [[dcp_rank]], dtype=torch.int32, device=seq_lens.device
+        )
+    seq_lens_tiled = (
+        seq_lens.to(torch.int32).unsqueeze(-1).repeat(1, rank_offsets.shape[1])
+    )
+    base = (
+        seq_lens_tiled
+        // cp_kv_cache_interleave_size
+        // dcp_size
+        * cp_kv_cache_interleave_size
+    )
+    remainder = seq_lens_tiled - base * dcp_size
+    remainder = torch.clip(
+        remainder - rank_offsets * cp_kv_cache_interleave_size,
+        0,
+        cp_kv_cache_interleave_size,
+    )
+    dcp_local_seq_lens = base + remainder
+    return dcp_local_seq_lens.squeeze(1)
+
+
+def mamba_get_block_table_tensor(
+    block_table: torch.Tensor,
+    seq_lens: torch.Tensor,
+    kv_cache_spec: KVCacheSpec,
+    mamba_cache_mode: str,
+) -> torch.Tensor:
+    """
+    Get the block table tensor for mamba kernels from the input
+    common_attn_metadata.block_table_tensor given different mamba cache modes.
+
+    - "all":   input  (#requests, cdiv(max_model_len, block_size));
+               output (#requests, cdiv(max_model_len, block_size)).
+
+    - "none":  input  (#requests, 1 + num_speculative_blocks);
+               output (#requests, 1 + num_speculative_blocks).
+
+    - "align": input  (#requests, cdiv(max_model_len, block_size));
+               output (#requests, 1 + num_speculative_blocks), which are the last
+               1 + num_speculative_blocks of each request.
+    """
+    if mamba_cache_mode in ("all", "none"):
+        return block_table
+    else:
+        assert isinstance(kv_cache_spec, MambaSpec)
+        # NOTE: For 0-length requests in CUDA graph, use a start_index of 0
+        # to handle the invalid block table.
+        start_indices = torch.clamp(
+            (seq_lens - 1) // kv_cache_spec.block_size,
+            min=0,
+        )
+        # Use int32 for arithmetic to avoid dtype promotion overhead,
+        # then convert to int64 for gather (which requires Long indices)
+        offsets = torch.arange(
+            1 + kv_cache_spec.num_speculative_blocks,
+            device=block_table.device,
+            dtype=torch.int32,
+        )
+        indices_to_gather = (start_indices.unsqueeze(1) + offsets).to(torch.int64)
+        return torch.gather(block_table, 1, indices_to_gather)

vllm/v1/attention/ops/chunked_prefill_paged_decode.py

@@ -0,0 +1,460 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Authors:
+#  - Burkhard Ringlein <ngl@zurich.ibm.com>
+#  - Jan van Lunteren <jvl@zurich.ibm.com>
+#  - Chih-Chieh Yang <chih.chieh.yang@ibm.com>
+#  - Thomas Parnell <tpa@zurich.ibm.com>
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
+
+from .prefix_prefill import context_attention_fwd
+
+float8_info = torch.finfo(current_platform.fp8_dtype())
+
+
+@triton.jit
+def cdiv_fn(x, y):
+    return (x + y - 1) // y
+
+
+@triton.jit
+def kernel_paged_attention_2d(
+    output_ptr,  # [num_tokens, num_query_heads, head_size]
+    query_ptr,  # [num_tokens, num_query_heads, head_size]
+    key_cache_ptr,  # [num_blks, num_kv_heads, head_size // x, blk_size, x]
+    value_cache_ptr,  # [num_blks, num_kv_heads, head_size, blk_size]
+    sink_ptr,  # [num_query_heads]
+    block_tables_ptr,  # [num_seqs, max_num_blocks_per_seq]
+    seq_lens_ptr,  # [num_seqs]
+    alibi_slopes_ptr,  # [num_query_heads]
+    scale,  # float32
+    k_scale,  # float32
+    v_scale,  # float32
+    out_scale_inv,
+    num_query_heads: tl.constexpr,  # int
+    num_queries_per_kv: tl.constexpr,  # int
+    num_queries_per_kv_padded: tl.constexpr,  # int
+    block_table_stride: tl.int64,  # int
+    query_stride_0: tl.int64,  # int
+    query_stride_1: tl.int64,  # int, should be equal to head_size
+    output_stride_0: tl.int64,  # int
+    output_stride_1: tl.int64,  # int, should be equal to head_size
+    BLOCK_SIZE: tl.constexpr,  # int
+    PHYSICAL_BLOCK_SIZE: tl.constexpr,  # int
+    HEAD_SIZE: tl.constexpr,  # int
+    HEAD_SIZE_PADDED: tl.constexpr,  # int, must be power of 2
+    USE_ALIBI_SLOPES: tl.constexpr,  # bool
+    SLIDING_WINDOW: tl.constexpr,  # int
+    x: tl.constexpr,  # int
+    stride_k_cache_0: tl.int64,  # int
+    stride_k_cache_1: tl.int64,  # int
+    stride_k_cache_2: tl.int64,  # int
+    stride_k_cache_3: tl.int64,  # int
+    stride_k_cache_4: tl.int64,  # int
+    stride_v_cache_0: tl.int64,  # int
+    stride_v_cache_1: tl.int64,  # int
+    stride_v_cache_2: tl.int64,  # int
+    stride_v_cache_3: tl.int64,  # int
+    filter_by_query_len: tl.constexpr,  # bool
+    query_start_len_ptr,  # [num_seqs+1]
+    USE_SINKS: tl.constexpr,  # bool
+    USE_FP8: tl.constexpr,
+    FP8_MIN: tl.constexpr = float8_info.min,
+    FP8_MAX: tl.constexpr = float8_info.max,
+):
+    seq_idx = tl.program_id(0)
+    kv_head_idx = tl.program_id(1)
+
+    if filter_by_query_len:
+        cur_batch_in_all_start_index = tl.load(query_start_len_ptr + seq_idx)
+        cur_batch_in_all_stop_index = tl.load(query_start_len_ptr + seq_idx + 1)
+        cur_batch_query_len = cur_batch_in_all_stop_index - cur_batch_in_all_start_index
+        if cur_batch_query_len > 1:
+            return
+    else:
+        cur_batch_in_all_start_index = seq_idx
+
+    query_head_idx = kv_head_idx * num_queries_per_kv + tl.arange(
+        0, num_queries_per_kv_padded
+    )
+
+    query_offset = (
+        cur_batch_in_all_start_index * query_stride_0
+        + query_head_idx[:, None] * query_stride_1
+    )
+
+    head_mask = query_head_idx < (kv_head_idx + 1) * num_queries_per_kv
+    head_mask = head_mask & (query_head_idx < num_query_heads)
+
+    dim_mask = tl.where(tl.arange(0, HEAD_SIZE_PADDED) < HEAD_SIZE, 1, 0).to(tl.int1)
+
+    # Q : (num_queries_per_kv, HEAD_SIZE,)
+    Q = tl.load(
+        query_ptr + query_offset + tl.arange(0, HEAD_SIZE_PADDED)[None, :],
+        mask=dim_mask[None, :] & head_mask[:, None],
+        other=0.0,
+    )
+
+    block_table_offset = seq_idx * block_table_stride
+
+    if not USE_SINKS:
+        M = tl.full([num_queries_per_kv_padded], float("-inf"), dtype=tl.float32)
+        L = tl.zeros([num_queries_per_kv_padded], dtype=tl.float32)
+    else:
+        M = tl.load(
+            sink_ptr + query_head_idx,
+            mask=head_mask,
+            other=float("-inf"),
+        ).to(dtype=tl.float32)
+        L = tl.where(float("-inf") < M, 1.0, 0.0)
+
+    acc = tl.zeros([num_queries_per_kv_padded, HEAD_SIZE_PADDED], dtype=tl.float32)
+
+    # sequence len for this particular sequence
+    seq_len = tl.load(seq_lens_ptr + seq_idx)
+
+    # alibi slope for this head
+    if USE_ALIBI_SLOPES:
+        alibi_slope = tl.load(
+            alibi_slopes_ptr + query_head_idx, mask=head_mask, other=0.0
+        )
+
+    num_blocks = cdiv_fn(seq_len, BLOCK_SIZE)
+
+    offs_n = tl.arange(0, BLOCK_SIZE)
+    offs_d = tl.arange(0, HEAD_SIZE_PADDED)
+    # iterate through tiles
+    for j in range(0, num_blocks):
+        start_n = j * BLOCK_SIZE
+        # Calculate the logical location within a non-standard physical block,
+        # such as 544 in Qwen/Qwen3-Next-80B-A3B-Thinking.
+        # Supports non-contiguous mapping
+        # from logical blocks to physical blocks
+        abs_token_idx = start_n + offs_n
+        l_block_idx = abs_token_idx // PHYSICAL_BLOCK_SIZE
+        # Vectorized loading of physical block IDs
+        p_block_idx = tl.load(block_tables_ptr + block_table_offset + l_block_idx)
+        internal_offsets = abs_token_idx % PHYSICAL_BLOCK_SIZE
+
+        # 5D addressing logic of K
+        k_offset = (
+            p_block_idx[None, :] * stride_k_cache_0
+            + kv_head_idx * stride_k_cache_1
+            + (offs_d[:, None] // x) * stride_k_cache_2
+            + internal_offsets[None, :] * stride_k_cache_3
+            + (offs_d[:, None] % x) * stride_k_cache_4
+        )
+
+        # 4D addressing logic of V (Slot is innermost)
+        v_offset = (
+            p_block_idx[:, None] * stride_v_cache_0
+            + kv_head_idx * stride_v_cache_1
+            + offs_d[None, :] * stride_v_cache_2
+            + internal_offsets[:, None] * stride_v_cache_3
+        )
+
+        # K : (HEAD_SIZE, BLOCK_SIZE)
+        K_load = tl.load(
+            key_cache_ptr + k_offset,
+            mask=dim_mask[:, None],
+            other=0.0,
+            eviction_policy="evict_last",
+        )
+
+        if K_load.dtype.is_fp8():
+            K = (K_load.to(tl.float32) * tl.load(k_scale)).to(Q.dtype)
+        else:
+            K = K_load
+
+        # V : (BLOCK_SIZE, HEAD_SIZE)
+        V_load = tl.load(
+            value_cache_ptr + v_offset,
+            mask=dim_mask[None, :],
+            other=0.0,
+            eviction_policy="evict_last",
+        )
+
+        if V_load.dtype.is_fp8():
+            V = (V_load.to(tl.float32) * tl.load(v_scale)).to(Q.dtype)
+        else:
+            V = V_load
+
+        seq_offset = j * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+        boundary = tl.full([BLOCK_SIZE], seq_len, dtype=tl.int32)
+        seq_mask = seq_offset[None, :] < boundary
+
+        # First calculate the dot, then apply the mask.
+        qk = scale * tl.dot(Q, K)
+        S = tl.where(head_mask[:, None] & seq_mask, qk, float("-inf"))
+
+        context_len = seq_len - 1
+
+        if SLIDING_WINDOW > 0:
+            S = tl.where((context_len - seq_offset) < SLIDING_WINDOW, S, -10000)
+
+        if USE_ALIBI_SLOPES:
+            S += alibi_slope[:, None] * (seq_offset - context_len)
+
+        # compute running maximum
+        # m_j : (num_queries_per_kv,)
+        m_j = tl.maximum(M, tl.max(S, axis=1))
+
+        # P : (num_queries_per_kv, BLOCK_SIZE,)
+        p = tl.exp(S - m_j[:, None])
+        p = tl.where(m_j[:, None] == float("-inf"), 0.0, p)
+
+        # l_j : (num_queries_per_kv,)
+        l_j = tl.sum(p, axis=1)
+
+        # alpha : (num_queries_per_kv, )
+        alpha = tl.exp(M - m_j)
+        alpha = tl.where(float("-inf") == M, 0.0, alpha)
+
+        # acc : (num_queries_per_kv, BLOCK_SIZE,)
+        acc = acc * alpha[:, None]
+
+        # update constants
+        L = L * alpha + l_j
+        M = m_j
+
+        # acc : (num_queries_per_kv, BLOCK_SIZE,)
+        acc += tl.dot(p.to(V.dtype), V)
+
+    # epilogue
+    acc = acc / (L[:, None] + 1e-10)
+    if USE_FP8:
+        acc = acc * tl.load(out_scale_inv)
+        acc = tl.clamp(acc, FP8_MIN, FP8_MAX)
+
+    output_offset = (
+        cur_batch_in_all_start_index * output_stride_0
+        + query_head_idx * output_stride_1
+    )
+
+    tl.store(
+        output_ptr + output_offset[:, None] + tl.arange(0, HEAD_SIZE_PADDED)[None, :],
+        acc,
+        mask=dim_mask[None, :] & head_mask[:, None],
+    )
+
+
+def chunked_prefill_paged_decode(
+    query,
+    key,
+    value,
+    output,
+    kv_cache_dtype,
+    key_cache,
+    value_cache,
+    block_table,
+    query_start_loc,
+    seq_lens,
+    max_seq_len,
+    max_query_len,
+    k_scale,
+    v_scale,
+    alibi_slopes=None,
+    sliding_window=None,
+    sm_scale=None,
+    output_scale=None,
+    # Optional tensor for sinks
+    sinks=None,
+    is_block_table_ptr: bool = False,
+):
+    if sm_scale is None:
+        sm_scale = 1.0 / (query.shape[2] ** 0.5)
+
+    use_alibi_slopes = alibi_slopes is not None
+
+    if sliding_window is None or sliding_window <= 0:
+        sliding_window = 0
+
+    if max_query_len > 1:
+        context_attention_fwd(
+            q=query,
+            k=key,
+            v=value,
+            o=output,
+            kv_cache_dtype=kv_cache_dtype,
+            k_cache=key_cache,
+            v_cache=value_cache,
+            b_loc=block_table,
+            b_start_loc=query_start_loc,
+            b_seq_len=seq_lens,
+            max_seq_len=max_seq_len,
+            max_input_len=max_query_len,
+            k_scale=k_scale,
+            v_scale=v_scale,
+            alibi_slopes=alibi_slopes,
+            sliding_window=sliding_window,
+            sm_scale=sm_scale,
+            skip_decode=True,
+            fp8_out_scale=output_scale,
+            sinks=sinks,
+        )
+
+    block_size = value_cache.shape[3]
+    num_seqs = len(seq_lens)
+    num_query_heads = query.shape[1]
+    # key may be None in cross-attention decode (already cached from encoder)
+    num_kv_heads = key.shape[1] if key is not None else key_cache.shape[1]
+    num_queries_per_kv = num_query_heads // num_kv_heads
+    head_size = query.shape[2]
+
+    # Conversion of FP8 Tensor from uint8 storage to
+    # appropriate torch.dtype for interpretation by Triton
+    if "fp8" in kv_cache_dtype:
+        assert key_cache.dtype in [torch.uint8, current_platform.fp8_dtype()]
+        assert value_cache.dtype in [torch.uint8, current_platform.fp8_dtype()]
+
+        if kv_cache_dtype in ("fp8", "fp8_e4m3"):
+            target_dtype = current_platform.fp8_dtype()
+        elif kv_cache_dtype == "fp8_e5m2":
+            target_dtype = torch.float8_e5m2
+        else:
+            raise ValueError(
+                f"Unsupported FP8 kv_cache_dtype {kv_cache_dtype}: "
+                f"should be one of 'fp8', 'fp8_e4m3', 'fp8_e5m2'."
+            )
+
+        key_cache = key_cache.view(target_dtype)
+        value_cache = value_cache.view(target_dtype)
+
+    num_queries_per_kv_padded = max(triton.next_power_of_2(num_queries_per_kv), 16)
+
+    from vllm.platforms.rocm import use_rocm_custom_paged_attention
+
+    use_custom = use_rocm_custom_paged_attention(
+        query.dtype,
+        head_size,
+        block_size,
+        num_queries_per_kv,
+        max_seq_len,
+        sliding_window,
+        kv_cache_dtype,
+        alibi_slopes,
+        sinks,
+    )
+    # Triton is only forced when encountering a non-standard block
+    # like Qwen3 with a size of 544.
+    # 1. Check if block_size is a power of 2 (16, 32, 64...)
+    # 2. If it's a power of 2, we trust the vLLM's native use_custom decision.
+    # 3. If it's not a power of 2 (such as Qwen3's 544),
+    # then our Triton path is forced.
+    is_pow2 = block_size > 0 and (block_size & (block_size - 1) == 0)
+    if not is_pow2:
+        use_custom = False
+
+    if use_custom:
+        _PARTITION_SIZE_ROCM = 256
+        max_num_partitions = (
+            max_seq_len + _PARTITION_SIZE_ROCM - 1
+        ) // _PARTITION_SIZE_ROCM
+        assert _PARTITION_SIZE_ROCM % block_size == 0
+        total_num_seq = block_table.shape[0]
+        tmp_output = torch.empty(
+            size=(total_num_seq, num_query_heads, max_num_partitions, head_size),
+            dtype=query.dtype,
+            device=output.device,
+        )
+        exp_sums = torch.empty(
+            size=(total_num_seq, num_query_heads, max_num_partitions),
+            dtype=torch.float32,
+            device=output.device,
+        )
+        max_logits = torch.empty_like(exp_sums)
+
+        ops.paged_attention_rocm(
+            output,
+            exp_sums,
+            max_logits,
+            tmp_output,
+            query,
+            key_cache,
+            value_cache,
+            num_kv_heads,
+            scale=sm_scale,
+            block_tables=block_table,
+            seq_lens=seq_lens,
+            query_start_loc=query_start_loc,
+            block_size=block_size,
+            max_seq_len=max_seq_len,
+            alibi_slopes=alibi_slopes,
+            kv_cache_dtype=kv_cache_dtype,
+            k_scale=k_scale,
+            v_scale=v_scale,
+            fp8_out_scale=output_scale,
+        )
+    else:
+        real_block_size = value_cache.shape[3]
+        # The standard model directly uses the original block_size.
+        # Non-standard 544 uses 32 to accommodate integer division logic.
+        TRITON_BLOCK_SIZE = block_size if is_pow2 else 32
+        if is_block_table_ptr:
+            # Using the physical base address of tensors
+            kv_element_size = key_cache.element_size()
+            block_byte_stride = key_cache.stride(0) * kv_element_size
+            # Get the starting physical address of the KV Cache
+            base_addr = key_cache.data_ptr()
+
+            # Normalization: Directly calculate the block offset
+            # of the pointer relative to the base address
+            processed_block_table = ((block_table - base_addr) // block_byte_stride).to(
+                torch.int32
+            )
+        else:
+            processed_block_table = block_table.to(torch.int32)
+
+        kernel_paged_attention_2d[
+            (
+                num_seqs,
+                num_kv_heads,
+            )
+        ](
+            output_ptr=output,
+            query_ptr=query,
+            key_cache_ptr=key_cache,
+            value_cache_ptr=value_cache,
+            sink_ptr=sinks,
+            block_tables_ptr=processed_block_table,
+            seq_lens_ptr=seq_lens,
+            alibi_slopes_ptr=alibi_slopes,
+            scale=sm_scale,
+            k_scale=k_scale,
+            v_scale=v_scale,
+            out_scale_inv=1.0 / output_scale if output_scale is not None else 1.0,
+            num_query_heads=num_query_heads,
+            num_queries_per_kv=num_queries_per_kv,
+            num_queries_per_kv_padded=num_queries_per_kv_padded,
+            block_table_stride=processed_block_table.stride(0),
+            query_stride_0=query.stride(0),
+            query_stride_1=query.stride(1),
+            output_stride_0=output.stride(0),
+            output_stride_1=output.stride(1),
+            BLOCK_SIZE=TRITON_BLOCK_SIZE,
+            PHYSICAL_BLOCK_SIZE=real_block_size,
+            HEAD_SIZE=head_size,
+            HEAD_SIZE_PADDED=triton.next_power_of_2(head_size),
+            USE_ALIBI_SLOPES=use_alibi_slopes,
+            SLIDING_WINDOW=sliding_window,
+            x=key_cache.shape[4],
+            stride_k_cache_0=key_cache.stride(0),
+            stride_k_cache_1=key_cache.stride(1),
+            stride_k_cache_2=key_cache.stride(2),
+            stride_k_cache_3=key_cache.stride(3),
+            stride_k_cache_4=key_cache.stride(4),
+            stride_v_cache_0=value_cache.stride(0),
+            stride_v_cache_1=value_cache.stride(1),
+            stride_v_cache_2=value_cache.stride(2),
+            stride_v_cache_3=value_cache.stride(3),
+            filter_by_query_len=True,
+            query_start_len_ptr=query_start_loc,
+            USE_SINKS=sinks is not None,
+            USE_FP8=output_scale is not None,
+        )

