F-lite first trial

.gitignore

@@ -0,0 +1,74 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Distribution / packaging
+.Python
+env/
+.coverage
+htmlcov/
+
+# IDE specific files
+.idea/
+.vscode/
+*.swp
+*.swo
+.DS_Store
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# ML specific
+*.onnx
+*.pt
+*.pth
+*.bin
+*.h5
+*.hdf5
+runs/
+wandb/
+checkpoints/
+outputs/
+logs/
+models/
+
+# Project specific
+.pytest_cache/
+.coverage
+htmlcov/
+
+# Large media files
+*.mp4
+*.tiff
+*.avi
+*.flv
+*.mov
+*.wmv 

app.py

@@ -2,26 +2,33 @@ import gradio as gr
 import numpy as np
 import random
 
-import spaces #[uncomment to use ZeroGPU]
-from diffusers import DiffusionPipeline
+import spaces
 import torch
+from pikigen import PikigenPipeline
+
+# Trick required because it is not a native diffusers model
+from diffusers.pipelines.pipeline_loading_utils import LOADABLE_CLASSES, ALL_IMPORTABLE_CLASSES
+LOADABLE_CLASSES.setdefault("pikigen", {}).setdefault("DiT", []).extend(["save_pretrained", "from_pretrained"])
+ALL_IMPORTABLE_CLASSES.setdefault("DiT", []).extend(["save_pretrained", "from_pretrained"])
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use
+model_repo_id = "Freepik/Pikigen-test"
 
 if torch.cuda.is_available():
-    torch_dtype = torch.float16
+    torch_dtype = torch.bfloat16
 else:
     torch_dtype = torch.float32
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
+pipe = PikigenPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
+pipe.enable_model_cpu_offload()  # For less memory consumption
+pipe.vae.enable_slicing()
+pipe.vae.enable_tiling()
 
 MAX_SEED = np.iinfo(np.int32).max
 MAX_IMAGE_SIZE = 1024
 
 
-@spaces.GPU #[uncomment to use ZeroGPU]
+@spaces.GPU
 def infer(
     prompt,
     negative_prompt,
@@ -52,9 +59,9 @@ def infer(
 
 
 examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
+    "A photorealistic 3D render of a charming, mischievous young boy with long, floppy donkey ears and a small pink pig nose",
+    "A landscape photograph showing a serene mountain lake at sunset with reflections in crystal clear water",
+    "A detailed digital painting of a futuristic cyberpunk city with neon lights and flying vehicles",
 ]
 
 css = """
@@ -66,7 +73,7 @@ css = """
 
 with gr.Blocks(css=css) as demo:
     with gr.Column(elem_id="col-container"):
-        gr.Markdown(" # Text-to-Image Gradio Template")
+        gr.Markdown(" # Pikigen Text-to-Image Demo")
 
         with gr.Row():
             prompt = gr.Text(
@@ -86,7 +93,7 @@ with gr.Blocks(css=css) as demo:
                 label="Negative prompt",
                 max_lines=1,
                 placeholder="Enter a negative prompt",
-                visible=False,
+                visible=True,
             )
 
             seed = gr.Slider(
@@ -105,7 +112,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=256,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024,  # Replace with defaults that work for your model
+                    value=1024,
                 )
 
                 height = gr.Slider(
@@ -113,7 +120,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=256,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024,  # Replace with defaults that work for your model
+                    value=1024,
                 )
 
             with gr.Row():
@@ -122,7 +129,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=0.0,
                     maximum=10.0,
                     step=0.1,
-                    value=0.0,  # Replace with defaults that work for your model
+                    value=3.5,
                 )
 
                 num_inference_steps = gr.Slider(
@@ -130,7 +137,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=1,
                     maximum=50,
                     step=1,
-                    value=2,  # Replace with defaults that work for your model
+                    value=30,
                 )
 
         gr.Examples(examples=examples, inputs=[prompt])

pikigen/__init__.py

@@ -0,0 +1,5 @@
+from .pipeline import PikigenPipeline, PikigenPipelineOutput, APGConfig
+from .model import DiT
+
+
+__all__ = ["PikigenPipeline", "PikigenPipelineOutput", "APGConfig", "DiT"]

