From 730a440734e1fb47c903c17e3231dac18e3e5fd6 Mon Sep 17 00:00:00 2001 From: Raushan Turganbay Date: Fri, 21 Jun 2024 09:28:14 +0500 Subject: [PATCH] Deprecate legacy cache + use cache position (#31491) * tmp * update models * revert utils * delete * Update src/transformers/models/dbrx/modeling_dbrx.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * modify warning msg --------- Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> --- src/transformers/models/dbrx/modeling_dbrx.py | 4 + .../models/llama/modeling_llama.py | 4 + .../models/mistral/modeling_mistral.py | 4 + .../models/mixtral/modeling_mixtral.py | 229 +++++++++++------ src/transformers/models/olmo/modeling_olmo.py | 4 + .../models/persimmon/modeling_persimmon.py | 199 +++++++++++---- src/transformers/models/phi/modeling_phi.py | 234 ++++++++++++----- src/transformers/models/phi3/modeling_phi3.py | 220 +++++++++++----- .../models/qwen2/modeling_qwen2.py | 238 ++++++++++++------ .../models/qwen2_moe/modeling_qwen2_moe.py | 238 ++++++++++++------ .../models/stablelm/modeling_stablelm.py | 233 ++++++++++++----- .../models/starcoder2/modeling_starcoder2.py | 238 ++++++++++++------ 12 files changed, 1296 insertions(+), 549 deletions(-) diff --git a/src/transformers/models/dbrx/modeling_dbrx.py b/src/transformers/models/dbrx/modeling_dbrx.py index cbb2231b5a0e..60b4700c91ed 100644 --- a/src/transformers/models/dbrx/modeling_dbrx.py +++ b/src/transformers/models/dbrx/modeling_dbrx.py @@ -1119,6 +1119,10 @@ def forward( if use_cache and not isinstance(past_key_values, Cache): # kept for BC (non `Cache` `past_key_values` inputs) return_legacy_cache = True past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" + ) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 diff --git a/src/transformers/models/llama/modeling_llama.py b/src/transformers/models/llama/modeling_llama.py index 77d95d1e89bd..7f3c30703304 100644 --- a/src/transformers/models/llama/modeling_llama.py +++ b/src/transformers/models/llama/modeling_llama.py @@ -926,6 +926,10 @@ def forward( if use_cache and not isinstance(past_key_values, Cache): # kept for BC (non `Cache` `past_key_values` inputs) return_legacy_cache = True past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" + ) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 diff --git a/src/transformers/models/mistral/modeling_mistral.py b/src/transformers/models/mistral/modeling_mistral.py index 788c5e6ee51b..73738dade42b 100644 --- a/src/transformers/models/mistral/modeling_mistral.py +++ b/src/transformers/models/mistral/modeling_mistral.py @@ -927,6 +927,10 @@ def forward( if use_cache and not isinstance(past_key_values, Cache): past_key_values = DynamicCache.from_legacy_cache(past_key_values) return_legacy_cache = True + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" + ) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 diff --git a/src/transformers/models/mixtral/modeling_mixtral.py b/src/transformers/models/mixtral/modeling_mixtral.py index 7680a28e2b9c..019f69d0eee2 100644 --- a/src/transformers/models/mixtral/modeling_mixtral.py +++ b/src/transformers/models/mixtral/modeling_mixtral.py @@ -30,10 +30,10 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN -from ...cache_utils import Cache, DynamicCache +from ...cache_utils import Cache, DynamicCache, StaticCache from ...modeling_attn_mask_utils import ( + AttentionMaskConverter, _prepare_4d_causal_attention_mask, - _prepare_4d_causal_attention_mask_for_sdpa, ) from ...modeling_outputs import ( MoeCausalLMOutputWithPast, @@ -327,6 +327,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() @@ -351,7 +352,7 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -366,13 +367,9 @@ def forward( f" {attn_weights.size()}" ) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - - attn_weights = attn_weights + attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) @@ -422,6 +419,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ): bsz, q_len, _ = hidden_states.size() @@ -487,7 +485,7 @@ def forward( attention_mask = attention_mask[:, slicing_tokens:] attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -702,6 +700,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. @@ -736,17 +735,15 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) + causal_mask = attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. @@ -758,13 +755,13 @@ def forward( # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. - is_causal = True if self.is_causal and attention_mask is None and q_len > 1 else False + is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, - attn_mask=attention_mask, + attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) @@ -890,6 +887,7 @@ def forward( output_attentions: Optional[bool] = False, output_router_logits: Optional[bool] = False, use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: @@ -906,6 +904,8 @@ def forward( use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. """ residual = hidden_states @@ -920,6 +920,7 @@ def forward( past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = residual + hidden_states @@ -1051,6 +1052,10 @@ def _init_weights(self, module): should not be returned during inference. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. """ @@ -1104,6 +1109,7 @@ def forward( output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, MoeModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_router_logits = ( @@ -1116,17 +1122,10 @@ def forward( return_dict = return_dict if return_dict is not None else self.config.use_return_dict - # retrieve input_ids and inputs_embeds - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape - elif inputs_embeds is not None: - batch_size, seq_length, _ = inputs_embeds.shape - else: - raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") - - past_key_values_length = 0 + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -1135,55 +1134,29 @@ def forward( ) use_cache = False - if use_cache: - use_legacy_cache = not isinstance(past_key_values, Cache) - if use_legacy_cache: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - past_key_values_length = past_key_values.get_usable_length(seq_length) - - if position_ids is None: - device = input_ids.device if input_ids is not None else inputs_embeds.device - position_ids = torch.arange( - past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + use_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + use_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" ) - position_ids = position_ids.unsqueeze(0).view(-1, seq_length) - else: - position_ids = position_ids.view(-1, seq_length).long() if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) - if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: - is_padding_right = attention_mask[:, -1].sum().item() != batch_size - if is_padding_right: - raise ValueError( - "You are attempting to perform batched generation with padding_side='right'" - " this may lead to unexpected behaviour for Flash Attention version of Mixtral. Make sure to " - " call `tokenizer.padding_side = 'left'` before tokenizing the input. " - ) - - if self._attn_implementation == "flash_attention_2": - # 2d mask is passed through the layers - attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None - elif self._attn_implementation == "sdpa" and not