flashmla.py

# adapted from: https://github.com/deepseek-ai/FlashMLA/blob/main/flash_mla/flash_mla_interface.py
from typing import Optional, Tuple

import torch

from aphrodite.platforms import current_platform

if current_platform.is_cuda():
    try:
        import aphrodite.extensions.cuda._flashmla_C  # noqa: F401
        _flashmla_C_AVAILABLE = True
    except ImportError:
        _flashmla_C_AVAILABLE = False
else:
    _flashmla_C_AVAILABLE = False


def is_flashmla_supported() -> Tuple[bool, Optional[str]]:
    """
    Return: is_supported_flag, unsupported_reason (optional).
    """
    if not current_platform.is_cuda():
        return False, "FlashMLA is only supported on CUDA devices."
    if current_platform.get_device_capability()[0] != 9:
        return False, "FlashMLA is only supported on Hopper devices."
    if not _flashmla_C_AVAILABLE:
        return False, "aphrodite.extensions.cuda._flashmla_C is not " \
            "available, likely was not compiled due to insufficient nvcc " \
            "version or a supported arch "\
            "(only sm90a currently) was not in the list of target arches to "\
            "compile for."
    return True, None


def get_mla_metadata(
    cache_seqlens: torch.Tensor,
    num_heads_per_head_k: int,
    num_heads_k: int,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Arguments:
        cache_seqlens: (batch_size), dtype torch.int32.
        num_heads_per_head_k: Equals to seq_len_q * num_heads_q // num_heads_k.
        num_heads_k: num_heads_k.

    Return:
        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), 
                                 dtype torch.int32.
        num_splits: (batch_size + 1), dtype torch.int32.
    """
    return torch.ops._flashmla_C.get_mla_metadata(cache_seqlens,
                                                  num_heads_per_head_k,
                                                  num_heads_k)


def flash_mla_with_kvcache(
    q: torch.Tensor,
    k_cache: torch.Tensor,
    block_table: torch.Tensor,
    cache_seqlens: torch.Tensor,
    head_dim_v: int,
    tile_scheduler_metadata: torch.Tensor,
    num_splits: torch.Tensor,
    softmax_scale: Optional[float] = None,
    causal: bool = False,
    descale_q: Optional[torch.Tensor] = None,
    descale_k: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Arguments:
        q: (batch_size, seq_len_q, num_heads_q, head_dim).
        k_cache: (num_blocks, page_block_size, num_heads_k, head_dim).
        block_table: (batch_size, max_num_blocks_per_seq), torch.int32.
        cache_seqlens: (batch_size), torch.int32.
        head_dim_v: Head_dim of v.
        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), 
                                 torch.int32, return by get_mla_metadata.
        num_splits: (batch_size + 1), torch.int32, return by get_mla_metadata.
        softmax_scale: float. The scaling of QK^T before applying softmax. 
                       Default to 1 / sqrt(head_dim).
        causal: bool. Whether to apply causal attention mask.
        descale_q: (batch_size), torch.float32. Descaling factors for Q.
        descale_k: (batch_size), torch.float32. Descaling factors for K.

    Return:
        out: (batch_size, seq_len_q, num_heads_q, head_dim_v).
        softmax_lse: (batch_size, num_heads_q, seq_len_q), torch.float32.
    """
    if softmax_scale is None:
        softmax_scale = q.shape[-1]**(-0.5)
    out, softmax_lse = torch.ops._flashmla_C.fwd_kvcache_mla(
        q,
        k_cache,
        head_dim_v,
        cache_seqlens,
        block_table,
        softmax_scale,
        causal,
        tile_scheduler_metadata,
        num_splits,
        descale_q,
        descale_k,
    )

    # Note(hc): need revisit when we support DCP with decode query_len > 1.
    return out.squeeze(1), softmax_lse.squeeze(-1)


#
# TODO: Add fake functions
#
# @register_fake("_flashmla_C::get_mla_metadata")
# def _get_mla_metadata_fake(....) -> Tuple[torch.Tensor, torch.Tensor]:
#     return ....
#
# @register_fake("_flashmla_C::fwd_kvcache_mla")
# def _fwd_kvcache_mla_fake(....) -> Tuple[torch.Tensor, torch.Tensor]:
#     return ....
#