upload file

README.md

@@ -1,8 +1,8 @@
 ---
 title: DeepGen Test
-emoji: 🐠
-colorFrom: yellow
-colorTo: yellow
+emoji: 📚
+colorFrom: red
+colorTo: gray
 sdk: gradio
 sdk_version: 6.6.0
 app_file: app.py

app.py

@@ -0,0 +1,479 @@
+import spaces
+import os
+import sys 
+import subprocess
+import importlib.util
+
+# --- CÀI ĐẶT ÉP BUỘC XTUNER (Bỏ qua kiểm tra xung đột) ---
+if importlib.util.find_spec("xtuner") is None:
+    print("Đang cài đặt xtuner bằng lệnh ngầm...")
+    subprocess.check_call([sys.executable, "-m", "pip", "install", "xtuner==0.2.0", "--no-deps"])
+
+import torch
+import torch.utils._pytree as _torch_pytree
+
+# 1. Vá lỗi PyTree cho Transformers mới
+def smart_pytree_patch():
+    orig_func = getattr(_torch_pytree, '_register_pytree_node', None)
+    if orig_func:
+        def patched_register(cls, to_iter, from_iter, serialized_type_name=None):
+            return orig_func(cls, to_iter, from_iter)
+        _torch_pytree.register_pytree_node = patched_register
+        _torch_pytree._register_pytree_node = patched_register
+
+smart_pytree_patch()
+
+import torch.distributed
+
+# 2. Vá lỗi Torch XPU
+if not hasattr(torch, 'xpu'):
+    class DummyXPU:
+        @staticmethod
+        def is_available(): return False
+        @staticmethod
+        def empty_cache(): pass
+        @staticmethod
+        def device_count(): return 0
+        @staticmethod
+        def current_device(): return 0
+        @staticmethod
+        def get_device_name(device=None): return "DummyXPU"
+        @staticmethod
+        def is_bf16_supported(): return False
+        @staticmethod
+        def synchronize(device=None): pass
+        @staticmethod
+        def set_device(device): pass
+        @staticmethod
+        def manual_seed(seed): pass
+        @staticmethod
+        def manual_seed_all(seed): pass
+        @staticmethod
+        def seed(): pass
+        @staticmethod
+        def seed_all(): pass
+        @staticmethod
+        def initial_seed(): return 0
+    torch.xpu = DummyXPU()
+
+# 3. Vá lỗi Device Mesh
+if not hasattr(torch.distributed, 'device_mesh'):
+    class DummyDeviceMesh: pass
+    class DummyDeviceMeshModule: DeviceMesh = DummyDeviceMesh
+    dummy_module = DummyDeviceMeshModule()
+    sys.modules['torch.distributed.device_mesh'] = dummy_module
+    torch.distributed.device_mesh = dummy_module
+
+# 4. Vá lỗi pad_sequence() không hỗ trợ padding_side (cần PyTorch >= 2.5)
+# ---------------------------------------------------------------
+import torch.nn.utils.rnn as _rnn
+_orig_pad_sequence = _rnn.pad_sequence
+
+def _patched_pad_sequence(sequences, batch_first=False, padding_value=0.0, padding_side='right'):
+    if padding_side == 'left':
+        sequences = [seq.flip(0) for seq in sequences]
+        padded = _orig_pad_sequence(sequences, batch_first=batch_first, padding_value=padding_value)
+        flip_dim = 1 if batch_first else 0
+        return padded.flip(flip_dim)
+    else:
+        return _orig_pad_sequence(sequences, batch_first=batch_first, padding_value=padding_value)
+
+_rnn.pad_sequence = _patched_pad_sequence
+torch.nn.utils.rnn.pad_sequence = _patched_pad_sequence
+
+import importlib, sys as _sys
+def _patch_pad_sequence_in_module(module_name):
+    mod = _sys.modules.get(module_name)
+    if mod and hasattr(mod, 'pad_sequence'):
+        mod.pad_sequence = _patched_pad_sequence
+
+import builtins as _builtins
+_orig_builtins_import = _builtins.__import__
+
+def _import_hook(name, *args, **kwargs):
+    module = _orig_builtins_import(name, *args, **kwargs)
+    for mod_name, mod in list(_sys.modules.items()):
+        if mod_name.startswith('src.') and hasattr(mod, 'pad_sequence'):
+            mod.pad_sequence = _patched_pad_sequence
+    return module
+
+_builtins.__import__ = _import_hook
+print("✅ Đã patch pad_sequence() để hỗ trợ padding_side")
+
+# 5. Vá lỗi F.interpolate bilinear nhận 3D input thay vì 4D
+# ---------------------------------------------------------------
+import torch.nn.functional as _F
+_orig_interpolate = _F.interpolate
+
+def _patched_interpolate(input, size=None, scale_factor=None, mode='nearest',
+                          align_corners=None, recompute_scale_factor=None, antialias=False):
+    squeezed = False
+    if mode in ('bilinear', 'bicubic') and input.dim() == 3:
+        input = input.unsqueeze(0)
+        squeezed = True
+    result = _orig_interpolate(
+        input, size=size, scale_factor=scale_factor, mode=mode,
+        align_corners=align_corners, recompute_scale_factor=recompute_scale_factor,
+        antialias=antialias
+    )
+    if squeezed:
+        result = result.squeeze(0)
+    return result
+
+_F.interpolate = _patched_interpolate
+torch.nn.functional.interpolate = _patched_interpolate
+print("✅ Đã patch F.interpolate() để hỗ trợ 3D input với bilinear mode")
+
+# 6. Vá lỗi vae.encode() / vae.decode() nhận 3D input thay vì 4D
+# ---------------------------------------------------------------
+try:
+    from diffusers.models.autoencoders.autoencoder_kl import AutoencoderKL as _AutoencoderKL
+    _orig_vae_encode = _AutoencoderKL.encode
+    _orig_vae_decode = _AutoencoderKL.decode
+
+    def _patched_vae_encode(self, x, *args, **kwargs):
+        if x.dim() == 3:
+            x = x.unsqueeze(0)
+        if x.dim() == 4 and x.shape[1] == 1:
+            x = x.repeat(1, 3, 1, 1)
+        elif x.dim() == 4 and x.shape[1] == 4:
+            x = x[:, :3, :, :]
+        result = _orig_vae_encode(self, x, *args, **kwargs)
+        return result
+
+    def _patched_vae_decode(self, z, *args, **kwargs):
+        if z.dim() == 3:
+            z = z.unsqueeze(0)
+        return _orig_vae_decode(self, z, *args, **kwargs)
+
+    _AutoencoderKL.encode = _patched_vae_encode
+    _AutoencoderKL.decode = _patched_vae_decode
+    print("✅ Đã patch AutoencoderKL.encode/decode() để hỗ trợ 3D input")
+except Exception as _e:
+    print(f"⚠️ Không thể patch AutoencoderKL: {_e}")
+
+# 7. Vá lỗi dtype string + flash_attention_2
+# ---------------------------------------------------------------
+STRING_TO_TORCH_DTYPE = {
+    "float16": torch.float16,
+    "float32": torch.float32,
+    "float64": torch.float64,
+    "bfloat16": torch.bfloat16,
+    "torch.bfloat16": torch.bfloat16,
+    "half": torch.float16,
+    "int8": torch.int8,
+}
+
+def str_to_torch_dtype(dtype):
+    if isinstance(dtype, str):
+        return STRING_TO_TORCH_DTYPE.get(dtype.lower(), torch.float32)
+    return dtype
+
+try:
+    from transformers import PretrainedConfig
+    original_pretrained_init = PretrainedConfig.__init__
+
+    def patched_pretrained_config_init(self, *args, **kwargs):
+        if 'torch_dtype' in kwargs and isinstance(kwargs['torch_dtype'], str):
+            kwargs['torch_dtype'] = str_to_torch_dtype(kwargs['torch_dtype'])
+        if kwargs.get('attn_implementation') == 'flash_attention_2':
+            kwargs['attn_implementation'] = 'sdpa'
+        original_pretrained_init(self, *args, **kwargs)
+        if hasattr(self, 'torch_dtype') and isinstance(self.torch_dtype, str):
+            self.torch_dtype = str_to_torch_dtype(self.torch_dtype)
+        if getattr(self, '_attn_implementation', None) == 'flash_attention_2':
+            self._attn_implementation = 'sdpa'
+        if getattr(self, '_attn_implementation_internal', None) == 'flash_attention_2':
+            self._attn_implementation_internal = 'sdpa'
+
+    PretrainedConfig.__init__ = patched_pretrained_config_init
+    print("✅ Đã patch PretrainedConfig.__init__")
+except Exception as e:
+    print(f"⚠️ Không thể patch PretrainedConfig: {e}")
+
+orig_is_floating_point = torch.is_floating_point
+def patched_is_floating_point(obj):
+    if isinstance(obj, str):
+        return obj.lower() in ["bfloat16", "float16", "float32", "float64", "half"]
+    return orig_is_floating_point(obj)
+torch.is_floating_point = patched_is_floating_point
+
+# ============================================================
+# IMPORTS CHÍNH
+# ============================================================
+import time
+import psutil
+import numpy as np
+import gradio as gr
+from PIL import Image
+from einops import rearrange
+from huggingface_hub import hf_hub_download
+
+from xtuner.registry import BUILDER
+from mmengine.config import Config
+
+# ============================================================
+# CONFIG PATCHES
+# ============================================================
+LOCAL_TO_HF_PATH = {
+    "model_zoo/Qwen2.5-VL-3B-Instruct":   "Qwen/Qwen2.5-VL-3B-Instruct",
+    "model_zoo/UniPic2-SD3.5M-Kontext-2B": "Skywork/UniPic2-SD3.5M-Kontext-2B",
+}
+
+def patch_config_paths(cfg):
+    cfg_text = cfg.pretty_text
+    changed = False
+    for local_path, hf_path in LOCAL_TO_HF_PATH.items():
+        if local_path in cfg_text:
+            cfg_text = cfg_text.replace(local_path, hf_path)
+            print(f"  → Đã thay path: '{local_path}' → '{hf_path}'")
+            changed = True
+    if changed:
+        return Config.fromstring(cfg_text, file_format='.py')
+    return cfg
+
+def patch_config_dtype(cfg):
+    if isinstance(cfg, dict):
+        for key in list(cfg.keys()):
+            val = cfg[key]
+            if key in ('torch_dtype', 'dtype', 'param_dtype', 'compute_dtype') and isinstance(val, str):
+                cfg[key] = str_to_torch_dtype(val)
+                print(f"  → Convert dtype cfg['{key}'] = '{val}' → {cfg[key]}")
+            elif key == 'attn_implementation' and val == 'flash_attention_2':
+                cfg[key] = 'sdpa'
+                print(f"  → Thay attn_implementation: flash_attention_2 → sdpa")
+            else:
+                patch_config_dtype(val)
+    elif isinstance(cfg, (list, tuple)):
+        for item in cfg:
+            patch_config_dtype(item)
+    elif hasattr(cfg, '_cfg_dict'):
+        patch_config_dtype(cfg._cfg_dict)
+    return cfg
+
+from xtuner.model.utils import guess_load_checkpoint
+
+REPO_ID = "oedevs/DeepGen-1.0"
+MODEL_WEIGHTS = {
+    "Pretrain (Alignment)": "iter_200000.pth",
+    "RL with MR-GRPO (Tốt nhất)": "model.pt"
+}
+
+current_loaded_method = None
+model = None
+
+def load_model(method_name):
+    global current_loaded_method, model
+    if current_loaded_method == method_name and model is not None:
+        return model
+    print(f"Đang chuẩn bị tải và nạp model: {method_name}...")
+    filename = MODEL_WEIGHTS[method_name]
+    weight_path = hf_hub_download(repo_id=REPO_ID, filename=filename)
+    config = Config.fromfile("configs/models/deepgen_scb.py")
+    print("Đang patch đường dẫn config...")
+    config = patch_config_paths(config)
+    print("Đang patch dtype và attn_implementation...")
+    patch_config_dtype(config)
+    new_model = BUILDER.build(config.model)
+    if weight_path.endswith('.pt'):
+        state_dict = torch.load(weight_path, map_location="cpu")
+    else:
+        state_dict = guess_load_checkpoint(weight_path)
+    new_model.load_state_dict(state_dict, strict=False)
+    model_dtype = new_model.dtype
+    if isinstance(model_dtype, str):
+        model_dtype = str_to_torch_dtype(model_dtype)
+    new_model = new_model.to(model_dtype)
+    new_model.eval()
+    model = new_model
+    current_loaded_method = method_name
+    return model
+
+def _process_image(image):
+    image = image.convert('RGB')
+    image = image.resize(size=(512, 512))
+    pixel_values = torch.from_numpy(np.array(image)).float()
+    pixel_values = pixel_values / 255
+    pixel_values = 2 * pixel_values - 1
+    pixel_values = rearrange(pixel_values, 'h w c -> c h w')
+    return pixel_values
+
+# ============================================================
+# HELPER: ĐO RAM CPU & VRAM GPU
+# ============================================================
+def _get_ram_mb():
+    return psutil.Process(os.getpid()).memory_info().rss / 1024 / 1024
+
+def _get_vram_mb():
+    return torch.cuda.memory_allocated() / 1024 / 1024 if torch.cuda.is_available() else 0.0
+
+def _get_vram_reserved_mb():
+    return torch.cuda.memory_reserved() / 1024 / 1024 if torch.cuda.is_available() else 0.0
+
+def _log_resources(label: str):
+    ram          = _get_ram_mb()
+    vram_alloc   = _get_vram_mb()
+    vram_reserved = _get_vram_reserved_mb()
+    print(f"  📊 [{label}]"
+          f"  RAM: {ram:.0f} MB"
+          f"  |  VRAM alloc: {vram_alloc:.0f} MB"
+          f"  |  VRAM reserved: {vram_reserved:.0f} MB")
+
+# ============================================================
+# INFERENCE
+# ============================================================
+@spaces.GPU(duration=120)
+def run_inference(task_type, prompt, cfg_prompt, cfg_scale, num_steps, seed, method, src_img=None):
+    t_total_start = time.perf_counter()
+    mode_label    = "Text-to-Image" if task_type == "t2i" else "Image-Editing"
+
+    print(f"\n{'='*60}")
+    print(f"🚀 BẮT ĐẦU [{mode_label}]  steps={num_steps}  cfg={cfg_scale}  seed={seed}")
+    print(f"   Prompt: {prompt[:100]}{'...' if len(prompt) > 100 else ''}")
+    print(f"{'='*60}")
+
+    # Reset peak VRAM counter cho lần chạy này
+    if torch.cuda.is_available():
+        torch.cuda.reset_peak_memory_stats()
+
+    _log_resources("Khởi đầu")
+
+    try:
+        # ── 1. Load model ──────────────────────────────────────
+        t0  = time.perf_counter()
+        net = load_model(method)
+        net = net.to("cuda")
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()
+        t_load = time.perf_counter() - t0
+        print(f"  ⏱️  Load/move model to CUDA : {t_load:.2f}s")
+        _log_resources("Sau load model")
+
+        generator   = torch.Generator(device=net.device).manual_seed(int(seed))
+        prompts     = [prompt.strip()]
+        cfg_prompts = [cfg_prompt]
+
+        # ── 2. Tiền xử lý ảnh nguồn (chỉ i2i) ────────────────
+        t_img = 0.0
+        pixel_values_src = None
+        if task_type == "i2i" and src_img is not None:
+            t0 = time.perf_counter()
+            processed_img    = _process_image(src_img).to(net.dtype).to(net.device)
+            pixel_values_src = processed_img.unsqueeze(0)
+            t_img = time.perf_counter() - t0
+            print(f"  ⏱️  Tiền xử lý ảnh nguồn  : {t_img:.3f}s")
+            _log_resources("Sau tiền xử lý ảnh")
+
+        # ── 3. Generate (bước nặng) ────────────────────────────
+        _log_resources("Trước generate")
+        t0 = time.perf_counter()
+        with torch.no_grad():
+            images = net.generate(
+                prompt=prompts,
+                cfg_prompt=cfg_prompts,
+                pixel_values_src=pixel_values_src,
+                cfg_scale=cfg_scale,
+                num_steps=int(num_steps),
+                progress_bar=False,
+                generator=generator,
+                height=512,
+                width=512
+            )
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()
+        t_gen = time.perf_counter() - t0
+        print(f"  ⏱️  Generate ({num_steps} steps)     : {t_gen:.2f}s  "
+              f"({t_gen / int(num_steps) * 1000:.1f} ms/step)")
+        _log_resources("Sau generate")
+
+        # ── 4. Post-process ────────────────────────────────────
+        t0 = time.perf_counter()
+        images = rearrange(images, 'b c h w -> b h w c')
+        images = torch.clamp(127.5 * images + 128.0, 0, 255).to("cpu", dtype=torch.uint8).numpy()
+        t_post = time.perf_counter() - t0
+        print(f"  ⏱️  Post-process             : {t_post:.3f}s")
+
+        # ── 5. Offload về CPU (ZeroGPU rule) ──────────────────
+        net = net.to("cpu")
+        torch.cuda.empty_cache()
+
+        # ── Tổng kết ──────────────────────────────────────────
+        t_total    = time.perf_counter() - t_total_start
+        vram_peak  = torch.cuda.max_memory_allocated() / 1024 / 1024 if torch.cuda.is_available() else 0
+        vram_now   = _get_vram_mb()
+        ram_now    = _get_ram_mb()
+
+        sep = "=" * 60
+        print(f"\n{sep}")
+        print(f"✅ KẾT QUẢ [{mode_label}]")
+        print(f"   ┌─ Thời gian ──────────────────────────────")
+        print(f"   │  Tổng              : {t_total:.2f}s")
+        print(f"   │  Load model        : {t_load:.2f}s")
+        if task_type == "i2i" and src_img is not None:
+            print(f"   │  Tiền xử lý ảnh   : {t_img:.3f}s")
+        print(f"   │  Generate          : {t_gen:.2f}s  ({t_gen/int(num_steps)*1000:.1f} ms/step)")
+        print(f"   │  Post-process      : {t_post:.3f}s")
+        print(f"   ├─ Bộ nhớ ────────────────────────────────")
+        print(f"   │  RAM CPU hiện tại  : {ram_now:.0f} MB")
+        print(f"   │  VRAM peak         : {vram_peak:.0f} MB")
+        print(f"   │  VRAM hiện tại     : {vram_now:.0f} MB")
+        print(f"   └──────────────────────────────────────────")
+        print(f"{sep}\n")
+
+        return Image.fromarray(images[0])
+
+    except Exception as e:
+        import traceback
+        t_total = time.perf_counter() - t_total_start
+        print(f"\n❌ LỖI sau {t_total:.2f}s [{mode_label}]: {str(e)}")
+        traceback.print_exc()
+        return None
+
+# --- GIAO DIỆN GRADIO ---
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# 🚀 DeepGen-1.0 Demo trên ZeroGPU")
+
+    method_dropdown = gr.Dropdown(
+        choices=list(MODEL_WEIGHTS.keys()),
+        value="RL with MR-GRPO (Tốt nhất)",
+        label="Cấu hình Model / Weights"
+    )
+
+    with gr.Tabs():
+        with gr.Tab("🖼️ Text-to-Image"):
+            with gr.Row():
+                with gr.Column():
+                    t2i_prompt     = gr.Textbox(label="Prompt", value="A quiet bookstore with a sign that says 'READ'. A coffee cup on the table with the word 'MORNING'.")
+                    t2i_cfg_prompt = gr.Textbox(label="Negative / CFG Prompt", value="")
+                    t2i_scale      = gr.Slider(1.0, 10.0, value=4.0, label="CFG Scale")
+                    t2i_steps      = gr.Slider(10, 100, value=50, step=1, label="Steps")
+                    t2i_seed       = gr.Number(value=42, label="Seed")
+                    t2i_btn        = gr.Button("Tạo ảnh", variant="primary")
+                with gr.Column():
+                    t2i_output = gr.Image(label="Kết quả")
+            t2i_btn.click(
+                fn=lambda *args: run_inference("t2i", *args),
+                inputs=[t2i_prompt, t2i_cfg_prompt, t2i_scale, t2i_steps, t2i_seed, method_dropdown],
+                outputs=t2i_output
+            )
+
+        with gr.Tab("🎨 Image Editing"):
+            with gr.Row():
+                with gr.Column():
+                    i2i_src        = gr.Image(label="Ảnh nguồn (Source Image)", type="pil")
+                    i2i_prompt     = gr.Textbox(label="Editing Prompt", value="Transform into a vibrant 1980s concert poster with bold colors. Keep the band members and text, add warm orange and yellow tones, retro typography style.")
+                    i2i_cfg_prompt = gr.Textbox(label="Negative Prompt", value="blurry, distorted faces, extra limbs, text errors, low quality, oversaturated")
+                    i2i_scale      = gr.Slider(1.0, 10.0, value=4.0, label="CFG Scale")
+                    i2i_steps      = gr.Slider(10, 100, value=20, step=1, label="Steps")
+                    i2i_seed       = gr.Number(value=42, label="Seed")
+                    i2i_btn        = gr.Button("Chỉnh sửa ảnh", variant="primary")
+                with gr.Column():
+                    i2i_output = gr.Image(label="Kết quả chỉnh sửa")
+            i2i_btn.click(
+                fn=lambda src, p, cfg, s, st, sd, m: run_inference("i2i", p, cfg, s, st, sd, m, src),
+                inputs=[i2i_src, i2i_prompt, i2i_cfg_prompt, i2i_scale, i2i_steps, i2i_seed, method_dropdown],
+                outputs=i2i_output
+            )
+
+# KHÔNG XÓA DÒNG NÀY
+demo.launch()

