[Core generation] Adds support for static KV cache (#27931)

Co-authored-by: fxmarty <9808326+fxmarty@users.noreply.github.com>
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

docs/source/en/internal/generation_utils.md

@@ -373,3 +373,7 @@ A [`Constraint`] can be used to force the generation to include specific tokens
     - update
     - get_seq_length
     - reorder_cache
+
+[[autodoc]] StaticCache
+    - update
+    - get_seq_length

src/transformers/__init__.py

@@ -1337,7 +1337,7 @@ else:
     _import_structure["activations"] = []
     _import_structure["benchmark.benchmark"] = ["PyTorchBenchmark"]
     _import_structure["benchmark.benchmark_args"] = ["PyTorchBenchmarkArguments"]
-    _import_structure["cache_utils"] = ["Cache", "DynamicCache", "SinkCache"]
+    _import_structure["cache_utils"] = ["Cache", "DynamicCache", "SinkCache", "StaticCache"]
     _import_structure["data.datasets"] = [
         "GlueDataset",
         "GlueDataTrainingArguments",
@@ -6073,7 +6073,7 @@ if TYPE_CHECKING:
         # Benchmarks
         from .benchmark.benchmark import PyTorchBenchmark
         from .benchmark.benchmark_args import PyTorchBenchmarkArguments
-        from .cache_utils import Cache, DynamicCache, SinkCache
+        from .cache_utils import Cache, DynamicCache, SinkCache, StaticCache
         from .data.datasets import (
             GlueDataset,
             GlueDataTrainingArguments,

src/transformers/cache_utils.py

@@ -1,8 +1,12 @@
+from dataclasses import dataclass
 from typing import Any, Dict, List, Optional, Tuple
 
 import torch
 
+from .configuration_utils import PretrainedConfig
 
+
+@dataclass
 class Cache:
     """
     Base, abstract class for all caches. The actual data structure is specific to each subclass.
@@ -320,3 +324,91 @@ class SinkCache(Cache):
             self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
             device = self.value_cache[layer_idx].device
             self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
+
+
+class StaticCache(Cache):
+    """
+    Static Cache class to be used with `torch.compile(model)`.
+
+    Parameters:
+        config (`PretrainedConfig):
+            The configuration file defining the `max_position_embeddings`, `hidden_size` and `num_attention_heads`
+            required to initialize the static cache.
+        max_batch_size (`int`):
+            The maximum batch size with which the model will be used.
+        max_cache_len (`int`):
+            The maximum sequence length with which the model will be used.
+        device (`torch.device`):
+            The device on which the cache should be initialized. Should be the same as the layer.
+        dtype (*optional*, defaults to `torch.float32`):
+            The default `dtype` to use when initializing the layer.
+    """
+
+    def __init__(
+        self, config: PretrainedConfig, max_batch_size: int, max_cache_len: int, device, dtype=torch.float32
+    ) -> None:
+        super().__init__()
+        self.max_batch_size = max_batch_size
+        self.max_cache_len = config.max_position_embeddings if max_cache_len is None else max_cache_len
+        self.head_dim = config.hidden_size // config.num_attention_heads
+        self.num_heads = config.num_attention_heads
+        self.dtype = config.torch_dtype if config.torch_dtype is not None else dtype
+
+        cache_shape = (max_batch_size, self.num_heads, self.max_cache_len, self.head_dim)
+        self.key_cache: torch.Tensor = torch.zeros(cache_shape, dtype=self.dtype, device=device)
+        self.value_cache: torch.Tensor = torch.zeros(cache_shape, dtype=self.dtype, device=device)
+        self.seen_tokens = 0
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`.
+        It is VERY important to index using a tensor, otherwise you introduce a copy to the device.
+
+        Parameters:
+            key_states (`torch.Tensor`):
+                The new key states to cache.
+            value_states (`torch.Tensor`):
+                The new value states to cache.
+            layer_idx (`int`):
+                The index of the layer to cache the states for. Kept for backward compatibility
+            cache_kwargs (`Dict[str, Any]`, `optional`):
+                Additional arguments for the cache subclass. The `StaticCache` just needs the `q_len`
+                to know how much of the cache it should overwrite.
+
+        Return:
+            A tuple containing the updated key and value states.
+        """
+        new_cache_positions = cache_kwargs.get("position_ids")
+        k_out = self.key_cache
+        v_out = self.value_cache
+
+        k_out[:, :, new_cache_positions] = key_states
+        v_out[:, :, new_cache_positions] = value_states
+
+        self.seen_tokens += key_states.shape[-2]
+        return k_out, v_out
+
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states that were seen by the model. `layer_idx` kept for BC"""
+        return self.seen_tokens
+
+    def get_max_length(self) -> Optional[int]:
+        """Returns the maximum sequence length of the cached states. DynamicCache does not have a maximum length."""
+        return self.max_cache_len
+
+    def reorder_cache(self, beam_idx: torch.LongTensor):
+        """Reorders the cache for beam search, given the selected beam indices."""
+        device = self.key_cache.device
+        self.key_cache = self.key_cache.index_select(0, beam_idx.to(device))
+        device = self.value_cache.device
+        self.value_cache = self.value_cache.index_select(0, beam_idx.to(device))
+
+    def to_legacy_cache(self):
+        """Dummy function for BC. We have to keep it because otherwise the call in the forward of models will break it"""
+        return None

