launcher/dcp2/baked/models/deepseek_v4_1/attention.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
DeepseekV4 MLA Attention Layer
"""

from abc import ABC, abstractmethod
from collections.abc import Callable
from functools import cached_property
from typing import TYPE_CHECKING, Any, ClassVar, cast

import regex as re
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import DeepseekV2Config, DeepseekV3Config

import vllm.envs as envs
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.model_executor.layers.fusion.quant_activation import QuantizedActivation
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    MergedColumnParallelLinear,
    ReplicatedLinear,
    RowParallelLinear,
)
from vllm.model_executor.layers.sparse_attn_indexer import SparseAttnIndexer
from vllm.models.common.ops import fused_q_kv_rmsnorm
from vllm.models.deepseek_v4_1.common.ops import (
    MXFP4_BLOCK_SIZE,
    fused_indexer_q_rope_quant,
    indexer_k_norm_rope_store,
)

if TYPE_CHECKING:
    from vllm.v1.attention.backends.mla.sparse_swa import (
        DeepseekSparseSWAMetadata,
    )

from vllm.config import (
    CacheConfig,
    VllmConfig,
    get_current_vllm_config,
)
from vllm.distributed import get_dcp_group, get_tensor_model_parallel_world_size
from vllm.forward_context import get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.models.utils import extract_layer_index
from vllm.models.deepseek_v4_1.common.rope import build_deepseek_v4_rope
from vllm.models.deepseek_v4_1.compressor import DeepseekCompressor
from vllm.triton_utils import tl, triton
from vllm.utils.multi_stream_utils import (
    execute_in_parallel,
    maybe_execute_in_parallel,
)
from vllm.v1.attention.backend import AttentionBackend, AttentionMetadata
from vllm.v1.attention.backends.mla.indexer import (
    DeepseekV4IndexerBackend,
    dsa_indexer_uses_fp4,
    get_max_prefill_buffer_size,
)
from vllm.v1.attention.backends.mla.sparse_swa import DeepseekV4SWACache
from vllm.v1.kv_cache_interface import (
    KVCacheSpec,
    MLAAttentionSpec,
    get_kv_quant_mode,
)

logger = init_logger(__name__)


def _replace_layer_index(prefix: str, layer_id: int) -> str:
    """Swap the layer index inside a ``...layers.<idx>...`` prefix."""
    new_prefix, n = re.subn(r"\.layers\.\d+\.", f".layers.{layer_id}.", prefix)
    assert n == 1, f"Cannot locate layer index in prefix {prefix}"
    return new_prefix


def _indexer_k_cache_head_dim(index_head_dim: int, use_fp4_kv: bool) -> int:
    """Per-token byte width of the paged indexer K cache row."""
    if use_fp4_kv:
        # MXFP4 stores two values per byte plus one UE8M0 byte per 32 values.
        # head_dim bytes = 64 packed values + 4 UE8M0 scales = 68.
        return index_head_dim // 2 + index_head_dim // MXFP4_BLOCK_SIZE
    # NOTE(yifan): FP8 indexer cache uses the same layout as V3.2:
    # head_dim bytes = 128 fp8 + 4 fp32 scale = 132.
    return index_head_dim + index_head_dim // 128 * 4


@triton.jit
def _fill_short_context_topk_indices(
    output,
    positions,
    TOP_K: tl.constexpr,
    COMPRESS_RATIO: tl.constexpr,
    PADDED_TOP_K: tl.constexpr,
):
    # small triton kernel that selects every candidate, -1 otherwise
    row = tl.program_id(0)
    offsets = tl.arange(0, PADDED_TOP_K)
    num_compressed = (tl.load(positions + row) + 1) // COMPRESS_RATIO
    tl.store(
        output + row * TOP_K + offsets,
        tl.where(offsets < num_compressed, offsets, -1),
        mask=offsets < TOP_K,
    )


def _resolve_dsv4_kv_cache_dtype(
    use_fp8_ds_mla_layout: bool,
    kv_cache_dtype: str,
    cache_config: CacheConfig | None,
) -> tuple[str, torch.dtype]:
    """Map ``(layout, --kv-cache-dtype)`` to ``(cache_dtype_str, torch_dtype)``.

    Both layouts are paged; they differ in the per-token block format. The
    ``fp8_ds_mla`` format is UE8M0 block-scaled fp8 packed as ``uint8`` (the
    canonical ``fp8_ds_mla`` string is written back onto ``cache_config`` so the
    page-size specs pick the 576B per-token slot). Plain-row backends store each
    token's KV row in its element dtype: bf16 or per-tensor FP8 E4M3.
    """
    if use_fp8_ds_mla_layout:
        # fp8_ds_mla block format: UE8M0 block-scaled fp8 packed as uint8.
        if kv_cache_dtype == "auto":
            kv_cache_dtype = "fp8"
        if not kv_cache_dtype.startswith("fp8"):
            raise ValueError(
                "DeepseekV4 fp8_ds_mla layout only supports fp8 "
                f"kv-cache, got {kv_cache_dtype}. Please set "
                "`--kv-cache-dtype fp8` or select a backend that supports "
                "bfloat16 KV cache."
            )
        if kv_cache_dtype != "fp8_ds_mla":
            if cache_config is not None:
                cache_config.cache_dtype = "fp8_ds_mla"
            kv_cache_dtype = "fp8_ds_mla"
            logger.info_once("Using DeepSeek's fp8_ds_mla KV cache format.")
        return kv_cache_dtype, torch.uint8

    # Plain bf16 / per-tensor fp8 KV row (FlashInfer).
    if kv_cache_dtype.startswith("fp8"):
        return kv_cache_dtype, torch.float8_e4m3fn
    # auto / bfloat16 -> plain bf16 KV row.
    return kv_cache_dtype, torch.bfloat16


class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
    """DeepseekV4 MLA attention layer.