vllm/v1/attention/ops/common.py

@@ -0,0 +1,465 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import torch
+
+from vllm.distributed.parallel_state import GroupCoordinator
+from vllm.triton_utils import tl, triton
+
+
+@triton.jit
+def _correct_attn_cp_out_kernel(
+    outputs_ptr,
+    new_output_ptr,
+    lses_ptr,
+    vlse_ptr,
+    outputs_stride_B,
+    outputs_stride_H,
+    outputs_stride_D,
+    lses_stride_N,
+    lses_stride_B,
+    lses_stride_H,
+    lse_idx,
+    HEAD_DIM: tl.constexpr,
+    N_ROUNDED: tl.constexpr,
+    IS_BASE_E: tl.constexpr,
+):
+    """
+    Apply the all-gathered lses to correct each local rank's attention
+    output. we still need perform a cross-rank reduction to obtain the
+    final attention output.
+
+    Args:
+        outputs_ptr (triton.PointerType):
+            Pointer to input tensor of shape [ B, H, D ]
+        lses_ptr (triton.PointerType):
+            Pointer to input tensor of shape [ N, B, H ]
+        new_output_ptr (triton.PointerType):
+            Pointer to output tensor of shape [ B, H, D ]
+        vlse_ptr (triton.PointerType):
+            Pointer to output tensor of shape [ B, H ]
+    """
+    batch_idx = tl.program_id(axis=0).to(tl.int64)
+    head_idx = tl.program_id(axis=1).to(tl.int64)
+    d_offsets = tl.arange(0, HEAD_DIM)
+    num_n_offsets = tl.arange(0, N_ROUNDED)
+
+    # shape = [N]
+    lse_offsets = (
+        num_n_offsets * lses_stride_N
+        + batch_idx * lses_stride_B
+        + head_idx * lses_stride_H
+    )
+
+    # calc final lse
+    lse = tl.load(lses_ptr + lse_offsets)
+    lse = tl.where((lse != lse) | (lse == float("inf")), -float("inf"), lse)
+    lse_max = tl.max(lse, axis=0)
+    lse_max = tl.where(lse_max == -float("inf"), 0, lse_max)
+    lse -= lse_max
+    if IS_BASE_E:
+        lse_exp = tl.exp(lse)
+        lse_acc = tl.sum(lse_exp, axis=0)
+        lse = tl.log(lse_acc)
+    else:
+        lse_exp = tl.exp2(lse)
+        lse_acc = tl.sum(lse_exp, axis=0)
+        lse = tl.log2(lse_acc)
+    lse += lse_max
+
+    lse_offsets = batch_idx * lses_stride_B + head_idx * lses_stride_H
+    tl.store(vlse_ptr + lse_offsets, lse)
+
+    # shape = [D]
+    output_offsets = (
+        batch_idx * outputs_stride_B
+        + head_idx * outputs_stride_H
+        + d_offsets * outputs_stride_D
+    )
+
+    # correct output
+    lse_offset = (
+        lse_idx * lses_stride_N + batch_idx * lses_stride_B + head_idx * lses_stride_H
+    )
+    lse_tmp = tl.load(lses_ptr + lse_offset)
+    lse_finally = lse_tmp - lse
+    lse_finally = tl.where(
+        (lse_finally != lse_finally) | (lse_finally == float("inf")),
+        -float("inf"),
+        lse_finally,
+    )
+    factor = tl.exp(lse_finally) if IS_BASE_E else tl.exp2(lse_finally)
+    output = tl.load(outputs_ptr + output_offsets)
+    output = output * factor
+
+    tl.store(new_output_ptr + output_offsets, output)
+
+
+class CPTritonContext:
+    """The CPTritonContext is used to avoid recompilation of the Triton JIT."""
+
+    def __init__(self):
+        self.inner_kernel = None
+
+    def call_kernel(self, kernel, grid, *regular_args, **const_args):
+        if self.inner_kernel is None:
+            self.inner_kernel = kernel[grid](*regular_args, **const_args)
+        else:
+            self.inner_kernel[grid](*regular_args)
+
+
+def correct_attn_out(
+    out: torch.Tensor,
+    lses: torch.Tensor,
+    cp_rank: int,
+    ctx: CPTritonContext,
+    is_lse_base_on_e: bool = True,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Correct the attention output using the all-gathered lses.
+
+    Args:
+        out: Tensor of shape [ B, H, D ]
+        lses: Tensor of shape [ N, B, H ]
+        cp_rank: Current rank in the context-parallel group
+        ctx: Triton context to avoid recompilation
+
+    Returns:
+        Tuple of (out, lse) with corrected attention and final log-sum-exp.
+    """
+    if ctx is None:
+        ctx = CPTritonContext()
+
+    # --- Normalize to 3D views ---
+    if out.ndim == 4 and out.shape[1] == 1:
+        out = out.squeeze(1)
+    assert out.ndim == 3, f"expected out [B,H,D] or [B,1,H,D], got {tuple(out.shape)}"
+
+    if lses.ndim == 4 and lses.shape[-1] == 1:
+        lses = lses.squeeze(-1)
+    if lses.ndim == 4 and lses.shape[1] == 1:
+        lses = lses.squeeze(1)
+    assert lses.ndim == 3, (
+        f"expected lses [N,B,H] (optionally with a 1-sized extra dim), "
+        f"got {tuple(lses.shape)}"
+    )
+
+    B, H, D = out.shape
+    N = lses.shape[0]
+
+    # Strides after we normalized shapes to 3-D views.  The kernel computes
+    # offsets for `vlse_ptr` using lses_stride_B/H, so the output buffer must
+    # have the same B/H stride layout as a slice of `lses`.
+    o_sB, o_sH, o_sD = out.stride()
+    l_sN, l_sB, l_sH = lses.stride()
+
+    # Allocate LSE with the same B/H strides as `lses` so writes land correctly
+    # even when `lses` is a non-contiguous view (e.g., 4-D to 3-D squeeze).
+    lse = torch.empty_strided(
+        (B, H), (l_sB, l_sH), device=lses.device, dtype=lses.dtype
+    )
+
+    # Kernel launch config
+    grid = (B, H, 1)
+
+    regular_args = (
+        out,
+        out,
+        lses,
+        lse,
+        o_sB,
+        o_sH,
+        o_sD,
+        l_sN,
+        l_sB,
+        l_sH,
+        cp_rank,
+    )
+    const_args = {"HEAD_DIM": D, "N_ROUNDED": N, "IS_BASE_E": is_lse_base_on_e}
+    ctx.call_kernel(_correct_attn_cp_out_kernel, grid, *regular_args, **const_args)
+    return out, lse
+
+
+def _cp_lse_common(
+    cp_attn_out: torch.Tensor,
+    cp_attn_lse: torch.Tensor,
+    cp_group: GroupCoordinator,
+    ctx: CPTritonContext | None = None,
+    is_lse_base_on_e=True,
+):
+    """
+    cp_attn_out: [ B, H, D ]
+    cp_attn_lse: [ B, H ]
+    """
+    if cp_group.world_size == 1:
+        return cp_attn_out
+
+    if ctx is None:
+        ctx = CPTritonContext()
+
+    cp_attn_lse = cp_attn_lse.contiguous()
+    lses = cp_group.all_gather(cp_attn_lse, dim=0).reshape(
+        (cp_group.world_size,) + cp_attn_lse.shape
+    )
+    out, lse = correct_attn_out(
+        cp_attn_out,
+        lses,
+        cp_group.rank_in_group,
+        ctx,
+        is_lse_base_on_e=is_lse_base_on_e,
+    )
+    return out, lse
+
+
+def cp_lse_ag_out_rs(
+    cp_attn_out: torch.Tensor,
+    cp_attn_lse: torch.Tensor,
+    cp_group: GroupCoordinator,
+    ctx: CPTritonContext | None = None,
+    return_lse: bool = False,
+    is_lse_base_on_e=True,
+):
+    """
+    cp_attn_out: [ B, H, D ]
+    cp_attn_lse: [ B, H ]
+    """
+    out, lse = _cp_lse_common(
+        cp_attn_out, cp_attn_lse, cp_group, ctx=ctx, is_lse_base_on_e=is_lse_base_on_e
+    )
+    out = cp_group.reduce_scatter(out, dim=1)
+
+    if return_lse:
+        cp_num_heads = lse.shape[1] // cp_group.world_size
+        cp_rank = cp_group.rank_in_group
+        lse = lse[:, cp_num_heads * cp_rank : cp_num_heads * (cp_rank + 1)]
+        return out, lse
+    return out
+
+
+def cp_lse_ag_out_ar(
+    cp_attn_out: torch.Tensor,
+    cp_attn_lse: torch.Tensor,
+    cp_group: GroupCoordinator,
+    ctx: CPTritonContext | None = None,
+    return_lse: bool = False,
+    is_lse_base_on_e=True,
+):
+    """
+    cp_attn_out: [ B, H, D ]
+    cp_attn_lse: [ B, H ]
+    """
+    out, lse = _cp_lse_common(
+        cp_attn_out, cp_attn_lse, cp_group, ctx=ctx, is_lse_base_on_e=is_lse_base_on_e
+    )
+    out = cp_group.all_reduce(out)
+
+    if return_lse:
+        return out, lse
+    return out
+
+
+@triton.jit
+def _pack_seq_kernel(
+    x_ptr,  # [N, D]
+    out_ptr,  # [B, Lmax, D]
+    lengths_ptr,  # *i32, [B]
+    N: tl.constexpr,
+    D: tl.constexpr,
+    Lmax: tl.constexpr,
+    PAD_VALUE: tl.constexpr,
+    BLOCK_T: tl.constexpr,  # timesteps per program
+    BLOCK_D: tl.constexpr,  # features per program
+):
+    pid_b = tl.program_id(0)  # batch id
+    pid_t = tl.program_id(1)  # block over time dimension
+    pid_d = tl.program_id(2)  # block over feature dimension
+    off_t = pid_t * BLOCK_T + tl.arange(0, BLOCK_T)  # [BLOCK_T]
+    off_d = pid_d * BLOCK_D + tl.arange(0, BLOCK_D)  # [BLOCK_D]
+
+    # Compute start index and sequence length from cumulative lengths
+    in_start = 0
+    for i in range(pid_b):
+        in_start += tl.load(lengths_ptr + i)
+    seq_len = tl.load(lengths_ptr + pid_b)
+
+    # valid time positions for this block
+    t_mask = off_t < Lmax
+
+    # compute input row indices for valid (b, t)
+    in_row = in_start + off_t
+    valid_row = (off_t < seq_len) & t_mask
+
+    # Pointers
+    # x_ptr: row-major [N, D]
+    x_row_ptr = x_ptr + in_row[:, None] * D + off_d[None, :]
+
+    # out_ptr: row-major [B, Lmax, D]
+    out_row_ptr = out_ptr + (pid_b * Lmax + off_t)[:, None] * D + off_d[None, :]
+
+    # Initialize with PAD (cast will occur as needed based on out_ptr dtype)
+    d_mask = off_d[None, :] < D
+    pad_vals = tl.full([BLOCK_T, BLOCK_D], PAD_VALUE, tl.float32)
+    tl.store(out_row_ptr, pad_vals, mask=t_mask[:, None] & d_mask)
+
+    # Load & write only where within seq_len
+    x_vals = tl.load(x_row_ptr, mask=valid_row[:, None] & d_mask)
+    tl.store(out_row_ptr, x_vals, mask=valid_row[:, None] & d_mask)
+
+
+def pack_seq_triton(
+    x: torch.Tensor,
+    lengths: torch.Tensor,
+    pad_value: float = -float("inf"),
+    block_t: int = 64,
+    block_d: int = 64,
+) -> torch.Tensor:
+    """
+    Pack sequences of different lengths into a batched tensor.
+
+    Args:
+        x: [N, ...] - input tensor where N is total number of tokens
+        lengths: [B] - sequence lengths for each batch
+        pad_value: value to use for padding
+        block_t: block size for time dimension
+        block_d: block size for feature dimension
+
+    Returns:
+        packed: [B, Lmax, ...] - packed tensor
+    """
+
+    # Handle multi-dimensional input by reshaping to (N, -1)
+    original_shape = x.shape
+    if len(original_shape) > 2:
+        N = original_shape[0]
+        x_reshaped = x.reshape(N, -1)
+        D = x_reshaped.shape[1]
+    else:
+        N, D = x.shape
+        x_reshaped = x
+
+    B = lengths.numel()
+    Lmax = int(lengths.max().item())
+
+    # Starts are computed inside the kernel from lengths
+
+    out = torch.empty((B, Lmax, D), device=x.device, dtype=x.dtype)
+
+    grid = (B, triton.cdiv(Lmax, block_t), triton.cdiv(D, block_d))
+    _pack_seq_kernel[grid](
+        x_reshaped,
+        out,
+        lengths.int(),
+        N,
+        D,
+        Lmax,
+        PAD_VALUE=float(pad_value),
+        BLOCK_T=block_t,
+        BLOCK_D=block_d,
+        num_warps=4,
+        num_stages=2,
+    )
+
+    # Reshape output back to original dimensions (except first dimension)
+    if len(original_shape) > 2:
+        output_shape = (B, Lmax) + original_shape[1:]
+        out = out.reshape(output_shape)
+
+    return out
+
+
+@triton.jit
+def _unpack_seq_triton_kernel(
+    packed_ptr,  # [B, Lmax, D]
+    out_ptr,  # [N, D]
+    lengths_ptr,  # *i32, [B]
+    B: tl.constexpr,
+    Lmax: tl.constexpr,
+    D: tl.constexpr,
+    BLOCK_T: tl.constexpr,  # timesteps per program
+    BLOCK_D: tl.constexpr,  # features per program
+):
+    pid_b = tl.program_id(0)  # batch id
+    pid_t = tl.program_id(1)  # block over time dimension
+    pid_d = tl.program_id(2)  # block over feature dimension
+    off_t = pid_t * BLOCK_T + tl.arange(0, BLOCK_T)  # [BLOCK_T]
+    off_d = pid_d * BLOCK_D + tl.arange(0, BLOCK_D)  # [BLOCK_D]
+
+    # bounds: compute start from cumulative lengths
+    in_start = 0
+    for i in range(pid_b):
+        in_start += tl.load(lengths_ptr + i)
+    seq_len = tl.load(lengths_ptr + pid_b)
+
+    # valid time positions for this block
+    t_mask = off_t < Lmax
+    valid_row = (off_t < seq_len) & t_mask
+
+    # compute output row indices for valid (b, t)
+    out_row = in_start + off_t
+
+    # Pointers
+    # packed_ptr: row-major [B, Lmax, D]
+    packed_row_ptr = packed_ptr + (pid_b * Lmax + off_t)[:, None] * D + off_d[None, :]
+
+    # out_ptr: row-major [N, D]
+    out_row_ptr = out_ptr + out_row[:, None] * D + off_d[None, :]
+
+    # Load from packed tensor and store to output
+    d_mask = off_d[None, :] < D
+    packed_vals = tl.load(packed_row_ptr, mask=valid_row[:, None] & d_mask)
+    tl.store(out_row_ptr, packed_vals, mask=valid_row[:, None] & d_mask)
+
+
+def unpack_seq_triton(
+    packed_tensor: torch.Tensor,
+    lengths: torch.Tensor,
+    block_t: int = 64,
+    block_d: int = 64,
+) -> torch.Tensor:
+    """
+    Unpack a packed decode query tensor back to the original format.
+    Efficient Triton implementation.
+
+    Args:
+        packed_tensor: [B, Lmax, ...] - packed tensor from pack_seq_triton
+        lengths: [B] - sequence lengths for each batch
+        block_t: block size for time dimension
+        block_d: block size for feature dimension
+
+    Returns:
+        unpacked_tensor: [N, ...] where N = sum(lengths)
+    """
+
+    # Handle multi-dimensional input by reshaping to (B, Lmax, -1)
+    original_shape = packed_tensor.shape
+    if len(original_shape) > 3:
+        B, Lmax = original_shape[:2]
+        packed_reshaped = packed_tensor.reshape(B, Lmax, -1)
+        D = packed_reshaped.shape[2]
+    else:
+        B, Lmax, D = packed_tensor.shape
+        packed_reshaped = packed_tensor
+
+    # Calculate total number of elements
+    N = int(lengths.sum().item())
+
+    out = torch.empty((N, D), device=packed_tensor.device, dtype=packed_tensor.dtype)
+
+    grid = (B, triton.cdiv(Lmax, block_t), triton.cdiv(D, block_d))
+    _unpack_seq_triton_kernel[grid](
+        packed_reshaped,
+        out,
+        lengths.int(),
+        B,
+        Lmax,
+        D,
+        BLOCK_T=block_t,
+        BLOCK_D=block_d,
+        num_warps=4,
+        num_stages=2,
+    )
+
+    # Reshape output back to original dimensions (except first dimension)
+    if len(original_shape) > 3:
+        output_shape = (N,) + original_shape[2:]
+        out = out.reshape(output_shape)
+
+    return out

vllm/v1/attention/ops/flashmla.py

@@ -0,0 +1,166 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# adapted from: https://github.com/deepseek-ai/FlashMLA/blob/main/flash_mla/flash_mla_interface.py
+
+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
+
+if current_platform.is_cuda():
+    try:
+        import vllm._flashmla_extension_C  # noqa: F401
+
+        _flashmla_extension_C_AVAILABLE = True
+    except ImportError:
+        _flashmla_extension_C_AVAILABLE = False
+else:
+    _flashmla_extension_C_AVAILABLE = False
+
+
+def _is_flashmla_available() -> tuple[bool, str | None]:
+    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 "
+            "was not in the list of target arches to compile for.",
+        )
+    if not _flashmla_extension_C_AVAILABLE:
+        return (
+            False,
+            "vllm._flashmla_extension_C is not available, likely "
+            "was not compiled due to a build error.",
+        )
+
+    return True, None
+
+
+def is_flashmla_dense_supported() -> tuple[bool, str | None]:
+    """
+    Return: is_supported_flag, unsupported_reason (optional).
+    """
+    is_available, maybe_reason = _is_flashmla_available()
+    if not is_available:
+        return False, maybe_reason
+    if not current_platform.is_device_capability_family(90):
+        return False, "FlashMLA Dense is only supported on Hopper devices."
+    return True, None
+
+
+def is_flashmla_sparse_supported() -> tuple[bool, str | None]:
+    """
+    Return: is_supported_flag, unsupported_reason (optional).
+    """
+    is_available, maybe_reason = _is_flashmla_available()
+    if not is_available:
+        return False, maybe_reason
+    if not (
+        current_platform.is_device_capability_family(90)
+        or current_platform.is_device_capability_family(100)
+    ):
+        return (
+            False,
+            "FlashMLA Sparse is only supported on Hopper and Blackwell devices.",
+        )
+    return True, None
+
+
+def _raise_flashmla_unavailable(*_args, **_kwargs):
+    _, reason = _is_flashmla_available()
+    raise RuntimeError(reason or "FlashMLA is not available")
+
+
+if _is_flashmla_available()[0]:
+    from vllm.third_party.flashmla.flash_mla_interface import (  # noqa: F401
+        FlashMLASchedMeta,
+        flash_attn_varlen_func,
+        flash_attn_varlen_kvpacked_func,
+        flash_attn_varlen_qkvpacked_func,
+        flash_mla_sparse_fwd,
+        flash_mla_with_kvcache,
+        get_mla_metadata,
+    )
+else:
+
+    class FlashMLASchedMeta:  # type: ignore[no-redef]
+        pass
+
+    flash_attn_varlen_func = _raise_flashmla_unavailable  # type: ignore[assignment]
+    flash_attn_varlen_kvpacked_func = _raise_flashmla_unavailable  # type: ignore[assignment]
+    flash_attn_varlen_qkvpacked_func = _raise_flashmla_unavailable  # type: ignore[assignment]
+    flash_mla_sparse_fwd = _raise_flashmla_unavailable  # type: ignore[assignment]
+    flash_mla_with_kvcache = _raise_flashmla_unavailable  # type: ignore[assignment]
+    get_mla_metadata = _raise_flashmla_unavailable  # type: ignore[assignment]
+
+
+def get_mla_metadata_dense_fp8(
+    cache_seqlens: torch.Tensor,
+    num_q_tokens_per_head_k: int,
+    num_heads_k: int,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    if not _is_flashmla_available()[0]:
+        _raise_flashmla_unavailable()
+    return torch.ops._flashmla_extension_C.get_mla_decoding_metadata_dense_fp8(
+        cache_seqlens,
+        num_q_tokens_per_head_k,
+        num_heads_k,
+    )
+
+
+def flash_mla_with_kvcache_fp8(
+    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: float | None = None,
+    causal: bool = False,
+    descale_q: torch.Tensor | None = None,
+    descale_k: torch.Tensor | None = None,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    if not _is_flashmla_available()[0]:
+        _raise_flashmla_unavailable()
+    if softmax_scale is None:
+        softmax_scale = q.shape[-1] ** (-0.5)
+    out, softmax_lse = torch.ops._flashmla_extension_C.fwd_kvcache_mla_fp8(
+        q,
+        k_cache,
+        head_dim_v,
+        cache_seqlens,
+        block_table,
+        softmax_scale,
+        causal,
+        tile_scheduler_metadata,
+        num_splits,
+        descale_q,
+        descale_k,
+    )
+    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/v1/attention/ops/merge_attn_states.py

