ProMoE-B-256/transformer/backbone_tcdit.py

import math

import torch
import torch.nn as nn
import torch.nn.functional as F

from .modeling_promoe_common import (
    Attention,
    FinalLayer,
    LabelEmbedder,
    Mlp,
    MoeMLP_DiffMoE as MoeMLP,
    PatchEmbed,
    TimestepEmbedder,
    get_2d_sincos_pos_embed,
    modulate,
)


class MoEGate(nn.Module):
    def __init__(self, embed_dim, num_experts=16, num_experts_per_tok=2, aux_loss_alpha=0.01):
        super().__init__()
        self.top_k = num_experts_per_tok
        self.n_routed_experts = num_experts
        self.scoring_func = "softmax"
        self.alpha = aux_loss_alpha
        self.seq_aux = False
        self.norm_topk_prob = False
        self.gating_dim = embed_dim
        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim)))
        self.reset_parameters()

    def reset_parameters(self):
        nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))

    def forward(self, hidden_states):
        bsz, seq_len, h = hidden_states.shape
        hidden_states = hidden_states.view(-1, h)
        logits = F.linear(hidden_states, self.weight, None)
        if self.scoring_func != "softmax":
            raise NotImplementedError(f"Unsupported gating scoring function: {self.scoring_func}")
        scores = logits.softmax(dim=-1)
        topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)
        if self.top_k > 1 and self.norm_topk_prob:
            topk_weight = topk_weight / (topk_weight.sum(dim=-1, keepdim=True) + 1e-20)

        if self.training and self.alpha > 0.0:
            scores_for_aux = scores
            topk_idx_for_aux_loss = topk_idx.view(bsz, -1)
            if self.seq_aux:
                scores_for_seq_aux = scores_for_aux.view(bsz, seq_len, -1)
                ce = torch.zeros(bsz, self.n_routed_experts, device=hidden_states.device)
                ce.scatter_add_(
                    1,
                    topk_idx_for_aux_loss,
                    torch.ones(bsz, seq_len * self.top_k, device=hidden_states.device),
                ).div_(seq_len * self.top_k / self.n_routed_experts)
                aux_loss = (ce * scores_for_seq_aux.mean(dim=1)).sum(dim=1).mean() * self.alpha
            else:
                mask_ce = F.one_hot(topk_idx_for_aux_loss.view(-1), num_classes=self.n_routed_experts)
                ce = mask_ce.float().mean(0)
                pi = scores_for_aux.mean(0)
                fi = ce * self.n_routed_experts
                aux_loss = (pi * fi).sum() * self.alpha
        else:
            aux_loss = None
        return topk_idx, topk_weight, aux_loss


class AddAuxiliaryLoss(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x, loss):
        ctx.dtype = loss.dtype
        ctx.required_aux_loss = loss.requires_grad
        return x

    @staticmethod
    def backward(ctx, grad_output):
        grad_loss = torch.ones(1, dtype=ctx.dtype, device=grad_output.device) if ctx.required_aux_loss else None
        return grad_output, grad_loss


class SparseMoEBlock(nn.Module):
    def __init__(
        self,
        experts,
        hidden_dim,
        mlp_ratio=4,
        num_experts=16,
        num_experts_per_tok=2,
        pretraining_tp=2,
        n_shared_experts=2,
    ):
        super().__init__()
        self.top_k = num_experts_per_tok
        self.experts = nn.ModuleList(experts)
        self.gate = MoEGate(embed_dim=hidden_dim, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok)
        self.n_shared_experts = n_shared_experts
        if self.n_shared_experts > 0:
            intermediate_size = hidden_dim * self.n_shared_experts
            self.shared_experts = MoeMLP(
                hidden_size=hidden_dim,
                intermediate_size=intermediate_size,
                pretraining_tp=pretraining_tp,
            )

    def forward(self, hidden_states):
        identity = hidden_states
        orig_shape = hidden_states.shape
        topk_idx, topk_weight, aux_loss = self.gate(hidden_states)

        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        flat_topk_idx = topk_idx.view(-1)
        if self.training:
            hidden_states = hidden_states.repeat_interleave(self.top_k, dim=0)
            y = torch.empty_like(hidden_states, dtype=hidden_states.dtype)
            for i, expert in enumerate(self.experts):
                y[flat_topk_idx == i] = expert(hidden_states[flat_topk_idx == i]).float()
            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
            y = y.view(*orig_shape)
            y = AddAuxiliaryLoss.apply(y, aux_loss)
        else:
            y = self.moe_infer(hidden_states, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape)
        if self.n_shared_experts > 0:
            y = y + self.shared_experts(identity)
        return y