    The platform-specific sparse-MLA forward (``forward_mqa`` /
    ``get_padded_num_q_heads`` / ``_o_proj`` / ``backend_cls``) is provided by a
    subclass — ``DeepseekV4FlashMLAAttention`` /
    ``DeepseekV4FlashInferSM120Attention`` /
    ``DeepseekV4FlashInferMLAAttention`` (CUDA) or
    ``DeepseekV41ROCMAiterMLAAttention`` (ROCm) — selected by the platform-specific
    deepseek_v4_1 model module. The base is never instantiated directly.
    """

    # Provided by the platform subclass.
    backend_cls: ClassVar[type[AttentionBackend]]
    # Backend for the SWA cache layer; None uses the default SWA backend.
    swa_backend_cls: ClassVar[type[AttentionBackend] | None] = None
    # KV-cache per-token block format (both layouts are paged). True (default)
    # = fp8_ds_mla (UE8M0 block-scaled fp8 packed as uint8); False = plain
    # bf16 / per-tensor fp8 KV row. Backends can override the instance hook when
    # a single attention class dispatches across arch-specific layouts.
    use_fp8_ds_mla_layout: ClassVar[bool] = True
    # Prefill is processed in fixed-size chunks; this bounds the bf16 kv-gather
    # workspace allocated in _forward_prefill and is also read by the dummy-run
    # path to pre-reserve that workspace.
    PREFILL_CHUNK_SIZE: ClassVar[int] = 4
    # DCP contract, mirroring AttentionImpl (v1/attention/backend.py): the
    # V4.1 layer runs its own forward (no AttentionImpl), so
    # check_attention_cp_compatibility reads these from the layer. A platform
    # subclass that implements the per-rank partial attention + LSE combine
    # sets can_return_lse_for_decode = True (SM120 FlashInfer).
    can_return_lse_for_decode: ClassVar[bool] = False
    supports_mtp_with_cp_non_trivial_interleave_size: ClassVar[bool] = False

    @classmethod
    @abstractmethod
    def get_padded_num_q_heads(cls, num_heads: int) -> int:
        """Q head count the q/output buffers are allocated at.

        The layer allocates the q/output buffers at
        ``[N, get_padded_num_q_heads(n_local_heads), head_dim]``. Must satisfy
        ``result >= num_heads``. Backends with no padding constraint return
        ``num_heads``.
        """
        raise NotImplementedError

    @abstractmethod
    def forward_mqa(
        self,
        q: torch.Tensor,
        kv: torch.Tensor,
        positions: torch.Tensor,
        output: torch.Tensor,
    ) -> None:
        """Platform-specific sparse MLA forward; writes attention into ``output``."""
        raise NotImplementedError

    @abstractmethod
    def _o_proj(self, o: torch.Tensor, positions: torch.Tensor) -> torch.Tensor:
        """Inverse-RoPE + wo_a + wo_b output projection (platform-specific)."""
        raise NotImplementedError

    def _uses_fp8_ds_mla_layout(self) -> bool:
        """Return whether this instance stores fp8 KV in fp8_ds_mla layout."""
        return self.use_fp8_ds_mla_layout

    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str,
        topk_indices_buffer: torch.Tensor | None = None,
        aux_stream_list: list[torch.cuda.Stream] | None = None,
        candidate_block_buffer: torch.Tensor | None = None,
    ) -> None:
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        cache_config = vllm_config.cache_config
        tp_size = get_tensor_model_parallel_world_size()
        layer_id = extract_layer_index(prefix)
        self.layer_id = layer_id

        self.prefix = prefix  # Alias for compatibility with compressor
        self.hidden_size = config.hidden_size
        self.n_heads = config.num_attention_heads
        assert self.n_heads % tp_size == 0
        self.n_local_heads = self.n_heads // tp_size
        # Decode context parallelism: compressed/indexer caches are sharded over
        # compressed states across the DCP group (a subset of the TP ranks),
        # SWA caches are replicated; the platform subclass gathers queries and
        # combines per-rank partial attention (DCP2-PLAN.md).
        parallel_config = vllm_config.parallel_config
        self.dcp_world_size = parallel_config.decode_context_parallel_size
        self.dcp_rank = get_dcp_group().rank_in_group if self.dcp_world_size > 1 else 0
        self.cp_kv_cache_interleave_size = parallel_config.cp_kv_cache_interleave_size
        self.need_to_return_lse_for_decode = (
            self.dcp_world_size > 1 and self.can_return_lse_for_decode
        )
        if self.dcp_world_size > 1 and self.cp_kv_cache_interleave_size != 1:
            raise NotImplementedError(
                "DeepSeek-V4.1 DCP supports cp_kv_cache_interleave_size=1 only "
                f"(got {self.cp_kv_cache_interleave_size})."
            )
        self.q_lora_rank = config.q_lora_rank
        self.o_lora_rank = config.o_lora_rank
        self.head_dim = config.head_dim
        self.rope_head_dim = config.qk_rope_head_dim
        self.nope_head_dim = self.head_dim - self.rope_head_dim
        self.n_groups = config.o_groups
        self.n_local_groups = self.n_groups // tp_size
        self.window_size = config.sliding_window
        # Vision variant: image spans are visible bidirectionally, widening
        # prefill SWA index rows by up to max_image_tokens columns.
        self.max_image_tokens = (
            getattr(config, "vision_max_n_token", 0)
            if getattr(config, "vision_n_layers", 0) > 0
            else 0
        )
        # ---- v4.1 sparse-attention topology ----
        # compress_ratios has one entry per layer (MTP layers included):
        # 0 = pure sliding window, 1 = full-length compressed cache,
        # 2 = ratio-2 compressed. Compressors and compressed-KV caches live
        # only on ``kv_source_layer_ids``; indexers only on
        # ``index_source_layer_ids``. Consumers reuse the most recently
        # published source below them.
        compress_ratios = getattr(config, "compress_ratios", None)
        if compress_ratios is not None and layer_id < len(compress_ratios):
            self.compress_ratio = int(compress_ratios[layer_id])
        else:
            # MTP layers past the configured list are pure sliding-window.
            self.compress_ratio = 0
        if self.compress_ratio not in (0, 1, 2):
            raise ValueError(
                f"DeepSeek V4.1 layer {layer_id} has compress_ratio="
                f"{self.compress_ratio}; only 0 (sliding window), 1 and 2 are "
                "supported."
            )
        self.kv_source_layers = tuple(
            getattr(config, "kv_source_layer_ids", None) or ()
        )
        self.index_source_layers = tuple(
            getattr(config, "index_source_layer_ids", None) or ()
        )
        self.candidate_source_layer = getattr(config, "candidate_source_layer_id", -1)
        self.candidate_topk_blocks = getattr(config, "candidate_topk_blocks", 0)
        self.candidate_block_size = getattr(config, "candidate_block_size", 0)

