common.py

# SPDX-License-Identifier: Apache-2.0
"""
This file implements common components for MLA implementations.

First we define:

Sq      as Q sequence length
Skv     as KV sequence length

MLA has two possible ways of computing, a data-movement friendly approach and a
compute friendly approach, we generally want to use the compute friendly
approach for "prefill" (i.e. the ratio Sq / Skv is "small", is near 1)
and the data-movement friendly approach for "decode" (i.e. the ratio
Sq / Skv is "large").

NOTE what we deem small and large is currently determined by if its labelled
prefill or decode by the scheduler, but this is something we should probably
tune.

Main reference: DeepseekV2 paper, and FlashInfer Implementation
(https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551).

Deepseek's MLA attention works the following way:
* Use a single latent vector to represent the per-token entry of the KV cache. 
* For decode (i.e. the memory friendly approach) the attention "simulates" a
multi-head attention, while the compute is similar to multi-query attention.

Below is example of both paths assuming batchsize = 1

## More Extent Definitions:

C           Context length, `Skv - Sq`
H           hidden size
N           number of attention heads
Lq          latent dimension for Q              1536 in DSV3
Lkv         latent dimension for K/V            512 in DSV3
P           nope dimension, no rope.            128 in DSV3
R           rope dimension, goes through rope.  64 in DSV3
V           V head dim.                         128 in DSV3

## Vector/Matrix Definitions

h_t         hidden states (input to attention)  shape [Sq, H]
q_c         latent/compressed Q                 shape [Sq, Lq]
q_nope      uncompressed Q (no-rope)            shape [Sq, N, P]
q_pe        uncompressed Q (rope)               shape [Sq, N, R]
kv_c        latent/compressed KV                shape [Skv, Lkv]
k_pe        decoupled k position embeddings     shape [Skv, R]
new_kv_c    new kv_c from current iter          shape [Sq, Lkv]
new_k_pe    new k_pe from current iter          shape [Sq, R]
cache_kv_c  cached k_c from previous iters      shape [C, Lkv]
cache_k_pe  cached k_pe from previous iters     shape [C, R]
W_DQ        project h_t to q_c                  shape [H, Lq]
W_UQ        project q_c to q_nope               shape [Lq, N * P]
W_QR        project q_c to q_pe                 shape [Lq, N * R]
W_DKV       project h_t to kv_c                 shape [H, Lkv]
W_UK        project kv_c to k_nope              shape [Lkv, N, P]
W_KR        project h_t to k_pe                 shape [H, R]
W_UV        project kv_c to v                   shape [Lkv, N, V]
W_O         project v to h_t                    shape [N * V, H]


## Compute Friendly Approach (i.e. "_forward_prefill"):

q_c      = h_t @ W_DQ
q_nope   = (q_c @ W_UQ).view(Sq, N, P)
q_pe     = RoPE(q_c @ W_QR).view(Sq, N, R)
new_kv_c = h_t @ W_DKV
new_k_pe = RoPE(h_t @ W_KR)
kv_c     = torch.cat([new_kv_c, cache_kv_c], dim=0)
k_pe     = torch.cat([new_k_pe, cache_k_pe], dim=0)
k_nope   = (kv_c @ W_UK.view(Lkv, N * P)).view(Skv, N, P)
v        = (kv_c @ W_UV.view(Lkv, N * V)).view(Skv, N, V)

// MHA with QK headdim = P + R
//           V headdim = V
//      spda_o shape [Sq, N, V]
spda_o = scaled_dot_product_attention(
    torch.cat([q_nope, q_pe], dim=-1),
    torch.cat([k_nope, k_pe.unsqueeze(1).expand(-1, N, -1)], dim=-1),
    v
) 
return spda_o @ W_O

NOTE: in the actual code,
    `kv_b_proj` is [W_UK; W_UV] concatenated per head
    `q_b_proj` is [W_UQ; W_QR] concatenated per head
    `out_proj` is W_O


## Data-Movement Friendly Approach (i.e. "_forward_decode"):

Runtime
q_c      = h_t @ W_DQ
q_nope   = (q_c @ W_UQ).view(-1, N, P)
ql_nope  = einsum("snh,lnh->snl", q, W_UK)
q_pe     = RoPE(q_c @ W_QR).view(Sq, N, R)
new_kv_c = h_t @ W_DKV
new_k_pe = RoPE(h_t @ W_KR)
kv_c     = torch.cat([new_kv_c, cache_kv_c], dim=0)
k_pe     = torch.cat([new_k_pe, cache_k_pe], dim=0)