configs/datasets/deepgen_512_fix_pixels/cc12m.py

@@ -0,0 +1,26 @@
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from src.datasets.collate_functions import collate_func_gen_txt_dynamic
+from src.datasets.text2image.caption_datasets import CaptionDataset
+
+
+with read_base():
+    from .processors import image_size, image_process
+
+dataset = dict(type=CaptionDataset,
+               image_size=image_size,
+               image_process=image_process,
+               cap_folder='data/cc12m/captions',
+               data_path='data/cc12m/data.json',
+               image_folder='data/cc12m/raw',
+               ceph_folder=None,
+               ceph_config=None)
+
+train_dataloader = dict(
+    batch_size=8,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_gen_txt_dynamic)
+)

configs/datasets/deepgen_512_fix_pixels/edit_pretrain.py

@@ -0,0 +1,128 @@
+from src.datasets.collate_functions import collate_func_img2img_txt_dynamic
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from xtuner.dataset import ConcatDataset
+from src.datasets.image2image.edit_datasets import ImageEditDataset, ReconstructDataset
+from PIL import Image
+with read_base():
+    from .processors import image_size, image_process
+
+
+
+
+
+
+
+
+Open4oEdit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/OpenGPT-4o-Image/editing.json',
+               image_folder = "data/OpenGPT-4o-Image/editing" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+Share4oEdit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/ShareGPT-4o-Image/text_and_image_to_image.json',
+               image_folder = "data/ShareGPT-4o-Image/editing" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+nano150Edit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/Nano-150k/data.json',
+               image_folder = "data/Nano-150k" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+picobananaEdit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/pico-banana/sft_with_local_source_image_path.json',
+               image_folder = "data/pico-banana" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+
+UniworldEdit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/UniWorld-V1-new/UniworldEdit.json',
+               image_folder = "data/UniWorld-V1-new" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+
+
+GPT4oEdit_hqedit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/GPT-Image-Edit-1.5M/gpt-edit/hqedit/hqedit.json',
+               image_folder = "data/GPT-Image-Edit-1.5M" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+GPT4oEdit_omniedit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/GPT-Image-Edit-1.5M/gpt-edit/omniedit/omniedit.json',
+               image_folder = "data/GPT-Image-Edit-1.5M" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+GPT4oEdit_ultraedit = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/GPT-Image-Edit-1.5M/gpt-edit/ultraedit/ultraedit.json',
+               image_folder = "data/GPT-Image-Edit-1.5M" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+Reason_edit = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/unireason/reson_edit.json',
+               image_folder = "data/unireason/edit" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+Omnigen_edit = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/X2I2/Omnigen.json',
+               image_folder = "data/X2I2/images" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+nhr_edit_1 = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/NHR-Edit/NHR-Edit_1.json',
+               image_folder = "data/NHR-Edit" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+nhr_edit_2 = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/NHR-Edit-part2/NHR-Edit_2.json',
+               image_folder = "data/NHR-Edit-part2" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+dataset = dict(
+    type=ConcatDataset,
+    datasets=[Open4oEdit,Share4oEdit,nano150Edit,picobananaEdit,UniworldEdit,GPT4oEdit_hqedit,GPT4oEdit_omniedit,GPT4oEdit_ultraedit,Reason_edit,Omnigen_edit,nhr_edit_1,nhr_edit_2],
+)
+
+train_dataloader = dict(
+    batch_size=16,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_img2img_txt_dynamic)
+)

configs/datasets/deepgen_512_fix_pixels/edit_sft_zh.py

@@ -0,0 +1,128 @@
+from src.datasets.collate_functions import collate_func_img2img_txt_dynamic
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from xtuner.dataset import ConcatDataset
+from src.datasets.image2image.edit_datasets import ImageEditDataset, ReconstructDataset
+from PIL import Image
+with read_base():
+    from .processors import image_size, image_process
+
+
+
+
+
+
+
+
+Open4oEdit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/OpenGPT-4o-Image/editing.json',
+               image_folder = "data/OpenGPT-4o-Image/editing" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+Share4oEdit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/ShareGPT-4o-Image/text_and_image_to_image.json',
+               image_folder = "data/ShareGPT-4o-Image/editing" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+nano150Edit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/Nano-150k/data.json',
+               image_folder = "data/Nano-150k" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+picobananaEdit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/pico-banana/sft_with_local_source_image_path.json',
+               image_folder = "data/pico-banana" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+
+UniworldEdit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/UniWorld-V1-new/UniworldEdit.json',
+               image_folder = "data/UniWorld-V1-new" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+
+
+GPT4oEdit_hqedit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/GPT-Image-Edit-1.5M/gpt-edit/hqedit/hqedit.json',
+               image_folder = "data/GPT-Image-Edit-1.5M" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+GPT4oEdit_omniedit  = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/GPT-Image-Edit-1.5M/gpt-edit/omniedit/omniedit.json',
+               image_folder = "data/GPT-Image-Edit-1.5M" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+GPT4oEdit_ultraedit = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/GPT-Image-Edit-1.5M/gpt-edit/ultraedit/ultraedit.json',
+               image_folder = "data/GPT-Image-Edit-1.5M" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+Reason_edit = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/unireason/reson_edit.json',
+               image_folder = "data/unireason/edit" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+Omnigen_edit = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/X2I2/Omnigen.json',
+               image_folder = "data/X2I2/images" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+nhr_edit_1 = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/NHR-Edit/NHR-Edit_1.json',
+               image_folder = "data/NHR-Edit" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+nhr_edit_2 = dict(type=ImageEditDataset,
+               image_size=image_size,
+               image_process=image_process,
+               data_path='data/NHR-Edit-part2/NHR-Edit_2.json',
+               image_folder = "data/NHR-Edit-part2" ,
+               ceph_folder=None,
+               ceph_config=None)
+
+dataset = dict(
+    type=ConcatDataset,
+    datasets=[Open4oEdit,Share4oEdit,nano150Edit,picobananaEdit,UniworldEdit,GPT4oEdit_hqedit,GPT4oEdit_omniedit,GPT4oEdit_ultraedit,Reason_edit,Omnigen_edit,nhr_edit_1,nhr_edit_2],
+)
+
+train_dataloader = dict(
+    batch_size=16,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_img2img_txt_dynamic)
+)

configs/datasets/deepgen_512_fix_pixels/joint_pretrain.py

@@ -0,0 +1,47 @@
+from src.datasets.collate_functions import (collate_func_img2img_txt_dynamic,
+                                            collate_func_gen_txt_dynamic, CollateConcat)
+from mmengine.config import read_base
+from src.datasets.samplers.multi_source_sampler import MultiSourceSampler, MultiSourceBatchSampler
+
+from xtuner.dataset import ConcatDataset
+
+
+with read_base():
+    from .processors import image_size, image_process
+    from .t2i_pretrain import dataset as t2i_pretrain_dataset
+    from .edit_pretrain import dataset as edit_pretrain_dataset
+
+
+dataset = dict(
+    type=ConcatDataset,
+    datasets=[edit_pretrain_dataset, t2i_pretrain_dataset]
+)
+
+group_keys = ['image2image', 'text2image']
+repeats = [1, 3] # the radio between editing and generation task 
+batch_sizes = [4, 4]
+batch_size = sum([repeat * batch_size for repeat, batch_size in zip(repeats, batch_sizes)]) // sum(repeats)
+
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=4,
+    prefetch_factor=1,
+    persistent_workers=False,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=MultiSourceSampler,
+                 repeats=repeats,
+                 batch_sizes=batch_sizes,  # fixed batch size for all sources
+                 shuffle=True),
+    batch_sampler=dict(type=MultiSourceBatchSampler,
+                       repeats=repeats,
+                       batch_sizes=batch_sizes,
+                       ),
+    collate_fn=dict(type=CollateConcat,
+                    collate_fns=[dict(type=collate_func_img2img_txt_dynamic),
+                                 dict(type=collate_func_gen_txt_dynamic),
+                                 ],
+                    keys=group_keys
+                    )
+)

configs/datasets/deepgen_512_fix_pixels/joint_sft_zh.py

@@ -0,0 +1,47 @@
+from src.datasets.collate_functions import (collate_func_img2img_txt_dynamic,
+                                            collate_func_gen_txt_dynamic, CollateConcat)
+from mmengine.config import read_base
+from src.datasets.samplers.multi_source_sampler import MultiSourceSampler, MultiSourceBatchSampler
+
+from xtuner.dataset import ConcatDataset
+
+
+with read_base():
+    from .processors import image_size, image_process
+    from .t2i_sft_zh import dataset as t2i_sft_dataset
+    from .edit_sft_zh import dataset as edit_sft_dataset
+
+
+dataset = dict(
+    type=ConcatDataset,
+    datasets=[edit_sft_dataset, t2i_sft_dataset]
+)
+
+group_keys = ['image2image', 'text2image']
+repeats = [1, 2] # the radio between editing and generation task
+batch_sizes = [4, 4]
+batch_size = sum([repeat * batch_size for repeat, batch_size in zip(repeats, batch_sizes)]) // sum(repeats)
+
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=4,
+    prefetch_factor=1,
+    persistent_workers=False,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=MultiSourceSampler,
+                 repeats=repeats,
+                 batch_sizes=batch_sizes,  # fixed batch size for all sources
+                 shuffle=True),
+    batch_sampler=dict(type=MultiSourceBatchSampler,
+                       repeats=repeats,
+                       batch_sizes=batch_sizes,
+                       ),
+    collate_fn=dict(type=CollateConcat,
+                    collate_fns=[dict(type=collate_func_img2img_txt_dynamic),
+                                 dict(type=collate_func_gen_txt_dynamic),
+                                 ],
+                    keys=group_keys
+                    )
+)

configs/datasets/deepgen_512_fix_pixels/laion6m.py

@@ -0,0 +1,27 @@
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from src.datasets.collate_functions import collate_func_gen_txt_dynamic
+from src.datasets.text2image.caption_datasets import CaptionDataset
+
+
+with read_base():
+    from .processors import image_size, image_process
+
+dataset = dict(type=CaptionDataset,
+               image_size=image_size,
+               cap_source='caption',
+               image_process=image_process,
+               cap_folder='data/laion6m/captions',
+               data_path='data/laion6m/data.json',
+               image_folder='data/laion6m/raw',
+               ceph_folder=None,
+               ceph_config=None)
+
+train_dataloader = dict(
+    batch_size=8,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_gen_txt_dynamic)
+)

configs/datasets/deepgen_512_fix_pixels/megalith10m.py

@@ -0,0 +1,27 @@
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from src.datasets.collate_functions import collate_func_gen_txt_dynamic
+from src.datasets.text2image.caption_datasets import CaptionDataset
+
+
+with read_base():
+    from .processors import image_size, image_process
+
+dataset = dict(type=CaptionDataset,
+               image_size=image_size,
+               image_process=image_process,
+               cap_source='caption_internlm2_short',
+               cap_folder='data/megalith-10m/captions',
+               data_path='data/megalith-10m/megalith10m_all.json',
+               image_folder='data/megalith-10m/raw',
+               ceph_folder=None,
+               ceph_config=None)
+
+train_dataloader = dict(
+    batch_size=8,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_gen_txt_dynamic)
+)