@@ -0,0 +1,47 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+
+from vllm.platforms import current_platform
+
+
+def merge_attn_states(
+    output: torch.Tensor,
+    prefix_output: torch.Tensor,
+    prefix_lse: torch.Tensor,
+    suffix_output: torch.Tensor,
+    suffix_lse: torch.Tensor,
+    output_lse: torch.Tensor | None = None,
+) -> None:
+    # NOTE(DefTruth): Currently, custom merge_attn_states CUDA kernel
+    # does not support FP8 dtype, fallback to use Triton kernel.
+    def supported_dtypes(o: torch.Tensor) -> bool:
+        return o.dtype in [torch.float32, torch.half, torch.bfloat16]
+
+    # NOTE(DefTruth): Currently, custom merge_attn_states CUDA
+    # kernel load/store 128b(16 bytes) per memory issue within
+    # thread. Namely, the headsize(headdim) must be multiple of
+    # pack_size (float32 -> 4, half/bfloat16 -> 8).
+    def supported_headdim(o: torch.Tensor) -> bool:
+        headdim = o.shape[2]  # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]
+        if o.dtype == torch.float32:
+            return headdim % 4 == 0
+        return headdim % 8 == 0
+
+    if (
+        current_platform.is_cuda()
+        and supported_dtypes(output)
+        and supported_headdim(output)
+    ):
+        from vllm._custom_ops import merge_attn_states
+
+        return merge_attn_states(
+            output, prefix_output, prefix_lse, suffix_output, suffix_lse, output_lse
+        )
+    else:
+        from vllm.v1.attention.ops.triton_merge_attn_states import merge_attn_states
+
+        return merge_attn_states(
+            output, prefix_output, prefix_lse, suffix_output, suffix_lse, output_lse
+        )

vllm/v1/attention/ops/paged_attn.py

@@ -0,0 +1,51 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+
+import torch
+
+from vllm.platforms import current_platform
+
+if current_platform.is_cuda_alike():
+    from vllm import _custom_ops as ops
+elif current_platform.is_xpu():
+    from vllm._xpu_ops import xpu_ops as ops  # type: ignore[no-redef]
+
+
+class PagedAttention:
+    @staticmethod
+    def split_kv_cache(
+        kv_cache: torch.Tensor,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        x = 16 // kv_cache.element_size()
+        num_blocks = kv_cache.shape[1]
+
+        key_cache = kv_cache[0]
+        key_cache = key_cache.view(num_blocks, num_kv_heads, head_size // x, -1, x)
+        value_cache = kv_cache[1]
+        value_cache = value_cache.view(num_blocks, num_kv_heads, head_size, -1)
+        return key_cache, value_cache
+
+    @staticmethod
+    def write_to_paged_cache(
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+        kv_cache_dtype: str,
+        k_scale: torch.Tensor,
+        v_scale: torch.Tensor,
+    ) -> None:
+        ops.reshape_and_cache(
+            key,
+            value,
+            key_cache,
+            value_cache,
+            slot_mapping.flatten(),
+            kv_cache_dtype,
+            k_scale,
+            v_scale,
+        )

vllm/v1/attention/ops/prefix_prefill.py

@@ -0,0 +1,862 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# The kernels in this file are adapted from LightLLM's context_attention_fwd:
+# https://github.com/ModelTC/lightllm/blob/main/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py
+
+import torch
+
+from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
+
+# Static kernels parameters
+BASE_BLOCK = 128 if current_platform.has_device_capability(80) else 64
+NUM_WARPS = 4 if current_platform.is_rocm() else 8
+
+# To check compatibility
+IS_TURING = current_platform.get_device_capability() == (7, 5)
+float8_info = torch.finfo(current_platform.fp8_dtype())
+
+
+# Here's an example autotuner config for this kernel. This config does provide
+# a performance improvement, but dramatically increases first call latency in
+# triton 3.2. Because of this tradeoff, it's currently commented out.
+# @triton.autotune(
+#     configs=[
+#         triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, \
+#                         "num_unroll_cache": 4, \
+#                         "num_unroll_request": 1 } | \
+#                         ({"kpack": 2, "waves_per_eu": 2} \
+#                             if current_platform.is_rocm() else {}), \
+#                         num_warps=4, \
+#                         num_stages=1)
+#     ],
+#     key=["BLOCK_SIZE", "MAX_Q_LEN", "MAX_CTX_LEN"]
+# )
+@triton.jit
+def _fwd_kernel(
+    Q,
+    K,
+    V,
+    K_cache,
+    V_cache,
+    sink_ptr,
+    B_Loc,
+    sm_scale,
+    k_scale,
+    v_scale,
+    out_scale_inv,
+    B_Start_Loc,
+    B_Seqlen,
+    x: tl.constexpr,
+    Out,
+    stride_b_loc_b,
+    stride_b_loc_s,
+    stride_qbs,
+    stride_qh,
+    stride_qd,
+    stride_kbs,
+    stride_kh,
+    stride_kd,
+    stride_vbs,
+    stride_vh,
+    stride_vd,
+    stride_obs,
+    stride_oh,
+    stride_od,
+    stride_k_cache_bs,
+    stride_k_cache_h,
+    stride_k_cache_d,
+    stride_k_cache_bl: tl.constexpr,
+    stride_k_cache_x,
+    stride_v_cache_bs,
+    stride_v_cache_h,
+    stride_v_cache_d,
+    stride_v_cache_bl,
+    num_queries_per_kv: tl.constexpr,
+    IN_PRECISION: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DMODEL_PADDED: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    PHYSICAL_BLOCK_SIZE: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    SLIDING_WINDOW: tl.constexpr,
+    num_unroll_cache: tl.constexpr,
+    num_unroll_request: tl.constexpr,
+    SKIP_DECODE: tl.constexpr,
+    USE_SINKS: tl.constexpr,
+    USE_FP8: tl.constexpr,
+    MAX_Q_LEN: tl.constexpr = 0,
+    MAX_CTX_LEN: tl.constexpr = 0,
+    FP8_MIN: tl.constexpr = float8_info.min,
+    FP8_MAX: tl.constexpr = float8_info.max,
+):
+    cur_batch = tl.program_id(0)
+    cur_head = tl.program_id(1)
+    start_m = tl.program_id(2)
+
+    cur_kv_head = cur_head // num_queries_per_kv
+
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+    cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
+    cur_batch_in_all_stop_index = tl.load(B_Start_Loc + cur_batch + 1)
+    cur_batch_query_len = cur_batch_in_all_stop_index - cur_batch_in_all_start_index
+    cur_batch_ctx_len = cur_batch_seq_len - cur_batch_query_len
+
+    if SKIP_DECODE and cur_batch_query_len == 1:
+        return
+
+    # start position inside of the query
+    # generally, N goes over kv, while M goes over query_len
+    block_start_loc = BLOCK_M * start_m
+
+    # initialize offsets
+    # [BLOCK_SIZE]; starts at 0
+    offs_bs_n = tl.arange(0, BLOCK_SIZE)
+    # [N]; starts at 0
+    offs_n = tl.arange(0, BLOCK_N)
+    # [D]; starts at 0
+    offs_d = tl.arange(0, BLOCK_DMODEL_PADDED)
+    # [M]; starts at current position in query
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    # [M,D]
+    off_q = (
+        (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs
+        + cur_head * stride_qh
+        + offs_d[None, :] * stride_qd
+    )
+
+    dim_mask = tl.where(tl.arange(0, BLOCK_DMODEL_PADDED) < BLOCK_DMODEL, 1, 0).to(
+        tl.int1
+    )  # [D]
+
+    q = tl.load(
+        Q + off_q,
+        mask=dim_mask[None, :] & (offs_m[:, None] < cur_batch_query_len),
+        other=0.0,
+    )  # [M,D]
+
+    # initialize pointer to m and l
+    if not USE_SINKS:
+        m_i = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+        l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
+    else:
+        m_i = tl.load(
+            sink_ptr + tl.full([BLOCK_M], cur_head, dtype=tl.int64),
+            mask=(offs_m < cur_batch_query_len),
+            other=float("-inf"),
+        ).to(dtype=tl.float32)
+        l_i = tl.where(m_i > float("-inf"), 1.0, 0.0)
+
+    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_PADDED], dtype=tl.float32)  # [M,D]
+
+    # compute query against context (no causal mask here)
+    for start_n in tl.range(
+        0, cur_batch_ctx_len, BLOCK_SIZE, loop_unroll_factor=num_unroll_cache
+    ):
+        # Under a block size of 544 (Qwen/Qwen3-Next-80B-A3B-Thinking),
+        # replace one physical block every 17 32-Tile blocks
+        # Calculate the logical block index of each of the 32 tokens
+        # in the current Tile (handling cross-block cases).
+        token_indices = start_n + offs_bs_n
+        bn_logical_indices = token_indices // PHYSICAL_BLOCK_SIZE
+
+        # 2. Vectorized loading of physical block IDs from B_Loc
+        bn = tl.load(
+            B_Loc + cur_batch * stride_b_loc_b + bn_logical_indices * stride_b_loc_s
+        ).to(tl.int64)
+
+        # 3. Calculate the exact offset of
+        # each token within its physical block.
+        internal_offsets = token_indices % PHYSICAL_BLOCK_SIZE
+
+        # Addressing of K (5D)
+        off_k = (
+            bn[None, :] * stride_k_cache_bs
+            + cur_kv_head * stride_k_cache_h
+            + (offs_d[:, None] // x) * stride_k_cache_d
+            + internal_offsets[None, :] * stride_k_cache_bl
+            + (offs_d[:, None] % x) * stride_k_cache_x
+        )
+
+        # Addressing of V (4D)
+        off_v = (
+            bn[:, None] * stride_v_cache_bs
+            + cur_kv_head * stride_v_cache_h
+            + offs_d[None, :] * stride_v_cache_d
+            + internal_offsets[:, None] * stride_v_cache_bl
+        )
+
+        if (
+            start_n + BLOCK_SIZE > cur_batch_ctx_len
+            or BLOCK_DMODEL != BLOCK_DMODEL_PADDED
+        ):
+            k_load = tl.load(
+                K_cache + off_k,
+                mask=dim_mask[:, None]
+                & ((start_n + offs_bs_n[None, :]) < cur_batch_ctx_len),
+                other=0.0,
+            )  # [D,N]
+        else:
+            k_load = tl.load(K_cache + off_k)
+
+        if k_load.dtype.is_fp8():
+            k = (k_load.to(tl.float32) * tl.load(k_scale)).to(q.dtype)
+        else:
+            k = k_load
+
+        # qk = tl.zeros([BLOCK_M, BLOCK_SIZE], dtype=tl.float32)  # [M,N]
+        qk = sm_scale * tl.dot(q, k, input_precision=IN_PRECISION)
+        qk = tl.where(
+            (start_n + offs_bs_n[None, :]) < cur_batch_ctx_len, qk, float("-inf")
+        )
+        # qk *= sm_scale
+        if SLIDING_WINDOW > 0:
+            # (cur_batch_ctx_len + offs_m[:, None]) are the positions of
+            # Q entries in sequence
+            # (start_n + offs_bs_n[None, :]) are the positions of
+            # KV entries in sequence
+            # So the condition makes sure each entry in Q only attends
+            # to KV entries not more than SLIDING_WINDOW away.
+            #
+            # We can't use -inf here, because the
+            # sliding window may lead to the entire row being masked.
+            # This then makes m_ij contain -inf, which causes NaNs in
+            # exp().
+            qk = tl.where(
+                (cur_batch_ctx_len + offs_m[:, None]) - (start_n + offs_bs_n[None, :])
+                < SLIDING_WINDOW,
+                qk,
+                float("-inf"),
+            )
+
+        # compute running maximum
+        m_ij = tl.maximum(m_i, tl.max(qk, axis=1))
+        p = tl.exp(qk - m_ij[:, None])
+        p = tl.where(m_ij[:, None] == float("-inf"), 0.0, p)
+        l_ij = tl.sum(p, axis=1)
+        alpha = tl.exp(m_i - m_ij)
+        alpha = tl.where(m_i == float("-inf"), 0.0, alpha)
+        acc = acc * alpha[:, None]
+
+        # update acc
+        if (
+            start_n + BLOCK_SIZE > cur_batch_ctx_len
+            or BLOCK_DMODEL != BLOCK_DMODEL_PADDED
+        ):
+            v_load = tl.load(
+                V_cache + off_v,
+                mask=dim_mask[None, :]
+                & ((start_n + offs_bs_n[:, None]) < cur_batch_ctx_len),
+                other=0.0,
+            )  # [N,D]
+        else:
+            v_load = tl.load(V_cache + off_v)
+
+        if v_load.dtype.is_fp8():
+            v = (v_load.to(tl.float32) * tl.load(v_scale)).to(q.dtype)
+        else:
+            v = v_load
+        p = p.to(v.dtype)
+
+        acc = tl.dot(p, v, acc=acc, input_precision=IN_PRECISION)
+        # # update m_i and l_i
+        l_i = l_i * alpha + l_ij
+        m_i = m_ij
+
+    off_k = (
+        offs_n[None, :] * stride_kbs
+        + cur_kv_head * stride_kh
+        + offs_d[:, None] * stride_kd
+    )
+    off_v = (
+        offs_n[:, None] * stride_vbs
+        + cur_kv_head * stride_vh
+        + offs_d[None, :] * stride_vd
+    )
+    k_ptrs = K + off_k
+    v_ptrs = V + off_v
+
+    # block_mask is 0 when we're already past the current query length
+    block_mask = tl.where(block_start_loc < cur_batch_query_len, 1, 0)
+
+    # compute query against itself (with causal mask)
+    for start_n in tl.range(
+        0,
+        block_mask * (start_m + 1) * BLOCK_M,
+        BLOCK_N,
+        loop_unroll_factor=num_unroll_request,
+    ):
+        start_n = tl.multiple_of(start_n, BLOCK_N)
+        # -- compute qk ----
+        k = tl.load(
+            k_ptrs + (cur_batch_in_all_start_index + start_n) * stride_kbs,
+            mask=dim_mask[:, None]
+            & ((start_n + offs_n[None, :]) < cur_batch_query_len),
+            other=0.0,
+        )
+
+        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+        qk = tl.dot(q, k, acc=qk, input_precision=IN_PRECISION)
+        qk *= sm_scale
+        # apply causal mask
+        qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk, float("-inf"))
+        if SLIDING_WINDOW > 0:
+            qk = tl.where(
+                offs_m[:, None] - (start_n + offs_n[None, :]) < SLIDING_WINDOW,
+                qk,
+                float("-inf"),
+            )
+
+        # compute running maximum
+        m_ij = tl.maximum(m_i, tl.max(qk, axis=1))
+        p = tl.exp(qk - m_ij[:, None])
+        p = tl.where(m_ij[:, None] == float("-inf"), 0.0, p)
+        l_ij = tl.sum(p, axis=1)
+        alpha = tl.exp(m_i - m_ij)
+        # To prevent NaN from appearing in the first round
+        alpha = tl.where(m_i == float("-inf"), 0.0, alpha)
+        acc = acc * alpha[:, None]
+
+        # update acc
+        v = tl.load(
+            v_ptrs + (cur_batch_in_all_start_index + start_n) * stride_vbs,
+            mask=dim_mask[None, :]
+            & ((start_n + offs_n[:, None]) < cur_batch_query_len),
+            other=0.0,
+        )
+        p = p.to(v.dtype)
+
+        acc = tl.dot(p, v, acc=acc, input_precision=IN_PRECISION)
+        # update m_i and l_i
+        l_i = l_i * alpha + l_ij
+        m_i = m_ij
+
+    acc = acc / (l_i[:, None] + 1e-10)
+
+    # initialize pointers to output
+    off_o = (
+        (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs
+        + cur_head * stride_oh
+        + offs_d[None, :] * stride_od
+    )
+    out_ptrs = Out + off_o
+    if USE_FP8:
+        acc = acc * tl.load(out_scale_inv)
+        acc = tl.clamp(acc, FP8_MIN, FP8_MAX)
+    tl.store(
+        out_ptrs, acc, mask=dim_mask[None, :] & (offs_m[:, None] < cur_batch_query_len)
+    )
+    return
+
+
+@triton.jit
+def _fwd_kernel_alibi(
+    Q,
+    K,
+    V,
+    K_cache,
+    V_cache,
+    B_Loc,
+    sm_scale,
+    k_scale,
+    v_scale,
+    B_Start_Loc,
+    B_Seqlen,
+    Alibi_slopes,
+    block_size,
+    x,
+    Out,
+    stride_b_loc_b,
+    stride_b_loc_s,
+    stride_qbs,
+    stride_qh,
+    stride_qd,
+    stride_kbs,
+    stride_kh,
+    stride_kd,
+    stride_vbs,
+    stride_vh,
+    stride_vd,
+    stride_obs,
+    stride_oh,
+    stride_od,
+    stride_k_cache_bs,
+    stride_k_cache_h,
+    stride_k_cache_d,
+    stride_k_cache_bl,
+    stride_k_cache_x,