// MQA with QK headdim = Lkv + R
//           V headdim = Lkv
//      spda_o shape [Sq, N, Lkv]
// NOTE: this is less compute-friendly since Lkv > P
//       but is more data-movement friendly since its MQA vs MHA
spda_o = scaled_dot_product_attention(
    torch.cat([ql_nope, q_pe], dim=-1),
    torch.cat([kv_c, k_pe], dim=-1),
    kv_c
)

o = einsum("snl,lnv->snv", spda_o.reshape(-1, N, Lkv), W_UV)
return o.view(-1, N * V) @ self.num_heads @ W_O


## Chunked Prefill

For chunked prefill we want to use the compute friendly algorithm. We are 
assuming sufficiently large Sq / Skv ratio, in the future may want to switch to 
the data-movement friendly approach if the chunk (i.e. `Sq`) is small.

However, the compute-friendly approach can potentially run out of memory if Skv
is large due to: `k_nope = (kv_c @ W_UK).view(Skv, N, P)`

To mitigate this, we chunk the computation of attention with respect to the 
current context (i.e. `cache_kv_c` and `cache_k_pe`) so that we can used a 
fixed workspace size.

The chunked prefill approach is as follows:

MCC        Max chunk of context to process per iter, computed dynamically, 
           used to bound the memory usage

q_c        = h_t @ W_DQ
q_nope     = (q_c @ W_UQ).view(Sq, N, P)
q_pe       = RoPE(q_c @ W_QR).view(Sq, N, R)
new_kv_c   = h_t @ W_DKV
new_k_pe   = RoPE(h_t @ W_KR)
new_k_nope = (new_kv_c @ W_UK.view(Lkv, N * P)).view(Sq, N, P)
new_v      = (new_kv_c @ W_UV.view(Lkv, N * V)).view(Sq, N, V)

// MHA between queries and new KV
//     with QK headdim = P + R
//           V headdim = V
//    curr_o   shape [Sq, N, V]
//    curr_lse shape [N, Sq], this is just order FA returns
curr_o, curr_lse = scaled_dot_product_attention(
    torch.cat([q_nope, q_pe], dim=-1),
    torch.cat([new_k_nope, new_k_pe.unsqueeze(1).expand(-1, N, -1)], dim=-1),
    new_v,
    casual=True,
    return_softmax_lse=True
) 

// Compute attention with the already existing context
for chunk_idx in range(cdiv(C, MCC)):
    chunk_start  = chunk_idx * MCC
    chunk_end    = min(chunk_start + MCC, C)
    Sc           = chunk_end - chunk_start
    cache_kv_c_chunk   = cache_kv_c[chunk_start:chunk_end]
    cache_k_pe_chunk   = cache_k_pe[chunk_start:chunk_end]
    cache_k_nope_chunk = (cache_kv_c_chunk @ W_UK).view(-1, N, P)
    cache_v_chunk      = (cache_kv_c_chunk @ W_UV).view(-1, N, V)

    chunk_o, chunk_lse = scaled_dot_product_attention(
        torch.cat([q_nope, q_pe], dim=-1),
        torch.cat([cache_k_nope_chunk,
                   cache_k_pe_chunk.unsqueeze(1).expand(-1, N, -1)],
                   dim=-1),
        cache_v_chunk,
        casual=False,
        return_softmax_lse=True
    )

    curr_o, curr_lse = merge_attn_states(
        suffix_output=curr_o,
        suffix_lse=curr_lse,
        prefix_output=chunk_o,
        prefix_lse=chunk_lse,
    )

return curr_o @ W_O
"""

import functools
from abc import abstractmethod
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Generic, Optional, TypeVar

import torch
import os

from vllm import _custom_ops as ops
from vllm.attention.backends.abstract import (AttentionBackend, AttentionLayer,
                                              AttentionMetadata,
                                              MLAAttentionImpl)
from vllm.attention.backends.utils import get_mla_dims
from vllm.attention.ops.merge_attn_states import merge_attn_states
from vllm.attention.utils.fa_utils import get_flash_attn_version
from vllm.logger import init_logger
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               LinearBase, RowParallelLinear,
                                               UnquantizedLinearMethod)
from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
from vllm.platforms import current_platform
from vllm.utils import cdiv, round_down

try:
    from vllm.vllm_flash_attn import flash_attn_varlen_func
    is_vllm_fa = True
except ImportError:
    # For rocm use upstream flash attention
    from flash_attn import flash_attn_varlen_func
    is_vllm_fa = False

if TYPE_CHECKING:
    from vllm.v1.core.sched.output import SchedulerOutput
    from vllm.v1.worker.gpu_input_batch import InputBatch
    from vllm.v1.worker.gpu_model_runner import GPUModelRunner

logger = init_logger(__name__)


class MLACommonBackend(AttentionBackend):

    accept_output_buffer: bool = True

    @staticmethod
    def get_name() -> str:
        return "TRITON_MLA_VLLM_V1"