        is_backbone = layer_id < config.num_hidden_layers
        self.is_kv_source = is_backbone and layer_id in self.kv_source_layers
        self.is_index_source = is_backbone and layer_id in self.index_source_layers
        if self.compress_ratio > 0:
            if not self.kv_source_layers or not self.index_source_layers:
                raise ValueError(
                    "DeepSeek V4.1 requires kv_source_layer_ids / "
                    "index_source_layer_ids in the config for compressed "
                    f"layers (layer {layer_id} has "
                    f"compress_ratio={self.compress_ratio})."
                )
            self.kv_source_layer_id = max(
                s for s in self.kv_source_layers if s <= layer_id
            )
            self.index_source_layer_id = max(
                s for s in self.index_source_layers if s <= layer_id
            )
        else:
            self.kv_source_layer_id = None
            self.index_source_layer_id = None
        self.eps = config.rms_norm_eps
        self.scale = self.head_dim**-0.5

        # Padded Q head count is dictated by the platform subclass.
        self.padded_heads = self.get_padded_num_q_heads(self.n_local_heads)
        # Sink padded to the same head count, initialized to -inf (no sink
        # effect). Weight loading fills the first n_local_heads slots.
        self.attn_sink = nn.Parameter(
            torch.full((self.padded_heads,), -float("inf"), dtype=torch.float32),
            requires_grad=False,
        )

        self.fused_wqa_wkv = MergedColumnParallelLinear(
            self.hidden_size,
            [self.q_lora_rank, self.head_dim],
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.fused_wqa_wkv",
            disable_tp=True,  # fused ReplicatedLinear
        )
        self.q_norm = RMSNorm(self.q_lora_rank, self.eps)
        self.wq_b = ColumnParallelLinear(
            self.q_lora_rank,
            self.n_heads * self.head_dim,
            bias=False,
            quant_config=quant_config,
            return_bias=False,
            prefix=f"{prefix}.wq_b",
        )

        self.kv_norm = RMSNorm(self.head_dim, self.eps)
        self.wo_a = ColumnParallelLinear(
            self.n_heads * self.head_dim // self.n_groups,
            self.n_groups * self.o_lora_rank,
            bias=False,
            quant_config=quant_config,
            return_bias=False,
            prefix=f"{prefix}.wo_a",
        )
        self.wo_a.is_bmm = True
        self.wo_a.bmm_batch_size = self.n_local_groups
        self._o_proj_block_size = (
            32 if getattr(self.wo_a, "weight_block_size", None) == [1, 32] else 128
        )
        self.wo_b = RowParallelLinear(
            self.n_groups * self.o_lora_rank,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            return_bias=False,
            prefix=f"{prefix}.wo_b",
        )

        # Initialize rotary embedding before the indexer/compressor consume it.
        self.rotary_emb = build_deepseek_v4_rope(
            config,
            head_dim=self.head_dim,
            rope_head_dim=self.rope_head_dim,
            max_position_embeddings=config.max_position_embeddings,
            compress_ratio=self.compress_ratio,
        )
        self.indexer_rotary_emb = self.rotary_emb
        self.topk_indices_buffer = topk_indices_buffer
        self.candidate_block_buffer = candidate_block_buffer

        # Register with compilation context for metadata lookup. Done before
        # indexer/compressor creation so consumers can resolve their source
        # layers through it.
        compilation_config = vllm_config.compilation_config
        if prefix and prefix in compilation_config.static_forward_context:
            raise ValueError(f"Duplicate layer name: {prefix}")
        if prefix:
            compilation_config.static_forward_context[prefix] = self
        self.kv_cache = torch.tensor([])
        self._static_forward_context = compilation_config.static_forward_context

        self.indexer = None
        if self.is_index_source:
            index_k_cache: DeepseekV4IndexerCache | None
            # Index K cache: owned only by kv-source layers (their indexer has
            # wk/k_norm); non-owning index sources share the K cache of the
            # latest kv source below them (its indexer produces the keys).
            if self.is_kv_source:
                index_k_cache = DeepseekV4IndexerCache(
                    head_dim=_indexer_k_cache_head_dim(
                        config.index_head_dim, dsa_indexer_uses_fp4(vllm_config)
                    ),
                    dtype=torch.uint8,
                    prefix=f"{prefix}.indexer.k_cache",
                    cache_config=cache_config,
                    compress_ratio=self.compress_ratio,
                )
            else:
                assert self.kv_source_layer_id is not None
                k_cache_prefix = (
                    f"{_replace_layer_index(prefix, self.kv_source_layer_id)}"
                    ".indexer.k_cache"
                )
                index_k_cache = self._static_forward_context.get(k_cache_prefix)
                if index_k_cache is None:
                    raise NotImplementedError(
                        f"Indexer K cache source {k_cache_prefix} not found on "
                        "this rank; PP splits inside a v4.1 kv-sharing group "
                        "are not supported."
                    )
            is_candidate_source = layer_id == self.candidate_source_layer
            uses_candidates = 0 <= self.candidate_source_layer < layer_id
            self.indexer = DeepseekV4Indexer(
                vllm_config,
                config=config,
                hidden_size=self.hidden_size,
                q_lora_rank=self.q_lora_rank,
                quant_config=quant_config,
                cache_config=cache_config,
                topk_indices_buffer=topk_indices_buffer,
                compress_ratio=self.compress_ratio,
                prefix=f"{prefix}.indexer",
                owns_k=self.is_kv_source,
                k_cache=index_k_cache,
                main_head_dim=self.head_dim,
                candidate_block_buffer=(
                    candidate_block_buffer
                    if (is_candidate_source or uses_candidates)
                    else None
                ),
                candidate_block_size=self.candidate_block_size,
                candidate_write=is_candidate_source,
            )

        self._prepare_and_attn_fn = self._prepare_and_attn
        if not vllm_config.use_v2_model_runner:
            # MRV1's piecewise capture only tolerates the wide eager region: with
            # the narrow one the attention input preparation stays in the captured
            # graph and MRV1 produces garbage (#51430).
            self._prepare_and_attn_fn = self._prepare_and_attn_eager