+    stride_v_cache_bs,
+    stride_v_cache_h,
+    stride_v_cache_d,
+    stride_v_cache_bl,
+    num_queries_per_kv: int,
+    IN_PRECISION: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,  # head size
+    BLOCK_DMODEL_PADDED: tl.constexpr,  # head size padded to a power of 2
+    BLOCK_N: tl.constexpr,
+    SKIP_DECODE: tl.constexpr,
+):
+    # attn_bias[]
+    cur_batch = tl.program_id(0)
+    cur_head = tl.program_id(1)
+    start_m = tl.program_id(2)
+
+    cur_kv_head = cur_head // num_queries_per_kv
+
+    # cur_batch_seq_len: the length of prompts
+    # cur_batch_ctx_len: the length of prefix
+    # cur_batch_in_all_start_index: the start id of the dim=0
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+    cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
+    cur_batch_in_all_stop_index = tl.load(B_Start_Loc + cur_batch + 1)
+    cur_batch_query_len = cur_batch_in_all_stop_index - cur_batch_in_all_start_index
+    cur_batch_ctx_len = cur_batch_seq_len - cur_batch_query_len
+
+    if SKIP_DECODE and cur_batch_query_len == 1:
+        return
+
+    block_start_loc = BLOCK_M * start_m
+
+    # initialize offsets
+    offs_n = tl.arange(0, BLOCK_N)
+    offs_d = tl.arange(0, BLOCK_DMODEL_PADDED)
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    off_q = (
+        (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs
+        + cur_head * stride_qh
+        + offs_d[None, :] * stride_qd
+    )
+
+    dim_mask = tl.where(tl.arange(0, BLOCK_DMODEL_PADDED) < BLOCK_DMODEL, 1, 0).to(
+        tl.int1
+    )
+
+    q = tl.load(
+        Q + off_q,
+        mask=dim_mask[None, :]
+        & (offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len),
+        other=0.0,
+    )
+
+    # # initialize pointer to m and l
+    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
+    l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
+    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_PADDED], dtype=tl.float32)
+
+    alibi_slope = tl.load(Alibi_slopes + cur_head)
+    alibi_start_q = tl.arange(0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
+    alibi_start_k = 0
+    for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
+        start_n = tl.multiple_of(start_n, BLOCK_N)
+        # -- compute qk ----
+        bn = tl.load(
+            B_Loc
+            + cur_batch * stride_b_loc_b
+            + ((start_n + offs_n) // block_size) * stride_b_loc_s,
+            mask=(start_n + offs_n) < cur_batch_ctx_len,
+            other=0,
+        ).to(tl.int64)
+        off_k = (
+            bn[None, :] * stride_k_cache_bs
+            + cur_kv_head * stride_k_cache_h
+            + (offs_d[:, None] // x) * stride_k_cache_d
+            + ((start_n + offs_n[None, :]) % block_size) * stride_k_cache_bl
+            + (offs_d[:, None] % x) * stride_k_cache_x
+        )
+        off_v = (
+            bn[:, None] * stride_v_cache_bs
+            + cur_kv_head * stride_v_cache_h
+            + offs_d[None, :] * stride_v_cache_d
+            + (start_n + offs_n[:, None]) % block_size * stride_v_cache_bl
+        )
+        k_load = tl.load(
+            K_cache + off_k,
+            mask=dim_mask[:, None] & ((start_n + offs_n[None, :]) < cur_batch_ctx_len),
+            other=0.0,
+        )  # [D,N]
+
+        if k_load.dtype.is_fp8():
+            k = (k_load.to(tl.float32) * tl.load(k_scale)).to(q.dtype)
+        else:
+            k = k_load
+
+        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+        qk = tl.dot(q, k, acc=qk, input_precision=IN_PRECISION)
+        qk = tl.where(
+            (start_n + offs_n[None, :]) < cur_batch_ctx_len, qk, float("-inf")
+        )
+        qk *= sm_scale
+
+        # load alibi
+        alibi = (
+            tl.arange(0, BLOCK_N)[None, :] + alibi_start_k - alibi_start_q[:, None]
+        ) * alibi_slope
+        alibi = tl.where(
+            (alibi <= 0) & (alibi_start_q[:, None] < cur_batch_seq_len),
+            alibi,
+            float("-inf"),
+        )
+        qk += alibi
+        alibi_start_k += BLOCK_N
+
+        # -- compute m_ij, p, l_ij
+        m_ij = tl.max(qk, 1)
+        m_i_new = tl.maximum(m_i, m_ij)
+        p = tl.math.exp(qk - m_i_new[:, None])
+        l_ij = tl.sum(p, 1)
+        # -- update m_i and l_i
+
+        alpha = tl.math.exp(m_i - m_i_new)
+        l_i_new = alpha * l_i + l_ij
+        # -- update output accumulator --
+        # scale p
+        # scale acc
+        acc_scale = alpha
+        # acc_scale = l_i / l_i_new * alpha
+        acc = acc * acc_scale[:, None]
+        # update acc
+        v_load = tl.load(
+            V_cache + off_v,
+            mask=dim_mask[None, :] & ((start_n + offs_n[:, None]) < cur_batch_ctx_len),
+            other=0.0,
+        )
+        if v_load.dtype.is_fp8():
+            v = (v_load.to(tl.float32) * tl.load(v_scale)).to(q.dtype)
+        else:
+            v = v_load
+        p = p.to(v.dtype)
+
+        acc = tl.dot(p, v, acc=acc, input_precision="ieee")
+        # update m_i and l_i
+        l_i = l_i_new
+        m_i = m_i_new
+
+    off_k = (
+        offs_n[None, :] * stride_kbs
+        + cur_kv_head * stride_kh
+        + offs_d[:, None] * stride_kd
+    )
+    off_v = (
+        offs_n[:, None] * stride_vbs
+        + cur_kv_head * stride_vh
+        + offs_d[None, :] * stride_vd
+    )
+    k_ptrs = K + off_k
+    v_ptrs = V + off_v
+
+    block_mask = tl.where(block_start_loc < cur_batch_seq_len - cur_batch_ctx_len, 1, 0)
+
+    # init alibi
+    alibi_slope = tl.load(Alibi_slopes + cur_head)
+    alibi_start_q = tl.arange(0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
+    alibi_start_k = cur_batch_ctx_len
+    # # init debugger
+    # offset_db_q = tl.arange(0, BLOCK_M) + block_start_loc
+    # offset_db_k = tl.arange(0, BLOCK_N)
+    # calc q[BLOCK_M, BLOCK_MODEL] mul k[prefix_len: , BLOCK_DMODEL]
+    for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
+        start_n = tl.multiple_of(start_n, BLOCK_N)
+        # -- compute qk ----
+        k = tl.load(
+            k_ptrs + (cur_batch_in_all_start_index + start_n) * stride_kbs,
+            mask=dim_mask[:, None]
+            & ((start_n + offs_n[None, :]) < cur_batch_seq_len - cur_batch_ctx_len),
+            other=0.0,
+        )
+
+        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+        qk = tl.dot(q, k, acc=qk, input_precision="ieee")
+        qk *= sm_scale
+        qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk, float("-inf"))
+
+        # load alibi
+        alibi = (
+            tl.arange(0, BLOCK_N)[None, :] + alibi_start_k - alibi_start_q[:, None]
+        ) * alibi_slope
+        alibi = tl.where(
+            (alibi <= 0) & (alibi_start_q[:, None] < cur_batch_seq_len),
+            alibi,
+            float("-inf"),
+        )
+        qk += alibi
+        alibi_start_k += BLOCK_N
+
+        # -- compute m_ij, p, l_ij
+        m_ij = tl.max(qk, 1)
+        m_i_new = tl.maximum(m_i, m_ij)
+        p = tl.math.exp(qk - m_i_new[:, None])
+        l_ij = tl.sum(p, 1)
+        # -- update m_i and l_i
+
+        alpha = tl.math.exp(m_i - m_i_new)
+        l_i_new = alpha * l_i + l_ij
+        # -- update output accumulator --
+        # scale p
+        # scale acc
+        acc_scale = alpha
+        # acc_scale = l_i / l_i_new * alpha
+        acc = acc * acc_scale[:, None]
+        # update acc
+        v = tl.load(
+            v_ptrs + (cur_batch_in_all_start_index + start_n) * stride_vbs,
+            mask=dim_mask[None, :]
+            & ((start_n + offs_n[:, None]) < cur_batch_seq_len - cur_batch_ctx_len),
+            other=0.0,
+        )
+        p = p.to(v.dtype)
+
+        acc = tl.dot(p, v, acc=acc, input_precision="ieee")
+        # update m_i and l_i
+        l_i = l_i_new
+        m_i = m_i_new
+
+    acc = acc / l_i[:, None]
+
+    # initialize pointers to output
+    off_o = (
+        (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs
+        + cur_head * stride_oh
+        + offs_d[None, :] * stride_od
+    )
+    out_ptrs = Out + off_o
+    tl.store(
+        out_ptrs,
+        acc,
+        mask=dim_mask[None, :]
+        & (offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len),
+    )
+    return
+
+
+@torch.inference_mode()
+def context_attention_fwd(
+    q,
+    k,
+    v,
+    o,
+    kv_cache_dtype: str,
+    k_cache,
+    v_cache,
+    b_loc,
+    b_start_loc,
+    b_seq_len,
+    max_seq_len,
+    max_input_len,
+    k_scale: torch.Tensor,
+    v_scale: torch.Tensor,
+    alibi_slopes=None,
+    sliding_window=None,
+    sm_scale=None,
+    skip_decode=False,
+    fp8_out_scale=None,
+    sinks=None,
+    is_block_table_ptr: bool = False,
+):
+    q_dtype_is_f32 = q.dtype is torch.float32
+
+    # Turing does have tensor core for float32 multiplication
+    # use ieee as fallback for triton kernels work. There is also
+    # warning on vllm/config.py to inform users this fallback
+    # implementation
+    IN_PRECISION = "ieee" if IS_TURING and q_dtype_is_f32 else None
+
+    # Conversion of FP8 Tensor from uint8 storage to
+    # appropriate torch.dtype for interpretation by Triton
+    if "fp8" in kv_cache_dtype:
+        assert k_cache.dtype in [torch.uint8, current_platform.fp8_dtype()]
+        assert v_cache.dtype in [torch.uint8, current_platform.fp8_dtype()]
+
+        if kv_cache_dtype in ("fp8", "fp8_e4m3"):
+            target_dtype = current_platform.fp8_dtype()
+        elif kv_cache_dtype == "fp8_e5m2":
+            target_dtype = torch.float8_e5m2
+        else:
+            raise ValueError("Unsupported FP8 dtype:", kv_cache_dtype)
+
+        k_cache = k_cache.view(target_dtype)
+        v_cache = v_cache.view(target_dtype)
+
+    if (
+        k_cache.dtype == torch.uint8
+        or v_cache.dtype == torch.uint8
+        and kv_cache_dtype == "auto"
+    ):
+        raise ValueError(
+            "kv_cache_dtype='auto' unsupported for\
+            FP8 KV Cache prefill kernel"
+        )
+
+    # shape constraints
+    Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
+    assert Lq == Lk and Lk == Lv
+    # round up Lk to a power of 2 - this is required for Triton block size
+    Lk_padded = triton.next_power_of_2(Lk)
+
+    if sm_scale is None:
+        sm_scale = 1.0 / (Lq**0.5)
+    batch, head = b_seq_len.shape[0], q.shape[1]
+    num_queries_per_kv = q.shape[1] // k.shape[1]
+
+    assert batch + 1 == len(b_start_loc)
+
+    # 0 means "disable"
+    if sliding_window is None or sliding_window <= 0:
+        sliding_window = 0
+
+    if is_block_table_ptr:
+        kv_element_size = k_cache.element_size()
+        block_byte_stride = k_cache.stride(0) * kv_element_size
+        # The physical starting point of the obtained KV Cache Pool
+        base_addr = k_cache.data_ptr()
+
+        mask = b_loc > 0
+        processed_b_loc = torch.where(
+            mask, (b_loc - base_addr) // block_byte_stride, b_loc
+        ).to(torch.int32)
+    else:
+        processed_b_loc = b_loc.to(torch.int32)
+
+    if alibi_slopes is not None:
+        assert sinks is None, "Sinks arg is not supported with alibi"
+        assert fp8_out_scale is None, "FP8 output not supported with alibi"
+        # need to reduce num. blocks when using fp32
+        # due to increased use of GPU shared memory
+        # if q.dtype is torch.float32:
+        BLOCK = BASE_BLOCK // 2 if q_dtype_is_f32 else BASE_BLOCK
+        # batch, head,
+        grid = (batch, head, triton.cdiv(max_input_len, BLOCK))
+        _fwd_kernel_alibi[grid](
+            q,
+            k,
+            v,
+            k_cache,
+            v_cache,
+            b_loc,
+            sm_scale,
+            k_scale,
+            v_scale,
+            b_start_loc,
+            b_seq_len,
+            alibi_slopes,
+            v_cache.shape[3],
+            k_cache.shape[4],
+            o,
+            b_loc.stride(0),
+            b_loc.stride(1),
+            q.stride(0),
+            q.stride(1),
+            q.stride(2),
+            k.stride(0),
+            k.stride(1),
+            k.stride(2),
+            v.stride(0),
+            v.stride(1),
+            v.stride(2),
+            o.stride(0),
+            o.stride(1),
+            o.stride(2),
+            k_cache.stride(0),
+            k_cache.stride(1),
+            k_cache.stride(2),
+            k_cache.stride(3),
+            k_cache.stride(4),  # [num_blocks, num_kv_heads, head_size/x, block_size, x]
+            v_cache.stride(0),
+            v_cache.stride(1),
+            v_cache.stride(2),
+            v_cache.stride(3),  # [num_blocks, num_kv_heads, head_size, block_size]
+            num_queries_per_kv=num_queries_per_kv,
+            IN_PRECISION=IN_PRECISION,
+            BLOCK_M=BLOCK,
+            BLOCK_DMODEL=Lk,
+            BLOCK_DMODEL_PADDED=Lk_padded,
+            BLOCK_N=BLOCK,
+            SKIP_DECODE=skip_decode,
+            num_warps=NUM_WARPS,
+            num_stages=1,
+        )
+        return
+
+    max_seq_len = 0 if max_seq_len is None else max_seq_len
+    extra_kargs = {}
+    if current_platform.is_rocm():
+        extra_kargs = {}
+
+    real_block_size = v_cache.shape[3]
+    is_pow2 = real_block_size > 0 and (real_block_size & (real_block_size - 1) == 0)
+    # For standard models involving powers of 2,
+    # follow the original logic (Llama 128/64)
+    # For non-standard models (Qwen3-next block_size 544), set to 32.
+    if is_pow2:
+        BLOCK_M = 128
+        BLOCK_N = 64
+    else:
+        BLOCK_M = 32
+        BLOCK_N = 32
+
+    # TRITON_BLOCK_SIZE is kept at 32 to ensure
+    # correct alignment logic when the kernel handles
+    # non-standard sizes (such as 544).
+    TRITON_BLOCK_SIZE = 32
+
+    grid_fn = lambda META: (batch, head, triton.cdiv(max_input_len, META["BLOCK_M"]))
+    _fwd_kernel[grid_fn](
+        q,
+        k,
+        v,
+        k_cache,
+        v_cache,
+        sinks,
+        processed_b_loc,
+        sm_scale,
+        k_scale,
+        v_scale,
+        1.0 / fp8_out_scale if fp8_out_scale is not None else 1.0,
+        b_start_loc,
+        b_seq_len,
+        k_cache.shape[4],
+        o,
+        processed_b_loc.stride(0),
+        processed_b_loc.stride(1),
+        q.stride(0),
+        q.stride(1),
+        q.stride(2),
+        k.stride(0),
+        k.stride(1),
+        k.stride(2),
+        v.stride(0),
+        v.stride(1),
+        v.stride(2),
+        o.stride(0),
+        o.stride(1),
+        o.stride(2),
+        stride_k_cache_bs=k_cache.stride(0),
+        stride_k_cache_h=k_cache.stride(1),
+        stride_k_cache_d=k_cache.stride(2),
+        stride_k_cache_bl=k_cache.stride(3),
+        stride_k_cache_x=k_cache.stride(4),
+        stride_v_cache_bs=v_cache.stride(0),
+        stride_v_cache_h=v_cache.stride(1),
+        stride_v_cache_d=v_cache.stride(2),
+        stride_v_cache_bl=v_cache.stride(3),
+        BLOCK_SIZE=TRITON_BLOCK_SIZE,
+        PHYSICAL_BLOCK_SIZE=real_block_size,
+        num_queries_per_kv=num_queries_per_kv,
+        IN_PRECISION=IN_PRECISION,
+        BLOCK_DMODEL=Lk,
+        BLOCK_DMODEL_PADDED=Lk_padded,
+        SLIDING_WINDOW=sliding_window,
+        SKIP_DECODE=skip_decode,
+        USE_FP8=fp8_out_scale is not None,
+        BLOCK_M=BLOCK_M,
+        BLOCK_N=BLOCK_N,
+        num_unroll_cache=4,
+        num_unroll_request=1,
+        num_warps=4,
+        num_stages=1,
+        USE_SINKS=sinks is not None,
+        **extra_kargs,
+    )
+    return