    @staticmethod
    def get_metadata_cls() -> type["AttentionMetadata"]:
        return MLACommonMetadata

    @staticmethod
    def get_builder_cls() -> type["MLACommonMetadataBuilder"]:
        return MLACommonMetadataBuilder

    @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,
    ) -> tuple[int, ...]:
        return (num_blocks, block_size, head_size)

    @staticmethod
    def get_supported_head_sizes() -> list[int]:
        return [576]


@dataclass
class MLACommonPrefillMetadata:
    """ Prefill Specific Metadata """

    @dataclass
    class ChunkedContextMetadata:
        # New for MLA (compared to FlashAttention)
        # For handling chunked prefill
        cu_seq_lens: torch.Tensor
        starts: torch.Tensor
        seq_tot: list[int]
        max_seq_lens: list[int]
        workspace: torch.Tensor

    # Input positions for rotrary embeddings since for MLA the rotary
    # position embeddings are applied inside the attention backend
    input_positions: torch.Tensor
    block_table: torch.Tensor
    query_start_loc: torch.Tensor
    max_query_len: int
    chunked_context: Optional[ChunkedContextMetadata] = None


@dataclass
class MLACommonDecodeMetadata:
    # Input positions for rotrary embeddings since for MLA the rotary
    # position embeddings are applied inside the attention backend
    input_positions: torch.Tensor
    block_table: torch.Tensor
    seq_lens: torch.Tensor


D = TypeVar("D", bound=MLACommonDecodeMetadata)


@dataclass
class MLACommonMetadata(Generic[D]):
    """Metadata for MLACommon.

    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """
    # NOTE(sang): Definition of context_len, query_len, and seq_len.
    # |---------- N-1 iteration --------|
    # |---------------- N iteration ---------------------|
    # |- tokenA -|......................|-- newTokens ---|
    # |---------- context_len ----------|
    # |-------------------- seq_len ---------------------|
    #                                   |-- query_len ---|

    num_actual_tokens: int  # Number of tokens excluding padding.
    query_start_loc: torch.Tensor
    slot_mapping: torch.Tensor

    # New for MLA (compared to FlashAttention)
    # For handling prefill decode split
    num_decodes: int
    num_decode_tokens: int
    num_prefills: int

    # For logging.
    num_input_tokens: int = 0  # Number of tokens including padding.

    # The dimension of the attention heads
    head_dim: Optional[int] = None

    decode: Optional[D] = None
    prefill: Optional[MLACommonPrefillMetadata] = None

    def __post_init__(self):
        supported_head_sizes = MLACommonBackend.get_supported_head_sizes()
        if self.head_dim is not None and self.head_dim \
                not in supported_head_sizes:
            raise ValueError(
                f"Only {supported_head_sizes} are supported for head_dim,",
                f"received {self.head_dim}.")


M = TypeVar("M", bound=MLACommonMetadata)


class MLACommonMetadataBuilder(Generic[M]):
    """
    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """

    def __init__(self,
                 runner: "GPUModelRunner",
                 metadata_cls: Optional[type[M]] = None):
        self.metadata_cls = metadata_cls \
            if metadata_cls is not None else MLACommonMetadata
        self.runner = runner
        scheduler_config = runner.scheduler_config
        model_config = runner.model_config
        cache_config = runner.cache_config
        self.chunked_prefill_enabled = scheduler_config.chunked_prefill_enabled
        self.num_heads = model_config.get_num_attention_heads(
            runner.parallel_config)
        self.mla_dims = get_mla_dims(model_config)
        self.aot_schedule = is_vllm_fa and (get_flash_attn_version() == 3)