        # Will be None on ROCm for now.
        self.aux_stream_list = aux_stream_list
        # [0]: GEMM start / post-GEMM event0. [1..3]: GEMM done events;
        # [1] doubles as post-GEMM event1. Reuse is safe: GEMM fully joins
        # before post-GEMM starts.
        self.ln_events = [torch.cuda.Event() for _ in range(4)]

        assert cache_config is not None, "DeepseekV4 attention requires cache_config"
        # ---- Attention / KV-cache setup ----
        self.max_num_batched_tokens = (
            vllm_config.scheduler_config.max_num_batched_tokens
        )
        self.max_model_len = vllm_config.model_config.max_model_len

        # Resolve the kv-cache dtype from this backend's block format. The same
        # resolution drives the SWA cache tensor dtype below.
        self.kv_cache_dtype, self.kv_cache_torch_dtype = _resolve_dsv4_kv_cache_dtype(
            self._uses_fp8_ds_mla_layout(), cache_config.cache_dtype, cache_config
        )

        self.swa_cache_layer = DeepseekV4SWACache(
            head_dim=self.head_dim,
            window_size=self.window_size,
            dtype=self.kv_cache_torch_dtype,
            prefix=f"{prefix}.swa_cache",
            cache_config=cache_config,
            backend_cls=self.swa_backend_cls,
            # Tech2Wild/Kai 2026-09-10: page size is a property of the SWA decode
            # kernel (FlashInfer SM120 sparse-MLA kernels are 64-token pages).
            block_size=getattr(self.swa_backend_cls, "get_swa_block_size", lambda: 32)(),
        )

        # The attention layer itself was already registered with the
        # compilation context above (before indexer/compressor creation).

        # Compressors live only on kv-source layers; consumers read the
        # source's compressed cache through the forward context.
        self.compressor = None
        if self.is_kv_source:
            self.compressor = DeepseekCompressor(
                vllm_config=vllm_config,
                compress_ratio=self.compress_ratio,
                hidden_size=self.hidden_size,
                head_dim=self.head_dim,
                rotate=True,
                prefix=f"{prefix}.compressor",
                k_cache_prefix=self.prefix,
            )
        # Prefix of the attention layer owning this layer's compressed KV
        # cache (self for kv sources).
        if self.compress_ratio > 0:
            assert self.kv_source_layer_id is not None
            self.compressed_cache_prefix: str | None = _replace_layer_index(
                prefix, self.kv_source_layer_id
            )
            if (
                not self.is_kv_source
                and self.compressed_cache_prefix not in self._static_forward_context
            ):
                raise NotImplementedError(
                    f"Compressed-KV source {self.compressed_cache_prefix} not "
                    "found on this rank; PP splits inside a v4.1 kv-sharing "
                    "group are not supported."
                )
        else:
            self.compressed_cache_prefix = None

        if vllm_config.kernel_config.enable_jit_warmup:
            from vllm.v1.attention.backends.mla.sparse_swa import (
                _COMPUTE_PREFILL_METADATA_KERNEL,
                _COMPUTE_SWA_INDICES_AND_LENS_KERNEL,
            )

            _COMPUTE_PREFILL_METADATA_KERNEL.register_warmup()
            _COMPUTE_SWA_INDICES_AND_LENS_KERNEL.register_warmup(
                window_size=self.window_size,
                block_size=self.swa_cache_layer.block_size,
                max_image_tokens=self.max_image_tokens,
            )

            if self.compress_ratio > 1:
                from vllm.v1.attention.backends.mla.compressor_utils import (
                    _COMPRESSED_SLOT_MAPPING_KERNEL,
                )

                _COMPRESSED_SLOT_MAPPING_KERNEL.register_warmup()

            if self.indexer is not None:
                from vllm.v1.attention.backends.mla.indexer import (
                    _BUILD_PREFILL_CHUNK_METADATA_KERNEL,
                    _PREPARE_UNIFORM_DECODE_KERNEL,
                )

                _PREPARE_UNIFORM_DECODE_KERNEL.register_warmup()
                _BUILD_PREFILL_CHUNK_METADATA_KERNEL.register_warmup()

            spec_config = vllm_config.speculative_config
            if spec_config is not None and spec_config.use_dspark():
                from vllm.v1.attention.backends.mla.sparse_swa import (
                    _COMPUTE_DSPARK_NONCAUSAL_SWA_INDICES_KERNEL,
                )

                _COMPUTE_DSPARK_NONCAUSAL_SWA_INDICES_KERNEL.register_warmup(
                    window_size=self.window_size,
                    num_speculative_tokens=spec_config.num_speculative_tokens,
                    block_size=self.swa_cache_layer.block_size,
                )

            if self.backend_cls.get_name() in (
                "FLASHMLA_SPARSE_DSV41",
                "ROCM_FLASHMLA_SPARSE_DSV4",
            ):
                from vllm.models.deepseek_v4_1.common.ops.cache_utils import (
                    _COMBINE_TOPK_SWA_INDICES_KERNEL,
                )

                _COMBINE_TOPK_SWA_INDICES_KERNEL.register_warmup()

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        llama_4_scaling: torch.Tensor | None = None,
    ) -> torch.Tensor:
        # Pre-allocate attention output with FlashMLA-padded head count.
        # The op writes into `o_padded`; we slice to n_local_heads after.
        num_tokens = hidden_states.shape[0]
        o_padded = torch.empty(
            (num_tokens, self.padded_heads, self.head_dim),
            dtype=hidden_states.dtype,
            device=hidden_states.device,
        )

