Initial support for hidream. Still a WIP

extensions_built_in/diffusion_models/hidream/src/models/attention.py

@@ -0,0 +1,106 @@
+import torch
+from torch import nn
+from typing import Optional
+from diffusers.models.attention_processor import Attention
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+
+@maybe_allow_in_graph
+class HiDreamAttention(Attention):
+    def __init__(
+        self,
+        query_dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        scale_qk: bool = True,
+        eps: float = 1e-5,
+        processor = None,
+        out_dim: int = None,
+        single: bool = False
+    ):
+        super(Attention, self).__init__()
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+        self.out_dim = out_dim if out_dim is not None else query_dim
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        self.sliceable_head_dim = heads
+        self.single = single
+
+        linear_cls = nn.Linear
+        self.linear_cls = linear_cls
+        self.to_q = linear_cls(query_dim, self.inner_dim)
+        self.to_k = linear_cls(self.inner_dim, self.inner_dim)
+        self.to_v = linear_cls(self.inner_dim, self.inner_dim)
+        self.to_out = linear_cls(self.inner_dim, self.out_dim)
+        self.q_rms_norm = nn.RMSNorm(self.inner_dim, eps)
+        self.k_rms_norm = nn.RMSNorm(self.inner_dim, eps)
+
+        if not single:
+            self.to_q_t = linear_cls(query_dim, self.inner_dim)
+            self.to_k_t = linear_cls(self.inner_dim, self.inner_dim)
+            self.to_v_t = linear_cls(self.inner_dim, self.inner_dim)
+            self.to_out_t = linear_cls(self.inner_dim, self.out_dim)
+            self.q_rms_norm_t = nn.RMSNorm(self.inner_dim, eps)
+            self.k_rms_norm_t = nn.RMSNorm(self.inner_dim, eps)
+
+        self.set_processor(processor)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.xavier_uniform_(m.weight)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(
+        self,
+        norm_image_tokens: torch.FloatTensor,
+        image_tokens_masks: torch.FloatTensor = None,
+        norm_text_tokens: torch.FloatTensor = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.Tensor:
+        return self.processor(
+            self,
+            image_tokens = norm_image_tokens,
+            image_tokens_masks = image_tokens_masks,
+            text_tokens = norm_text_tokens,
+            rope = rope,
+        )
+
+class FeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+    ):
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * (
+            (hidden_dim + multiple_of - 1) // multiple_of
+        )
+
+        self.w1 = nn.Linear(dim, hidden_dim, bias=False)
+        self.w2 = nn.Linear(hidden_dim, dim, bias=False)
+        self.w3 = nn.Linear(dim, hidden_dim, bias=False)
+        self.apply(self._init_weights)
+    
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.xavier_uniform_(m.weight)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        return self.w2(torch.nn.functional.silu(self.w1(x)) * self.w3(x))

extensions_built_in/diffusion_models/hidream/src/models/attention_processor.py

@@ -0,0 +1,95 @@
+from typing import Optional
+import torch
+from .attention import HiDreamAttention
+
+try:
+    from flash_attn_interface import flash_attn_func
+    USE_FLASH_ATTN3 = True
+except:
+    from flash_attn import flash_attn_func
+    USE_FLASH_ATTN3 = False
+
+# Copied from https://github.com/black-forest-labs/flux/blob/main/src/flux/math.py
+def apply_rope(xq: torch.Tensor, xk: torch.Tensor, freqs_cis: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
+    xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