        # Dont try to access the runner on AMD
        if self.aot_schedule:
            self.page_size = self.runner.block_size

        if self.chunked_prefill_enabled:
            self.chunked_prefill_workspace_size = min(
                # Max sure there is enough for 8 full length request or at least
                # 4 pages of cache per request
                max(
                    8 * model_config.max_model_len, 4 *
                    scheduler_config.max_num_seqs * cache_config.block_size),
                # For long-context models try not to over-allocate limiting
                # kv-cache space, limiting it to 64k tokens,
                # which would result in the workspace being:
                #   2*(576)*(64*1024) = 144mb
                # (assuming 576 MLA head dim, and fp16)
                # which would result in up-projected context being
                #   2*(192*128)*(64*1024) = 3gb
                # (assuming 192 QK head dim, 128 heads, and fp16)
                128 * 1024)
            assert self.chunked_prefill_workspace_size >= \
                scheduler_config.max_num_seqs * cache_config.block_size
            self.chunked_prefill_workspace = torch.empty(
                (self.chunked_prefill_workspace_size,
                 model_config.get_head_size()),
                dtype=model_config.dtype,
                device=runner.device,
            )

    def reorder_batch(self, input_batch: "InputBatch",
                      scheduler_output: "SchedulerOutput") -> bool:
        # We now want to reorder the batch so that the "decode" requests are and
        # the front and the "prefill" requests are at the using the least amount
        # swaps possible. (NOTE for now we loosely use "decode" to mean requests
        # where attention is likely memory-bound and "prefill" to mean requests
        # where attention is likely compute-bound, TODO(lucas): figure out a
        # better naming here)
        decodes = []
        prefills = []
        num_decode_tokens = 0
        num_prefill_tokens = 0

        for i, req_id in enumerate(input_batch.req_ids):
            num_tokens = scheduler_output.num_scheduled_tokens[req_id]
            # for now treat 1 scheduled token as "decode" even if its not,
            # we should update this to something like < 8 in the future but
            # currently the TritonMLA._forward_decode only supports
            # num_tokens = 1
            if num_tokens == 1:
                decodes.append(i)
                num_decode_tokens += num_tokens
            else:
                prefills.append(i)
                num_prefill_tokens += num_tokens

        # We hope that this is fairly minimal since decodes
        # should be around for a number of iterations so hopefully they are
        # relatively stationary (and new request are generally appended to the
        # persistent batch so already should be at the back)
        # To achieve this we loop over the decodes in descending order and
        # the prefills in ascending order. We swap decodes from the  "back"
        # i.e. past where the last decode should be in the reodorered with
        # prefills from the front of the batch.
        # `decodes` and `prefills` are already in ascending order just based on
        # the above loop
        num_decodes = len(decodes)
        num_prefills = len(prefills)
        modified_batch = False

        for i in range(1, min(num_decodes, num_prefills) + 1):
            # If the decode is at the "back" of the batch, i, we can swap it
            # with the prefill closest to the front of the batch
            decode_idx = decodes[num_decodes - i]
            if decode_idx < num_decodes:
                break

            input_batch.swap_states(prefills[i - 1], decode_idx)
            modified_batch = True

        # Save for next `build` call
        # TODO(lucas): this is a bit of a hack, we should probably have a
        # better way of doing this
        self._num_decodes = num_decodes
        self._num_prefills = num_prefills
        self._num_decode_tokens = num_decode_tokens
        self._num_prefill_tokens = num_prefill_tokens

        return modified_batch

    def _build_decode(self, input_positions: torch.Tensor,
                      block_table: torch.Tensor, seq_lens: torch.Tensor):
        return MLACommonDecodeMetadata(
            input_positions=input_positions,
            block_table=block_table,
            seq_lens=seq_lens,
        )

    def build(self, num_reqs: int, num_actual_tokens: int, max_query_len: int,
              common_prefix_len: int) -> M:
        assert self._num_decodes + self._num_prefills == num_reqs

        # Note(simon): be careful about the CPU <> GPU memory movement in this
        # function. We should avoid GPU -> CPU sync as much as possible because
        # it blocks on all previous kernels.
        device = self.runner.device
        block_table = (
            self.runner.input_batch.block_table.get_device_tensor()[:num_reqs])
        query_start_loc = self.runner.query_start_loc_cpu[:num_reqs + 1].to(
            device, non_blocking=True)
        slot_mapping = self.runner.slot_mapping_cpu[:num_actual_tokens].to(
            device, non_blocking=True).long()
        input_positions = self.runner.positions_cpu[:num_actual_tokens].to(
            device, non_blocking=True).long()

        seq_lens_cpu = self.runner.seq_lens_cpu[:num_reqs]
        seq_lens = seq_lens_cpu.to(device, non_blocking=True)

        prefill_metadata = None
        if self._num_prefills > 0:
            reqs_start = self._num_decodes  # prefill_start
            tokens_start = self._num_decode_tokens

            context_lens_cpu = self.runner.input_batch.\
                num_computed_tokens_cpu_tensor[reqs_start:num_reqs]
            max_context_len_cpu = context_lens_cpu.max().item()
            num_prefills_with_context_cpu = (context_lens_cpu > 0).sum().item()
            prefill_query_start_loc = query_start_loc[
                reqs_start:] - query_start_loc[reqs_start]

            chunked_context_metadata = None
            if self.chunked_prefill_enabled and self._num_prefills > 0 \
                and max_context_len_cpu > 0:
                # NOTE: it is recommend you read the `Chunked Prefill` section
                # in the comment at the top of the file before trying to
                # understand the following code

