v1

meteor/arch/modeling_internlm2.py

@@ -43,18 +43,18 @@ from .configuration_internlm2 import InternLM2Config
 logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = 'InternLM2Config'
-# flash_attn_func, flash_attn_varlen_func = None, None
-# pad_input, index_first_axis, unpad_input = None, None, None
-# def _import_flash_attn():
-#     global flash_attn_func, flash_attn_varlen_func
-#     global pad_input, index_first_axis, unpad_input
-#     try:
-#         from flash_attn import flash_attn_func as _flash_attn_func, flash_attn_varlen_func as _flash_attn_varlen_func
-#         from flash_attn.bert_padding import pad_input as _pad_input, index_first_axis as _index_first_axis, unpad_input as _unpad_input
-#         flash_attn_func, flash_attn_varlen_func = _flash_attn_func, _flash_attn_varlen_func
-#         pad_input, index_first_axis, unpad_input = _pad_input, _index_first_axis, _unpad_input
-#     except ImportError:
-#         raise ImportError("flash_attn is not installed.")
+flash_attn_func, flash_attn_varlen_func = None, None
+pad_input, index_first_axis, unpad_input = None, None, None
+def _import_flash_attn():
+    global flash_attn_func, flash_attn_varlen_func
+    global pad_input, index_first_axis, unpad_input
+    try:
+        from flash_attn import flash_attn_func as _flash_attn_func, flash_attn_varlen_func as _flash_attn_varlen_func
+        from flash_attn.bert_padding import pad_input as _pad_input, index_first_axis as _index_first_axis, unpad_input as _unpad_input
+        flash_attn_func, flash_attn_varlen_func = _flash_attn_func, _flash_attn_varlen_func
+        pad_input, index_first_axis, unpad_input = _pad_input, _index_first_axis, _unpad_input
+    except ImportError:
+        raise ImportError("flash_attn is not installed.")
 
 # Copied from transformers.models.llama.modeling_llama._get_unpad_data
 def _get_unpad_data(attention_mask):
@@ -492,12 +492,196 @@ class InternLM2Attention(nn.Module):
 
         return attn_output, attn_weights, past_key_value
 
+
+class InternLM2FlashAttention2(InternLM2Attention):
+    """InternLM2 flash attention module.
+
+    This module inherits from `InternLM2Attention` as the weights of the module
+    stays untouched. The only required change would be on the forward pass
+    where it needs to correctly call the public API of flash attention and deal
+    with padding tokens in case the input contains any of them.
+    """
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        im_mask: Optional[Tuple[torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor],
+               Optional[Tuple[torch.Tensor]]]:
+        # InternLM2FlashAttention2 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()
+
+        qkv_states = self.wqkv(hidden_states, im_mask)
+
+        qkv_states = rearrange(
+            qkv_states,
+            'b q (h gs d) -> b q h gs d',
+            gs=2 + self.num_key_value_groups,
+            d=self.head_dim,
+            q=q_len,
+        )
+
+        query_states = qkv_states[..., :self.num_key_value_groups, :]
+        query_states = rearrange(query_states, 'b q h gs d -> b q (h gs) d')
+        key_states = qkv_states[..., -2, :]
+        value_states = qkv_states[..., -1, :]
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        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)
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len)
+
+        attn_output = attn_output.reshape(bsz, q_len,
+                                          self.hidden_size).contiguous()
+        attn_output = self.wo(attn_output, im_mask)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        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
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        # Contains at least one padding token in the sequence
+        causal = self.is_causal and query_length != 1
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._unpad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    def _unpad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q.to(torch.int64),
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
 class InternLM2DecoderLayer(nn.Module):
 
     def __init__(self, config: InternLM2Config):
         super().__init__()
         self.hidden_size = config.hidden_size
-        self.attention = InternLM2Attention(config=config)
+        self.attention = (
+            InternLM2Attention(config=config)
+            if not getattr(config, 'attn_implementation')=="flash_attention_2" else
+            InternLM2FlashAttention2(config=config))
         self.feed_forward = InternLM2MLP(config)
         self.attention_norm = InternLM2RMSNorm(
             config.hidden_size, eps=config.rms_norm_eps)
@@ -762,7 +946,7 @@ class InternLM2Model(InternLM2PreTrainedModel):
 
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
-        # if self.config.attn_implementation: _import_flash_attn()
+        if self.config.attn_implementation: _import_flash_attn()
 
         # retrieve input_ids and inputs_embeds
         if input_ids is not None and inputs_embeds is not None: