from abc import ABC, abstractmethod from collections.abc import Iterable import torch from .configuration_utils import PreTrainedConfig from .utils import ( is_hqq_available, is_optimum_quanto_available, is_quanto_greater, is_torch_greater_or_equal, is_torchdynamo_compiling, logging, ) if is_hqq_available(): from hqq.core.quantize import Quantizer as HQQQuantizer _is_torch_greater_or_equal_than_2_7 = is_torch_greater_or_equal("2.7", accept_dev=True) logger = logging.get_logger(__name__) class CacheLayerMixin(ABC): """Base, abstract class for a single layer's cache.""" is_compileable = False def __init__(self): self.keys: torch.Tensor | None = None self.values: torch.Tensor | None = None self.is_initialized = False def __repr__(self): return f"{self.__class__.__name__}" @abstractmethod def lazy_initialization(self, key_states: torch.Tensor, value_states: torch.Tensor) -> None: ... @abstractmethod def update( self, key_states: torch.Tensor, value_states: torch.Tensor, *args, **kwargs ) -> tuple[torch.Tensor, torch.Tensor]: ... @abstractmethod def get_mask_sizes(self, query_length: int) -> tuple[int, int]: ... @abstractmethod def get_seq_length(self) -> int: ... @abstractmethod def get_max_cache_shape(self) -> int: ... def offload(self): """Offload this layer's data to CPU device.""" if self.is_initialized: self.keys = self.keys.to("cpu", non_blocking=True) self.values = self.values.to("cpu", non_blocking=True) def prefetch(self): """In case of layer offloading, this allows to move the data back to the layer's device ahead of time.""" if self.is_initialized and self.keys.device != self.device: self.keys = self.keys.to(self.device, non_blocking=True) self.values = self.values.to(self.device, non_blocking=True) def reset(self) -> None: """Resets the cache values while preserving the objects""" if self.is_initialized: self.keys.zero_() self.values.zero_() # This attribute is set on several Layers if hasattr(self, "cumulative_length"): # It can either be an int for dynamic layers, or a tensor for static layers if isinstance(self.cumulative_length, int): self.cumulative_length = 0 else: self.cumulative_length.zero_() def reorder_cache(self, beam_idx: torch.LongTensor) -> None: """Reorders this layer's cache for beam search.""" if self.get_seq_length() > 0: self.keys = self.keys.index_select(0, beam_idx.to(self.keys.device)) self.values = self.values.index_select(0, beam_idx.to(self.values.device)) class DynamicLayer(CacheLayerMixin): """ A cache layer that grows dynamically as more tokens are generated. This is the default for generative models. It stores the key and value states as tensors of shape `[batch_size, num_heads, seq_len, head_dim]`. """ is_sliding = False def lazy_initialization(self, key_states: torch.Tensor, value_states: torch.Tensor) -> None: self.dtype, self.device = key_states.dtype, key_states.device self.keys = torch.tensor([], dtype=self.dtype, device=self.device) self.values = torch.tensor([], dtype=self.dtype, device=self.device) self.is_initialized = True def update( self, key_states: torch.Tensor, value_states: torch.Tensor, *args, **kwargs ) -> tuple[torch.Tensor, torch.Tensor]: """ Update the key and value caches in-place, and return the necessary keys and value states. Args: key_states (`torch.Tensor`): The new key states to cache. value_states (`torch.Tensor`): The new value states to cache. Returns: tuple[`torch.Tensor`, `torch.Tensor`]: The key and value states. """ # Lazy initialization if not self.is_initialized: self.lazy_initialization(key_states, value_states) self.keys = torch.cat([self.keys, key_states], dim=-2) self.values = torch.cat([self.values, value_states], dim=-2) return self.keys, self.values def get_mask_sizes(self, query_length: int) -> tuple[int, int]: """Return the length and offset of the cache, used to generate the mask""" kv_offset = 0 kv_length = self.get_seq_length() + query_length return kv_length, kv_offset def get_seq_length(self) -> int: """Returns the sequence length of the cached states.""" if not self.is_initialized or self.keys.numel() == 0: return 0 return self.keys.shape[-2] def get_max_cache_shape(self) -> int: """Returns the maximum sequence length of the cache object. DynamicLayer does not have a maximum length.""" return -1 def crop(self, max_length: int) -> None: """ Crop the past key values up to a new `max_length` in terms of tokens. `max_length` can also be negative to remove `max_length` tokens. """ if max_length < 0: max_length = self.get_seq_length() - abs(max_length) if self.get_seq_length() <= max_length: return self.keys = self.keys[..., :max_length, :] self.values = self.values[..., :max_length, :] def batch_repeat_interleave(self, repeats: int) -> None: """Repeat the cache `repeats` times in the batch dimension.""" if self.get_seq_length() > 0: self.keys = self.keys.repeat_interleave(repeats, dim=0) self.values = self.values.repeat_interleave(repeats, dim=0) def batch_select_indices(self, indices: torch.Tensor) -> None: """Only keep the `indices` in the batch dimension of the cache.""" if self.get_seq_length() > 0: self.keys = self.keys[indices, ...] self.values = self.values[indices, ...] class DynamicSlidingWindowLayer(DynamicLayer): """ A cache layer that grows dynamically as more tokens are generated, up until the sliding window size. It stores the key and value states as tensors of shape `[batch_size, num_heads, min(seq_len, sliding_window), head_dim]`. """ is_sliding = True def __init__(self, sliding_window: int): super().__init__() self.sliding_window = sliding_window self.cumulative_length = 0 self._sliding_window_tensor = torch.tensor(self.sliding_window, dtype=torch.long) def lazy_initialization(self, key_states: torch.Tensor, value_states: torch.Tensor) -> None: super().lazy_initialization(key_states, value_states) self._sliding_window_tensor = self._sliding_window_tensor.to(self.device) def update( self, key_states: torch.Tensor, value_states: torch.Tensor, *args, **kwargs ) -> tuple[torch.Tensor, torch.Tensor]: """ Update the key and value caches in-place, and return the necessary keys and value states. Args: key_states (`torch.Tensor`): The new key states to cache. value_states (`torch.Tensor`): The new value states to cache. Returns: tuple[`torch.Tensor`, `torch.Tensor`]: The key and value states. """ # Lazy initialization if not self.is_initialized: self.lazy_initialization(key_states, value_states) self.cumulative_length += key_states.shape[-2] # Compute the full states full_key_states = torch.cat([self.keys, key_states], dim=-2) full_value_states = torch.cat([self.values, value_states], dim=-2) # Only cache the last `self.sliding_window - 1` tokens (or all of them if lower than that) self.keys = full_key_states[:, :, -self.sliding_window + 1 :, :] self.values = full_value_states[:, :, -self.sliding_window + 1 :, :] # Return the full states return full_key_states, full_value_states def get_mask_sizes(self, query_length: int) -> tuple[int, int]: """Return the length and offset of the cache, used to generate the attention mask""" is_full = self.cumulative_length >= self.sliding_window kv_offset = max(self.cumulative_length - self.sliding_window + 1, 0) if is_full: kv_length = self.sliding_window - 1 + query_length else: kv_length = self.cumulative_length + query_length return kv_length, kv_offset def get_seq_length(self) -> int: """Returns the sequence length of the cached states.""" return self.cumulative_length def get_max_cache_shape(self) -> int: """Return the maximum cache shape of the cache""" return self.sliding_window def crop(self, max_length: int) -> None: """ Crop the past key values up to a new `max_length` in terms of tokens. `max_length` can also be negative to remove `max_length` tokens. """ if self.get_seq_length() >= self.sliding_window: raise ValueError( "Cannot `crop` a `DynamicSlidingWindowLayer` after it has seen more tokens than its" "sliding window (otherwise some states are lost)" ) super().crop(max_length) self.cumulative_length = self.keys.shape[-2] class StaticLayer(CacheLayerMixin): """ A static cache layer that stores the key and value states as static tensors of shape `[batch_size, num_heads, max_cache_len), head_dim]`. It lazily allocates its full backing tensors, and then mutates them in-place. Built for `torch.compile` support. Args: max_cache_len (`int`): Maximum number of tokens that can be stored, used for tensor preallocation. """ is_compileable = True is_sliding = False def __init__(self, max_cache_len: int): super().