+    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
+    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
+    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
+
+def attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor):
+    if USE_FLASH_ATTN3:
+        hidden_states = flash_attn_func(query, key, value, causal=False, deterministic=False)[0]
+    else:
+        hidden_states = flash_attn_func(query, key, value, dropout_p=0., causal=False)
+    hidden_states = hidden_states.flatten(-2)
+    hidden_states = hidden_states.to(query.dtype)
+    return hidden_states
+
+class HiDreamAttnProcessor_flashattn:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __call__(
+        self,
+        attn: HiDreamAttention,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        dtype = image_tokens.dtype
+        batch_size = image_tokens.shape[0]
+
+        query_i = attn.q_rms_norm(attn.to_q(image_tokens)).to(dtype=dtype)
+        key_i = attn.k_rms_norm(attn.to_k(image_tokens)).to(dtype=dtype)
+        value_i = attn.to_v(image_tokens)
+
+        inner_dim = key_i.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query_i = query_i.view(batch_size, -1, attn.heads, head_dim)
+        key_i = key_i.view(batch_size, -1, attn.heads, head_dim)
+        value_i = value_i.view(batch_size, -1, attn.heads, head_dim)
+        if image_tokens_masks is not None:
+            key_i = key_i * image_tokens_masks.view(batch_size, -1, 1, 1)
+
+        if not attn.single:
+            query_t = attn.q_rms_norm_t(attn.to_q_t(text_tokens)).to(dtype=dtype)
+            key_t = attn.k_rms_norm_t(attn.to_k_t(text_tokens)).to(dtype=dtype)
+            value_t = attn.to_v_t(text_tokens)
+
+            query_t = query_t.view(batch_size, -1, attn.heads, head_dim)
+            key_t = key_t.view(batch_size, -1, attn.heads, head_dim)
+            value_t = value_t.view(batch_size, -1, attn.heads, head_dim)
+
+            num_image_tokens = query_i.shape[1]
+            num_text_tokens = query_t.shape[1]
+            query = torch.cat([query_i, query_t], dim=1)
+            key = torch.cat([key_i, key_t], dim=1)
+            value = torch.cat([value_i, value_t], dim=1)
+        else:
+            query = query_i
+            key = key_i
+            value = value_i
+        
+        if query.shape[-1] == rope.shape[-3] * 2:
+            query, key = apply_rope(query, key, rope)
+        else:
+            query_1, query_2 = query.chunk(2, dim=-1)
+            key_1, key_2 = key.chunk(2, dim=-1)
+            query_1, key_1 = apply_rope(query_1, key_1, rope)
+            query = torch.cat([query_1, query_2], dim=-1)
+            key = torch.cat([key_1, key_2], dim=-1)
+
+        hidden_states = attention(query, key, value)
+
+        if not attn.single:
+            hidden_states_i, hidden_states_t = torch.split(hidden_states, [num_image_tokens, num_text_tokens], dim=1)
+            hidden_states_i = attn.to_out(hidden_states_i)
+            hidden_states_t = attn.to_out_t(hidden_states_t)
+            return hidden_states_i, hidden_states_t
+        else:
+            hidden_states = attn.to_out(hidden_states)
+            return hidden_states

extensions_built_in/diffusion_models/hidream/src/models/embeddings.py

@@ -0,0 +1,114 @@
+import torch
+from torch import nn
+from typing import List
+from diffusers.models.embeddings import Timesteps, TimestepEmbedding
+
+# Copied from https://github.com/black-forest-labs/flux/blob/main/src/flux/math.py
+def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
+    assert dim % 2 == 0, "The dimension must be even."
+
+    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
+    omega = 1.0 / (theta**scale)
+
+    batch_size, seq_length = pos.shape
+    out = torch.einsum("...n,d->...nd", pos, omega)
+    cos_out = torch.cos(out)
+    sin_out = torch.sin(out)
+
+    stacked_out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
+    out = stacked_out.view(batch_size, -1, dim // 2, 2, 2)
+    return out.float()
+
+# Copied from https://github.com/black-forest-labs/flux/blob/main/src/flux/modules/layers.py
+class EmbedND(nn.Module):
+    def __init__(self, theta: int, axes_dim: List[int]):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        n_axes = ids.shape[-1]
+        emb = torch.cat(
+            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
+            dim=-3,
+        )
+        return emb.unsqueeze(2)
+    
+class PatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size=2,
+        in_channels=4,
+        out_channels=1024,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.out_channels = out_channels
+        self.proj = nn.Linear(in_channels * patch_size * patch_size, out_channels, bias=True)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.xavier_uniform_(m.weight)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, latent):
+        latent = self.proj(latent)
+        return latent
+    
+class PooledEmbed(nn.Module):
+    def __init__(self, text_emb_dim, hidden_size): 
+        super().__init__()
+        self.pooled_embedder = TimestepEmbedding(in_channels=text_emb_dim, time_embed_dim=hidden_size)
+        self.apply(self._init_weights)
+    
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.normal_(m.weight, std=0.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, pooled_embed):
+        return self.pooled_embedder(pooled_embed)
+    
+class TimestepEmbed(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=frequency_embedding_size, time_embed_dim=hidden_size)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.normal_(m.weight, std=0.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, timesteps, wdtype):
+        t_emb = self.time_proj(timesteps).to(dtype=wdtype)
+        t_emb = self.timestep_embedder(t_emb)
+        return t_emb
+    
+class OutEmbed(nn.Module):
+    def __init__(self, hidden_size, patch_size, out_channels):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(hidden_size, 2 * hidden_size, bias=True)
+        )
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.zeros_(m.weight)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x, adaln_input):
+        shift, scale = self.adaLN_modulation(adaln_input).chunk(2, dim=1)
+        x = self.norm_final(x) * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+        x = self.linear(x)
+        return x

extensions_built_in/diffusion_models/hidream/src/models/moe.py

@@ -0,0 +1,154 @@
+import math
+import torch
+from torch import nn
+import torch.nn.functional as F
+from .attention import FeedForwardSwiGLU
+from torch.distributed.nn.functional import all_gather
+
+_LOAD_BALANCING_LOSS = []
+def save_load_balancing_loss(loss):
+    global _LOAD_BALANCING_LOSS
+    _LOAD_BALANCING_LOSS.append(loss)
+
+def clear_load_balancing_loss():
+    global _LOAD_BALANCING_LOSS
+    _LOAD_BALANCING_LOSS.clear()
+
+def get_load_balancing_loss():
+    global _LOAD_BALANCING_LOSS
+    return _LOAD_BALANCING_LOSS
+
+def batched_load_balancing_loss():
+    aux_losses_arr = get_load_balancing_loss()
+    alpha = aux_losses_arr[0][-1]
+    Pi = torch.stack([ent[1] for ent in aux_losses_arr], dim=0)
+    fi = torch.stack([ent[2] for ent in aux_losses_arr], dim=0)
+
+    fi_list = all_gather(fi)
+    fi = torch.stack(fi_list, 0).mean(0)
+
+    aux_loss = (Pi * fi).sum(-1).mean() * alpha
+    return aux_loss
+
+# Modified from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+class MoEGate(nn.Module):
+    def __init__(self, embed_dim, num_routed_experts=4, num_activated_experts=2, aux_loss_alpha=0.01):
+        super().__init__()
+        self.top_k = num_activated_experts
+        self.n_routed_experts = num_routed_experts
+
+        self.scoring_func = 'softmax'
+        self.alpha = aux_loss_alpha
+        self.seq_aux = False
+
+        # topk selection algorithm
+        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) -> None:
+        import torch.nn.init  as init
+        init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+    
+    def forward(self, hidden_states):
+        bsz, seq_len, h = hidden_states.shape    
+        # print(bsz, seq_len, h)    