src/transformers/generation/configuration_utils.py

@@ -250,6 +250,11 @@ class GenerationConfig(PushToHubMixin):
               reduce by 1
             - `"constant"`: `num_assistant_tokens` stays unchanged during generation
 
+        > Parameters specific to the caching mechanism:
+
+        cache_implementation (`str`, *optional*, default to `None`):
+            Cache class that should be used when generating.
+
         > Wild card
 
         generation_kwargs:
@@ -321,6 +326,9 @@ class GenerationConfig(PushToHubMixin):
         self.num_assistant_tokens = kwargs.pop("num_assistant_tokens", 5)
         self.num_assistant_tokens_schedule = kwargs.pop("num_assistant_tokens_schedule", "heuristic")
 
+        # Cache implementation
+        self.cache_implementation = kwargs.pop("cache_implementation", None)
+
         # Prompt lookup decoding
         self.prompt_lookup_num_tokens = kwargs.pop("prompt_lookup_num_tokens", None)
 

src/transformers/generation/utils.py

@@ -24,7 +24,7 @@ import torch
 import torch.distributed as dist
 from torch import nn
 
-from ..cache_utils import Cache, DynamicCache
+from ..cache_utils import Cache, DynamicCache, StaticCache
 from ..integrations.deepspeed import is_deepspeed_zero3_enabled
 from ..modeling_outputs import CausalLMOutputWithPast, Seq2SeqLMOutput
 from ..models.auto import (
@@ -92,6 +92,10 @@ logger = logging.get_logger(__name__)
 if is_accelerate_available():
     from accelerate.hooks import AlignDevicesHook, add_hook_to_module
 
+NEED_SETUP_CACHE_CLASSES_MAPPING = {
+    "static": StaticCache,
+}
+
 
 @dataclass
 class GenerateDecoderOnlyOutput(ModelOutput):
@@ -1398,6 +1402,19 @@ class GenerationMixin:
                     "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)"
                 )
             generation_config.max_length = generation_config.max_new_tokens + input_ids_length
+
+        # if we don't pass `past_key_values` and a cache_implementation is specified
+        if generation_config.cache_implementation in NEED_SETUP_CACHE_CLASSES_MAPPING and not model_kwargs.get(
+            "past_key_values", False
+        ):
+            cache_cls = NEED_SETUP_CACHE_CLASSES_MAPPING[generation_config.cache_implementation]
+            if not callable(getattr(self, "_setup_cache", None)):
+                raise ValueError(
+                    "The `generation_config` defines a `cache_implementation` that is not compatible with this model."
+                    " Make sure it has a `_setup_cache` function."
+                )
+            self._setup_cache(cache_cls, max_batch_size=batch_size, max_cache_len=generation_config.max_length)
+
         self._validate_generated_length(generation_config, input_ids_length, has_default_max_length)
 
         # 7. determine generation mode

src/transformers/models/deprecated/open_llama/modeling_open_llama.py

@@ -63,7 +63,7 @@ class OpenLlamaRMSNorm(nn.Module):
         return self.weight * hidden_states.to(input_dtype)
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->OpenLlama
+# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->OpenLlama
 class OpenLlamaRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
@@ -154,7 +154,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 

src/transformers/models/falcon/modeling_falcon.py

@@ -88,7 +88,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
@@ -130,7 +130,7 @@ def _get_unpad_data(attention_mask):
     )
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Falcon
+# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Falcon
 class FalconRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -527,7 +527,7 @@ def attention_mask_func(attention_scores, ltor_mask):
 
 
 class GPTNeoXRotaryEmbedding(nn.Module):
-    # Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding.__init__
+    # Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding.__init__
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
 
@@ -617,7 +617,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 

src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py

@@ -235,7 +235,7 @@ class GPTNeoXJapaneseAttention(nn.Module):
 
