# mypy: allow-untyped-defs # flake8: noqa C101 """This module implements the user facing API for flex_attention in PyTorch.""" import functools import inspect import itertools import math import operator import warnings from enum import Enum from typing import Any, Callable, Optional, Union import torch from torch import Tensor from torch._dynamo._trace_wrapped_higher_order_op import TransformGetItemToIndex from torch._higher_order_ops.flex_attention import flex_attention as flex_attention_hop from torch._higher_order_ops.utils import _set_compilation_env from torch.fx.experimental.proxy_tensor import ( _temp_remove_metadata_torch_function_mode, _temp_remove_pre_dispatch_torch_function_mode, ) from torch.nn.attention._utils import _supported_head_dim, _validate_sdpa_input from torch.utils._pytree import tree_map_only __all__ = [ "BlockMask", "flex_attention", "create_block_mask", "create_mask", "create_nested_block_mask", "or_masks", "and_masks", "noop_mask", ] _score_mod_signature = Callable[[Tensor, Tensor, Tensor, Tensor, Tensor], Tensor] _mask_mod_signature = Callable[[Tensor, Tensor, Tensor, Tensor], Tensor] class _ModificationType(Enum): """Enum for the type of modification function. - SCORE_MOD: score_mod function which accepts a score as the first argument - mask_mod: mask function which does not accept a score and is only used for generating block mask """ SCORE_MOD = 1 MASK_MOD = 2 UNKNOWN = 3 def _get_mod_type(fn: Callable) -> _ModificationType: """Get the type of modification function. This function inspects the number of positional arguments of the function to determine the type of modification function. If the function has 5 positional arguments, it is considered as a score_mod function. If the function has 4 positional arguments, it is considered as a mask function. """ num_positional_args = sum( 1 for param in inspect.signature(fn).parameters.values() if param.default == inspect.Parameter.empty ) assert num_positional_args == 5 or num_positional_args == 4 if num_positional_args == 5: return _ModificationType.SCORE_MOD elif num_positional_args == 4: return _ModificationType.MASK_MOD else: return _ModificationType.UNKNOWN # Need to define it here so that Dynamo doesn't skip it def _vmap_for_bhqkv( fn: Callable, prefix: tuple[Optional[int], ...], suffix: tuple[Optional[int], ...] = (), out_dims: Union[int, list[Optional[int]]] = 0, group_dim: bool = False, ): """Used to vmap both score_mods and mask_mods over 4-dimensional/5-dimension inputs. Mapping over the [b, hq, q_idx, kv_idx] or [b, hkv, g, q_idx, kv_idx] dimensions. Args: fn (callable): The function to vmap. prefix (tuple): The prefix of the vmap. For score mod functions, this should be set to (0,). For mask_mods = () suffix (tuple): We need to add (0,) if gradOut is being mapped over, and (None,) * len(other_buffers). out_dims (tuple): For forward cases, keep this as the default 0 since we are only returning 1 output. For backwards, the joint graph returns grads for B, H, Q_idx, KV_idx and other_buffers, so we set this to (0, None, None, None, None) + (None,) * len(other_buffers). Returns: callable: The vmapped function. """ # We vamp a function 4 times, broadcasting the [b, h, q_idx, kv_idx] dimensions dimensions: list[tuple[None | int, None | int, None | int, None | int]] = [] dimensions = [ (None, None, None, 0), (None, None, 0, None), (None, 0, None, None), ] if group_dim: dimensions += [ (None, 0, None, None), ] dimensions += [ (0, None, None, None), ] for dims in dimensions: fn = torch.vmap(fn, in_dims=prefix + dims + suffix, out_dims=out_dims) # type: ignore[arg-type] return fn def _identity( score: Tensor, batch: Tensor, head: Tensor, token_q: Tensor, token_kv: Tensor, ) -> Tensor: return score def noop_mask( batch: Tensor, head: Tensor, token_q: Tensor, token_kv: Tensor, ) -> Tensor: """Returns a noop mask_mod""" return batch.new_ones(size=(), dtype=torch.bool, device=batch.device) _DEFAULT_SPARSE_BLOCK_SIZE = 128 _LARGE_SPARSE_BLOCK_SIZE = 1 << 30 def _ordered_to_dense(num_blocks_in_row: Tensor, col_indices: Tensor): num_rows = col_indices.shape[-2] num_cols = col_indices.shape[-1] batch_dims = num_blocks_in_row.shape[:-1] device = num_blocks_in_row.device def create_dense_one(kv_num_blocks, kv_indices): dense_mask = kv_indices.new_zeros(num_rows, num_cols + 1, dtype=torch.int32) row_indices = torch.arange(num_rows, dtype=torch.int, device=device).unsqueeze( -1 ) col_range = torch.arange(num_cols, dtype=torch.int, device=device) index_mask = col_range < kv_num_blocks.unsqueeze(-1) # We write to one spot "out of bounds" valid_indices = torch.where(index_mask, kv_indices, num_cols) # set the values in 'a' to 1 where the indices are valid dense_mask[row_indices, valid_indices] = dense_mask.new_ones(()) return dense_mask[:, :num_cols].contiguous() create_dense_batched = create_dense_one for _ in range(len(batch_dims)): create_dense_batched = torch.vmap(create_dense_batched, in_dims=(0, 0)) out = create_dense_batched(num_blocks_in_row, col_indices) return out def _dense_to_ordered(dense_mask) -> tuple[Tensor, Tensor]: dense_mask = dense_mask.to(dtype=torch.int32) num_blocks_in_row = dense_mask.sum(dim=-1) col_indices = torch.argsort(dense_mask, dim=-1, descending=True, stable=True) return ( num_blocks_in_row.to(torch.int32, memory_format=torch.contiguous_format), col_indices.to(torch.int32, memory_format=torch.contiguous_format), ) def _transpose_ordered(num_blocks_in_row: Tensor, col_indices: Tensor): dense = _ordered_to_dense(num_blocks_in_row, col_indices) return _dense_to_ordered(dense.transpose(-2, -1)) def _adjust_num_blocks_and_indices( num_blocks: Tensor, indices: Tensor, new_num_rows: int, new_num_cols: int, ): indices = indices[:, :, :new_num_rows, :new_num_cols] num_blocks = num_blocks[:, :, :new_num_rows] num_blocks = torch.where(num_blocks < new_num_cols, num_blocks, new_num_cols) num_blocks = torch.sum(indices < num_blocks[:, :, :, None], dim=-1).to(torch.int32) return num_blocks, indices class BlockMask: r""" BlockMask is our format for representing a block-sparse attention mask. It is somewhat of a cross in-between BCSR and a non-sparse format. Basics ------ A block-sparse mask means that instead of representing the sparsity of individual elements in the mask, a KV_BLOCK_SIZE x Q_BLOCK_SIZE block is considered sparse only if every element within that block is sparse. This aligns well with hardware, which generally expects to perform contiguous loads and computation. This format is primarily optimized for 1. simplicity, and 2. kernel efficiency. Notably, it is *not* optimized for size, as this mask is always reduced by a factor of KV_BLOCK_SIZE * Q_BLOCK_SIZE. If the size is a concern, the tensors can be reduced in size by increasing the block size. The essentials of our format are: num_blocks_in_row: Tensor[ROWS]: Describes the number of blocks present in each row. col_indices: Tensor[ROWS, MAX_BLOCKS_IN_COL]: `col_indices[i]` is the sequence of block positions for row i. The values of this row after `col_indices[i][num_blocks_in_row[i]]` are undefined. For example, to reconstruct the original tensor from this format: .. code-block:: python dense_mask = torch.zeros(ROWS, COLS) for row in range(ROWS): for block_idx in range(num_blocks_in_row[row]): dense_mask[row, col_indices[row, block_idx]] = 1 Notably, this format makes it easier to implement a reduction along the *rows* of the mask. Details ------- The basics of our format require only kv_num_blocks and kv_indices. But, we have up to 8 tensors on this object. This represents 4 pairs: 1. (kv_num_blocks, kv_indices): Used for the forwards pass of attention, as we reduce along the KV dimension. 2. [OPTIONAL] (full_kv_num_blocks, full_kv_indices): This is optional and purely an optimization. As it turns out, applying masking to every block is quite expensive! If we specifically know which blocks are "full" and don't require masking at all, then we can skip applying mask_mod to these blocks. This requires the user to split out a separate mask_mod from the score_mod. For causal masks, this is about a 15% speedup. 3. [GENERATED] (q_num_blocks, q_indices): Required for the backwards pass, as computing dKV requires iterating along the mask along the Q dimension. These are autogenerated from 1. 4. [GENERATED] (full_q_num_blocks, full_q_indices): Same as above, but for the backwards pass. These are autogenerated from 2. """ seq_lengths: tuple[int, int] kv_num_blocks: Tensor kv_indices: Tensor full_kv_num_blocks: Optional[Tensor] full_kv_indices: Optional[Tensor] q_num_blocks: Optional[Tensor] q_indices: Optional[Tensor] full_q_num_blocks: Optional[Tensor] full_q_indices: Optional[Tensor] BLOCK_SIZE: tuple[int, int] mask_mod: _mask_mod_signature def __init__( self, seq_lengths: tuple[int, int], kv_num_blocks: Tensor, kv_indices: Tensor, full_kv_num_blocks: Optional[Tensor], full_kv_indices: Optional[Tensor], q_num_blocks: Optional[Tensor], q_indices: Optional[Tensor], full_q_num_blocks: Optional[Tensor], full_q_indices: Optional[Tensor], BLOCK_SIZE: tuple[int, int], mask_mod: _mask_mod_signature, ): if kv_indices.dim() < 2: raise RuntimeError("BlockMask must have at least 2 dimensions") assert kv_num_blocks is not None, "kv_num_blocks must be provided" assert kv_indices is not None, "kv_indices must be provided" assert q_num_blocks is not None, "q_num_blocks must be provided" assert q_indices is not None, "q_indices must be provided" assert (full_kv_num_blocks is None) == ( full_kv_indices is None ), "full_kv_num_blocks and full_kv_indices must be both provided or omitted" assert (full_q_num_blocks is None) == ( full_q_indices is None ), "full_q_num_blocks and full_q_indices must be both provided or omitted" self.seq_lengths = seq_lengths self.kv_num_blocks = kv_num_blocks self.kv_indices = kv_indices self.full_kv_num_blocks = full_kv_num_blocks self.full_kv_indices = full_kv_indices self.q_num_blocks = q_num_blocks self.q_indices = q_indices self.full_q_num_blocks = full_q_num_blocks self.full_q_indices = full_q_indices self.BLOCK_SIZE = BLOCK_SIZE self.mask_mod = mask_mod @classmethod def from_kv_blocks( cls, kv_num_blocks: Tensor, kv_indices: Tensor, full_kv_num_blocks: Optional[Tensor] = None, full_kv_indices: Optional[Tensor] = None, BLOCK_SIZE: Union[int, tuple[int, int]] = _DEFAULT_SPARSE_BLOCK_SIZE, mask_mod: Optional[_mask_mod_signature] = None, seq_lengths: Optional[tuple[int, int]] = None, ): """ Creates a BlockMask instance from key-value block information. Args: kv_num_blocks (Tensor): Number of kv_blocks in each Q_BLOCK_SIZE row tile. kv_indices (Tensor): Indices of key-value blocks in each Q_BLOCK_SIZE row tile. full_kv_num_blocks (Optional[Tensor]): Number of full kv_blocks in each Q_BLOCK_SIZE row tile. full_kv_indices (Optional[Tensor]): Indices of full key-value blocks in each Q_BLOCK_SIZE row tile. BLOCK_SIZE (Union[int, tuple[int, int]]): Size of KV_BLOCK_SIZE x Q_BLOCK_SIZE tiles. mask_mod (Optional[Callable]): Function to modify the mask. Returns: BlockMask: Instance with full Q information generated via _transposed_ordered Raises: RuntimeError: If kv_indices has < 2 dimensions. AssertionError: If only one of full_kv_* args is provided. """ if kv_indices.dim() < 2: raise RuntimeError("BlockMask must have at least 2 dimensions") assert (full_kv_num_blocks is None) == ( full_kv_indices is None ), "full_kv_num_blocks and full_kv_indices must