__init__() self.max_cache_len = max_cache_len # Very important that it's a tensor here, to avoid recompiling when we update it and use it to create positions self.cumulative_length = torch.tensor([0], dtype=int) def lazy_initialization(self, key_states: torch.Tensor, value_states: torch.Tensor) -> None: """ Lazy initialization of the keys and values tensors. This allows to get all properties (dtype, device, num_heads in case of TP etc...) at runtime directly, which is extremely practical as it avoids moving devices, dtypes etc later on for each `update` (which could break the static dynamo addresses as well). If this is unwanted, one can call `early_initialization(...)` on the Cache directly, which will call this function ahead-of-time (this is required for `torch.export` for example). Note that for `compile`, as we internally don't compile the prefill, this is guaranteed to have been called already when compiling. If compiling the prefill as well, e.g. calling `model.compile(...)` before `generate` with a static cache, it is still supported in general, but without guarantees depending on the compilation options (e.g. cuda graphs, i.e. `mode="reduce-overhead"` is known to fail). But it will in general work correctly, and prefill should not be compiled anyway for performances! """ self.dtype, self.device = key_states.dtype, key_states.device self.max_batch_size, self.num_heads = key_states.shape[:2] self.v_head_dim = value_states.shape[-1] self.k_head_dim = key_states.shape[-1] self.keys = torch.zeros( (self.max_batch_size, self.num_heads, self.max_cache_len, self.k_head_dim), dtype=self.dtype, device=self.device, ) self.values = torch.zeros( (self.max_batch_size, self.num_heads, self.max_cache_len, self.v_head_dim), dtype=self.dtype, device=self.device, ) self.cumulative_length = self.cumulative_length.to(self.device) # Note: `mark_static_address` is used to tag the tensors as a fixed data pointer, preventing compiled graph # breaks or cudagraph skips due to inplace mutations when updating the cache. However, it is not supported when # tracing the graph, so we skip it in this case. As prefill should never be compiled, this is not an issue and it # will still be run (except when users compile prefill explicitly, but this should be avoided!) # Without this, we cannot use cudagraphs!! if not is_torchdynamo_compiling(): torch._dynamo.mark_static_address(self.keys) torch._dynamo.mark_static_address(self.values) torch._dynamo.mark_static_address(self.cumulative_length) self.is_initialized = True def update( self, key_states: torch.Tensor, value_states: torch.Tensor, *args, **kwargs ) -> tuple[torch.Tensor, torch.Tensor]: """ Update the key and value caches in-place, and return the necessary keys and value states. Args: key_states (`torch.Tensor`): The new key states to cache. value_states (`torch.Tensor`): The new value states to cache. Returns: tuple[`torch.Tensor`, `torch.Tensor`]: The key and value states. """ # Lazy initialization if not self.is_initialized: self.lazy_initialization(key_states, value_states) # Create a tensor to slice the static kv at the correct indices kv_length = key_states.shape[-2] cache_position = torch.arange(kv_length, device=self.device) + self.cumulative_length # Note that has to be performed in-place, as we have a static address that we need to keep self.cumulative_length.add_(kv_length) # Update the cache try: self.keys.index_copy_(2, cache_position, key_states) self.values.index_copy_(2, cache_position, value_states) except NotImplementedError: # Fallback for devices like MPS where index_copy_ might not be supported. self.keys[:, :, cache_position] = key_states self.values[:, :, cache_position] = value_states return self.keys, self.values def get_mask_sizes(self, query_length: int) -> tuple[int, int]: """Return the length and offset of the cache, used to generate the attention mask""" kv_offset = 0 kv_length = self.max_cache_len return kv_length, kv_offset def get_seq_length(self) -> int: """Returns the sequence length of the cached states.""" return self.cumulative_length if self.is_initialized else 0 def get_max_cache_shape(self) -> int: """Return the maximum cache shape of the cache""" return self.max_cache_len class StaticSlidingWindowLayer(StaticLayer): """ A static cache layer that stores the key and value states as static tensors of shape `[batch_size, num_heads, min(max_cache_len, sliding_window), head_dim]`. It lazily allocates its full backing tensors, and then mutates them in-place. Built for `torch.compile` support. Args: max_cache_len (`int`): Maximum number of tokens that can be stored, used for tensor preallocation. sliding_window (`int`): The size of the sliding window. """ is_sliding = True def __init__(self, max_cache_len: int, sliding_window: int): effective_max_cache_len = min(sliding_window, max_cache_len) super().__init__(max_cache_len=effective_max_cache_len) # Here, to avoid data-dependent control flows, we also need to use a python int to keep track of the cumulative length self.cumulative_length_int = 0 def update( self, key_states: torch.Tensor, value_states: torch.Tensor, *args, **kwargs ) -> tuple[torch.Tensor, torch.Tensor]: """ Update the key and value caches in-place, and return the necessary keys and value states. Args: key_states (`torch.Tensor`): The new key states to cache. value_states (`torch.Tensor`): The new value states to cache. Returns: tuple[`torch.Tensor`, `torch.Tensor`]: The key and value states. """ # Lazy initialization if not self.is_initialized: self.lazy_initialization(key_states, value_states) kv_length = key_states.shape[-2] current_length = self.cumulative_length_int is_full = current_length >= self.max_cache_len # Update it now that we saved the value above self.cumulative_length_int += kv_length if is_full: # In general, we should use a much simpler `cat` here as well, independently of the states size. However, # dynamo is currently bugged when doing it - see https://github.com/pytorch/pytorch/issues/159855 for more details if key_states.shape[-2] == 1: # Roll all values to the left by 1 position new_keys = self.keys.roll(-1, dims=-2) new_values = self.values.roll(-1, dims=-2) # Overwrite the last position with new states # (note: very important to use a tensor to index here, see https://github.com/pytorch/pytorch/issues/159855) index = torch.tensor([-1], dtype=int, device=self.device) new_keys[:, :, index] = key_states new_values[:, :, index] = value_states # Copy back into `self` (do not just assign again) in order to keep the static dynamo address self.keys.copy_(new_keys) self.values.copy_(new_values) # Very important to return the `self` tensors here, as they have the static dynamo address return self.keys, self.values # Already full but using more than 1 new token (e.g. prefill caching, chat continuation, etc...) else: full_key_states = torch.cat((self.keys[:, :, 1:, :], key_states), dim=-2) full_value_states = torch.cat((self.values[:, :, 1:, :], value_states), dim=-2) # Not yet full, but becoming full on this update elif current_length + kv_length > self.max_cache_len: # Fast prefill path, no need to cat() in this case, as the cache is currently empty if current_length == 0: full_key_states = key_states full_value_states = value_states else: full_key_states = torch.cat((self.keys[:, :, :current_length, :], key_states), dim=-2) full_value_states = torch.cat((self.values[:, :, :current_length, :], value_states), dim=-2) else: # Note: very important to use the tensor version of the cumulative length here, as otherwise cudagraphs # (triggered by mode="reduced_overhead") will lead to random crashes, as the int would be overwritten cache_position = torch.arange(kv_length, device=self.device) + self.cumulative_length try: self.keys.index_copy_(2, cache_position, key_states) self.values.index_copy_(2, cache_position, value_states) except NotImplementedError: self.keys[:, :, cache_position] = key_states self.values[:, :, cache_position] = value_states # Update the tensor version of the length in-place (we don't need to update it if we are already outside # of this branch, as we don't need the tensor anymore) self.cumulative_length.add_(kv_length) # Very important to return the `self` tensors here, as they have the static dynamo address return self.keys, self.values # We only cache the last `sliding_window` tokens self.keys.copy_(full_key_states[:, :, -self.max_cache_len :, :]) self.values.copy_(full_value_states[:, :, -self.max_cache_len :, :]) # we should return the whole states instead of `self.keys/values` here, as otherwise we lose some context return full_key_states, full_value_states def get_mask_sizes(self, query_length: int) -> tuple[int, int]: """Return the length and offset of the cache, used to generate the attention mask""" sliding_window = self.max_cache_len is_full = self.cumulative_length_int >= self.max_cache_len kv_offset = max(self.cumulative_length_int - sliding_window + 1, 0) # The cache is already full if is_full: kv_length = sliding_window + query_length - 1 # Not yet full, but becoming full on this update elif self.cumulative_length_int + query_length > sliding_window: kv_length = self.cumulative_length_int + query_length # Here the Cache is still smaller than the local size, but we return the local size as it's static else: kv_length = sliding_window return kv_length, kv_offset def get_seq_length(self) -> int: """Returns the sequence length of the cached states.""" return self.cumulative_length_int def reset(self): super().reset() self.cumulative_length_int = 0 class QuantizedLayer(DynamicLayer): """ A quantized layer similar to what is described in the [KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache paper](https://huggingface.co/papers/2402.02750). It allows the model to generate longer sequence length without allocating too much memory for the key and value caches by applying quantization. The cache has two types of storage, one for original precision and one for the quantized cache. A `residual length` is set as a maximum capacity for the original precision cache. When the length goes beyond maximum capacity, the original precision cache is discarded and moved into the quantized cache. The quantization is done per-channel with a set `q_group_size` for both Keys and Values, in contrast to what was described in the paper. """ def __init__( self, nbits: int = 4, axis_key: int = 0, axis_value: int = 0, q_group_size: int = 64, residual_length: int = 128, ): super().__init__() self.nbits = nbits self.axis_key = axis_key self.axis_value = axis_value self.q_group_size = q_group_size self.residual_length = residual_length self.cumulative_length = 0 def update( self, key_states: torch.Tensor, value_states: torch.Tensor, *args, **kwargs ) -> tuple[torch.Tensor, torch.Tensor]: """ Update the key and value caches in-place, and return the necessary keys and value states. Args: key_states (`torch.Tensor`): The new key states to cache. value_states (`torch.Tensor`): The new value states to cache. Returns: tuple[`torch.Tensor`, `torch.Tensor`]: The key and value states. """ self.cumulative_length += key_states.shape[-2] # Lazy initialization if not self.is_initialized: self.lazy_initialization(key_states, value_states) self._quantized_keys = self._quantize(key_states.contiguous(), axis=self.axis_key) self._quantized_values = self._quantize(value_states.contiguous(), axis=self.axis_value) return key_states, value_states dequant_keys = self._dequantize(self._quantized_keys) dequant_values = self._dequantize(self._quantized_values) keys_to_return = torch.cat([dequant_keys, self.keys, key_states], dim=-2) values_to_return = torch.cat([dequant_values, self.values, value_states], dim=-2) if self.keys.dim() == 4 and self.keys.shape[-2] + 1 >= self.residual_length: self._quantized_keys = self._quantize(keys_to_return.contiguous(), axis=self.axis_key) self._quantized_values = self._quantize(values_to_return.contiguous(), axis=self.axis_value) self.keys = torch.tensor([], dtype=key_states.dtype, device=key_states.device) self.values = torch.tensor([], dtype=key_states.dtype, device=key_states.device) else: self.keys = torch.cat([self.keys, key_states], dim=-2) self.values = torch.cat([self.values, value_states], dim=-2) return keys_to_return, values_to_return @abstractmethod def _quantize(self, tensor, axis): ... @abstractmethod def _dequantize(self, q_tensor): ... def get_seq_length(self) -> int: """Returns the sequence length of the cached states.""" return self.cumulative_length class QuantoQuantizedLayer(QuantizedLayer): def __init__( self, nbits: int = 4, axis_key: int = 0, axis_value: int = 0, q_group_size: int = 64, residual_length: int = 128, ): super().__init__( nbits=nbits, axis_key=axis_key, axis_value=axis_value, q_group_size=q_group_size, residual_length=residual_length, ) # We need to import quanto here to avoid circular imports due to optimum/quanto/models/transformers_models.py if not is_optimum_quanto_available(): raise ImportError( "You need to install optimum-quanto in order to use KV cache quantization with optimum-quanto " "backend. Please install it via with `pip install optimum-quanto`" ) elif is_quanto_greater("0.2.5", accept_dev=True): from optimum.quanto import MaxOptimizer, qint2, qint4 else: raise ImportError( "You need optimum-quanto package version to be greater or equal than 0.2.5 to use `QuantoQuantizedLayer`. " ) if self.nbits not in [2, 4]: raise ValueError(f"`nbits` for `quanto` backend has to be one of [`2`, `4`] but got {self.nbits}") if self.axis_key not in [0, -1]: raise ValueError(f"`axis_key` for `quanto` backend has to be one of [`0`, `-1`] but got {self.axis_key}") if self.axis_value not in [0, -1]: raise ValueError( f"`axis_value` for `quanto` backend has to be one of [`0`, `-1`] but got {self.axis_value}" ) self.qtype = qint4 if self.nbits == 4 else qint2 self.optimizer = MaxOptimizer() # hardcode as it's the only one for per-channel quantization def _quantize(self, tensor, axis): from optimum.quanto import quantize_weight scale, zeropoint = self.optimizer(tensor, self.qtype, axis, self.q_group_size) qtensor = quantize_weight(tensor, self.qtype, axis, scale, zeropoint, self.q_group_size) return qtensor def _dequantize(self, qtensor): return qtensor.dequantize() class HQQQuantizedLayer(QuantizedLayer): def __init__( self, nbits: int = 4, axis_key: int = 0, axis_value: int = 0, q_group_size: int = 64, residual_length: int = 128, ): super().