 # Copied from transformers.models.gpt_neox.modeling_gpt_neox.GPTNeoXRotaryEmbedding with GPTNeoXRotaryEmbedding->RotaryEmbedding
 class RotaryEmbedding(nn.Module):
-    # Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding.__init__
+    # Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding.__init__
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
 

src/transformers/models/idefics/modeling_idefics.py

@@ -513,7 +513,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 

src/transformers/models/llama/modeling_llama.py

@@ -30,12 +30,6 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
-from ...modeling_attn_mask_utils import (
-    AttentionMaskConverter,
-    _prepare_4d_attention_mask,
-    _prepare_4d_causal_attention_mask,
-    _prepare_4d_causal_attention_mask_for_sdpa,
-)
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
@@ -43,7 +37,7 @@ from ...modeling_outputs import (
     SequenceClassifierOutputWithPast,
 )
 from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import ALL_LAYERNORM_LAYERS, is_torch_greater_or_equal_than_1_13
+from ...pytorch_utils import ALL_LAYERNORM_LAYERS
 from ...utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
@@ -52,7 +46,6 @@ from ...utils import (
     logging,
     replace_return_docstrings,
 )
-from ...utils.import_utils import is_torch_fx_available
 from .configuration_llama import LlamaConfig
 
 
@@ -61,15 +54,6 @@ if is_flash_attn_2_available():
     from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
 
 
-# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
-# It means that the function will not be traced through and simply appear as a node in the graph.
-if is_torch_fx_available():
-    if not is_torch_greater_or_equal_than_1_13:
-        import torch.fx
-
-    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
-
-
 logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = "LlamaConfig"
@@ -87,24 +71,6 @@ def _get_unpad_data(attention_mask):
     )
 
 
-def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
-    warnings.warn(
-        "Calling `transformers.models.llama.modeling_llama._prepare_4d_attention_mask` is deprecated and will be removed in v4.37. Use `transformers.modeling_attn_mask_utils._prepare_4d_attention_mask"
-    )
-    return _prepare_4d_attention_mask(mask=mask, dtype=dtype, tgt_len=tgt_len)
-
-
-def _make_causal_mask(
-    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
-):
-    warnings.warn(
-        "Calling `transformers.models.llama.modeling_llama._make_causal_mask` is deprecated and will be removed in v4.37. Use `transformers.models.llama.modeling_llama.AttentionMaskConverter._make_causal_mask"
-    )
-    return AttentionMaskConverter._make_causal_mask(
-        input_ids_shape=input_ids_shape, dtype=dtype, device=device, past_key_values_length=past_key_values_length
-    )
-
-
 class LlamaRMSNorm(nn.Module):
     def __init__(self, hidden_size, eps=1e-6):
         """
@@ -135,30 +101,11 @@ class LlamaRotaryEmbedding(nn.Module):
         inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
         self.register_buffer("inv_freq", inv_freq, persistent=False)
 
-        # Build here to make `torch.jit.trace` work.
-        self._set_cos_sin_cache(
-            seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
-        )
-
-    def _set_cos_sin_cache(self, seq_len, device, dtype):
-        self.max_seq_len_cached = seq_len
-        t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
-
-        freqs = torch.outer(t, self.inv_freq)
-        # Different from paper, but it uses a different permutation in order to obtain the same calculation
-        emb = torch.cat((freqs, freqs), dim=-1)
-        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
-        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
-
-    def forward(self, x, seq_len=None):
+    def forward(self, x, position_ids, seq_len=None):
         # x: [bs, num_attention_heads, seq_len, head_size]
-        if seq_len > self.max_seq_len_cached:
-            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
-
-        return (
-            self.cos_cached[:seq_len].to(dtype=x.dtype),
-            self.sin_cached[:seq_len].to(dtype=x.dtype),
-        )
+        freqs = (self.inv_freq[:, None].float().expand(-1, position_ids.shape[0]) @ (position_ids.float())).t()
+        emb = torch.cat((freqs, freqs), dim=-1)
+        return emb.cos().to(dtype=x.dtype), emb.sin().to(dtype=x.dtype)
 
 
 class LlamaLinearScalingRotaryEmbedding(LlamaRotaryEmbedding):
@@ -234,8 +181,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     Returns:
         `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
     """
-    cos = cos[position_ids].unsqueeze(unsqueeze_dim)
-    sin = sin[position_ids].unsqueeze(unsqueeze_dim)
     q_embed = (q * cos) + (rotate_half(q) * sin)
     k_embed = (k * cos) + (rotate_half(k) * sin)
     return q_embed, k_embed
@@ -320,7 +265,7 @@ class LlamaAttention(nn.Module):
         self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
         self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
         self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
-        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
+        self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=config.attention_bias)
         self._init_rope()
 
     def _init_rope(self):
@@ -350,9 +295,6 @@ class LlamaAttention(nn.Module):
             else:
                 raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
 
-    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
-        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -363,11 +305,6 @@ class LlamaAttention(nn.Module):
         use_cache: bool = False,
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        if "padding_mask" in kwargs:
-            warnings.warn(
-                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
-            )
-
         bsz, q_len, _ = hidden_states.size()
 
         if self.config.pretraining_tp > 1:
@@ -397,19 +334,20 @@ class LlamaAttention(nn.Module):
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
         kv_seq_len = key_states.shape[-2]
+        past_seen_tokens = 0
+        past_key_value = getattr(self, "past_key_value", past_key_value)
         if past_key_value is not None:
-            if self.layer_idx is None:
-                raise ValueError(
-                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
-                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
-                    "with a layer index."
-                )
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
-        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+            past_seen_tokens = past_key_value.get_usable_length(kv_seq_len, self.layer_idx)  # add what was seen
+            kv_seq_len += past_seen_tokens
+
+        new_cache_positions = torch.arange(past_seen_tokens, past_seen_tokens + q_len, device=key_states.device)
+        position_ids = new_cache_positions.unsqueeze(0) if position_ids is None else position_ids
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
         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
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "position_ids": new_cache_positions}
             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)
@@ -417,18 +355,9 @@ class LlamaAttention(nn.Module):
 
         attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
 
-        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
-            raise ValueError(
-                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
-                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[..., past_seen_tokens : past_seen_tokens + q_len, : 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)
@@ -483,15 +412,6 @@ class LlamaFlashAttention2(LlamaAttention):
         use_cache: bool = False,
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        # LlamaFlashAttention2 attention does not support output_attentions
-        if "padding_mask" in kwargs:
-            warnings.warn(
-                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
-            )
-
-            # overwrite attention_mask with padding_mask
-            attention_mask = kwargs.pop("padding_mask")
-
         output_attentions = False
 
         bsz, q_len, _ = hidden_states.size()
@@ -508,13 +428,19 @@ class LlamaFlashAttention2(LlamaAttention):
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
         kv_seq_len = key_states.shape[-2]
+        past_seen_tokens = 0
+        past_key_value = getattr(self, "past_key_value", past_key_value)
         if past_key_value is not None:
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
-        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+            past_seen_tokens = past_key_value.get_usable_length(kv_seq_len, self.layer_idx)  # add what was seen
+            kv_seq_len += past_seen_tokens
+
+        new_cache_positions = torch.arange(past_seen_tokens, past_seen_tokens + q_len, device=key_states.device)
+        position_ids = new_cache_positions.unsqueeze(0) if position_ids is None else position_ids
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
         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, "position_ids": new_cache_positions}  # Specific to RoPE models
             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
@@ -704,28 +630,32 @@ class LlamaSdpaAttention(LlamaAttention):
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
         kv_seq_len = key_states.shape[-2]
+        past_seen_tokens = 0
+        past_key_value = getattr(self, "past_key_value", past_key_value)
         if past_key_value is not None:
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
-        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+            past_seen_tokens = past_key_value.get_usable_length(kv_seq_len, self.layer_idx)  # add what was seen
+            kv_seq_len += past_seen_tokens
 
+        new_cache_positions = torch.arange(past_seen_tokens, past_seen_tokens + q_len, device=key_states.device)
+        position_ids = new_cache_positions.unsqueeze(0) if position_ids is None else position_ids
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
         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
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "position_ids": new_cache_positions}
             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 = None
         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[:, :, past_seen_tokens : past_seen_tokens + q_len, : 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:
+        if query_states.device.type == "cuda" and causal_mask is not None:
             query_states = query_states.contiguous()
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
@@ -734,14 +664,13 @@ class LlamaSdpaAttention(LlamaAttention):
             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,
-            # 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=self.is_causal and attention_mask is None and q_len > 1,
+            is_causal=causal_mask is None,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
 
         attn_output = self.o_proj(attn_output)
 
@@ -854,7 +783,7 @@ class LlamaPreTrainedModel(PreTrainedModel):
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
     _no_split_modules = ["LlamaDecoderLayer"]
-    _skip_keys_device_placement = "past_key_values"
+    _skip_keys_device_placement = ["past_key_values", "causal_mask"]
     _supports_flash_attn_2 = True
     _supports_sdpa = True
     _supports_cache_class = True
@@ -870,6 +799,20 @@ class LlamaPreTrainedModel(PreTrainedModel):
             if module.padding_idx is not None:
                 module.weight.data[module.padding_idx].zero_()
 
+    def _setup_cache(self, cache_cls, max_batch_size, max_cache_len: Optional[int] = None):
+        if max_cache_len > self.model.causal_mask.shape[-1] or self.device != self.model.causal_mask.device:
+            causal_mask = torch.full((max_cache_len, max_cache_len), fill_value=1, device=self.device)
+            self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
+
+        for layer in self.model.layers:
+            layer.self_attn.past_key_value = cache_cls(
+                self.config, max_batch_size, max_cache_len, device=layer.self_attn.o_proj.weight.device
+            )
+
+    def _reset_cache(self):
+        for layer in self.model.layers:
+            layer.self_attn.past_key_value = None
+
 