vllm/v1/attention/ops/rocm_aiter_mla_sparse.py

@@ -0,0 +1,648 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import functools
+import importlib
+
+import torch
+
+from vllm.forward_context import get_forward_context
+from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
+from vllm.v1.attention.backends.mla.indexer import DeepseekV32IndexerMetadata
+from vllm.v1.attention.ops.common import pack_seq_triton, unpack_seq_triton
+
+if current_platform.is_cuda_alike():
+    from vllm import _custom_ops as ops
+
+
+@triton.jit
+def _indexer_k_quant_and_cache_kernel(
+    k_ptr,  # [num_tokens, head_dim]
+    kv_cache_ptr,  # [n_blks, blk_size//tile_block, head_dim // 16B, tile_block, 16B]
+    # [n_blocks, blk_size, head_dim]
+    kv_cache_scale_ptr,  # [n_blks, blk_size]
+    slot_mapping_ptr,  # [num_tokens]
+    kv_cache_scale_stride,
+    kv_cache_value_stride,
+    block_size,
+    num_tokens,
+    head_dim: tl.constexpr,
+    LAYOUT: tl.constexpr,
+    BLOCK_TILE_SIZE: tl.constexpr,
+    HEAD_TILE_SIZE: tl.constexpr,
+    IS_FNUZ: tl.constexpr,
+    USE_UE8M0: tl.constexpr,
+):
+    tid = tl.program_id(0)
+    offset = tl.arange(0, head_dim)
+    if LAYOUT == "SHUFFLE":
+        tile_offset = (
+            offset // HEAD_TILE_SIZE * BLOCK_TILE_SIZE * HEAD_TILE_SIZE
+            + offset % HEAD_TILE_SIZE
+        )
+    else:
+        tile_offset = offset
+    tile_store_offset = tile_offset
+    # for idx in tl.range(tid, num_tokens, n_program):
+    src_ptr = k_ptr + tid * head_dim
+    slot_id = tl.load(slot_mapping_ptr + tid)
+    if slot_id < 0:
+        return
+    block_id = slot_id // block_size
+    block_offset = slot_id % block_size
+    tile_block_id = block_offset // BLOCK_TILE_SIZE
+    tile_block_offset = block_offset % BLOCK_TILE_SIZE
+    val = tl.load(src_ptr + offset)
+    amax = tl.max(val.abs(), axis=-1).to(tl.float32)
+    if IS_FNUZ:
+        scale = tl.maximum(1e-4, amax) / 224.0
+    else:
+        scale = tl.maximum(1e-4, amax) / 448.0
+
+    if USE_UE8M0:
+        scale = tl.exp2(tl.ceil(tl.log2(scale)))
+
+    fp8_val = (val.to(tl.float32) / scale).to(kv_cache_ptr.type.element_ty)
+    if LAYOUT == "SHUFFLE":
+        dst_ptr = (
+            kv_cache_ptr
+            + block_id * kv_cache_value_stride
+            + tile_block_id * BLOCK_TILE_SIZE * head_dim
+            + tile_block_offset * HEAD_TILE_SIZE
+        )
+    else:
+        dst_ptr = (
+            kv_cache_ptr + block_id * kv_cache_value_stride + block_offset * head_dim
+        )
+    tl.store(dst_ptr + tile_store_offset, fp8_val)
+    dst_scale_ptr = kv_cache_scale_ptr + block_id * kv_cache_scale_stride + block_offset
+    tl.store(dst_scale_ptr, scale)
+
+
+def indexer_k_quant_and_cache_triton(
+    k: torch.Tensor,
+    kv_cache: torch.Tensor,  # [num_blocks, block_size, head_dim + 4]
+    slot_mapping: torch.Tensor,
+    quant_block_size,
+    scale_fmt,
+    block_tile_size=16,
+    head_tile_size=16,
+):
+    num_blocks = kv_cache.shape[0]
+    head_dim = k.shape[-1]
+    num_tokens = slot_mapping.shape[0]
+    block_size = kv_cache.shape[1]
+    # In real layout, we store the first portion as kv cache value
+    # and second portion as kv cache scale
+    kv_cache = kv_cache.view(num_blocks, -1)
+    kv_cache_value = kv_cache[:, : block_size * head_dim]
+    kv_cache_scale = kv_cache[:, block_size * head_dim :].view(torch.float32)
+    head_tile_size = head_tile_size // kv_cache.element_size()
+    grid = (num_tokens,)
+    _indexer_k_quant_and_cache_kernel[grid](
+        k,
+        kv_cache_value,
+        kv_cache_scale,
+        slot_mapping,
+        kv_cache_scale.stride(0),
+        kv_cache_value.stride(0),
+        block_size,
+        num_tokens,
+        head_dim,
+        "NHD",
+        block_tile_size,
+        head_tile_size,
+        IS_FNUZ=current_platform.fp8_dtype() == torch.float8_e4m3fnuz,
+        USE_UE8M0=scale_fmt == "ue8m0",
+    )
+
+
+@triton.jit
+def _cp_gather_indexer_quant_cache_kernel(
+    kv_cache_ptr,  # [n_blks,blk_size//tile_blk,head_dim//16B,tile_blk,16B]
+    # [n_blks, blk_size, head_dim]
+    kv_cache_scale_ptr,  # [n_blks, blk_size]
+    k_fp8_ptr,  # [num_tokens, head_dim]
+    k_scale_ptr,  # [num_tokens]
+    block_table_ptr,  # [batch_size, block_table_stride]
+    cu_seqlen_ptr,  # [batch_size + 1]
+    token_to_seq_ptr,  # [num_tokens]
+    block_size,
+    block_table_stride,
+    kv_cache_stride,
+    kv_cache_scale_stride,
+    LAYOUT: tl.constexpr,
+    HEAD_DIM: tl.constexpr,
+    BLOCK_TILE_SIZE: tl.constexpr,
+    HEAD_TILE_SIZE: tl.constexpr,
+):
+    tid = tl.program_id(0)
+    offset = tl.arange(0, HEAD_DIM)
+    batch_id = tl.load(token_to_seq_ptr + tid)
+    batch_start = tl.load(cu_seqlen_ptr + batch_id)
+    batch_end = tl.load(cu_seqlen_ptr + batch_id + 1)
+    batch_offset = tid - batch_start
+    if tid >= batch_end:
+        return
+    block_table_id = batch_offset // block_size
+    block_offset = batch_offset % block_size
+    block_table_offset = batch_id * block_table_stride + block_table_id
+    block_id = tl.load(block_table_ptr + block_table_offset)
+    tiled_block_id = block_offset // BLOCK_TILE_SIZE
+    tiled_block_offset = block_offset % BLOCK_TILE_SIZE
+    if LAYOUT == "SHUFFLE":
+        src_cache_offset = (
+            block_id * kv_cache_stride
+            + tiled_block_id * HEAD_DIM * BLOCK_TILE_SIZE
+            + tiled_block_offset * HEAD_TILE_SIZE
+        )
+    else:
+        src_cache_offset = block_id * kv_cache_stride + block_offset * HEAD_DIM
+    src_scale_offset = block_id * kv_cache_scale_stride + block_offset
+    dst_offset = tid * HEAD_DIM
+    src_scale_ptr = kv_cache_scale_ptr + src_scale_offset
+    src_cache_ptr = kv_cache_ptr + src_cache_offset
+    dst_k_ptr = k_fp8_ptr + dst_offset
+    scale_val = tl.load(src_scale_ptr)
+    tl.store(k_scale_ptr + tid, scale_val)
+    if LAYOUT == "SHUFFLE":
+        tiled_src_offset = (
+            offset // HEAD_TILE_SIZE * HEAD_TILE_SIZE * BLOCK_TILE_SIZE
+            + offset % HEAD_TILE_SIZE
+        )
+    else:
+        tiled_src_offset = offset
+    val = tl.load(src_cache_ptr + tiled_src_offset)
+    tl.store(dst_k_ptr + offset, val)
+
+
+def cp_gather_indexer_k_quant_cache_triton(
+    k_cache: torch.Tensor,  # [num_blocks, block_size, head_dim + 4]
+    k_fp8: torch.Tensor,
+    k_fp8_scale: torch.Tensor,
+    block_table: torch.Tensor,
+    cu_seqlen: torch.Tensor,
+    token_to_seq: torch.Tensor,
+    block_tile_size: int = 16,
+    head_tile_size: int = 16,
+):
+    num_tokens = k_fp8.size(0)
+    block_size = k_cache.size(1)
+    block_table_stride = block_table.stride(0)
+    head_dim = k_fp8.shape[-1]
+    num_blocks = k_cache.shape[0]
+    # we assume the kv cache already been split to 2 portion
+    k_cache = k_cache.view(num_blocks, -1)
+    fp8_dtype = current_platform.fp8_dtype()
+    k_cache_value = k_cache[:, : block_size * head_dim].view(fp8_dtype)
+    k_cache_scale = k_cache[:, block_size * head_dim :].view(torch.float32)
+    grid = (num_tokens,)
+    k_fp8_scale = k_fp8_scale.view(torch.float32)
+    _cp_gather_indexer_quant_cache_kernel[grid](
+        k_cache_value,
+        k_cache_scale,
+        k_fp8,
+        k_fp8_scale,
+        block_table,
+        cu_seqlen,
+        token_to_seq,
+        block_size,
+        block_table_stride,
+        k_cache_value.stride(0),
+        k_cache_scale.stride(0),
+        "NHD",
+        head_dim,
+        block_tile_size,
+        head_tile_size,
+    )
+
+
+# Taken from https://github.com/deepseek-ai/DeepGEMM/blob/main/tests/test_attention.py#L156
+def fp8_paged_mqa_logits_torch(
+    q: torch.Tensor,
+    kv_cache: torch.Tensor,
+    weights: torch.Tensor,
+    context_lens: torch.Tensor,
+    block_tables: torch.Tensor,
+    max_model_len: int,
+):
+    from vllm.utils.math_utils import cdiv
+
+    fp8_dtype = current_platform.fp8_dtype()
+    batch_size, next_n, _, dim = q.size()
+    kv_cache, scale = kv_cache[..., :dim], kv_cache[..., dim:]
+    scale = scale.contiguous().view(torch.float)
+    q = q.float()
+    kv_cache = kv_cache.view(fp8_dtype).float() * scale
+    num_block, block_size, _, dim = kv_cache.size()
+    logits = torch.full(
+        [batch_size * next_n, max_model_len],
+        float("-inf"),
+        device=q.device,
+        dtype=torch.float32,
+    )
+    context_lens = context_lens.tolist()
+    for i in range(batch_size):
+        context_len = context_lens[i]
+        q_offsets = torch.arange(context_len - next_n, context_len, device="cuda")
+        weight_slice = (
+            weights[i * next_n : (i + 1) * next_n, :].transpose(0, 1).contiguous()
+        )
+        for block_rk in range(cdiv(context_len, block_size)):
+            block_idx = block_tables[i][block_rk]
+            qx, kx = q[i], kv_cache[block_idx]
+            k_offsets = torch.arange(
+                block_rk * block_size, (block_rk + 1) * block_size, device="cuda"
+            )
+            mask = (k_offsets[None, :] < context_len) & (
+                k_offsets[None, :] <= q_offsets[:, None]
+            )
+            s = torch.where(
+                mask[None, :, :],
+                (qx.transpose(0, 1) @ kx.transpose(0, 1).transpose(1, 2)).to(
+                    logits.dtype
+                ),
+                float("-inf"),
+            )
+            s = torch.relu(s) * weight_slice[..., None]
+            s = s.sum(dim=0)
+            logits[
+                i * next_n : (i + 1) * next_n,
+                block_rk * block_size : (block_rk + 1) * block_size,
+            ] = torch.where(k_offsets[None, :] <= q_offsets[:, None], s, float("-inf"))
+    return logits
+
+
+def rocm_fp8_paged_mqa_logits(
+    q_fp8: torch.Tensor,
+    kv_cache_fp8: torch.Tensor,
+    weights: torch.Tensor,
+    context_lens: torch.Tensor,
+    block_tables: torch.Tensor,
+    schedule_metadata: torch.Tensor,
+    max_model_len: int,
+) -> torch.Tensor:
+    """Compute FP8 MQA logits using paged KV-cache.
+
+    Args:
+        q_fp8: Query tensor of shape [B, next_n, H, D]. Casted to
+            `torch.float8_e4m3fn` by caller.
+        kv_cache_fp8: Paged KV-cache in packed FP8+scale layout with shape
+            [num_blocks, block_size, 1, D+4], dtype `torch.uint8`. The last
+            4 bytes per (block,pos) store the `float` dequant scale.
+        weights: Tensor of shape [B * next_n, H], dtype `torch.float32`.
+        context_lens: Tensor of shape [B], dtype int32; effective context length
+            for each batch element.
+        block_tables: Tensor of shape [B, max_blocks], dtype int32; maps logical
+            block indices to physical blocks in the paged cache.
+        schedule_metadata: Returned by `get_paged_mqa_logits_metadata`;
+            used to distribute work across SMs.
+        max_model_len: Maximum sequence length used to size the logits output.
+
+    Returns:
+        Logits tensor of shape [B * next_n, max_model_len], dtype
+        `torch.float32`.
+    """
+    from vllm._aiter_ops import rocm_aiter_ops
+
+    @functools.lru_cache
+    def paged_mqa_logits_module():
+        paged_mqa_logits_module_path = None
+        if importlib.util.find_spec("aiter.ops.triton.pa_mqa_logits") is not None:
+            paged_mqa_logits_module_path = "aiter.ops.triton.pa_mqa_logits"
+        elif (
+            importlib.util.find_spec("aiter.ops.triton.attention.pa_mqa_logits")
+            is not None
+        ):
+            paged_mqa_logits_module_path = "aiter.ops.triton.attention.pa_mqa_logits"
+
+        if paged_mqa_logits_module_path is not None:
+            try:
+                module = importlib.import_module(paged_mqa_logits_module_path)
+                return module
+            except ImportError:
+                return None
+        return None
+
+    aiter_paged_mqa_logits_module = None
+    if rocm_aiter_ops.is_enabled():
+        aiter_paged_mqa_logits_module = paged_mqa_logits_module()
+    # FIXME(ganyi): Temporarily disable the aiter path until nightly docker
+    # update aiter to the fix PR.
+    aiter_paged_mqa_logits_module = None
+
+    if aiter_paged_mqa_logits_module is not None:
+        deepgemm_fp8_paged_mqa_logits_stage1 = (
+            aiter_paged_mqa_logits_module.deepgemm_fp8_paged_mqa_logits_stage1
+        )
+        batch_size, next_n, heads, _ = q_fp8.shape
+        out_qk = torch.full(
+            (heads, batch_size * next_n, max_model_len),
+            float("-inf"),
+            device="cuda",
+            dtype=torch.float32,
+        )
+        deepgemm_fp8_paged_mqa_logits_stage1(
+            q_fp8,
+            kv_cache_fp8,
+            weights,
+            out_qk,
+            context_lens,
+            block_tables,
+            max_model_len,
+        )
+        return out_qk.sum(dim=0)
+    else:
+        return fp8_paged_mqa_logits_torch(
+            q_fp8, kv_cache_fp8, weights, context_lens, block_tables, max_model_len
+        )
+
+
+# Take from https://github.com/deepseek-ai/DeepGEMM/blob/main/tests/test_attention.py#L84
+def fp8_mqa_logits_torch(
+    q: torch.Tensor,
+    kv: tuple[torch.Tensor, torch.Tensor],
+    weights: torch.Tensor,
+    cu_seqlen_ks: torch.Tensor,
+    cu_seqlen_ke: torch.Tensor,
+) -> torch.Tensor:
+    """Compute FP8 MQA logits for a single sequence without KV paging.
+
+    Args:
+        q: Query tensor of shape [M, H, D]. Casted to
+            `torch.float8_e4m3fn` by caller.
+        kv: Tuple `(k_fp8, k_scales)` where `k_fp8` has shape [N, D] with
+            dtype `torch.float8_e4m3fn` and `k_scales` has shape [N] (or
+            [N, 1]) with dtype `torch.float32`.
+        weights: weights of shape [M, H], dtype `torch.float32`.
+        cu_seqlen_ks: Start indices (inclusive) for valid K per query position,
+            shape [M], dtype int32.
+        cu_seqlen_ke: End indices (exclusive) for valid K per query position,
+            shape [M], dtype int32.
+
+    Returns:
+        Logits tensor of shape [M, N], dtype `torch.float32`.
+    """
+    kv, scale = kv
+    seq_len_kv = kv.shape[0]
+    k = kv.to(torch.bfloat16)
+    q = q.to(torch.bfloat16)
+
+    mask_lo = (
+        torch.arange(0, seq_len_kv, device="cuda")[None, :] >= cu_seqlen_ks[:, None]
+    )
+    mask_hi = (
+        torch.arange(0, seq_len_kv, device="cuda")[None, :] < cu_seqlen_ke[:, None]
+    )
+    mask = mask_lo & mask_hi
+
+    score = torch.einsum("mhd,nd->hmn", q, k).float() * scale
+    logits = (score.relu() * weights.unsqueeze(-1).transpose(0, 1)).sum(dim=0)
+    logits = logits.masked_fill(~mask, float("-inf"))
+
+    return logits
+
+
+def rocm_fp8_mqa_logits(
+    q: torch.Tensor,
+    kv: tuple[torch.Tensor, torch.Tensor],
+    weights: torch.Tensor,
+    cu_seqlen_ks: torch.Tensor,
+    cu_seqlen_ke: torch.Tensor,
+) -> torch.Tensor:
+    """Compute FP8 MQA logits for a single sequence without KV paging.
+
+    Args:
+        q: Query tensor of shape [M, H, D]. Casted to
+            `torch.float8_e4m3fn` by caller.
+        kv: Tuple `(k_fp8, k_scales)` where `k_fp8` has shape [N, D] with
+            dtype `torch.float8_e4m3fn` and `k_scales` has shape [N] (or
+            [N, 1]) with dtype `torch.float32`.
+        weights: weights of shape [M, H], dtype `torch.float32`.
+        cu_seqlen_ks: Start indices (inclusive) for valid K per query position,
+            shape [M], dtype int32.
+        cu_seqlen_ke: End indices (exclusive) for valid K per query position,
+            shape [M], dtype int32.
+
+    Returns:
+        Logits tensor of shape [M, N], dtype `torch.float32`.
+    """
+
+    # TODO(ganyi): Temporarily workaround, will remove the module check and reference
+    # path after aiter merge this kernel into main
+    from vllm._aiter_ops import rocm_aiter_ops
+
+    @functools.lru_cache
+    def mqa_logits_module():
+        mqa_logits_module_path = None
+        if importlib.util.find_spec("aiter.ops.triton.fp8_mqa_logits") is not None:
+            mqa_logits_module_path = "aiter.ops.triton.fp8_mqa_logits"
+        elif (
+            importlib.util.find_spec("aiter.ops.triton.attention.fp8_mqa_logits")
+            is not None
+        ):
+            mqa_logits_module_path = "aiter.ops.triton.attention.fp8_mqa_logits"
+
+        if mqa_logits_module_path is not None:
+            try:
+                module = importlib.import_module(mqa_logits_module_path)
+                return module
+            except ImportError:
+                return None
+        return None
+
+    aiter_mqa_logits_module = None
+    if rocm_aiter_ops.is_enabled():
+        aiter_mqa_logits_module = mqa_logits_module()
+
+    if aiter_mqa_logits_module is not None:
+        fp8_mqa_logits = aiter_mqa_logits_module.fp8_mqa_logits
+        kv, scale = kv
+        return fp8_mqa_logits(q, kv, scale, weights, cu_seqlen_ks, cu_seqlen_ke)
+    else:
+        return fp8_mqa_logits_torch(q, kv, weights, cu_seqlen_ks, cu_seqlen_ke)
+
+
+def rocm_aiter_sparse_attn_indexer_fake(
+    hidden_states: torch.Tensor,
+    k_cache_prefix: str,
+    kv_cache: torch.Tensor,
+    q_fp8: torch.Tensor,
+    k: torch.Tensor,
+    weights: torch.Tensor,
+    quant_block_size: int,
+    scale_fmt: str | None,
+    topk_tokens: int,
+    head_dim: int,
+    max_model_len: int,
+    total_seq_lens: int,
+    topk_indices_buffer: torch.Tensor | None,
+) -> torch.Tensor:
+    # profile run
+    # NOTE(Chen): create the max possible flattened_kv. So that
+    # profile_run can get correct memory usage.
+    _flattened_kv = torch.empty(
+        [total_seq_lens, head_dim + 4], device=k.device, dtype=torch.uint8
+    )
+    fp8_dtype = current_platform.fp8_dtype()
+    _k_fp8 = _flattened_kv[..., :head_dim].view(fp8_dtype).contiguous()
+    _k_scale = _flattened_kv[..., head_dim:].view(torch.float32).contiguous()
+    return topk_indices_buffer
+
+
+def rocm_aiter_sparse_attn_indexer(
+    hidden_states: torch.Tensor,
+    k_cache_prefix: str,
+    kv_cache: torch.Tensor,
+    q_fp8: torch.Tensor,
+    k: torch.Tensor,
+    weights: torch.Tensor,
+    quant_block_size: int,
+    scale_fmt: str | None,
+    topk_tokens: int,
+    head_dim: int,
+    max_model_len: int,
+    total_seq_lens: int,
+    topk_indices_buffer: torch.Tensor | None,
+) -> torch.Tensor:
+    # careful! this will be None in dummy run
+    attn_metadata = get_forward_context().attn_metadata
+    fp8_dtype = current_platform.fp8_dtype()
+    # assert isinstance(attn_metadata, dict)
+    if not isinstance(attn_metadata, dict):
+        return rocm_aiter_sparse_attn_indexer_fake(
+            hidden_states,
+            k_cache_prefix,
+            kv_cache,
+            q_fp8,
+            k,
+            weights,
+            quant_block_size,
+            scale_fmt,
+            topk_tokens,
+            head_dim,
+            max_model_len,
+            total_seq_lens,
+            topk_indices_buffer,
+        )
+    attn_metadata = attn_metadata[k_cache_prefix]
+    assert isinstance(attn_metadata, DeepseekV32IndexerMetadata)
+    slot_mapping = attn_metadata.slot_mapping
+    has_decode = attn_metadata.num_decodes > 0
+    has_prefill = attn_metadata.num_prefills > 0
+    num_decode_tokens = attn_metadata.num_decode_tokens
+
+    ops.indexer_k_quant_and_cache(
+        k,
+        kv_cache,
+        slot_mapping,
+        quant_block_size,
+        scale_fmt,
+    )
+
+    topk_indices_buffer[: hidden_states.shape[0]] = -1
+    if has_prefill:
+        prefill_metadata = attn_metadata.prefill
+        for chunk in prefill_metadata.chunks:
+            k_fp8 = torch.empty(
+                [chunk.total_seq_lens, head_dim],
+                device=k.device,
+                dtype=fp8_dtype,
+            )
+            k_scale = torch.empty(
+                [chunk.total_seq_lens, 4],
+                device=k.device,
+                dtype=torch.uint8,
+            )
+
+            ops.cp_gather_indexer_k_quant_cache(
+                kv_cache,
+                k_fp8,
+                k_scale,
+                chunk.block_table,
+                chunk.cu_seq_lens,
+            )
+
+            logits = rocm_fp8_mqa_logits(
+                q_fp8[chunk.token_start : chunk.token_end],
+                (k_fp8, k_scale.view(torch.float32)),
+                weights[chunk.token_start : chunk.token_end],
+                chunk.cu_seqlen_ks,
+                chunk.cu_seqlen_ke,
+            )
+            num_rows = logits.shape[0]
+            assert topk_tokens == 2048, "top_k_per_row assumes size 2048"
+            topk_indices = topk_indices_buffer[
+                chunk.token_start : chunk.token_end, :topk_tokens
+            ]
+            torch.ops._C.top_k_per_row_prefill(
+                logits,
+                chunk.cu_seqlen_ks,
+                chunk.cu_seqlen_ke,
+                topk_indices,
+                num_rows,
+                logits.stride(0),
+                logits.stride(1),
+                topk_tokens,
+            )
+
+    if has_decode:
+        decode_metadata = attn_metadata.decode
+        # kv_cache size requirement [num_block, block_size, n_head, head_dim],
+        # we only have [num_block, block_size, head_dim],
+        kv_cache = kv_cache.unsqueeze(-2)
+        decode_lens = decode_metadata.decode_lens
+        if decode_metadata.requires_padding:
+            # pad in edge case where we have short chunked prefill length <
+            # decode_threshold since we unstrictly split
+            # prefill and decode by decode_threshold
+            # (currently set to 1 + speculative tokens)
+            padded_q_fp8_decode_tokens = pack_seq_triton(
+                q_fp8[:num_decode_tokens], decode_lens
+            )
+        else:
+            padded_q_fp8_decode_tokens = q_fp8[:num_decode_tokens].reshape(
+                decode_lens.shape[0], -1, *q_fp8.shape[1:]
+            )
+        # TODO: move and optimize below logic with triton kernels
+        batch_size = padded_q_fp8_decode_tokens.shape[0]
+        next_n = padded_q_fp8_decode_tokens.shape[1]
+        assert batch_size == decode_metadata.seq_lens.shape[0]
+        num_padded_tokens = batch_size * next_n
+
+        logits = rocm_fp8_paged_mqa_logits(
+            padded_q_fp8_decode_tokens,
+            kv_cache,
+            weights[:num_padded_tokens],
+            decode_metadata.seq_lens,
+            decode_metadata.block_table,
+            decode_metadata.schedule_metadata,
+            max_model_len=max_model_len,
+        )
+
+        num_rows = logits.shape[0]
+        assert topk_tokens == 2048, "top_k_per_row assumes size 2048"
+        topk_indices = topk_indices_buffer[:num_decode_tokens, :topk_tokens]
+        torch.ops._C.top_k_per_row_decode(
+            logits,
+            next_n,
+            decode_metadata.seq_lens,
+            topk_indices,
+            num_rows,
+            logits.stride(0),
+            logits.stride(1),
+            topk_tokens,
+        )
+
+        if decode_metadata.requires_padding:
+            # if padded, we need to unpack
+            # the topk indices removing padded tokens
+            topk_indices = unpack_seq_triton(
+                topk_indices.reshape(batch_size, -1, topk_indices.shape[-1]),
+                decode_lens,
+            )
+            topk_indices_buffer[:num_decode_tokens, : topk_indices.shape[-1]] = (
+                topk_indices
+            )
+
+    return topk_indices_buffer

vllm/v1/attention/ops/triton_decode_attention.py

@@ -0,0 +1,709 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Adapted from
+# https://github.com/sgl-project/sglang/blob/9f635ea50de920aa507f486daafba26a5b837574/python/sglang/srt/layers/attention/triton_ops/decode_attention.py
+# which was originally adapted from
+# https://github.com/ModelTC/lightllm/blob/96353e868a840db4d103138caf15ed9dbea8c186/lightllm/models/deepseek2/triton_kernel/gqa_flash_decoding_stage1.py
+# https://github.com/ModelTC/lightllm/blob/96353e868a840db4d103138caf15ed9dbea8c186/lightllm/models/deepseek2/triton_kernel/gqa_flash_decoding_stage2.py
+
+# Changes:
+# - Add support for page size >= 1.
+
+# Copyright 2025 vLLM Team
+# Copyright 2023-2024 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""
+Memory-efficient attention for decoding.
+It supports page size >= 1.
+"""
+
+import logging
+
+from packaging import version
+
+from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
+
+is_hip_ = current_platform.is_rocm()
+
+logger = logging.getLogger(__name__)
+
+# Only print the following warnings when triton version < 3.2.0.
+# The issue won't affect performance or accuracy.
+if version.parse(triton.__version__) < version.parse("3.2.0"):
+    logger.warning(
+        "The following error message 'operation scheduled before its operands' "
+        "can be ignored."
+    )
+
+
+@triton.jit
+def tanh(x):
+    # Tanh is just a scaled sigmoid
+    return 2 * tl.sigmoid(2 * x) - 1
+
+
+@triton.jit
+def _fwd_kernel_stage1(
+    Q,
+    K_Buffer,
+    V_Buffer,
+    sm_scale,
+    Req_to_tokens,
+    B_Seqlen,
+    Att_Out,
+    stride_req_to_tokens_b,
+    stride_qbs,
+    stride_qh,
+    stride_buf_kbs,
+    stride_buf_kh,
+    stride_buf_vbs,
+    stride_buf_vh,
+    stride_mid_ob,
+    stride_mid_oh,
+    stride_mid_os,
+    kv_group_num: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DV: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    NUM_KV_SPLITS: tl.constexpr,
+    PAGE_SIZE: tl.constexpr,
+    logit_cap: tl.constexpr,
+    Lk: tl.constexpr,
+    Lv: tl.constexpr,
+):
+    cur_batch = tl.program_id(0)
+    cur_head = tl.program_id(1)
+    split_kv_id = tl.program_id(2)
+
+    cur_kv_head = cur_head // kv_group_num
+
+    offs_d = tl.arange(0, BLOCK_DMODEL)
+    offs_dv = tl.arange(0, BLOCK_DV)
+    mask_d = offs_d < Lk
+    mask_dv = offs_dv < Lv
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+    cur_batch_req_idx = cur_batch
+
+    off_q = cur_batch * stride_qbs + cur_head * stride_qh + offs_d
+    q = tl.load(Q + off_q, mask=mask_d, other=0.0)
+
+    kv_len_per_split = tl.cdiv(cur_batch_seq_len, NUM_KV_SPLITS)
+    split_kv_start = kv_len_per_split * split_kv_id
+    split_kv_end = tl.minimum(split_kv_start + kv_len_per_split, cur_batch_seq_len)
+
+    e_max = -float("inf")
+    e_sum = 0.0
+    acc = tl.zeros([BLOCK_DV], dtype=tl.float32)
+
+    if split_kv_end > split_kv_start:
+        for start_n in range(split_kv_start, split_kv_end, BLOCK_N):
+            offs_n = start_n + tl.arange(0, BLOCK_N)
+            kv_page_number = tl.load(
+                Req_to_tokens
+                + stride_req_to_tokens_b * cur_batch_req_idx
+                + offs_n // PAGE_SIZE,
+                mask=offs_n < split_kv_end,
+                other=0,
+            )
+            kv_loc = kv_page_number * PAGE_SIZE + offs_n % PAGE_SIZE
+            offs_buf_k = (
+                kv_loc[:, None] * stride_buf_kbs
+                + cur_kv_head * stride_buf_kh
+                + offs_d[None, :]
+            )
+            k = tl.load(
+                K_Buffer + offs_buf_k,
+                mask=(offs_n[:, None] < split_kv_end) & (mask_d[None, :]),
+                other=0.0,
+            )
+            qk = tl.sum(q[None, :] * k, 1)
+            qk *= sm_scale
+
+            if logit_cap > 0:
+                qk = logit_cap * tanh(qk / logit_cap)
+
+            qk = tl.where(offs_n < split_kv_end, qk, float("-inf"))
+
+            offs_buf_v = (
+                kv_loc[:, None] * stride_buf_vbs
+                + cur_kv_head * stride_buf_vh
+                + offs_dv[None, :]
+            )
+            v = tl.load(
+                V_Buffer + offs_buf_v,
+                mask=(offs_n[:, None] < split_kv_end) & (mask_dv[None, :]),
+                other=0.0,
+            )
+
+            n_e_max = tl.maximum(tl.max(qk, 0), e_max)
+            re_scale = tl.exp(e_max - n_e_max)
+            p = tl.exp(qk - n_e_max)
+            acc *= re_scale
+            acc += tl.sum(p[:, None] * v, 0)
+
+            e_sum = e_sum * re_scale + tl.sum(p, 0)
+            e_max = n_e_max
+
+        offs_mid_o = (
+            cur_batch * stride_mid_ob
+            + cur_head * stride_mid_oh
+            + split_kv_id * stride_mid_os
+            + offs_dv
+        )
+
+        tl.store(
+            Att_Out + offs_mid_o,
+            acc / e_sum,
+            mask=(mask_dv),
+        )
+
+        offs_mid_o_1 = (
+            cur_batch * stride_mid_ob
+            + cur_head * stride_mid_oh
+            + split_kv_id * stride_mid_os
+            + Lv
+        )
+
+        tl.store(
+            Att_Out + offs_mid_o_1,
+            e_max + tl.log(e_sum),
+        )
+
+
+def _decode_att_m_fwd(
+    q,
+    k_buffer,
+    v_buffer,
+    att_out,
+    Req_to_tokens,
+    B_Seqlen,
+    num_kv_splits,
+    sm_scale,
+    page_size,
+    logit_cap,
+):
+    BLOCK = 64 if not is_hip_ else 8
+
+    NUM_KV_SPLITS = num_kv_splits
+    Lk = k_buffer.shape[-1]
+    Lv = v_buffer.shape[-1]
+
+    batch, head_num = q.shape[0], q.shape[1]
+
+    grid = (batch, head_num, NUM_KV_SPLITS)
+    kv_group_num = q.shape[1] // k_buffer.shape[-2]
+
+    num_warps = 4
+    if kv_group_num != 1:
+        num_warps = 1 if is_hip_ else 2
+
+    BLOCK_DMODEL = triton.next_power_of_2(Lk)
+    BLOCK_DV = triton.next_power_of_2(Lv)
+
+    _fwd_kernel_stage1[grid](
+        q,
+        k_buffer,
+        v_buffer,
+        sm_scale,
+        Req_to_tokens,
+        B_Seqlen,
+        att_out,
+        Req_to_tokens.stride(0),
+        q.stride(0),
+        q.stride(1),
+        k_buffer.stride(-3),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        k_buffer.stride(-2),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        v_buffer.stride(-3),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        v_buffer.stride(-2),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        att_out.stride(0),
+        att_out.stride(1),
+        att_out.stride(2),
+        kv_group_num=kv_group_num,
+        BLOCK_DMODEL=BLOCK_DMODEL,
+        BLOCK_DV=BLOCK_DV,
+        BLOCK_N=BLOCK,
+        NUM_KV_SPLITS=NUM_KV_SPLITS,
+        PAGE_SIZE=page_size,
+        logit_cap=logit_cap,
+        num_warps=num_warps,
+        num_stages=2,
+        Lk=Lk,
+        Lv=Lv,
+    )
+
+
+@triton.jit
+def _fwd_grouped_kernel_stage1(
+    Q,
+    K_Buffer,
+    V_Buffer,
+    sm_scale,
+    Req_to_tokens,
+    B_Seqlen,
+    Att_Out,
+    stride_req_to_tokens_b,
+    stride_qbs,
+    stride_qh,
+    stride_buf_kbs,
+    stride_buf_kh,
+    stride_buf_vbs,
+    stride_buf_vh,
+    stride_mid_ob,
+    stride_mid_oh,
+    stride_mid_os,
+    kv_group_num: tl.constexpr,
+    q_head_num: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DPE: tl.constexpr,
+    BLOCK_DV: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_H: tl.constexpr,
+    NUM_KV_SPLITS: tl.constexpr,
+    PAGE_SIZE: tl.constexpr,
+    logit_cap: tl.constexpr,
+    Lk: tl.constexpr,
+    Lv: tl.constexpr,
+):
+    cur_batch = tl.program_id(0)
+    cur_head_id = tl.program_id(1)
+    cur_kv_head = cur_head_id // tl.cdiv(kv_group_num, BLOCK_H)
+    split_kv_id = tl.program_id(2)
+
+    VALID_BLOCK_H: tl.constexpr = BLOCK_H if kv_group_num > BLOCK_H else kv_group_num
+    cur_head = cur_head_id * VALID_BLOCK_H + tl.arange(0, BLOCK_H)
+    mask_h = cur_head < (cur_head_id + 1) * VALID_BLOCK_H
+    mask_h = mask_h & (cur_head < q_head_num)
+
+    offs_d = tl.arange(0, BLOCK_DMODEL)
+    offs_dv = tl.arange(0, BLOCK_DV)
+    mask_d = offs_d < Lk
+    mask_dv = offs_dv < Lv
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+    cur_batch_req_idx = cur_batch
+
+    offs_q = cur_batch * stride_qbs + cur_head[:, None] * stride_qh + offs_d[None, :]
+    q = tl.load(Q + offs_q, mask=(mask_h[:, None]) & (mask_d[None, :]), other=0.0)
+
+    if BLOCK_DPE > 0:
+        offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
+        mask_dpe = offs_dpe < Lk
+        off_qpe = (
+            cur_batch * stride_qbs + cur_head[:, None] * stride_qh + offs_dpe[None, :]
+        )
+        qpe = tl.load(
+            Q + off_qpe, mask=(mask_h[:, None]) & (mask_dpe[None, :]), other=0.0
+        )
+
+    kv_len_per_split = tl.cdiv(cur_batch_seq_len, NUM_KV_SPLITS)
+    split_kv_start = kv_len_per_split * split_kv_id
+    split_kv_end = tl.minimum(split_kv_start + kv_len_per_split, cur_batch_seq_len)
+
+    e_max = tl.zeros([BLOCK_H], dtype=tl.float32) - float("inf")
+    e_sum = tl.zeros([BLOCK_H], dtype=tl.float32)
+    acc = tl.zeros([BLOCK_H, BLOCK_DV], dtype=tl.float32)
+
+    if split_kv_end > split_kv_start:
+        for start_n in range(split_kv_start, split_kv_end, BLOCK_N):
+            offs_n = start_n + tl.arange(0, BLOCK_N)
+            kv_page_number = tl.load(
+                Req_to_tokens
+                + stride_req_to_tokens_b * cur_batch_req_idx
+                + offs_n // PAGE_SIZE,
+                mask=offs_n < split_kv_end,
+                other=0,
+            )
+            kv_loc = kv_page_number * PAGE_SIZE + offs_n % PAGE_SIZE
+            offs_buf_k = (
+                kv_loc[None, :] * stride_buf_kbs
+                + cur_kv_head * stride_buf_kh
+                + offs_d[:, None]
+            )
+            k = tl.load(
+                K_Buffer + offs_buf_k,
+                mask=(offs_n[None, :] < split_kv_end) & (mask_d[:, None]),
+                other=0.0,
+            )
+            qk = tl.dot(q, k.to(q.dtype))
+            if BLOCK_DPE > 0:
+                offs_buf_kpe = (
+                    kv_loc[None, :] * stride_buf_kbs
+                    + cur_kv_head * stride_buf_kh
+                    + offs_dpe[:, None]
+                )
+                kpe = tl.load(
+                    K_Buffer + offs_buf_kpe,
+                    mask=(offs_n[None, :] < split_kv_end) & (mask_dpe[:, None]),
+                    other=0.0,
+                )
+                qk += tl.dot(qpe, kpe.to(qpe.dtype))
+            qk *= sm_scale
+
+            if logit_cap > 0:
+                qk = logit_cap * tanh(qk / logit_cap)
+
+            qk = tl.where(
+                mask_h[:, None] & (offs_n[None, :] < split_kv_end), qk, float("-inf")
+            )
+
+            offs_buf_v = (
+                kv_loc[:, None] * stride_buf_vbs
+                + cur_kv_head * stride_buf_vh
+                + offs_dv[None, :]
+            )
+            v = tl.load(
+                V_Buffer + offs_buf_v,
+                mask=(offs_n[:, None] < split_kv_end) & (mask_dv[None, :]),
+                other=0.0,
+            )
+
+            n_e_max = tl.maximum(tl.max(qk, 1), e_max)
+            re_scale = tl.exp(e_max - n_e_max)
+            p = tl.exp(qk - n_e_max[:, None])
+            acc *= re_scale[:, None]
+            acc += tl.dot(p.to(v.dtype), v)
+
+            e_sum = e_sum * re_scale + tl.sum(p, 1)
+            e_max = n_e_max
+
+        offs_mid_o = (
+            cur_batch * stride_mid_ob
+            + cur_head[:, None] * stride_mid_oh
+            + split_kv_id * stride_mid_os
+            + offs_dv[None, :]
+        )
+
+        tl.store(
+            Att_Out + offs_mid_o,
+            acc / e_sum[:, None],
+            mask=(mask_h[:, None]) & (mask_dv[None, :]),
+        )
+
+        offs_mid_o_1 = (
+            cur_batch * stride_mid_ob
+            + cur_head * stride_mid_oh
+            + split_kv_id * stride_mid_os
+            + Lv
+        )
+
+        tl.store(
+            Att_Out + offs_mid_o_1,
+            e_max + tl.log(e_sum),
+            mask=mask_h,
+        )
+
+
+def _decode_grouped_att_m_fwd(
+    q,
+    k_buffer,
+    v_buffer,
+    att_out,
+    Req_to_tokens,
+    B_Seqlen,
+    num_kv_splits,
+    sm_scale,
+    page_size,
+    logit_cap,
+):
+    BLOCK = 32
+    Lk = k_buffer.shape[-1]
+    Lv = v_buffer.shape[-1]
+
+    # [TODO] work around shmem limit on MI3xx
+    if is_hip_ and Lk >= 576:
+        BLOCK = 16
+
+    if Lk == 576:
+        BLOCK_DMODEL = 512
+        BLOCK_DPE = 64
+    elif Lk == 288:
+        BLOCK_DMODEL = 256
+        BLOCK_DPE = 32
+    else:
+        BLOCK_DMODEL = triton.next_power_of_2(Lk)
+        BLOCK_DPE = 0
+    BLOCK_DV = triton.next_power_of_2(Lv)
+
+    batch, head_num = q.shape[0], q.shape[1]
+    kv_group_num = q.shape[1] // k_buffer.shape[-2]
+
+    BLOCK_H = 16
+    NUM_KV_SPLITS = num_kv_splits
+    grid = (
+        batch,
+        triton.cdiv(head_num, min(BLOCK_H, kv_group_num)),
+        NUM_KV_SPLITS,
+    )
+
+    extra_kargs = {}
+    num_stages = 2
+    if is_hip_:
+        # https://rocm.docs.amd.com/en/latest/how-to/rocm-for-ai/inference-optimization/workload.html#mi300x-triton-kernel-performance-optimization
+        # https://github.com/triton-lang/triton/blob/main/third_party/amd/backend/compiler.py
+        extra_kargs = {"waves_per_eu": 1, "matrix_instr_nonkdim": 16, "kpack": 2}
+        num_stages = 1
+
+    _fwd_grouped_kernel_stage1[grid](
+        q,
+        k_buffer,
+        v_buffer,
+        sm_scale,
+        Req_to_tokens,
+        B_Seqlen,
+        att_out,
+        Req_to_tokens.stride(0),
+        q.stride(0),
+        q.stride(1),
+        k_buffer.stride(-3),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        k_buffer.stride(-2),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        v_buffer.stride(-3),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        v_buffer.stride(-2),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        att_out.stride(0),
+        att_out.stride(1),
+        att_out.stride(2),
+        kv_group_num=kv_group_num,
+        q_head_num=head_num,
+        BLOCK_DMODEL=BLOCK_DMODEL,
+        BLOCK_DPE=BLOCK_DPE,
+        BLOCK_DV=BLOCK_DV,
+        BLOCK_N=BLOCK,
+        BLOCK_H=BLOCK_H,
+        NUM_KV_SPLITS=NUM_KV_SPLITS,
+        PAGE_SIZE=page_size,
+        logit_cap=logit_cap,
+        num_warps=4,
+        num_stages=num_stages,
+        Lk=Lk,
+        Lv=Lv,
+        **extra_kargs,
+    )
+
+
+@triton.jit
+def _fwd_kernel_stage2(
+    Mid_O,
+    o,
+    lse,
+    B_Seqlen,
+    stride_mid_ob,
+    stride_mid_oh,
+    stride_mid_os,
+    stride_obs,
+    stride_oh,
+    stride_lse_bs,
+    NUM_KV_SPLITS: tl.constexpr,
+    BLOCK_DV: tl.constexpr,
+    Lv: tl.constexpr,
+):
+    cur_batch = tl.program_id(0)
+    cur_head = tl.program_id(1)
+
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+
+    offs_d = tl.arange(0, BLOCK_DV)
+    mask_d = offs_d < Lv
+
+    e_sum = 0.0
+    e_max = -float("inf")
+    acc = tl.zeros([BLOCK_DV], dtype=tl.float32)
+
+    offs_v = cur_batch * stride_mid_ob + cur_head * stride_mid_oh + offs_d
+    offs_logic = cur_batch * stride_mid_ob + cur_head * stride_mid_oh + Lv
+
+    for split_kv_id in range(0, NUM_KV_SPLITS):
+        kv_len_per_split = tl.cdiv(cur_batch_seq_len, NUM_KV_SPLITS)
+        split_kv_start = kv_len_per_split * split_kv_id
+        split_kv_end = tl.minimum(split_kv_start + kv_len_per_split, cur_batch_seq_len)
+
+        if split_kv_end > split_kv_start:
+            tv = tl.load(
+                Mid_O + offs_v + split_kv_id * stride_mid_os, mask=mask_d, other=0.0
+            )
+            tlogic = tl.load(Mid_O + offs_logic + split_kv_id * stride_mid_os)
+            n_e_max = tl.maximum(tlogic, e_max)
+
+            old_scale = tl.exp(e_max - n_e_max)
+            acc *= old_scale
+            exp_logic = tl.exp(tlogic - n_e_max)
+            acc += exp_logic * tv
+
+            e_sum = e_sum * old_scale + exp_logic
+            e_max = n_e_max
+
+    tl.store(
+        o + cur_batch * stride_obs + cur_head * stride_oh + offs_d,
+        acc / e_sum,
+        mask=mask_d,
+    )
+    lse_val = e_max + tl.log(e_sum)
+    tl.store(
+        lse + cur_batch * stride_lse_bs + cur_head,
+        lse_val,
+    )
+
+
+def _decode_softmax_reducev_fwd(
+    logits,
+    q,
+    o,
+    lse,
+    v_buffer,
+    b_seq_len,
+    num_kv_splits,
+):
+    batch, head_num = q.shape[0], q.shape[1]
+    Lv = v_buffer.shape[-1]
+    BLOCK_DV = triton.next_power_of_2(Lv)
+
+    NUM_KV_SPLITS = num_kv_splits
+
+    extra_kargs = {}
+    if is_hip_:
+        # https://rocm.docs.amd.com/en/docs-6.2.0/how-to/llm-fine-tuning-optimization/optimizing-triton-kernel.html
+        # https://github.com/triton-lang/triton/blob/main/third_party/amd/backend/compiler.py
+        extra_kargs = {"waves_per_eu": 4, "matrix_instr_nonkdim": 16, "kpack": 2}
+
+    grid = (batch, head_num)
+    _fwd_kernel_stage2[grid](
+        logits,
+        o,
+        lse,
+        b_seq_len,
+        logits.stride(0),
+        logits.stride(1),
+        logits.stride(2),
+        o.stride(0),
+        o.stride(1),
+        lse.stride(0),
+        NUM_KV_SPLITS=NUM_KV_SPLITS,
+        BLOCK_DV=BLOCK_DV,
+        Lv=Lv,
+        num_warps=4,
+        num_stages=2,
+        **extra_kargs,
+    )
+
+
+def decode_attention_fwd_normal(
+    q,
+    k_buffer,
+    v_buffer,
+    o,
+    lse,
+    req_to_token,
+    b_seq_len,
+    attn_logits,
+    num_kv_splits,
+    sm_scale,
+    page_size,
+    logit_cap=0.0,
+):
+    _decode_att_m_fwd(
+        q,
+        k_buffer,
+        v_buffer,
+        attn_logits,
+        req_to_token,
+        b_seq_len,
+        num_kv_splits,
+        sm_scale,
+        page_size,
+        logit_cap,
+    )
+    _decode_softmax_reducev_fwd(
+        attn_logits, q, o, lse, v_buffer, b_seq_len, num_kv_splits
+    )
+
+
+def decode_attention_fwd_grouped(
+    q,
+    k_buffer,
+    v_buffer,
+    o,
+    lse,
+    req_to_token,
+    b_seq_len,
+    attn_logits,
+    num_kv_splits,
+    sm_scale,
+    page_size,
+    logit_cap=0.0,
+):
+    _decode_grouped_att_m_fwd(
+        q,
+        k_buffer,
+        v_buffer,
+        attn_logits,
+        req_to_token,
+        b_seq_len,
+        num_kv_splits,
+        sm_scale,
+        page_size,
+        logit_cap,
+    )
+    _decode_softmax_reducev_fwd(
+        attn_logits, q, o, lse, v_buffer, b_seq_len, num_kv_splits
+    )
+
+
+def decode_attention_fwd(
+    q,
+    k_buffer,
+    v_buffer,
+    o,
+    lse,
+    req_to_token,
+    b_seq_len,
+    attn_logits,
+    num_kv_splits,
+    sm_scale,
+    page_size=1,
+    logit_cap=0.0,
+):
+    assert num_kv_splits == attn_logits.shape[2]
+    kv_group_num = q.shape[1] // v_buffer.shape[-2]
+
+    if kv_group_num == 1:
+        # MHA
+        decode_attention_fwd_normal(
+            q,
+            k_buffer,
+            v_buffer,
+            o,
+            lse,
+            req_to_token,
+            b_seq_len,
+            attn_logits,
+            num_kv_splits,
+            sm_scale,
+            page_size,
+            logit_cap,
+        )
+    else:
+        # GQA/MQA/MLA
+        decode_attention_fwd_grouped(
+            q,
+            k_buffer,
+            v_buffer,
+            o,
+            lse,
+            req_to_token,
+            b_seq_len,
+            attn_logits,
+            num_kv_splits,
+            sm_scale,
+            page_size,
+            logit_cap,
+        )