output_attentions: - # output_attentions=True can not be supported when using SDPA, and we fall back on - # the manual implementation that requires a 4D causal mask in all cases. - attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, - ) - else: - # 4d mask is passed through the layers - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) hidden_states = inputs_embeds @@ -1201,22 +1174,24 @@ def forward( layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, - attention_mask, + causal_mask, position_ids, past_key_values, output_attentions, output_router_logits, use_cache, + cache_position, ) else: layer_outputs = decoder_layer( hidden_states, - attention_mask=attention_mask, + attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, output_router_logits=output_router_logits, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = layer_outputs[0] @@ -1254,6 +1229,87 @@ def forward( router_logits=all_router_logits, ) + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + class MixtralForCausalLM(MixtralPreTrainedModel): _tied_weights_keys = ["lm_head.weight"] @@ -1303,6 +1359,7 @@ def forward( output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, MoeCausalLMOutputWithPast]: r""" Args: @@ -1352,6 +1409,7 @@ def forward( output_hidden_states=output_hidden_states, output_router_logits=output_router_logits, return_dict=return_dict, + cache_position=cache_position, ) hidden_states = outputs[0] @@ -1405,15 +1463,21 @@ def prepare_inputs_for_generation( attention_mask=None, inputs_embeds=None, output_router_logits=False, + cache_position=None, + use_cache=True, **kwargs, ): past_length = 0 # Omit tokens covered by past_key_values if past_key_values is not None: # Past key values are always initialized with a `Cache` object -> no need for if-else anymore - cache_length = past_key_values.get_seq_length() - past_length = past_key_values.seen_tokens - max_cache_length = past_key_values.get_max_length() + past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() + max_cache_length = ( + torch.tensor(past_key_values.get_max_length(), device=input_ids.device) + if past_key_values.get_max_length() is not None + else None + ) + cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where @@ -1449,13 +1513,20 @@ def prepare_inputs_for_generation( else: model_inputs = {"input_ids": input_ids} + input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] + if cache_position is None: + cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) + elif use_cache: + cache_position = cache_position[-input_length:] + model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, - "use_cache": kwargs.get("use_cache"), + "use_cache": use_cache, "attention_mask": attention_mask, "output_router_logits": output_router_logits, + "cache_position": cache_position, } ) return model_inputs diff --git a/src/transformers/models/olmo/modeling_olmo.py b/src/transformers/models/olmo/modeling_olmo.py index 50ec015e521c..0458f916d375 100644 --- a/src/transformers/models/olmo/modeling_olmo.py +++ b/src/transformers/models/olmo/modeling_olmo.py @@ -907,6 +907,10 @@ def forward( if use_cache and not isinstance(past_key_values, Cache): # kept for BC (non `Cache` `past_key_values` inputs) return_legacy_cache = True past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" + ) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 diff --git a/src/transformers/models/persimmon/modeling_persimmon.py b/src/transformers/models/persimmon/modeling_persimmon.py index 7ccfdd49e87e..fc7bcb74f6bb 100644 --- a/src/transformers/models/persimmon/modeling_persimmon.py +++ b/src/transformers/models/persimmon/modeling_persimmon.py @@ -28,8 +28,10 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN -from ...cache_utils import Cache, DynamicCache -from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask +from ...cache_utils import Cache, DynamicCache, StaticCache +from ...modeling_attn_mask_utils import ( + AttentionMaskConverter, +) from ...modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, @@ -275,6 +277,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() @@ -322,7 +325,12 @@ def forward( if past_key_value is not None: # Specific to RoPE models with partial rotation - cache_kwargs = {"sin": sin, "cos": cos, "partial_rotation_size": self.rotary_emb.dim} + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_emb.dim, + "cache_position": cache_position, + } key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) @@ -333,12 +341,9 @@ def forward( f" {attn_weights.size()}" ) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - attn_weights = attn_weights + attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dtype=torch.float32, dim=-1).to(query_states.dtype) @@ -381,6 +386,7 @@ def forward( past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: @@ -414,6 +420,7 @@ def forward( past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = residual + hidden_states @@ -544,6 +551,10 @@ def _init_weights(self, module): more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. """ @@ -592,6 +603,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( @@ -601,18 +613,10 @@ def forward( return_dict = return_dict if return_dict is not None else self.config.use_return_dict - # retrieve input_ids and inputs_embeds - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape - elif inputs_embeds is not None: - batch_size, seq_length, _ = inputs_embeds.shape - else: - raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") - - seq_length_with_past = seq_length - past_key_values_length = 0 + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -621,29 +625,28 @@ def forward( ) use_cache = False - if use_cache: - use_legacy_cache = not isinstance(past_key_values, Cache) - if use_legacy_cache: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - past_key_values_length = past_key_values.get_usable_length(seq_length) - seq_length_with_past = seq_length_with_past + past_key_values_length - - if position_ids is None: - device = input_ids.device if input_ids is not None else inputs_embeds.device - position_ids = torch.arange( - past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + use_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + use_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" ) - position_ids = position_ids.unsqueeze(0) if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) - # embed positions - if attention_mask is None: - attention_mask = torch.ones( - (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions ) hidden_states = inputs_embeds @@ -661,19 +664,22 @@ def forward( layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, - attention_mask, + causal_mask, position_ids, past_key_values, output_attentions, + use_cache, + cache_position, ) else: layer_outputs = decoder_layer( hidden_states, - attention_mask=attention_mask, + attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = layer_outputs[0] @@ -703,6 +709,87 @@ def forward( attentions=all_self_attns, ) + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + class PersimmonForCausalLM(PersimmonPreTrainedModel): _tied_weights_keys = ["lm_head.weight"] @@ -755,6 +842,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: r""" Args: @@ -799,6 +887,7 @@ def forward( output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, + cache_position=cache_position, ) hidden_states = outputs[0] @@ -830,14 +919,25 @@ def forward( ) def prepare_inputs_for_generation( - self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + use_cache=True, + **kwargs, ): past_length = 0 if past_key_values is not None: # Past key values are always initialized with a `Cache` object -> no need for if-else anymore - cache_length = past_key_values.get_seq_length() - past_length = past_key_values.seen_tokens - max_cache_length = past_key_values.get_max_length() + past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() + max_cache_length = ( + torch.tensor(past_key_values.get_max_length(), device=input_ids.device) + if past_key_values.get_max_length() is not None + else None + ) + cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where @@ -873,12 +973,19 @@ def prepare_inputs_for_generation( else: model_inputs = {"input_ids": input_ids} + input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] + if cache_position is None: + cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) + elif use_cache: + cache_position = cache_position[-input_length:] + model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, - "use_cache": kwargs.get("use_cache"), + "use_cache": use_cache, "attention_mask": attention_mask, + "cache_position": cache_position, } ) return model_inputs diff --git a/src/transformers/models/phi/modeling_phi.py b/src/transformers/models/phi/modeling_phi.py index a2c3793c0119..0c68de968d16 100644 --- a/src/transformers/models/phi/modeling_phi.py +++ b/src/transformers/models/phi/modeling_phi.py @@ -26,10 +26,9 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN -from ...cache_utils import Cache, DynamicCache +from ...cache_utils import Cache, DynamicCache, StaticCache from ...modeling_attn_mask_utils import ( - _prepare_4d_causal_attention_mask, - _prepare_4d_causal_attention_mask_for_sdpa, + AttentionMaskConverter, ) from ...modeling_outputs import ( BaseModelOutputWithPast, @@ -307,6 +306,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() @@ -350,7 +350,12 @@ def forward( key_states = torch.cat((key_rot, key_pass), dim=-1) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos, "partial_rotation_size": self.rotary_emb.dim} + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_emb.dim, + "cache_position": cache_position, + } key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) @@ -368,11 +373,8 @@ def forward( ) if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - attn_weights = attn_weights + attention_mask + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights += causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype) @@ -421,6 +423,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: # PhiFlashAttention2 attention does not support output_attentions @@ -466,7 +469,12 @@ def forward( key_states = torch.cat((key_rot, key_pass), dim=-1) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos, "partial_rotation_size": self.rotary_emb.dim} + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_emb.dim, + "cache_position": cache_position, + } key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache @@ -634,6 +642,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. @@ -694,12 +703,21 @@ def forward( key_states = torch.cat((key_rot, key_pass), dim=-1) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos, "partial_rotation_size": self.rotary_emb.dim} + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_emb.dim, + "cache_position": cache_position, + } key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) + causal_mask = attention_mask + if attention_mask is not None: + causal_mask = causal_mask[:, :, :, : key_states.shape[-2]] + # SDPA with memory-efficient backend is broken in torch==2.1.2 when using non-contiguous inputs and a custom # attn_mask, so we need to call `.contiguous()` here. This was fixed in torch==2.2.0. # Reference: https://github.com/pytorch/pytorch/issues/112577 @@ -710,13 +728,13 @@ def forward( # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. - is_causal = True if self.is_causal and attention_mask is None and q_len > 1 else False + is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, - attn_mask=attention_mask, + attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) @@ -752,6 +770,7 @@ def forward( output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, past_key_value: Optional[Tuple[torch.Tensor]] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: @@ -769,6 +788,8 @@ def forward( If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence """ residual = hidden_states @@ -783,6 +804,7 @@ def forward( past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) attn_outputs = self.resid_dropout(attn_outputs) @@ -909,6 +931,10 @@ def _init_weights(self, module): more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. """ @@ -961,6 +987,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( @@ -970,17 +997,10 @@ def forward( return_dict = return_dict if return_dict is not None else self.config.use_return_dict - # retrieve input_ids and inputs_embeds - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape[:2] - elif inputs_embeds is not None: - batch_size, seq_length = inputs_embeds.shape[:2] - else: - raise ValueError("You have to specify either input_ids or inputs_embeds") - - past_key_values_length = 0 + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -989,41 +1009,31 @@ def forward( ) use_cache = False - if use_cache: - use_legacy_cache = not isinstance(past_key_values, Cache) - if use_legacy_cache: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - past_key_values_length = past_key_values.get_usable_length(seq_length) - - if position_ids is None: - device = input_ids.device if input_ids is not None else inputs_embeds.device - position_ids = torch.arange( - past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + use_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + use_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" ) - position_ids = position_ids.unsqueeze(0) if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) - inputs_embeds = self.embed_dropout(inputs_embeds) - - # Attention mask. - if self._use_flash_attention_2: - # 2d mask is passed through the layers - attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None - elif self._use_sdpa and not output_attentions: - attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - ) - else: - # 4d mask is passed through the layers - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) + + inputs_embeds = self.embed_dropout(inputs_embeds) hidden_states = inputs_embeds # decoder layers @@ -1039,19 +1049,22 @@ def forward( layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, - attention_mask, + causal_mask, position_ids, - past_key_values, output_attentions, + use_cache, + past_key_values, + cache_position, ) else: layer_outputs = decoder_layer( hidden_states, - attention_mask=attention_mask, + attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = layer_outputs[0] @@ -1080,6 +1093,87 @@ def forward( attentions=all_self_attns, ) + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + class PhiForCausalLM(PhiPreTrainedModel): _tied_weights_keys = ["lm_head.weight"] @@ -1132,6 +1226,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: r""" Args: @@ -1176,6 +1271,7 @@ def forward( output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, + cache_position=cache_position, ) hidden_states = outputs[0] @@ -1209,14 +1305,25 @@ def forward( # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation def prepare_inputs_for_generation( - self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + use_cache=True, + **kwargs, ): past_length = 0 if past_key_values is not None: # Past key values are always initialized with a `Cache` object -> no need for if-else anymore - cache_length = past_key_values.get_seq_length() - past_length = past_key_values.seen_tokens - max_cache_length = past_key_values.get_max_length() + past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() + max_cache_length = ( + torch.tensor(past_key_values.get_max_length(), device=input_ids.device) + if past_key_values.get_max_length() is not None + else None + ) + cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where @@ -1252,12 +1359,19 @@ def prepare_inputs_for_generation( else: model_inputs = {"input_ids": input_ids} + input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] + if cache_position is None: + cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) + elif use_cache: + cache_position = cache_position[-input_length:] + model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, - "use_cache": kwargs.get("use_cache"), + "use_cache": use_cache, "attention_mask": attention_mask, + "cache_position": cache_position, } ) return model_inputs diff --git a/src/transformers/models/phi3/modeling_phi3.py b/src/transformers/models/phi3/modeling_phi3.py index e14785bd1f8b..05becfe3a436 100644 --- a/src/transformers/models/phi3/modeling_phi3.py +++ b/src/transformers/models/phi3/modeling_phi3.py @@ -26,8 +26,10 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN -from ...cache_utils import Cache, DynamicCache -from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask +from ...cache_utils import Cache, DynamicCache, StaticCache +from ...modeling_attn_mask_utils import ( + AttentionMaskConverter, +) from ...modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, @@ -331,6 +333,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.") @@ -360,7 +363,7 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -376,11 +379,8 @@ def forward( ) if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - attn_weights = attn_weights + attention_mask + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights += causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype) @@ -429,6 +429,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: # Phi3FlashAttention2 attention does not support output_attentions @@ -503,7 +504,7 @@ def forward( attention_mask = attention_mask[:, slicing_tokens:] attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -723,6 +724,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. @@ -759,17 +761,15 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) + causal_mask = attention_mask if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. @@ -781,13 +781,13 @@ def forward( # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. - is_causal = True if self.is_causal and attention_mask is None and q_len > 1 else False + is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, - attn_mask=attention_mask, + attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) @@ -829,6 +829,7 @@ def forward( past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: @@ -860,6 +861,7 @@ def forward( past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = residual + self.resid_attn_dropout(attn_outputs) @@ -992,6 +994,10 @@ def _init_weights(self, module): more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. """ @@ -1042,6 +1048,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( @@ -1051,17 +1058,10 @@ def forward( return_dict = return_dict if return_dict is not None else self.config.use_return_dict - # retrieve input_ids and inputs_embeds - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape[:2] - elif inputs_embeds is not None: - batch_size, seq_length = inputs_embeds.shape[:2] - else: - raise ValueError("You have to specify either input_ids or inputs_embeds") - - past_key_values_length = 0 + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -1070,45 +1070,29 @@ def forward( ) use_cache = False - if use_cache: - use_legacy_cache = not isinstance(past_key_values, Cache) - if use_legacy_cache: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - past_key_values_length = past_key_values.get_usable_length(seq_length) - - if position_ids is None: - device = input_ids.device if input_ids is not None else inputs_embeds.device - position_ids = torch.arange( - past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + use_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + use_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" ) - position_ids = position_ids.unsqueeze(0).view(-1, seq_length) - else: - position_ids = position_ids.view(-1, seq_length).long() if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) - if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: - is_padding_right = attention_mask[:, -1].sum().item() != batch_size - if is_padding_right: - raise ValueError( - "You are attempting to perform batched generation with padding_side='right'" - " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to " - " call `tokenizer.padding_side = 'left'` before tokenizing the input. " - ) - - if self._attn_implementation == "flash_attention_2": - # 2d mask is passed through the layers - attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None - else: - # 4d mask is passed through the layers - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) hidden_states = inputs_embeds @@ -1125,20 +1109,22 @@ def forward( layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, - attention_mask, + causal_mask, position_ids, past_key_values, output_attentions, use_cache, + cache_position, ) else: layer_outputs = decoder_layer( hidden_states, - attention_mask=attention_mask, + attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = layer_outputs[0] @@ -1167,6 +1153,87 @@ def forward( attentions=all_self_attns, ) + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + class Phi3ForCausalLM(Phi3PreTrainedModel): _tied_weights_keys = ["lm_head.weight"] @@ -1220,6 +1287,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: r""" Args: @@ -1297,14 +1365,25 @@ def forward( # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation def prepare_inputs_for_generation( - self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + use_cache=True, + **kwargs, ): past_length = 0 if past_key_values is not None: # Past key values are always initialized with a `Cache` object -> no need for if-else anymore - cache_length = past_key_values.get_seq_length() - past_length = past_key_values.seen_tokens - max_cache_length = past_key_values.get_max_length() + past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() + max_cache_length = ( + torch.tensor(past_key_values.get_max_length(), device=input_ids.device) + if past_key_values.get_max_length() is not None + else None + ) + cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where @@ -1340,12 +1419,19 @@ def prepare_inputs_for_generation( else: model_inputs = {"input_ids": input_ids} + input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] + if cache_position is None: + cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) + elif use_cache: + cache_position = cache_position[-input_length:] + model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, - "use_cache": kwargs.get("use_cache"), + "use_cache": use_cache, "attention_mask": attention_mask, + "cache_position": cache_position, } ) return model_inputs diff --git a/src/transformers/models/qwen2/modeling_qwen2.py b/src/transformers/models/qwen2/modeling_qwen2.py index aca2f30d8c94..fceee705fdc7 100644 --- a/src/transformers/models/qwen2/modeling_qwen2.py +++ b/src/transformers/models/qwen2/modeling_qwen2.py @@ -30,8 +30,10 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN -from ...cache_utils import Cache, DynamicCache -from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa +from ...cache_utils import Cache, DynamicCache, StaticCache +from ...modeling_attn_mask_utils import ( + AttentionMaskConverter, +) from ...modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, @@ -248,6 +250,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() @@ -272,7 +275,7 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -287,13 +290,9 @@ def forward( f" {attn_weights.size()}" ) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - - attn_weights = attn_weights + attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) @@ -343,6 +342,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ): bsz, q_len, _ = hidden_states.size() @@ -409,7 +409,7 @@ def forward( attention_mask = attention_mask[:, slicing_tokens:] attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -628,6 +628,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. @@ -662,17 +663,15 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) + causal_mask = attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. @@ -684,13 +683,13 @@ def forward( # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. - is_causal = True if self.is_causal and attention_mask is None and q_len > 1 else False + is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, - attn_mask=attention_mask, + attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) @@ -734,6 +733,7 @@ def forward( past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: @@ -747,6 +747,8 @@ def forward( If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. """ residual = hidden_states @@ -761,6 +763,7 @@ def forward( past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = residual + hidden_states @@ -891,6 +894,10 @@ def _init_weights(self, module): more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. """ @@ -940,6 +947,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( @@ -949,15 +957,10 @@ def forward( return_dict = return_dict if return_dict is not None else self.config.use_return_dict - # retrieve input_ids and inputs_embeds - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape - elif inputs_embeds is not None: - batch_size, seq_length, _ = inputs_embeds.shape - else: - raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -966,57 +969,29 @@ def forward( ) use_cache = False - past_key_values_length = 0 - - if use_cache: - use_legacy_cache = not isinstance(past_key_values, Cache) - if use_legacy_cache: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - past_key_values_length = past_key_values.get_usable_length(seq_length) - - if position_ids is None: - device = input_ids.device if input_ids is not None else inputs_embeds.device - position_ids = torch.arange( - past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + use_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + use_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" ) - position_ids = position_ids.unsqueeze(0).view(-1, seq_length) - else: - position_ids = position_ids.view(-1, seq_length).long() if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) - if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: - is_padding_right = attention_mask[:, -1].sum().item() != batch_size - if is_padding_right: - raise ValueError( - "You are attempting to perform batched generation with padding_side='right'" - " this may lead to unexpected behaviour for Flash Attention version of Qwen2. Make sure to " - " call `tokenizer.padding_side = 'left'` before tokenizing the input. " - ) - - if self._attn_implementation == "flash_attention_2": - # 2d mask is passed through the layers - attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None - elif self._attn_implementation == "sdpa" and not output_attentions: - # output_attentions=True can not be supported when using SDPA, and we fall back on - # the manual implementation that requires a 4D causal mask in all cases. - attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, - ) - else: - # 4d mask is passed through the layers - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) hidden_states = inputs_embeds @@ -1033,20 +1008,22 @@ def forward( layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, - attention_mask, + causal_mask, position_ids, past_key_values, output_attentions, use_cache, + cache_position, ) else: layer_outputs = decoder_layer( hidden_states, - attention_mask=attention_mask, + attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = layer_outputs[0] @@ -1076,6 +1053,87 @@ def forward( attentions=all_self_attns, ) + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + class Qwen2ForCausalLM(Qwen2PreTrainedModel): _tied_weights_keys = ["lm_head.weight"] @@ -1121,6 +1179,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: r""" Args: @@ -1165,6 +1224,7 @@ def forward( output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, + cache_position=cache_position, ) hidden_states = outputs[0] @@ -1197,15 +1257,26 @@ def forward( ) def prepare_inputs_for_generation( - self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + use_cache=True, + **kwargs, ): past_length = 0 # Omit tokens covered by past_key_values if past_key_values is not None: # Past key values are always initialized with a `Cache` object -> no need for if-else anymore - cache_length = past_key_values.get_seq_length() - past_length = past_key_values.seen_tokens - max_cache_length = past_key_values.get_max_length() + past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() + max_cache_length = ( + torch.tensor(past_key_values.get_max_length(), device=input_ids.device) + if past_key_values.get_max_length() is not None + else None + ) + cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where @@ -1241,12 +1312,19 @@ def prepare_inputs_for_generation( else: model_inputs = {"input_ids": input_ids} + input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] + if cache_position is None: + cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) + elif use_cache: + cache_position = cache_position[-input_length:] + model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, - "use_cache": kwargs.get("use_cache"), + "use_cache": use_cache, "attention_mask": attention_mask, + "cache_position": cache_position, } ) return model_inputs diff --git a/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py b/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py index c32c06b5e7c2..951b1c35815d 100644 --- a/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py +++ b/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py @@ -30,8 +30,10 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN -from ...cache_utils import Cache, DynamicCache -from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa +from ...cache_utils import Cache, DynamicCache, StaticCache +from ...modeling_attn_mask_utils import ( + AttentionMaskConverter, +) from ...modeling_outputs import ( MoeCausalLMOutputWithPast, MoeModelOutputWithPast, @@ -325,6 +327,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() @@ -349,7 +352,7 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -364,13 +367,9 @@ def forward( f" {attn_weights.size()}" ) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - - attn_weights = attn_weights + attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) @@ -421,6 +420,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ): bsz, q_len, _ = hidden_states.size() @@ -487,7 +487,7 @@ def forward( attention_mask = attention_mask[:, slicing_tokens:] attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -706,6 +706,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. @@ -740,17 +741,15 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) + causal_mask = attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. @@ -762,13 +761,13 @@ def forward( # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. - is_causal = True if self.is_causal and attention_mask is None and q_len > 1 else False + is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, - attn_mask=attention_mask, + attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) @@ -876,6 +875,7 @@ def forward( output_attentions: Optional[bool] = False, output_router_logits: Optional[bool] = False, use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: @@ -892,6 +892,8 @@ def forward( If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. """ residual = hidden_states @@ -906,6 +908,7 @@ def forward( past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = residual + hidden_states @@ -1048,6 +1051,10 @@ def _init_weights(self, module): should not be returned during inference. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. """ @@ -1098,6 +1105,7 @@ def forward( output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, MoeModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_router_logits = ( @@ -1110,15 +1118,10 @@ def forward( return_dict = return_dict if return_dict is not None else self.config.use_return_dict - # retrieve input_ids and inputs_embeds - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape - elif inputs_embeds is not None: - batch_size, seq_length, _ = inputs_embeds.shape - else: - raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -1127,57 +1130,29 @@ def forward( ) use_cache = False - past_key_values_length = 0 - - if use_cache: - use_legacy_cache = not isinstance(past_key_values, Cache) - if use_legacy_cache: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - past_key_values_length = past_key_values.get_usable_length(seq_length) - - if position_ids is None: - device = input_ids.device if input_ids is not None else inputs_embeds.device - position_ids = torch.arange( - past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + use_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + use_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" ) - position_ids = position_ids.unsqueeze(0).view(-1, seq_length) - else: - position_ids = position_ids.view(-1, seq_length).long() if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) - if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: - is_padding_right = attention_mask[:, -1].sum().item() != batch_size - if is_padding_right: - raise ValueError( - "You are attempting to perform batched generation with padding_side='right'" - " this may lead to unexpected behaviour for Flash Attention version of Qwen2MoE. Make sure to " - " call `tokenizer.padding_side = 'left'` before tokenizing the input. " - ) - - if self._attn_implementation == "flash_attention_2": - # 2d mask is passed through the layers - attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None - elif self._attn_implementation == "sdpa" and not output_attentions: - # output_attentions=True can not be supported when using SDPA, and we fall back on - # the manual implementation that requires a 4D causal mask in all cases. - attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, - ) - else: - # 4d mask is passed through the layers - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) hidden_states = inputs_embeds @@ -1195,22 +1170,24 @@ def forward( layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, - attention_mask, + causal_mask, position_ids, past_key_values, output_attentions, output_router_logits, use_cache, + cache_position, ) else: layer_outputs = decoder_layer( hidden_states, - attention_mask=attention_mask, + attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, output_router_logits=output_router_logits, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = layer_outputs[0] @@ -1248,6 +1225,87 @@ def forward( router_logits=all_router_logits, ) + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + class Qwen2MoeForCausalLM(Qwen2MoePreTrainedModel): _tied_weights_keys = ["lm_head.weight"] @@ -1297,6 +1355,7 @@ def forward( output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, MoeCausalLMOutputWithPast]: r""" Args: @@ -1345,6 +1404,7 @@ def forward( output_hidden_states=output_hidden_states, output_router_logits=output_router_logits, return_dict=return_dict, + cache_position=cache_position, ) hidden_states = outputs[0] @@ -1392,15 +1452,26 @@ def forward( ) def prepare_inputs_for_generation( - self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + use_cache=True, + **kwargs, ): past_length = 0 # Omit tokens covered by past_key_values if past_key_values is not None: # Past key values are always initialized with a `Cache` object -> no need for if-else anymore - cache_length = past_key_values.get_seq_length() - past_length = past_key_values.seen_tokens - max_cache_length = past_key_values.get_max_length() + past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() + max_cache_length = ( + torch.tensor(past_key_values.get_max_length(), device=input_ids.device) + if past_key_values.get_max_length() is not None + else None + ) + cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where @@ -1436,12 +1507,19 @@ def prepare_inputs_for_generation( else: model_inputs = {"input_ids": input_ids} + input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] + if cache_position is None: + cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) + elif use_cache: + cache_position = cache_position[-input_length:] + model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, - "use_cache": kwargs.get("use_cache"), + "use_cache": use_cache, "attention_mask": attention_mask, + "cache_position": cache_position, } ) return model_inputs diff --git a/src/transformers/models/stablelm/modeling_stablelm.py b/src/transformers/models/stablelm/modeling_stablelm.py index 920c45800e61..e3cc57642af3 100755 --- a/src/transformers/models/stablelm/modeling_stablelm.py +++ b/src/transformers/models/stablelm/modeling_stablelm.py @@ -29,8 +29,10 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN -from ...cache_utils import