be both provided or omitted" # Generate q_num_blocks and q_indices q_num_blocks, q_indices = _transpose_ordered(kv_num_blocks, kv_indices) if full_kv_num_blocks is not None: assert full_kv_indices is not None full_q_num_blocks, full_q_indices = _transpose_ordered( full_kv_num_blocks, full_kv_indices ) else: full_q_num_blocks, full_q_indices = None, None if isinstance(BLOCK_SIZE, int): BLOCK_SIZE = (BLOCK_SIZE, BLOCK_SIZE) mask_mod = mask_mod if mask_mod is not None else noop_mask if seq_lengths is None: q_length = kv_indices.shape[-2] * BLOCK_SIZE[0] kv_length = q_indices.shape[-2] * BLOCK_SIZE[1] seq_lengths = (q_length, kv_length) return cls( seq_lengths=seq_lengths, kv_num_blocks=kv_num_blocks, kv_indices=kv_indices, full_kv_num_blocks=full_kv_num_blocks, full_kv_indices=full_kv_indices, q_num_blocks=q_num_blocks, q_indices=q_indices, full_q_num_blocks=full_q_num_blocks, full_q_indices=full_q_indices, BLOCK_SIZE=BLOCK_SIZE, mask_mod=mask_mod, ) def as_tuple(self, flatten: bool = True): """ Returns a tuple of the attributes of the BlockMask. Args: flatten (bool): If True, it will flatten the tuple of (KV_BLOCK_SIZE, Q_BLOCK_SIZE) """ if flatten: block_size = (self.BLOCK_SIZE[0], self.BLOCK_SIZE[1]) # type: ignore[assignment] seq_lengths = (self.seq_lengths[0], self.seq_lengths[1]) # type: ignore[assignment] else: block_size = (self.BLOCK_SIZE,) # type: ignore[assignment] seq_lengths = (self.seq_lengths,) # type: ignore[assignment] return ( *seq_lengths, self.kv_num_blocks, self.kv_indices, self.full_kv_num_blocks, self.full_kv_indices, self.q_num_blocks, self.q_indices, self.full_q_num_blocks, self.full_q_indices, *block_size, self.mask_mod, ) @property def shape(self): *batch_dims, _, _ = self.kv_indices.shape return tuple(batch_dims) + self.seq_lengths def __str__(self): s = f"BlockMask(shape={self.shape}, sparsity={self.sparsity():.2f}%, \n" mask_str = self.to_string().strip() s += mask_str s += "\n)" return s def __getitem__(self, index) -> "BlockMask": """ Returns a new BlockMask instance by getting the mask for the given index position. Args: index: Index to apply to all attributes. Example Usage: .. code-block:: python def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx block_mask = create_block_mask(causal_mask, 4, 2, 512, 512, device="cuda") assert block_mask.kv_num_blocks.shape == (4,2,4) assert block_mask.kv_indices.shape == (4,2,4,4) # Index on batch dimension new_block_mask = block_mask[0] assert new_block_mask.kv_num_blocks.shape == (2,4) assert new_block_mask.kv_indices.shape == (2,4,4) # Index on batch and head dimension new_block_mask = block_mask[0, 1] assert new_block_mask.kv_num_blocks.shape == (4,) assert new_block_mask.kv_indices.shape == (4,4) # slicing on batch and head dimension new_block_mask = block_mask[0:2, 1:2] assert new_block_mask.kv_num_blocks.shape == (2,1,4) assert new_block_mask.kv_indices.shape == (2,1,4,4) # slicing on batch, head, and query dimension new_block_mask = block_mask[0:2, 1:2, torch.tensor([1], dtype=torch.int32)] assert new_block_mask.kv_num_blocks.shape == (2,1,1) assert new_block_mask.kv_indices.shape == (2,1,1,4) """ new_kv_num_blocks = self.kv_num_blocks[index] new_kv_indices = self.kv_indices[index] if self.full_kv_num_blocks is not None: assert self.full_kv_indices is not None new_full_kv_num_blocks = self.full_kv_num_blocks[index] new_full_kv_indices = self.full_kv_indices[index] else: new_full_kv_num_blocks = None new_full_kv_indices = None return BlockMask.from_kv_blocks( new_kv_num_blocks, new_kv_indices, new_full_kv_num_blocks, new_full_kv_indices, BLOCK_SIZE=self.BLOCK_SIZE, mask_mod=None, seq_lengths=self.seq_lengths, ) def __repr__(self): def shape_or_none(x: Optional[torch.Tensor]): return x.shape if x is not None else None return ( f"BlockMask(\n" f" kv_num_blocks={self.kv_num_blocks.shape},\n" f" kv_indices={self.kv_indices.shape},\n" f" full_kv_num_blocks={shape_or_none(self.full_kv_num_blocks )},\n" f" full_kv_indices={shape_or_none(self.full_kv_indices)},\n" f" q_num_blocks={shape_or_none(self.q_num_blocks)},\n" f" q_indices={shape_or_none(self.q_indices)},\n" f" full_q_num_blocks={shape_or_none(self.full_q_num_blocks)},\n" f" full_q_indices={shape_or_none(self.full_q_indices)},\n" f" BLOCK_SIZE={self.BLOCK_SIZE},\n" f" shape={self.shape},\n" f" sparsity={self.sparsity():.2f}%,\n" f" mask_mod={self.mask_mod.__name__ if hasattr(self.mask_mod, '__name__') else self.mask_mod}\n" f")" ) def _adjust(self, new_q_len: int, new_kv_len: int): new_num_rows = (new_q_len + self.BLOCK_SIZE[0] - 1) // self.BLOCK_SIZE[0] new_num_cols = (new_kv_len + self.BLOCK_SIZE[1] - 1) // self.BLOCK_SIZE[1] new_kv_num_blocks, new_kv_indices = _adjust_num_blocks_and_indices( self.kv_num_blocks, self.kv_indices, new_num_rows, new_num_cols ) if self.full_kv_num_blocks is not None: assert self.full_kv_indices is not None ( new_full_kv_num_blocks, new_full_kv_indices, ) = _adjust_num_blocks_and_indices( self.full_kv_num_blocks, self.full_kv_indices, new_num_rows, new_num_cols, ) else: new_full_kv_num_blocks = None new_full_kv_indices = None return self.from_kv_blocks( new_kv_num_blocks, new_kv_indices, new_full_kv_num_blocks, new_full_kv_indices, self.BLOCK_SIZE, self.mask_mod, ) def numel(self): """Returns the number of elements (not accounting for sparsity) in the mask.""" shape = self.shape def _prod(xs): return functools.reduce(operator.mul, xs, 1) return _prod(shape) def sparsity(self) -> float: """Computes the percentage of blocks that are sparse (i.e. not computed)""" total_size = self.numel() computed_blocks = self.kv_num_blocks.sum() if self.full_kv_num_blocks is not None: computed_blocks += self.full_kv_num_blocks.sum() computed_size = computed_blocks.item() * self.BLOCK_SIZE[0] * self.BLOCK_SIZE[1] dense_ratio = computed_size / total_size return 100 * (1 - dense_ratio) def to_dense(self) -> Tensor: """Returns a dense block that is equivalent to the block mask.""" partial_dense = _ordered_to_dense(self.kv_num_blocks, self.kv_indices) if self.full_kv_num_blocks is not None: assert self.full_kv_indices is not None return partial_dense | _ordered_to_dense( self.full_kv_num_blocks, self.full_kv_indices ) return partial_dense def to_string(self, grid_size=(20, 20), limit=4): """Returns a string representation of the block mask. Quite nifty. If grid_size is None, prints out an uncompressed version. Warning, it can be quite big! """ dense_mask = self.to_dense() *batch_dims, num_rows, num_cols = dense_mask.shape if isinstance(grid_size, int): max_rows = grid_size max_cols = grid_size elif grid_size == -1: max_rows = num_rows max_cols = num_cols else: max_rows, max_cols = grid_size def create_block_vis(*batch_idx): descriptors = [] descriptors.append(f"{batch_idx}") vis = ", ".join(reversed(descriptors)) + "\n" def summarize_section(section): percentage = section.float().mean().item() if percentage == 1: return "█" elif percentage == 0: return " " else: return "░" def cdiv(a, b): return (a + (b - 1)) // b row_step = max(1, cdiv(num_rows, max_rows)) col_step = max(1, cdiv(num_cols, max_cols)) for r in range(0, num_rows, row_step): for c in range(0, num_cols, col_step): cur_mask = dense_mask for idx in batch_idx: cur_mask = cur_mask[idx] char = summarize_section( cur_mask[r : r + row_step, c : c + col_step] ) vis += char * 2 vis += "\n" return vis total_vis = [] for idx, batch_idx in enumerate( itertools.product(*[range(i) for i in batch_dims]) ): if idx == limit: total_vis.append("...") total_vis.append("To print out more, set BlockMask.to_string(limit=N)") total_vis.append( "You can also index (BlockMask[batch, head]) to choose a specific batch or head" ) break block_vis = create_block_vis(*batch_idx) total_vis.append(block_vis) return "\n".join(total_vis) def to(self, device: Union[torch.device, str]) -> "BlockMask": """Moves the BlockMask to the specified device. Args: device (torch.device or str): The target device to move the BlockMask to. Can be a torch.device object or a string (e.g., 'cpu', 'cuda:0'). Returns: BlockMask: A new BlockMask instance with all tensor components moved to the specified device. Note: This method does not modify the original BlockMask in-place. Instead, it returns a new BlockMask instance where invidual tensor attributes may or may not be moved to the specified device, depending on their current device placement. """ mapped_attributes = tree_map_only( torch.Tensor, lambda x: x.to(device), self.as_tuple(flatten=False), ) return BlockMask(*mapped_attributes) def _broadcast_to_dim(x, dim): while x.dim() < dim: x = x.unsqueeze(0) return x def _round_up_to_multiple(x, multiple): return (x + multiple - 1) // multiple * multiple def _convert_mask_to_block_mask( mask: Tensor, Q_BLOCK_SIZE=_DEFAULT_SPARSE_BLOCK_SIZE, KV_BLOCK_SIZE=_DEFAULT_SPARSE_BLOCK_SIZE, separate_full_blocks: bool = False, ) -> tuple[Tensor, Optional[Tensor]]: assert mask.dtype == torch.bool mask = _broadcast_to_dim(mask, 4) def padding_needed_for_multiple(x, multiple): return _round_up_to_multiple(x, multiple) - x mask = torch.nn.functional.pad( mask, ( 0, padding_needed_for_multiple(mask.shape[-1], KV_BLOCK_SIZE), 0, padding_needed_for_multiple(mask.shape[-2], Q_BLOCK_SIZE), ), ) B, H, Q, KV = mask.shape assert Q % Q_BLOCK_SIZE == 0 assert KV % KV_BLOCK_SIZE == 0 mask = mask.view( B, H, Q // Q_BLOCK_SIZE, Q_BLOCK_SIZE, KV // KV_BLOCK_SIZE, KV_BLOCK_SIZE ) # [B, H, Q//Q_BLOCK_SIZE, Q_BLOCK_SIZE, KV//KV_BLOCK_SIZE, KV_BLOCK_SIZE] mask = mask.permute( 0, 1, 2, 4, 3, 5 ) # [B, H, Q//Q_BLOCK_SIZE, KV//KV_BLOCK_SIZE, Q_BLOCK_SIZE, KV_BLOCK_SIZE] mask_block_sum = mask.sum( dim=[-2, -1] ) # [B, H, Q//Q_BLOCK_SIZE, KV//KV_BLOCK_SIZE] if separate_full_blocks: full_block_sum = Q_BLOCK_SIZE * KV_BLOCK_SIZE full_blocks = mask_block_sum == full_block_sum partial_blocks = (mask_block_sum > 0) & (mask_block_sum < full_block_sum) partial_blocks = partial_blocks.to(dtype=torch.int8) full_blocks = full_blocks.to(dtype=torch.int8) return partial_blocks, full_blocks else: partial_blocks = mask_block_sum > 0 partial_blocks = partial_blocks.to(dtype=torch.int8) return partial_blocks, None def or_masks(*mask_mods: _mask_mod_signature) -> _mask_mod_signature: """Returns a mask_mod that's the union of provided mask_mods""" if not all(callable(arg) for arg in mask_mods): raise RuntimeError(f"All inputs should be callable mask_mods: {mask_mods}") def or_mask(b, h, q_idx, kv_idx): result = b.new_zeros((), dtype=torch.bool) for mask in mask_mods: result = result | mask(b, h, q_idx, kv_idx) return result return or_mask def and_masks(*mask_mods: _mask_mod_signature) -> _mask_mod_signature: """Returns a mask_mod that's the intersection of provided mask_mods""" if not all(callable(arg) for arg in mask_mods): raise RuntimeError(f"All inputs should be callable mask_mods: {mask_mods}") def and_mask(b, h, q_idx, kv_idx): result = b.new_ones((), dtype=torch.bool) for mask in mask_mods: result = result & mask(b, h, q_idx, kv_idx) return result return and_mask def _convert_block_mask_to_mask( block_mask, KV_BLOCK_SIZE=_DEFAULT_SPARSE_BLOCK_SIZE, Q_BLOCK_SIZE=_DEFAULT_SPARSE_BLOCK_SIZE, ) -> Tensor: assert block_mask.dim() == 4 B, H, Q, KV = block_mask.shape block_mask = block_mask.expand(Q_BLOCK_SIZE, KV_BLOCK_SIZE, *block_mask.shape) block_mask = block_mask.permute(2, 3, 4, 0, 5, 1).reshape( B, H, Q * Q_BLOCK_SIZE, KV * KV_BLOCK_SIZE ) return block_mask def _create_sparse_block_from_block_mask( block_mask: tuple[Tensor, Optional[Tensor]], mask_mod: Optional[Callable], seq_lengths: tuple[int, int], Q_BLOCK_SIZE: int = _DEFAULT_SPARSE_BLOCK_SIZE, KV_BLOCK_SIZE: int = _DEFAULT_SPARSE_BLOCK_SIZE, ) -> BlockMask: partial_blocks, full_blocks = block_mask partial_bm = _dense_to_ordered(partial_blocks) if full_blocks is not None: full_bm: tuple[Optional[Tensor], Optional[Tensor]] = _dense_to_ordered( full_blocks ) else: full_bm = (None, None) return BlockMask.from_kv_blocks( partial_bm[0], partial_bm[1], full_bm[0], full_bm[1], BLOCK_SIZE=(Q_BLOCK_SIZE, KV_BLOCK_SIZE), mask_mod=mask_mod, seq_lengths=seq_lengths, ) def create_mask( mod_fn: Union[_score_mod_signature, _mask_mod_signature], B: Optional[int], H: Optional[int], Q_LEN: int, KV_LEN: int, device: str = "cuda", ) -> Tensor: r"""This function creates a mask tensor from a mod_fn function. Args: mod_fn (Union[_score_mod_signature, _mask_mod_signature]): Function to modify attention scores. B (int): Batch size. H (int): Number of query heads. Q_LEN (int): Sequence length of query. KV_LEN (int): Sequence length of key/value. device (str): Device to run the mask creation on. Returns: mask (Tensor): A mask tensor with shape (B, H, M, N). """ if B is None: B = 1 if H is None: H = 1 b = torch.arange(0, B, device=device) h = torch.arange(0, H, device=device) m = torch.arange(0, Q_LEN, device=device) n = torch.arange(0, KV_LEN, device=device) mod_type = _get_mod_type(mod_fn) with TransformGetItemToIndex(): if mod_type == _ModificationType.SCORE_MOD: score_mod = mod_fn score_mod = _vmap_for_bhqkv(score_mod, prefix=(0,)) # first input is score out = score_mod(torch.zeros(B, H, Q_LEN, KV_LEN, device=device), b, h, m, n) mask = torch.where(torch.isneginf(out), False, True) return mask elif mod_type == _ModificationType.MASK_MOD: mask_mod = mod_fn mask_mod = _vmap_for_bhqkv(mask_mod, prefix=()) mask = mask_mod(b, h, m, n) return mask else: raise AssertionError def create_block_mask( mask_mod: _mask_mod_signature, B: Optional[int], H: Optional[int], Q_LEN: int, KV_LEN: int, device: str = "cuda", BLOCK_SIZE: Union[int, tuple[int, int]] = _DEFAULT_SPARSE_BLOCK_SIZE, _compile=False, ) -> BlockMask: r"""This function creates a block mask tuple from a mask_mod function. Args: mask_mod (Callable): mask_mod function. This is a callable that defines the masking pattern for the attention mechanism. It takes four arguments: b (batch size), h (number of heads), q_idx (query index), and kv_idx (key/value index). It should return a boolean tensor indicating which attention connections are allowed (True) or masked out (False). B (int): Batch size. H (int): Number of query heads. Q_LEN (int): Sequence length of query. KV_LEN (int): Sequence length of key/value. device (str): Device to run the mask creation on. BLOCK_SIZE (int or tuple[int, int]): Block size for the block mask. If a single int is provided it is used for both query and key/value. Returns: BlockMask: A BlockMask object that contains the block mask information. Example Usage: .. code-block:: python def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx block_mask = create_block_mask(causal_mask, 1, 1, 8192, 8192, device="cuda") query = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) key = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) value = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) output = flex_attention(query, key, value, block_mask=block_mask) """ mod_type = _get_mod_type(mask_mod) assert ( mod_type == _ModificationType.MASK_MOD ), f"create-block_mask requires a mask_mod function! Got {mask_mod}" if B is None: B = 1 if H is None: H = 1 if isinstance(BLOCK_SIZE, int): Q_BLOCK_SIZE = BLOCK_SIZE KV_BLOCK_SIZE = BLOCK_SIZE else: Q_BLOCK_SIZE, KV_BLOCK_SIZE = BLOCK_SIZE if _compile: warnings.warn( "_compile flag on create_block_mask was originally added to work around a torch.compile limitation. That limitation has since been addressed. So, to compile create_block_mask, we suggest doing torch.compile(create_block_mask). This still works for now, but will be removed in the future.", DeprecationWarning, ) return torch.compile(create_block_mask)( mask_mod, B, H, Q_LEN, KV_LEN, device, BLOCK_SIZE ) mask_tensor = create_mask(mask_mod, B, H, Q_LEN, KV_LEN, device) partial_block_mask, full_block_mask = _convert_mask_to_block_mask( mask_tensor, Q_BLOCK_SIZE=Q_BLOCK_SIZE, KV_BLOCK_SIZE=KV_BLOCK_SIZE, separate_full_blocks=True, ) block_mask = _create_sparse_block_from_block_mask( (partial_block_mask, full_block_mask), mask_mod, (Q_LEN, KV_LEN), Q_BLOCK_SIZE, KV_BLOCK_SIZE, ) return block_mask def _create_empty_block_mask(query: Tensor, key: Tensor) -> BlockMask: r"""Default block mask for flex attention. If users don't specify any block sparse mask info, we create this empty block sparse mask. Which creates a BlockMask with 1 block that is the full length of the query and key tensors. """ device = query.device return BlockMask.from_kv_blocks( kv_num_blocks=torch.ones([1, 1, 1], dtype=torch.int32, device=device), kv_indices=torch.zeros([1, 1, 1, 1], dtype=torch.int32, device=device), BLOCK_SIZE=_LARGE_SPARSE_BLOCK_SIZE, seq_lengths=(1, 1), ) def _nested_mod_func_adapter( orig_mod_func: Union[_score_mod_signature, _mask_mod_signature], q_nt: torch.Tensor, kv_nt: torch.Tensor, is_score_mod: bool, ) -> Union[_score_mod_signature, _mask_mod_signature]: r"""Adapter to convert a score_mod / mask_mod to be NJT-compatible. The given mod func should be written as if operating over a single sequence at a item. This adapter will handle conversion from indices operating over a "stacked sequence" of length ``sum(S)`` for sequence length ``S`` in the NJT to "sequence relative" indices in range ``[0, S)``. Args: orig_mod_func (Callable): Function to modify attention scores. It takes four or five arguments, depending on whether a mask_mod or score_mod func is passed. q_nt (torch.Tensor): Jagged layout nested tensor (NJT) that defines the sequence length structure for query. kv_nt (torch.Tensor): Jagged layout nested tensor (NJT) that defines the sequence length structure for key / value. is_score_mod (bool): Indicates whether the mod function is a score_mod. Returns: nt_score_mod: An NJT-compatible version of orig_score_mod """ # Used to convert indices within the "stacked" sequence (range [0, sum(*))) # to "sequence local" indices (range [0, S) for each S). def _build_seq_idx(offsets, total_length): range_tensor = torch.arange( total_length, device=offsets.device, dtype=torch.int32 ) # Use searchsorted to find the index for each position # NB: This assumes offsets[0] to offsets[-1] spans the packed dim of values. # If we ever loosen this restriction, this logic will need to be updated. seq_idx = torch.searchsorted(offsets, range_tensor, right=True) - 1 return seq_idx q_offsets = q_nt._offsets # type: ignore[attr-defined] kv_offsets = kv_nt._offsets # type: ignore[attr-defined] q_seq_idx = _build_seq_idx(q_offsets, q_nt._values.shape[q_nt._ragged_idx - 1]) # type: ignore[attr-defined] if q_nt is kv_nt: kv_seq_idx = q_seq_idx else: # cross attention case kv_seq_idx = _build_seq_idx(kv_offsets, kv_nt._values.shape[kv_nt._ragged_idx - 1]) # type: ignore[attr-defined] # Converts q_idx / kv_idx from [0, total_length) -> [0, S), where S refers # to the sequence length for each sequence in the NJT, for use in given # score_mod. This allows the user to write a score_mod as if it were # operating on a single sequence and the "stacked sequence" is split # automatically into individual sequences for them. if is_score_mod: def nt_score_mod(score, b, h, q_idx, kv_idx): b_nested = q_seq_idx[q_idx] q_nested = q_idx - q_offsets[q_seq_idx[q_idx]] kv_nested = kv_idx - kv_offsets[kv_seq_idx[kv_idx]] is_same_sequence = q_seq_idx[q_idx] == kv_seq_idx[kv_idx] return torch.where( is_same_sequence, orig_mod_func(score, b_nested, h, q_nested, kv_nested), # type: ignore[call-arg] # don't allow inter-sequence attention float("-inf"), ) return nt_score_mod else: def nt_mask_mod(b, h, q_idx, kv_idx): b_nested = q_seq_idx[q_idx] q_nested = q_idx - q_offsets[q_seq_idx[q_idx]] kv_nested = kv_idx - kv_offsets[kv_seq_idx[kv_idx]] # don't allow inter-sequence attention is_same_sequence = q_seq_idx[q_idx] == kv_seq_idx[kv_idx] return orig_mod_func(b_nested, h, q_nested, kv_nested) & is_same_sequence # type: ignore[call-arg] return nt_mask_mod def create_nested_block_mask( mask_mod: _mask_mod_signature, B: Optional[int], H: Optional[int], q_nt: torch.Tensor, kv_nt: Optional[torch.Tensor] = None, BLOCK_SIZE: Union[int, tuple[int, int]] = _DEFAULT_SPARSE_BLOCK_SIZE, _compile=False, ) -> BlockMask: r"""This function creates a nested tensor compatible block mask tuple from a mask_mod function. The returned BlockMask will be on the device specified by the input nested tensor. Args: mask_mod (Callable): mask_mod function. This is a callable that defines the masking pattern for the attention mechanism. It takes four arguments: b (batch size), h (number of heads), q_idx (query index), and kv_idx (key/value index). It should return a boolean tensor indicating which attention connections are allowed (True) or masked out (False). B (int): Batch size. H (int): Number of query heads. q_nt (torch.Tensor): Jagged layout nested tensor (NJT) that defines the sequence length structure for query. The block mask will be constructed to operate on a "stacked sequence" of length ``sum(S)`` for sequence length ``S`` from the NJT. kv_nt (torch.Tensor): Jagged layout nested tensor (NJT) that defines the sequence length structure for key / value, allowing for cross attention. The block mask will be constructed to operate on a "stacked sequence" of length ``sum(S)`` for sequence length ``S`` from the NJT. If this is None, ``q_nt`` is used to define the structure for key / value as well. Default: None BLOCK_SIZE (int or tuple[int, int]): Block size for the block mask. If a single int is provided it is used for both query and key/value. Returns: BlockMask: A BlockMask object that contains the block mask information. Example Usage: .. code-block:: python # shape (B, num_heads, seq_len*, D) where seq_len* varies across the batch query = torch.nested.nested_tensor(..., layout=torch.jagged) key = torch.nested.nested_tensor(..., layout=torch.jagged) value = torch.nested.nested_tensor(..., layout=torch.jagged) def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx block_mask = create_nested_block_mask(causal_mask, 1, 1, query, _compile=True) output = flex_attention(query, key, value, block_mask=block_mask) .. code-block:: python # shape (B, num_heads, seq_len*, D) where seq_len* varies across the batch query = torch.nested.nested_tensor(..., layout=torch.jagged) key = torch.nested.nested_tensor(..., layout=torch.jagged) value = torch.nested.nested_tensor(..., layout=torch.jagged) def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx # cross attention case: pass both query and key/value NJTs block_mask = create_nested_block_mask(causal_mask, 1, 1, query, key, _compile=True) output = flex_attention(query, key, value, block_mask=block_mask) """ # use same structure for kv as for q by default if kv_nt is None: kv_nt = q_nt if q_nt.device != kv_nt.device: raise ValueError( "create_nested_block_mask(): Expected q_nt and kv_nt to be on the same device" ) return create_block_mask( _nested_mod_func_adapter(mask_mod, q_nt, kv_nt, is_score_mod=False), # type: ignore[arg-type] B, H, q_nt._values.shape[q_nt._ragged_idx - 1], # type: ignore[attr-defined] kv_nt._values.shape[kv_nt._ragged_idx - 1], # type: ignore[attr-defined] device=q_nt.device, # type: ignore[arg-type] # compile is important so we don't materialize a mask_tensor of # shape (1, 1, total_seqlen, total_seqlen) BLOCK_SIZE=BLOCK_SIZE, _compile=_compile, ) def _apply_kernel_options( query: Tensor, key: Tensor, value: Tensor, return_lse: bool, kernel_options ): kernel_options = {} if kernel_options is None else dict(kernel_options) kernel_options.setdefault("PRESCALE_QK", False) kernel_options.setdefault("ROWS_GUARANTEED_SAFE", False) kernel_options.setdefault("BLOCKS_ARE_CONTIGUOUS", False) # This forces all biases grad scatters to be done in the DQ iteration loop of the backwards kernel_options.setdefault("WRITE_DQ", True) # If forward kernel needs to return logsumexp is decided by this rule internally. assert "OUTPUT_LOGSUMEXP" not in kernel_options kernel_options["OUTPUT_LOGSUMEXP"] = True if not return_lse: # We used to check if q,k,v required grads but since captured buffers can require grad # we always write unless in no_grad output_logsumexp = torch.is_grad_enabled() kernel_options["OUTPUT_LOGSUMEXP"] = output_logsumexp any_inputs_on_cpu_device = ( query.device.type == "cpu" or key.device.type == "cpu" or value.device.type == "cpu" ) if any_inputs_on_cpu_device: # CPU with torch.compile now supports infernece, and will not return lse # TODO: support CPU for training and return lse kernel_options["OUTPUT_LOGSUMEXP"] = False return kernel_options def _validate_embed_dim(query: Tensor, key: Tensor, value: Tensor): if query.size(-1) != key.size(-1): raise ValueError( f"Expect query and key/value to have the same embedding dimension " f"but got E={query.size(-1)} and E={key.size(-1)}." ) return # TODO this config segfaults with Triton without: # https://github.com/triton-lang/triton/pull/4540 if not ( _supported_head_dim(query.size(-1)) and _supported_head_dim(value.size(-1)) ): raise ValueError( f"NYI: Currently non power of 2 embedding dimension are not supported. " f"Got E={query.size(-1)} and Ev={value.size(-1)}." ) def _validate_device(query: Tensor, key: Tensor, value: Tensor): """TODO: Remove once non cuda/cpu devices support is added We only need to check query since we have already that q,k,v are on the same device """ if query.device.type != "cuda" and query.device.type != "cpu": raise ValueError( "FlexAttention is only supported on CUDA or CPU devices. " f"Found input tensors on {query.device.type} device." ) def _validate_nestedness(query: Tensor, key: Tensor, value: Tensor): # Currently, inputs can only be all nested or no nested. if query.is_nested != key.is_nested or key.is_nested != value.is_nested: raise ValueError( "FlexAttention does not support mixed nested tensor / non-nested tensor inputs. " "Please file an issue requesting this if it is important to you." ) if ( (query.is_nested and query._lengths is not None) # type: ignore[attr-defined] or (key.is_nested and key._lengths is not None) # type: ignore[attr-defined] or (value.is_nested and value._lengths is not None) # type: ignore[attr-defined] ): raise ValueError( "FlexAttention does not support nested tensors that are non-contiguous with holes. " "Please file an issue requesting this if it is important to you." ) def flex_attention( query: Tensor, key: Tensor, value: Tensor, score_mod: Optional[_score_mod_signature] = None, block_mask: Optional[BlockMask] = None, scale: Optional[float] = None, enable_gqa: bool = False, return_lse: bool = False, kernel_options: Optional[dict[str, Any]] = None, ) -> Union[Tensor, tuple[Tensor, Tensor]]: r"""This function implements scaled dot product attention with an arbitrary attention score modification function. This function computes the scaled dot product attention between query, key, and value tensors with a user-defined attention score modification function. The attention score modification function will be applied after the attention scores have been calculated between the query and key tensors. The attention scores are calculated as follows: The ``score_mod`` function should have the following signature: .. code-block:: python def score_mod( score: Tensor, batch: Tensor, head: Tensor, q_idx: Tensor, k_idx: Tensor ) -> Tensor: Where: - ``score``: A scalar tensor representing the attention score, with the same data type and device as the query, key, and value tensors. - ``batch``, ``head``, ``q_idx``, ``k_idx``: Scalar tensors indicating the batch index, query head index, query index, and key/value index, respectively. These should have the ``torch.int`` data type and be located on the same device as the score tensor. Args: query (Tensor): Query tensor; shape :math:`(B, Hq, L, E)`. key (Tensor): Key tensor; shape :math:`(B, Hkv, S, E)`. value (Tensor): Value tensor; shape :math:`(B, Hkv, S, Ev)`. score_mod (Optional[Callable]): Function to modify attention scores. By default no score_mod is applied. block_mask (Optional[BlockMask]): BlockMask object that controls the blocksparsity pattern of the attention. scale (Optional[float]): Scaling factor applied prior to softmax. If none, the default value is set to :math:`\frac{1}{\sqrt{E}}`. enable_gqa (bool): If set to True, enables Grouped Query Attention (GQA) and broadcasts key/value heads to query heads. return_lse (bool): Whether to return the logsumexp of the attention scores. Default is False. kernel_options (Optional[Dict[str, Any]]): Options to pass into the Triton kernels. Returns: output (Tensor): Attention output; shape :math:`(B, Hq, L, Ev)`. Shape legend: - :math:`N: \text{Batch size} ... : \text{Any number of other batch dimensions (optional)}` - :math:`S: \text{Source sequence length}` - :math:`L: \text{Target sequence length}` - :math:`E: \text{Embedding dimension of the query and key}` - :math:`Ev: \text{Embedding dimension of the value}` .. warning:: `torch.nn.attention.flex_attention` is a prototype feature in PyTorch. Please look forward to a more stable implementation in a future version of PyTorch. Read more about feature classification at: https://pytorch.org/blog/pytorch-feature-classification-changes/#prototype """ # Some basic input validation _validate_sdpa_input(query, key, value) _validate_embed_dim(query, key, value) _validate_device(query, key, value) _validate_nestedness(query, key, value) if query.dim() != 4 or key.dim() != 4 or value.dim() != 4: raise NotImplementedError("NYI: query, key, and value must be 4D tensors") if (not enable_gqa) and query.size(-3) != key.size(-3): raise ValueError( f"Expect query and key/value to have the same number of heads " f"but got Hq={query.size(-3)} and Hkv={key.size(-3)}. " f"Try setting enable_gqa=True for GQA." ) if enable_gqa: Hq = query.size(1) Hkv = key.size(1) if Hq % Hkv != 0: raise ValueError( f"Expect number of query heads to be a multiple of kv heads for GQA " f"but got Hq={Hq} and Hkv={Hkv}." ) if query.size(0) != key.size(0): if block_mask is None: raise ValueError( f"Expect query and key/value to have the same batch size, " f"or non-none block_mask, " f"but got block_mask=None, Bq={query.size(0)}, and Bkv={key.size(0)}." ) if block_mask.kv_num_blocks.size(0) != query.size(0): raise ValueError( f"Expect query and key/value to have the same batch size, " f"or block_mask and query to have the same batch size, " f"but got Bq={query.size(0)}, Bkv={key.size(0)}, B_block_mask={block_mask.kv_num_blocks.size(0)}." ) if score_mod is None: score_mod = _identity elif query.is_nested: # use same NJT if the ragged structures for sequence lengths match between q and kv kv = ( query if query.size(query._ragged_idx) == key.size(query._ragged_idx) # type: ignore[attr-defined] else key ) score_mod = _nested_mod_func_adapter(score_mod, query, kv, is_score_mod=True) # type: ignore[assignment] if block_mask is None: block_mask = _create_empty_block_mask(query, key) if ( block_mask.BLOCK_SIZE[0] == _LARGE_SPARSE_BLOCK_SIZE and block_mask.BLOCK_SIZE[1] == _LARGE_SPARSE_BLOCK_SIZE ): # This corresponds to the case where we essentially have a "no-op" block mask. pass elif query.is_nested: if block_mask.shape[-2] != query._values.size(query._ragged_idx - 1): # type: ignore[attr-defined] raise RuntimeError( f"block_mask of shape {block_mask.shape} is not compatible with nested tensor input " f"with total sequence length of {query._values.size(query._ragged_idx - 1)}" # type: ignore[attr-defined] ) else: block_mask_q_len = block_mask.shape[-2] block_mask_kv_len = block_mask.shape[-1] if query.size(-2) > block_mask_q_len or key.size(-2) > block_mask_kv_len: raise ValueError( f"block_mask was created for block_mask.shape={block_mask.shape} but got q_len={query.size(-2)} and kv_len={key.size(-2)}. " "As the block mask was created for a smaller length than you're using it for, you likely need to create a new block mask." ) elif ( query.size(-2) < block_mask_q_len and key.size(-2) <= block_mask_kv_len ) or (query.size(-2) <= block_mask_q_len and key.size(-2) < block_mask_kv_len): raise ValueError( f"block_mask was created for block_mask.shape={block_mask.shape} but got q_len={query.size(-2)} and kv_len={key.size(-2)}. " "As the block mask was created for a larger length than you're using it for, you can either 1. create a new block mask with the correct length, or 2. 'adjust' the existing block mask to the correct length by calling block_mask._adjust(q_len, kv_len). This essentially 'crops' the block mask to the upper left corner, which does not work for all mask_mods!" ) assert query.size(-2) == block_mask_q_len assert key.size(-2) == block_mask_kv_len if scale is None: scale = 1.0 / math.sqrt(query.size(-1)) if query.device != block_mask.kv_num_blocks.device: # type: ignore[union-attr] raise RuntimeError( f"Expect q/k/v and block_mask to be on the same device " f"but got {query.device} and {block_mask.kv_num_blocks.device}." # type: ignore[union-attr] ) kernel_options = _apply_kernel_options( query, key, value, return_lse, kernel_options, ) if torch.compiler.is_dynamo_compiling(): # mark head_dim and number of heads to be static for x in [query, key, value]: torch._dynamo.mark_static(x, -3) torch._dynamo.mark_static(x, -1) out, lse = flex_attention_hop( query, key, value, score_mod, block_mask.as_tuple(), scale, kernel_options # type: ignore[union-attr] ) if return_lse: return out, lse * math.log(2) else: return out if not torch._dynamo.is_dynamo_supported(): raise RuntimeError("flex_attention requires dynamo support") from torch._dynamo.backends.debugging import ( make_eager_backend_with_torch_function_mode, ) # Dynamo is expecting a callable with "__code__" attribute. # We cannot directly pass hop to it. So we wrap it in a dummy function. def _flex_attention_hop_wrapper(*args, **kwargs): return flex_attention_hop(*args, **kwargs) with _set_compilation_env(): with torch._dynamo.utils.disable_cache_limit(): with _temp_remove_pre_dispatch_torch_function_mode(): with _temp_remove_metadata_torch_function_mode() as metadata_mode: if metadata_mode: backend = make_eager_backend_with_torch_function_mode( metadata_mode ) else: backend = "eager" out, lse = torch.compile( _flex_attention_hop_wrapper, backend=backend, fullgraph=True )( query, key, value, score_mod, block_mask.as_tuple(), # type: ignore[union-attr] scale, kernel_options, ) if return_lse: return out, lse * math.log(2) else: return out