# Copyright 2025 HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import itertools from collections.abc import Callable import torch import torch.nn.functional as F from .cache_utils import Cache from .configuration_utils import PreTrainedConfig from .utils import is_torch_xpu_available, logging from .utils.deprecation import deprecate_kwarg from .utils.generic import GeneralInterface, is_flash_attention_requested from .utils.import_utils import is_torch_flex_attn_available, is_torch_greater_or_equal, is_tracing if is_torch_flex_attn_available(): from torch.nn.attention.flex_attention import _DEFAULT_SPARSE_BLOCK_SIZE as flex_default_block_size from torch.nn.attention.flex_attention import BlockMask, create_block_mask else: # Register a fake type to avoid crashing for annotations and `isinstance` checks BlockMask = torch.Tensor _is_torch_greater_or_equal_than_2_5 = is_torch_greater_or_equal("2.5", accept_dev=True) _is_torch_greater_or_equal_than_2_6 = is_torch_greater_or_equal("2.6", accept_dev=True) _is_torch_xpu_available = is_torch_xpu_available() if _is_torch_greater_or_equal_than_2_6: from torch._dynamo._trace_wrapped_higher_order_op import TransformGetItemToIndex logger = logging.get_logger(__name__) def and_masks(*mask_functions: Callable) -> Callable: """Returns a mask function that is the intersection of provided mask functions""" if not all(callable(arg) for arg in mask_functions): raise RuntimeError(f"All inputs should be callable mask_functions: {mask_functions}") def and_mask(batch_idx, head_idx, q_idx, kv_idx): result = q_idx.new_ones((), dtype=torch.bool) for mask in mask_functions: result = result & mask(batch_idx, head_idx, q_idx, kv_idx).to(result.device) return result return and_mask def or_masks(*mask_functions: Callable) -> Callable: """Returns a mask function that is the union of provided mask functions""" if not all(callable(arg) for arg in mask_functions): raise RuntimeError(f"All inputs should be callable mask_functions: {mask_functions}") def or_mask(batch_idx, head_idx, q_idx, kv_idx): result = q_idx.new_zeros((), dtype=torch.bool) for mask in mask_functions: result = result | mask(batch_idx, head_idx, q_idx, kv_idx).to(result.device) return result return or_mask def causal_mask_function(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: """ This creates a basic lower-diagonal causal mask. """ return kv_idx <= q_idx def bidirectional_mask_function(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: """ This creates a full bidirectional mask. NOTE: It is important to keep an index-based version for non-vmap expansion. """ return q_idx >= 0 def sliding_window_overlay(sliding_window: int) -> Callable: """ This is an overlay depicting a sliding window pattern. Add it on top of a causal mask for a proper sliding window mask. """ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: return kv_idx > q_idx - sliding_window return inner_mask def chunked_overlay(chunk_size: int, left_padding: torch.Tensor) -> Callable: """ This is an overlay depicting a chunked attention pattern. Add it on top of a causal mask for a proper chunked attention mask. """ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: return (kv_idx - left_padding[batch_idx]) // chunk_size == (q_idx - left_padding[batch_idx]) // chunk_size return inner_mask def sliding_window_causal_mask_function(sliding_window: int) -> Callable: """ This return the mask_function function to create a sliding window mask. """ return and_masks(sliding_window_overlay(sliding_window), causal_mask_function) def sliding_window_bidirectional_overlay(sliding_window: int) -> Callable: """ This is an overlay depicting a bidirectional sliding window pattern. """ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: """A token can attend to any other token if their absolute distance is within the (inclusive) sliding window size (distance <= sliding_window).""" return abs(q_idx - kv_idx) <= sliding_window return inner_mask def sliding_window_bidirectional_mask_function(sliding_window: int) -> Callable: """ This return the mask_function function to create a bidirectional sliding window mask. """ return and_masks(sliding_window_bidirectional_overlay(sliding_window), bidirectional_mask_function) def chunked_causal_mask_function(chunk_size: int, left_padding: torch.Tensor) -> Callable: """ This return the mask_function function to create a chunked attention mask. """ return and_masks(chunked_overlay(chunk_size, left_padding), causal_mask_function) def padding_mask_function(padding_mask: torch.Tensor) -> Callable: """ This return the mask_function function corresponding to a 2D padding mask. """ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: # Note that here the mask should ALWAYS be at least of the max `kv_index` size in the dimension 1. This is because # we cannot pad it here in the mask_function as we don't know the final size, and we cannot try/except, as it is not # vectorizable on accelerator devices return padding_mask[batch_idx, kv_idx] return inner_mask def packed_sequence_mask_function(packed_sequence_mask: torch.Tensor) -> Callable: """ This return the mask_function function corresponding to a 2D packed sequence mask. """ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: return packed_sequence_mask[batch_idx, q_idx] == packed_sequence_mask[batch_idx, kv_idx] return inner_mask def add_offsets_to_mask_function(mask_function: Callable, q_offset: int, kv_offset: int) -> Callable: """ This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths, not start and end indices. """ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: return mask_function(batch_idx, head_idx, q_idx + q_offset, kv_idx + kv_offset) return inner_mask def prepare_padding_mask(attention_mask: torch.Tensor | None, kv_length: int, kv_offset: int) -> torch.Tensor | None: """ From the 2D attention mask, prepare the correct padding mask to use by potentially padding it. """ local_padding_mask = attention_mask if attention_mask is not None: # Pad it if necessary if (padding_length := kv_length + kv_offset - attention_mask.shape[-1]) > 0: local_padding_mask = torch.nn.functional.pad(attention_mask, (0, padding_length)) return local_padding_mask def _can_skip_causal_mask_xpu( padding_mask: torch.Tensor | None, query_length: int, kv_length: int, local_attention_size: int | None, ) -> bool: """ XPU-specific logic for determining if we can skip causal mask creation. For XPU devices, we have special handling: - Single query tokens (query_length == 1) use the same logic as CUDA - Multi-query tokens can skip if padding_mask is provided and correctly structured The mask must have all True values in the query window and all False after """ if is_tracing(padding_mask): return False # Check local attention constraint (same as CUDA) if local_attention_size is not None and kv_length >= local_attention_size: return False if padding_mask is None: # Without padding mask, can skip if single query token or full causal attention return query_length == 1 or kv_length == query_length # XPU allows skipping under additional conditions when padding_mask is provided if query_length == 1: # Single query token: skip only if no padding tokens present return padding_mask.all() # XPU-specific: check if query window is all True and rest is all False # This allows XPU to