                # currently we allocate an equal amount of workspace for each
                # prefill in the batch, we could probably use a more advanced
                # algorithm here and allocate more workspace to prefills with
                # longer context lengths
                max_context_chunk = (self.chunked_prefill_workspace_size //
                                     num_prefills_with_context_cpu)

                # align max_context_chunk to page_size by rounding down,
                # currently the `gather_cache` kernel cannot handle
                # `context_chunk_starts` that are not aligned to page_size
                max_context_chunk = round_down(max_context_chunk,
                                               self.page_size)

                assert max_context_chunk > 0
                num_chunks = cdiv(max_context_len_cpu, max_context_chunk)

                # if `max_context_chunk = 256`, `num_chunks = 3`, and
                #   `num_prefills_with_context = 4`, create a tensor that looks
                # like
                #  [[0, 0, 0, 0], [256, 256, 256, 256], [512, 512, 512, 512]]
                # Note(simon): this is done in CPU because of downstream's
                # of `to_list`.
                chunk_starts = \
                    torch.arange(num_chunks, dtype=torch.int32) \
                    .unsqueeze(1).expand(-1, self._num_prefills) \
                    * max_context_chunk
                chunk_ends = torch.min(context_lens_cpu.unsqueeze(0),
                                       chunk_starts + max_context_chunk)
                chunk_seq_lens = (chunk_ends - chunk_starts).clamp(min=0)

                cu_seq_lens_cpu = torch.zeros(num_chunks,
                                              self._num_prefills + 1,
                                              dtype=torch.int32,
                                              pin_memory=True)
                torch.cumsum(chunk_seq_lens,
                             dim=1,
                             out=cu_seq_lens_cpu[:, 1:],
                             dtype=torch.int32)

                chunked_context_metadata = \
                    MLACommonPrefillMetadata.ChunkedContextMetadata(
                    cu_seq_lens=cu_seq_lens_cpu.to(device, non_blocking=True),
                    starts=chunk_starts.to(device, non_blocking=True),
                    seq_tot=chunk_seq_lens.sum(dim=1).tolist(),
                    max_seq_lens=chunk_seq_lens.max(dim=1).values.tolist(),
                    workspace=self.chunked_prefill_workspace,
                )

                assert max(chunked_context_metadata.max_seq_lens) <= \
                    self.chunked_prefill_workspace_size

            prefill_metadata = MLACommonPrefillMetadata(
                input_positions=input_positions[tokens_start:],
                block_table=block_table[reqs_start:, ...],
                query_start_loc=prefill_query_start_loc,
                max_query_len=max_query_len,
                chunked_context=chunked_context_metadata,
            )

        decode_metadata = None
        if self._num_decodes > 0:
            decode_metadata = self._build_decode(
                input_positions=input_positions[:self._num_decode_tokens],
                block_table=block_table[:self._num_decodes, ...],
                seq_lens=seq_lens[:self._num_decodes],
            )

        return self.metadata_cls(
            num_actual_tokens=num_actual_tokens,
            query_start_loc=query_start_loc,
            slot_mapping=slot_mapping,
            head_dim=self.runner.model_config.get_head_size(),
            # MLACommonMetadata Chunk prefill specific
            num_decodes=self._num_decodes,
            num_decode_tokens=self._num_decode_tokens,
            num_prefills=self._num_prefills,
            prefill=prefill_metadata,
            decode=decode_metadata,
        )

    def use_cascade_attention(self, *args, **kwargs) -> bool:
        return False


class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
    """
    NOTE: Please read the comment at the top of the file before trying to
    understand this class
    """

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float,
        num_kv_heads: int,
        alibi_slopes: Optional[list[float]],
        sliding_window: Optional[int],
        kv_cache_dtype: str,
        blocksparse_params: Optional[dict[str, Any]],
        logits_soft_cap: Optional[float],
        attn_type: str,
        # MLA Specific Arguments
        q_lora_rank: Optional[int],
        kv_lora_rank: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        qk_head_dim: int,
        v_head_dim: int,
        rotary_emb: RotaryEmbedding,
        # q_proj should be q_b_proj if q_lora_rank is not None, but from an
        # attention backend perspective we rely on the layer to pass in the
        # correct matrix
        q_proj: ColumnParallelLinear,
        kv_b_proj: ColumnParallelLinear,
        o_proj: RowParallelLinear,
    ) -> 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.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_head_dim
        self.v_head_dim = v_head_dim