pikigen/model.py

@@ -0,0 +1,455 @@
+# DiT with cross attention
+
+import math
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+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.accelerate_utils import apply_forward_hook
+from einops import rearrange
+from peft import get_peft_model_state_dict, set_peft_model_state_dict
+from torch import nn
+
+
+def timestep_embedding(t, dim, max_period=10000):
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
+        device=t.device
+    )
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+
+    return embedding
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-6, trainable=False):
+        super().__init__()
+        self.eps = eps
+        if trainable:
+            self.weight = nn.Parameter(torch.ones(dim))
+        else:
+            self.weight = None
+
+    def forward(self, x):
+        x_dtype = x.dtype
+        x = x.float()
+        norm = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+        if self.weight is not None:
+            return (x * norm * self.weight).to(dtype=x_dtype)
+        else:
+            return (x * norm).to(dtype=x_dtype)
+
+
+class QKNorm(nn.Module):
+    """Normalizing the query and the key independently, as Flux proposes"""
+
+    def __init__(self, dim, trainable=False):
+        super().__init__()
+        self.query_norm = RMSNorm(dim, trainable=trainable)
+        self.key_norm = RMSNorm(dim, trainable=trainable)
+
+    def forward(self, q, k):
+        q = self.query_norm(q)
+        k = self.key_norm(k)
+        return q, k
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        qkv_bias=False,
+        is_self_attn=True,
+        cross_attn_input_size=None,
+        residual_v=False,
+        dynamic_softmax_temperature=False,
+    ):
+        super().__init__()
+        assert dim % num_heads == 0
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim**-0.5
+        self.is_self_attn = is_self_attn
+        self.residual_v = residual_v
+        self.dynamic_softmax_temperature = dynamic_softmax_temperature
+
+        if is_self_attn:
+            self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        else:
+            self.q = nn.Linear(dim, dim, bias=qkv_bias)
+            self.context_kv = nn.Linear(cross_attn_input_size, dim * 2, bias=qkv_bias)
+
+        self.proj = nn.Linear(dim, dim, bias=False)
+
+        if residual_v:
+            self.lambda_param = nn.Parameter(torch.tensor(0.5).reshape(1))
+
+        self.qk_norm = QKNorm(self.head_dim)
+
+    def forward(self, x, context=None, v_0=None, rope=None):
+        if self.is_self_attn:
+            qkv = self.qkv(x)
+            qkv = rearrange(qkv, "b l (k h d) -> k b h l d", k=3, h=self.num_heads)
+            q, k, v = qkv.unbind(0)
+
+            if self.residual_v and v_0 is not None:
+                v = self.lambda_param * v + (1 - self.lambda_param) * v_0
+
+            if rope is not None:
+                # print(q.shape, rope[0].shape, rope[1].shape)
+                q = apply_rotary_emb(q, rope[0], rope[1])
+                k = apply_rotary_emb(k, rope[0], rope[1])
+
+                # https://arxiv.org/abs/2306.08645
+                # https://arxiv.org/abs/2410.01104
+                # ratioonale is that if tokens get larger, categorical distribution get more uniform
+                # so you want to enlargen entropy.
+
+                token_length = q.shape[2]
+                if self.dynamic_softmax_temperature:
+                    ratio = math.sqrt(math.log(token_length) / math.log(1040.0))  # 1024 + 16
+                    k = k * ratio
+            q, k = self.qk_norm(q, k)
+
+        else:
+            q = rearrange(self.q(x), "b l (h d) -> b h l d", h=self.num_heads)
+            kv = rearrange(
+                self.context_kv(context),
+                "b l (k h d) -> k b h l d",
+                k=2,
+                h=self.num_heads,
+            )
+            k, v = kv.unbind(0)
+            q, k = self.qk_norm(q, k)
+
+        x = F.scaled_dot_product_attention(q, k, v)
+        x = rearrange(x, "b h l d -> b l (h d)")
+        x = self.proj(x)
+        return x, v if self.is_self_attn else None
+
+
+class DiTBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        cross_attn_input_size,
+        num_heads,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        residual_v=False,
+        dynamic_softmax_temperature=False,
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.norm1 = RMSNorm(hidden_size, trainable=qkv_bias)
+        self.self_attn = Attention(
+            hidden_size,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            is_self_attn=True,
+            residual_v=residual_v,
+            dynamic_softmax_temperature=dynamic_softmax_temperature,
+        )
+
+        if cross_attn_input_size is not None:
+            self.norm2 = RMSNorm(hidden_size, trainable=qkv_bias)
+            self.cross_attn = Attention(