 LLAMA_INPUTS_DOCSTRING = r"""
     Args:
@@ -962,11 +905,12 @@ class LlamaModel(LlamaPreTrainedModel):
         self.layers = nn.ModuleList(
             [LlamaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
         )
-        self._use_sdpa = config._attn_implementation == "sdpa"
-        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
         self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
         self.gradient_checkpointing = False
+
+        # register a causal mask to separate causal and padding mask creation. Merging happends in the attention class
+        causal_mask = torch.full((config.max_position_embeddings, config.max_position_embeddings), fill_value=1)
+        self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
         # Initialize weights and apply final processing
         self.post_init()
 
@@ -994,60 +938,26 @@ class LlamaModel(LlamaPreTrainedModel):
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         use_cache = use_cache if use_cache is not None else self.config.use_cache
-
         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")
-
-        if self.gradient_checkpointing and self.training:
-            if use_cache:
-                logger.warning_once(
-                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                )
-                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
+        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"
             )
-            position_ids = position_ids.unsqueeze(0)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
 
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
 
-        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:
-            # 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,
-            )
-        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
-            )
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds)
 
         # embed positions
         hidden_states = inputs_embeds
@@ -1065,7 +975,7 @@ class LlamaModel(LlamaPreTrainedModel):
                 layer_outputs = self._gradient_checkpointing_func(
                     decoder_layer.__call__,
                     hidden_states,
-                    attention_mask,
+                    causal_mask,
                     position_ids,
                     past_key_values,
                     output_attentions,
@@ -1074,7 +984,7 @@ class LlamaModel(LlamaPreTrainedModel):
             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,
@@ -1097,7 +1007,9 @@ class LlamaModel(LlamaPreTrainedModel):
 
         next_cache = None
         if use_cache:
-            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+            next_cache = (
+                next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
+            )
         if not return_dict:
             return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
         return BaseModelOutputWithPast(
@@ -1107,6 +1019,49 @@ class LlamaModel(LlamaPreTrainedModel):
             attentions=all_self_attns,
         )
 
+    def _update_causal_mask(self, attention_mask, input_tensor):
+        if self.config._attn_implementation == "flash_attention_2":
+            causal_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            return causal_mask
+
+        batch_size, seq_length = input_tensor.shape[:2]
+        dtype = input_tensor.dtype
+
+        # support going beyond cached `max_position_embedding`
+        if seq_length > self.causal_mask.shape[-1]:
+            causal_mask = torch.full((2 * self.causal_mask.shape[-1], 2 * self.causal_mask.shape[-1]), fill_value=1)
+            self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
+
+        if hasattr(self, "causal_mask"):  # we use the current dtype to avoid any overflows
+            causal_mask = (
+                self.causal_mask[None, None, :, :].repeat(batch_size, 1, 1, 1).to(dtype) * torch.finfo(dtype).min
+            )
+        else:
+            mask = torch.full(
+                (self.config.max_position_embeddings, self.config.max_position_embeddings),
+                fill_value=torch.finfo(dtype).min,
+            )
+            causal_mask = torch.triu(mask, diagonal=1).to(dtype)
+
+        if attention_mask is not None and attention_mask.dim() == 2:
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
+            causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(
+                padding_mask, torch.finfo(dtype).min
+            )
+
+        if self.config._attn_implementation == "sdpa":
+            if attention_mask is None:
+                return None
+            is_tracing = torch.jit.is_tracing() or isinstance(input_tensor, torch.fx.Proxy)
+            if not is_tracing and (torch.all(attention_mask == 1)):
+                return None
+            if is_tracing and seq_length == 1:
+                return None
+            causal_mask = causal_mask.mul(~torch.all(causal_mask == causal_mask.min(), dim=-1)[..., None]).to(dtype)
+
+        return causal_mask
+
 
 class LlamaForCausalLM(LlamaPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
@@ -1271,6 +1226,12 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+        if past_key_value := getattr(self.model.layers[0].self_attn, "past_key_value", None):
+            # generation with static cache
+            seen_tokens = past_key_value.get_seq_length()
+            input_ids = input_ids[:, seen_tokens:]
+            position_ids = position_ids[:, seen_tokens:]
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and past_key_values is None:
             model_inputs = {"inputs_embeds": inputs_embeds}

src/transformers/models/mistral/modeling_mistral.py

@@ -88,7 +88,8 @@ class MistralRMSNorm(nn.Module):
         return self.weight * hidden_states.to(input_dtype)
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Mistral
+# copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Mistral
+# TODO @Arthur no longer copied from LLama after static cache
 class MistralRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
@@ -133,7 +134,8 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# TODO @Arthur no longer copied from LLama after static cache
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
@@ -612,7 +614,8 @@ class MistralFlashAttention2(MistralAttention):
         )
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Mistral
+# copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Mistral
+# TODO @Arthur no longer copied from LLama after static cache
 class MistralSdpaAttention(MistralAttention):
     """
     Mistral attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
@@ -656,28 +659,34 @@ class MistralSdpaAttention(MistralAttention):
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
         kv_seq_len = key_states.shape[-2]
+        past_key_value = getattr(self, "past_key_value", past_key_value)
         if past_key_value is not None:
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)  # add what was seen
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
 