optimize the 1st token in static cache return padding_mask[:, :query_length].all() and not padding_mask[:, query_length:].any() def _ignore_causal_mask_sdpa( padding_mask: torch.Tensor | None, query_length: int, kv_length: int, kv_offset: int, local_attention_size: int | None = None, ) -> bool: """ Detects whether the causal mask can be ignored in case PyTorch's SDPA is used, rather relying on SDPA's `is_causal` argument. In case no token is masked in the 2D `padding_mask` argument, if `query_length == 1` or `key_value_length == query_length`, we rather rely on SDPA `is_causal` argument to use causal/non-causal masks, allowing to dispatch to the flash attention kernel (that can otherwise not be used if a custom `attn_mask` is passed). """ if padding_mask is not None and padding_mask.shape[-1] > kv_length: mask_indices = torch.arange(kv_length, device=padding_mask.device) mask_indices += kv_offset padding_mask = padding_mask[:, mask_indices] if _is_torch_xpu_available: # XPU devices have special handling for mask skipping: # - Single query tokens use the same logic as CUDA # - Multi-query tokens can skip if padding_mask is provided and correctly structured # (all True in query window, all False after) return _can_skip_causal_mask_xpu(padding_mask, query_length, kv_length, local_attention_size) # When using `torch.export` or `torch.onnx.dynamo_export`, we must pass an example input, and `is_causal` behavior is # hard-coded to the forward. If a user exports a model with query_length > 1, the exported model will hard-code `is_causal=True` # which is in general wrong (see https://github.com/pytorch/pytorch/issues/108108). Thus, we only set # `ignore_causal_mask = True` if we are not tracing if ( not is_tracing(padding_mask) # only cases when lower and upper diags are the same, see https://github.com/pytorch/pytorch/issues/108108 and (query_length == 1 or kv_length == query_length) # in this case we need to add special patterns to the mask so cannot be skipped otherwise and (local_attention_size is None or kv_length < local_attention_size) # In this case, we need to add padding to the mask, so cannot be skipped otherwise and (padding_mask is None or padding_mask.all()) ): return True return False def _can_skip_bidirectional_mask_xpu( padding_mask: torch.Tensor | None, kv_length: int, local_attention_size: int | None, ) -> bool: """ XPU-specific logic for determining if we can skip bidirectional mask creation. For XPU devices, we have special handling: - Skip if no padding and no local attention constraint """ if is_tracing(padding_mask): return False # Check local attention constraint (same as CUDA) if local_attention_size is not None and kv_length >= local_attention_size: return False if padding_mask is None: # Without padding mask, can always skip for full bidirectional attention return True # Skip only if no padding tokens present return padding_mask.all() def _ignore_bidirectional_mask_sdpa( padding_mask: torch.Tensor | None, kv_length: int, local_attention_size: int | None = None, ) -> bool: """ Detects whether the bidirectional mask can be ignored in case PyTorch's SDPA is used. In case no token is masked in the 2D `padding_mask` argument and no local attention constraint applies (i.e. `local_attention_size` is None or `kv_length < local_attention_size`), we skip mask creation, allowing to dispatch to the flash attention kernel (that can otherwise not be used if a custom `attn_mask` is passed). """ if _is_torch_xpu_available: # XPU devices have special handling for mask skipping: # - Skip if no padding and no local attention constraint return _can_skip_bidirectional_mask_xpu(padding_mask, kv_length, local_attention_size) # When using `torch.export` or `torch.onnx.dynamo_export`, we need to avoid to check the contents of the mask; # otherwise, we will encounter dynamic control flows if ( not is_tracing(padding_mask) and (padding_mask is None or padding_mask.all()) # in this case we need to add special patterns to the mask so cannot be skipped otherwise and (local_attention_size is None or kv_length < local_attention_size) ): return True return False def _vmap_expansion_sdpa(mask_function: Callable) -> Callable: """ Used to vmap our mask_functions over the all 4 dimensions (b_idx, h_idx, q_idx, kv_idx) of the inputs. Using vmap here allows us to keep the performance of vectorized ops, while having a single set of primitive functions between attention interfaces (i.e. between flex and sdpa/eager, FA2 being a bit different). """ # We vmap the function over all 4 dimensions, broadcasting [b_idx, h_idx, q_idx, kv_idx] dimensions = [(None, None, None, 0), (None, None, 0, None), (None, 0, None, None), (0, None, None, None)] for dims in dimensions: mask_function = torch.vmap(mask_function, in_dims=dims, out_dims=0) return mask_function def _non_vmap_expansion_sdpa( batch_indices: torch.Tensor, head_indices: torch.Tensor, q_indices: torch.Tensor, kv_indices: torch.Tensor ): """ Used to broadcast our mask_functions over the all 4 dimensions (b_idx, h_idx, q_idx, kv_idx) of the inputs. Allows the usage of any index-based mask function without relying on vmap. NOTE: This is limited to index based functions only and is not guaranteed to work otherwise. Reference: - https://github.com/huggingface/optimum-onnx/blob/c123e8f4fab61b54a8e0e31ce74462bcacca576e/optimum/exporters/onnx/model_patcher.py#L362-L365 """ batch_indices = batch_indices[:, None, None, None] head_indices = head_indices[None, :, None, None] q_indices = q_indices[None, None, :, None] kv_indices = kv_indices[None, None, None, :] return batch_indices, head_indices, q_indices, kv_indices def sdpa_mask( batch_size: int, cache_position: torch.Tensor, kv_length: int, kv_offset: int = 0, mask_function: Callable = causal_mask_function, attention_mask: torch.Tensor | None = None, local_size: int | None = None, allow_is_causal_skip: bool = True, allow_is_bidirectional_skip: bool = False, allow_torch_fix: bool = True, use_vmap: bool = False, **kwargs, ) -> torch.Tensor | None: """ Create a 4D boolean mask of shape `(batch_size, 1, query_length, kv_length)` where a value of True indicates that the element should take part in the attention computation, and False that it should not. This function can only be used with torch>=2.5, as the context manager is otherwise not available. Args: batch_size (`int`): The batch size of the input sequence. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. kv_length (`int`): The size that the key and value states will have during the attention computation. kv_offset (`int`, optional): An optional offset to indicate at which first position the key and values states will refer to. mask_function (`Callable`): The mask factory function describing the mask pattern. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length) local_size (`int`, optional): The size of the local attention, if we do not use full attention. This is used only if `allow_is_causal_skip=True` to try to skip mask creation if possible. allow_is_causal_skip (`bool`, optional): Whether to allow to return `None` for the mask under conditions where we can use the `is_causal` argument in `torch.sdpa` instead. Default to `True`. allow_is_bidirectional_skip (`bool`, optional): Whether to allow to return `None` for the mask under conditions where we do not have to add any bias, i.e. full attention without any padding. Default to `False`. allow_torch_fix (`bool`, optional): Whether to update the mask in case a query is not attending to any tokens, to solve a bug in torch's older versions. We need an arg to skip it when using eager. By default `True`. use_vmap (`bool`, optional): Whether to use `vmap` during the mask construction or not. Allows powerful custom patterns that may not be index-based (for the cost of speed performance). By default `False`. ## Creating a simple causal mask: To create the following causal mask: 0 ■ ⬚ ⬚ ⬚ ⬚ 1 ■ ■ ⬚ ⬚ ⬚ 2 ■ ■ ■ ⬚ ⬚ 3 ■ ■ ■ ■ ⬚ 4 ■ ■ ■ ■ ■ You can do ```python >>> sdpa_mask(batch_size=1, cache_position=torch.arange(5), kv_length=5) >>> tensor([[[[ True, False, False, False, False], [ True, True, False, False, False], [ True, True, True, False, False], [ True, True, True, True, False], [ True, True, True, True, True]]]]) ``` ## Creating a sliding window mask: To create the following sliding window mask (`sliding_window=3`): 0 ■ ⬚ ⬚ ⬚ ⬚ 1 ■ ■ ⬚ ⬚ ⬚ 2 ■ ■ ■ ⬚ ⬚ 3 ⬚ ■ ■ ■ ⬚ 4 ⬚ ⬚ ■ ■ ■ You can do ```python >>> sdpa_mask(batch_size=1, cache_position=torch.arange(5), kv_length=5, mask_function=sliding_window_causal_mask_function(3)) >>> tensor([[[[ True, False, False, False, False], [ True, True, False, False, False], [ True, True, True, False, False], [False, True, True, True, False], [False, False, True, True, True]]]]) ``` ## Creating a chunked attention mask To create the following chunked attention mask (`chunk_size=3`): 0 ■ ⬚ ⬚ ⬚ ⬚ 1 ■ ■ ⬚ ⬚ ⬚ 2 ■ ■ ■ ⬚ ⬚ 3 ⬚ ⬚ ⬚ ■ ⬚ 4 ⬚ ⬚ ⬚ ■ ■ You can do ```python >>> sdpa_mask(batch_size=1, cache_position=torch.arange(5), kv_length=5, mask_function=chunked_causal_mask_function(3, torch.zeros(1, dtype=int))) >>> tensor([[[[ True, False, False, False, False], [ True, True, False, False, False], [ True, True, True, False, False], [False, False, False, True, False], [False, False, False, True, True]]]]) ``` """ q_length = cache_position.shape[0] # Potentially pad the 2D mask padding_mask = prepare_padding_mask(attention_mask, kv_length, kv_offset) # Under specific conditions, we can avoid materializing the mask # 1. Causal masks can rely on the `is_causal` argument # 2. Bidirectional do not need any further processing (no bias) if allow_is_causal_skip and _ignore_causal_mask_sdpa(padding_mask, q_length, kv_length, kv_offset, local_size): return None if allow_is_bidirectional_skip and _ignore_bidirectional_mask_sdpa(padding_mask, kv_length, local_size): return None # Potentially add the padding 2D mask if padding_mask is not None: mask_function = and_masks(mask_function, padding_mask_function(padding_mask)) batch_arange = torch.arange(batch_size, device=cache_position.device) head_arange = torch.arange(1, device=cache_position.device) # Similar to `kv_arange = torch.arange(start=kv_offset, end=kv_offset + kv_length, device=cache_position.device)` # but without data-dependent slicing (i.e. torch.compile friendly) kv_arange = torch.arange(kv_length, device=cache_position.device) + kv_offset # Actual mask creation # Option 1: Fast non-vmap mask creation (default) if not use_vmap: # Apply mask function element-wise through broadcasting attention_mask = mask_function(*_non_vmap_expansion_sdpa(batch_arange, head_arange, cache_position, kv_arange)) # Expand the mask to match batch size and query length if they weren't used in the mask function attention_mask = attention_mask.expand(batch_size, -1, q_length, kv_length) # Option 2: Vmap mask creation (torch>=2.6 and custom patterns) elif _is_torch_greater_or_equal_than_2_6: # This creates the 4D mask easily. Note that we need this context manager as vmap cannot handle slicing a tensor from # scalar tensor (it internally calls `.item()` which vmap does not allow, but this context works around it # We don't need to add an offset to the mask_function either, as we vmap directly the correct indices for k and kv indices with TransformGetItemToIndex(): attention_mask = _vmap_expansion_sdpa(mask_function)(batch_arange, head_arange, cache_position, kv_arange) # Option 3: Error out since it indicates that the user did something custom, which they shouldn't have (torch<2.6) else: raise ValueError( "The vmap functionality for mask creation is only supported from torch>=2.6. " "Please update your torch version or use `use_vmap=False` with index-based masks." ) # Due to a bug in versions of torch<2.5, we need to update the mask in case a query is not attending to any # tokens (due to padding). See details in https://github.com/pytorch/pytorch/issues/110213 if not _is_torch_greater_or_equal_than_2_5 and allow_torch_fix: attention_mask = attention_mask | torch.all(~attention_mask, dim=-1, keepdim=True) return attention_mask def eager_mask( batch_size: int, cache_position: torch.Tensor, kv_length: int, kv_offset: int = 0, mask_function: Callable = causal_mask_function, attention_mask: torch.Tensor | None = None, dtype: torch.dtype = torch.float32, allow_is_bidirectional_skip: bool = False, use_vmap: bool = False, **kwargs, ) -> torch.Tensor: """ Create a 4D float mask of shape `(batch_size, 1, query_length, kv_length)` where a value of 0 indicates that the element should take part in the attention computation, and -inf (minimum value for the given `dtype`) that it should not. Args: batch_size (`int`): The batch size of the input sequence. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. kv_length (`int`): The size that the key and value states will have during the attention computation. kv_offset (`int`, optional): An optional offset to indicate at which first position the key and values states will refer to. mask_function (`Callable`): The mask factory function describing the mask pattern. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length) dtype (`torch.dtype`, optional): The dtype to use for the mask. By default, `torch.float32`. allow_is_bidirectional_skip (`bool`, optional): Whether to allow to return `None` for the mask under conditions where we do not have to add any bias, i.e. full attention without any padding. Default to `False`. use_vmap (`bool`, optional): Whether to use `vmap` during the mask construction or not. Allows powerful custom patterns that may not be index-based (for the cost of speed performance). By default `False`. """ # The masks for eager attention are simply boolean mask from sdpa, casted to 0 and -inf _ = kwargs.pop("allow_is_causal_skip", None) _ = kwargs.pop("allow_torch_fix", None) mask = sdpa_mask( batch_size=batch_size, cache_position=cache_position, kv_length=kv_length, kv_offset=kv_offset, mask_function=mask_function, attention_mask=attention_mask, allow_is_causal_skip=False, allow_is_bidirectional_skip=allow_is_bidirectional_skip, allow_torch_fix=False, use_vmap=use_vmap, **kwargs, ) # only bidirectional masks can be skipped, otherwise we convert bool -> float if mask is not None: min_dtype = torch.finfo(dtype).min # we need 0s where the tokens should be taken into account, and -inf otherwise (mask is already of boolean type) mask = torch.where(mask, torch.tensor(0.0, device=mask.device, dtype=dtype), min_dtype) return mask def flash_attention_mask( batch_size: int, cache_position: torch.Tensor, kv_length: int, kv_offset: int = 0, mask_function: Callable = causal_mask_function, attention_mask: torch.Tensor | None = None, **kwargs, ): """ Create the attention mask necessary to use FA2. Since FA2 is un-padded by definition, here we simply return `None` if the mask is fully causal, or we return the 2D mask which will then be used to extract the seq_lens. We just slice it in case of sliding window. Args: batch_size (`int`): The batch size of the input sequence. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. kv_length (`int`): The size that the key and value states will have during the attention computation. kv_offset (`int`, optional): An optional offset to indicate at which first position the key and values states will refer to. mask_function (`Callable`): The mask factory function describing the mask pattern. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length) """ if attention_mask is not None: # Here we need to slice from the right if using sliding or chunked (for full attention, this is equivalent to doing nothing) attention_mask = attention_mask[:, -kv_length:] # We only return an actual mask if there is at least 1 padding token, otherwise we return `None` and use `is_causal` in FA2 # (note that the attention_mask is a boolean dtype here) if attention_mask.all(): attention_mask = None return attention_mask def flex_attention_mask( batch_size: int, cache_position: torch.Tensor, kv_length: int, kv_offset: int = 0, mask_function: Callable = causal_mask_function, attention_mask: torch.Tensor | None = None, **kwargs, ) -> BlockMask: """ Create a 4D block mask which is a compressed representation of the full 4D block causal mask. BlockMask is essential for performant computation of flex attention. See: https://pytorch.org/blog/flexattention/ Args: batch_size (`int`): The batch size of the input sequence. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. kv_length (`int`): The size that the key and value states will have during the attention computation. kv_offset (`int`, optional): An optional offset to indicate at which first position the key and values states will refer to. mask_function (`Callable`): The mask factory function describing the mask pattern. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length) """ q_length, q_offset = cache_position.shape[0], cache_position[0] # Potentially add the padding 2D mask if attention_mask is not None: # Older torch (2.5.x) cannot handle sequences not in multiples of 128 (default block size) # Hence we pad to multiples of this as a minimum to ensure this pad_len = ((attention_mask.shape[1] // flex_default_block_size) + 1) * flex_default_block_size pad_len = pad_len - attention_mask.shape[1] if not _is_torch_greater_or_equal_than_2_6 and pad_len > 0: attention_mask = torch.nn.functional.pad(attention_mask, value=0, pad=(0, pad_len)) padding_mask = prepare_padding_mask(attention_mask, kv_length, kv_offset) mask_function = and_masks(mask_function, padding_mask_function(padding_mask)) # Add the offsets on top (because flex interface only allows length, not start and end indices) mask_function = add_offsets_to_mask_function(mask_function, q_offset, kv_offset) # Finally create the block mask block_mask = create_block_mask( mask_mod=mask_function, B=batch_size, H=None, Q_LEN=q_length, KV_LEN=kv_length, device=cache_position.device, _compile=_is_torch_greater_or_equal_than_2_6, ) return block_mask class AttentionMaskInterface(GeneralInterface): # Class instance object, so that a call to `register` can be reflected into all other files correctly, even if # a new instance is created (in order to locally override a given function) _global_mapping = { "sdpa": sdpa_mask, "eager": eager_mask, "flash_attention_2": flash_attention_mask, "flash_attention_3": flash_attention_mask, "flex_attention": flex_attention_mask, } # Global AttentionMaskInterface shared by all models which do not need to overwrite any of the existing ones ALL_MASK_ATTENTION_FUNCTIONS: AttentionMaskInterface = AttentionMaskInterface() def find_packed_sequence_indices(position_ids: torch.Tensor) -> torch.Tensor | None: """ Find the indices of the sequence to which each new query token in the sequence belongs when using packed tensor format (i.e. several sequences packed in the same batch dimension). Args: position_ids (`torch.Tensor`) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. Returns: A 2D tensor where each similar integer indicates that the tokens belong to the same sequence. For example, if we pack 3 sequences of 2, 3 and 1 tokens respectively along a single batch dim, this will return [[0, 0, 1, 1, 1, 2]]. If the there is only one sequence in each batch item (and we don't compile), then we return `None` indicating no packed sequences. This is the same as [[0, 0, 0, 0, 0, 0]] for the example above. """ # What separate different sequences is when 2 consecutive positions_ids are separated by more than 1. So # taking the diff (by prepending the first value - 1 to keep correct indexing) and applying cumsum to the result # gives exactly the sequence indices # Note that we assume that a single sequence cannot span several batch dimensions, i.e. 1 single sequence # cannot be part of the end of the first batch dim and the start of the 2nd one for example first_dummy_value = position_ids[:, :1] - 1 # We just need the diff on this first value to be 1 position_diff = torch.diff(position_ids, prepend=first_dummy_value, dim=-1) packed_sequence_mask = (position_diff != 1).cumsum(-1) # Sadly this is a dynamic control flow, so we cannot enable this check on anything compile related if not is_tracing(packed_sequence_mask) and (packed_sequence_mask[:, -1] == 0).all(): return None return packed_sequence_mask @deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds") def _preprocess_mask_arguments( config: PreTrainedConfig, inputs_embeds: torch.Tensor, attention_mask: torch.Tensor | BlockMask | None, cache_position: torch.Tensor, past_key_values: Cache | None, position_ids: torch.Tensor | None, layer_idx: int | None, ) -> tuple[bool, torch.Tensor | BlockMask | None, int, int]: """ Perform some common pre-processing of the mask arguments we get from the modeling code. Mostly determine the key-value length and offsets, and if we should early exit or not. Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. layer_idx (`int`, optional): If `past_key_values` is not None, this is the layer index of the cache from which to get the key-value length and offset. Indeed, for hybrid caches, different layers may return different lengths. Returns: early_exit (`bool`): Whether we should early exit mask creation, and return the mask as-is. attention_mask (`torch.Tensor` or `BlockMask` or `None`): The attention mask to either return immediately, or to use in downstream mask creation. packed_sequence_mask (`torch.Tensor`, optional): In case we detected packed sequence format, this is a tensor where each similar integer indicates that the tokens belong to the same sequence. kv_length (`int`): The size that the key and value states will have during the attention computation. kv_offset (`int`): An offset to indicate at which first position the key and values states will refer to. """ # If the mask is already 4D, simply return as-is (it was already prepared, or it is custom) if isinstance(attention_mask, (torch.Tensor, BlockMask)) and len(attention_mask.shape) == 4: return True, attention_mask, None, None, None # For TGI/vLLM backends, or other custom attention without equivalent mask creation: we don't need a mask! # Note: it's not ideal to check the `_global_mapping` attribute instead of the object itself, however otherwise # full graph dynamo tracing (i.e. torch.export or compile with `fullgraph=True`) will fail on Python<3.11 # with `torch._dynamo.exc.Unsupported: 'inline in skipfiles:Mapping.__contains__ | __contains__, skipped # according trace_rules.lookup SKIP_DIRS'` -- can be removed when we require Python>=3.11 if config._attn_implementation not in ALL_MASK_ATTENTION_FUNCTIONS._global_mapping: return True, None, None, None, None # Move the mask to correct device, and potentially switch dtype for efficiency if attention_mask is not None and attention_mask.ndim == 2: attention_mask = attention_mask.to(device=cache_position.device, dtype=torch.bool) # If using a cache, it can give all information about mask sizes based on seen tokens if past_key_values is not None: kv_length, kv_offset = past_key_values.get_mask_sizes(cache_position, layer_idx) # Otherwise, we infer based on our input else: # 1. Rely on input directly if attention_mask is None: kv_length, kv_offset = inputs_embeds.shape[1], 0 # 2. Rely on the mask instead - needed for special cases like prefix tuning in PEFT # # This is a very unique and special case where an encoder utilizes a cache and expects its length # to be accounted for (usually, they should never use a cache). In general, the mask should always # match with the input sizes nonetheless (i.e. it does not affect others). # Conclusion: "prefix tuning is evil" else: kv_length, kv_offset = attention_mask.shape[-1], 0 # We check the position_ids for potential packed sequence format (only if the 2D attention mask is explicitly None, # and we don't have past_key_values, i.e. generally a training setup) packed_sequence_mask = None if position_ids is not None and attention_mask is None and past_key_values is None: batch_size = inputs_embeds.shape[0] # The position ids are sometimes just unsqueezed, without being expanded if batch_size != position_ids.shape[0]: position_ids = position_ids.expand(batch_size, -1) packed_sequence_mask = find_packed_sequence_indices(position_ids) return False, attention_mask, packed_sequence_mask, kv_length, kv_offset @deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds") def create_causal_mask( config: PreTrainedConfig, inputs_embeds: torch.Tensor, attention_mask: torch.Tensor | None, cache_position: torch.Tensor, past_key_values: Cache | None, position_ids: torch.Tensor | None = None, or_mask_function: Callable | None = None, and_mask_function: Callable | None = None, ) -> torch.Tensor | BlockMask | None: """ Create a standard causal mask based on the attention implementation used (stored in the config). If `past_key_values` has an hybrid cache structure, this function will return the mask corresponding to one of the "full_attention" layers (to align to what is needed in the `modeling_xxx.py` files). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. or_mask_function (`Callable`, optional): An optional mask function to combine with the causal mask function (by doing the union of both). This is useful to easily overlay another mask on top of the causal one, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the causal mask function (by doing the intersection of both). This is useful to easily overlay another mask on top of the causal one, for example for image tokens handling. """ # Power feature: if `is_causal` is False, then fallback to bi-directional mask for bi-directional attention. # It allows to use decoder-only models with bi-directional attention as well if not getattr(config, "is_causal", True): return create_bidirectional_mask( config, inputs_embeds, attention_mask, or_mask_function=or_mask_function, and_mask_function=and_mask_function, ) # If we have an hybrid cache structure, here we want to create the mask for the full layers if hasattr(past_key_values, "is_sliding") and False in past_key_values.is_sliding: layer_idx = past_key_values.is_sliding.index(False) else: layer_idx = 0 early_exit, attention_mask, packed_sequence_mask, kv_length, kv_offset = _preprocess_mask_arguments( config, inputs_embeds, attention_mask, cache_position, past_key_values, position_ids, layer_idx ) if early_exit: return attention_mask batch_size, dtype = inputs_embeds.shape[0], inputs_embeds.dtype mask_factory_function = causal_mask_function mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation] # Defaulting to using non-vmap based mask creations except when detecting # users passing custom mask functions (as we cannot guarantee that they # are properly index-based as required by our implementation). use_vmap = False # Do not allow skip if we are compiling (this is to match BC) # TODO: cyril -> probably revisit and remove this, but a lot of tests rely on it if _is_torch_xpu_available: # Do not allow skip if we are compiling for decoding, but for prefill, we still allow skip to optimization the perf of 1st token generation allow_is_causal_skip = not (getattr(past_key_values, "is_compileable", False) and cache_position.shape[0] == 1) else: allow_is_causal_skip = not getattr(past_key_values, "is_compileable", False) # Allow slight deviations from causal mask # Note that it is very important to apply this before any other deviations of the mask (such as packed sequence mask, # padding mask, etc) as the resulting mask may otherwise not be correct! if or_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = or_masks(mask_factory_function, or_mask_function) allow_is_causal_skip = False use_vmap = True if and_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = and_masks(mask_factory_function, and_mask_function) allow_is_causal_skip = False use_vmap = True # If we detected packing format if packed_sequence_mask is not None: mask_factory_function = and_masks(mask_factory_function, packed_sequence_mask_function(packed_sequence_mask)) allow_is_causal_skip = False # We now create the mask causal_mask = mask_interface( batch_size=batch_size, cache_position=cache_position, kv_length=kv_length, kv_offset=kv_offset, mask_function=mask_factory_function, attention_mask=attention_mask, allow_is_causal_skip=allow_is_causal_skip, # additional kwarg for sdpa dtype=dtype, # Additional kwarg for eager config=config, # Pass the config as well, in case someone wants to easily have their own mask_interface use_vmap=use_vmap, # Short-circuit to non-vmap expansions for the mask ) return causal_mask @deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds") def create_bidirectional_mask( config: PreTrainedConfig, inputs_embeds: torch.Tensor, attention_mask: torch.Tensor | None, encoder_hidden_states: torch.Tensor | None = None, or_mask_function: Callable | None = None, and_mask_function: Callable | None = None, ) -> torch.Tensor | BlockMask | None: """ Create a standard bidirectional mask based on the attention implementation used (stored in the config). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is only used to infer metadata such as the batch size, query length, dtype, and device. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, kv_length). It can also be an already prepared 4D mask of shape (batch_size, 1, query_length, kv_length), in which case it is returned as-is. encoder_hidden_states (`torch.Tensor`, optional): The input embeddings of shape (batch_size, kv_length, hidden_dim). If provided, it is used instead of `inputs_embeds` to infer the batch size, kv length and dtype. or_mask_function (`Callable`, optional): An optional mask function to combine with the base mask function (by doing the union of both). This is useful to easily overlay another mask on top, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the base mask function (by doing the intersection of both). This is useful to easily overlay another mask on top, for example for image tokens handling. """ # Due to the logic surrounding `cache_position` in inferring query-related information, we # construct a dummy tensor imitating initial positions cache_position = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device, dtype=torch.long) embeds = encoder_hidden_states if encoder_hidden_states is not None else inputs_embeds # We ignore a few irrelevant arguments at the end as we do not have a (growing) cache here early_exit, attention_mask, _, kv_length, kv_offset = _preprocess_mask_arguments( config, embeds, attention_mask, cache_position, None, None, 0 ) if early_exit: return attention_mask batch_size, dtype = embeds.shape[0], embeds.dtype mask_factory_function = bidirectional_mask_function mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation] # Allow skipping the mask creation except we have additional masking operators (and/or masks) allow_is_bidirectional_skip = True # Defaulting to using non-vmap based mask creations except when detecting # users passing custom mask functions (as we cannot guarantee that they # are properly index-based as required by our implementation). use_vmap = False # Allow slight deviations from the base mask # Note that it is very important to apply this before any other deviations of the mask (such as packed sequence mask, # padding mask, etc) as the resulting mask may otherwise not be correct! if or_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = or_masks(mask_factory_function, or_mask_function) allow_is_bidirectional_skip = False use_vmap = True if and_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = and_masks(mask_factory_function, and_mask_function) allow_is_bidirectional_skip = False use_vmap = True # We now create the mask attention_mask = mask_interface( batch_size=batch_size, cache_position=cache_position, kv_length=kv_length, kv_offset=kv_offset, mask_function=mask_factory_function, attention_mask=attention_mask, # Additional kwargs for sdpa allow_is_causal_skip=False, allow_is_bidirectional_skip=allow_is_bidirectional_skip, dtype=dtype, # Additional kwarg for eager config=config, # Pass the config as well, in case someone wants to easily have their own mask_interface use_vmap=use_vmap, # Short-circuit to non-vmap expansions for the mask ) return attention_mask @deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds") def create_sliding_window_causal_mask( config: PreTrainedConfig, inputs_embeds: torch.Tensor, attention_mask: torch.Tensor | None, cache_position: torch.Tensor, past_key_values: Cache | None, position_ids: torch.Tensor | None = None, or_mask_function: Callable | None = None, and_mask_function: Callable | None = None, ) -> torch.Tensor | BlockMask | None: """ Create a sliding window causal mask based on the attention implementation used (stored in the config). This type of attention pattern was mostly democratized by Mistral. If `past_key_values` has an hybrid cache structure, this function will return the mask corresponding to one of the "sliding_attention" layers (to align to what is needed in the `modeling_xxx.py` files). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. or_mask_function (`Callable`, optional): An optional mask function to combine with the sliding causal mask function (by doing the union of both). This is useful to easily overlay another mask on top of the sliding causal one, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the sliding causal mask function (by doing the intersection of both). This is useful to easily overlay another mask on top of the sliding causal one, for example for image tokens handling. """ # Power feature: if `is_causal` is False, then fallback to bi-directional mask for bi-directional attention # It allows to use decoder-only models with bi-directional attention as well if not getattr(config, "is_causal", True): return create_bidirectional_sliding_window_mask( config, inputs_embeds, attention_mask, or_mask_function=or_mask_function, and_mask_function=and_mask_function, ) # If we have an hybrid cache structure, here we want to create the mask for the sliding layers if hasattr(past_key_values, "is_sliding") and True in past_key_values.is_sliding: layer_idx = past_key_values.is_sliding.index(True) else: layer_idx = 0 early_exit, attention_mask, packed_sequence_mask, kv_length, kv_offset = _preprocess_mask_arguments( config, inputs_embeds, attention_mask, cache_position, past_key_values, position_ids, layer_idx ) if early_exit: return attention_mask sliding_window = getattr(config, "sliding_window", None) if sliding_window is None: raise ValueError("Could not find a `sliding_window` argument in the config, or it is not set") batch_size, dtype = inputs_embeds.shape[0], inputs_embeds.dtype mask_factory_function = sliding_window_causal_mask_function(sliding_window) mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation] # Defaulting to using non-vmap based mask creations except when detecting # users passing custom mask functions (as we cannot guarantee that they # are properly index-based as required by our implementation). use_vmap = False # Do not allow skip if we are compiling (this is to match BC) # TODO: cyril -> probably revisit and remove this, but a lot of tests rely on it allow_is_causal_skip = not getattr(past_key_values, "is_compileable", False) # Allow slight deviations from causal mask # Note that it is very important to apply this before any other deviations of the mask (such as packed sequence mask, # padding mask, etc) as the resulting mask may otherwise not be correct! if or_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = or_masks(mask_factory_function, or_mask_function) allow_is_causal_skip = False use_vmap = True if and_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = and_masks(mask_factory_function, and_mask_function) allow_is_causal_skip = False use_vmap = True # If we detected packing format if packed_sequence_mask is not None: mask_factory_function = and_masks(mask_factory_function, packed_sequence_mask_function(packed_sequence_mask)) allow_is_causal_skip = False # We now create the mask causal_mask = mask_interface( batch_size=batch_size, cache_position=cache_position, kv_length=kv_length, kv_offset=kv_offset, mask_function=mask_factory_function, attention_mask=attention_mask, allow_is_causal_skip=allow_is_causal_skip, # additional kwarg for sdpa local_size=sliding_window, # Additional kwarg for sdpa dtype=dtype, # Additional kwarg for eager config=config, # Pass the config as well, in case someone wants to easily have their own mask_interface use_vmap=use_vmap, # Short-circuit to non-vmap expansions for the mask ) return causal_mask @deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds") def create_bidirectional_sliding_window_mask( config: PreTrainedConfig, inputs_embeds: torch.Tensor, attention_mask: torch.Tensor | None, or_mask_function: Callable | None = None, and_mask_function: Callable | None = None, ) -> torch.Tensor | BlockMask | None: """ Create a standard bidirectional sliding window mask based on the attention implementation used (stored in the config). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is only used to infer metadata such as the batch size, query length, dtype, and device. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, kv_length). It can also be an already prepared 4D mask of shape (batch_size, 1, query_length, kv_length), in which case it is returned as-is. or_mask_function (`Callable`, optional): An optional mask function to combine with the base mask function (by doing the union of both). This is useful to easily overlay another mask on top, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the base mask function (by doing the intersection of both). This is useful to easily overlay another mask on top, for example for image tokens handling. """ # Due to the logic surrounding `cache_position` in inferring query-related information, we # construct a dummy tensor imitating initial positions cache_position = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device, dtype=torch.long) # We ignore a few irrelevant arguments at the end as we do not have a (growing) cache here early_exit, attention_mask, _, kv_length, kv_offset = _preprocess_mask_arguments( config, inputs_embeds, attention_mask, cache_position, None, None, 0 ) if early_exit: return attention_mask sliding_window = getattr(config, "sliding_window", None) if sliding_window is None: raise ValueError("Could not find a `sliding_window` argument in the config, or it is not set") batch_size, dtype = inputs_embeds.shape[0], inputs_embeds.dtype mask_factory_function = sliding_window_bidirectional_mask_function(sliding_window) mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation] use_vmap = False allow_is_bidirectional_skip = True if or_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = or_masks(mask_factory_function, or_mask_function) allow_is_bidirectional_skip = False use_vmap = True if and_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = and_masks(mask_factory_function, and_mask_function) allow_is_bidirectional_skip = False use_vmap = True attention_mask = mask_interface( batch_size=batch_size, cache_position=cache_position, kv_length=kv_length, kv_offset=kv_offset, mask_function=mask_factory_function, attention_mask=attention_mask, allow_is_causal_skip=False, allow_is_bidirectional_skip=allow_is_bidirectional_skip, local_size=sliding_window, # Additional kwarg for sdpa dtype=dtype, # Additional kwarg for eager config=config, # Pass the config as well, in case someone wants to easily have their own mask_interface use_vmap=use_vmap, # Short-circuit to non-vmap expansions for the mask ) return attention_mask @deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds") def create_chunked_causal_mask( config: PreTrainedConfig, inputs_embeds: torch.Tensor, attention_mask: torch.Tensor | None, cache_position: torch.Tensor, past_key_values: Cache | None, position_ids: torch.Tensor | None = None, or_mask_function: Callable | None = None, and_mask_function: Callable | None = None, ) -> torch.Tensor | BlockMask | None: """ Create a chunked attention causal mask based on the attention implementation used (stored in the config). This type of attention pattern was mostly democratized by Llama4. If `past_key_values` has an hybrid cache structure, this function will return the mask corresponding to one of the "chunked_attention" layers (to align to what is needed in the `modeling_xxx.py` files). Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. or_mask_function (`Callable`, optional): An optional mask function to combine with the chunked causal mask function (by doing the union of both). This is useful to easily overlay another mask on top of the chunked causal one, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the chunked causal mask function (by doing the intersection of both). This is useful to easily overlay another mask on top of the chunked causal one, for example for image tokens handling. """ # If we have an hybrid cache structure, here we want to create the mask for the sliding layers if hasattr(past_key_values, "is_sliding") and True in past_key_values.is_sliding: layer_idx = past_key_values.is_sliding.index(True) else: layer_idx = 0 early_exit, attention_mask, packed_sequence_mask, kv_length, kv_offset = _preprocess_mask_arguments( config, inputs_embeds, attention_mask, cache_position, past_key_values, position_ids, layer_idx ) if early_exit: return attention_mask chunk_size = getattr(config, "attention_chunk_size", None) if chunk_size is None: raise ValueError("Could not find an `attention_chunk_size` argument in the config, or it is not set") # Raise if using chunked attention on context too large with FA if is_flash_attention_requested(config) and kv_length + kv_offset > chunk_size: raise ValueError( "Flash attention cannot handle chunked attention, and the key-value length is larger than the chunk size so the " "chunked pattern cannot be respected. You should use another `attn_implementation` when instantiating the model" ) batch_size, dtype = inputs_embeds.shape[0], inputs_embeds.dtype # For chunked attention and batched inputs, we need to take the number of left padding tokens into account # to start the chunk from the actual start of the sequence for the padded sequence if attention_mask is not None: # Only count the left padding tokens, not all of them left_padding_tokens = (attention_mask.cumsum(dim=-1) == torch.zeros_like(attention_mask)).sum(dim=-1) else: left_padding_tokens = torch.zeros(batch_size, device=cache_position.device, dtype=int) mask_factory_function = chunked_causal_mask_function(chunk_size, left_padding_tokens) mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation] # Defaulting to using non-vmap based mask creations except when detecting # users passing custom mask functions (as we cannot guarantee that they # are properly index-based as required by our implementation). use_vmap = False # Do not allow skip if we are compiling (this is to match BC) # TODO: cyril -> probably revisit and remove this, but a lot of tests rely on it allow_is_causal_skip = not getattr(past_key_values, "is_compileable", False) # Allow slight deviations from causal mask # Note that it is very important to apply this before any other deviations of the mask (such as packed sequence mask, # padding mask, etc) as the resulting mask may otherwise not be correct! if or_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = or_masks(mask_factory_function, or_mask_function) allow_is_causal_skip = False use_vmap = True if and_mask_function is not None: if not _is_torch_greater_or_equal_than_2_6: raise ValueError("Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6") mask_factory_function = and_masks(mask_factory_function, and_mask_function) allow_is_causal_skip = False use_vmap = True # If we detected packing format if packed_sequence_mask is not None: mask_factory_function = and_masks(mask_factory_function, packed_sequence_mask_function(packed_sequence_mask)) allow_is_causal_skip = False # We now create the mask causal_mask = mask_interface( batch_size=batch_size, cache_position=cache_position, kv_length=kv_length, kv_offset=kv_offset, mask_function=mask_factory_function, attention_mask=attention_mask, allow_is_causal_skip=allow_is_causal_skip, # additional kwarg for sdpa local_size=chunk_size, # Additional kwarg for sdpa dtype=dtype, # Additional kwarg for eager config=config, # Pass the config as well, in case someone wants to easily have their own mask_interface use_vmap=use_vmap, # Short-circuit to non-vmap expansions for the mask ) return causal_mask LAYER_PATTERN_TO_MASK_FUNCTION_MAPPING = { "full_attention": create_causal_mask, "sliding_attention": create_sliding_window_causal_mask, "chunked_attention": create_chunked_causal_mask, } @deprecate_kwarg("input_embeds", version="5.6.0", new_name="inputs_embeds") def create_masks_for_generate( config: PreTrainedConfig, inputs_embeds: torch.Tensor, attention_mask: torch.Tensor | None, cache_position: torch.Tensor, past_key_values: Cache | None, position_ids: torch.Tensor | None = None, or_mask_function: Callable | None = None, and_mask_function: Callable | None = None, **kwargs, ): """ This function mimics how we create the masks in the `modeling_xxx.py` files, and is used in places like `generate` in order to easily create the masks in advance, when we compile the forwards with Static caches. Args: config (`PreTrainedConfig`): The model config. inputs_embeds (`torch.Tensor`): The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the batch size, query length and dtype. attention_mask (`torch.Tensor`, optional): The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length). It can also be an already prepared 4D mask, in which case it is returned as-is. cache_position (`torch.Tensor`): A tensor of shape (query_length,) indicating the current indices of the input sequence elements. past_key_values (`Cache`, optional): The past key values, if we use a cache. position_ids (`torch.Tensor`, optional) A 2D tensor of shape (batch_size, query_length) indicating the positions of each token in the sequences. or_mask_function (`Callable`, optional): An optional mask function to combine with the other mask function (by doing the union of both). This is useful to easily overlay another mask on top of the causal one, for example for image tokens handling. and_mask_function (`Callable`, optional): An optional mask function to combine with the other mask function (by doing the intersection of both). This is useful to easily overlay another mask on top of the causal one, for example for image tokens handling. """ # The attribute reside in the text config for composite models effective_config = config.get_text_config() # Prepare the mask args mask_kwargs = { "config": effective_config, "inputs_embeds": inputs_embeds, "attention_mask": attention_mask, "cache_position": cache_position, "past_key_values": past_key_values, "position_ids": position_ids, "or_mask_function": or_mask_function, "and_mask_function": and_mask_function, } # If the attribute exist, we need several masks if hasattr(effective_config, "layer_types"): causal_masks = {} for layer_pattern in set(effective_config.layer_types): causal_masks[layer_pattern] = LAYER_PATTERN_TO_MASK_FUNCTION_MAPPING[layer_pattern](**mask_kwargs) return causal_masks # In this case, all layers are sliding elif getattr(effective_config, "sliding_window", None) is not None: return create_sliding_window_causal_mask(**mask_kwargs) # In this case, all layers are chunked elif getattr(effective_config, "attention_chunk_size", None) is not None: return create_chunked_causal_mask(**mask_kwargs) # All layers use standard causal attention return create_causal_mask(**mask_kwargs) # Below are utilities to pretty-print the different masks # Print the matrix with words as row labels GREEN = "\033[92m" YELLOW = "\033[93m" RESET = "\033[0m" BLACK_SQUARE = "■" WHITE_SQUARE = "⬚" GREY_SQUARE = "∙" LOW_TRIANGLE = "⬕" UPPER_TRIANGLE = "⬔" def get_style(style): if style == "majong": BLACK_SQUARE = "🀞" # Full block (represents "on" or active) BLACK_SQUARE = "🀙" # Full block (represents "on" or active) WHITE_SQUARE = "🀆" # "▒" # Light shade (represents "off" or inactive) LOW_TRIANGLE = "🀛" # Lower left triangle (stylized indication) UPPER_TRIANGLE = "🀛" # Upper left triangle (stylized indication) else: BLACK_SQUARE = "█" # Full block (represents "on" or active) WHITE_SQUARE = "░" # "▒" # Light shade (represents "off" or inactive) LOW_TRIANGLE = "▙" # Lower left triangle (stylized indication)) UPPER_TRIANGLE = "▜" # Upper left triangle (stylized indication) return BLACK_SQUARE, WHITE_SQUARE, LOW_TRIANGLE, UPPER_TRIANGLE # LOW_TRIANGLE = UPPER_TRIANGLE = "⟍" # Upper right triangle (stylized indication) YELLOW_SQUARE = f"{YELLOW}{BLACK_SQUARE}{RESET}" GREEN_SQUARE = f"{GREEN}{BLACK_SQUARE}{RESET}" def tensor_to_mask_visual(original_tensor: torch.Tensor, grid_size=(20, 40), style="majong") -> str: BLACK_SQUARE, WHITE_SQUARE, LOW_TRIANGLE, UPPER_TRIANGLE = get_style(style) h, w = original_tensor.shape max_h, max_w = grid_size if not (h < max_h and w < max_w): # Preserve aspect ratio within max grid size aspect_ratio = 2 * w / h if aspect_ratio > 1: w = max_w h = min(max_h, max(1, round(max_w / aspect_ratio))) else: h = max_h w = max(1, round(max_h * aspect_ratio)) # Step 1: Rescale tensor by average pooling tensor = original_tensor.unsqueeze(0).unsqueeze(0) # Add batch and channel dimensions tensor = F.adaptive_avg_pool2d(tensor, output_size=(h, w))[0, 0] # Remove extra dims else: tensor = original_tensor # Step 3: Build the string representation result = [] for i in range(h): row = "" for j in range(w): if tensor[i, j] == 1: row += BLACK_SQUARE elif tensor[i, j] == 0: row += WHITE_SQUARE else: if j > 0: if tensor[i, j - 1] == 1: row += LOW_TRIANGLE elif tensor[i, j - 1] == 0: row += UPPER_TRIANGLE else: row += BLACK_SQUARE if tensor[i, j] == 1 else WHITE_SQUARE else: row += ( BLACK_SQUARE if tensor[i, j] == 1 else ( WHITE_SQUARE if tensor[i, j] == 0 else (UPPER_TRIANGLE if tensor[i, j + 1] == 1 else LOW_TRIANGLE) ) ) result.append(row) return "\n".join(result) class AttentionMask(torch.Tensor): def __new__(cls, data, style=None): # Create a new instance of AttentionMask as a Tensor cls.style = style return torch.Tensor._make_subclass(cls, data, require_grad=False) def __init__(self, data): # You can initialize any additional metadata here if needed pass def to_string(self, grid_size=(20, 40), limit=4): """Returns a string representation of the block mask.""" dense_mask = self *batch_dims, num_rows, num_cols = dense_mask.shape 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 AttentionMask.to_string(limit=N)") total_vis.append("You can also index (AttentionMask[batch, head]) to choose a specific batch or head") break block_vis = tensor_to_mask_visual(dense_mask[batch_idx], grid_size=grid_size, style=self.style) total_vis.append(block_vis) total_vis.append(f"torch.Tensor(shape={tuple(self.shape)}, dtype={self.dtype})") return "\n".join(total_vis) def __repr__(self): return self.to_string() def __str__(self): return self.to_string() @classmethod def from_tensor(cls, tensor: torch.Tensor, style: str | None = None) -> "AttentionMask": res = cls(tensor) res.style = style return res