Cache, DynamicCache -from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa +from ...cache_utils import Cache, DynamicCache, StaticCache +from ...modeling_attn_mask_utils import ( + AttentionMaskConverter, +) from ...modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, @@ -316,6 +318,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() @@ -360,7 +363,12 @@ def forward( if past_key_value is not None: # Specific to RoPE models with partial rotation - cache_kwargs = {"sin": sin, "cos": cos, "partial_rotation_size": self.rotary_emb.dim} + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_emb.dim, + "cache_position": cache_position, + } key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # Repeat k/v heads if n_kv_heads < n_heads @@ -375,12 +383,9 @@ def forward( f" {attn_weights.size()}" ) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - attn_weights = attn_weights + attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights += causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dtype=torch.float32, dim=-1).to(query_states.dtype) @@ -414,6 +419,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. @@ -473,13 +479,22 @@ def forward( if past_key_value is not None: # Specific to RoPE models with partial rotation - cache_kwargs = {"sin": sin, "cos": cos, "partial_rotation_size": self.rotary_emb.dim} + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_emb.dim, + "cache_position": cache_position, + } key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # Repeat k/v heads if n_kv_heads < n_heads key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) + causal_mask = attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. if query_states.device.type == "cuda" and attention_mask is not None: @@ -490,13 +505,13 @@ def forward( # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. - is_causal = True if self.is_causal and attention_mask is None and q_len > 1 else False + is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, - attn_mask=attention_mask, + attn_mask=causal_mask, dropout_p=self.attention_dropout.p if self.training else 0.0, is_causal=is_causal, ) @@ -533,6 +548,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: # StableLmFlashAttention2 attention does not support output_attentions @@ -583,7 +599,12 @@ def forward( key_states = torch.cat((key_rot, key_pass), dim=-1) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos, "partial_rotation_size": self.rotary_emb.dim} + cache_kwargs = { + "sin": sin, + "cos": cos, + "partial_rotation_size": self.rotary_emb.dim, + "cache_position": cache_position, + } key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache @@ -739,6 +760,7 @@ def forward( past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: @@ -758,6 +780,8 @@ def forward( use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. """ residual = hidden_states @@ -772,6 +796,7 @@ def forward( past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) # copied from transformers.models.gpt_neox.modeling_gpt_neox.GPTNeoXLayer.forward @@ -912,6 +937,10 @@ def _init_weights(self, module): more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. """ @@ -961,6 +990,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( @@ -970,18 +1000,10 @@ def forward( return_dict = return_dict if return_dict is not None else self.config.use_return_dict - # retrieve input_ids and inputs_embeds - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape - elif inputs_embeds is not None: - batch_size, seq_length, _ = inputs_embeds.shape - else: - raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") - - seq_length_with_past = seq_length - past_key_values_length = 0 + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -990,36 +1012,29 @@ def forward( ) use_cache = False - if use_cache: - use_legacy_cache = not isinstance(past_key_values, Cache) - if use_legacy_cache: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - past_key_values_length = past_key_values.get_usable_length(seq_length) - seq_length_with_past = seq_length_with_past + past_key_values_length - - if position_ids is None: - device = input_ids.device if input_ids is not None else inputs_embeds.device - position_ids = torch.arange( - past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + use_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + use_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" ) - position_ids = position_ids.unsqueeze(0) if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) - # embed positions - if self._attn_implementation == "flash_attention_2": - # 2d mask is passed through the layers - attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None - # for output_attentions case used fallback to eager attention realization - elif self._attn_implementation == "sdpa" and not output_attentions: - attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( - attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length - ) - else: - # 4d mask is passed through the layers - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) hidden_states = inputs_embeds @@ -1036,19 +1051,22 @@ def forward( layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, - attention_mask, + causal_mask, position_ids, past_key_values, output_attentions, + use_cache, + cache_position, ) else: layer_outputs = decoder_layer( hidden_states, - attention_mask=attention_mask, + attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = layer_outputs[0] @@ -1078,6 +1096,87 @@ def forward( attentions=all_self_attns, ) + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM with PERSIMMON->STABLELM,Persimmon->StableLm class StableLmForCausalLM(StableLmPreTrainedModel): @@ -1132,6 +1231,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: r""" Args: @@ -1175,6 +1275,7 @@ def forward( output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, + cache_position=cache_position, ) hidden_states = outputs[0] @@ -1206,14 +1307,25 @@ def forward( ) def prepare_inputs_for_generation( - self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + use_cache=True, + **kwargs, ): past_length = 0 if past_key_values is not None: # Past key values are always initialized with a `Cache` object -> no need for if-else anymore - cache_length = past_key_values.get_seq_length() - past_length = past_key_values.seen_tokens - max_cache_length = past_key_values.get_max_length() + past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() + max_cache_length = ( + torch.tensor(past_key_values.get_max_length(), device=input_ids.device) + if past_key_values.get_max_length() is not None + else None + ) + cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where @@ -1249,12 +1361,19 @@ def prepare_inputs_for_generation( else: model_inputs = {"input_ids": input_ids} + input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] + if cache_position is None: + cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) + elif use_cache: + cache_position = cache_position[-input_length:] + model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, - "use_cache": kwargs.get("use_cache"), + "use_cache": use_cache, "attention_mask": attention_mask, + "cache_position": cache_position, } ) return model_inputs diff --git a/src/transformers/models/starcoder2/modeling_starcoder2.py b/src/transformers/models/starcoder2/modeling_starcoder2.py