+        ### compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        logits = F.linear(hidden_states, self.weight, None)
+        if self.scoring_func == 'softmax':
+            scores = logits.softmax(dim=-1)
+        else:
+            raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}')
+        
+        ### select top-k experts
+        topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)
+        
+        ### norm gate to sum 1
+        if self.top_k > 1 and self.norm_topk_prob:
+            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weight = topk_weight / denominator
+
+        ### expert-level computation auxiliary loss
+        if self.training and self.alpha > 0.0:
+            scores_for_aux = scores
+            aux_topk = self.top_k
+            # always compute aux loss based on the naive greedy topk method
+            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 * aux_topk, device=hidden_states.device)).div_(seq_len * aux_topk / 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
+                save_load_balancing_loss((aux_loss, Pi, fi, self.alpha))
+        else:
+            aux_loss = None
+        return topk_idx, topk_weight, aux_loss
+
+# Modified from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+class MOEFeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        num_routed_experts: int,
+        num_activated_experts: int,
+    ):
+        super().__init__()
+        self.shared_experts = FeedForwardSwiGLU(dim, hidden_dim // 2)
+        self.experts = nn.ModuleList([FeedForwardSwiGLU(dim, hidden_dim) for i in range(num_routed_experts)])
+        self.gate = MoEGate(
+            embed_dim = dim, 
+            num_routed_experts = num_routed_experts, 
+            num_activated_experts = num_activated_experts
+        )
+        self.num_activated_experts = num_activated_experts
+
+    def forward(self, x):
+        wtype = x.dtype
+        identity = x
+        orig_shape = x.shape
+        topk_idx, topk_weight, aux_loss = self.gate(x) 
+        x = x.view(-1, x.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+        if self.training:
+            x = x.repeat_interleave(self.num_activated_experts, dim=0)
+            y = torch.empty_like(x, dtype=wtype)
+            for i, expert in enumerate(self.experts): 
+                y[flat_topk_idx == i] = expert(x[flat_topk_idx == i]).to(dtype=wtype)
+            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+            y =  y.view(*orig_shape).to(dtype=wtype)
+            #y = AddAuxiliaryLoss.apply(y, aux_loss)
+        else:
+            y = self.moe_infer(x, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape)
+        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.num_activated_experts 
+        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]]) 
+            
+            # for fp16 and other dtype
+            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

extensions_built_in/diffusion_models/hidream/src/models/transformers/transformer_hidream_image.py

@@ -0,0 +1,543 @@
+from typing import Any, Callable, Dict, Optional, Tuple, List
+
+import torch
+import torch.nn as nn
+import einops
+from einops import repeat
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from ..embeddings import PatchEmbed, PooledEmbed, TimestepEmbed, EmbedND, OutEmbed
+from ..attention import HiDreamAttention, FeedForwardSwiGLU
+from ..attention_processor import HiDreamAttnProcessor_flashattn
+from ..moe import MOEFeedForwardSwiGLU
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+class TextProjection(nn.Module):
+    def __init__(self, in_features, hidden_size):
+        super().__init__()
+        self.linear = nn.Linear(in_features=in_features, out_features=hidden_size, bias=False)
+
+    def forward(self, caption):
+        hidden_states = self.linear(caption)
+        return hidden_states
+    
+class BlockType:
+    TransformerBlock = 1
+    SingleTransformerBlock = 2
+
+@maybe_allow_in_graph
+class HiDreamImageSingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(dim, 6 * dim, bias=True)
+        )
+        nn.init.zeros_(self.adaLN_modulation[1].weight)
+        nn.init.zeros_(self.adaLN_modulation[1].bias)
+
+        # 1. Attention
+        self.norm1_i = nn.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        self.attn1 = HiDreamAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            processor = HiDreamAttnProcessor_flashattn(),
+            single = True
+        )
+
+        # 3. Feed-forward
+        self.norm3_i = nn.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        if num_routed_experts > 0:
+            self.ff_i = MOEFeedForwardSwiGLU(
+                dim = dim, 
+                hidden_dim = 4 * dim,
+                num_routed_experts = num_routed_experts,
+                num_activated_experts = num_activated_experts,
+            )
+        else:
+            self.ff_i = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim)
+    
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+
+    ) -> torch.FloatTensor:
+        wtype = image_tokens.dtype
+        shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i = \
+            self.adaLN_modulation(adaln_input)[:,None].chunk(6, dim=-1)
+        
+        # 1. MM-Attention
+        norm_image_tokens = self.norm1_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_msa_i) + shift_msa_i
+        attn_output_i = self.attn1(
+            norm_image_tokens,
+            image_tokens_masks,
+            rope = rope,
+        )
+        image_tokens = gate_msa_i * attn_output_i + image_tokens
+        
+        # 2. Feed-forward
+        norm_image_tokens = self.norm3_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_mlp_i) + shift_mlp_i
+        ff_output_i = gate_mlp_i * self.ff_i(norm_image_tokens.to(dtype=wtype))
+        image_tokens = ff_output_i + image_tokens
+        return image_tokens
+
+@maybe_allow_in_graph
+class HiDreamImageTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(dim, 12 * dim, bias=True)
+        )
+        nn.init.zeros_(self.adaLN_modulation[1].weight)
+        nn.init.zeros_(self.adaLN_modulation[1].bias)
+
+        # 1. Attention
+        self.norm1_i = nn.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        self.norm1_t = nn.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        self.attn1 = HiDreamAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            processor = HiDreamAttnProcessor_flashattn(),
+            single = False
+        )
+
+        # 3. Feed-forward
+        self.norm3_i = nn.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        if num_routed_experts > 0:
+            self.ff_i = MOEFeedForwardSwiGLU(
+                dim = dim, 
+                hidden_dim = 4 * dim,
+                num_routed_experts = num_routed_experts,
+                num_activated_experts = num_activated_experts,
+            )
+        else:
+            self.ff_i = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim)
+        self.norm3_t = nn.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        self.ff_t = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim)
+    
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        wtype = image_tokens.dtype
+        shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i, \
+        shift_msa_t, scale_msa_t, gate_msa_t, shift_mlp_t, scale_mlp_t, gate_mlp_t = \
+            self.adaLN_modulation(adaln_input)[:,None].chunk(12, dim=-1)
+        
+        # 1. MM-Attention
+        norm_image_tokens = self.norm1_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_msa_i) + shift_msa_i
+        norm_text_tokens = self.norm1_t(text_tokens).to(dtype=wtype)
+        norm_text_tokens = norm_text_tokens * (1 + scale_msa_t) + shift_msa_t
+
+        attn_output_i, attn_output_t = self.attn1(
+            norm_image_tokens,
+            image_tokens_masks,
+            norm_text_tokens,
+            rope = rope,
+        )
+
+        image_tokens = gate_msa_i * attn_output_i + image_tokens
+        text_tokens = gate_msa_t * attn_output_t + text_tokens
+        
+        # 2. Feed-forward
+        norm_image_tokens = self.norm3_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_mlp_i) + shift_mlp_i
+        norm_text_tokens = self.norm3_t(text_tokens).to(dtype=wtype)
+        norm_text_tokens = norm_text_tokens * (1 + scale_mlp_t) + shift_mlp_t
+
+        ff_output_i = gate_mlp_i * self.ff_i(norm_image_tokens)
+        ff_output_t = gate_mlp_t * self.ff_t(norm_text_tokens)
+        image_tokens = ff_output_i + image_tokens
+        text_tokens = ff_output_t + text_tokens
+        return image_tokens, text_tokens
+    
+@maybe_allow_in_graph
+class HiDreamImageBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        block_type: BlockType = BlockType.TransformerBlock,
+    ):
+        super().__init__()
+        block_classes = {
+            BlockType.TransformerBlock: HiDreamImageTransformerBlock,
+            BlockType.SingleTransformerBlock: HiDreamImageSingleTransformerBlock,
+        }
+        self.block = block_classes[block_type](
+            dim,