        # Hack for V1 for now to avoid torch library overhead (since we are
        # already inside an attention custom op), pull out the forward
        # method from the rotary embedding and call it directly
        # TODO(lucas): we should probably find a cleaner way to do this
        self.rotary_emb = rotary_emb.forward_native
        if current_platform.is_cuda():
            self.rotary_emb = rotary_emb.forward_cuda

        self.q_proj = q_proj
        self.kv_b_proj = kv_b_proj
        self.o_proj = o_proj
        self.vllm_flash_attn_version = get_flash_attn_version()

        # Handle the differences between the flash_attn_varlen from flash_attn
        # and the one from vllm_flash_attn. The former is used on RoCM and the
        # latter has an additional parameter to control FA2 vs FA3
        self.flash_attn_varlen_func = flash_attn_varlen_func
        self.vllm_flash_attn_version = get_flash_attn_version()
        if self.vllm_flash_attn_version is not None:
            self.flash_attn_varlen_func = \
                functools.partial(flash_attn_varlen_func,
                                  fa_version=self.vllm_flash_attn_version)
        
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'

        # For MLA the v head dim is smaller than qk head dim so we pad out
        # v with 0s to match the qk head dim for attention backends that do
        # not support different headdims
        # We don't need to pad V if we are on a hopper system with FA3
        self._pad_v = self.vllm_flash_attn_version is None or not (
            self.vllm_flash_attn_version == 3
            and current_platform.get_device_capability()[0] == 9
            and torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count == 120 )

    def _flash_attn_varlen_diff_headdims(self,
                                         q,
                                         k,
                                         v,
                                         return_softmax_lse=False,
                                         softmax_scale=None,
                                         **kwargs):
        maybe_padded_v = v
        if self._pad_v:
            # maybe_padded_v = torch.nn.functional.pad(
            #     v, [0, q.shape[-1] - v.shape[-1]], value=0)
            maybe_padded_v = torch.nn.functional.pad(
                    v, [0, q.shape[-1] - v.shape[-1]- 32], value=0)
            maybe_padded_v = maybe_padded_v[..., :-32].reshape(v.shape[0], v.shape[1],v.shape[2])

        attn_out = self.flash_attn_varlen_func(
            q=q,
            k=k,
            v=maybe_padded_v,
            return_softmax_lse=return_softmax_lse,
            softmax_scale=softmax_scale,
            **kwargs,
        )

        # Unpack the output if there is multiple results
        lse = None
        if isinstance(attn_out, tuple):
            attn_out, lse = attn_out[0], attn_out[1]

        # unpad if necessary
        if self._pad_v:
            attn_out = attn_out[..., :v.shape[-1]]

        # Remain consistent with old `flash_attn_varlen_func` where there
        # is only one output tensor if `return_softmax_lse` is False.
        if return_softmax_lse:
            return attn_out, lse
        return attn_out

    def _v_up_proj_and_o_proj(self, x):
        # Convert from (B, N, L) to (N, B, L)
        x = x.view(-1, self.num_heads, self.kv_lora_rank).transpose(0, 1)
        # Multiply (N, B, L) x (N, L, V) -> (N, B, V)
        x = torch.bmm(x, self.W_UV)
        # Convert from (N, B, V) to (B, N * V)
        x = x.transpose(0, 1).reshape(-1, self.num_heads * self.v_head_dim)
        return self.o_proj(x)[0]

    # Return `ql_nope`, `q_pe`
    def _q_proj_and_k_up_proj(self, x):
        q_nope, q_pe = self.q_proj(x)[0]\
            .view(-1, self.num_heads, self.qk_head_dim)\
            .split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

        # Convert from (B, N, P) to (N, B, P)
        q_nope = q_nope.transpose(0, 1)
        # Multiply (N, B, P) x (N, P, L) -> (N, B, L)
        ql_nope = torch.bmm(q_nope, self.W_UK_T)
        # Convert from (N, B, L) to (B, N, L)
        return ql_nope.transpose(0, 1), q_pe

    def process_weights_after_loading(self, act_dtype: torch.dtype):

        def get_layer_weight(layer):
            WEIGHT_NAMES = ("weight", "qweight", "weight_packed")
            for attr in WEIGHT_NAMES:
                if hasattr(layer, attr):
                    return getattr(layer, attr)
            raise AttributeError(
                f"Layer '{layer}' has no recognized weight attribute:"
                f" {WEIGHT_NAMES}.")

