flash_neox_modeling.py

import torch
import torch.distributed

from torch import nn
from transformers.activations import ACT2FN
from transformers.modeling_utils import PreTrainedModel
from transformers.models.gpt_neox import GPTNeoXConfig

# Flash attention imports
import rotary_emb
import flash_attn_cuda
import dropout_layer_norm

from flash_attn.layers.rotary import RotaryEmbedding


class FastLinear(nn.Linear):
    def __init__(
        self,
        in_features: int,
        out_features: int,
        bias: bool = True,
        device=None,
        dtype=None,
    ) -> None:
        super(FastLinear, self).__init__(in_features, out_features, bias, device, dtype)
        self.swap_dims = True

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        if self.swap_dims:
            self.weight = nn.Parameter(self.weight.T)
            self.swap_dims = False

        if self.bias is not None:
            return torch.addmm(self.bias, input, self.weight)
        return torch.matmul(input, self.weight)


class TensorParallelColumnLinear(FastLinear):
    def __init__(
        self,
        in_features,
        out_features,
        process_group: torch.distributed.ProcessGroup,
        bias=True,
        device=None,
        dtype=None,
    ):
        self.process_group = process_group
        self.tp_world_size = process_group.size()
        assert out_features % self.tp_world_size == 0
        out_features = out_features // self.tp_world_size

        super().__init__(
            in_features=in_features,
            out_features=out_features,
            bias=bias,
            device=device,
            dtype=dtype,
        )

    def forward(self, input):
        return super(TensorParallelColumnLinear, self).forward(input)


class TensorParallelRowLinear(FastLinear):
    def __init__(
        self,
        in_features,
        out_features,
        process_group: torch.distributed.ProcessGroup,
        bias=True,
        device=None,
        dtype=None,
    ):
        self.process_group = process_group
        self.tp_world_size = process_group.size()
        assert in_features % self.tp_world_size == 0
        in_features = in_features // self.tp_world_size

        super().__init__(
            in_features=in_features,
            out_features=out_features,
            bias=bias,
            device=device,
            dtype=dtype,
        )

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        out = super(TensorParallelRowLinear, self).forward(input)
        torch.distributed.all_reduce(out, group=self.process_group)

        return out


class TensorParallelEmbedding(nn.Embedding):
    def __init__(
        self,
        num_embeddings,
        embedding_dim,
        process_group: torch.distributed.ProcessGroup,
        padding_idx=None,
        max_norm=None,
        norm_type=2.0,
        scale_grad_by_freq=False,
        sparse=False,
        _weight=None,
        device=None,
        dtype=None,
    ):
        self.process_group = process_group
        self.tp_rank = process_group.rank()
        self.tp_world_size = process_group.size()

        self.original_num_embeddings = num_embeddings

        assert num_embeddings % self.tp_world_size == 0
        block_size = num_embeddings // self.tp_world_size
        # inputs in `[min_id, max_id[` are handled by `self` to get embeddings
        self.min_id = self.tp_rank * block_size
        self.max_id = (self.tp_rank + 1) * block_size

        super().__init__(
            block_size,
            embedding_dim,
            padding_idx=padding_idx,
            max_norm=max_norm,
            norm_type=norm_type,
            scale_grad_by_freq=scale_grad_by_freq,
            sparse=sparse,
            _weight=_weight,
            device=device,
            dtype=dtype,
        )

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        # `0` if input is in the correct interval, else `1`
        input_mask = torch.logical_or(self.min_id > input, input >= self.max_id)
        # translate for [0, self.max_id - self.min_id[
        input = input - self.min_id
        # default all out of bounds values to `0`
        input[input_mask] = 0
        out = super().forward(input)
        out[input_mask] = 0.0
        torch.distributed.all_reduce(out, group=self.process_group)
        return out


class PositionRotaryEmbedding(RotaryEmbedding):
    def _update_cos_sin_cache(self, dtype, device, seqlen):
        # Reset the tables if the sequence length has changed,
        # or if we're on a new device (possibly due to tracing for instance)
        if (
            seqlen > self._seq_len_cached
            or self._cos_cached.device != device
            or self._cos_cached.dtype != dtype
        ):
            self._seq_len_cached = seqlen
            t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
            # Don't do einsum, it converts fp32 to fp16
            # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
            freqs = torch.outer(t, self.inv_freq.to(device=t.device))
            self._cos_cached = torch.cos(freqs).to(dtype)
            self._sin_cached = torch.sin(freqs).to(dtype)