configs/datasets/deepgen_512_fix_pixels/processors.py

@@ -0,0 +1,41 @@
+from transformers import AutoImageProcessor, AutoTokenizer
+
+IMAGE_MEAN = (0.48145466, 0.4578275, 0.40821073)
+IMAGE_STD = (0.26862954, 0.26130258, 0.27577711)
+
+
+qwen2_5_vl_model_name_or_path = "model_zoo/Qwen2.5-VL-3B-Instruct"
+
+prompt_template = dict(
+    IMG_START_TOKEN='<|vision_start|>',
+    IMG_END_TOKEN='<|vision_end|>',
+    IMG_CONTEXT_TOKEN='<|image_pad|>',
+    IMG_START_TOKEN_FOR_GENERATION=False,
+    SYSTEM=('<|im_start|>system\n{system}<|im_end|>\n'),
+    INSTRUCTION=('<|im_start|>user\n{input}<|im_end|>\n'
+                 '<|im_start|>assistant\n'),
+    SUFFIX='<|im_end|>',
+    SUFFIX_AS_EOS=True,
+    SEP='\n',
+    STOP_WORDS=['<|im_end|>', '<|endoftext|>'],
+    GENERATION='Generate an image: {input}',
+    CFG='Generate an image.'
+)
+
+tokenizer_kwargs = dict(add_special_tokens=True)
+
+pad_index = 0
+image_size = 512
+image_process = 'fix_pixels'
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=qwen2_5_vl_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+
+image_processor = dict(type=AutoImageProcessor.from_pretrained,
+                       pretrained_model_name_or_path=qwen2_5_vl_model_name_or_path)

configs/datasets/deepgen_512_fix_pixels/redcaps5m.py

@@ -0,0 +1,27 @@
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from src.datasets.collate_functions import collate_func_gen_txt_dynamic
+from src.datasets.text2image.caption_datasets import CaptionDataset
+
+
+with read_base():
+    from .processors import image_size, image_process
+
+dataset = dict(type=CaptionDataset,
+               image_size=image_size,
+               image_process=image_process,
+               cap_source='caption',
+               cap_folder='data/redcaps5m_resized/raw',
+               data_path='data/redcaps5m_resized/redcaps5m_data.json',
+               image_folder='data/redcaps5m_resized/raw',
+               ceph_folder=None,
+               ceph_config=None)
+
+train_dataloader = dict(
+    batch_size=8,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_gen_txt_dynamic)
+)

configs/datasets/deepgen_512_fix_pixels/t2i_pretrain.py

@@ -0,0 +1,35 @@
+from src.datasets.collate_functions import collate_func_gen_txt_dynamic
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from xtuner.dataset import ConcatDataset
+from src.datasets.text2image.blip3_o import BLIP3oDataset
+
+
+
+with read_base():
+    from .processors import image_size, image_process
+    from .redcaps5m import dataset as redcaps5m_datasets
+    from .laion6m import dataset as laion6m_dataset
+    from .text2image2m import dataset as text2image2m_dataset
+    from .megalith10m import dataset as megalith10m_dataset
+    from .cc12m import dataset as cc12m_dataset
+
+
+
+
+
+dataset = dict(
+    type=ConcatDataset,
+    datasets=[redcaps5m_datasets, laion6m_dataset, text2image2m_dataset, megalith10m_dataset,cc12m_dataset],
+)
+
+
+train_dataloader = dict(
+    batch_size=4,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_gen_txt_dynamic)
+)
+

configs/datasets/deepgen_512_fix_pixels/t2i_sft_zh.py

@@ -0,0 +1,97 @@
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from src.datasets.collate_functions import collate_func_gen_txt_dynamic
+from src.datasets.text2image.blip3_o import BLIP3oDataset
+from xtuner.dataset import ConcatDataset
+
+with read_base():
+    from .processors import image_size, image_process
+
+image_process = 'fix_pixels'
+dataset_blip3o60k = dict(type=BLIP3oDataset,
+               image_size=image_size,
+               data_path='data/BLIP3o/blip3o_60k.json',
+               image_folder ='data/BLIP3o',
+               image_process = image_process,
+               ceph_folder=None,
+               ceph_config=None)
+
+dataset_share4oimg = dict(type=BLIP3oDataset,
+               image_size=image_size,
+               data_path='data/ShareGPT-4o-Image/share_4o_img.json',
+               image_folder = "data/ShareGPT-4o-Image/t2i" ,
+               image_process =image_process,
+               ceph_folder=None,
+               ceph_config=None)
+
+dataset_echo4oimg = dict(type=BLIP3oDataset,
+               image_size=image_size,
+               data_path='data/Echo4o/echo-4o-image_t2i.json',
+               image_folder = "data/Echo4o" ,
+               image_process =image_process,
+               ceph_folder=None,
+               ceph_config=None)
+
+dataset_open4oimg = dict(type=BLIP3oDataset,
+               image_size=image_size,
+               data_path='data/OpenGPT-4o-Image/OpenGPT-4o-Image.json',
+               image_folder = "data/OpenGPT-4o-Image/t2i" ,
+               image_process= image_process,
+               ceph_folder=None,
+               ceph_config=None)
+
+
+
+
+
+
+
+dataset_reason_img = dict(type=BLIP3oDataset,
+               image_size=image_size,
+               data_path='data/unireason/reson_t2i.json',
+               image_folder = "data/unireason/t2i" ,
+               image_process = image_process,
+               ceph_folder=None,
+               ceph_config=None)
+
+dataset_banana = dict(type=BLIP3oDataset,
+               image_size=image_size,
+               data_path='data/banana/banana-50k.json',
+               image_folder = "data/banana" ,
+               image_process = image_process,
+               ceph_folder=None,
+               ceph_config=None)
+
+dataset_text = dict(type=BLIP3oDataset,
+               image_size=image_size,
+               data_path='data/text_render/text.json',
+               image_folder = "data/text_render" ,
+               image_process = image_process,
+               ceph_folder=None,
+               ceph_config=None)
+
+
+
+dataset_poster = dict(type=BLIP3oDataset,
+               image_size=image_size,
+               data_path='data/poster/data.json',
+               image_folder = "data/poster" ,
+               image_process = image_process,
+               ceph_folder=None,
+               ceph_config=None)
+
+
+
+dataset = dict(
+    type=ConcatDataset,
+    datasets=[dataset_blip3o60k,dataset_share4oimg,dataset_echo4oimg,dataset_open4oimg,dataset_reason_img,dataset_banana,dataset_text,dataset_poster],
+)
+
+train_dataloader = dict(
+    batch_size=2,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_gen_txt_dynamic)
+)

configs/datasets/deepgen_512_fix_pixels/text2image2m.py

@@ -0,0 +1,42 @@
+from src.datasets.text2image.caption_datasets import CaptionDataset
+from src.datasets.collate_functions import collate_func_gen_txt_dynamic
+from mmengine.config import read_base
+from mmengine.dataset import InfiniteSampler
+from xtuner.dataset import ConcatDataset
+
+with read_base():
+    from .processors import image_size, image_process
+
+t2i_2m = dict(type=CaptionDataset,
+              image_size=image_size,
+              image_process=image_process,
+              cap_source='prompt',
+              data_path='data/text-to-image-2M/data/data_512_2M.json',
+              cap_folder='data/text-to-image-2M/raw/data_512_2M',
+              image_folder='data/text-to-image-2M/raw/data_512_2M',
+              ceph_folder=None,
+              ceph_config=None,)
+
+t2i_10k = dict(type=CaptionDataset,
+               image_size=image_size,
+               image_process=image_process,
+               cap_source='prompt',
+               data_path='data/text-to-image-2M/data/data_1024_10K.json',
+               cap_folder='data/text-to-image-2M/raw/data_1024_10K',
+               image_folder='data/text-to-image-2M/raw/data_1024_10K',
+               ceph_folder=None,
+               ceph_config=None)
+
+dataset = dict(
+    type=ConcatDataset,
+    datasets=[t2i_2m, t2i_10k]
+)
+
+train_dataloader = dict(
+    batch_size=8,
+    num_workers=4,
+    pin_memory=True,
+    dataset=dataset,
+    sampler=dict(type=InfiniteSampler, shuffle=True),
+    collate_fn=dict(type=collate_func_gen_txt_dynamic)
+)

configs/finetune/deepgen_joint_sft.py

@@ -0,0 +1,115 @@
+from mmengine.config import read_base
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from xtuner.engine.runner import TrainLoop
+from src.optimisers.custom_adamw import CustomAdamW
+# from torch.optim import AdamW
+
+with read_base():
+    from ..models.deepgen import model
+    from ..datasets.deepgen_512_fix_pixels.joint_sft import train_dataloader
+
+
+model.num_queries = 128
+model.use_activation_checkpointing = False
+model.freeze_transformer = False
+model.lora_modules = 'auto'
+model.lora_rank = 64
+model.lora_alpha = 128
+model.pretrained_pth = "your_path_to_pretrained_pth"
+# Scheduler & Optimizer
+accumulative_counts = 3
+dataloader_num_workers = 4
+max_iters = 400000
+optim_type = CustomAdamW
+lr = 5e-5
+betas = (0.9, 0.95)
+weight_decay = 0.05
+max_norm = 1.0  # grad clip
+warmup_ratio = 0.01
+
+
+# Save
+save_steps = 40000
+save_total_limit = 5  # Maximum checkpoints to keep (-1 means unlimited)
+
+
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale="dynamic",
+    dtype="bfloat16",
+)
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=False,
+        begin=0,
+        end=warmup_ratio * max_iters),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=False,
+        begin=warmup_ratio * max_iters,
+        end=max_iters)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_iters=max_iters)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)

configs/finetune/deepgen_joint_sft_scb.py

@@ -0,0 +1,114 @@
+from mmengine.config import read_base
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from xtuner.engine.runner import TrainLoop
+from src.optimisers.custom_adamw import CustomAdamW
+# from torch.optim import AdamW
+
+with read_base():
+    from ..models.deepgen_scb import model
+    from ..datasets.deepgen_512_fix_pixels.joint_sft_zh import train_dataloader
+
+model.num_queries = 128
+model.use_activation_checkpointing = False
+model.freeze_transformer = False
+model.lora_modules = 'auto'
+model.lora_rank = 64
+model.lora_alpha = 128
+model.pretrained_pth = 'your_path_to_pretrained_pth'
+# Scheduler & Optimizer
+accumulative_counts = 3
+dataloader_num_workers = 4
+max_iters = 400000
+optim_type = CustomAdamW
+lr = 5e-5
+betas = (0.9, 0.95)
+weight_decay = 0.05
+max_norm = 1.0  # grad clip
+warmup_ratio = 0.01
+
+
+# Save
+save_steps = 40000
+save_total_limit = 5  # Maximum checkpoints to keep (-1 means unlimited)
+
+
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale="dynamic",
+    dtype="bfloat16",
+)
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=False,
+        begin=0,
+        end=warmup_ratio * max_iters),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=False,
+        begin=warmup_ratio * max_iters,
+        end=max_iters)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_iters=max_iters)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)

configs/models/deepgen.py

@@ -0,0 +1,71 @@
+import torch
+from src.models.sd3_kontext.qwen2_5_vl_sd3_hf_dynamic import Qwen2p5VLStableDiffusion3HF
+from diffusers import FlowMatchEulerDiscreteScheduler, AutoencoderKL
+from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer
+from src.models.sd3_kontext.transformer_sd3_dynamic import SD3Transformer2DModel
+
+
+sd3_5_model_name_or_path = "model_zoo/UniPic2-SD3.5M-Kontext-2B"
+qwen2_5_vl_model_name_or_path = "model_zoo/Qwen2.5-VL-3B-Instruct"
+
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=qwen2_5_vl_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+prompt_template = dict(
+    IMG_START_TOKEN='<|vision_start|>',
+    IMG_END_TOKEN='<|vision_end|>',
+    IMG_CONTEXT_TOKEN='<|image_pad|>',
+    IMG_START_TOKEN_FOR_GENERATION=False,
+    SYSTEM=('<|im_start|>system\n{system}<|im_end|>\n'),
+    INSTRUCTION=('<|im_start|>user\n{input}<|im_end|>\n'
+                 '<|im_start|>assistant\n'),
+    SUFFIX='<|im_end|>',
+    SUFFIX_AS_EOS=True,
+    SEP='\n',
+    STOP_WORDS=['<|im_end|>', '<|endoftext|>'],
+    GENERATION='Generate an image: {input}',
+    CFG='Generate an image.'
+)
+
+
+model = dict(
+    type=Qwen2p5VLStableDiffusion3HF,
+    num_queries=128,
+    connector=dict(
+        hidden_size=2048,
+        intermediate_size=11946,
+        num_hidden_layers=6,
+        _attn_implementation='flash_attention_2',
+        num_attention_heads=32, ),
+    lmm=dict(type=Qwen2_5_VLForConditionalGeneration.from_pretrained,
+             pretrained_model_name_or_path=qwen2_5_vl_model_name_or_path,
+             torch_dtype=torch.bfloat16,
+             attn_implementation="flash_attention_2", ),
+    tokenizer=tokenizer,
+    prompt_template=prompt_template,
+    freeze_lmm=True,
+    transformer=dict(
+        type=SD3Transformer2DModel.from_pretrained,
+        pretrained_model_name_or_path=sd3_5_model_name_or_path,
+        subfolder="transformer",
+        torch_dtype=torch.bfloat16),
+    test_scheduler=dict(
+        type=FlowMatchEulerDiscreteScheduler.from_pretrained,
+        pretrained_model_name_or_path=sd3_5_model_name_or_path,
+        subfolder="scheduler"),
+    train_scheduler=dict(
+        type=FlowMatchEulerDiscreteScheduler.from_pretrained,
+        pretrained_model_name_or_path=sd3_5_model_name_or_path,
+        subfolder="scheduler"),
+    vae=dict(
+        type=AutoencoderKL.from_pretrained,
+        pretrained_model_name_or_path=sd3_5_model_name_or_path,
+        subfolder="vae",
+        torch_dtype=torch.bfloat16),
+    pretrained_pth=None,
+    use_activation_checkpointing=False,
+    freeze_transformer=True,
+)

configs/models/deepgen_scb.py

@@ -0,0 +1,71 @@
+import torch
+from src.models.sd3_kontext.qwen2_5_vl_sd3_hf_dynamic_fusion import Qwen2p5VLStableDiffusion3HF
+from diffusers import FlowMatchEulerDiscreteScheduler, AutoencoderKL
+from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer
+from src.models.sd3_kontext.transformer_sd3_dynamic import SD3Transformer2DModel
+
+
+sd3_5_model_name_or_path = "model_zoo/UniPic2-SD3.5M-Kontext-2B"
+qwen2_5_vl_model_name_or_path = "model_zoo/Qwen2.5-VL-3B-Instruct"
+
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=qwen2_5_vl_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+prompt_template = dict(
+    IMG_START_TOKEN='<|vision_start|>',
+    IMG_END_TOKEN='<|vision_end|>',
+    IMG_CONTEXT_TOKEN='<|image_pad|>',
+    IMG_START_TOKEN_FOR_GENERATION=False,
+    SYSTEM=('<|im_start|>system\n{system}<|im_end|>\n'),
+    INSTRUCTION=('<|im_start|>user\n{input}<|im_end|>\n'
+                 '<|im_start|>assistant\n'),
+    SUFFIX='<|im_end|>',
+    SUFFIX_AS_EOS=True,
+    SEP='\n',
+    STOP_WORDS=['<|im_end|>', '<|endoftext|>'],
+    GENERATION='Generate an image: {input}',
+    CFG='Generate an image.'
+)
+
+
+model = dict(
+    type=Qwen2p5VLStableDiffusion3HF,
+    num_queries=128,
+    connector=dict(
+        hidden_size=2048,
+        intermediate_size=11946,
+        num_hidden_layers=6,
+        _attn_implementation='flash_attention_2',
+        num_attention_heads=32, ),
+    lmm=dict(type=Qwen2_5_VLForConditionalGeneration.from_pretrained,
+             pretrained_model_name_or_path=qwen2_5_vl_model_name_or_path,
+             torch_dtype=torch.bfloat16,
+             attn_implementation="flash_attention_2", ),
+    tokenizer=tokenizer,
+    prompt_template=prompt_template,
+    freeze_lmm=True,
+    transformer=dict(
+        type=SD3Transformer2DModel.from_pretrained,
+        pretrained_model_name_or_path=sd3_5_model_name_or_path,
+        subfolder="transformer",
+        torch_dtype=torch.bfloat16),
+    test_scheduler=dict(
+        type=FlowMatchEulerDiscreteScheduler.from_pretrained,
+        pretrained_model_name_or_path=sd3_5_model_name_or_path,
+        subfolder="scheduler"),
+    train_scheduler=dict(
+        type=FlowMatchEulerDiscreteScheduler.from_pretrained,
+        pretrained_model_name_or_path=sd3_5_model_name_or_path,
+        subfolder="scheduler"),
+    vae=dict(
+        type=AutoencoderKL.from_pretrained,
+        pretrained_model_name_or_path=sd3_5_model_name_or_path,
+        subfolder="vae",
+        torch_dtype=torch.bfloat16),
+    pretrained_pth=None,
+    use_activation_checkpointing=False,
+    freeze_transformer=True,
+)