        def get_and_maybe_dequant_weights(layer: LinearBase):
            if not isinstance(layer.quant_method, UnquantizedLinearMethod):
                # NOTE: This should only be used offline, since it's O(N^3)
                eye = torch.eye(layer.input_size_per_partition,
                                dtype=act_dtype,
                                device=get_layer_weight(layer).device)
                dequant_weights = layer.quant_method.apply(layer,
                                                           eye,
                                                           bias=None)
                del eye
                # standardize to (output, input)
                return dequant_weights.T
            return layer.weight

        # we currently do not have quantized bmm's which are needed for
        # `W_UV` and `W_UK_T`, we we just store fp16/bf16 copies and perform
        # the bmm's in 16-bit, the extra memory overhead of this is fairly low
        if self.use_llama_nn and isinstance(self.kv_b_proj.quant_method, UnquantizedLinearMethod):
            kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj)
        else:
            kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T
        assert kv_b_proj_weight.shape == (
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim)), (
                f"{kv_b_proj_weight.shape=}, "
                f"{self.kv_lora_rank=}, "
                f"{self.num_heads=}, "
                f"{self.qk_nope_head_dim=}, "
                f"{self.v_head_dim=}")
        kv_b_proj_weight = kv_b_proj_weight.view(
            self.kv_lora_rank,
            self.num_heads,
            self.qk_nope_head_dim + self.v_head_dim,
        )

        W_UK, W_UV = kv_b_proj_weight.split(
            [self.qk_nope_head_dim, self.v_head_dim], dim=-1)

        # Convert from (L, N, V) to (N, L, V)
        self.W_UV = W_UV.transpose(0, 1)
        # Convert from (L, N, P) to (N, P, L)
        self.W_UK_T = W_UK.permute(1, 2, 0)

    def _compute_prefill_context(
        self,
        q: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
        attn_metadata: MLACommonMetadata,
    ):
        assert attn_metadata.prefill is not None
        prefill_metadata = attn_metadata.prefill
        assert prefill_metadata.chunked_context is not None

        output = None
        iters = len(prefill_metadata.chunked_context.seq_tot)
        workspace = prefill_metadata.chunked_context.workspace

        for i in range(iters):
            toks = prefill_metadata.chunked_context.seq_tot[i]

            ops.gather_cache(
                src_cache=kv_c_and_k_pe_cache,
                dst=workspace,
                block_table=prefill_metadata.block_table,
                cu_seq_lens=prefill_metadata.chunked_context.cu_seq_lens[i],
                batch_size=attn_metadata.num_prefills,
                seq_starts=prefill_metadata.chunked_context.starts[i],
            )

            kv_c_normed = workspace[:toks]\
                [..., :self.kv_lora_rank]
            k_pe = workspace[:toks]\
                [..., self.kv_lora_rank:].unsqueeze(1)

            kv_nope = self.kv_b_proj(kv_c_normed)[0].view( \
                -1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
            k_nope, v = kv_nope\
                .split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)

            k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))),
                          dim=-1)

            attn_output, attn_softmax_lse = \
                self._flash_attn_varlen_diff_headdims(
                q=q,
                k=k,
                v=v,
                cu_seqlens_q=prefill_metadata.query_start_loc,
                cu_seqlens_k=prefill_metadata.chunked_context.cu_seq_lens[i],
                max_seqlen_q=prefill_metadata.max_query_len,
                max_seqlen_k=prefill_metadata.chunked_context.max_seq_lens[i],
                softmax_scale=self.scale,
                causal=False,  # Context is unmasked
                return_softmax_lse=True,
            )

            if output is None:
                output = attn_output
                output_lse = attn_softmax_lse
            else:
                output_tmp = torch.empty_like(output)
                output_lse_tmp = torch.empty_like(output_lse)
                merge_attn_states(
                    output=output_tmp,
                    output_lse=output_lse_tmp,
                    prefix_output=output,
                    prefix_lse=output_lse,
                    suffix_output=attn_output,
                    suffix_lse=attn_softmax_lse,
                )
                output = output_tmp
                output_lse = output_lse_tmp

        return output, output_lse

    def _forward_prefill(
        self,
        q: torch.Tensor,
        kv_c_normed: torch.Tensor,
        k_pe: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
        attn_metadata: MLACommonMetadata,
    ) -> torch.Tensor:
        assert attn_metadata.prefill is not None

        has_context = attn_metadata.prefill.chunked_context is not None
        kv_nope = self.kv_b_proj(kv_c_normed)[0].view(\
            -1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
        k_nope, v = kv_nope\
            .split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)