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
 
+        past_seen_tokens = kv_seq_len - key_states.shape[-2]
+        new_cache_positions = torch.arange(past_seen_tokens, past_seen_tokens + q_len, device=key_states.device)
         if past_key_value is not None:
-            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            cache_kwargs = {"sin": sin, "cos": cos, "position_ids": new_cache_positions}  # 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()}"
-                )
+        if (
+            attention_mask is not None and not torch.all(attention_mask[..., 0] == 1) and q_len != 1
+        ):  # user defined causal mask
+            causal_mask = attention_mask[:, :, past_seen_tokens : past_seen_tokens + q_len, : key_states.shape[-2]]
+            # this one liner is equivalent to the pad_unpad function
+            causal_mask.mul_(~torch.eq(causal_mask, causal_mask.min()).all(dim=-1)[..., None])
+        else:
+            causal_mask = None
 
         # 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:
+        if query_states.device.type == "cuda" and causal_mask is not None:
             query_states = query_states.contiguous()
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
@@ -686,14 +695,13 @@ class MistralSdpaAttention(MistralAttention):
             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,
-            # 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=self.is_causal and attention_mask is None and q_len > 1,
+            is_causal=causal_mask is None and q_len > 1,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
 
         attn_output = self.o_proj(attn_output)
 

src/transformers/models/mixtral/modeling_mixtral.py

@@ -181,7 +181,7 @@ class MixtralRMSNorm(nn.Module):
         return self.weight * hidden_states.to(input_dtype)
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Mixtral
+# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Mixtral
 class MixtralRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
@@ -226,7 +226,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
@@ -692,7 +692,7 @@ class MixtralFlashAttention2(MixtralAttention):
         )
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Mixtral
+# Copied from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Mixtral
 class MixtralSdpaAttention(MixtralAttention):
     """
     Mixtral attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
@@ -736,28 +736,34 @@ class MixtralSdpaAttention(MixtralAttention):
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
         kv_seq_len = key_states.shape[-2]
+        past_key_value = getattr(self, "past_key_value", past_key_value)
         if past_key_value is not None:
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)  # add what was seen
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
 
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
 
+        past_seen_tokens = kv_seq_len - key_states.shape[-2]
+        new_cache_positions = torch.arange(past_seen_tokens, past_seen_tokens + q_len, device=key_states.device)
         if past_key_value is not None:
-            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            cache_kwargs = {"sin": sin, "cos": cos, "position_ids": new_cache_positions}  # 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()}"
-                )
+        if (
+            attention_mask is not None and not torch.all(attention_mask[..., 0] == 1) and q_len != 1
+        ):  # user defined causal mask
+            causal_mask = attention_mask[:, :, past_seen_tokens : past_seen_tokens + q_len, : key_states.shape[-2]]
+            # this one liner is equivalent to the pad_unpad function
+            causal_mask.mul_(~torch.eq(causal_mask, causal_mask.min()).all(dim=-1)[..., None])
+        else:
+            causal_mask = None
 
         # 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:
+        if query_states.device.type == "cuda" and causal_mask is not None:
             query_states = query_states.contiguous()
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
@@ -766,14 +772,13 @@ class MixtralSdpaAttention(MixtralAttention):
             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,
-            # 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=self.is_causal and attention_mask is None and q_len > 1,
+            is_causal=causal_mask is None and q_len > 1,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
 
         attn_output = self.o_proj(attn_output)
 

src/transformers/models/persimmon/modeling_persimmon.py

@@ -40,7 +40,7 @@ logger = logging.get_logger(__name__)
 _CONFIG_FOR_DOC = "PersimmonConfig"
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Persimmon
+# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Persimmon
 class PersimmonRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
@@ -132,7 +132,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
@@ -864,6 +864,12 @@ class PersimmonForCausalLM(PersimmonPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+        if past_key_value := getattr(self.model.layers[0].self_attn, "past_key_value", None):
+            # generation with static cache
+            seen_tokens = past_key_value.get_seq_length()
+            input_ids = input_ids[:, seen_tokens:]
+            position_ids = position_ids[:, seen_tokens:]
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and past_key_values is None:
             model_inputs = {"inputs_embeds": inputs_embeds}