__init__( nbits=nbits, axis_key=axis_key, axis_value=axis_value, q_group_size=q_group_size, residual_length=residual_length, ) if not is_hqq_available(): raise ImportError( "You need to install `HQQ` in order to use KV cache quantization with HQQ backend. " "Please install it via with `pip install hqq`" ) if self.nbits not in [1, 2, 3, 4, 8]: raise ValueError( f"`nbits` for `HQQ` backend has to be one of [`1`, `2`, `3`, `4`, `8`] but got {self.nbits}" ) if self.axis_key not in [0, 1]: raise ValueError(f"`axis_key` for `HQQ` backend has to be one of [`0`, `1`] but got {self.axis_key}") if self.axis_value not in [0, 1]: raise ValueError(f"`axis_value` for `HQQ` backend has to be one of [`0`, `1`] but got {self.axis_value}") self.quantizer = HQQQuantizer def _quantize(self, tensor, axis): qtensor, meta = self.quantizer.quantize( tensor, axis=axis, device=self.keys.device, compute_dtype=self.keys.dtype, nbits=self.nbits, group_size=self.q_group_size, ) meta["compute_dtype"] = self.keys.dtype self.quantizer.cuda(qtensor, meta=meta, device=self.keys.device) # Move to device and cast to dtype meta["scale"] = meta["scale"].to(qtensor.device) meta["zero"] = meta["zero"].to(qtensor.device) return qtensor, meta def _dequantize(self, qtensor): quant_tensor, meta = qtensor tensor = self.quantizer.dequantize(quant_tensor, meta) return tensor class LinearAttentionCacheLayerMixin(ABC): """Base, abstract class for a linear attention single layer's cache.""" # All shapes are static by essence in a LinearAttention layer, so it is compileable is_compileable = True def __init__(self): self.conv_states: torch.Tensor | None = None self.recurrent_states: torch.Tensor | None = None self.is_conv_states_initialized = False self.is_recurrent_states_initialized = False self.has_previous_state = False def __repr__(self): return f"{self.__class__.__name__}" @abstractmethod def lazy_initialization( self, conv_states: torch.Tensor | None = None, recurrent_states: torch.Tensor | None = None ) -> None: ... @abstractmethod def update_conv_state(self, conv_states: torch.Tensor) -> torch.Tensor: ... @abstractmethod def update_recurrent_state(self, recurrent_states: torch.Tensor) -> torch.Tensor: ... def offload(self): """Offload this layer's data to CPU device.""" if self.is_conv_states_initialized: self.conv_states = self.conv_states.to("cpu", non_blocking=True) if self.is_recurrent_states_initialized: self.recurrent_states = self.recurrent_states.to("cpu", non_blocking=True) def prefetch(self): """In case of layer offloading, this allows to move the data back to the layer's device ahead of time.""" if self.is_conv_states_initialized and self.conv_states.device != self.device: self.conv_states = self.conv_states.to(self.device, non_blocking=True) if self.is_recurrent_states_initialized and self.recurrent_states.device != self.device: self.recurrent_states = self.recurrent_states.to(self.device, non_blocking=True) def reset(self) -> None: """Resets the cache values while preserving the objects""" if self.is_conv_states_initialized: self.conv_states.zero_() if self.is_recurrent_states_initialized: self.recurrent_states.zero_() self.has_previous_state = False def reorder_cache(self, beam_idx: torch.LongTensor): """Reorders the cache for beam search, given the selected beam indices.""" if self.is_conv_states_initialized: self.conv_states = self.conv_states.index_select(0, beam_idx.to(self.device)) # recurrent_states can stay empty sometimes, see e.g. lfm2 which only uses the conv_states if self.is_recurrent_states_initialized: self.recurrent_states = self.recurrent_states.index_select(0, beam_idx.to(self.device)) def crop(self, max_length: int): # We don't crop the linear attention cache, so simply do nothing here pass class LinearAttentionLayer(LinearAttentionCacheLayerMixin): def lazy_initialization( self, conv_states: torch.Tensor | None = None, recurrent_states: torch.Tensor | None = None ) -> None: # Here, we will lazy init both states separately, each in their own update function if conv_states is not None: self.dtype, self.device = conv_states.dtype, conv_states.device # Even if prefill is larfer/shorter than the conv_size, the tensor is always either padded or truncated self.max_batch_size, self.conv_kernel_size = conv_states.shape[0], conv_states.shape[-1] # The shape is always static, so we init as such self.conv_states = torch.zeros_like(conv_states, dtype=self.dtype, device=self.device) # Mark as static address to be able to use cudagraphs if not is_torchdynamo_compiling(): torch._dynamo.mark_static_address(self.conv_states) self.is_conv_states_initialized = True if recurrent_states is not None: # The shape is always static, so we init as such self.recurrent_states = torch.zeros_like(recurrent_states, dtype=self.dtype, device=self.device) # Mark as static address to be able to use cudagraphs if not is_torchdynamo_compiling(): torch._dynamo.mark_static_address(self.recurrent_states) self.is_recurrent_states_initialized = True def update_conv_state(self, conv_states: torch.Tensor, **kwargs) -> torch.Tensor: """ Update the linear attention cache in-place, and return the necessary conv states. Args: conv_states (`torch.Tensor`): The new conv states to cache. Returns: `torch.Tensor`: The updated conv states. """ # Lazy initialization if not self.is_conv_states_initialized: self.lazy_initialization(conv_states=conv_states) if not self.has_previous_state: # Note that we copy instead of assigning, to preserve the static address for cudagraphs self.conv_states.copy_(conv_states) self.has_previous_state = True # Technically, this update is not logically correct if the prefill is smaller than `conv_kernel_size`, # as it will `roll` anyway in the first decoding step, even though it should `roll` ONLY if the cache is already full. # But since `conv_kernel_size=4` in practice, it's almost impossible to have a smaller prefill so it's mostly fine for now else: # Note that we copy instead of assigning, to preserve the static address for cudagraphs num_new_tokens = conv_states.shape[-1] if num_new_tokens >= self.conv_kernel_size: self.conv_states.copy_(conv_states[..., -self.conv_kernel_size :]) else: new_conv_states = self.conv_states.roll(shifts=-num_new_tokens, dims=-1) new_conv_states[:, :, -num_new_tokens:] = conv_states self.conv_states.copy_(new_conv_states) return self.conv_states def update_recurrent_state(self, recurrent_states: torch.Tensor, **kwargs) -> torch.Tensor: """ Update the linear attention cache in-place, and return the necessary ssm states. Args: smm_states (`torch.Tensor`): The new ssm states to cache. Returns: `torch.Tensor`: The updated ssm states. """ if not self.is_recurrent_states_initialized: self.lazy_initialization(recurrent_states=recurrent_states) # Note that we copy instead of assigning, to preserve the static address for cudagraphs self.recurrent_states.copy_(recurrent_states) return self.recurrent_states class LinearAttentionAndFullAttentionLayer(LinearAttentionLayer, DynamicLayer): # The dynamic Attention part makes it non-compileable is_compileable = False def __init__(self): DynamicLayer.__init__(self) LinearAttentionLayer.