        k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))), dim=-1)

        output = self._flash_attn_varlen_diff_headdims(
            q=q,
            k=k,
            v=v,
            cu_seqlens_q=attn_metadata.prefill.query_start_loc,
            cu_seqlens_k=attn_metadata.prefill.query_start_loc,
            max_seqlen_q=attn_metadata.prefill.max_query_len,
            max_seqlen_k=attn_metadata.prefill.max_query_len,
            softmax_scale=self.scale,
            causal=True,
            return_softmax_lse=has_context,
        )

        if has_context:
            suffix_output, suffix_lse = output
            context_output, context_lse = self._compute_prefill_context( \
                q, kv_c_and_k_pe_cache, attn_metadata)

            output = torch.empty_like(suffix_output)
            merge_attn_states(
                output=output,
                prefix_output=context_output,
                prefix_lse=context_lse,
                suffix_output=suffix_output,
                suffix_lse=suffix_lse,
            )

        return self.o_proj(output.flatten(start_dim=-2))[0]

    @abstractmethod
    def _forward_decode(
        self,
        ql_nope: torch.Tensor,
        q_pe: torch.Tensor,
        kv_c_and_k_pe_cache: torch.Tensor,
        attn_metadata: M,
    ) -> torch.Tensor:
        raise NotImplementedError

    def forward(
        self,
        layer: AttentionLayer,
        hidden_states_or_q_c: torch.Tensor,  # query in unified attn
        k_c_normed: torch.Tensor,  # key in unified attn
        k_pe: torch.Tensor,  # value in unified attn
        kv_cache: torch.Tensor,
        attn_metadata: M,
        output: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:

        assert output is not None, "Output tensor must be provided."

        if attn_metadata is None:
            # Profiling run.
            return output

        num_actual_toks = attn_metadata.num_actual_tokens

        # Inputs and outputs may be padded for CUDA graphs
        output_padded = output
        output = output[:num_actual_toks, ...]
        hidden_states_or_q_c = hidden_states_or_q_c[:num_actual_toks, ...]
        k_c_normed = k_c_normed[:num_actual_toks, ...]
        k_pe = k_pe[:num_actual_toks, ...]

        # Restore head dim (for rotary embedding)
        k_pe = k_pe.unsqueeze(1)

        assert attn_metadata.num_decodes is not None and \
            attn_metadata.num_prefills is not None and \
            attn_metadata.num_decode_tokens is not None

        has_decode = attn_metadata.num_decodes > 0
        has_prefill = attn_metadata.num_prefills > 0
        num_decode_tokens = attn_metadata.num_decode_tokens

        decode_hs_or_q_c = hidden_states_or_q_c[:num_decode_tokens]
        decode_k_pe = k_pe[:num_decode_tokens]

        prefill_hs_or_q_c = hidden_states_or_q_c[num_decode_tokens:]
        prefill_k_pe = k_pe[num_decode_tokens:]
        prefill_k_c_normed = k_c_normed[num_decode_tokens:]

        if has_decode:
            assert attn_metadata.decode is not None
            decode_ql_nope, decode_q_pe = \
                self._q_proj_and_k_up_proj(decode_hs_or_q_c)
            decode_q_pe[...], decode_k_pe[...] = self.rotary_emb(
                attn_metadata.decode.input_positions, decode_q_pe.contiguous(),
                decode_k_pe)

        if has_prefill:
            assert attn_metadata.prefill is not None
            prefill_q = self.q_proj(prefill_hs_or_q_c)[0]\
                .view(-1, self.num_heads, self.qk_head_dim)
            prefill_q_pe = prefill_q[..., self.qk_nope_head_dim:]

            prefill_q_pe[...], prefill_k_pe[...] = self.rotary_emb(
                attn_metadata.prefill.input_positions,
                prefill_q_pe.contiguous(), prefill_k_pe)

        # write the latent and rope to kv cache
        if kv_cache.numel() > 0:
            ops.concat_and_cache_mla(
                k_c_normed,
                k_pe.squeeze(1),
                kv_cache,
                attn_metadata.slot_mapping.flatten(),
                kv_cache_dtype=self.kv_cache_dtype,
                scale=layer._k_scale,
            )

        if has_prefill:
            output[num_decode_tokens:] = self._forward_prefill(
                prefill_q, prefill_k_c_normed, prefill_k_pe, kv_cache,
                attn_metadata)

        if has_decode:
            output[:num_decode_tokens] = self._forward_decode(
                decode_ql_nope, decode_q_pe, kv_cache, attn_metadata)

        return output_padded