    def get_cos_sin(self, position_ids: torch.Tensor, max_s: int, dtype: torch.dtype):
        """
        Return cos and sin for the asked position ids
        """

        self._update_cos_sin_cache(dtype, position_ids.device, max_s)

        cos = torch.index_select(self._cos_cached, 0, position_ids)
        sin = torch.index_select(self._sin_cached, 0, position_ids)
        return cos.unsqueeze(1), sin.unsqueeze(1)

    def forward(self, qkv: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor):
        rotary_dim = cos.shape[-1]
        q1 = qkv[:, 0, :, :rotary_dim]
        q2 = qkv[:, 0, :, rotary_dim : 2 * rotary_dim]
        k1 = qkv[:, 1, :, :rotary_dim]
        k2 = qkv[:, 1, :, rotary_dim : 2 * rotary_dim]

        rotary_emb.apply_rotary(q1, q2, cos, sin, q1, q2, False)
        rotary_emb.apply_rotary(k1, k2, cos, sin, k1, k2, False)
        return qkv


class FlashNeoxAttention(torch.nn.Module):
    def __init__(
        self, num_heads, hidden_size, rotary_pct, rotary_emb_base, process_group=None
    ):
        super().__init__()
        self.num_heads = num_heads
        self.hidden_size = hidden_size
        self.head_size = hidden_size // num_heads

        rotary_ndims = int(self.head_size * rotary_pct)
        self.rotary_emb = PositionRotaryEmbedding(rotary_ndims, base=rotary_emb_base)
        self.softmax_scale = self.head_size ** (-0.5)

        if process_group is None:
            self.query_key_value = FastLinear(hidden_size, 3 * hidden_size)
            self.dense = FastLinear(hidden_size, hidden_size)
        else:
            self.num_heads = self.num_heads // process_group.size()
            self.query_key_value = TensorParallelColumnLinear(
                hidden_size,
                3 * hidden_size,
                process_group=process_group,
            )
            self.dense = TensorParallelRowLinear(
                hidden_size,
                hidden_size,
                process_group=process_group,
            )
        self.swap_dims = True

    # TODO: remove and swap dims when loading weights
    def _swap_dims(self):
        """Swap dims for the first inference to avoid an additional permute"""
        self.query_key_value.weight = torch.nn.Parameter(
            self.query_key_value.weight.view(
                self.num_heads, 3, self.head_size, self.hidden_size
            )
            .permute(1, 0, 2, 3)
            .reshape(-1, self.hidden_size)
        )
        self.query_key_value.bias = torch.nn.Parameter(
            self.query_key_value.bias.view(self.num_heads, 3, self.head_size)
            .permute(1, 0, 2)
            .reshape(-1)
        )
        self.swap_dims = False

    def forward(
        self,
        hidden_states,
        cos,
        sin,
        cu_seqlens,
        max_s,
        layer_past,
        layer_past_present_indices,
        cu_seqlens_q,
    ):
        if self.swap_dims:
            self._swap_dims()

        qkv = self.query_key_value(hidden_states)
        qkv = qkv.view(-1, 3, self.num_heads, self.head_size)
        qkv_rot = self.rotary_emb(qkv, cos, sin)

        # Prefill
        if layer_past_present_indices is None:
            # Copy to layer past
            layer_past[...] = qkv_rot[:, 1:]

            # output
            attn_output = torch.empty_like(qkv[:, 0])
            # flash attention
            flash_attn_cuda.fwd(
                qkv[:, 0],
                qkv[:, 1],
                qkv[:, 2],
                attn_output,
                cu_seqlens,
                cu_seqlens,
                max_s,
                max_s,
                0.0,
                self.softmax_scale,
                False,
                True,
                False,
                0,
                None,
            )
        # Decode
        else:
            query = qkv_rot[:, 0]
            # Add present to the layer_past tensor at the correct indices
            layer_past[layer_past_present_indices] = qkv_rot[:, 1:]