        # Keep the attention input preparation in the captured graph. Only the
        # sparse indexer and MLA attention run in the eager break below.
        qr_kv, kv_score, indexer_weights = self._run_parallel_input_projections(
            hidden_states
        )
        qr, qr_scale, kv = self._split_qkv_and_norm(qr_kv)

        self._prepare_and_attn_fn(
            hidden_states,
            qr,
            kv,
            qr_scale,
            kv_score,
            indexer_weights,
            positions,
            o_padded,
        )
        o = o_padded[:, : self.n_local_heads, :]

        # Inverse-RoPE + wo_a + wo_b output projection (platform-specific).
        return self._o_proj(o, positions)

    @cached_property
    def _can_fuse_query_quant(self) -> bool:
        from vllm.models.deepseek_v4_1.common.ops.query_quant import (
            can_fuse_query_quant,
        )

        linears = [self.wq_b]
        if self.indexer is not None:
            linears.append(self.indexer.wq_b)
        return can_fuse_query_quant(linears)

    def _split_qkv_and_norm(
        self, qr_kv: torch.Tensor
    ) -> tuple[torch.Tensor | QuantizedActivation, torch.Tensor | None, torch.Tensor]:
        """Split the fused q-lora / kv projection and RMSNorm both halves.

        Compatible MXFP8 projections share the quantized Q and scales;
        other projection backends consume the normalized Q directly.
        """
        qr, kv = qr_kv.split([self.q_lora_rank, self.head_dim], dim=-1)
        if self.q_lora_rank % 32 == 0 and self._can_fuse_query_quant:
            from vllm.models.deepseek_v4_1.common.ops.query_quant import (
                fused_q_kv_rmsnorm_quant,
            )

            qr_quant, kv = fused_q_kv_rmsnorm_quant(
                qr,
                kv,
                self.q_norm.weight.data,
                self.kv_norm.weight.data,
                self.eps,
            )
            return qr_quant, None, kv
        qr, kv = fused_q_kv_rmsnorm(
            qr,
            kv,
            self.q_norm.weight.data,
            self.kv_norm.weight.data,
            self.eps,
        )
        return qr, None, kv

    @eager_break_during_capture
    def _prepare_and_attn_eager(
        self,
        hidden_states: torch.Tensor,
        qr: torch.Tensor | QuantizedActivation,
        kv: torch.Tensor,
        qr_scale: torch.Tensor | None,
        kv_score: torch.Tensor,
        indexer_weights: torch.Tensor,
        positions: torch.Tensor,
        o_padded: torch.Tensor,
    ) -> None:
        """Wide eager region: the whole of ``_prepare_and_attn`` runs eagerly.

        The nested ``_sparse_indexer_and_attn`` break runs inline, since
        ``add_eager`` clears ``_capturing`` before invoking this.
        """
        self._prepare_and_attn(
            hidden_states,
            qr,
            kv,
            qr_scale,
            kv_score,
            indexer_weights,
            positions,
            o_padded,
        )

    def _prepare_and_attn(
        self,
        hidden_states: torch.Tensor,
        qr: torch.Tensor | QuantizedActivation,
        kv: torch.Tensor,
        qr_scale: torch.Tensor | None,
        kv_score: torch.Tensor,
        indexer_weights: torch.Tensor,
        positions: torch.Tensor,
        o_padded: torch.Tensor,
    ) -> None:
        """Attention input preparation followed by the sparse indexer and MLA.

        Only the latter runs in the eager break.

        Q/SWA preparation overlaps state saving and compression. Once the
        latent is ready, main-cache insertion overlaps indexer preparation;
        both cache writes finish before sparse attention reads either cache.
        """
        attn_metadata = get_forward_context().attn_metadata
        indexer = self.indexer
        compressor = self.compressor
        aux_streams = self.aux_stream_list

        def project_query_and_cache_kv() -> torch.Tensor:
            q = self._wq_b_proj(qr, qr_scale).view(
                -1, self.n_local_heads, self.head_dim
            )
            return self._fused_qnorm_rope_kv_insert(q, kv, positions, attn_metadata)

        index_q: torch.Tensor | None = None
        index_q_scale: torch.Tensor | None = None
        index_weights_out: torch.Tensor | None = None
        latent: torch.Tensor | None = None
        aux_stream = aux_streams[0] if aux_streams is not None else None

        if compressor is not None:
            # Q projection / KV insertion on the default stream overlaps the
            # compressor on aux stream 0 (sequential on ROCm).
            q, latent = maybe_execute_in_parallel(
                project_query_and_cache_kv,
                lambda: compressor(kv_score, positions),
                self.ln_events[0],
                self.ln_events[1],
                aux_stream,
            )
        else:
            q = project_query_and_cache_kv()

        def prepare_indexer():
            if indexer is None:
                return None, None, None
            return indexer(
                qr,
                latent,
                indexer_weights,
                positions,
                self.indexer_rotary_emb,
                qr_scale,
            )

        if compressor is not None:
            indexer_result, _ = maybe_execute_in_parallel(
                prepare_indexer,
                lambda: compressor.insert_cache(latent, positions, self.rotary_emb),
                self.ln_events[0],
                self.ln_events[1],
                aux_stream,
            )
        else:
            indexer_result = prepare_indexer()
        index_q, index_q_scale, index_weights_out = indexer_result

        self._sparse_indexer_and_attn(
            hidden_states,
            index_q,
            index_q_scale,
            index_weights_out,
            q,
            kv,
            positions,
            o_padded,
        )

    def _fused_wqa_wkv_gemm(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # Override point: the ROCm layer preshuffles this weight in place, so
        # it cannot go through fused_wqa_wkv directly.
        # MergedColumnParallelLinear returns (output, bias); bias is None.
        qr_kv, _ = self.fused_wqa_wkv(hidden_states)
        return qr_kv

    def _wq_b_proj(
        self,
        qr: torch.Tensor | QuantizedActivation,
        qr_scale: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """Project normalized Q, bypassing quantization when already fused."""
        assert qr_scale is None, "ROCm-only path"
        return self.wq_b(qr)

    def _run_parallel_input_projections(
        self, hidden_states: torch.Tensor
    ) -> tuple[
        torch.Tensor,
        torch.Tensor | None,
        torch.Tensor | None,
    ]:
        aux_streams = self.aux_stream_list
        if aux_streams is not None:
            aux_streams = aux_streams[:2]