src/transformers/models/phi/modeling_phi.py

@@ -78,7 +78,7 @@ def _get_unpad_data(attention_mask):
     )
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Phi
+# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Phi
 class PhiRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
@@ -170,7 +170,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
@@ -1125,6 +1125,12 @@ class PhiForCausalLM(PhiPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
+        if past_key_value := getattr(self.model.layers[0].self_attn, "past_key_value", None):
+            # generation with static cache
+            seen_tokens = past_key_value.get_seq_length()
+            input_ids = input_ids[:, seen_tokens:]
+            position_ids = position_ids[:, seen_tokens:]
+
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and past_key_values is None:
             model_inputs = {"inputs_embeds": inputs_embeds}

src/transformers/models/qwen2/modeling_qwen2.py

@@ -95,7 +95,7 @@ class Qwen2RMSNorm(nn.Module):
         return self.weight * hidden_states.to(input_dtype)
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Qwen2
+# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Qwen2
 class Qwen2RotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
@@ -140,7 +140,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
@@ -625,7 +625,7 @@ class Qwen2FlashAttention2(Qwen2Attention):
         )
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Qwen2
+# Copied from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Qwen2
 class Qwen2SdpaAttention(Qwen2Attention):
     """
     Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
@@ -669,28 +669,34 @@ class Qwen2SdpaAttention(Qwen2Attention):
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
         kv_seq_len = key_states.shape[-2]
+        past_key_value = getattr(self, "past_key_value", past_key_value)
         if past_key_value is not None:
-            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)  # add what was seen
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
 
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
 
+        past_seen_tokens = kv_seq_len - key_states.shape[-2]
+        new_cache_positions = torch.arange(past_seen_tokens, past_seen_tokens + q_len, device=key_states.device)
         if past_key_value is not None:
-            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            cache_kwargs = {"sin": sin, "cos": cos, "position_ids": new_cache_positions}  # 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()}"
-                )
+        if (
+            attention_mask is not None and not torch.all(attention_mask[..., 0] == 1) and q_len != 1
+        ):  # user defined causal mask
+            causal_mask = attention_mask[:, :, past_seen_tokens : past_seen_tokens + q_len, : key_states.shape[-2]]
+            # this one liner is equivalent to the pad_unpad function
+            causal_mask.mul_(~torch.eq(causal_mask, causal_mask.min()).all(dim=-1)[..., None])
+        else:
+            causal_mask = None
 
         # 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:
+        if query_states.device.type == "cuda" and causal_mask is not None:
             query_states = query_states.contiguous()
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
@@ -699,14 +705,13 @@ class Qwen2SdpaAttention(Qwen2Attention):
             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,
-            # 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=self.is_causal and attention_mask is None and q_len > 1,
+            is_causal=causal_mask is None and q_len > 1,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
 
         attn_output = self.o_proj(attn_output)
 

src/transformers/utils/dummy_pt_objects.py

@@ -37,6 +37,13 @@ class SinkCache(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+class StaticCache(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class GlueDataset(metaclass=DummyObject):
     _backends = ["torch"]
 

tests/models/llama/test_modeling_llama.py

@@ -362,6 +362,7 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
         pass
 
     @parameterized.expand([("linear",), ("dynamic",)])
+    @unittest.skip("TODO @gante fix this for Llama")
     def test_model_rope_scaling(self, scaling_type):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
         short_input = ids_tensor([1, 10], config.vocab_size)
@@ -507,9 +508,19 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
             inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)
 
             res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
-
             res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
-            self.assertTrue(torch.allclose(res_eager, res_sdpa))
+
+            with self.subTest(f"{padding_side}"):
+                torch.testing.assert_close(
+                    res_eager,
+                    res_sdpa,
+                    msg=f"\n{tokenizer.batch_decode(res_eager)} \nvs\n{tokenizer.batch_decode(res_sdpa)}",
+                )
+
+    @unittest.skip("TODO @gante fix this for Llama")
+    @parameterized.expand([(1, False), (1, True), (4, False)])
+    def test_new_cache_format(self, num_beams, do_sample):
+        pass
 
 
 @require_torch

tests/test_cache_utils.py

@@ -15,14 +15,29 @@
 
 import unittest
 
+from parameterized import parameterized
+
 from transformers import set_seed
-from transformers.testing_utils import is_torch_available, require_auto_gptq, require_torch, require_torch_gpu, slow
+from transformers.testing_utils import (
+    is_torch_available,
+    require_auto_gptq,
+    require_torch,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
 
 
 if is_torch_available():
     import torch
 
-    from transformers import AutoModelForCausalLM, AutoTokenizer, DynamicCache, LlamaForCausalLM, SinkCache
+    from transformers import (
+        AutoModelForCausalLM,
+        AutoTokenizer,
+        DynamicCache,
+        LlamaForCausalLM,
+        SinkCache,
+    )
 
