Update modeling_llama.py

modeling_llama.py

@@ -49,6 +49,7 @@ from transformers.utils import (
     replace_return_docstrings,
 )
 from .configuration_llama import LlamaConfig
+from transformers.models.llama.modeling_llama import LlamaRMSNorm
 
 
 if is_flash_attn_2_available():
@@ -56,6 +57,7 @@ if is_flash_attn_2_available():
     from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
 
 
+
 logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = "LlamaConfig"
@@ -72,24 +74,6 @@ def _get_unpad_data(attention_mask):
         max_seqlen_in_batch,
     )
 
-
-class LlamaRMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps=1e-6):
-        """
-        LlamaRMSNorm is equivalent to T5LayerNorm
-        """
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-
-
 ALL_LAYERNORM_LAYERS.append(LlamaRMSNorm)
 
 
@@ -183,7 +167,6 @@ def rotate_half(x):
 
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
-
     Args:
         q (`torch.Tensor`): The query tensor.
         k (`torch.Tensor`): The key tensor.
@@ -505,7 +488,6 @@ class LlamaFlashAttention2(LlamaAttention):
         """
         Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
         first unpad the input, then computes the attention scores and pad the final attention scores.
-
         Args:
             query_states (`torch.Tensor`):
                 Input query states to be passed to Flash Attention API
@@ -656,7 +638,6 @@ class LlamaSdpaAttention(LlamaAttention):
         value_states = repeat_kv(value_states, self.num_key_value_groups)
 
         causal_mask = attention_mask
-        # if attention_mask is not None and cache_position is not None:
         if attention_mask is not None:
             causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
 
@@ -667,12 +648,15 @@ class LlamaSdpaAttention(LlamaAttention):
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
 
+        # In case we are not compiling, we may set `causal_mask` to None, which is required to dispatch to SDPA's Flash Attention 2 backend, rather
+        # relying on the `is_causal` argument.
         attn_output = torch.nn.functional.scaled_dot_product_attention(
             query_states,
             key_states,
             value_states,
             attn_mask=causal_mask,
             dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=causal_mask is None and q_len > 1,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
@@ -769,11 +753,9 @@ LLAMA_START_DOCSTRING = r"""
     This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
     library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
     etc.)
-
     This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
     Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
     and behavior.
-
     Parameters:
         config ([`LlamaConfig`]):
             Model configuration class with all the parameters of the model. Initializing with a config file does not
@@ -834,50 +816,38 @@ LLAMA_INPUTS_DOCSTRING = r"""
         input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
             Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
             it.
-
             Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
             [`PreTrainedTokenizer.__call__`] for details.
-
             [What are input IDs?](../glossary#input-ids)
         attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
             Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
-
             - 1 for tokens that are **not masked**,
             - 0 for tokens that are **masked**.
-
             [What are attention masks?](../glossary#attention-mask)
-
             Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
             [`PreTrainedTokenizer.__call__`] for details.
-
             If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
             `past_key_values`).
-
             If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
             and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
             information on the default strategy.
-
             - 1 indicates the head is **not masked**,
             - 0 indicates the head is **masked**.
         position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
             config.n_positions - 1]`.
-
             [What are position IDs?](../glossary#position-ids)
         past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
             Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
             blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
             returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
-
             Two formats are allowed:
             - a [`~cache_utils.Cache`] instance;
             - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
             shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
             cache format.
-
             The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
             legacy cache format will be returned.
-
             If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
             have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
             of shape `(batch_size, sequence_length)`.
@@ -910,7 +880,6 @@ LLAMA_INPUTS_DOCSTRING = r"""
 class LlamaModel(LlamaPreTrainedModel):
     """
     Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
-
     Args:
         config: LlamaConfig
     """
@@ -987,7 +956,7 @@ class LlamaModel(LlamaPreTrainedModel):
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
 
-        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_seen_tokens)
 
         # embed positions
         hidden_states = inputs_embeds
@@ -1051,16 +1020,31 @@ class LlamaModel(LlamaPreTrainedModel):
             attentions=all_self_attns,
         )
 
-    # 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
-    def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_seen_tokens: int,
+    ):
+        # 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
 
+        if self.config._attn_implementation == "sdpa":
+            # 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.
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens
+            ):
+                return None
+
         dtype, device = input_tensor.dtype, input_tensor.device
         min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
@@ -1068,7 +1052,9 @@ class LlamaModel(LlamaPreTrainedModel):
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (
-                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else cache_position[-1] + 1
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
             )
 
         causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
@@ -1160,20 +1146,14 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
                 Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                 config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                 (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
-
         Returns:
-
         Example:
-
         ```python
         >>> from transformers import AutoTokenizer, LlamaForCausalLM
-
         >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
         >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
-
         >>> prompt = "Hey, are you conscious? Can you talk to me?"
         >>> inputs = tokenizer(prompt, return_tensors="pt")
-
         >>> # Generate
         >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
         >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
@@ -1328,10 +1308,8 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
 @add_start_docstrings(
     """
     The LLaMa Model transformer with a sequence classification head on top (linear layer).
-
     [`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
     (e.g. GPT-2) do.
-
     Since it does classification on the last token, it requires to know the position of the last token. If a
     `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
     no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
@@ -1545,3 +1523,4 @@ class LlamaForQuestionAnswering(LlamaPreTrainedModel):
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
+