vllm/v1/attention/ops/triton_merge_attn_states.py

@@ -0,0 +1,116 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+
+# Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
+# can be used to combine partial attention results (in the split-KV case)
+def merge_attn_states(
+    output: torch.Tensor,
+    prefix_output: torch.Tensor,
+    prefix_lse: torch.Tensor,
+    suffix_output: torch.Tensor,
+    suffix_lse: torch.Tensor,
+    output_lse: torch.Tensor | None = None,
+) -> None:
+    num_tokens = output.shape[0]
+    num_query_heads = output.shape[1]
+    head_size = output.shape[2]
+    padded_head_size = triton.next_power_of_2(head_size)
+    # We assume the output stride on num_head is not always as same as the
+    # `suffix_output` and `prefix_output`, as them might be padded by the attention
+    # backend.
+    prefix_head_stride = prefix_output.stride(1)
+    output_head_stride = output.stride(1)
+    # TODO(woosuk): Use CUDA kernel instead of Triton to minimize CPU overhead.
+    merge_attn_states_kernel[(num_tokens, num_query_heads)](
+        output,
+        output_lse,
+        prefix_output,
+        prefix_lse,
+        suffix_output,
+        suffix_lse,
+        prefix_head_stride,
+        output_head_stride,
+        head_size,
+        padded_head_size,
+        output_lse is not None,
+    )
+
+
+@triton.jit
+def merge_attn_states_kernel(
+    output,  # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]
+    output_lse,  # [NUM_HEADS, NUM_TOKENS]
+    prefix_output,  # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]
+    prefix_lse,  # [NUM_HEADS, NUM_TOKENS]
+    suffix_output,  # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]
+    suffix_lse,  # [NUM_HEADS, NUM_TOKENS]
+    prefix_head_stride,
+    output_head_stride,
+    HEAD_SIZE: tl.constexpr,
+    PADDED_HEAD_SIZE: tl.constexpr,
+    OUTPUT_LSE: tl.constexpr,
+):
+    token_idx = tl.program_id(0)
+    num_tokens = tl.num_programs(0)
+    head_idx = tl.program_id(1)
+    num_heads = tl.num_programs(1)
+
+    p_lse = tl.load(prefix_lse + head_idx * num_tokens + token_idx)
+    s_lse = tl.load(suffix_lse + head_idx * num_tokens + token_idx)
+
+    # FA2 and FA3 have different behavior for when the sum-exp is 0, this namely
+    # arises with 0 len seqlens. FA3 returns -inf here while FA2 returns inf.
+    # If we see an inf assume FA2 and convert inf to -inf for consistency
+    # and correctness. Inf generally doesn't make sense in this context outside
+    # of undefined-behavior/FA2-case, so I think this a safe assumption.
+    p_lse = float("-inf") if p_lse == float("inf") else p_lse
+    s_lse = float("-inf") if s_lse == float("inf") else s_lse
+
+    max_lse = tl.maximum(p_lse, s_lse)
+    p_lse = p_lse - max_lse
+    s_lse = s_lse - max_lse
+    # Will reuse precomputed Exp values for scale factor computation.
+    p_se = tl.exp(p_lse)
+    s_se = tl.exp(s_lse)
+    out_se = p_se + s_se
+
+    if OUTPUT_LSE:
+        out_lse = tl.log(out_se) + max_lse
+        tl.store(output_lse + head_idx * num_tokens + token_idx, out_lse)
+
+    head_arange = tl.arange(0, PADDED_HEAD_SIZE)
+    head_mask = head_arange < HEAD_SIZE
+    p_out = tl.load(
+        prefix_output
+        + token_idx * num_heads * prefix_head_stride
+        + head_idx * prefix_head_stride
+        + head_arange,
+        mask=head_mask,
+    )
+    s_out = tl.load(
+        suffix_output
+        + token_idx * num_heads * prefix_head_stride
+        + head_idx * prefix_head_stride
+        + head_arange,
+        mask=head_mask,
+    )
+
+    # NOTE(woosuk): Be careful with the numerical stability.
+    # We should compute the scale first, and then multiply it with the output.
+    # Do not multiply the output with tl.exp(p_lse) or tl.exp(s_lse) directly.
+    p_scale = p_se / out_se
+    s_scale = s_se / out_se
+    out = p_out * p_scale + s_out * s_scale
+    tl.store(
+        output
+        + token_idx * num_heads * output_head_stride
+        + head_idx * output_head_stride
+        + head_arange,
+        out,
+        mask=head_mask,
+    )

vllm/v1/attention/ops/triton_prefill_attention.py

@@ -0,0 +1,253 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Adapted from
+# https://github.com/sgl-project/sglang/blob/97cb762bb65ebf05025eb342de03c184660427a3/python/sglang/srt/layers/attention/triton_ops/prefill_attention.py
+# Changes:
+# - Add support for sliding window attention
+
+# Copyright 2023-2024 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""
+Memory-efficient attention for prefill.
+It supports page size = 1.
+"""
+
+# Adapted from
+# https://github.com/ModelTC/lightllm/blob/f2a54f0912293f683bf1d1695fd12c4098a5bf82/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py#L1
+import torch
+
+from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
+from vllm.utils.math_utils import RCP_LN2
+
+
+@triton.jit
+def _fwd_kernel(
+    Q,
+    K,
+    V,
+    sm_scale,
+    B_Start_Loc,
+    B_Seqlen,
+    Out,
+    stride_qbs,
+    stride_qh,
+    stride_kbs,
+    stride_kh,
+    stride_vbs,
+    stride_vh,
+    stride_obs,
+    stride_oh,
+    kv_group_num: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    IS_CAUSAL: tl.constexpr,
+    SLIDING_WINDOW_Q: tl.constexpr,
+    SLIDING_WINDOW_K: tl.constexpr,
+    Lk: tl.constexpr,
+):
+    cur_batch = tl.program_id(0)
+    cur_head = tl.program_id(1)
+    start_m = tl.program_id(2)
+
+    cur_kv_head = cur_head // kv_group_num
+
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+    cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
+
+    block_start_loc = BLOCK_M * start_m
+
+    # initialize offsets
+    offs_n = tl.arange(0, BLOCK_N)
+    offs_d = tl.arange(0, BLOCK_DMODEL)
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    off_q = (
+        (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs
+        + cur_head * stride_qh
+        + offs_d[None, :]
+    )
+    off_k = offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh + offs_d[:, None]
+    off_v = offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh + offs_d[None, :]
+
+    mask_d = offs_d < Lk
+
+    q = tl.load(
+        Q + off_q,
+        mask=(offs_m[:, None] < cur_batch_seq_len) & (mask_d[None, :]),
+        other=0.0,
+    )
+
+    k_ptrs = K + off_k
+    v_ptrs = V + off_v
+
+    # initialize pointer to m and l
+    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
+    l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
+    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
+
+    block_mask = tl.where(block_start_loc < cur_batch_seq_len, 1, 0)
+
+    # Calculate the end position for attention computation
+    end_n = cur_batch_seq_len
+
+    # Apply causal attention pruning and sliding window attention pruning
+    end_n = tl.minimum(end_n, (start_m + 1) * BLOCK_M) if IS_CAUSAL else end_n
+
+    # Calculate the start position for backward sliding window
+    start_n_limit = 0
+    end_n_limit = block_mask * end_n
+
+    for start_n in range(start_n_limit, end_n_limit, BLOCK_N):
+        # -- prepare attention mask ----
+        # Position indices in the sequence
+        pos_q = offs_m[:, None]  # Query positions [BLOCK_M, 1]
+        pos_k = start_n + offs_n[None, :]  # Key positions [1, BLOCK_N]
+
+        # Valid sequence mask
+        mask = pos_k < cur_batch_seq_len
+        # Causal mask
+        if IS_CAUSAL:
+            mask &= pos_q >= pos_k
+
+        # Bidirectional sliding window masks
+        sliding_mask_q = (
+            pos_q - pos_k <= SLIDING_WINDOW_Q if SLIDING_WINDOW_Q > 0 else None
+        )
+        sliding_mask_k = (
+            pos_k - pos_q <= SLIDING_WINDOW_K if SLIDING_WINDOW_K > 0 else None
+        )
+        if sliding_mask_q is not None:
+            mask &= sliding_mask_q
+        if sliding_mask_k is not None:
+            mask &= sliding_mask_k
+
+        start_n = tl.multiple_of(start_n, BLOCK_N)
+        # -- compute qk ----
+        k = tl.load(
+            k_ptrs + (cur_batch_in_all_start_index + start_n) * stride_kbs,
+            mask=(pos_k < cur_batch_seq_len) & (mask_d[:, None]),
+            other=0.0,
+        )
+
+        qk = tl.dot(q, k)
+        qk = tl.where(mask, qk * sm_scale, -1.0e8)
+        m_ij = tl.maximum(m_i, tl.max(qk, 1))
+        qk -= m_ij[:, None]
+        p = tl.math.exp2(qk)
+        l_ij = tl.sum(p, 1)
+
+        # -- update m_i and l_i
+        alpha = tl.math.exp2(m_i - m_ij)
+        l_i = l_i * alpha + l_ij
+        # -- update output accumulator --
+        acc = acc * alpha[:, None]
+        # update acc
+        v = tl.load(
+            v_ptrs + (cur_batch_in_all_start_index + start_n) * stride_vbs,
+            mask=((start_n + offs_n[:, None]) < cur_batch_seq_len) & (mask_d[None, :]),
+            other=0.0,
+        )
+        p = p.to(v.dtype)
+        acc = tl.dot(p, v, acc)
+        # update m_i
+        m_i = m_ij
+
+    acc = acc / l_i[:, None]
+    off_o = (
+        (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs
+        + cur_head * stride_oh
+        + offs_d[None, :]
+    )
+    out_ptrs = Out + off_o
+    tl.store(
+        out_ptrs, acc, mask=(offs_m[:, None] < cur_batch_seq_len) & (mask_d[None, :])
+    )
+
+
+def get_block_size(dtype: torch.dtype) -> int:
+    if dtype == torch.float32:
+        return 32
+    elif current_platform.is_cuda_alike() and current_platform.has_device_capability(
+        80
+    ):
+        return 128
+    else:
+        return 64
+
+
+def context_attention_fwd(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    o: torch.Tensor,
+    b_start_loc: torch.Tensor,
+    b_seq_len: torch.Tensor,
+    max_input_len: int,
+    is_causal: bool = True,
+    softmax_scale: float | None = None,
+    sliding_window_q: int | None = None,
+    sliding_window_k: int | None = None,
+):
+    """
+    q, k, v: [b * s, head, head_dim]
+    b_start_loc: [b]
+    b_seq_len: [b]
+    out: [b * s, head, head_dim]
+    """
+    BLOCK = get_block_size(q.dtype)
+
+    Lq, Lk, _ = q.shape[-1], k.shape[-1], v.shape[-1]
+
+    sm_scale = 1.0 / (Lq**0.5) if softmax_scale is None else softmax_scale
+    # rescale with 1/ln(2) for triton exp2
+    sm_scale *= RCP_LN2
+
+    batch, head = b_seq_len.shape[0], q.shape[1]
+    kv_group_num = q.shape[1] // k.shape[1]
+
+    grid = (batch, head, triton.cdiv(max_input_len, BLOCK))
+    num_warps = 4 if Lk <= 64 else 8
+
+    sliding_window_q = sliding_window_q if sliding_window_q is not None else 0
+    sliding_window_k = sliding_window_k if sliding_window_k is not None else 0
+
+    _fwd_kernel[grid](
+        q,
+        k,
+        v,
+        sm_scale,
+        b_start_loc,
+        b_seq_len,
+        o,
+        q.stride(0),
+        q.stride(1),
+        k.stride(0),
+        k.stride(1),
+        v.stride(0),
+        v.stride(1),
+        o.stride(0),
+        o.stride(1),
+        kv_group_num=kv_group_num,
+        BLOCK_M=BLOCK,
+        BLOCK_DMODEL=triton.next_power_of_2(Lk),
+        BLOCK_N=BLOCK,
+        IS_CAUSAL=is_causal,
+        SLIDING_WINDOW_Q=sliding_window_q,
+        SLIDING_WINDOW_K=sliding_window_k,
+        num_warps=num_warps,
+        num_stages=1,
+        Lk=Lk,
+    )