            # output
            attn_output = torch.empty_like(query)
            # flash attention
            flash_attn_cuda.fwd(
                query,
                layer_past[:, 0],
                layer_past[:, 1],
                attn_output,
                cu_seqlens_q,
                cu_seqlens,
                1,
                max_s,
                0.0,
                self.softmax_scale,
                False,
                False,
                False,
                0,
                None,
            )

        return self.dense(attn_output.view(-1, self.num_heads * self.head_size))


class FlashMLP(nn.Module):
    def __init__(self, act, hidden_size, intermediate_size, process_group=None):
        super().__init__()
        self.act = (
            ACT2FN[act]
            if "gelu" not in act
            else lambda x: torch.nn.functional.gelu(x, approximate="tanh")
        )

        if process_group is None:
            self.dense_h_to_4h = FastLinear(hidden_size, intermediate_size)
            self.dense_4h_to_h = FastLinear(intermediate_size, hidden_size)
        else:
            self.dense_h_to_4h = TensorParallelColumnLinear(
                hidden_size,
                intermediate_size,
                process_group=process_group,
            )
            self.dense_4h_to_h = TensorParallelRowLinear(
                intermediate_size,
                hidden_size,
                process_group=process_group,
            )
        self.heuristic = "auto"
        self.process_group = process_group

    def forward(self, hidden_states):
        hidden_states = self.dense_h_to_4h(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.dense_4h_to_h(hidden_states)
        return hidden_states


class FlashNeoXLayer(nn.Module):
    def __init__(
        self,
        num_heads,
        act,
        hidden_size,
        intermediate_size,
        rotary_pct,
        rotary_emb_base,
        layer_norm_eps,
        use_parallel_residual,
        process_group=None,
    ):
        super().__init__()
        self.use_parallel_residual = use_parallel_residual
        self.input_layernorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
        self.post_attention_layernorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
        self.attention = FlashNeoxAttention(
            num_heads, hidden_size, rotary_pct, rotary_emb_base, process_group
        )
        self.mlp = FlashMLP(act, hidden_size, intermediate_size, process_group)

    def forward(
        self,
        hidden_states,
        residual,
        cos,
        sin,
        cu_seqlens,
        max_s,
        layer_past,
        layer_past_present_indices,
        cu_seqlens_q,
    ):
        if self.use_parallel_residual:
            # faster input layer norm
            ln1_hidden_states, *rest = dropout_layer_norm.dropout_add_ln_fwd(
                hidden_states,
                None,
                self.input_layernorm.weight,
                self.input_layernorm.bias,
                None,
                None,
                None,
                None,
                0.0,
                self.input_layernorm.eps,
                1.0,
                0,
                None,
                False,
                False,
            )

            attn_output = self.attention(
                ln1_hidden_states,
                cos,
                sin,
                cu_seqlens,
                max_s,
                layer_past,
                layer_past_present_indices,
                cu_seqlens_q,
            )

            # faster post attention layer norm
            ln2_hidden_states, *rest = dropout_layer_norm.dropout_add_ln_fwd(
                hidden_states,
                None,
                self.post_attention_layernorm.weight,
                self.post_attention_layernorm.bias,
                None,
                None,
                None,
                None,
                0.0,
                self.post_attention_layernorm.eps,
                1.0,
                0,
                None,
                False,
                False,
            )

            mlp_output = self.mlp(ln2_hidden_states)
            return mlp_output + attn_output + hidden_states, None
        else:
            # faster input layer norm
            hidden_states, residual, *rest = dropout_layer_norm.dropout_add_ln_fwd(
                hidden_states,
                residual,
                self.input_layernorm.weight,
                self.input_layernorm.bias,
                None,
                None,
                None,
                None,
                0.0,
                self.input_layernorm.eps,
                1.0,
                0,
                None,
                False,
                False,
            )

            hidden_states = self.attention(
                hidden_states,
                cos,
                sin,
                cu_seqlens,
                max_s,
                layer_past,
                layer_past_present_indices,
                cu_seqlens_q,
            )