index 1f9cb3e2a318..c02a90f6a582 100644 --- a/src/transformers/models/starcoder2/modeling_starcoder2.py +++ b/src/transformers/models/starcoder2/modeling_starcoder2.py @@ -30,8 +30,10 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN -from ...cache_utils import Cache, DynamicCache -from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa +from ...cache_utils import Cache, DynamicCache, StaticCache +from ...modeling_attn_mask_utils import ( + AttentionMaskConverter, +) from ...modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, @@ -232,6 +234,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() @@ -256,7 +259,7 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -271,13 +274,9 @@ def forward( f" {attn_weights.size()}" ) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - - attn_weights = attn_weights + attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights += causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) @@ -328,6 +327,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ): bsz, q_len, _ = hidden_states.size() @@ -393,7 +393,7 @@ def forward( attention_mask = attention_mask[:, slicing_tokens:] attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads @@ -608,6 +608,7 @@ def forward( past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. @@ -642,17 +643,15 @@ def forward( query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: - cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) + causal_mask = attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. @@ -664,13 +663,13 @@ def forward( # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. # # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. - is_causal = True if self.is_causal and attention_mask is None and q_len > 1 else False + is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, - attn_mask=attention_mask, + attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) @@ -713,6 +712,7 @@ def forward( past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: @@ -726,6 +726,8 @@ def forward( If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. """ residual = hidden_states @@ -740,6 +742,7 @@ def forward( past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = residual + hidden_states @@ -871,6 +874,10 @@ def _init_weights(self, module): more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. """ @@ -920,6 +927,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( @@ -929,15 +937,10 @@ def forward( return_dict = return_dict if return_dict is not None else self.config.use_return_dict - # retrieve input_ids and inputs_embeds - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape - elif inputs_embeds is not None: - batch_size, seq_length, _ = inputs_embeds.shape - else: - raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -946,57 +949,29 @@ def forward( ) use_cache = False - past_key_values_length = 0 - - if use_cache: - use_legacy_cache = not isinstance(past_key_values, Cache) - if use_legacy_cache: - past_key_values = DynamicCache.from_legacy_cache(past_key_values) - past_key_values_length = past_key_values.get_usable_length(seq_length) - - if position_ids is None: - device = input_ids.device if input_ids is not None else inputs_embeds.device - position_ids = torch.arange( - past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + use_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + use_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" ) - position_ids = position_ids.unsqueeze(0).view(-1, seq_length) - else: - position_ids = position_ids.view(-1, seq_length).long() if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) - if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: - is_padding_right = attention_mask[:, -1].sum().item() != batch_size - if is_padding_right: - raise ValueError( - "You are attempting to perform batched generation with padding_side='right'" - " this may lead to unexpected behaviour for Flash Attention version of Starcoder2. Make sure to " - " call `tokenizer.padding_side = 'left'` before tokenizing the input. " - ) - - if self._attn_implementation == "flash_attention_2": - # 2d mask is passed through the layers - attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None - elif self._attn_implementation == "sdpa" and not output_attentions: - # output_attentions=True can not be supported when using SDPA, and we fall back on - # the manual implementation that requires a 4D causal mask in all cases. - attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, - ) - else: - # 4d mask is passed through the layers - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, - (batch_size, seq_length), - inputs_embeds, - past_key_values_length, - sliding_window=self.config.sliding_window, + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) hidden_states = inputs_embeds hidden_states = nn.functional.dropout(hidden_states, p=self.embedding_dropout, training=self.training) @@ -1014,20 +989,22 @@ def forward( layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, - attention_mask, + causal_mask, position_ids, past_key_values, output_attentions, use_cache, + cache_position, ) else: layer_outputs = decoder_layer( hidden_states, - attention_mask=attention_mask, + attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, + cache_position=cache_position, ) hidden_states = layer_outputs[0] @@ -1057,6 +1034,87 @@ def forward( attentions=all_self_attns, ) + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + # Copied from transformers.models.qwen2.modeling_qwen2.Qwen2ForCausalLM with QWEN2->STARCODER2,Qwen2->Starcoder2 class Starcoder2ForCausalLM(Starcoder2PreTrainedModel): @@ -1104,6 +1162,7 @@ def forward( output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: r""" Args: @@ -1148,6 +1207,7 @@ def forward( output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, + cache_position=cache_position, ) hidden_states = outputs[0] @@ -1180,15 +1240,26 @@ def forward( ) def prepare_inputs_for_generation( - self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + use_cache=True, + **kwargs, ): past_length = 0 # Omit tokens covered by past_key_values if past_key_values is not None: # Past key values are always initialized with a `Cache` object -> no need for if-else anymore - cache_length = past_key_values.get_seq_length() - past_length = past_key_values.seen_tokens - max_cache_length = past_key_values.get_max_length() + past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length() + max_cache_length = ( + torch.tensor(past_key_values.get_max_length(), device=input_ids.device) + if past_key_values.get_max_length() is not None + else None + ) + cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length) # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where @@ -1224,12 +1295,19 @@ def prepare_inputs_for_generation( else: model_inputs = {"input_ids": input_ids} + input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1] + if cache_position is None: + cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) + elif use_cache: + cache_position = cache_position[-input_length:] + model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, - "use_cache": kwargs.get("use_cache"), + "use_cache": use_cache, "attention_mask": attention_mask, + "cache_position": cache_position, } ) return model_inputs