vllm/v1/attention/ops/triton_reshape_and_cache_flash.py

@@ -0,0 +1,395 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+
+from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
+
+
+@triton.jit
+def reshape_and_cache_kernel_flash(
+    key_ptr,  # [num_tokens, num_heads, head_size]
+    value_ptr,  # [num_tokens, num_heads, head_size]
+    key_cache_ptr,  # [num_blocks, block_size, num_heads, head_size]
+    value_cache_ptr,  # [num_blocks, block_size, num_heads, head_size]
+    slot_mapping_ptr,  # [num_tokens]
+    k_scale,  # float32
+    v_scale,  # float32
+    # strides
+    key_stride: tl.int64,
+    value_stride: tl.int64,
+    block_stride: tl.int64,
+    head_stride: tl.int64,
+    dim_stride_k: tl.int64,
+    dim_stride_v: tl.int64,
+    page_stride: tl.int64,
+    num_heads: tl.constexpr,
+    head_size: tl.constexpr,
+    block_size: tl.constexpr,
+    x: tl.constexpr,
+    USE_HEAD_MAJOR_LAYOUT: tl.constexpr,
+    # FP8 flags
+    FP8_KV_CACHE: tl.constexpr,
+    # tune parameters
+    TILE_SIZE: tl.constexpr,
+):
+    token_idx = tl.program_id(axis=0)
+    slot_idx = tl.load(slot_mapping_ptr + token_idx).to(tl.int64)
+    if slot_idx < 0:
+        # Padding token that should be ignored.
+        return
+
+    block_idx = slot_idx // block_size
+    block_offset = slot_idx % block_size
+
+    tile_i = tl.program_id(axis=1)
+    tile_offs = tl.arange(0, TILE_SIZE)
+    tile_pos = tile_i * TILE_SIZE + tile_offs
+    src_key_idx = token_idx * key_stride
+    src_value_idx = token_idx * value_stride
+
+    if USE_HEAD_MAJOR_LAYOUT:
+        # Decompose the tile index back into head and dim coordinates.
+        cur_head = tile_pos // head_size
+        cur_dim = tile_pos % head_size
+        # Value addressing (4D): [Block, Head, Dim, Slot]
+        tgt_idx_v = (
+            block_idx * block_stride
+            + cur_head * head_stride
+            + cur_dim * dim_stride_v
+            + block_offset * 1
+        )
+        # Key addressing (5D): [Block, Head, Dim//8, Slot, 8]
+        tgt_idx_k = (
+            block_idx * block_stride
+            + cur_head * head_stride
+            + (cur_dim // x) * dim_stride_k
+            + block_offset * x
+            + (cur_dim % x)
+        )
+    else:
+        tgt_base = block_idx * block_stride + block_offset * page_stride
+        tgt_idx_k = tgt_base + tile_pos
+        tgt_idx_v = tgt_base + tile_pos
+
+    # [TILE_SIZE]
+    key_load = tl.load(
+        key_ptr + src_key_idx + tile_pos, mask=tile_pos < (num_heads * head_size)
+    )
+    if FP8_KV_CACHE:
+        # tl.store will do the correct implicit cast to fp8,
+        # based on the key_cache_ptr.dtype.element_ty
+        key_tile = key_load if key_load.dtype.is_fp8() else key_load / tl.load(k_scale)
+    else:
+        key_tile = key_load
+
+    # [TILE_SIZE]
+    value_load = tl.load(
+        value_ptr + src_value_idx + tile_pos, mask=tile_pos < (num_heads * head_size)
+    )
+    if FP8_KV_CACHE:
+        if value_load.dtype.is_fp8():
+            value_tile = value_load
+        else:
+            # tl.store will do the correct implicit cast to fp8,
+            #  based on the value_cache_ptr.dtype.element_ty
+            value_tile = value_load / tl.load(v_scale)
+    else:
+        value_tile = value_load
+
+    tl.store(
+        key_cache_ptr + tgt_idx_k,
+        key_tile,
+        mask=tile_pos < (num_heads * head_size),
+    )
+    tl.store(
+        value_cache_ptr + tgt_idx_v,
+        value_tile,
+        mask=tile_pos < (num_heads * head_size),
+    )
+    return
+
+
+def triton_reshape_and_cache_flash(
+    key: torch.Tensor,  # [num_tokens, num_heads, head_size]
+    value: torch.Tensor,  # [num_tokens, num_heads, head_size]
+    # [num_blocks, block_size, num_heads, head_size]
+    key_cache: torch.Tensor,
+    # [num_blocks, block_size, num_heads, head_size]
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,  # [num_tokens]
+    kv_cache_dtype: str,  # "auto", "fp8"
+    k_scale: torch.Tensor,  # float32
+    v_scale: torch.Tensor,  # float32
+):
+    num_heads = key.shape[1]
+    head_size = key.shape[2]
+
+    use_head_major_layout = key_cache.ndim == 5
+    if use_head_major_layout:
+        block_size = key_cache.shape[3]
+        x = key_cache.shape[4]
+        head_stride = key_cache.stride(1)
+        dim_stride_k = key_cache.stride(2)
+        dim_stride_v = value_cache.stride(2)
+    else:
+        block_size = key_cache.shape[1]
+        x = 1
+        dim_stride_k = 0
+        dim_stride_v = 0
+        head_stride = key_cache.stride()[2]
+    n = num_heads * head_size
+    key_stride = key.stride()[0]
+    value_stride = value.stride()[0]
+    block_stride = key_cache.stride()[0]
+    page_stride = key_cache.stride()[1]
+
+    assert kv_cache_dtype == "auto" or kv_cache_dtype.startswith("fp8"), (
+        f"unsupported kv_cache_dtype (str), got {kv_cache_dtype}."
+    )
+    kv_cache_torch_dtype = (
+        current_platform.fp8_dtype()
+        if kv_cache_dtype.startswith("fp8")
+        else key_cache.dtype
+    )
+
+    if key_cache.dtype != kv_cache_torch_dtype and kv_cache_dtype.startswith("fp8"):
+        # to avoid erounous implicit cast in triton kernel (tl.store to uint8)
+        # (e.g. explicit cast to fp8e4m3fnuz is not supported in triton 3.4)
+        key_cache = key_cache.view(kv_cache_torch_dtype)
+        value_cache = value_cache.view(kv_cache_torch_dtype)
+    assert kv_cache_dtype != torch.uint8, (
+        "explicit fp8 cast and store to "
+        "uint8 is not supported by triton reshape_and_cache_flash"
+    )
+
+    FP8_KV_CACHE = kv_cache_dtype.startswith("fp8")
+    assert (not FP8_KV_CACHE) or kv_cache_torch_dtype in [
+        torch.float8_e4m3fn,
+        torch.float8_e5m2,
+        torch.uint8,
+        torch.float8_e4m3fnuz,
+    ], (
+        "unsupported dtype of KV cache tensor, got "
+        "{kv_cache_torch_dtype}. Supported kv cache dtypes: fp8e4m3fn, "
+        "fp8e5m2, uint8, bfloat16, float16, float32, fp8e4m3fnuz."
+    )
+
+    # heuristics instead of autotuning
+    TILE_SIZE = min(2048, triton.next_power_of_2(n))
+    if current_platform.is_rocm() or current_platform.is_xpu():
+        num_stages = 4
+        num_warps = 8
+    else:  # cuda
+        num_stages = 10
+        num_warps = 16
+        if torch.cuda.get_device_capability(key.device)[0] < 9:
+            TILE_SIZE = min(512, TILE_SIZE)
+
+    # TODO(ngl): maybe replace with static launch grid to avoid overhead if
+    #   using cudagraphs
+    grid = lambda meta: (
+        slot_mapping.shape[0],
+        triton.cdiv(n, meta["TILE_SIZE"]),
+    )
+
+    reshape_and_cache_kernel_flash[grid](
+        key_ptr=key,
+        value_ptr=value,
+        key_cache_ptr=key_cache,
+        value_cache_ptr=value_cache,
+        slot_mapping_ptr=slot_mapping,
+        k_scale=k_scale,
+        v_scale=v_scale,
+        # strides
+        key_stride=key_stride,
+        value_stride=value_stride,
+        block_stride=block_stride,
+        head_stride=head_stride,
+        dim_stride_k=dim_stride_k,
+        dim_stride_v=dim_stride_v,
+        page_stride=page_stride,
+        num_heads=num_heads,
+        head_size=head_size,
+        block_size=block_size,
+        x=x,
+        USE_HEAD_MAJOR_LAYOUT=use_head_major_layout,
+        # FP8 flags
+        FP8_KV_CACHE=FP8_KV_CACHE,
+        # autotune parameters
+        TILE_SIZE=TILE_SIZE,
+        num_warps=num_warps,
+        num_stages=num_stages,
+    )
+
+
+@triton.jit
+def reshape_and_cache_kernel_flash_diffkv(
+    key_ptr,  # [num_tokens, num_heads, head_size]
+    value_ptr,  # [num_tokens, num_heads, head_size_v]
+    kv_cache_ptr,  # [num_blocks, block_size, num_heads, head_size + head_size_v]
+    slot_mapping_ptr,  # [num_tokens]
+    k_scale,  # float32
+    v_scale,  # float32
+    # strides
+    key_stride: tl.int64,
+    value_stride: tl.int64,
+    block_stride: tl.int64,
+    page_stride: tl.int64,
+    num_heads: tl.constexpr,
+    head_size_k: tl.constexpr,
+    head_size_v: tl.constexpr,
+    block_size: tl.constexpr,
+    # FP8 flags
+    FP8_KV_CACHE: tl.constexpr,
+    # tune parameters
+    TILE_SIZE: tl.constexpr,
+):
+    token_idx = tl.program_id(axis=0)
+    slot_idx = tl.load(slot_mapping_ptr + token_idx).to(tl.int64)
+    if slot_idx < 0:
+        # Padding token that should be ignored.
+        return
+
+    tile_i = tl.program_id(axis=1)
+    tile_offs = tl.arange(0, TILE_SIZE)
+
+    block_idx = slot_idx // block_size
+    block_offset = slot_idx % block_size
+
+    src_key_idx = token_idx * key_stride + tile_i * head_size_k
+    src_value_idx = token_idx * value_stride + tile_i * head_size_v
+
+    tgt_idx = (
+        block_idx * block_stride
+        + block_offset * page_stride
+        + tile_i * (head_size_k + head_size_v)
+    )
+
+    # [TILE_SIZE]
+    key_load = tl.load(key_ptr + src_key_idx + tile_offs, mask=tile_offs < head_size_k)
+    if FP8_KV_CACHE:
+        # tl.store will do the correct implicit cast to fp8,
+        # based on the key_cache_ptr.dtype.element_ty
+        key_tile = key_load if key_load.dtype.is_fp8() else key_load / tl.load(k_scale)
+    else:
+        key_tile = key_load
+
+    # [TILE_SIZE]
+    value_load = tl.load(
+        value_ptr + src_value_idx + tile_offs, mask=tile_offs < head_size_v
+    )
+    if FP8_KV_CACHE:
+        if value_load.dtype.is_fp8():
+            value_tile = value_load
+        else:
+            # tl.store will do the correct implicit cast to fp8,
+            #  based on the value_cache_ptr.dtype.element_ty
+            value_tile = value_load / tl.load(v_scale)
+    else:
+        value_tile = value_load
+
+    tl.store(
+        kv_cache_ptr + tgt_idx + tile_offs,
+        key_tile,
+        mask=tile_offs < head_size_k,
+    )
+    tl.store(
+        kv_cache_ptr + tgt_idx + head_size_k + tile_offs,
+        value_tile,
+        mask=tile_offs < head_size_v,
+    )
+    return
+
+
+def triton_reshape_and_cache_flash_diffkv(
+    key: torch.Tensor,  # [num_tokens, num_heads, head_size]
+    value: torch.Tensor,  # [num_tokens, num_heads, head_size_v]
+    # [num_blocks, block_size, num_heads, head_size + head_size_v]
+    kv_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,  # [num_tokens]
+    kv_cache_dtype: str,  # "auto", "fp8"
+    k_scale: torch.Tensor,  # float32
+    v_scale: torch.Tensor,  # float32
+):
+    num_heads = key.shape[1]
+    head_size_k = key.shape[2]
+    head_size_v = value.shape[2]
+    block_size = kv_cache.shape[1]
+
+    k_stride = key.stride()[0]
+    v_stride = value.stride()[0]
+    block_stride = kv_cache.stride()[0]
+    page_stride = kv_cache.stride()[1]
+
+    assert kv_cache_dtype == "auto" or kv_cache_dtype.startswith("fp8"), (
+        f"unsupported kv_cache_dtype (str), got {kv_cache_dtype}."
+    )
+    kv_cache_torch_dtype = (
+        current_platform.fp8_dtype()
+        if kv_cache_dtype.startswith("fp8")
+        else kv_cache.dtype
+    )
+
+    if kv_cache.dtype != kv_cache_torch_dtype and kv_cache_dtype.startswith("fp8"):
+        # to avoid erounous implicit cast in triton kernel (tl.store to uint8)
+        # (e.g. explicit cast to fp8e4m3fnuz is not supported in triton 3.4)
+        kv_cache = kv_cache.view(kv_cache_torch_dtype)
+    assert kv_cache_dtype != torch.uint8, (
+        "explicit fp8 cast and store to "
+        "uint8 is not supported by triton reshape_and_cache_flash_diffkv"
+    )
+
+    FP8_KV_CACHE = kv_cache_dtype.startswith("fp8")
+    assert (not FP8_KV_CACHE) or kv_cache_torch_dtype in [
+        torch.float8_e4m3fn,
+        torch.float8_e5m2,
+        torch.uint8,
+        torch.float8_e4m3fnuz,
+    ], (
+        "unsupported dtype of KV cache tensor, got "
+        "{kv_cache_torch_dtype}. Supported kv cache dtypes: fp8e4m3fn, "
+        "fp8e5m2, uint8, bfloat16, float16, float32, fp8e4m3fnuz."
+    )
+
+    # heuristics instead of autotuning
+    TILE_SIZE = max(head_size_k, head_size_v)
+    TILE_SIZE = triton.next_power_of_2(TILE_SIZE)
+    if current_platform.is_rocm() or current_platform.is_xpu():
+        num_stages = 4
+        num_warps = 8
+    else:  # cuda
+        num_stages = 10
+        num_warps = 16
+
+    # TODO(ngl): maybe replace with static launch grid to avoid overhead if
+    #   using cudagraphs
+    grid = lambda meta: (
+        slot_mapping.shape[0],
+        num_heads,
+    )
+
+    reshape_and_cache_kernel_flash_diffkv[grid](
+        key_ptr=key,
+        value_ptr=value,
+        kv_cache_ptr=kv_cache,
+        slot_mapping_ptr=slot_mapping,
+        k_scale=k_scale,
+        v_scale=v_scale,
+        # strides
+        key_stride=k_stride,
+        value_stride=v_stride,
+        block_stride=block_stride,
+        page_stride=page_stride,
+        num_heads=num_heads,
+        head_size_k=head_size_k,
+        head_size_v=head_size_v,
+        block_size=block_size,
+        # FP8 flags
+        FP8_KV_CACHE=FP8_KV_CACHE,
+        # autotune parameters
+        TILE_SIZE=TILE_SIZE,
+        num_warps=num_warps,
+        num_stages=num_stages,
+    )

vllm/v1/attention/ops/vit_attn_wrappers.py

@@ -0,0 +1,270 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+This file contains ops for ViT attention to be compatible with torch.compile
+as there are operations here not supported by torch.compile (for instance,
+`.item()` in flash attention)
+
+Using these ops and wrapping vision blocks with `torch.compile` can speed up
+throughput in vision models by ~5% relative on H100, and improve token
+latencies by ~7% (see qwen2_5_vl for example usage)
+
+To use these ops, you must have a recent version of PyTorch installed (>= 2.4.0)
+"""
+
+import einops
+import torch
+import torch.nn.functional as F
+
+from vllm.platforms import current_platform
+from vllm.utils.torch_utils import direct_register_custom_op
+
+
+def flash_attn_maxseqlen_wrapper(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    batch_size: int,
+    is_rocm_aiter: bool,
+    fa_version: int | None,
+    scale: float | None = None,
+    cu_seqlens: torch.Tensor | None = None,
+    max_seqlen: torch.Tensor | None = None,
+) -> torch.Tensor:
+    kwargs = {}
+    if is_rocm_aiter:
+        from aiter import flash_attn_varlen_func
+    else:
+        from vllm.v1.attention.backends.fa_utils import flash_attn_varlen_func
+
+        # if not current_platform.is_rocm() and fa_version is not None:
+        #     kwargs["fa_version"] = fa_version
+
+    q_len = q.size(1)
+    if cu_seqlens is None:
+        cu_seqlens = torch.arange(
+            0, (batch_size + 1) * q_len, step=q_len, dtype=torch.int32, device=q.device
+        )
+    max_seqlen = q_len if max_seqlen is None else max_seqlen.item()
+
+    q, k, v = (einops.rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
+    output = flash_attn_varlen_func(
+        q,
+        k,
+        v,
+        cu_seqlens_q=cu_seqlens,
+        cu_seqlens_k=cu_seqlens,
+        max_seqlen_q=max_seqlen,
+        max_seqlen_k=max_seqlen,
+        dropout_p=0.0,
+        causal=False,
+        softmax_scale=scale,
+        **kwargs,
+    )
+    context_layer = einops.rearrange(output, "(b s) h d -> b s h d", b=batch_size)
+    return context_layer
+
+
+def flash_attn_maxseqlen_wrapper_fake(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    batch_size: int,
+    is_rocm_aiter: bool,
+    fa_version: int | None,
+    scale: float | None = None,
+    cu_seqlens: torch.Tensor | None = None,
+    max_seqlen: torch.Tensor | None = None,
+) -> torch.Tensor:
+    return torch.empty_like(q)
+
+
+direct_register_custom_op(
+    op_name="flash_attn_maxseqlen_wrapper",
+    op_func=flash_attn_maxseqlen_wrapper,
+    fake_impl=flash_attn_maxseqlen_wrapper_fake,
+)
+
+
+def vit_flash_attn_wrapper(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    batch_size: int,
+    is_rocm_aiter: bool,
+    fa_version: int | None,
+    scale: float | None = None,
+    cu_seqlens: torch.Tensor | None = None,
+    max_seqlen: torch.Tensor | None = None,
+) -> torch.Tensor:
+    return torch.ops.vllm.flash_attn_maxseqlen_wrapper(
+        q,
+        k,
+        v,
+        batch_size,
+        is_rocm_aiter,
+        fa_version,
+        scale,
+        cu_seqlens,
+        max_seqlen,
+    )
+
+
+def triton_attn_wrapper(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    batch_size: int,
+    scale: float | None = None,
+    cu_seqlens: torch.Tensor | None = None,
+    max_seqlen: torch.Tensor | None = None,
+) -> torch.Tensor:
+    from vllm.v1.attention.ops.triton_prefill_attention import context_attention_fwd
+
+    q_len = q.size(1)
+    if cu_seqlens is None:
+        cu_seqlens = torch.arange(
+            0, (batch_size + 1) * q_len, step=q_len, dtype=torch.int32, device=q.device
+        )
+    max_seqlen = q_len if max_seqlen is None else max_seqlen.item()
+
+    q, k, v = (einops.rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
+    output = torch.empty_like(q)
+    context_attention_fwd(
+        q,
+        k,
+        v,
+        output,
+        b_start_loc=cu_seqlens[:-1],
+        b_seq_len=cu_seqlens[1:] - cu_seqlens[:-1],
+        max_input_len=max_seqlen,
+        is_causal=False,
+        sliding_window_q=None,
+        sliding_window_k=None,
+        softmax_scale=scale,
+    )
+
+    context_layer = einops.rearrange(output, "(b s) h d -> b s h d", b=batch_size)
+    return context_layer
+
+
+def triton_attn_wrapper_fake(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    batch_size: int,
+    scale: float | None = None,
+    cu_seqlens: torch.Tensor | None = None,
+    max_seqlen: torch.Tensor | None = None,
+) -> torch.Tensor:
+    return torch.empty_like(q)
+
+
+direct_register_custom_op(
+    op_name="triton_attn_wrapper",
+    op_func=triton_attn_wrapper,
+    fake_impl=triton_attn_wrapper_fake,
+)
+
+
+def vit_triton_attn_wrapper(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    batch_size: int,
+    scale: float | None = None,
+    cu_seqlens: torch.Tensor | None = None,
+    max_seqlen: torch.Tensor | None = None,
+) -> torch.Tensor:
+    return torch.ops.vllm.triton_attn_wrapper(
+        q,
+        k,
+        v,
+        batch_size,
+        scale,
+        cu_seqlens,
+        max_seqlen,
+    )
+
+
+def apply_sdpa(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    scale: float | None = None,
+    enable_gqa: bool = False,
+) -> torch.Tensor:
+    """
+    Input shape:
+    (batch_size x seq_len x num_heads x head_size)
+    """
+    q, k, v = (einops.rearrange(x, "b s h d -> b h s d") for x in [q, k, v])
+    output = F.scaled_dot_product_attention(
+        q, k, v, dropout_p=0.0, scale=scale, enable_gqa=enable_gqa
+    )
+    output = einops.rearrange(output, "b h s d -> b s h d ")
+    return output
+
+
+# TODO: Once we have a torch 2.10, we can use tensor slices
+# so we won't need to wrap this in custom ops
+def torch_sdpa_wrapper(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    scale: float | None = None,
+    cu_seqlens: torch.Tensor | None = None,
+    enable_gqa: bool = False,
+) -> torch.Tensor:
+    # Never remove the contiguous logic for ROCm
+    # Without it, hallucinations occur with the backend
+    if current_platform.is_rocm():
+        q = q.contiguous()
+        k = k.contiguous()
+        v = v.contiguous()
+
+    if cu_seqlens is None:
+        return apply_sdpa(q, k, v, scale=scale, enable_gqa=enable_gqa)
+
+    outputs = []
+
+    lens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
+    q_chunks = torch.split(q, lens, dim=1)
+    k_chunks = torch.split(k, lens, dim=1)
+    v_chunks = torch.split(v, lens, dim=1)
+    for q_i, k_i, v_i in zip(q_chunks, k_chunks, v_chunks):
+        output_i = apply_sdpa(q_i, k_i, v_i, scale=scale, enable_gqa=enable_gqa)
+        outputs.append(output_i)
+    context_layer = torch.cat(outputs, dim=1)
+    return context_layer
+
+
+def torch_sdpa_wrapper_fake(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    scale: float | None,
+    cu_seqlens: torch.Tensor | None,
+    enable_gqa: bool = False,
+) -> torch.Tensor:
+    return torch.empty_like(q)
+
+
+direct_register_custom_op(
+    op_name="torch_sdpa_wrapper",
+    op_func=torch_sdpa_wrapper,
+    fake_impl=torch_sdpa_wrapper_fake,
+)
+
+
+def vit_torch_sdpa_wrapper(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    scale: float | None = None,
+    cu_seqlens: torch.Tensor | None = None,
+    enable_gqa: bool = False,
+) -> torch.Tensor:
+    return torch.ops.vllm.torch_sdpa_wrapper(
+        q, k, v, scale, cu_seqlens, enable_gqa=enable_gqa
+    )

vllm/v1/attention/selector.py

@@ -0,0 +1,152 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from functools import cache
+from typing import NamedTuple, cast, get_args
+
+import torch
+
+from vllm.config.cache import CacheDType
+from vllm.logger import init_logger
+from vllm.utils.import_utils import resolve_obj_by_qualname
+from vllm.v1.attention.backend import AttentionBackend, AttentionType
+from vllm.v1.attention.backends.registry import (
+    MAMBA_TYPE_TO_BACKEND_MAP,
+    MambaAttentionBackendEnum,
+)
+
+logger = init_logger(__name__)
+
+
+class AttentionSelectorConfig(NamedTuple):
+    head_size: int
+    dtype: torch.dtype
+    kv_cache_dtype: CacheDType | None
+    block_size: int | None
+    use_mla: bool = False
+    has_sink: bool = False
+    use_sparse: bool = False
+    use_mm_prefix: bool = False
+    use_per_head_quant_scales: bool = False
+    attn_type: str = AttentionType.DECODER
+
+    def __repr__(self):
+        return (
+            f"AttentionSelectorConfig(head_size={self.head_size}, "
+            f"dtype={self.dtype}, "
+            f"kv_cache_dtype={self.kv_cache_dtype}, "
+            f"block_size={self.block_size}, "
+            f"use_mla={self.use_mla}, "
+            f"has_sink={self.has_sink}, "
+            f"use_sparse={self.use_sparse}, "
+            f"use_mm_prefix={self.use_mm_prefix}, "
+            f"use_per_head_quant_scales={self.use_per_head_quant_scales}, "
+            f"attn_type={self.attn_type})"
+        )
+
+
+def get_attn_backend(
+    head_size: int,
+    dtype: torch.dtype,
+    kv_cache_dtype: str | None,
+    block_size: int | None,
+    use_mla: bool = False,
+    has_sink: bool = False,
+    use_sparse: bool = False,
+    use_mm_prefix: bool = False,
+    use_per_head_quant_scales: bool = False,
+    attn_type: str | None = None,
+    num_heads: int | None = None,
+) -> type[AttentionBackend]:
+    """Selects which attention backend to use and lazily imports it."""
+
+    if kv_cache_dtype is not None:
+        valid_cache_dtypes = get_args(CacheDType)
+        assert kv_cache_dtype in valid_cache_dtypes, (
+            f"Invalid kv_cache_dtype: {kv_cache_dtype}. "
+            f"Valid values are: {valid_cache_dtypes}"
+        )
+
+    from vllm.config import get_current_vllm_config
+
+    vllm_config = get_current_vllm_config()
+
+    attn_selector_config = AttentionSelectorConfig(
+        head_size=head_size,
+        dtype=dtype,
+        kv_cache_dtype=cast(CacheDType | None, kv_cache_dtype),
+        block_size=block_size,
+        use_mla=use_mla,
+        has_sink=has_sink,
+        use_sparse=use_sparse,
+        use_mm_prefix=use_mm_prefix,
+        use_per_head_quant_scales=use_per_head_quant_scales,
+        attn_type=attn_type or AttentionType.DECODER,
+    )
+
+    return _cached_get_attn_backend(
+        backend=vllm_config.attention_config.backend,
+        attn_selector_config=attn_selector_config,
+        num_heads=num_heads,
+    )
+
+
+@cache
+def _cached_get_attn_backend(
+    backend,
+    attn_selector_config: AttentionSelectorConfig,
+    num_heads: int | None = None,
+) -> type[AttentionBackend]:
+    from vllm.platforms import current_platform
+
+    attention_cls = current_platform.get_attn_backend_cls(
+        backend,
+        attn_selector_config=attn_selector_config,
+        num_heads=num_heads,
+    )
+    if not attention_cls:
+        raise ValueError(
+            f"Invalid attention backend for {current_platform.device_name}"
+        )
+    backend = resolve_obj_by_qualname(attention_cls)
+
+    # Adjust kv cache layout if the selected backend requires a specific one
+    required_layout = backend.get_required_kv_cache_layout()
+    if required_layout is not None:
+        from vllm.v1.attention.backends.utils import set_kv_cache_layout
+
+        set_kv_cache_layout(required_layout)
+        logger.info(
+            "Using %s KV cache layout for %s backend.",
+            required_layout,
+            backend.get_name(),
+        )
+
+    return backend
+
+
+def get_mamba_attn_backend(
+    mamba_type: str,
+) -> type[AttentionBackend]:
+    """Select which mamba attention backend to use and lazily import it."""
+    return _cached_get_mamba_attn_backend(mamba_type)
+
+
+@cache
+def _cached_get_mamba_attn_backend(
+    mamba_type: str,
+) -> type[AttentionBackend]:
+    assert mamba_type and isinstance(mamba_type, str)
+
+    selected_backend = None
+    try:
+        backend_name = MAMBA_TYPE_TO_BACKEND_MAP[mamba_type]
+        selected_backend = MambaAttentionBackendEnum[backend_name]
+    except KeyError as e:
+        raise ValueError(
+            f"Invalid mamba attention backend type: '{backend_name}'. Valid "
+            f"backends are: {list(MambaAttentionBackendEnum.__members__.keys())}"
+        ) from e
+
+    mamba_attn_backend = selected_backend.get_class()
+    return mamba_attn_backend