configs/pretrain/deepgen_joint_pretrain.py

@@ -0,0 +1,114 @@
+from mmengine.config import read_base
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from xtuner.engine.runner import TrainLoop
+from src.optimisers.custom_adamw import CustomAdamW
+# from torch.optim import AdamW
+
+with read_base():
+    from ..models.deepgen import model
+    from ..datasets.deepgen_512_fix_pixels.joint_pretrain import train_dataloader
+
+
+
+model.num_queries = 128
+model.use_activation_checkpointing = False
+model.freeze_transformer = True
+model.lora_modules = None
+
+# Scheduler & Optimizer
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_iters = 200000
+optim_type = CustomAdamW
+lr = 1e-4
+betas = (0.9, 0.95)
+weight_decay = 0.05
+max_norm = 1.0  # grad clip
+warmup_ratio = 0.01
+
+
+# Save
+save_steps = 40000
+save_total_limit = 5  # Maximum checkpoints to keep (-1 means unlimited)
+
+
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale="dynamic",
+    dtype="bfloat16",
+)
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=False,
+        begin=0,
+        end=warmup_ratio * max_iters),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=False,
+        begin=warmup_ratio * max_iters,
+        end=max_iters)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_iters=max_iters)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)

configs/pretrain/deepgen_joint_pretrain_scb.py

@@ -0,0 +1,112 @@
+from mmengine.config import read_base
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from xtuner.engine.runner import TrainLoop
+from src.optimisers.custom_adamw import CustomAdamW
+# from torch.optim import AdamW
+
+with read_base():
+    from ..models.deepgen_scb import model
+    from ..datasets.deepgen_512_fix_pixels.joint_pretrain import train_dataloader
+
+model.num_queries = 128
+model.use_activation_checkpointing = False
+model.freeze_transformer = True
+model.lora_modules = None
+
+# Scheduler & Optimizer
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_iters = 200000
+optim_type = CustomAdamW
+lr = 1e-4
+betas = (0.9, 0.95)
+weight_decay = 0.05
+max_norm = 1.0  # grad clip
+warmup_ratio = 0.01
+
+
+# Save
+save_steps = 40000
+save_total_limit = 5  # Maximum checkpoints to keep (-1 means unlimited)
+
+
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale="dynamic",
+    dtype="bfloat16",
+)
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=False,
+        begin=0,
+        end=warmup_ratio * max_iters),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=False,
+        begin=warmup_ratio * max_iters,
+        end=max_iters)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_iters=max_iters)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)

requirements.txt

@@ -0,0 +1,13 @@
+numpy<2
+gradio
+huggingface_hub
+accelerate
+einops
+inflect
+peft
+opencv-python-headless
+spaces
+diffusers
+triton==2.1.0
+transformers==4.56.1
+mmengine

src/datasets/collate_functions.py

@@ -0,0 +1,228 @@
+import torch
+from xtuner.utils import DEFAULT_PAD_TOKEN_INDEX, IGNORE_INDEX
+from typing import Dict, Sequence
+from torch.nn.utils.rnn import pad_sequence
+from functools import partial
+from dataclasses import dataclass
+
+
+def collate_func_img2img(instances: Sequence[Dict],
+                         pad_index: int = DEFAULT_PAD_TOKEN_INDEX):
+    pixel_values_src_list, pixel_values_list, input_ids_list, texts = [], [], [], []
+    for instance in instances:
+        pixel_values_src_ = instance.pop('pixel_values_src')
+        if isinstance(pixel_values_src_, torch.Tensor):
+            pixel_values_src_ = [pixel_values_src_]
+        pixel_values_src_list += pixel_values_src_
+        pixel_values_list.append(instance.pop('pixel_values'))
+        input_ids_list.append(instance.pop('input_ids'))
+        texts.append(instance.pop('text', None))
+
+    ori_length = [len(ids) for ids in input_ids_list]
+    pad_length = max(ori_length)
+    attention_mask = torch.zeros(len(instances), pad_length, dtype=torch.bool)
+    input_ids = torch.full(size=(len(instances), pad_length),
+                           fill_value=pad_index, dtype=torch.long)
+
+    # left padding for editing
+    for i, length in enumerate(ori_length):
+        attention_mask[i, -length:] = True
+        input_ids_i = input_ids_list[i]
+        if not isinstance(input_ids_i, torch.Tensor):
+            input_ids_i = torch.tensor(input_ids_i, dtype=torch.long)
+        input_ids[i, -length:] = input_ids_i
+
+    pixel_values = torch.stack(pixel_values_list)
+    pixel_values_src = torch.stack(pixel_values_src_list)
+
+    data_dict = dict(input_ids=input_ids, attention_mask=attention_mask,
+                     pixel_values=pixel_values, pixel_values_src=pixel_values_src, texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+def collate_func_img2img_text(instances: Sequence[Dict]):
+    pixel_values_src_list, pixel_values_list, texts = [], [], []
+    for instance in instances:
+        pixel_values_src_ = instance.pop('pixel_values_src')
+        if isinstance(pixel_values_src_, torch.Tensor):
+            pixel_values_src_ = [pixel_values_src_]
+        pixel_values_src_list += pixel_values_src_
+        pixel_values_list.append(instance.pop('pixel_values'))
+        texts.append(instance.pop('text'))
+
+    pixel_values = torch.stack(pixel_values_list)
+    pixel_values_src = torch.stack(pixel_values_src_list)
+
+    data_dict = dict(pixel_values=pixel_values, pixel_values_src=pixel_values_src, texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+def collate_func_img2img_txt_dynamic(instances: Sequence[Dict]):
+    pixel_values_src, pixel_values, texts = [], [], []
+    for instance in instances:
+        pixel_values_src_ = instance.pop('pixel_values_src')
+        if isinstance(pixel_values_src_, torch.Tensor):    # only has one ref image
+            pixel_values_src_ = [pixel_values_src_]
+        pixel_values_src.append(pixel_values_src_)
+        pixel_values.append(instance.pop('pixel_values'))
+        texts.append(instance.pop('text'))
+
+    data_dict = dict(pixel_values=pixel_values, pixel_values_src=pixel_values_src, texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+def collate_func_gen_txt_dynamic(instances: Sequence[Dict]):
+    pixel_values, texts = [], []
+    for example in instances:
+        pixel_values.append(example.pop('pixel_values'))
+        texts.append(example.pop('text'))
+
+    data_dict = dict(pixel_values=pixel_values, texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+
+def collate_func_gen(instances: Sequence[Dict],
+                     pad_index: int = DEFAULT_PAD_TOKEN_INDEX):
+    pixel_values, input_ids, input_lengths, texts, pixel_init = [], [], [], [], []
+    for example in instances:
+        pixel_values.append(example.pop('pixel_values'))
+        input_lengths.append(len(example['input_ids']))
+        input_ids.append(example.pop('input_ids'))
+        texts.append(example.pop('text', None))
+        pixel_init.append(example.pop('pixel_init'))
+
+    input_ids = pad_sequence(input_ids, batch_first=True, padding_value=pad_index)
+    attention_mask = torch.zeros_like(input_ids).bool()
+    for i in range(len(input_ids)):
+        attention_mask[i, :input_lengths[i]] = True
+
+    data_dict = dict(pixel_values=torch.stack(pixel_values),
+                     pixel_init = pixel_init,
+                     input_ids=input_ids,
+                     attention_mask=attention_mask,
+                     texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+def collate_func_gen_text(instances: Sequence[Dict]):
+    pixel_values, texts = [], []
+    for example in instances:
+        pixel_values.append(example.pop('pixel_values'))
+        texts.append(example.pop('text'))
+
+    data_dict = dict(pixel_values=torch.stack(pixel_values), texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+def collate_func_gen_tokens(instances: Sequence[Dict],
+                            pad_index: int = DEFAULT_PAD_TOKEN_INDEX):
+    image_tokens, input_ids, input_lengths, texts = [], [], [], []
+    for example in instances:
+        image_tokens.append(example.pop('image_tokens'))
+        input_lengths.append(len(example['input_ids']))
+        input_ids.append(example.pop('input_ids'))
+        texts.append(example.pop('text', None))
+
+    input_ids = pad_sequence(input_ids, batch_first=True, padding_value=pad_index)
+    attention_mask = torch.zeros_like(input_ids).bool()
+    for i in range(len(input_ids)):
+        attention_mask[i, :input_lengths[i]] = True
+
+    data_dict = dict(image_tokens=torch.stack(image_tokens),
+                     input_ids=input_ids,
+                     attention_mask=attention_mask,
+                     texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+def collate_func_gen_latents(instances: Sequence[Dict],
+                             pad_index: int = DEFAULT_PAD_TOKEN_INDEX):
+    image_latents, input_ids, input_lengths, texts = [], [], [], []
+    for example in instances:
+        image_latents.append(example.pop('image_latents'))
+        input_lengths.append(len(example['input_ids']))
+        input_ids.append(example.pop('input_ids'))
+        texts.append(example.pop('text', None))
+
+    input_ids = pad_sequence(input_ids, batch_first=True, padding_value=pad_index)
+    attention_mask = torch.zeros_like(input_ids).bool()
+    for i in range(len(input_ids)):
+        attention_mask[i, :input_lengths[i]] = True
+
+    data_dict = dict(image_latents=torch.stack(image_latents),
+                     input_ids=input_ids,
+                     attention_mask=attention_mask,
+                     texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+def collate_func_gen_text_latents(instances: Sequence[Dict]):
+    image_latents, texts = [], []
+    for example in instances:
+        image_latents.append(example.pop('image_latents'))
+        texts.append(example.pop('text', None))
+
+    data_dict = dict(image_latents=torch.stack(image_latents), texts=texts)
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+def collate_func_und(instances, pad_index=DEFAULT_PAD_TOKEN_INDEX):
+    input_ids_list, labels_list, pixel_values_list = [], [], []
+
+    for sample in instances:
+        input_ids_list.append(torch.LongTensor(sample['input_ids']))
+        labels_list.append(torch.LongTensor(sample['labels']))
+
+        if 'pixel_values' in sample:
+            pixel_values_list.append(sample['pixel_values'])
+
+    ori_length = [len(input_ids_) for input_ids_ in input_ids_list]
+    # right padding
+    if len(instances) > 1:
+        input_ids = pad_sequence(
+            input_ids_list, batch_first=True, padding_value=pad_index)
+        labels = pad_sequence(
+            labels_list, batch_first=True, padding_value=IGNORE_INDEX)
+    else:
+        input_ids = torch.stack(input_ids_list)
+        labels = torch.stack(labels_list)
+
+    attention_mask = torch.zeros_like(input_ids).bool()
+    for i, length in enumerate(ori_length):
+        attention_mask[i, :length] = True        # right padding
+
+    data_dict = {
+        'input_ids': input_ids,
+        'attention_mask': attention_mask,
+        'labels': labels,
+        'pixel_values': torch.stack(pixel_values_list) if len(pixel_values_list) > 0 else None
+    }
+
+    return {'data': data_dict, 'data_samples': None}
+
+
+class CollateConcat(object):
+    def __init__(self, collate_fns, keys):
+        self.keys = keys
+        self.collate_fns = {}
+        for key, collate_fn in zip(keys, collate_fns):
+            func = collate_fn.pop('type')
+            self.collate_fns[key] = partial(func, **collate_fn)
+
+    def __call__(self, data_samples):
+        data_samples = [data_sample for data_sample in data_samples if len(data_sample) > 0]
+        data_dict = {}
+        key = data_samples[0]['type']
+        data_dict[key] = self.collate_fns[key](data_samples)['data']
+
+        return {'data': data_dict, 'data_samples': None}

src/datasets/image2image/edit_datasets.py

@@ -0,0 +1,88 @@
+import torch
+import numpy as np
+from einops import rearrange
+from src.datasets.utils import crop2square
+from src.datasets.text2image.caption_datasets import CaptionDataset
+from PIL import Image
+import os
+
+class ImageEditDataset(CaptionDataset):
+    def _process_image(self, image):
+        assert self.image_process != 'crop2square'
+        return super()._process_image(image)['pixel_values']
+        # image = image.resize(size=(self.image_size, self.image_size))
+        # pixel_values = torch.from_numpy(np.array(image)).float()
+        # pixel_values = pixel_values / 255
+        # pixel_values = 2 * pixel_values - 1
+        # pixel_values = rearrange(pixel_values, 'h w c -> c h w')
+        # return pixel_values
+
+    def _process_text(self, text):
+        prompt_template = self.prompt_template
+        image_tokens = prompt_template['IMG_START_TOKEN'] + \
+                       prompt_template['IMG_CONTEXT_TOKEN'] * self.image_length + \
+                       prompt_template['IMG_END_TOKEN']
+        prompt = f'{image_tokens}\n{text}'
+        prompt = self.prompt_template['INSTRUCTION'].format(input=prompt)
+        if self.prompt_template.get('IMG_START_TOKEN_FOR_GENERATION', True):
+            prompt += prompt_template['IMG_START_TOKEN']
+        input_ids = self.tokenizer.encode(prompt, return_tensors='pt', **self.tokenizer_kwargs)[0]
+
+        return dict(input_ids=input_ids)
+
+    def __getitem__(self, idx):
+        if self.debug:
+            idx = 0
+        try:
+            data_sample = self.data_list[idx]
+            if self.image_folder is not None:
+                source_image = Image.open(os.path.join(self.image_folder,data_sample['input_image'][0])).convert('RGB')
+                target_image = Image.open(os.path.join(self.image_folder,data_sample['output_image'])).convert('RGB')
+            else:
+                source_image = Image.open(data_sample['input_image'][0]).convert('RGB')
+                target_image = Image.open(data_sample['output_image']).convert('RGB')
+            # prompt = self._read_json(data_sample['annotation'])[self.cap_source]
+            prompt = data_sample['instruction']
+
+            pixel_values_src = self._process_image(source_image)
+            pixel_values = self._process_image(target_image)
+
+            data = self._process_text(prompt) if self.tokenizer is not None else dict()
+
+            data.update(
+                pixel_values_src=pixel_values_src, pixel_values=pixel_values,
+                image_dir=self.image_folder,type='image2image', text=prompt)
+
+            return data
+
+        except Exception as e:
+            print(f"Error when reading {self.data_path}:{self.data_list[idx]}: {e}", flush=True)
+            return self._retry()
+
+
+class ReconstructDataset(CaptionDataset):
+    def _process_image(self, image):
+        assert self.image_process != 'crop2square'
+        return super()._process_image(image)['pixel_values']
+
+    def __getitem__(self, idx):
+        if self.debug:
+            idx = 0
+        try:
+            data_sample = self.data_list[idx]
+            image = self._read_image(data_sample['image']).convert('RGB')
+            prompt = "Keep the image as it is."
+            pixel_values = pixel_values_src = self._process_image(image)
+
+            data = self._process_text(prompt) if self.tokenizer is not None else dict()
+
+            data.update(
+                pixel_values_src=pixel_values_src, pixel_values=pixel_values,
+                image_dir=self.image_folder, image_file=data_sample['image'],
+                type='image2image', text=prompt)
+
+            return data
+
+        except Exception as e:
+            print(f"Error when reading {self.data_path}:{self.data_list[idx]}: {e}", flush=True)
+            return self._retry()