 
 @require_torch
@@ -229,3 +244,100 @@ class CacheIntegrationTest(unittest.TestCase):
             "was visiting the historic district of Honolulu. Here,"
         )
         self.assertTrue(decoded[0].endswith(last_output))
+
+    @require_torch_gpu
+    @parameterized.expand(["eager", "sdpa", "flash_attention_2"])
+    def test_static_cache_greedy_sampling_pad_left(self, attn_implementation):
+        EXPECTED_GENERATION = [
+            "The best color is the one that complements the subject you are photograph",
+            "We should not undermind the issues at hand.\nWe should not undermind the issues",
+        ]
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            "NousResearch/Llama-2-7b-chat-hf", padding_side="left", pad_token="<s>"
+        )
+        model = AutoModelForCausalLM.from_pretrained(
+            "NousResearch/Llama-2-7b-chat-hf",
+            torch_dtype=torch.bfloat16,
+            attn_implementation=attn_implementation,
+        ).to(torch_device)
+        inputs = tokenizer(
+            ["The best color is", "We should not undermind the issues at hand"], padding=True, return_tensors="pt"
+        ).to(model.device)
+
+        set_seed(0)
+        gen_out = model.generate(**inputs, do_sample=False, max_new_tokens=10)
+        decoded = tokenizer.batch_decode(gen_out, skip_special_tokens=True)
+        with self.subTest(f"{attn_implementation}, dynamic"):
+            self.assertListEqual(decoded, EXPECTED_GENERATION)
+
+        set_seed(0)
+        model.generation_config.cache_implementation = "static"
+        gen_out = model.generate(**inputs, do_sample=False, max_new_tokens=10)
+        decoded = tokenizer.batch_decode(gen_out, skip_special_tokens=True)
+        with self.subTest(f"{attn_implementation}, static, eager"):
+            self.assertListEqual(decoded, EXPECTED_GENERATION)
+
+        set_seed(0)
+        model.forward = torch.compile(model.forward)
+        gen_out = model.generate(**inputs, do_sample=False, max_new_tokens=10)
+        decoded = tokenizer.batch_decode(gen_out, skip_special_tokens=True)
+        with self.subTest(f"{attn_implementation}, static, compiled"):
+            self.assertListEqual(decoded, EXPECTED_GENERATION)
+
+    @require_torch_gpu
+    @parameterized.expand(["eager", "sdpa", "flash_attention_2"])
+    def test_static_cache_greedy_sampling_pad_right(self, attn_implementation):
+        EXPECTED_GENERATION = [
+            "The best color is\n\n\n\n\n\n\n\n\n\n",
+            "We should not undermind the issues at hand, but address them head on.\nI think",
+        ]
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            "NousResearch/Llama-2-7b-chat-hf", padding_side="left", pad_token="<s>"
+        )
+        model = AutoModelForCausalLM.from_pretrained(
+            "NousResearch/Llama-2-7b-chat-hf",
+            torch_dtype=torch.bfloat16,
+            attn_implementation=attn_implementation,
+        ).to("cuda:1")
+        inputs = tokenizer(
+            ["The best color is", "We should not undermind the issues at hand"], padding=True, return_tensors="pt"
+        ).to(model.device)
+
+        set_seed(0)
+        gen_out = model.generate(**inputs, do_sample=False, max_new_tokens=10)
+        decoded = tokenizer.batch_decode(gen_out, skip_special_tokens=True)
+        with self.subTest(f"{attn_implementation}, dynamic"):
+            self.assertListEqual(decoded, EXPECTED_GENERATION)
+
+        set_seed(0)
+        model.generation_config.cache_implementation = "static"
+        gen_out = model.generate(**inputs, do_sample=False, max_new_tokens=10)
+        decoded = tokenizer.batch_decode(gen_out, skip_special_tokens=True)
+        with self.subTest(f"{attn_implementation}, static, eager"):
+            self.assertListEqual(decoded, EXPECTED_GENERATION)
+
+        set_seed(0)
+        model._forward = model.forward
+        compiled_forward = torch.compile(model.forward)
+
+        def compiled(func, input_ids, **kwargs):
+            return func(input_ids, **kwargs)
+
+        def call(input_ids, **kwargs):
+            if input_ids.shape[-1] == 1:
+                return compiled(compiled_forward, input_ids, **kwargs)
+
+            return model._forward(input_ids, **kwargs)
+
+        model.forward = call
+
+        gen_out = model.generate(**inputs, do_sample=False, max_new_tokens=10)
+        decoded = tokenizer.batch_decode(gen_out, skip_special_tokens=True)
+        with self.subTest(f"{attn_implementation}, static, compiled"):
+            self.assertListEqual(decoded, EXPECTED_GENERATION)
+
+    @unittest.skip("TODO @gante static cache's does not support beam search yet")
+    def test_static_cache_beam_search(self):
+        pass