__init__(self) def lazy_initialization(self, *args, **kwargs) -> None: # When the Attention cache is used with `update`, `lazy_initialization` is called with 2 positional args if len(args) == 2 and len(kwargs) == 0: DynamicLayer.lazy_initialization(self, *args) # Otherwise, for the LinearAttention cache, when it's called in `update_conv_state` or `update_recurrent_state`, it's # always called with 1 single kwarg (cause it needs to know if it's for the conv or ssm states) if len(args) == 0 and len(kwargs) == 1: LinearAttentionLayer.lazy_initialization(self, **kwargs) def reset(self) -> None: LinearAttentionLayer.reset(self) DynamicLayer.reset(self) def reorder_cache(self, beam_idx: torch.LongTensor): """Reorders the cache for beam search, given the selected beam indices.""" LinearAttentionLayer.reorder_cache(self, beam_idx) DynamicLayer.reorder_cache(self, beam_idx) class Cache: """ A `Cache` is mostly a list of `CacheLayerMixin` objects, one per model layer. It serves as a container for the Cache of each layer. Args: layers (`Optional`, *optional*): A list of pre-created `CacheLayerMixin` or `LinearAttentionCacheLayerMixin`. If omitted (`None`), then `layer_class_to_replicate` will be used. layer_class_to_replicate (`type[CacheLayerMixin | LinearAttentionCacheLayerMixin]`, *optional*): Only used if `layers` is omitted (`None`), in which case it will be used as the base class for each layer, and the layers will be added lazily as soon as `update` is called with a `layer_idx` greater than the current list of layers. offloading (`bool`, *optional*, defaults to `False`): Whether to perform offloading of the layers to `cpu`, to save GPU memory. offload_only_non_sliding (`bool`, *optional*, defaults to `True`): If `offloading` is `True`, this further decides if only the non-sliding layers will be offloaded (because usually the sliding layers are small in size, so there is no need to offload them, and skipping it is faster). """ def __init__( self, layers: list[CacheLayerMixin | LinearAttentionCacheLayerMixin] | None = None, layer_class_to_replicate: type[CacheLayerMixin | LinearAttentionCacheLayerMixin] | None = None, offloading: bool = False, offload_only_non_sliding: bool = True, ): if layers is not None and layer_class_to_replicate is not None: raise ValueError( "You can construct a Cache either from a list `layers` of all the predefined `CacheLayer`, or from a " "`layer_class_to_replicate`, in which case the Cache will append a new layer corresponding to " "`layer_class_to_replicate` for each new call to `update` with an idx not already in the Cache." ) if layers is None and layer_class_to_replicate is None: raise ValueError( "You should provide exactly one of `layers` or `layer_class_to_replicate` to initialize a Cache." ) self.layers = layers if layers is not None else [] self.layer_class_to_replicate = layer_class_to_replicate self.offloading = offloading if self.offloading: self.only_non_sliding = offload_only_non_sliding self.prefetch_stream = torch.Stream() if _is_torch_greater_or_equal_than_2_7 else torch.cuda.Stream() def __repr__(self): return f"{self.__class__.__name__}(layers={self.layers})" def prefetch(self, layer_idx: int, only_non_sliding: bool = True): """ Prefetch a given layer on its device. If `only_non_sliding` is True, it will try to prefetch only the layers which are non-sliding. If the `layer_idx` is outside the range, this will circle back to the first layers. Note that we use a non-default stream for this, to avoid blocking. """ if only_non_sliding: # Try to find next non-sliding, starting at `layer_idx` try: layer_idx = layer_idx + self.is_sliding[layer_idx:].index(False) # In this case, we need to circle back to the beginning except ValueError: layer_idx = self.is_sliding.index(False) else: layer_idx = layer_idx if layer_idx < len(self.layers) else 0 # Prefetch with self.prefetch_stream if _is_torch_greater_or_equal_than_2_7 else torch.cuda.stream(self.prefetch_stream): self.layers[layer_idx].prefetch() def offload(self, layer_idx: int, only_non_sliding: bool = True): """ Offload a given `layer_idx`. If `only_non_sliding` is True, it will offload `layer_idx` only if it is a non-sliding layer. Note that we do it on the default stream, so that we ensure all earlier computation in the layer's `update` methods are finished. """ if not (only_non_sliding and self.is_sliding[layer_idx]): self.layers[layer_idx].offload() def update( self, key_states: torch.Tensor, value_states: torch.Tensor, layer_idx: int, *args, **kwargs ) -> tuple[torch.Tensor, torch.Tensor]: """ Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`. Parameters: key_states (`torch.Tensor`): The new key states to cache. value_states (`torch.Tensor`): The new value states to cache. layer_idx (`int`): The index of the layer to cache the states for. Return: A tuple containing the updated key and value states. """ # In this case, the `layers` were not provided, and we must append as much as `layer_idx` if self.layer_class_to_replicate is not None: while len(self.layers) <= layer_idx: self.layers.append(self.layer_class_to_replicate()) if self.offloading: # Wait for the stream to finish if needed, and start prefetching the next layer torch.cuda.default_stream(key_states.device).wait_stream(self.prefetch_stream) self.prefetch(layer_idx + 1, self.only_non_sliding) keys, values = self.layers[layer_idx].update(key_states, value_states, *args, **kwargs) if self.offloading: self.offload(layer_idx, self.only_non_sliding) return keys, values def update_conv_state(self, conv_states: torch.Tensor, layer_idx: int, **kwargs) -> torch.Tensor: """ Updates the cache with the new `conv_states` for the layer `layer_idx`. Parameters: conv_states (`torch.Tensor`): The new conv states to cache. layer_idx (`int`): The index of the layer to cache the states for. Return: `torch.Tensor`: The updated conv states. """ # NOTE: if we slightly break `update` arg order, we could combine this with it, and allow offloading support # out of the box if not isinstance(self.layers[layer_idx], LinearAttentionCacheLayerMixin): raise ValueError("Cannot call `update_conv_state` on a non-LinearAttention layer!") conv_states = self.layers[layer_idx].update_conv_state(conv_states, **kwargs) return conv_states def update_recurrent_state(self, recurrent_states: torch.Tensor, layer_idx: int, **kwargs) -> torch.Tensor: """ Updates the cache with the new `recurrent_states` for the layer `layer_idx`. Parameters: smm_states (`torch.Tensor`): The new ssm states to cache. layer_idx (`int`): The index of the layer to cache the states for. Return: `torch.Tensor`: The updated ssm states. """ # NOTE: if we slightly break `update` arg order, we could combine this with it, and allow offloading support # out of the box if not isinstance(self.layers[layer_idx], LinearAttentionCacheLayerMixin): raise ValueError("Cannot call `update_conv_state` on a non-LinearAttention layer!") recurrent_states = self.layers[layer_idx].update_recurrent_state(recurrent_states, **kwargs) return recurrent_states def early_initialization( self, batch_size: int, num_heads: int | list[int], head_dim: int | list[int], dtype: torch.dtype, device: torch.device, ): """ Initialize all the layers in advance (it's otherwise lazily initialized on the first `update` call). This is useful for our `export` recipes, as `export` needs everything in advance. """ # To allow different num_heads and head_dim depending on layers, we accept lists if isinstance(num_heads, int): num_heads = [num_heads] * len(self) if isinstance(head_dim, int): head_dim = [head_dim] * len(self) if len(num_heads) != len(self.layers): raise ValueError( f"`num_head` was provided as a list of length {len(num_heads)}, but the Cache currently has {len(self.layers)} layers" ) if len(head_dim) != len(self.layers): raise ValueError( f"`head_dim` was provided as a list of length {len(num_heads)}, but the Cache currently has {len(self.layers)} layers" ) for layer, layer_num_heads, layer_head_dim in zip(self.layers, num_heads, head_dim): # Note that the initialization needs all dimensions (except -2), as well as device and dtype, so we use # this fake tensor approach. It has size 0 on the -2 dimension, so it does not allocate any data (it only # creates an empty tensor with correct shape, dtype and device), which is very efficient and practical fake_kv_tensor = torch.zeros((batch_size, layer_num_heads, 0, layer_head_dim), dtype=dtype, device=device) # Init the layer layer.lazy_initialization(fake_kv_tensor, fake_kv_tensor) def get_seq_length(self, layer_idx: int = 0) -> int: """Returns the sequence length of the cache for the given layer.""" if layer_idx >= len(self.layers): return 0 # For alternating attention/linear attention caches, `get_seq_length` needs to use attention layer idx when called with default layer_idx if not isinstance(self.layers[layer_idx], CacheLayerMixin): # If