src/datasets/samplers/multi_source_sampler.py

@@ -0,0 +1,203 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import itertools
+from typing import Iterator, List, Optional, Sized, Union
+import torch
+from mmengine.dist import get_dist_info, sync_random_seed
+from torch.utils.data import Sampler
+
+
+class FixedBatchMultiSourceSampler(Sampler):
+    r"""Multi-Source Infinite Sampler.
+
+    According to the sampling ratio, sample data from different
+    datasets to form batches.
+
+    Args:
+        repeat (tuple): repeat factor
+        dataset (Sized): The dataset.
+        batch_size (int): Size of mini-batch.
+        shuffle (bool): Whether shuffle the dataset or not. Defaults to True.
+        seed (int, optional): Random seed. If None, set a random seed.
+            Defaults to None.
+    """
+
+    def __init__(self,
+                 repeat,
+                 dataset: Sized,
+                 batch_size: int,
+                 shuffle: bool = True,
+                 seed: Optional[int] = None) -> None:
+
+        assert hasattr(dataset, 'cumulative_sizes'),\
+            f'The dataset must be ConcatDataset, but get {dataset}'
+        assert isinstance(batch_size, int) and batch_size > 0, \
+            'batch_size must be a positive integer value, ' \
+            f'but got batch_size={batch_size}'
+        assert len(repeat) == len(dataset.cumulative_sizes), \
+            'The length of repeat must be equal to ' \
+            f'the number of datasets, but got repeat={repeat}'
+
+        rank, world_size = get_dist_info()
+        self.rank = rank
+        self.world_size = world_size
+
+        self.dataset = dataset
+        self.repeat = repeat
+        self.cumulative_sizes = [0] + dataset.cumulative_sizes
+        self.batch_size = batch_size
+
+        self.seed = sync_random_seed() if seed is None else seed
+        self.shuffle = shuffle
+        self.source2inds = {
+            source: self._indices_of_rank(len(ds))
+            for source, ds in enumerate(dataset.datasets)
+        }
+
+    def _infinite_indices(self, sample_size: int) -> Iterator[int]:
+        """Infinitely yield a sequence of indices."""
+        g = torch.Generator()
+        g.manual_seed(self.seed)
+        while True:
+            if self.shuffle:
+                yield from torch.randperm(sample_size, generator=g).tolist()
+            else:
+                yield from torch.arange(sample_size).tolist()
+
+    def _indices_of_rank(self, sample_size: int) -> Iterator[int]:
+        """Slice the infinite indices by rank."""
+        yield from itertools.islice(
+            self._infinite_indices(sample_size), self.rank, None,
+            self.world_size)
+
+    def __len__(self) -> int:
+        return len(self.dataset)
+
+    def set_epoch(self, epoch: int) -> None:
+        """Not supported in `epoch-based runner."""
+        pass
+
+    def __iter__(self) -> Iterator[int]:
+        while True:
+            for source, repeat in enumerate(self.repeat):
+                for _ in range(repeat):
+                    batch_buffer_per_source = []
+                    while len(batch_buffer_per_source) < self.batch_size:
+                        idx = next(self.source2inds[source])
+                        idx += self.cumulative_sizes[source]
+                        batch_buffer_per_source.append(idx)
+
+                    yield from batch_buffer_per_source
+
+
+
+class MultiSourceSampler(Sampler):
+    def __init__(self,
+                 repeats,
+                 dataset: Sized,
+                 batch_sizes: list[int],
+                 shuffle: bool = True,
+                 seed: Optional[int] = None) -> None:
+
+        assert hasattr(dataset, 'cumulative_sizes'),\
+            f'The dataset must be ConcatDataset, but get {dataset}'
+
+        assert isinstance(batch_sizes, list), \
+            f'source_ratio must be a list, but got batch_sizes={batch_sizes}'
+        assert len(batch_sizes) == len(dataset.cumulative_sizes), \
+            'The length of batch_sizes must be equal to ' \
+            f'the number of datasets, but got batch_sizes={batch_sizes}'
+
+        rank, world_size = get_dist_info()
+        self.rank = rank
+        self.world_size = world_size
+
+        self.dataset = dataset
+        self.cumulative_sizes = [0] + dataset.cumulative_sizes
+        self.batch_sizes = batch_sizes
+
+
+        self.seed = sync_random_seed() if seed is None else seed
+        self.shuffle = shuffle
+        self.source2inds = {
+            source: self._indices_of_rank(len(ds))
+            for source, ds in enumerate(dataset.datasets)
+        }
+
+        self.repeats = repeats
+        assert len(self.repeats) == len(self.batch_sizes)
+
+    def _infinite_indices(self, sample_size: int) -> Iterator[int]:
+        """Infinitely yield a sequence of indices."""
+        g = torch.Generator()
+        g.manual_seed(self.seed)
+        while True:
+            if self.shuffle:
+                yield from torch.randperm(sample_size, generator=g).tolist()
+            else:
+                yield from torch.arange(sample_size).tolist()
+
+    def _indices_of_rank(self, sample_size: int) -> Iterator[int]:
+        """Slice the infinite indices by rank."""
+        yield from itertools.islice(
+            self._infinite_indices(sample_size), self.rank, None,
+            self.world_size)
+
+
+    def __len__(self) -> int:
+        return len(self.dataset)
+
+    def set_epoch(self, epoch: int) -> None:
+        """Not supported in `epoch-based runner."""
+        pass
+
+    def __iter__(self) -> Iterator[int]:
+        while True:
+            for source, (batch_size, repeat) in enumerate(zip(self.batch_sizes, self.repeats)):
+                for _ in range(repeat):
+                    batch_buffer_per_source = []
+                    while len(batch_buffer_per_source) < batch_size:
+                        idx = next(self.source2inds[source])
+                        idx += self.cumulative_sizes[source]
+                        batch_buffer_per_source.append(idx)
+
+                    yield from batch_buffer_per_source
+
+    @property
+    def batch_size(self):
+        batch_size_sum = sum([batch_size * repeat for batch_size, repeat in zip(self.batch_sizes, self.repeats)])
+        batch_size_ave = batch_size_sum // sum(self.repeats)
+
+        return batch_size_ave
+
+
+class MultiSourceBatchSampler(Sampler[list[int]]):
+    def __init__(
+        self,
+        sampler: Union[FixedBatchMultiSourceSampler, MultiSourceSampler],
+        batch_sizes: list[int],
+        repeats: list[int],
+        **kwargs
+    ) -> None:
+        self.sampler = sampler
+        self.batch_sizes = batch_sizes
+        self.repeats = repeats
+
+    def __iter__(self) -> Iterator[list[int]]:
+        # Implemented based on the benchmarking in https://github.com/pytorch/pytorch/pull/76951
+        sampler_iter = iter(self.sampler)
+
+        while True:
+            for source, (batch_size, repeat) in enumerate(zip(self.batch_sizes, self.repeats)):
+                for _ in range(repeat):
+                    batch = [*itertools.islice(sampler_iter, batch_size)]
+                    yield batch
+
+    @property
+    def batch_size(self):
+        batch_size_sum = sum([batch_size * repeat for batch_size, repeat in zip(self.batch_sizes, self.repeats)])
+        batch_size_ave = batch_size_sum // sum(self.repeats)
+
+        return batch_size_ave
+
+    def __len__(self) -> int:
+        return len(self.sampler) // self.batch_size

src/datasets/text2image/blip3_o.py

@@ -0,0 +1,49 @@
+import torch
+import os
+from src.datasets.text2image.caption_datasets import CaptionDataset
+from PIL import Image
+
+class BLIP3oDataset(CaptionDataset):
+    def __getitem__(self, idx):
+        if self.debug:
+            idx = 0
+        try:
+            data_sample = self.data_list[idx]
+
+            if self.image_tokens_folder is not None:
+                image_tokens = torch.load(os.path.join(self.image_tokens_folder,
+                                                       data_sample['image'] + '.pt')).long()
+                data = dict(image_tokens=image_tokens)
+            elif self.latents_ceph_folder is not None:
+                image_latents = torch.load(
+                    self._read_ceph(
+                        os.path.join(
+                            self.latents_ceph_folder, data_sample['image'] + '.pt'
+                        )
+                    )
+                )
+                data = dict(image_latents=image_latents)
+            elif self.image_latents_folder is not None:
+                image_latents = torch.load(os.path.join(self.image_latents_folder,
+                                                        data_sample['image'] + '.pt'))
+                data = dict(image_latents=image_latents)
+            else:
+                if self.image_folder is not None:
+                    image = Image.open(os.path.join(self.image_folder,data_sample['image_path'])).convert('RGB')
+                else:
+                    image = Image.open(data_sample['image_path']).convert('RGB')
+                data = self._process_image(image)
+
+                caption = data_sample['txt']
+
+            # print(caption)
+            data["pixel_init"] = image
+            data.update(self._process_text(caption))
+            data.update(image_dir=self.image_folder, image_file=None,
+                        type='text2image',text=caption)
+
+            return data
+
+        except Exception as e:
+            print(f"Error when reading {self.data_path}:{self.data_list[idx]}: {e}", flush=True)
+            return self._retry()

src/datasets/text2image/caption_datasets.py

@@ -0,0 +1,226 @@
+from torch.utils.data import Dataset
+from PIL import Image
+import os
+import io
+import json
+import random
+import torch
+try:
+    from aoss_client.client import Client
+except:
+    try:
+        from petrel_client.client import Client
+    except:
+        Client = None
+from glob import glob
+from xtuner.registry import BUILDER
+from src.datasets.utils import crop2square, resize_image_fix_pixels, resize_image_dynamic
+from einops import rearrange
+import numpy as np
+
+
+class CaptionDataset(Dataset):
+    def __init__(self,
+                 data_path,
+                 image_folder=None,
+                 debug=False,
+                 image_processor=None,
+                 image_process='crop2square',
+                 ceph_folder=None,
+                 latents_ceph_folder=None,
+                 ceph_config=None,
+                 tokenizer=None,
+                 prompt_template=None,
+                 max_length=2048,
+                 min_image_size=80,
+                 image_size=256,
+                 image_length=256,
+                 unit_image_size=32,
+                 image_tokens_folder=None,
+                 image_latents_folder=None,
+                 cap_folder=None,
+                 cap_source='caption',
+                 tokenizer_kwargs=dict(add_special_tokens=True),
+                 unconditional=0.1
+                 ):
+        super().__init__()
+        self.data_path = data_path
+        self._load_data(data_path)
+        self.image_folder = image_folder
+        self.cap_folder = cap_folder
+        self.cap_source = cap_source
+        self.debug = debug
+
+        if image_processor is not None:
+            self.image_processor = BUILDER.build(image_processor)
+        else:
+            self.image_processor = None
+
+        if tokenizer is not None:
+            self.tokenizer = BUILDER.build(tokenizer)
+        else:
+            self.tokenizer = None
+        self.prompt_template = prompt_template
+
+        self.max_length = max_length
+        self.image_process = image_process
+        self.image_length = image_length
+        self.image_tokens_folder = image_tokens_folder
+        self.image_latents_folder = image_latents_folder
+        self.min_image_size = min_image_size
+        self.image_size = image_size
+        self.unit_image_size = unit_image_size
+        self.unconditional = unconditional
+        self.tokenizer_kwargs = tokenizer_kwargs
+
+        self.FILE_CLIENT = None
+        self.ceph_folder = ceph_folder
+        self.ceph_config = ceph_config
+        self.latents_ceph_folder = latents_ceph_folder
+        self.use_ceph = ((Client is not None) and (ceph_config is not None) and os.path.exists(ceph_config))
+
+    def _load_data(self, data_path: str):      # image path and annotation path are saved in a json file
+        if data_path.endswith('.json'):
+            with open(data_path, 'r') as f:
+                self.data_list = json.load(f)
+        else:
+            json_files = glob(f"{data_path}/*.json")
+            data_list = []
+            for json_file in json_files:
+                with open(json_file, 'r') as f:
+                    data_list += json.load(f)
+
+            self.data_list = data_list
+
+        print(f"Load {len(self.data_list)} data samples from {data_path}", flush=True)
+
+    def __len__(self):
+        return len(self.data_list)
+
+    def _read_ceph(self, ceph_path):
+        if self.FILE_CLIENT is None:
+            self.FILE_CLIENT = Client(self.ceph_config)
+        data_bytes = self.FILE_CLIENT.get(ceph_path)
+
+        return io.BytesIO(data_bytes)
+
+    def _read_image(self, image_file):
+        if self.image_folder is None:
+            assert self.use_ceph
+            assert self.ceph_folder is not None
+            image = Image.open(
+                self._read_ceph(
+                    os.path.join(self.ceph_folder, image_file)
+                )
+            )
+        else:
+            image = Image.open(
+                os.path.join(self.image_folder, image_file)
+            )
+        assert image.width > self.min_image_size and image.height > self.min_image_size, f"Image: {image.size}"
+        assert image.width / image.height > 0.1, f"Image: {image.size}"
+        assert image.width / image.height < 10, f"Image: {image.size}"
+        return image.convert('RGB')
+
+    def _read_json(self, annotation_file):
+        if self.cap_folder is None:
+            assert self.use_ceph
+            assert self.ceph_folder is not None
+            annotation = json.load(
+                self._read_ceph(
+                    os.path.join(self.ceph_folder, annotation_file)
+                )
+            )
+        else:
+            with open(os.path.join(self.cap_folder, annotation_file), 'r') as f:
+                annotation = json.load(f)
+
+        return annotation
+
+    def _process_image(self, image):
+        data = dict()
+        if self.image_process == 'crop2square':
+            image = crop2square(image)
+            image = image.resize(size=(self.image_size, self.image_size))
+        elif self.image_process == 'dynamic':   # dynamic and make sure the largest edge <= self.image_size
+            image = resize_image_dynamic(x=image, image_size=self.image_size, unit_image_size=self.unit_image_size)
+        elif self.image_process == 'fix_pixels': # fix pixels contain radio of image
+            # import pdb; pdb.set_trace()
+            image = resize_image_fix_pixels(x=image, image_size=self.image_size, unit_image_size=self.unit_image_size)
+        elif self.image_process == 'resize2square':
+            image = image.resize(size=(self.image_size, self.image_size))
+        else:
+            raise NotImplementedError
+
+        # assert image.width <= self.image_size
+        # assert image.height <= self.image_size
+        assert image.width % self.unit_image_size == 0
+        assert image.height % self.unit_image_size == 0
+
+        pixel_values = torch.from_numpy(np.array(image)).float()
+        pixel_values = pixel_values / 255
+        pixel_values = 2 * pixel_values - 1
+        pixel_values = rearrange(pixel_values, 'h w c -> c h w')
+
+        data.update(pixel_values=pixel_values)
+        return data
+
+    def _process_text(self, text):
+        if self.tokenizer is None:
+            return {}
+        if random.uniform(0, 1) < self.unconditional:
+            prompt = self.prompt_template['CFG']
+        else:
+            prompt = self.prompt_template['GENERATION'].format(input=text.strip())
+
+        prompt = self.prompt_template['INSTRUCTION'].format(input=prompt)
+        if self.prompt_template.get('IMG_START_TOKEN_FOR_GENERATION', True):
+            prompt += self.prompt_template['IMG_START_TOKEN']
+        input_ids = self.tokenizer.encode(prompt, return_tensors='pt', **self.tokenizer_kwargs)[0]
+
+        return dict(input_ids=input_ids[:self.max_length])
+
+    def _retry(self):
+        return self.__getitem__(random.choice(range(self.__len__())))
+
+    def __getitem__(self, idx):
+        if self.debug:
+            idx = 0
+        try:
+            data_sample = self.data_list[idx]
+
+            if self.image_tokens_folder is not None:
+                image_tokens = torch.load(os.path.join(self.image_tokens_folder,
+                                                       data_sample['image'] + '.pt')).long()
+                data = dict(image_tokens=image_tokens)
+            elif self.latents_ceph_folder is not None:
+                image_latents = torch.load(
+                    self._read_ceph(
+                        os.path.join(
+                            self.latents_ceph_folder, data_sample['image'] + '.pt'
+                        )
+                    )
+                )
+                data = dict(image_latents=image_latents)
+            elif self.image_latents_folder is not None:
+                image_latents = torch.load(os.path.join(self.image_latents_folder,
+                                                        data_sample['image'] + '.pt'))
+                data = dict(image_latents=image_latents)
+            else:
+                image = self._read_image(data_sample['image']).convert('RGB')
+                data = self._process_image(image)
+
+            caption = self._read_json(data_sample['annotation'])[self.cap_source]
+            # caption = self._read_json(data_sample['annotation'])
+            # print(caption)
+
+            data.update(self._process_text(caption))
+            data['pixel_init'] = image
+            data.update(image_dir=self.image_folder, image_file=data_sample['image'],
+                        type='text2image',text=caption)
+
+            return data
+
+        except Exception as e:
+            print(f"Error when reading {self.data_path}:{self.data_list[idx]}: {e}", flush=True)
+            return self._retry()