        # fused_wqa_wkv (heaviest) on default; the two lighter input GEMMs on
        # aux streams 0/1 when their owning module exists. ln_events[0] is the
        # fan-out start event; ln_events[1..2] are per-aux done events. The
        # v4.1 indexer derives K from the kv-source compressor's latent, so
        # unlike v4.0 there is no indexer K GEMM over hidden_states here.
        aux_fns: list[Callable[[], Any] | None] = [None, None]

        if self.compressor is not None:
            # Local ref so the closure keeps a non-None type for mypy.
            compressor = self.compressor

            def compressor_kv_score() -> torch.Tensor:
                return torch.mm(
                    hidden_states,
                    compressor.fused_wkv_wgate.weight.T,
                    out_dtype=torch.float32,
                )

            aux_fns[0] = compressor_kv_score

        if self.indexer is not None:
            indexer = self.indexer

            def indexer_weights_proj() -> torch.Tensor:
                # ReplicatedLinear returns (output, bias); bias is None.
                weights, _ = indexer.weights_proj(hidden_states)
                return weights

            aux_fns[1] = indexer_weights_proj

        qr_kv, (kv_score, indexer_weights) = execute_in_parallel(
            lambda: self._fused_wqa_wkv_gemm(hidden_states),
            aux_fns,
            self.ln_events[0],
            self.ln_events[1:3],
            aux_streams,
            enable=hidden_states.shape[0]
            <= envs.VLLM_MULTI_STREAM_GEMM_TOKEN_THRESHOLD,
        )

        return qr_kv, kv_score, indexer_weights

    @eager_break_during_capture
    def _sparse_indexer_and_attn(
        self,
        hidden_states: torch.Tensor,
        index_q: torch.Tensor | None,
        index_q_scale: torch.Tensor | None,
        index_weights: torch.Tensor | None,
        q: torch.Tensor,
        kv: torch.Tensor,
        positions: torch.Tensor,
        out: torch.Tensor,
    ) -> None:
        if self.indexer is not None and index_q is not None:
            assert index_weights is not None
            q_quant = (index_q, index_q_scale) if index_q_scale is not None else index_q
            self.indexer.indexer_op(
                hidden_states,
                q_quant,
                None,
                index_weights,
            )

        # MLA attention writes into the pre-allocated `out` buffer
        # ([num_tokens, padded_heads, head_dim]).
        self.forward_mqa(q, kv, positions, out)

    def _fused_qnorm_rope_kv_insert(
        self,
        q: torch.Tensor,
        kv: torch.Tensor,
        positions: torch.Tensor,
        attn_metadata: (
            dict[str, AttentionMetadata] | list[dict[str, AttentionMetadata]] | None
        ),
    ) -> torch.Tensor:
        if not isinstance(attn_metadata, dict):
            # Profile run: kernel doesn't fire; produce a padded tensor so
            # downstream FlashMLA gets the right shape.
            if self.n_local_heads < self.padded_heads:
                return F.pad(
                    q,
                    (0, 0, 0, self.padded_heads - self.n_local_heads),
                    value=0.0,
                )
            return q

        swa_metadata = cast(
            "DeepseekSparseSWAMetadata | None",
            attn_metadata.get(self.swa_cache_layer.prefix),
        )
        assert swa_metadata is not None

        swa_kv_cache = self.swa_cache_layer.kv_cache
        # The fused insert ops require int64 position_ids; the runner's positions
        # buffer is already int64, so no cast is needed.
        assert positions.dtype == torch.int64
        cos_sin_cache = self.rotary_emb.cos_sin_cache
        cache_dtype = swa_kv_cache.dtype

        # kv is unchanged; attention reads kv solely via swa_kv_cache.
        if cache_dtype == torch.uint8:
            # fp8_ds_mla UE8M0 paged path. Horizontally fused:
            #   Q side: GPT-J RoPE, zero-filling the padding head slots; the
            #           kernel allocates and returns the padded q tensor.
            #   KV side: GPT-J RoPE + UE8M0 FP8 quant + paged cache insert.
            swa_kv_cache_2d = swa_kv_cache.view(swa_kv_cache.shape[0], -1)
            return torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
                q,
                kv,
                swa_kv_cache_2d,
                swa_metadata.slot_mapping,
                positions,
                cos_sin_cache,
                self.padded_heads,
                self.eps,
                swa_metadata.block_size,
                False,
            )

        # Plain-row path: the [num_blocks, block_size, 512] cache stores the KV
        # row in its element dtype (no Q padding). bf16 rewrites q in place;
        # per-tensor fp8 writes a separately-allocated fp8 q and quantizes the
        # KV row.
        block_size = swa_metadata.block_size
        assert swa_kv_cache.shape[1:] == (block_size, self.head_dim)
        swa_kv_cache_3d = swa_kv_cache
        if cache_dtype == torch.bfloat16:
            torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert(
                q,
                kv,
                swa_kv_cache_3d,
                swa_metadata.slot_mapping,
                positions,
                cos_sin_cache,
                self.eps,
                block_size,
                False,
            )
            return q

        # per-tensor fp8 (torch.float8_e4m3fn)
        q_fp8 = torch.empty_like(q, dtype=torch.float8_e4m3fn)
        torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert(
            q,
            kv,
            q_fp8,
            swa_kv_cache_3d,
            swa_metadata.slot_mapping,
            positions,
            cos_sin_cache,
            self._flashinfer_fp8_kv_scale,
            self._flashinfer_fp8_q_scale_inv,
            self.eps,
            block_size,
            False,
        )
        return q_fp8

    def bind_kv_cache(self, kv_cache: torch.Tensor) -> None:
        # [B, H=1, N, C] -> [B, N, C]
        self.kv_cache = kv_cache.squeeze(1)

    def get_attn_backend(self) -> type[AttentionBackend]:
        return self.backend_cls