    @torch.no_grad()
    def moe_infer(self, x, flat_expert_indices, flat_expert_weights):
        expert_cache = torch.zeros_like(x)
        idxs = flat_expert_indices.argsort()
        tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0)
        token_idxs = idxs // self.top_k
        for i, end_idx in enumerate(tokens_per_expert):
            start_idx = 0 if i == 0 else tokens_per_expert[i - 1]
            if start_idx == end_idx:
                continue
            expert = self.experts[i]
            exp_token_idx = token_idxs[start_idx:end_idx]
            expert_tokens = x[exp_token_idx]
            expert_out = expert(expert_tokens)
            expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
            expert_cache = expert_cache.to(expert_out.dtype)
            expert_cache.scatter_reduce_(
                0,
                exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]),
                expert_out,
                reduce="sum",
            )
        return expert_cache


class DiTBlock(nn.Module):
    def __init__(
        self,
        hidden_size,
        num_heads,
        mlp_ratio=4,
        pretraining_tp=2,
        use_swiglu=False,
        MoE_config=None,
        use_moe=True,
        **block_kwargs,
    ):
        super().__init__()
        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)
        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        self.use_moe = use_moe
        if use_moe:
            if not use_swiglu:
                approx_gelu = lambda: nn.GELU(approximate="tanh")
                experts = [
                    Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)
                    for _ in range(MoE_config.num_experts)
                ]
            else:
                experts = [
                    MoeMLP(
                        hidden_size=hidden_size,
                        intermediate_size=mlp_hidden_dim,
                        pretraining_tp=pretraining_tp,
                    )
                    for _ in range(MoE_config.num_experts)
                ]
            self.mlp = SparseMoEBlock(
                experts=experts,
                hidden_dim=hidden_size,
                num_experts=MoE_config.num_experts,
                num_experts_per_tok=MoE_config.capacity,
                n_shared_experts=MoE_config.n_shared_experts,
            )
        else:
            if not use_swiglu:
                approx_gelu = lambda: nn.GELU(approximate="tanh")
                self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)
            else:
                self.mlp = MoeMLP(hidden_size=hidden_size, intermediate_size=mlp_hidden_dim)

        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True))

    def forward(self, x, c):
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
        x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
        x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
        return x


class DiT(nn.Module):
    def __init__(
        self,
        input_size=32,
        patch_size=2,
        in_channels=4,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4,
        qk_norm=False,
        class_dropout_prob=0.1,
        num_classes=1000,
        pretraining_tp=1,
        learn_sigma=True,
        use_swiglu=False,
        MoE_config=None,
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = in_channels
        self.out_channels = in_channels * 2 if learn_sigma else in_channels
        self.patch_size = patch_size
        self.num_heads = num_heads
        self.MoE_config = MoE_config
        use_moe_flag = [i % 2 == 1 for i in range(depth)] if self.MoE_config.interleave else [True] * depth

        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
        self.t_embedder = TimestepEmbedder(hidden_size)
        self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob)
        num_patches = self.x_embedder.num_patches
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size), requires_grad=False)

        self.blocks = nn.ModuleList(
            [
                DiTBlock(
                    hidden_size,
                    num_heads,
                    mlp_ratio=mlp_ratio,
                    qk_norm=qk_norm,
                    use_swiglu=use_swiglu,
                    pretraining_tp=pretraining_tp,
                    MoE_config=MoE_config,
                    use_moe=use_moe_flag[i],
                )
                for i in range(depth)
            ]
        )
        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)
        self.initialize_weights()

    def initialize_weights(self):
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)

        self.apply(_basic_init)
        pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches**0.5))
        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
        w = self.x_embedder.proj.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
        nn.init.constant_(self.x_embedder.proj.bias, 0)
        nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
        for block in self.blocks:
            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
        nn.init.constant_(self.final_layer.linear.weight, 0)
        nn.init.constant_(self.final_layer.linear.bias, 0)

    def unpatchify(self, x):
        c = self.out_channels
        p = self.x_embedder.patch_size[0]
        h = w = int(x.shape[1] ** 0.5)
        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
        x = torch.einsum("nhwpqc->nchpwq", x)
        return x.reshape(shape=(x.shape[0], c, h * p, h * p))

    def forward(self, x, t, context, **kwargs):
        y = context
        if len(x.shape) != 4:
            x = x.squeeze(2)
        x = self.x_embedder(x) + self.pos_embed
        t = self.t_embedder(t)
        y = self.y_embedder(y, self.training)
        c = t + y
        for block in self.blocks:
            x = block(x, c)
        x = self.final_layer(x, c)
        return self.unpatchify(x)