+            num_attention_heads,
+            attention_head_dim,
+            num_routed_experts,
+            num_activated_experts
+        )
+    
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: torch.FloatTensor = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        return self.block(
+            image_tokens,
+            image_tokens_masks,
+            text_tokens,
+            adaln_input,
+            rope,
+        )
+
+class HiDreamImageTransformer2DModel(
+    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin
+):
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["HiDreamImageBlock"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: Optional[int] = None,
+        in_channels: int = 64,
+        out_channels: Optional[int] = None,
+        num_layers: int = 16,
+        num_single_layers: int = 32,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 20,
+        caption_channels: List[int] = None,
+        text_emb_dim: int = 2048,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        axes_dims_rope: Tuple[int, int] = (32, 32),
+        max_resolution: Tuple[int, int] = (128, 128),
+        llama_layers: List[int] = None, 
+    ):
+        super().__init__()
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.llama_layers = llama_layers
+
+        self.t_embedder = TimestepEmbed(self.inner_dim)
+        self.p_embedder = PooledEmbed(text_emb_dim, self.inner_dim)
+        self.x_embedder = PatchEmbed(
+            patch_size = patch_size,
+            in_channels = in_channels,
+            out_channels = self.inner_dim,
+        )
+        self.pe_embedder = EmbedND(theta=10000, axes_dim=axes_dims_rope)
+
+        self.double_stream_blocks = nn.ModuleList(
+            [
+                HiDreamImageBlock(
+                    dim = self.inner_dim,
+                    num_attention_heads = self.config.num_attention_heads,
+                    attention_head_dim = self.config.attention_head_dim,
+                    num_routed_experts = num_routed_experts,
+                    num_activated_experts = num_activated_experts,
+                    block_type = BlockType.TransformerBlock
+                )
+                for i in range(self.config.num_layers)
+            ]
+        )
+
+        self.single_stream_blocks = nn.ModuleList(
+            [
+                HiDreamImageBlock(
+                    dim = self.inner_dim,
+                    num_attention_heads = self.config.num_attention_heads,
+                    attention_head_dim = self.config.attention_head_dim,
+                    num_routed_experts = num_routed_experts,
+                    num_activated_experts = num_activated_experts,
+                    block_type = BlockType.SingleTransformerBlock
+                )
+                for i in range(self.config.num_single_layers)
+            ]
+        )
+
+        self.final_layer = OutEmbed(self.inner_dim, patch_size, self.out_channels)
+
+        caption_channels = [caption_channels[1], ] * (num_layers + num_single_layers) + [caption_channels[0], ]
+        caption_projection = []
+        for caption_channel in caption_channels:
+            caption_projection.append(TextProjection(in_features = caption_channel, hidden_size = self.inner_dim))
+        self.caption_projection = nn.ModuleList(caption_projection)
+        self.max_seq = max_resolution[0] * max_resolution[1] // (patch_size * patch_size)
+
+        self.gradient_checkpointing = False
+
+    # def _set_gradient_checkpointing(self, module, value=False):
+    #     if hasattr(module, "gradient_checkpointing"):
+    #         module.gradient_checkpointing = value
+    def _set_gradient_checkpointing(
+        self, enable: bool = True, gradient_checkpointing_func: Callable = torch.utils.checkpoint.checkpoint
+    ) -> None:
+        is_gradient_checkpointing_set = False
+
+        for name, module in self.named_modules():
+            if hasattr(module, "gradient_checkpointing"):
+                logger.debug(f"Setting `gradient_checkpointing={enable}` for '{name}'")
+                module._gradient_checkpointing_func = gradient_checkpointing_func
+                module.gradient_checkpointing = enable
+                is_gradient_checkpointing_set = True
+
+        if not is_gradient_checkpointing_set:
+            raise ValueError(
+                f"The module {self.__class__.__name__} does not support gradient checkpointing. Please make sure to "
+                f"use a module that supports gradient checkpointing by creating a boolean attribute `gradient_checkpointing`."