    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None:
        # Only kv-source layers own a compressed-KV cache; consumers read the
        # source's cache through the forward context, and cr==0 layers are
        # pure SWA. The SWA cache is allocated separately as
        # DeepseekV4SWACache.
        if not self.is_kv_source:
            return None
        # fp8_ds_mla is a UE8M0 block-scaled uint8 layout and needs 576B
        # alignment; plain bf16 / per-tensor fp8 rows use natural element-size
        # pages.
        uses_fp8_ds_mla_layout = self.kv_cache_dtype == "fp8_ds_mla"
        # Tech2Wild/Kai 2026-09-10: SM120 FlashInfer sparse-MLA kernels are
        # instantiated for 64-STATE compressed pages, so a kv-source layer's page
        # is 64 * compress_ratio tokens there (ratio 1 -> 64, ratio 2 -> 128).
        _hook = getattr(self.backend_cls, "get_compressed_block_size", None)
        _bs = _hook(self.compress_ratio) if _hook is not None else None
        return MLAAttentionSpec(
            block_size=_bs or vllm_config.cache_config.block_size,
            num_kv_heads=1,
            head_size=self.head_dim,
            dtype=torch.uint8 if uses_fp8_ds_mla_layout else self.kv_cache_torch_dtype,
            tokens_per_state=self.compress_ratio,
            cache_dtype_str=self.kv_cache_dtype,
            alignment=576 if uses_fp8_ds_mla_layout else 512,
            model_version="deepseek_v4",
            kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype),
            # DeepseekV4: 448B NoPE + 128B RoPE + 8B fp8 scale = 584B per token;
            # head_size stays semantic (512).
            state_content_bytes=584 if uses_fp8_ds_mla_layout else None,
        )

    def _compressed_kv_cache(self) -> torch.Tensor:
        """The compressed-KV cache tensor of this layer's kv source (own
        cache for kv-source layers)."""
        if self.is_kv_source:
            return self.kv_cache
        assert self.compressed_cache_prefix is not None
        source = self._static_forward_context[self.compressed_cache_prefix]
        return source.kv_cache


def _dsv41_indexer_sm12x() -> bool:
    from vllm.platforms import current_platform

    return current_platform.is_cuda() and current_platform.is_device_capability_family(120)


class DeepseekV4IndexerSM12xBackend(DeepseekV4IndexerBackend):
    """Tech2Wild 2026-09-10: SM12x indexer pages hold 64 states (64 tokens at
    compress ratio 1, 128 at ratio 2), so both sizes are kernel block sizes."""

    @staticmethod
    def get_supported_kernel_block_sizes() -> list[int]:
        return [128, 64]


class DeepseekV4IndexerCache(torch.nn.Module, AttentionLayerBase):
    def __init__(
        self,
        head_dim: int,
        dtype: torch.dtype,
        prefix: str,
        cache_config: CacheConfig,
        compress_ratio: int = 1,
    ):
        super().__init__()
        self.kv_cache = torch.tensor([])
        self.head_dim = head_dim
        self.prefix = prefix
        self.cache_config = cache_config
        self.dtype = dtype
        self.compress_ratio = compress_ratio
        compilation_config = get_current_vllm_config().compilation_config
        if prefix in compilation_config.static_forward_context:
            raise ValueError(f"Duplicate layer name: {prefix}")
        compilation_config.static_forward_context[prefix] = self

    def bind_kv_cache(self, kv_cache: torch.Tensor) -> None:
        # [B, H=1, N, C] -> [B, N, C]
        self.kv_cache = kv_cache.squeeze(1)

    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
        # head_dim already carries the fp8 scale padding
        # tokens_per_state=1 for V3.2, >1 for DeepseekV4; same cache layout.
        uses_fp8_ds_mla_layout = vllm_config.cache_config.cache_dtype == "fp8_ds_mla"
        # Tech2Wild 2026-09-10 (SM12x): DeepGEMM's paged MQA logits (indexer decode)
        # takes 32 or 64 states per block, and the indexer metadata builder passes
        # this spec's num_states, so pages hold 64 states = 64 * compress_ratio tokens.
        block_size = self.cache_config.block_size
        if _dsv41_indexer_sm12x():
            block_size = 64 * max(1, self.compress_ratio)
        return MLAAttentionSpec(
            block_size=block_size,
            num_kv_heads=1,
            head_size=self.head_dim,
            dtype=self.dtype,
            tokens_per_state=self.compress_ratio,
            # 576B for FlashMLA packing; 512B for FlashInfer sparse (#44577).
            alignment=576 if uses_fp8_ds_mla_layout else 512,
        )

    def forward(self): ...

    def get_attn_backend(self) -> type[AttentionBackend]:
        if _dsv41_indexer_sm12x():
            return DeepseekV4IndexerSM12xBackend
        return DeepseekV4IndexerBackend


class DeepseekV4Indexer(nn.Module):
    """DeepSeek V4.1 sparse-attention indexer.

    Exists only on ``index_source_layer_ids``; consumers reuse the topk
    indices it publishes into the shared ``topk_indices_buffer``. Unlike v4.0
    the index key is derived from the kv-source layer's compressor latent
    (``k = k_norm(wk(latent))``, ``owns_k``) instead of an indexer-local
    compressor over hidden states, so there is no hidden-state K GEMM here.
    Non-owning index sources share the kv source's paged K cache.

    Two-level candidate filtering: the indexer at ``candidate_source_layer_id``
    additionally publishes the top ``candidate_topk_blocks`` blocks of
    ``candidate_block_size`` compressed positions (``candidate_write``);
    later indexers mask their scores to those blocks before their own top-k.
    """

    def __init__(
        self,
        vllm_config: VllmConfig,
        config: DeepseekV2Config | DeepseekV3Config,
        hidden_size: int,
        q_lora_rank: int,
        quant_config: QuantizationConfig | None,
        cache_config: CacheConfig | None,
        topk_indices_buffer: torch.Tensor | None,
        compress_ratio: int,
        prefix: str,
        *,
        owns_k: bool,
        k_cache: DeepseekV4IndexerCache,
        main_head_dim: int,
        candidate_block_buffer: torch.Tensor | None = None,
        candidate_block_size: int = 0,
        candidate_write: bool = False,
    ):
        super().__init__()
        self.vllm_config = vllm_config
        self.config = config
        self.quant_config = quant_config
        self.topk_tokens = config.index_topk
        self.n_head = config.index_n_heads  # 32
        self.head_dim = config.index_head_dim  # 128
        self.rope_dim = config.qk_rope_head_dim  # 64
        self.q_lora_rank = q_lora_rank  # 1280
        self.compress_ratio = compress_ratio
        self.owns_k = owns_k
        self.use_fp4_kv = dsa_indexer_uses_fp4(vllm_config)
        logger.info_once(
            "Using %s indexer cache for Lightning Indexer.",
            "MXFP4" if self.use_fp4_kv else "FP8",
        )