+                hidden_size,
+                num_heads=num_heads,
+                qkv_bias=qkv_bias,
+                is_self_attn=False,
+                cross_attn_input_size=cross_attn_input_size,
+                dynamic_softmax_temperature=dynamic_softmax_temperature,
+            )
+        else:
+            self.norm2 = None
+            self.cross_attn = None
+
+        self.norm3 = RMSNorm(hidden_size, trainable=qkv_bias)
+        mlp_hidden = int(hidden_size * mlp_ratio)
+        self.mlp = nn.Sequential(
+            nn.Linear(hidden_size, mlp_hidden),
+            nn.GELU(),
+            nn.Linear(mlp_hidden, hidden_size),
+        )
+
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 9 * hidden_size, bias=True))
+
+        self.adaLN_modulation[-1].weight.data.zero_()
+        self.adaLN_modulation[-1].bias.data.zero_()
+
+    # @torch.compile(mode='reduce-overhead')
+    def forward(self, x, context, c, v_0=None, rope=None):
+        (
+            shift_sa,
+            scale_sa,
+            gate_sa,
+            shift_ca,
+            scale_ca,
+            gate_ca,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+        ) = self.adaLN_modulation(c).chunk(9, dim=1)
+
+        scale_sa = scale_sa[:, None, :]
+        scale_ca = scale_ca[:, None, :]
+        scale_mlp = scale_mlp[:, None, :]
+
+        shift_sa = shift_sa[:, None, :]
+        shift_ca = shift_ca[:, None, :]
+        shift_mlp = shift_mlp[:, None, :]
+
+        gate_sa = gate_sa[:, None, :]
+        gate_ca = gate_ca[:, None, :]
+        gate_mlp = gate_mlp[:, None, :]
+
+        norm_x = self.norm1(x.clone())
+        norm_x = norm_x * (1 + scale_sa) + shift_sa
+        attn_out, v = self.self_attn(norm_x, v_0=v_0, rope=rope)
+        x = x + attn_out * gate_sa
+
+        if self.norm2 is not None:
+            norm_x = self.norm2(x)
+            norm_x = norm_x * (1 + scale_ca) + shift_ca
+            x = x + self.cross_attn(norm_x, context)[0] * gate_ca
+
+        norm_x = self.norm3(x)
+        norm_x = norm_x * (1 + scale_mlp) + shift_mlp
+        x = x + self.mlp(norm_x) * gate_mlp
+
+        return x, v
+
+
+class PatchEmbed(nn.Module):
+    def __init__(self, patch_size=16, in_channels=3, embed_dim=768):
+        super().__init__()
+        self.patch_proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
+        self.patch_size = patch_size
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        x = self.patch_proj(x)
+        x = rearrange(x, "b c h w -> b (h w) c")
+        return x
+
+
+class TwoDimRotary(torch.nn.Module):
+    def __init__(self, dim, base=10000, h=256, w=256):
+        super().__init__()
+        self.inv_freq = torch.FloatTensor([1.0 / (base ** (i / dim)) for i in range(0, dim, 2)])
+        self.h = h
+        self.w = w
+
+        t_h = torch.arange(h, dtype=torch.float32)
+        t_w = torch.arange(w, dtype=torch.float32)
+
+        freqs_h = torch.outer(t_h, self.inv_freq).unsqueeze(1)  # h, 1, d / 2
+        freqs_w = torch.outer(t_w, self.inv_freq).unsqueeze(0)  # 1, w, d / 2
+        freqs_h = freqs_h.repeat(1, w, 1)  # h, w, d / 2
+        freqs_w = freqs_w.repeat(h, 1, 1)  # h, w, d / 2
+        freqs_hw = torch.cat([freqs_h, freqs_w], 2)  # h, w, d
+
+        self.register_buffer("freqs_hw_cos", freqs_hw.cos())
+        self.register_buffer("freqs_hw_sin", freqs_hw.sin())
+
+    def forward(self, x, height_width=None, extend_with_register_tokens=0):
+        if height_width is not None:
+            this_h, this_w = height_width
+        else:
+            this_hw = x.shape[1]
+            this_h, this_w = int(this_hw**0.5), int(this_hw**0.5)
+
+        cos = self.freqs_hw_cos[0 : this_h, 0 : this_w]
+        sin = self.freqs_hw_sin[0 : this_h, 0 : this_w]
+
+        cos = cos.clone().reshape(this_h * this_w, -1)
+        sin = sin.clone().reshape(this_h * this_w, -1)
+
+        # append N of zero-attn tokens
+        if extend_with_register_tokens > 0:
+            cos = torch.cat(
+                [
+                    torch.ones(extend_with_register_tokens, cos.shape[1]).to(cos.device),
+                    cos,
+                ],
+                0,
+            )
+            sin = torch.cat(
+                [
+                    torch.zeros(extend_with_register_tokens, sin.shape[1]).to(sin.device),
+                    sin,
+                ],
+                0,
+            )
+
+        return cos[None, None, :, :], sin[None, None, :, :]  # [1, 1, T + N, Attn-dim]
+
+
+def apply_rotary_emb(x, cos, sin):
+    orig_dtype = x.dtype
+    x = x.to(dtype=torch.float32)
+    assert x.ndim == 4  # multihead attention
+    d = x.shape[3] // 2
+    x1 = x[..., :d]
+    x2 = x[..., d:]
+    y1 = x1 * cos + x2 * sin
+    y2 = x1 * (-sin) + x2 * cos
+    return torch.cat([y1, y2], 3).to(dtype=orig_dtype)
+
+
+class DiT(ModelMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapterMixin):  # type: ignore[misc]
+    @register_to_config
+    def __init__(
+        self,
+        in_channels=4,
+        patch_size=2,
+        hidden_size=1152,
+        depth=28,
+        num_heads=16,
+        mlp_ratio=4.0,
+        cross_attn_input_size=128,
+        residual_v=False,
+        train_bias_and_rms=True,
+        use_rope=True,
+        gradient_checkpoint=False,
+        dynamic_softmax_temperature=False,
+        rope_base=10000,
+    ):
+        super().__init__()
+
+        self.patch_embed = PatchEmbed(patch_size, in_channels, hidden_size)
+
+        if use_rope:
+            self.rope = TwoDimRotary(hidden_size // (2 * num_heads), base=rope_base, h=512, w=512)
+        else:
+            self.positional_embedding = nn.Parameter(torch.zeros(1, 2048, hidden_size))
+
+        self.register_tokens = nn.Parameter(torch.randn(1, 16, hidden_size))
+
+        self.time_embed = nn.Sequential(
+            nn.Linear(hidden_size, 4 * hidden_size),
+            nn.SiLU(),
+            nn.Linear(4 * hidden_size, hidden_size),
+        )
+
+        self.blocks = nn.ModuleList(
+            [
+                DiTBlock(
+                    hidden_size=hidden_size,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    cross_attn_input_size=cross_attn_input_size,
+                    residual_v=residual_v,
+                    qkv_bias=train_bias_and_rms,
+                    dynamic_softmax_temperature=dynamic_softmax_temperature,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.final_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
+
+        self.final_norm = RMSNorm(hidden_size, trainable=train_bias_and_rms)
+        self.final_proj = nn.Linear(hidden_size, patch_size * patch_size * in_channels)
+        nn.init.zeros_(self.final_modulation[-1].weight)
+        nn.init.zeros_(self.final_modulation[-1].bias)
+        nn.init.zeros_(self.final_proj.weight)
+        nn.init.zeros_(self.final_proj.bias)
+        self.paramstatus = {}
+        for n, p in self.named_parameters():
+            self.paramstatus[n] = {
+                "shape": p.shape,
+                "requires_grad": p.requires_grad,
+            }
+
+    def save_lora_weights(self, save_directory):
+        """Save LoRA weights to a file"""
+        lora_state_dict = get_peft_model_state_dict(self)
+        torch.save(lora_state_dict, f"{save_directory}/lora_weights.pt")
+
+    def load_lora_weights(self, load_directory):
+        """Load LoRA weights from a file"""
+        lora_state_dict = torch.load(f"{load_directory}/lora_weights.pt")
+        set_peft_model_state_dict(self, lora_state_dict)
+
+    @apply_forward_hook
+    def forward(self, x, context, timesteps):
+        b, c, h, w = x.shape
+        x = self.patch_embed(x)  # b, T, d
+
+        x = torch.cat([self.register_tokens.repeat(b, 1, 1), x], 1)  # b, T + N, d
+
+        if self.config.use_rope:
+            cos, sin = self.rope(
+                x,
+                extend_with_register_tokens=16,
+                height_width=(h // self.config.patch_size, w // self.config.patch_size),
+            )
+        else:
+            x = x + self.positional_embedding.repeat(b, 1, 1)[:, : x.shape[1], :]
+            cos, sin = None, None
+
+        t_emb = timestep_embedding(timesteps * 1000, self.config.hidden_size).to(x.device, dtype=x.dtype)
+        t_emb = self.time_embed(t_emb)
+
+        v_0 = None
+
+        for _idx, block in enumerate(self.blocks):
+            if self.config.gradient_checkpoint:
+                x, v = torch.utils.checkpoint.checkpoint(
+                    block,
+                    x,
+                    context,
+                    t_emb,
+                    v_0,
+                    (cos, sin),
+                    use_reentrant=True,
+                )
+            else:
+                x, v = block(x, context, t_emb, v_0, (cos, sin))
+            if v_0 is None:
+                v_0 = v
+
+        x = x[:, 16:, :]
+        final_shift, final_scale = self.final_modulation(t_emb).chunk(2, dim=1)
+        x = self.final_norm(x)
+        x = x * (1 + final_scale[:, None, :]) + final_shift[:, None, :]
+        x = self.final_proj(x)
+
+        x = rearrange(
+            x,
+            "b (h w) (p1 p2 c) -> b c (h p1) (w p2)",
+            h=h // self.config.patch_size,
+            w=w // self.config.patch_size,
+            p1=self.config.patch_size,
+            p2=self.config.patch_size,
+        )
+        return x
+
+
+if __name__ == "__main__":
+    model = DiT(
+        in_channels=4,
+        patch_size=2,
+        hidden_size=1152,
+        depth=28,
+        num_heads=16,
+        mlp_ratio=4.0,
+        cross_attn_input_size=128,
+        residual_v=False,
+        train_bias_and_rms=True,
+        use_rope=True,
+    ).cuda()
+    print(
+        model(
+            torch.randn(1, 4, 64, 64).cuda(),
+            torch.randn(1, 37, 128).cuda(),
+            torch.tensor([1.0]).cuda(),
+        )
+    )

