update

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