+            )
+
+    def expand_timesteps(self, timesteps, batch_size, device):
+        if not torch.is_tensor(timesteps):
+            is_mps = device.type == "mps"
+            if isinstance(timesteps, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(device)
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(batch_size)
+        return timesteps
+
+    def unpatchify(self, x: torch.Tensor, img_sizes: List[Tuple[int, int]], is_training: bool) -> List[torch.Tensor]:            
+        if is_training:
+            x = einops.rearrange(x, 'B S (p1 p2 C) -> B C S (p1 p2)', p1=self.config.patch_size, p2=self.config.patch_size)
+        else:
+            x_arr = []
+            for i, img_size in enumerate(img_sizes):
+                pH, pW = img_size
+                x_arr.append(
+                    einops.rearrange(x[i, :pH*pW].reshape(1, pH, pW, -1), 'B H W (p1 p2 C) -> B C (H p1) (W p2)', 
+                        p1=self.config.patch_size, p2=self.config.patch_size)
+                )
+            x = torch.cat(x_arr, dim=0)
+        return x
+
+    def patchify(self, x, max_seq, img_sizes=None):
+        pz2 = self.config.patch_size * self.config.patch_size
+        if isinstance(x, torch.Tensor):
+            B, C = x.shape[0], x.shape[1]
+            device = x.device
+            dtype = x.dtype
+        else:
+            B, C = len(x), x[0].shape[0]
+            device = x[0].device
+            dtype = x[0].dtype
+        x_masks = torch.zeros((B, max_seq), dtype=dtype, device=device)
+
+        if img_sizes is not None:
+            for i, img_size in enumerate(img_sizes):
+                x_masks[i, 0:img_size[0] * img_size[1]] = 1
+            x = einops.rearrange(x, 'B C S p -> B S (p C)', p=pz2)
+        elif isinstance(x, torch.Tensor):
+            pH, pW = x.shape[-2] // self.config.patch_size, x.shape[-1] // self.config.patch_size
+            x = einops.rearrange(x, 'B C (H p1) (W p2) -> B (H W) (p1 p2 C)', p1=self.config.patch_size, p2=self.config.patch_size)
+            img_sizes = [[pH, pW]] * B
+            x_masks = None
+        else:
+            raise NotImplementedError
+        return x, x_masks, img_sizes
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timesteps: torch.LongTensor = None,
+        encoder_hidden_states: torch.Tensor = None,
+        pooled_embeds: torch.Tensor = None,
+        img_sizes: Optional[List[Tuple[int, int]]] = None,
+        img_ids: Optional[torch.Tensor] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ):
+        if joint_attention_kwargs is not None:
+            joint_attention_kwargs = joint_attention_kwargs.copy()
+            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        # spatial forward
+        batch_size = hidden_states.shape[0]
+        hidden_states_type = hidden_states.dtype
+
+        # 0. time
+        timesteps = self.expand_timesteps(timesteps, batch_size, hidden_states.device)
+        timesteps = self.t_embedder(timesteps, hidden_states_type)
+        p_embedder = self.p_embedder(pooled_embeds)
+        adaln_input = timesteps + p_embedder
+
+        hidden_states, image_tokens_masks, img_sizes = self.patchify(hidden_states, self.max_seq, img_sizes)
+        if image_tokens_masks is None:
+            pH, pW = img_sizes[0]
+            img_ids = torch.zeros(pH, pW, 3, device=hidden_states.device)
+            img_ids[..., 1] = img_ids[..., 1] + torch.arange(pH, device=hidden_states.device)[:, None]
+            img_ids[..., 2] = img_ids[..., 2] + torch.arange(pW, device=hidden_states.device)[None, :]
+            img_ids = repeat(img_ids, "h w c -> b (h w) c", b=batch_size)
+        hidden_states = self.x_embedder(hidden_states)
+
+        T5_encoder_hidden_states = encoder_hidden_states[0]
+        encoder_hidden_states = encoder_hidden_states[-1]
+        encoder_hidden_states = [encoder_hidden_states[k] for k in self.llama_layers]
+
+        if self.caption_projection is not None:
+            new_encoder_hidden_states = []
+            for i, enc_hidden_state in enumerate(encoder_hidden_states):
+                enc_hidden_state = self.caption_projection[i](enc_hidden_state)
+                enc_hidden_state = enc_hidden_state.view(batch_size, -1, hidden_states.shape[-1])