pikigen/pipeline.py

@@ -0,0 +1,296 @@
+import logging
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from diffusers import AutoencoderKL, DiffusionPipeline
+from diffusers.utils import BaseOutput
+from diffusers.utils.torch_utils import randn_tensor
+from PIL import Image
+from torch import FloatTensor
+from tqdm.auto import tqdm
+from transformers import T5EncoderModel, T5TokenizerFast
+
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class APGConfig:
+    """APG (Augmented Parallel Guidance) configuration"""
+
+    enabled: bool = True
+    orthogonal_threshold: float = 0.03
+
+
+@dataclass
+class PikigenPipelineOutput(BaseOutput):
+    """
+    Output class for PikigenPipeline pipeline.
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+    """
+
+    images: Union[List[Image.Image], np.ndarray]
+
+
+class PikigenPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Pikigen model.
+    This model inherits from [`DiffusionPipeline`].
+    """
+
+    model_cpu_offload_seq = "text_encoder->dit_model->vae" 
+
+    dit_model: torch.nn.Module
+    vae: AutoencoderKL
+    text_encoder: T5EncoderModel
+    tokenizer: T5TokenizerFast
+    _progress_bar_config: Dict[str, Any]
+
+    def __init__(
+        self, dit_model: torch.nn.Module, vae: AutoencoderKL, text_encoder: T5EncoderModel, tokenizer: T5TokenizerFast
+    ):
+        super().__init__()
+        # Register all modules for the pipeline
+        # Access DiffusionPipeline's register_modules directly to avoid mypy error
+        DiffusionPipeline.register_modules(
+            self, dit_model=dit_model, vae=vae, text_encoder=text_encoder, tokenizer=tokenizer
+        )
+
+        # Move models to channels last for better performance
+        # AutoencoderKL inherits from torch.nn.Module which has these methods
+        if hasattr(self.vae, "to"):
+            self.vae.to(memory_format=torch.channels_last)
+        if hasattr(self.vae, "requires_grad_"):
+            self.vae.requires_grad_(False)
+        if hasattr(self.text_encoder, "requires_grad_"):
+            self.text_encoder.requires_grad_(False)
+
+        # Constants
+        self.vae_scale_factor = 8
+        self.return_index = -8  # T5 hidden state index to use
+
+    def enable_vae_slicing(self):
+        """Enable VAE slicing for memory efficiency."""
+        if hasattr(self.vae, "enable_slicing"):
+            self.vae.enable_slicing()
+
+    def enable_vae_tiling(self):
+        """Enable VAE tiling for memory efficiency."""
+        if hasattr(self.vae, "enable_tiling"):
+            self.vae.enable_tiling()
+
+    def set_progress_bar_config(self, **kwargs):
+        """Set progress bar configuration."""
+        self._progress_bar_config = kwargs
+
+    def progress_bar(self, iterable=None, **kwargs):
+        """Create progress bar for iterations."""
+        self._progress_bar_config = getattr(self, "_progress_bar_config", None) or {}
+        config = {**self._progress_bar_config, **kwargs}
+        return tqdm(iterable, **config)
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 512,
+        return_index: int = -8,
+    ) -> Tuple[FloatTensor, FloatTensor]:
+        """Encodes the prompt and negative prompt."""
+        if isinstance(prompt, str):
+            prompt = [prompt]
+        device = device or self.text_encoder.device
+        # Text encoder forward pass
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids.to(device)
+        prompt_embeds = self.text_encoder(text_input_ids, return_dict=True, output_hidden_states=True)
+        prompt_embeds_tensor = prompt_embeds.hidden_states[return_index]
+        if return_index != -1:
+            prompt_embeds_tensor = self.text_encoder.encoder.final_layer_norm(prompt_embeds_tensor)
+            prompt_embeds_tensor = self.text_encoder.encoder.dropout(prompt_embeds_tensor)
+