this is called with non-default arg, raise if layer_idx != 0: raise ValueError( f"You called `get_seq_length` on layer index {layer_idx}, but this layer is a LinearAttention layer, which " "does not track sequence length." ) try: # Use the first attention layer layer_idx = next(idx for idx in range(len(self)) if isinstance(self.layers[idx], CacheLayerMixin)) except StopIteration: raise ValueError( "`get_seq_length` can only be called on Attention layers, and the current Cache seem to only contain " "LinearAttention layers." ) return self.layers[layer_idx].get_seq_length() def has_previous_state(self, layer_idx: int | None = None) -> bool: """Returns whether the LinearAttention layer at index `layer_idx` has previous state or not.""" if layer_idx is not None and layer_idx >= len(self.layers): return False # In this case, use last LinearAttention layer if layer_idx is None: try: layer_idx = next( idx for idx in range(len(self) - 1, -1, -1) if isinstance(self.layers[idx], LinearAttentionCacheLayerMixin) ) except StopIteration: raise ValueError( "`has_previous_state` can only be called on LinearAttention layers, and the current Cache seem to " "only contain Attention layers." ) elif not isinstance(self.layers[layer_idx], LinearAttentionCacheLayerMixin): raise ValueError( f"You called `has_previous_state` on layer index {layer_idx}, but this layer is an Attention layer, which " "does not support calling it." ) return self.layers[layer_idx].has_previous_state def get_mask_sizes(self, query_length: int, layer_idx: int) -> tuple[int, int]: """ Return a tuple (kv_length, kv_offset) corresponding to the length and offset that will be returned for the given layer at `layer_idx`. The masks are then prepared according to the given lengths (kv_length, kv_offset) and patterns for each layer. """ # For DynamicCache, where the layers are created at runtime -> if it was not yet created, the size is # simply the query_length if layer_idx >= len(self.layers): return query_length, 0 # For alternating attention/linear attention caches, `get_mask_sizes` needs to use attention layer idx when called with default layer_idx if not isinstance(self.layers[layer_idx], CacheLayerMixin): # If this is called with non-default arg, raise if layer_idx != 0: raise ValueError( f"You called `get_mask_sizes` on layer index {layer_idx}, but this layer is a LinearAttention layer, which " "does not track sequence length." ) try: # Use the first attention layer layer_idx = next(idx for idx in range(len(self)) if isinstance(self.layers[idx], CacheLayerMixin)) except StopIteration: raise ValueError( "`get_mask_sizes` can only be called on Attention layers, and the current Cache seem to only contain " "LinearAttention layers." ) return self.layers[layer_idx].get_mask_sizes(query_length) def get_max_cache_shape(self, layer_idx: int = 0) -> int: """Returns maximum sequence length of the cache object. Dynamic caches do not have a maximum length.""" # For DynamicCache, where the layers are created at runtime -> if it was not yet created, return -1 # as DynamicLayer does if layer_idx >= len(self.layers): return -1 return self.layers[layer_idx].get_max_cache_shape() def reset(self): """Recursively reset all layers tensors""" for layer_idx in range(len(self.layers)): self.layers[layer_idx].reset() def reorder_cache(self, beam_idx: torch.LongTensor): """Reorder the cache for beam search""" for layer_idx in range(len(self.layers)): self.layers[layer_idx].reorder_cache(beam_idx) def crop(self, max_length: int): """Crop the cache to the given length""" for layer_idx in range(len(self.layers)): self.layers[layer_idx].crop(max_length) def batch_repeat_interleave(self, repeats: int): """Repeat and interleave the cache""" for layer_idx in range(len(self.layers)): self.layers[layer_idx].batch_repeat_interleave(repeats) def batch_select_indices(self, indices: torch.Tensor): """Select indices from the cache""" for layer_idx in range(len(self.layers)): self.layers[layer_idx].batch_select_indices(indices) @property def max_batch_size(self) -> int: """Return the maximum batch size of the cache""" values = [layer.max_batch_size for layer in self.layers] if len(set(values)) > 1: raise ValueError(f"Max batch size is not consistent across layers: {values}") return values[0] @property def max_cache_len(self) -> int: """Return the maximum cache length of the cache""" values = [layer.max_cache_len for layer in self.layers] return max(values) @property def is_compileable(self) -> bool: """Return whether the cache is compilable""" # For DynamicCache dispatching the layers lazily (otherwise, all([]) is True) if len(self.layers) == 0: return False return all(layer.is_compileable for layer in self.layers) @property def is_initialized(self) -> bool: """Return whether the cache data is initialized""" return len(self.layers) > 0 and all(layer.is_initialized for layer in self.layers) @property def is_sliding(self) -> list[bool]: """Return whether the layers of the cache are sliding window""" return [getattr(layer, "is_sliding", False) for layer in self.layers] def __len__(self): """ This value corresponds to the number of layers in the model. """ # Note: for DynamicCache, layers are initialized lazily, so this will not be accurate before the first # forward through all the layers return len(self.layers) class DynamicCache(Cache): """ A cache that grows dynamically as more tokens are generated. This is the default for generative models. It stores the key and value states as a list of `CacheLayer`, one for each layer. The expected shape for each tensor in the `CacheLayer`s is `[batch_size, num_heads, seq_len, head_dim]`. If a config is passed, it will additionally check for sliding or hybrid cache structure, greatly reducing the memory requirement of the cached tensors to `[batch_size, num_heads, min(seq_len, sliding_window), head_dim]`. See `Cache` for details on common methods that are implemented by all cache classes. Args: ddp_cache_data (`Iterable[tuple[torch.Tensor, torch.Tensor]]`, *optional*): It was originally added for compatibility with `torch.distributed` (DDP). In a nutshell, it is `map(gather_map, zip(*caches))`, i.e. each item in the iterable contains the key and value states for a layer gathered across replicas by torch.distributed (shape=[global batch size, num_heads, seq_len, head_dim]). Note: it needs to be the 1st arg as well to work correctly config (`PreTrainedConfig`, *optional*): The config of the model for which this Cache will be used. If passed, it will be used to check for sliding or hybrid layer structure, greatly reducing the memory requirement of the cached tensors to `[batch_size, num_heads, min(seq_len, sliding_window), head_dim]`. offloading (`bool`, *optional*, defaults to `False`): Whether to perform offloading of the layers to `cpu`, to save GPU memory. offload_only_non_sliding (`bool`, *optional*, defaults to `False`): If `offloading` is `True`, this further decides if only the non-sliding layers will be offloaded (because usually the sliding layers are small in size, so there is no need to offload them, and skipping it is faster). Example: ```python >>> from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache >>> model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2-0.5B-Instruct") >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-0.5B-Instruct") >>> inputs = tokenizer(text="My name is Qwen2", return_tensors="pt") >>> # Prepare a cache class and pass it to model's forward >>> past_key_values = DynamicCache(config=model.config) >>> outputs = model(**inputs, past_key_values=past_key_values, use_cache=True) >>> outputs.past_key_values # access cache filled with key/values from generation ``` """ def __init__( self, ddp_cache_data: Iterable[tuple[torch.Tensor | None, ...]] | None = None, config: PreTrainedConfig | None = None, offloading: bool = False, offload_only_non_sliding: bool = False, ): layers = [] # If a config is passed, use it to infer the layer types and initialize accordingly if config is not None: decoder_config = config.get_text_config(decoder=True) sliding_window = getattr(decoder_config, "sliding_window", None) or getattr( decoder_config, "attention_chunk_size", None ) layer_types = getattr(decoder_config, "layer_types", None) if layer_types is None: layer_types = [] for _ in range(decoder_config.num_hidden_layers): if sliding_window is not None: layer_types.append("sliding_attention") else: layer_types.append("full_attention") # Some models have shared layers thus no cache is needed for them (e.g. Gemma3n) if hasattr(decoder_config, "num_kv_shared_layers"): layer_types = layer_types[: -decoder_config.num_kv_shared_layers] for layer_type in layer_types: # From a cache point of view, both sliding and chunked are the same in how they should behave and how many # states they should return - only the mask changes to make them different at the end! if layer_type in ("sliding_attention", "chunked_attention"): layers.append(DynamicSlidingWindowLayer(sliding_window=sliding_window)) # Note: we want moe layers to be LinearAttentionLayer, so that we can correctly grab sequence length etc from attention layers. # Since moe layers will stay empty (they don't need any cache), we don't want them to collide for mask creation etc # TODO: maybe use a dummy layer in those cases, or a dictionary {idx: Layer} for self.layers, so that we can skip # the indices we don't need elif layer_type in ("mamba", "conv", "linear_attention", "moe"): layers.append(LinearAttentionLayer()) elif layer_type == "hybrid": layers.append(LinearAttentionAndFullAttentionLayer()) else: layers.append(DynamicLayer()) # In this case, use the passed data to already fill in the Cache if ddp_cache_data is not None: # Init all the layers with the data for layer_idx, kv_and_optional_sliding in enumerate(ddp_cache_data): # If the config was not passed above, initialize a new cache layer for each entry of the ddp_data if config is None: # kv_and_optional_sliding contains at least two elements: the key and value states. It can also # contain a third element, which is an optional sliding window tensor. sliding_window_tensor = kv_and_optional_sliding[2] if len(kv_and_optional_sliding) == 3 else None # If there is a sliding window tensor, use it to initialize the layer if sliding_window_tensor is not None: # Since the same layer is dispatched across replicas, sliding_window is the same for all sliding_window = sliding_window_tensor[0].item() layers.append(DynamicSlidingWindowLayer(sliding_window=sliding_window)) else: layers.append(DynamicLayer()) # Update the layer with the data _, _ = layers[layer_idx].update(kv_and_optional_sliding[0], kv_and_optional_sliding[1]) # If neither of config nor ddp_data was passed, then simply lazy init a full cache of DynamicLayer if len(layers) == 0: super().__init__( layer_class_to_replicate=DynamicLayer, offloading=offloading, offload_only_non_sliding=offload_only_non_sliding, ) else: super().__init__(layers=layers, offloading=offloading, offload_only_non_sliding=offload_only_non_sliding) def __iter__(self): for layer in self.layers: yield layer.keys, layer.values, getattr(layer, "_sliding_window_tensor", None) class StaticCache(Cache): """ Static Cache class to be used with `torch.compile(model)` and `torch.export()`. It will check the `config` for potential hybrid cache structure, and initialize each layer accordingly. See `Cache` for details on common methods that are implemented by all cache classes. Args: config (`PreTrainedConfig`): The config of the model for which this Cache will be used. It will be used to check for sliding or hybrid layer structure, and initialize each layer accordingly. max_cache_len (`int`): The maximum number of tokens that this Cache should hold. offloading (`bool`, *optional*, defaults to `False`): Whether to perform offloading of the layers to `cpu`, to save GPU memory. offload_only_non_sliding (`bool`, *optional*, defaults to `True`): If `offloading` is `True`, this further decides if only the non-sliding layers will be offloaded (because usually the sliding layers are small in size, so there is no need to offload them, and skipping it is faster). Example: ```python >>> from transformers import AutoTokenizer, AutoModelForCausalLM, StaticCache >>> model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf") >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf") >>> inputs = tokenizer(text="My name is Llama", return_tensors="pt") >>> # Prepare a cache class and pass it to model's forward >>> # Leave empty space for 10 new tokens, which can be used when calling forward iteratively 10 times to generate >>> max_generated_length = inputs.input_ids.shape[1] + 10 >>> past_key_values = StaticCache(config=model.config, max_cache_len=max_generated_length) >>> outputs = model(**inputs, past_key_values=past_key_values, use_cache=True) >>> outputs.past_key_values # access cache filled with key/values from generation StaticCache() ``` """ # Pass-in kwargs as well to avoid crashing for BC (it used more arguments before) def __init__( self, config: PreTrainedConfig, max_cache_len: int, offloading: bool = False, offload_only_non_sliding: bool = True, **kwargs, ): config = config.get_text_config(decoder=True) layer_types = getattr(config, "layer_types", None) # If `layer_types` is not explicitly provided, infer if the model is fully sliding if layer_types is None: if getattr(config, "sliding_window", None) is not None: layer_types = ["sliding_attention" for _ in range(config.num_hidden_layers)] elif getattr(config, "attention_chunk_size", None) is not None: layer_types = ["chunked_attention" for _ in range(config.num_hidden_layers)] else: layer_types = ["full_attention" for _ in range(config.num_hidden_layers)] # Some models have shared layers thus no cache is needed for them (e.g. Gemma3n) if hasattr(config, "num_kv_shared_layers"): layer_types = layer_types[: -config.num_kv_shared_layers] layers = [] for layer_type in layer_types: if layer_type == "sliding_attention": layer = StaticSlidingWindowLayer(max_cache_len=max_cache_len, sliding_window=config.sliding_window) elif layer_type == "chunked_attention": # From a cache point of view, both sliding and chunked are the same in how they should behave and how many # states they should return - only the mask changes to make them different at the end! layer = StaticSlidingWindowLayer( max_cache_len=max_cache_len, sliding_window=config.attention_chunk_size ) # LinearAttention layers are static by essence - using `"moe"` as well is a trick, see the comment about it on DynamicCache elif layer_type in ("mamba", "conv", "linear_attention", "moe"): layer = LinearAttentionLayer() else: layer = StaticLayer(max_cache_len=max_cache_len) layers.append(layer) super().