            # faster post attention layer norm
            hidden_states, residual, *rest = dropout_layer_norm.dropout_add_ln_fwd(
                hidden_states,
                residual,
                self.post_attention_layernorm.weight,
                self.post_attention_layernorm.bias,
                None,
                None,
                None,
                None,
                0.0,
                self.post_attention_layernorm.eps,
                1.0,
                0,
                None,
                False,
                False,
            )

            mlp_output = self.mlp(hidden_states)

            return mlp_output, residual


class FlashGPTNeoXPreTrainedModel(PreTrainedModel):
    config_class = GPTNeoXConfig
    base_model_prefix = "gpt_neox"
    supports_gradient_checkpointing = False
    _no_split_modules = None


class FlashGPTNeoXModel(FlashGPTNeoXPreTrainedModel):
    def __init__(self, config, process_group=None):
        super().__init__(config)
        self.config = config

        self.tp_embeddings = False
        if process_group is not None:
            self.tp_rank = process_group.rank()
            self.tp_world_size = process_group.size()
            if config.vocab_size % self.tp_world_size == 0:
                self.tp_embeddings = True

        if self.tp_embeddings:
            self.embed_in = TensorParallelEmbedding(
                config.vocab_size, config.hidden_size, process_group=process_group
            )
        else:
            self.embed_in = nn.Embedding(config.vocab_size, config.hidden_size)

        self.layers = nn.ModuleList(
            [
                FlashNeoXLayer(
                    config.num_attention_heads,
                    config.hidden_act,
                    config.hidden_size,
                    config.intermediate_size,
                    config.rotary_pct,
                    config.rotary_emb_base,
                    config.layer_norm_eps,
                    config.use_parallel_residual,
                    process_group,
                )
                for _ in range(config.num_hidden_layers)
            ]
        )
        self.final_layer_norm = nn.LayerNorm(
            config.hidden_size, eps=config.layer_norm_eps
        )

        self.gradient_checkpointing = False

        self.head_size = self.layers[0].attention.head_size
        self.num_heads = self.layers[0].attention.num_heads

    def forward(
        self,
        input_ids,
        position_ids,
        cu_seqlens,
        max_s,
        past_key_values=None,
    ):
        hidden_states = self.embed_in(input_ids)

        # Prefill
        if past_key_values is None:
            # Create past tensor
            past_key_values = hidden_states.new_empty(
                (
                    len(self.layers),
                    len(hidden_states),
                    2,
                    self.num_heads,
                    self.head_size,
                )
            )
            layer_past_present_indices = None
            cu_seqlens_q = None
        # Decode
        else:
            # Create indices from cumulative sequence lengths
            layer_past_present_indices = cu_seqlens[1:] - 1
            cu_seqlens_q = torch.arange(
                cu_seqlens.shape[0], dtype=torch.int32, device=hidden_states.device
            )

        # Get rotary cos and sin for this forward
        # Avoid to index in each layer
        cos, sin = self.layers[0].attention.rotary_emb.get_cos_sin(
            position_ids, max_s, hidden_states.dtype
        )

        residual = None
        for i, layer in enumerate(self.layers):
            hidden_states, residual = layer(
                hidden_states,
                residual,
                cos,
                sin,
                cu_seqlens,
                max_s,
                past_key_values[i],
                layer_past_present_indices,
                cu_seqlens_q,
            )

        # Faster final layer norm
        hidden_states, *rest = dropout_layer_norm.dropout_add_ln_fwd(
            hidden_states,
            residual,
            self.final_layer_norm.weight,
            self.final_layer_norm.bias,
            None,
            None,
            None,
            None,
            0.0,
            self.final_layer_norm.eps,
            1.0,
            0,
            None,
            False,
            False,
        )

        return hidden_states, past_key_values


class FlashGPTNeoXForCausalLM(FlashGPTNeoXPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        if config.tp_parallel:
            process_group = torch.distributed.distributed_c10d._get_default_group()
        else:
            process_group = None

        self.gpt_neox = FlashGPTNeoXModel(config, process_group)

        if self.gpt_neox.tp_embeddings:
            self.embed_out = FastLinear(
                config.hidden_size,
                config.vocab_size // process_group.size(),
                bias=False,
            )
        else:
            self.embed_out = FastLinear(
                config.hidden_size, config.vocab_size, bias=False
            )

    def forward(
        self,
        input_ids,
        position_ids,
        cu_seqlens,
        max_s,
        past_key_values=None,
    ):
        hidden_states, present = self.gpt_neox(
            input_ids, position_ids, cu_seqlens, max_s, past_key_values
        )
        return self.embed_out(hidden_states), present