+        dtype = dtype or next(self.text_encoder.parameters()).dtype
+        prompt_embeds_tensor = prompt_embeds_tensor.to(dtype=dtype, device=device)
+
+        # Handle negative prompts
+        if negative_prompt is None:
+            negative_embeds = torch.zeros_like(prompt_embeds_tensor)
+        else:
+            if isinstance(negative_prompt, str):
+                negative_prompt = [negative_prompt]
+            negative_result = self.encode_prompt(
+                prompt=negative_prompt, device=device, dtype=dtype, return_index=return_index
+            )
+            negative_embeds = negative_result[0]
+
+        # Explicitly cast both tensors to FloatTensor for mypy
+        from typing import cast
+
+        prompt_tensor = cast(FloatTensor, prompt_embeds_tensor.to(dtype=dtype))
+        negative_tensor = cast(FloatTensor, negative_embeds.to(dtype=dtype))
+        return (prompt_tensor, negative_tensor)
+
+    def to(self, torch_device=None, torch_dtype=None, silence_dtype_warnings=False):
+        """Move pipeline components to specified device and dtype."""
+        if hasattr(self, "vae"):
+            self.vae.to(device=torch_device, dtype=torch_dtype)
+        if hasattr(self, "text_encoder"):
+            self.text_encoder.to(device=torch_device, dtype=torch_dtype)
+        if hasattr(self, "dit_model"):
+            self.dit_model.to(device=torch_device, dtype=torch_dtype)
+        return self
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: Optional[int] = 1024,
+        width: Optional[int] = 1024,
+        num_inference_steps: int = 30,
+        guidance_scale: float = 3.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        dtype: Optional[torch.dtype] = None,
+        alpha: Optional[float] = None,
+        apg_config: Optional[APGConfig] = None,
+        **kwargs,
+    ):
+        """Generate images from text prompt."""
+        batch_size = 1  # TODO: Make this method support batch generation
+        # Ensure height and width are not None for calculation
+        if height is None:
+            height = 1024
+        if width is None:
+            width = 1024
+
+        dtype = dtype or next(self.dit_model.parameters()).dtype
+        apg_config = apg_config or APGConfig()
+
+        device = self._execution_device
+
+        # 2. Encode prompts
+        prompt_embeds, negative_embeds = self.encode_prompt(
+            prompt=prompt, negative_prompt=negative_prompt, device=self.text_encoder.device, dtype=dtype
+        )
+
+        # 3. Initialize latents
+        latent_height = height // self.vae_scale_factor
+        latent_width = width // self.vae_scale_factor
+
+        if isinstance(generator, list):
+            if len(generator) != batch_size:
+                raise ValueError(f"Got {len(generator)} generators for {batch_size} samples")
+
+        latents = randn_tensor((batch_size, 16, latent_height, latent_width), generator=generator, device=device, dtype=dtype)
+        acc_latents = latents.clone()
+
+        # 4. Calculate alpha if not provided
+        if alpha is None:
+            image_token_size = latent_height * latent_width
+            alpha = 2 * math.sqrt(image_token_size / (64 * 64))
+
+        # 6. Sampling loop
+        self.dit_model.eval()
+        
+        # Check if guidance is needed
+        do_classifier_free_guidance = guidance_scale >= 1.0
+        
+        for i in self.progress_bar(range(num_inference_steps, 0, -1)):
+            # Calculate timesteps
+            t = i / num_inference_steps
+            t_next = (i - 1) / num_inference_steps
+            # Scale timesteps according to alpha
+            t = t * alpha / (1 + (alpha - 1) * t)
+            t_next = t_next * alpha / (1 + (alpha - 1) * t_next)
+            dt = t - t_next
+            
+            # Create tensor with proper device
