Class FlashSelfAttention

models/12B/modeling_telechat.py:156–215

Implement the scaled dot product attention with softmax. Arguments --------- softmax_scale: The temperature to use for the softmax attention. (default: 1/sqrt(d_keys) where d_keys is computed at runtime) attention_dropout: The

class FlashSelfAttention(torch.nn.Module):
    # Extracted from https://github.com/microsoft/Megatron-DeepSpeed/blob/main/megatron/model/transformer.py
    """Implement the scaled dot product attention with softmax.
    Arguments
    ---------
        softmax_scale: The temperature to use for the softmax attention.
                      (default: 1/sqrt(d_keys) where d_keys is computed at
                      runtime)
        attention_dropout: The dropout rate to apply to the attention
                           (default: 0.0)
    """

    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0,
                 device=None, dtype=None):
        super().__init__()
        assert flash_attn_unpadded_func is not None, ('Please install FlashAttention first, '
                                                      'e.g., with pip install flash-attn')
        assert rearrange is not None, 'Please install einops first, e.g., with pip install einops'
        self.causal = causal
        self.softmax_scale = softmax_scale
        self.dropout_p = attention_dropout

    def forward(self, q, k, v):
        """Implements the multihead softmax attention.
        Arguments
        ---------
            q, k, v: The tensor containing the query, key, and value. (B, S, H, D)
        """
        assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
        assert all((i.is_cuda for i in (q, k, v)))

        batch_size, seqlen_q = q.shape[0], q.shape[1]
        seqlen_k = k.shape[1]

        q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [q, k, v]]
        cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32,
                                    device=q.device)
        if self.training:
            # during training q,k,v always have same seqlen
            assert seqlen_k == seqlen_q

            is_causal = self.causal
            cu_seqlens_k = cu_seqlens_q
            dropout_p = self.dropout_p
        else:
            # turn off FA causal mask after first inference autoregressive iteration
            # only on first autoregressive step q,k,v have same seqlen
            is_causal = seqlen_q == seqlen_k
            cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32,
                                        device=q.device)
            dropout_p = 0

        output = flash_attn_unpadded_func(
            q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k,
            dropout_p=dropout_p,
            softmax_scale=self.softmax_scale, causal=is_causal
        )