__init__(layers=layers, offloading=offloading, offload_only_non_sliding=offload_only_non_sliding) class QuantizedCache(Cache): """ A quantizer cache similar to what is described in the [KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache paper](https://huggingface.co/papers/2402.02750). It allows the model to generate longer sequence length without allocating too much memory for keys and values by applying quantization. The cache has two types of storage, one for original precision and one for the quantized cache. A `residual length` is set as a maximum capacity for the original precision cache. When the length goes beyond maximum capacity, the original precision cache is discarded and moved into the quantized cache. The quantization is done per-channel with a set `q_group_size` for both keys and values, in contrast to what was described in the paper. See `Cache` for details on common methods that are implemented by all cache classes. Args: backend (`str`): The quantization backend to use. One of `("quanto", "hqq"). config (`PreTrainedConfig`): The config of the model for which this Cache will be used. nbits (`int`, *optional*, defaults to 4): The number of bits for quantization. axis_key (`int`, *optional*, defaults to 0): The axis on which to quantize the keys. axis_value (`int`, *optional*, defaults to 0): The axis on which to quantize the values. q_group_size (`int`, *optional*, defaults to 64): Quantization is done per-channel according to a set `q_group_size` for both keys and values. residual_length (`int`, *optional*, defaults to 128): Maximum capacity for the original precision cache """ def __init__( self, backend: str, config: PreTrainedConfig, nbits: int = 4, axis_key: int = 0, axis_value: int = 0, q_group_size: int = 64, residual_length: int = 128, ): if backend == "quanto": layer_class = QuantoQuantizedLayer elif backend == "hqq": layer_class = HQQQuantizedLayer else: raise ValueError(f"Unknown quantization backend `{backend}`") config = config.get_text_config(decoder=True) layers = [ layer_class(nbits, axis_key, axis_value, q_group_size, residual_length) for _ in range(config.num_hidden_layers) ] super().__init__(layers=layers) class EncoderDecoderCache(Cache): """ Base, abstract class for all encoder-decoder caches. Can be used to hold combinations of self-attention and cross-attention caches. See `Cache` for details on common methods that are implemented by all cache classes. Args: caches (`Iterable`): Usually an iterable of length 2, containing 2 `Cache` objects, the first one for self-attention, the second one for cross-attention. Can optionally also be an iterable of length 1, containing a `tuple[tuple[torch.Tensor]]` (usually used for compatibility with torch dp and ddp). Example: ```python >>> from transformers import AutoProcessor, AutoModelForCausalLM, DynamicCache, EncoderDecoderCache >>> model = AutoModelForCausalLM.from_pretrained("openai/whisper-small") >>> processor = AutoProcessor.from_pretrained("openai/whisper-small") >>> inputs = processor(audio=YOUR-AUDIO, return_tensors="pt") >>> # Prepare cache classes for encoder and decoder and pass it to model's forward >>> self_attention_cache = DynamicCache(config=self.config) >>> cross_attention_cache = DynamicCache(config=self.config) >>> past_key_values = EncoderDecoderCache(self_attention_cache, cross_attention_cache) >>> outputs = model(**inputs, past_key_values=past_key_values, use_cache=True) >>> outputs.past_key_values # access cache filled with key/values from generation EncoderDecoderCache() ``` """ def __init__(self, *caches) -> None: # For dp and ddp support, if only one argument is passed, it should be an iterable of DynamicCache ddp data if len(caches) == 1: self_attention_cache_data, cross_attention_cache_data = [], [] for combined_cache_data in caches[0]: if len(combined_cache_data) == 6: # two tuple of style (self_attn_k, self_attn_v, self_attn_sliding) self_attention_cache_data.append(combined_cache_data[:3]) cross_attention_cache_data.append(combined_cache_data[3:]) # To support old DDP-style init, we handle the case where the tuple has no sliding window tensor elif len(combined_cache_data) == 4: # two tuple of style (self_attn_k, self_attn_v) self_attention_cache_data.append(combined_cache_data[:2]) cross_attention_cache_data.append(combined_cache_data[2:]) else: raise ValueError(f"Expected {len(combined_cache_data) = } to be 4 or 6.\n{combined_cache_data = }") self.self_attention_cache = DynamicCache(self_attention_cache_data) self.cross_attention_cache = DynamicCache(cross_attention_cache_data) # Otherwise, we should get two arguments, a self-attention cache and a cross-attention cache elif len(caches) == 2: if not isinstance(caches[0], Cache) or not isinstance(caches[1], Cache): raise TypeError(f"One of the two arguments is not a Cache: {type(caches[0]) = }, {type(caches[1]) = }") self.self_attention_cache = caches[0] self.cross_attention_cache = caches[1] # Error case else: raise ValueError(f"Expected 1 or 2 arguments, got {len(caches)}") self.is_updated = {} for layer_idx in range(len(self.cross_attention_cache)): self.is_updated[layer_idx] = bool(self.cross_attention_cache.get_seq_length(layer_idx) > 0) def __iter__(self): """Returns tuples of style (self_attn_k, self_attn_v, self_attn_sliding, cross_attn_k, cross_attn_v, cross_attn_sliding)""" for self_attention_layer, cross_attention_layer in zip(self.self_attention_cache, self.cross_attention_cache): yield self_attention_layer + cross_attention_layer def __repr__(self) -> str: return ( f"{self.__class__.__name__}(self_attention_cache={self.self_attention_cache}, cross_attention_cache=" f"{self.cross_attention_cache})" ) def __len__(self): """ Support for backwards-compatible `past_key_values` length, e.g. `len(past_key_values)`. This value corresponds to the number of layers in the model. """ return len(self.self_attention_cache) def get_seq_length(self, layer_idx: int = 0) -> int: """Returns the sequence length of the cached states. A layer index can be optionally passed.""" return self.self_attention_cache.get_seq_length(layer_idx) def reset(self): self.self_attention_cache.reset() self.cross_attention_cache.reset() for layer_idx in self.is_updated: self.is_updated[layer_idx] = False def reorder_cache(self, beam_idx: torch.LongTensor): """Reorders the cache for beam search, given the selected beam indices.""" self.self_attention_cache.reorder_cache(beam_idx) self.cross_attention_cache.reorder_cache(beam_idx) def check_dynamic_cache(self, method: str): if not ( isinstance(self.self_attention_cache, DynamicCache) and isinstance(self.cross_attention_cache, DynamicCache) ): raise TypeError( f"`{method}` is only defined for dynamic cache, got {self.self_attention_cache.__str__()} for the self " f"attention cache and {self.cross_attention_cache.__str__()} for the cross attention cache." ) # TODO(gante, sanchit-gandhi): move following functionality into `.generate` def crop(self, maximum_length: int): """ Crop the past key values up to a new `maximum_length` in terms of tokens. `maximum_length` can also be negative to remove `maximum_length` tokens. This is used in assisted decoding and contrastive search (on the Hub). """ self.check_dynamic_cache(self.crop.__name__) self.self_attention_cache.crop(maximum_length) def batch_repeat_interleave(self, repeats: int): """Repeat the cache `repeats` times in the batch dimension. Used in contrastive search (on the Hub).""" self.check_dynamic_cache(self.batch_repeat_interleave.__name__) self.self_attention_cache.batch_repeat_interleave(repeats) self.cross_attention_cache.batch_repeat_interleave(repeats) def batch_select_indices(self, indices: torch.Tensor): """Only keep the `indices` in the batch dimension of the cache. Used in contrastive search (on the Hub).""" self.check_dynamic_cache(self.batch_select_indices.__name__) self.self_attention_cache.batch_select_indices(indices) self.cross_attention_cache.batch_select_indices(indices) def get_max_cache_shape(self) -> int: """Returns the maximum sequence length (i.e. max capacity) of the cache object""" return self.self_attention_cache.get_max_cache_shape() def get_mask_sizes(self, query_length: int, layer_idx: int) -> tuple[int, int]: return self.self_attention_cache.get_mask_sizes(query_length, layer_idx) @property def is_sliding(self): return self.self_attention_cache.is_sliding @property def is_compileable(self) -> bool: return self.self_attention_cache.is_compileable # Deprecated alias: SlidingWindowCache was removed in transformers v5. StaticCache is the replacement. SlidingWindowCache = StaticCache