+            t_tensor = torch.tensor([t] * batch_size, device=device, dtype=dtype)
+            
+            if do_classifier_free_guidance:
+                # Duplicate latents for both conditional and unconditional inputs
+                latents_input = torch.cat([latents] * 2)
+                # Concatenate negative and positive prompt embeddings
+                context_input = torch.cat([negative_embeds, prompt_embeds])
+                # Duplicate timesteps for the batch
+                t_input = torch.cat([t_tensor] * 2)
+                
+                # Get model predictions in a single pass
+                model_outputs = self.dit_model(latents_input, context_input, t_input)
+                
+                # Split outputs back into unconditional and conditional predictions
+                uncond_output, cond_output = model_outputs.chunk(2)
+                
+                if apg_config.enabled:
+                    # Augmented Parallel Guidance
+                    dy = cond_output
+                    dd = cond_output - uncond_output
+                    # Find parallel direction
+                    parallel_direction = (dy * dd).sum() / (dy * dy).sum() * dy
+                    orthogonal_direction = dd - parallel_direction
+                    # Scale orthogonal component
+                    orthogonal_std = orthogonal_direction.std()
+                    orthogonal_scale = min(1, apg_config.orthogonal_threshold / orthogonal_std)
+                    orthogonal_direction = orthogonal_direction * orthogonal_scale
+                    model_output = dy + (guidance_scale - 1) * orthogonal_direction
+                else:
+                    # Standard classifier-free guidance
+                    model_output = uncond_output + guidance_scale * (cond_output - uncond_output)
+            else:
+                # If no guidance needed, just run the model normally
+                model_output = self.dit_model(latents, prompt_embeds, t_tensor)
+            
+            # Update latents
+            acc_latents = acc_latents + dt * model_output.to(device)
+            latents = acc_latents.clone()
+
+        # 7. Decode latents
+        # These checks handle the case where mypy doesn't recognize these attributes
+        scaling_factor = getattr(self.vae.config, "scaling_factor", 0.18215) if hasattr(self.vae, "config") else 0.18215
+        shift_factor = getattr(self.vae.config, "shift_factor", 0) if hasattr(self.vae, "config") else 0
+
+        latents = latents / scaling_factor + shift_factor
+
+        vae_dtype = self.vae.dtype if hasattr(self.vae, "dtype") else dtype
+        decoded_images = self.vae.decode(latents.to(vae_dtype)).sample if hasattr(self.vae, "decode") else latents
+
+        # Offload all models
+        try:
+            self.maybe_free_model_hooks()
+        except AttributeError as e:
+            if "OptimizedModule" in str(e):
+                import warnings
+                warnings.warn(
+                    "Encountered 'OptimizedModule' error when offloading models. "
+                    "This issue might be fixed in the future by: "
+                    "https://github.com/huggingface/diffusers/pull/10730"
+                )
+            else:
+                raise
+            
+        # 8. Post-process images
+        images = (decoded_images / 2 + 0.5).clamp(0, 1)
+        # Convert to PIL Images
+        images = (images * 255).round().clamp(0, 255).to(torch.uint8).cpu()
+        pil_images = [Image.fromarray(img.permute(1, 2, 0).numpy()) for img in images]
+
+        return PikigenPipelineOutput(
+            images=pil_images,
+        )
+
+

requirements.txt

@@ -1,6 +1,9 @@
-accelerate
 diffusers
-invisible_watermark
+einops
+jsonargparse
+peft
+protobuf
+rich
+sentencepiece
 torch
-transformers
-xformers
+transformers