+                new_encoder_hidden_states.append(enc_hidden_state)
+            encoder_hidden_states = new_encoder_hidden_states
+            T5_encoder_hidden_states = self.caption_projection[-1](T5_encoder_hidden_states)
+            T5_encoder_hidden_states = T5_encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+            encoder_hidden_states.append(T5_encoder_hidden_states)
+
+        txt_ids = torch.zeros(
+            batch_size, 
+            encoder_hidden_states[-1].shape[1] + encoder_hidden_states[-2].shape[1] + encoder_hidden_states[0].shape[1], 
+            3, 
+            device=img_ids.device, dtype=img_ids.dtype
+        )
+        ids = torch.cat((img_ids, txt_ids), dim=1)
+        rope = self.pe_embedder(ids)
+
+        # 2. Blocks
+        block_id = 0
+        initial_encoder_hidden_states = torch.cat([encoder_hidden_states[-1], encoder_hidden_states[-2]], dim=1)
+        initial_encoder_hidden_states_seq_len = initial_encoder_hidden_states.shape[1]
+        for bid, block in enumerate(self.double_stream_blocks):
+            cur_llama31_encoder_hidden_states = encoder_hidden_states[block_id]
+            cur_encoder_hidden_states = torch.cat([initial_encoder_hidden_states, cur_llama31_encoder_hidden_states], dim=1)
+            if self.training and self.gradient_checkpointing:
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+                    return custom_forward
+                
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states, initial_encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    image_tokens_masks,
+                    cur_encoder_hidden_states,
+                    adaln_input,
+                    rope,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, initial_encoder_hidden_states = block(
+                    image_tokens = hidden_states,
+                    image_tokens_masks = image_tokens_masks,
+                    text_tokens = cur_encoder_hidden_states,
+                    adaln_input = adaln_input,
+                    rope = rope,
+                )
+            initial_encoder_hidden_states = initial_encoder_hidden_states[:, :initial_encoder_hidden_states_seq_len]
+            block_id += 1
+
+        image_tokens_seq_len = hidden_states.shape[1]
+        hidden_states = torch.cat([hidden_states, initial_encoder_hidden_states], dim=1)
+        hidden_states_seq_len = hidden_states.shape[1]
+        if image_tokens_masks is not None:
+            encoder_attention_mask_ones = torch.ones(
+                (batch_size, initial_encoder_hidden_states.shape[1] + cur_llama31_encoder_hidden_states.shape[1]),
+                device=image_tokens_masks.device, dtype=image_tokens_masks.dtype
+            )
+            image_tokens_masks = torch.cat([image_tokens_masks, encoder_attention_mask_ones], dim=1)
+
+        for bid, block in enumerate(self.single_stream_blocks):
+            cur_llama31_encoder_hidden_states = encoder_hidden_states[block_id]
+            hidden_states = torch.cat([hidden_states, cur_llama31_encoder_hidden_states], dim=1)
+            if self.training and self.gradient_checkpointing:
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+                    return custom_forward
+                
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    image_tokens_masks,
+                    None,
+                    adaln_input,
+                    rope,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    image_tokens = hidden_states,
+                    image_tokens_masks = image_tokens_masks,
+                    text_tokens = None,
+                    adaln_input = adaln_input,
+                    rope = rope,
+                )
+            hidden_states = hidden_states[:, :hidden_states_seq_len]
+            block_id += 1
+        
+        hidden_states = hidden_states[:, :image_tokens_seq_len, ...]
+        output = self.final_layer(hidden_states, adaln_input)
+        output = self.unpatchify(output, img_sizes, self.training)
+        if image_tokens_masks is not None:
+            image_tokens_masks = image_tokens_masks[:, :image_tokens_seq_len]
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output, image_tokens_masks)
+        return Transformer2DModelOutput(sample=output, mask=image_tokens_masks)
+