        # no tensor parallel, just replicated
        self.wq_b = ReplicatedLinear(
            self.q_lora_rank,
            self.head_dim * self.n_head,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.wq_b",
        )
        self.weights_proj = ReplicatedLinear(
            hidden_size,
            self.n_head,
            bias=False,
            quant_config=None,
            prefix=f"{prefix}.weights_proj",
        )
        self.softmax_scale = self.head_dim**-0.5

        self.scale_fmt = "ue8m0"
        self.quant_block_size = 128  # TODO: get from config
        self.topk_indices_buffer = topk_indices_buffer

        self.max_model_len = (
            vllm_config.model_config.max_model_len // self.compress_ratio
        )
        self.prefix = prefix

        self.max_total_seq_len = (
            get_max_prefill_buffer_size(vllm_config) // self.compress_ratio
        )

        assert cache_config is not None, "Deepseek V4 indexer requires cache_config"
        if owns_k:
            # wk maps the main compressor's (pre-RoPE) latent to index keys.
            # The checkpoint stores it in bf16 with no quantization scales.
            self.wk = ReplicatedLinear(
                main_head_dim,
                self.head_dim,
                bias=False,
                quant_config=None,
                prefix=f"{prefix}.wk",
            )
            self.k_norm = RMSNorm(self.head_dim, config.rms_norm_eps)
        self.k_cache = k_cache

        self.indexer_op = SparseAttnIndexer(
            self.k_cache,
            self.quant_block_size,
            self.scale_fmt,
            self.topk_tokens,
            self.head_dim,
            self.max_model_len,
            self.max_total_seq_len,
            self.topk_indices_buffer,
            skip_k_cache_insert=True,
            use_fp4_cache=self.use_fp4_kv,
            compress_ratio=self.compress_ratio,
            candidate_blocks=candidate_block_buffer,
            candidate_block_size=candidate_block_size,
            candidate_write=candidate_write,
        )

    def _produce_k(
        self,
        latent: torch.Tensor | None,
        positions: torch.Tensor,
        rotary_emb: nn.Module,
    ) -> None:
        """Turn the compressor's pre-RoPE latent into paged index keys.

        ``k_norm(wk(latent))`` at group-boundary tokens, RoPE'd at the group
        position and MXFP4/FP8-quantized into the indexer K cache.
        """
        attn_metadata = get_forward_context().attn_metadata
        if not isinstance(attn_metadata, dict) or latent is None:
            # Profile run: the indexer K cache is not bound yet and the
            # compressor skipped its latent output.
            return
        assert self.owns_k
        indexer_metadata = cast(Any, attn_metadata[self.k_cache.prefix])
        # ReplicatedLinear returns (output, bias); bias is None. Rows at
        # non-boundary tokens hold garbage latent and are skipped by the
        # store kernel.
        k_pre, _ = self.wk(latent)
        indexer_k_norm_rope_store(
            k_pre,
            positions,
            rotary_emb.cos_sin_cache,
            self.k_norm.weight,
            self.k_norm.variance_epsilon,
            self.k_cache.kv_cache,
            indexer_metadata.slot_mapping,
            self.compress_ratio,
            self.use_fp4_kv,
        )

    def forward(
        self,
        qr: torch.Tensor | QuantizedActivation,
        latent: torch.Tensor | None,
        indexer_weights: torch.Tensor,
        positions: torch.Tensor,
        rotary_emb: nn.Module,
        qr_scale: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor | None]:
        attn_metadata = get_forward_context().attn_metadata
        if isinstance(attn_metadata, dict):
            indexer_metadata = cast(Any, attn_metadata[self.k_cache.prefix])
            if (
                indexer_metadata.max_seq_len // self.compress_ratio <= self.topk_tokens
                and not torch.cuda.is_current_stream_capturing()
            ):
                # candidates num smaller than topk, every candidate is selected
                # but we still need to build k cache
                if self.owns_k:
                    self._produce_k(latent, positions, rotary_emb)
                assert self.topk_indices_buffer is not None
                num_tokens = (
                    indexer_metadata.num_decode_tokens
                    + indexer_metadata.num_prefill_tokens
                )
                if num_tokens > 0:
                    _fill_short_context_topk_indices[(num_tokens,)](
                        self.topk_indices_buffer,
                        positions,
                        TOP_K=self.topk_tokens,
                        COMPRESS_RATIO=self.compress_ratio,
                        PADDED_TOP_K=triton.next_power_of_2(self.topk_tokens),
                        num_warps=8,
                    )
                return None, None, None

        if self.owns_k:
            # K write must land before indexer_op reads the cache
            # (skip_k_cache_insert=True).
            self._produce_k(latent, positions, rotary_emb)

        q = self._wq_b_proj(qr, qr_scale)
        q = q.view(-1, self.n_head, self.head_dim)
        q_quant, weights = fused_indexer_q_rope_quant(
            positions,
            q,
            rotary_emb.cos_sin_cache,
            indexer_weights,
            self.softmax_scale,
            self.n_head**-0.5,
            use_fp4=self.use_fp4_kv,
        )
        if isinstance(q_quant, tuple):
            q, q_scale = q_quant
        else:
            q, q_scale = q_quant, None
        return q, q_scale, weights

    def _wq_b_proj(
        self,
        qr: torch.Tensor | QuantizedActivation,
        qr_scale: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """Project normalized Q with the main attention's shared quantization."""
        assert qr_scale is None, "ROCm-only path"
        # ReplicatedLinear returns (output, bias); bias is None.
        q, _ = self.wq_b(qr)
        return q