src/datasets/utils.py

@@ -0,0 +1,186 @@
+import copy
+import random
+from xtuner.dataset.utils import get_bos_eos_token_ids
+from xtuner.utils import DEFAULT_IMAGE_TOKEN, IGNORE_INDEX, IMAGE_TOKEN_INDEX
+import json
+import math
+
+INPUT_IMAGE_TOKEN_INDEX = IMAGE_TOKEN_INDEX
+OUTPUT_IMAGE_TOKEN_INDEX = -300
+
+
+def resize_image_fix_pixels(x, image_size, unit_image_size=32):
+    # import pdb; pdb.set_trace()
+    w, h = x.size
+    ratio = image_size / ((h * w) ** 0.5)
+
+    target_h = math.ceil(h * ratio / unit_image_size) * unit_image_size
+    target_w = math.ceil(w * ratio / unit_image_size) * unit_image_size
+
+    x = x.resize(size=(target_w, target_h))
+
+    return x
+
+
+def resize_image_dynamic(x, image_size, unit_image_size=32):
+    w, h = x.size
+    if w >= h and w >= image_size:
+        target_w = image_size
+        target_h = h * (target_w / w)
+        target_h = math.ceil(target_h / unit_image_size) * unit_image_size
+
+    elif h >= w and h >= image_size:
+        target_h = image_size
+        target_w = w * (target_h / h)
+        target_w = math.ceil(target_w / unit_image_size) * unit_image_size
+
+    else:
+        target_h = math.ceil(h / unit_image_size) * unit_image_size
+        target_w = math.ceil(w / unit_image_size) * unit_image_size
+
+    x = x.resize(size=(target_w, target_h))
+
+    return x
+
+
+
+def crop2square(pil_img):
+    width, height = pil_img.width, pil_img.height
+
+    if width > height:
+        y0, y1 = 0, height
+        x0 = random.randint(0, width - height)    # [0, w - h]
+        x1 = x0 + height    # [h, w]
+    else:
+        x0, x1 = 0, width
+        y0 = random.randint(0, height - width)   # [0, h - w]
+        y1 = y0 + width     # [w, h]
+
+    return pil_img.crop(box=(x0, y0, x1, y1))
+
+
+def load_jsonl(json_file):
+    with open(json_file) as f:
+        lines = f.readlines()
+    data = []
+    for line in lines:
+        data.append(json.loads(line))
+    return data
+
+
+def encode_fn(example,
+              tokenizer,
+              max_length=None,
+              image_length=1,
+              input_ids_with_output=True,
+              with_image_token=False,
+              truncation='right'):
+    """We only support the following three scenarios:
+
+    1. Incremental pretraining dataset.
+        example['conversation'] = [
+                {
+                    'input': '',
+                    'output': '### Human: Can you write xxx'
+                }
+            ]
+
+    2. Single-turn conversation dataset.
+        example['conversation'] = [
+                {
+                    'input': 'Give three tips for staying healthy.',
+                    'output': '1.Eat a balanced diet xxx'
+                }
+            ]
+
+    3. Multi-turn conversation dataset.
+        example['conversation'] = [
+                {
+                    'input': 'Give three tips for staying healthy.',
+                    'output': '1.Eat a balanced diet xxx'
+                },
+                {
+                    'input': 'Please expand on the second point.',
+                    'output': 'Here is an expanded explanation of the xxx'
+                }
+            ]
+    """
+    bos_token_id, eos_token_id = get_bos_eos_token_ids(tokenizer)
+    is_multi_turn_conversation = len(example['conversation']) > 1
+    if is_multi_turn_conversation:
+        assert input_ids_with_output
+
+    input_ids, labels = [], []
+    next_needs_bos_token = True
+    for single_turn_conversation in example['conversation']:
+        input = single_turn_conversation['input']
+        if DEFAULT_IMAGE_TOKEN in input and with_image_token:
+            chunk_encode = [
+                tokenizer.encode(chunk, add_special_tokens=False)
+                for chunk in input.split(DEFAULT_IMAGE_TOKEN)
+            ]
+            assert len(chunk_encode) == 2
+            input_encode = []
+            for idx, cur_chunk_encode in enumerate(chunk_encode):
+                input_encode.extend(cur_chunk_encode)
+                if idx != len(chunk_encode) - 1:
+                    # input_encode.append(IMAGE_TOKEN_INDEX)
+                    input_encode += [IMAGE_TOKEN_INDEX] * image_length
+        else:
+            input_encode = tokenizer.encode(input, add_special_tokens=False)
+        if next_needs_bos_token:
+            input_ids += bos_token_id
+            labels += [IGNORE_INDEX] * len(bos_token_id)
+        input_ids += input_encode
+        labels += [IGNORE_INDEX] * len(input_encode)
+        if input_ids_with_output and 'output' in single_turn_conversation:
+            # Add output
+            output_with_loss = single_turn_conversation.get(
+                'output_with_loss', True)
+            output = single_turn_conversation['output']
+            if DEFAULT_IMAGE_TOKEN in output and with_image_token:
+                chunk_encode = [
+                    tokenizer.encode(chunk, add_special_tokens=False)
+                    for chunk in output.split(DEFAULT_IMAGE_TOKEN)
+                ]
+                assert len(chunk_encode) == 2
+                output_encode = []
+                for idx, cur_chunk_encode in enumerate(chunk_encode):
+                    output_encode.extend(cur_chunk_encode)
+                    if idx != len(chunk_encode) - 1:
+                        output_encode += [IMAGE_TOKEN_INDEX] * image_length
+            else:
+                output_encode = tokenizer.encode(output, add_special_tokens=False)
+            # output_encode = tokenizer.encode(output, add_special_tokens=False)
+            input_ids += output_encode
+            if output_with_loss:
+                labels += copy.deepcopy(output_encode)
+            else:
+                labels += [IGNORE_INDEX] * len(output_encode)
+            # Add EOS_TOKEN (with loss)
+            if single_turn_conversation.get('need_eos_token', True):
+                next_needs_bos_token = True
+                input_ids += eos_token_id
+                if output_with_loss:
+                    labels += copy.deepcopy(eos_token_id)
+                else:
+                    labels += [IGNORE_INDEX] * len(eos_token_id)
+            else:
+                next_needs_bos_token = False
+            # Add SEP (without loss)
+            sep = single_turn_conversation.get('sep', '')
+            if sep != '':
+                sep_encode = tokenizer.encode(sep, add_special_tokens=False)
+                input_ids += sep_encode
+                labels += [IGNORE_INDEX] * len(sep_encode)
+
+    if max_length is not None and len(input_ids) > max_length:
+        if truncation == 'right':
+            input_ids = input_ids[:max_length]
+            labels = labels[:max_length]
+        elif truncation == 'left':
+            input_ids = input_ids[-max_length:]
+            labels = labels[-max_length:]
+        else:
+            assert truncation is None
+    return {'input_ids': input_ids, 'labels': labels}

src/models/connector/__init__.py

@@ -0,0 +1,2 @@
+from .configuration_connector import ConnectorConfig
+from .modeling_connector import ConnectorEncoder

src/models/connector/configuration_connector.py

@@ -0,0 +1,27 @@
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class ConnectorConfig(PretrainedConfig):
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act

src/models/connector/modeling_connector.py

@@ -0,0 +1,507 @@
+# coding=utf-8
+# Copyright 2024 Google AI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch Connector model."""
+
+import math
+import warnings
+from typing import Any, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
+from transformers.activations import ACT2FN
+from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    ModelOutput,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+    torch_int,
+)
+from .configuration_connector import ConnectorConfig
+
+
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
+
+logger = logging.get_logger(__name__)
+
+
+def init_weights(module):
+    """Initialize the weights"""
+    if isinstance(module, nn.Embedding):
+        default_flax_embed_init(module.weight)
+    elif isinstance(module, ConnectorAttention):
+        nn.init.xavier_uniform_(module.q_proj.weight)
+        nn.init.xavier_uniform_(module.k_proj.weight)
+        nn.init.xavier_uniform_(module.v_proj.weight)
+        nn.init.xavier_uniform_(module.out_proj.weight)
+        nn.init.zeros_(module.q_proj.bias)
+        nn.init.zeros_(module.k_proj.bias)
+        nn.init.zeros_(module.v_proj.bias)
+        nn.init.zeros_(module.out_proj.bias)
+    elif isinstance(module, ConnectorMLP):
+        nn.init.xavier_uniform_(module.fc1.weight)
+        nn.init.xavier_uniform_(module.fc2.weight)
+        nn.init.normal_(module.fc1.bias, std=1e-6)
+        nn.init.normal_(module.fc2.bias, std=1e-6)
+    elif isinstance(module, (nn.Linear, nn.Conv2d)):
+        lecun_normal_(module.weight)
+        if module.bias is not None:
+            nn.init.zeros_(module.bias)
+    elif isinstance(module, nn.LayerNorm):
+        module.bias.data.zero_()
+        module.weight.data.fill_(1.0)
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    tensor.clamp_(min=a, max=b)
+
+
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+) -> torch.Tensor:
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsequently scaled and shifted by the mean and std args.
+
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+
+class ConnectorAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Input shape: Batch x Time x Channel"""
+
+        batch_size, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        k_v_seq_len = key_states.shape[-2]
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale
+
+        if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class ConnectorFlashAttention2(ConnectorAttention):
+    """
+    ConnectorAttention flash attention module. This module inherits from `ConnectorAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    is_causal = False
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    # Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        output_attentions = False
+
+        batch_size, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32.
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            is_causal=self.is_causal,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+        )
+
+        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights
+
+
+class ConnectorSdpaAttention(ConnectorAttention):
+    """
+    Connector attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `ConnectorAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    is_causal = False
+
+    # Adapted from ConnectorAttention.forward and transformers.models.llama.modeling_llama.LlamaSdpaAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "ConnectorModel is using ConnectorSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                output_attentions=output_attentions,
+            )
+
+        batch_size, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        is_causal = True if self.is_causal and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(batch_size, q_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None
+
+
+CONNECTOR_ATTENTION_CLASSES = {
+    "eager": ConnectorAttention,
+    "flash_attention_2": ConnectorFlashAttention2,
+    "sdpa": ConnectorSdpaAttention,
+}
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Connector
+class ConnectorMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class ConnectorEncoderLayer(nn.Module):
+    def __init__(self, config: ConnectorConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = CONNECTOR_ATTENTION_CLASSES[config._attn_implementation](config=config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = ConnectorMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    # Ignore copy
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, embed_dim)`.
+            attention_mask (`torch.FloatTensor`):
+                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.altclip.modeling_altclip.AltCLIPEncoder with AltCLIP->Connector
+class ConnectorEncoder(nn.Module):
+    def __init__(self, config: ConnectorConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([ConnectorEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+        self.apply(init_weights)
+
+    def forward(self, inputs_embeds):
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers:
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    encoder_layer.__call__,
+                    hidden_states,
+                    None,
+                    False,
+                    use_reentrant=False
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    None,
+                    output_attentions=False,
+                )
+
+            hidden_states = layer_outputs[0]
+
+        return hidden_states

src/models/connector/modeling_qwen2.py

@@ -0,0 +1,50 @@
+import torch
+import torch.nn as nn
+from transformers import Qwen2PreTrainedModel, Qwen2Config
+from transformers.models.qwen2.modeling_qwen2 import Qwen2RMSNorm, Qwen2DecoderLayer
+
+
+class Qwen2Connector(Qwen2PreTrainedModel):
+    def __init__(self, config: Qwen2Config):
+        super().__init__(config)
+        self.layers = nn.ModuleList(
+            [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+
+        for layer in self.layers:
+            layer.self_attn.is_causal = False
+
+        self._attn_implementation = config._attn_implementation
+        assert self._attn_implementation == 'flash_attention_2'
+        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(self, inputs_embeds):
+        position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device)
+        position_ids = position_ids.expand(inputs_embeds.shape[0], -1)
+        hidden_states = inputs_embeds
+
+        for encoder_layer in self.layers:
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    encoder_layer.__call__,
+                    hidden_states,
+                    None,
+                    position_ids,
+                    use_reentrant=False
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask=None,
+                    position_ids=position_ids,
+                )
+
+            hidden_states = layer_outputs[0]
+
+        hidden_states = self.norm(hidden_states)
+
+        return hidden_states

src/models/sd3_kontext/pipeline_stable_diffusion_3.py

@@ -0,0 +1,1256 @@
+# Copyright 2025 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import (
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    SiglipImageProcessor,
+    SiglipVisionModel,
+    T5EncoderModel,
+    T5TokenizerFast,
+)
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin
+from diffusers.models.autoencoders import AutoencoderKL
+from diffusers.models.transformers import SD3Transformer2DModel
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import StableDiffusion3Pipeline
+
+        >>> pipe = StableDiffusion3Pipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16
+        ... )
+        >>> pipe.to("cuda")
+        >>> prompt = "A cat holding a sign that says hello world"
+        >>> image = pipe(prompt).images[0]
+        >>> image.save("sd3.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin):
+    r"""
+    Args:
+        transformer ([`SD3Transformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant,
+            with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size`
+            as its dimension.
+        text_encoder_2 ([`CLIPTextModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
+        text_encoder_3 ([`T5EncoderModel`]):
+            Frozen text-encoder. Stable Diffusion 3 uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+            [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_3 (`T5TokenizerFast`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        image_encoder (`SiglipVisionModel`, *optional*):
+            Pre-trained Vision Model for IP Adapter.
+        feature_extractor (`SiglipImageProcessor`, *optional*):
+            Image processor for IP Adapter.
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae"
+    _optional_components = ["image_encoder", "feature_extractor"]
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"]
+
+    def __init__(
+        self,
+        transformer: SD3Transformer2DModel,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer_2: CLIPTokenizer,
+        text_encoder_3: T5EncoderModel,
+        tokenizer_3: T5TokenizerFast,
+        image_encoder: SiglipVisionModel = None,
+        feature_extractor: SiglipImageProcessor = None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            text_encoder_3=text_encoder_3,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            tokenizer_3=tokenizer_3,
+            transformer=transformer,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
+        )
+        self.default_sample_size = (
+            self.transformer.config.sample_size
+            if hasattr(self, "transformer") and self.transformer is not None
+            else 128
+        )
+        self.patch_size = (
+            self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2
+        )
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_images_per_prompt: int = 1,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        if self.text_encoder_3 is None:
+            return torch.zeros(
+                (
+                    batch_size * num_images_per_prompt,
+                    self.tokenizer_max_length,
+                    self.transformer.config.joint_attention_dim,
+                ),
+                device=device,
+                dtype=dtype,
+            )
+
+        text_inputs = self.tokenizer_3(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            # logger.warning(
+            #     "The following part of your input was truncated because `max_sequence_length` is set to "
+            #     f" {max_sequence_length} tokens: {removed_text}"
+            # )
+
+        prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0]
+
+        dtype = self.text_encoder_3.dtype
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def _get_clip_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        clip_skip: Optional[int] = None,
+        clip_model_index: int = 0,
+    ):
+        device = device or self._execution_device
+
+        clip_tokenizers = [self.tokenizer, self.tokenizer_2]
+        clip_text_encoders = [self.text_encoder, self.text_encoder_2]
+
+        tokenizer = clip_tokenizers[clip_model_index]
+        text_encoder = clip_text_encoders[clip_model_index]
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            # logger.warning(
+            #     "The following part of your input was truncated because CLIP can only handle sequences up to"
+            #     f" {self.tokenizer_max_length} tokens: {removed_text}"
+            # )
+        prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+        pooled_prompt_embeds = prompt_embeds[0]
+
+        if clip_skip is None:
+            prompt_embeds = prompt_embeds.hidden_states[-2]
+        else:
+            prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1)
+
+        return prompt_embeds, pooled_prompt_embeds
+
+    def encode_pooled_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        clip_skip: Optional[int] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder, lora_scale)
+            if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            _, pooled_prompt_embed = self._get_clip_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=0,
+            )
+            _, pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                prompt=prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=1,
+            )
+
+            pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            _, negative_pooled_prompt_embed = self._get_clip_prompt_embeds(
+                negative_prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=0,
+            )
+            _, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                negative_prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=1,
+            )
+
+            negative_pooled_prompt_embeds = torch.cat(
+                [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]],
+        prompt_3: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        clip_skip: Optional[int] = None,
+        max_sequence_length: int = 256,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in all text-encoders
+            prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
+                used in all text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
+            negative_prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
+                `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder, lora_scale)
+            if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            prompt_3 = prompt_3 or prompt
+            prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3
+
+            prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=0,
+            )
+            prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                prompt=prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=1,
+            )
+            clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1)
+
+            t5_prompt_embed = self._get_t5_prompt_embeds(
+                prompt=prompt_3,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+            clip_prompt_embeds = torch.nn.functional.pad(
+                clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1])
+            )
+
+            prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2)
+            pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+            negative_prompt_3 = negative_prompt_3 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+            negative_prompt_3 = (
+                batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3
+            )
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds(
+                negative_prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=0,
+            )
+            negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                negative_prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=1,
+            )
+            negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1)
+
+            t5_negative_prompt_embed = self._get_t5_prompt_embeds(
+                prompt=negative_prompt_3,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+            negative_clip_prompt_embeds = torch.nn.functional.pad(
+                negative_clip_prompt_embeds,
+                (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]),
+            )
+
+            negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2)
+            negative_pooled_prompt_embeds = torch.cat(
+                [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        prompt_3,
+        height,
+        width,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        negative_prompt_3=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        max_sequence_length=None,
+    ):
+        if (
+            height % (self.vae_scale_factor * self.patch_size) != 0
+            or width % (self.vae_scale_factor * self.patch_size) != 0
+        ):
+            raise ValueError(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}."
+                f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}."
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_3 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+        elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)):
+            raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_3 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if max_sequence_length is not None and max_sequence_length > 512:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def skip_guidance_layers(self):
+        return self._skip_guidance_layers
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def joint_attention_kwargs(self):
+        return self._joint_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_image
+    def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor:
+        """Encodes the given image into a feature representation using a pre-trained image encoder.
+
+        Args:
+            image (`PipelineImageInput`):
+                Input image to be encoded.
+            device: (`torch.device`):
+                Torch device.
+
+        Returns:
+            `torch.Tensor`: The encoded image feature representation.
+        """
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=self.dtype)
+
+        return self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+
+    # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.prepare_ip_adapter_image_embeds
+    def prepare_ip_adapter_image_embeds(
+        self,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+    ) -> torch.Tensor:
+        """Prepares image embeddings for use in the IP-Adapter.
+
+        Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed.
+
+        Args:
+            ip_adapter_image (`PipelineImageInput`, *optional*):
+                The input image to extract features from for IP-Adapter.
+            ip_adapter_image_embeds (`torch.Tensor`, *optional*):
+                Precomputed image embeddings.
+            device: (`torch.device`, *optional*):
+                Torch device.
+            num_images_per_prompt (`int`, defaults to 1):
+                Number of images that should be generated per prompt.
+            do_classifier_free_guidance (`bool`, defaults to True):
+                Whether to use classifier free guidance or not.
+        """
+        device = device or self._execution_device
+
+        if ip_adapter_image_embeds is not None:
+            if do_classifier_free_guidance:
+                single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2)
+            else:
+                single_image_embeds = ip_adapter_image_embeds
+        elif ip_adapter_image is not None:
+            single_image_embeds = self.encode_image(ip_adapter_image, device)
+            if do_classifier_free_guidance:
+                single_negative_image_embeds = torch.zeros_like(single_image_embeds)
+        else:
+            raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.")
+
+        image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
+
+        if do_classifier_free_guidance:
+            negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0)
+            image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0)
+
+        return image_embeds.to(device=device)
+
+    def enable_sequential_cpu_offload(self, *args, **kwargs):
+        if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload:
+            logger.warning(
+                "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses "
+                "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling "
+                "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`."
+            )
+
+        super().enable_sequential_cpu_offload(*args, **kwargs)
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        prompt_3: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 28,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: float = 7.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        cond_latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 256,
+        skip_guidance_layers: List[int] = None,
+        skip_layer_guidance_scale: float = 2.8,
+        skip_layer_guidance_stop: float = 0.2,
+        skip_layer_guidance_start: float = 0.01,
+        mu: Optional[float] = None,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                will be used instead
+            prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
+                will be used instead
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
+                `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
+                the text `prompt`, usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used instead
+            negative_prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
+                `text_encoder_3`. If not defined, `negative_prompt` is used instead
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            ip_adapter_image (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`torch.Tensor`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images,
+                emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to
+                `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of
+                a plain tuple.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`.
+            skip_guidance_layers (`List[int]`, *optional*):
+                A list of integers that specify layers to skip during guidance. If not provided, all layers will be
+                used for guidance. If provided, the guidance will only be applied to the layers specified in the list.
+                Recommended value by StabiltyAI for Stable Diffusion 3.5 Medium is [7, 8, 9].
+            skip_layer_guidance_scale (`int`, *optional*): The scale of the guidance for the layers specified in
+                `skip_guidance_layers`. The guidance will be applied to the layers specified in `skip_guidance_layers`
+                with a scale of `skip_layer_guidance_scale`. The guidance will be applied to the rest of the layers
+                with a scale of `1`.
+            skip_layer_guidance_stop (`int`, *optional*): The step at which the guidance for the layers specified in
+                `skip_guidance_layers` will stop. The guidance will be applied to the layers specified in
+                `skip_guidance_layers` until the fraction specified in `skip_layer_guidance_stop`. Recommended value by
+                StabiltyAI for Stable Diffusion 3.5 Medium is 0.2.
+            skip_layer_guidance_start (`int`, *optional*): The step at which the guidance for the layers specified in
+                `skip_guidance_layers` will start. The guidance will be applied to the layers specified in
+                `skip_guidance_layers` from the fraction specified in `skip_layer_guidance_start`. Recommended value by
+                StabiltyAI for Stable Diffusion 3.5 Medium is 0.01.
+            mu (`float`, *optional*): `mu` value used for `dynamic_shifting`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            prompt_3,
+            height,
+            width,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            negative_prompt_3=negative_prompt_3,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._skip_layer_guidance_scale = skip_layer_guidance_scale
+        self._clip_skip = clip_skip
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_3=prompt_3,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            negative_prompt_3=negative_prompt_3,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            device=device,
+            clip_skip=self.clip_skip,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+
+        if self.do_classifier_free_guidance:
+            if skip_guidance_layers is not None:
+                original_prompt_embeds = prompt_embeds
+                original_pooled_prompt_embeds = pooled_prompt_embeds
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 5. Prepare timesteps
+        scheduler_kwargs = {}
+        if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None:
+            _, _, height, width = latents.shape
+            image_seq_len = (height // self.transformer.config.patch_size) * (
+                width // self.transformer.config.patch_size
+            )
+            mu = calculate_shift(
+                image_seq_len,
+                self.scheduler.config.get("base_image_seq_len", 256),
+                self.scheduler.config.get("max_image_seq_len", 4096),
+                self.scheduler.config.get("base_shift", 0.5),
+                self.scheduler.config.get("max_shift", 1.16),
+            )
+            scheduler_kwargs["mu"] = mu
+        elif mu is not None:
+            scheduler_kwargs["mu"] = mu
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            **scheduler_kwargs,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare image embeddings
+        if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None:
+            ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+            if self.joint_attention_kwargs is None:
+                self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds}
+            else:
+                self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds)
+
+        if cond_latents is not None and self.do_classifier_free_guidance:
+            if cond_latents.shape[0] == latents.shape[0]:
+                cond_latents = torch.cat([cond_latents]*2)
+
+        # 7. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    cond_hidden_states=cond_latents,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    pooled_projections=pooled_prompt_embeds,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                    should_skip_layers = (
+                        True
+                        if i > num_inference_steps * skip_layer_guidance_start
+                        and i < num_inference_steps * skip_layer_guidance_stop
+                        else False
+                    )
+                    if skip_guidance_layers is not None and should_skip_layers:
+                        timestep = t.expand(latents.shape[0])
+                        latent_model_input = latents
+                        noise_pred_skip_layers = self.transformer(
+                            hidden_states=latent_model_input,
+                            timestep=timestep,
+                            encoder_hidden_states=original_prompt_embeds,
+                            pooled_projections=original_pooled_prompt_embeds,
+                            joint_attention_kwargs=self.joint_attention_kwargs,
+                            return_dict=False,
+                            skip_layers=skip_guidance_layers,
+                        )[0]
+                        noise_pred = (
+                            noise_pred + (noise_pred_text - noise_pred_skip_layers) * self._skip_layer_guidance_scale
+                        )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            image = latents
+
+        else:
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusion3PipelineOutput(images=image)

src/models/sd3_kontext/pipeline_stable_diffusion_3_dynamic.py

@@ -0,0 +1,1257 @@
+# Copyright 2024 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import (
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    SiglipImageProcessor,
+    SiglipVisionModel,
+    T5EncoderModel,
+    T5TokenizerFast,
+)
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin
+from diffusers.models.autoencoders import AutoencoderKL
+from diffusers.models.transformers import SD3Transformer2DModel
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import StableDiffusion3Pipeline
+
+        >>> pipe = StableDiffusion3Pipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16
+        ... )
+        >>> pipe.to("cuda")
+        >>> prompt = "A cat holding a sign that says hello world"
+        >>> image = pipe(prompt).images[0]
+        >>> image.save("sd3.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin):
+    r"""
+    Args:
+        transformer ([`SD3Transformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant,
+            with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size`
+            as its dimension.
+        text_encoder_2 ([`CLIPTextModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
+        text_encoder_3 ([`T5EncoderModel`]):
+            Frozen text-encoder. Stable Diffusion 3 uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+            [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_3 (`T5TokenizerFast`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        image_encoder (`SiglipVisionModel`, *optional*):
+            Pre-trained Vision Model for IP Adapter.
+        feature_extractor (`SiglipImageProcessor`, *optional*):
+            Image processor for IP Adapter.
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae"
+    _optional_components = ["image_encoder", "feature_extractor"]
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"]
+
+    def __init__(
+        self,
+        transformer: SD3Transformer2DModel,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer_2: CLIPTokenizer,
+        text_encoder_3: T5EncoderModel,
+        tokenizer_3: T5TokenizerFast,
+        image_encoder: SiglipVisionModel = None,
+        feature_extractor: SiglipImageProcessor = None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            text_encoder_3=text_encoder_3,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            tokenizer_3=tokenizer_3,
+            transformer=transformer,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
+        )
+        self.default_sample_size = (
+            self.transformer.config.sample_size
+            if hasattr(self, "transformer") and self.transformer is not None
+            else 128
+        )
+        self.patch_size = (
+            self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2
+        )
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_images_per_prompt: int = 1,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        if self.text_encoder_3 is None:
+            return torch.zeros(
+                (
+                    batch_size * num_images_per_prompt,
+                    self.tokenizer_max_length,
+                    self.transformer.config.joint_attention_dim,
+                ),
+                device=device,
+                dtype=dtype,
+            )
+
+        text_inputs = self.tokenizer_3(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            # logger.warning(
+            #     "The following part of your input was truncated because `max_sequence_length` is set to "
+            #     f" {max_sequence_length} tokens: {removed_text}"
+            # )
+
+        prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0]
+
+        dtype = self.text_encoder_3.dtype
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def _get_clip_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        clip_skip: Optional[int] = None,
+        clip_model_index: int = 0,
+    ):
+        device = device or self._execution_device
+
+        clip_tokenizers = [self.tokenizer, self.tokenizer_2]
+        clip_text_encoders = [self.text_encoder, self.text_encoder_2]
+
+        tokenizer = clip_tokenizers[clip_model_index]
+        text_encoder = clip_text_encoders[clip_model_index]
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            # logger.warning(
+            #     "The following part of your input was truncated because CLIP can only handle sequences up to"
+            #     f" {self.tokenizer_max_length} tokens: {removed_text}"
+            # )
+        prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+        pooled_prompt_embeds = prompt_embeds[0]
+
+        if clip_skip is None:
+            prompt_embeds = prompt_embeds.hidden_states[-2]
+        else:
+            prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1)
+
+        return prompt_embeds, pooled_prompt_embeds
+
+    def encode_pooled_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        clip_skip: Optional[int] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder, lora_scale)
+            if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            _, pooled_prompt_embed = self._get_clip_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=0,
+            )
+            _, pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                prompt=prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=1,
+            )
+
+            pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            _, negative_pooled_prompt_embed = self._get_clip_prompt_embeds(
+                negative_prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=0,
+            )
+            _, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                negative_prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=1,
+            )
+
+            negative_pooled_prompt_embeds = torch.cat(
+                [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]],
+        prompt_3: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        clip_skip: Optional[int] = None,
+        max_sequence_length: int = 256,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in all text-encoders
+            prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
+                used in all text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
+            negative_prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
+                `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder, lora_scale)
+            if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            prompt_3 = prompt_3 or prompt
+            prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3
+
+            prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=0,
+            )
+            prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                prompt=prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=1,
+            )
+            clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1)
+
+            t5_prompt_embed = self._get_t5_prompt_embeds(
+                prompt=prompt_3,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+            clip_prompt_embeds = torch.nn.functional.pad(
+                clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1])
+            )
+
+            prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2)
+            pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+            negative_prompt_3 = negative_prompt_3 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+            negative_prompt_3 = (
+                batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3
+            )
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds(
+                negative_prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=0,
+            )
+            negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                negative_prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=1,
+            )
+            negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1)
+
+            t5_negative_prompt_embed = self._get_t5_prompt_embeds(
+                prompt=negative_prompt_3,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+            negative_clip_prompt_embeds = torch.nn.functional.pad(
+                negative_clip_prompt_embeds,
+                (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]),
+            )
+
+            negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2)
+            negative_pooled_prompt_embeds = torch.cat(
+                [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        prompt_3,
+        height,
+        width,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        negative_prompt_3=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        max_sequence_length=None,
+    ):
+        if (
+            height % (self.vae_scale_factor * self.patch_size) != 0
+            or width % (self.vae_scale_factor * self.patch_size) != 0
+        ):
+            raise ValueError(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}."
+                f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}."
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_3 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+        elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)):
+            raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_3 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if max_sequence_length is not None and max_sequence_length > 512:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def skip_guidance_layers(self):
+        return self._skip_guidance_layers
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def joint_attention_kwargs(self):
+        return self._joint_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_image
+    def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor:
+        """Encodes the given image into a feature representation using a pre-trained image encoder.
+
+        Args:
+            image (`PipelineImageInput`):
+                Input image to be encoded.
+            device: (`torch.device`):
+                Torch device.
+
+        Returns:
+            `torch.Tensor`: The encoded image feature representation.
+        """
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=self.dtype)
+
+        return self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+
+    # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.prepare_ip_adapter_image_embeds
+    def prepare_ip_adapter_image_embeds(
+        self,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+    ) -> torch.Tensor:
+        """Prepares image embeddings for use in the IP-Adapter.
+
+        Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed.
+
+        Args:
+            ip_adapter_image (`PipelineImageInput`, *optional*):
+                The input image to extract features from for IP-Adapter.
+            ip_adapter_image_embeds (`torch.Tensor`, *optional*):
+                Precomputed image embeddings.
+            device: (`torch.device`, *optional*):
+                Torch device.
+            num_images_per_prompt (`int`, defaults to 1):
+                Number of images that should be generated per prompt.
+            do_classifier_free_guidance (`bool`, defaults to True):
+                Whether to use classifier free guidance or not.
+        """
+        device = device or self._execution_device
+
+        if ip_adapter_image_embeds is not None:
+            if do_classifier_free_guidance:
+                single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2)
+            else:
+                single_image_embeds = ip_adapter_image_embeds
+        elif ip_adapter_image is not None:
+            single_image_embeds = self.encode_image(ip_adapter_image, device)
+            if do_classifier_free_guidance:
+                single_negative_image_embeds = torch.zeros_like(single_image_embeds)
+        else:
+            raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.")
+
+        image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
+
+        if do_classifier_free_guidance:
+            negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0)
+            image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0)
+
+        return image_embeds.to(device=device)
+
+    def enable_sequential_cpu_offload(self, *args, **kwargs):
+        if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload:
+            logger.warning(
+                "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses "
+                "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling "
+                "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`."
+            )
+
+        super().enable_sequential_cpu_offload(*args, **kwargs)
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        prompt_3: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 28,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: float = 7.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        cond_latents: Optional[list[torch.FloatTensor]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 256,
+        skip_guidance_layers: List[int] = None,
+        skip_layer_guidance_scale: float = 2.8,
+        skip_layer_guidance_stop: float = 0.2,
+        skip_layer_guidance_start: float = 0.01,
+        mu: Optional[float] = None,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                will be used instead
+            prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
+                will be used instead
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used instead
+            negative_prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
+                `text_encoder_3`. If not defined, `negative_prompt` is used instead
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            ip_adapter_image (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`torch.Tensor`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images,
+                emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to
+                `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of
+                a plain tuple.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`.
+            skip_guidance_layers (`List[int]`, *optional*):
+                A list of integers that specify layers to skip during guidance. If not provided, all layers will be
+                used for guidance. If provided, the guidance will only be applied to the layers specified in the list.
+                Recommended value by StabiltyAI for Stable Diffusion 3.5 Medium is [7, 8, 9].
+            skip_layer_guidance_scale (`int`, *optional*): The scale of the guidance for the layers specified in
+                `skip_guidance_layers`. The guidance will be applied to the layers specified in `skip_guidance_layers`
+                with a scale of `skip_layer_guidance_scale`. The guidance will be applied to the rest of the layers
+                with a scale of `1`.
+            skip_layer_guidance_stop (`int`, *optional*): The step at which the guidance for the layers specified in
+                `skip_guidance_layers` will stop. The guidance will be applied to the layers specified in
+                `skip_guidance_layers` until the fraction specified in `skip_layer_guidance_stop`. Recommended value by
+                StabiltyAI for Stable Diffusion 3.5 Medium is 0.2.
+            skip_layer_guidance_start (`int`, *optional*): The step at which the guidance for the layers specified in
+                `skip_guidance_layers` will start. The guidance will be applied to the layers specified in
+                `skip_guidance_layers` from the fraction specified in `skip_layer_guidance_start`. Recommended value by
+                StabiltyAI for Stable Diffusion 3.5 Medium is 0.01.
+            mu (`float`, *optional*): `mu` value used for `dynamic_shifting`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            prompt_3,
+            height,
+            width,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            negative_prompt_3=negative_prompt_3,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._skip_layer_guidance_scale = skip_layer_guidance_scale
+        self._clip_skip = clip_skip
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_3=prompt_3,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            negative_prompt_3=negative_prompt_3,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            device=device,
+            clip_skip=self.clip_skip,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+
+        if self.do_classifier_free_guidance:
+            if skip_guidance_layers is not None:
+                original_prompt_embeds = prompt_embeds
+                original_pooled_prompt_embeds = pooled_prompt_embeds
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 5. Prepare timesteps
+        scheduler_kwargs = {}
+        if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None:
+            _, _, height, width = latents.shape
+            image_seq_len = (height // self.transformer.config.patch_size) * (
+                width // self.transformer.config.patch_size
+            )
+            mu = calculate_shift(
+                image_seq_len,
+                self.scheduler.config.get("base_image_seq_len", 256),
+                self.scheduler.config.get("max_image_seq_len", 4096),
+                self.scheduler.config.get("base_shift", 0.5),
+                self.scheduler.config.get("max_shift", 1.16),
+            )
+            scheduler_kwargs["mu"] = mu
+        elif mu is not None:
+            scheduler_kwargs["mu"] = mu
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            **scheduler_kwargs,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare image embeddings
+        if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None:
+            ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+            if self.joint_attention_kwargs is None:
+                self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds}
+            else:
+                self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds)
+
+
+        if cond_latents is not None and self.do_classifier_free_guidance:
+            if len(cond_latents) == latents.shape[0]:
+                cond_latents = cond_latents * 2
+
+        # 7. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    cond_hidden_states=cond_latents,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    pooled_projections=pooled_prompt_embeds,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                    should_skip_layers = (
+                        True
+                        if i > num_inference_steps * skip_layer_guidance_start
+                        and i < num_inference_steps * skip_layer_guidance_stop
+                        else False
+                    )
+                    if skip_guidance_layers is not None and should_skip_layers:
+                        timestep = t.expand(latents.shape[0])
+                        latent_model_input = latents
+                        noise_pred_skip_layers = self.transformer(
+                            hidden_states=latent_model_input,
+                            timestep=timestep,
+                            encoder_hidden_states=original_prompt_embeds,
+                            pooled_projections=original_pooled_prompt_embeds,
+                            joint_attention_kwargs=self.joint_attention_kwargs,
+                            return_dict=False,
+                            skip_layers=skip_guidance_layers,
+                        )[0]
+                        noise_pred = (
+                            noise_pred + (noise_pred_text - noise_pred_skip_layers) * self._skip_layer_guidance_scale
+                        )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            image = latents
+
+        else:
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusion3PipelineOutput(images=image)

src/models/sd3_kontext/qwen2_5_vl_sd3_hf_dynamic.py

@@ -0,0 +1,792 @@
+import random
+import torch
+import math
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.modules.module import T
+import torch.distributed as dist
+from mmengine.logging import print_log
+from src.models.connector import ConnectorConfig, ConnectorEncoder
+from xtuner.model.utils import guess_load_checkpoint
+from xtuner.registry import BUILDER
+from diffusers.training_utils import compute_density_for_timestep_sampling, compute_loss_weighting_for_sd3
+from peft import LoraConfig
+from src.models.sd3_kontext.pipeline_stable_diffusion_3_dynamic import StableDiffusion3Pipeline, calculate_shift
+from mmengine.model import BaseModel
+from functools import partial
+from six.moves import map, zip
+from copy import deepcopy
+from einops import rearrange
+
+
+IMAGE_MEAN = (0.48145466, 0.4578275, 0.40821073)
+IMAGE_STD = (0.26862954, 0.26130258, 0.27577711)
+
+
+def multi_apply(func, *args, **kwargs):
+    pfunc = partial(func, **kwargs) if kwargs else func
+    map_results = map(pfunc, *args)
+    return tuple(map(list, zip(*map_results)))
+
+
+def find_target_linear_names(model, num_lora_modules=-1, lora_namespan_exclude=[], verbose=True):
+    linear_cls = torch.nn.modules.Linear
+    embedding_cls = torch.nn.modules.Embedding
+    lora_module_names = []
+
+    for name, module in model.named_modules():
+        if any(ex_keyword in name for ex_keyword in lora_namespan_exclude):
+            continue
+        if isinstance(module, (linear_cls, embedding_cls)):
+            lora_module_names.append(name)
+    
+    if num_lora_modules > 0:
+        lora_module_names = lora_module_names[-num_lora_modules:]
+    if verbose:
+        print(f"Found {len(lora_module_names)} lora modules: {lora_module_names}")
+    return lora_module_names
+
+
+class Qwen2p5VLStableDiffusion3HF(BaseModel):
+    def __init__(self,
+                 transformer,
+                 train_scheduler,
+                 test_scheduler,
+                 vae,
+                 lmm,
+                 tokenizer,
+                 prompt_template,
+                 connector,
+                 num_queries=64,
+                 vit_input_size=448,
+                 max_length=1024,
+                 freeze_lmm=True,
+                 freeze_mq=False,
+                 res_vit=False,
+                 pretrained_pth=None,
+                 use_activation_checkpointing=False,
+                 lora_modules='auto',  # ["to_k", "to_q", "to_v", "to_out.0"], 'auto'
+                 lora_rank=64,
+                 lora_alpha=128,
+                 freeze_transformer=True,
+                 unconditional=0.1,
+                 ema_cfg=None,
+                 weighting_scheme='none',
+                 logit_mean=0.0,
+                 logit_std=1.0,
+                 ):
+        super().__init__()
+
+        self.lmm = BUILDER.build(lmm)
+        if freeze_lmm:
+            self.lmm.requires_grad_(False)
+        self.freeze_lmm = freeze_lmm
+
+        self.transformer = BUILDER.build(transformer)
+        if freeze_transformer:
+            self.transformer.requires_grad_(False)
+        self.freeze_transformer = freeze_transformer
+        self.res_vit = res_vit
+
+
+
+        self.weighting_scheme = weighting_scheme
+        self.logit_mean = logit_mean
+        self.logit_std = logit_std
+
+        self.vae = BUILDER.build(vae)
+        self.vae.requires_grad_(False)
+
+        self.use_activation_checkpointing = use_activation_checkpointing
+        self.tokenizer = BUILDER.build(tokenizer)
+
+        self.prompt_template = prompt_template
+        self.vit_input_size = vit_input_size
+        self.max_length = max_length
+        self.image_token_id = self.tokenizer.convert_tokens_to_ids(prompt_template['IMG_CONTEXT_TOKEN'])
+        self.register_buffer('vit_mean', torch.tensor(IMAGE_MEAN), persistent=False)
+        self.register_buffer('vit_std', torch.tensor(IMAGE_STD), persistent=False)
+
+        self.num_queries = num_queries
+        self.connector = ConnectorEncoder(ConnectorConfig(**connector))
+
+
+        self.projector_1 = nn.Linear(self.llm.config.hidden_size, self.connector.config.hidden_size)
+        self.projector_2 = nn.Linear(self.connector.config.hidden_size, self.transformer.config.pooled_projection_dim)
+        self.projector_3 = nn.Linear(self.connector.config.hidden_size, self.transformer.config.joint_attention_dim)
+
+        # zero out
+        nn.init.zeros_(self.projector_2.weight)
+        nn.init.zeros_(self.projector_3.weight)
+        nn.init.zeros_(self.projector_2.bias)
+        nn.init.zeros_(self.projector_3.bias)
+
+        self.meta_queries = nn.Parameter(
+            torch.zeros(num_queries, self.llm.config.hidden_size))
+        nn.init.normal_(self.meta_queries, std=1 / math.sqrt(self.llm.config.hidden_size))
+
+        if freeze_mq:
+            self.projector_1.requires_grad_(False)
+            self.projector_2.requires_grad_(False)
+            self.projector_3.requires_grad_(False)
+            self.connector.requires_grad_(False)
+            self.meta_queries.requires_grad_(False)
+        self.freeze_mq = freeze_mq
+
+        self.unconditional = unconditional
+
+        self.train_scheduler = BUILDER.build(train_scheduler)
+        self.test_scheduler = BUILDER.build(test_scheduler)
+
+        if use_activation_checkpointing:
+            self.gradient_checkpointing_enable()
+
+        if lora_modules is not None:
+            assert self.freeze_lmm
+            self.llm.config.tie_word_embeddings = False
+            if lora_modules == 'auto':
+                lora_modules = find_target_linear_names(self.lmm)
+            # now we will add new LoRA weights the transformer layers
+            transformer_lora_config = LoraConfig(
+                r=lora_rank,
+                lora_alpha=lora_alpha,
+                init_lora_weights="gaussian",
+                target_modules=lora_modules,
+                lora_dropout=0.05,
+            )
+            self.lmm.add_adapter(transformer_lora_config)
+
+        if pretrained_pth is not None:
+            pretrained_state_dict = guess_load_checkpoint(pretrained_pth)
+            info = self.load_state_dict(pretrained_state_dict, strict=False)
+            print_log(f'Load pretrained weight from {pretrained_pth}')
+
+        self.ema_cfg = ema_cfg
+        if ema_cfg is not None:
+            self.ema = nn.ModuleDict()
+            self.ema.steps = 0
+            if not self.freeze_transformer:
+                self.ema.update(dict(transformer=deepcopy(self.transformer)))
+
+            if not self.freeze_mq:
+                self.ema.update(dict(projector_1=deepcopy(self.projector_1),
+                                     projector_2=deepcopy(self.projector_2),
+                                     projector_3=deepcopy(self.projector_3),
+                                     connector=deepcopy(self.connector)
+                                     )
+                                )
+                self.ema.register_buffer('meta_queries', deepcopy(self.meta_queries.data))
+
+            self.ema.requires_grad_(False)   # parameters in ema are not learnable
+
+            if 'checkpoint' in ema_cfg:
+                ema_state_dict = guess_load_checkpoint(ema_cfg['checkpoint'])
+                info = self.ema.load_state_dict(ema_state_dict, strict=False)
+                print_log(f"Load ema weight from {ema_cfg['checkpoint']}")
+
+    @torch.no_grad()
+    def ema_step(self, ):
+        if self.ema_cfg is None:
+            return
+
+        steps = self.ema.steps
+        update_interval = self.ema_cfg.get('update_interval', 1)
+        save_interval = self.ema_cfg.get('save_interval', 1000)
+        momentum = self.ema_cfg.get('momentum', 0.99)
+
+        if steps % update_interval == 0 and steps > 0:
+            if not self.freeze_mq:
+                for ema_param, base_param in zip(self.ema.projector_1.parameters(), self.projector_1.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+                for ema_param, base_param in zip(self.ema.projector_2.parameters(), self.projector_2.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+                for ema_param, base_param in zip(self.ema.projector_3.parameters(), self.projector_3.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+                for ema_param, base_param in zip(self.ema.connector.parameters(), self.connector.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+                self.ema.meta_queries.data.lerp_(self.meta_queries.data.detach(), 1.0 - momentum)
+
+            if not self.freeze_transformer:
+                for ema_param, base_param in zip(self.ema.transformer.parameters(), self.transformer.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+
+        # print(f"steps: {steps}, rank: {dist.get_rank()},", flush=True)
+
+        if steps % save_interval == 0 and steps > 0:
+            is_ddp = dist.is_available() and dist.is_initialized()
+            is_primary_proc = (not is_ddp) or dist.get_rank() == 0
+            print(f"steps: {steps}, rank: {dist.get_rank()}, is_ddp:{is_ddp}, is_primary_proc: {is_primary_proc}.", flush=True)
+            if is_primary_proc:
+                save_path = self.ema_cfg.get('save_path')
+                torch.save(self.ema.state_dict(), save_path)
+            if is_ddp:
+                dist.barrier()
+
+        self.ema.steps = self.ema.steps + 1
+
+    def llm2dit(self, x):
+        x = self.connector(self.projector_1(x))
+        pooled_out = self.projector_2(x.mean(1))
+        seq_out = self.projector_3(x)
+
+        return pooled_out, seq_out
+
+    @property
+    def llm(self):
+        return self.lmm.language_model
+
+    def gradient_checkpointing_enable(self):
+        self.activation_checkpointing_enable()
+
+    def activation_checkpointing_enable(self):
+        self.llm.gradient_checkpointing_enable()
+        self.transformer.enable_gradient_checkpointing()
+        self.connector.gradient_checkpointing = True
+
+    def gradient_checkpointing_disable(self):
+        self.activation_checkpointing_disable()
+
+    def activation_checkpointing_disable(self):
+        self.llm.gradient_checkpointing_disable()
+        self.transformer.disable_gradient_checkpointing()
+        self.connector.gradient_checkpointing = False
+
+    @property
+    def device(self):
+        return self.llm.device
+
+    @property
+    def dtype(self):
+        return self.llm.dtype
+
+    def train(self: T, mode: bool = True) -> T:
+        super().train(mode=mode)
+        if self.vae is not None:
+            self.vae.train(mode=False)
+        if not mode:
+            self.gradient_checkpointing_disable()
+
+        return self
+
+    def state_dict(self, *args, **kwargs) -> dict:
+        state_dict = super().state_dict(*args, **kwargs)
+        state_dict = {k: v for k, v in state_dict.items()
+                      if 'vae.' not in k and 'lmm.' not in k and 'ema.' not in k}
+        return state_dict
+
+    @torch.no_grad()
+    def pixels_to_latents(self, x):
+        z = self.vae.encode(x).latent_dist.sample()
+        z = (z - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+        return z
+
+    @torch.no_grad()
+    def latents_to_pixels(self, z):
+        z = (z / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+        x_rec = self.vae.decode(z).sample
+        return x_rec
+
+    def forward(self, data, data_samples=None, mode='loss'):
+        if mode == 'loss':
+            self.ema_step()
+            return self.compute_loss(data_dict=data)
+        else:
+            raise NotImplementedError
+
+    def compute_loss(self, data_dict):
+        losses = {}
+        for data_type in ['text2image', 'image2image']:
+            if data_type in data_dict:
+                losses[f'loss_{data_type}'] = getattr(self, f'{data_type}_loss')(data_dict[data_type])
+        if len(losses) == 0:
+            if 'pixel_values_src' in data_dict:
+                losses[f'loss_image2image'] = self.image2image_loss(data_dict)
+            else:
+                losses[f'loss_text2image'] = self.text2image_loss(data_dict)
+
+        return losses
+
+
+    def prepare_forward_input(self,
+                              query_embeds,
+                              input_ids=None,
+                              image_embeds=None,
+                              image_grid_thw=None,
+                              attention_mask=None,
+                              past_key_values=None):
+        b, l, _ = query_embeds.shape
+        assert l > 0
+        attention_mask = attention_mask.to(device=self.device, dtype=torch.bool)
+
+        assert l == self.num_queries
+
+        input_ids = torch.cat([input_ids, input_ids.new_zeros(b, l)], dim=1)
+        attention_mask = torch.cat([attention_mask, attention_mask.new_ones(b, l)], dim=1)
+
+        position_ids, _ = self.lmm.model.get_rope_index(
+            input_ids=input_ids,
+            image_grid_thw=image_grid_thw,
+            video_grid_thw=None,
+            second_per_grid_ts=None,
+            attention_mask=attention_mask,
+        )
+
+        # prepare context
+        if past_key_values is not None:
+            inputs_embeds = query_embeds
+            position_ids = position_ids[..., -l:]
+        else:
+            input_ids = input_ids[:, :-l]   # context input_ids
+
+            if image_embeds is None:
+                inputs_embeds = self.llm.get_input_embeddings()(input_ids)
+            else:
+                inputs_embeds = torch.zeros(*input_ids.shape, self.llm.config.hidden_size,
+                                            device=self.device, dtype=self.dtype)
+                inputs_embeds[input_ids == self.image_token_id] = \
+                    image_embeds.contiguous().view(-1, self.llm.config.hidden_size)
+                inputs_embeds[input_ids != self.image_token_id] = self.llm.get_input_embeddings()(
+                    input_ids[input_ids != self.image_token_id]
+                )
+
+            inputs_embeds = torch.cat([inputs_embeds, query_embeds], dim=1)
+
+        inputs = dict(inputs_embeds=inputs_embeds,
+                      attention_mask=attention_mask,
+                      position_ids=position_ids,
+                      past_key_values=past_key_values)
+
+        return inputs
+
+
+    @torch.no_grad()
+    def get_semantic_features_dynamic(self, pixel_values):
+        # e.g., 512 -> 448
+        pixel_values = [F.interpolate(p[None], scale_factor=28 / 32, mode='bilinear') for p in pixel_values]
+        image_embeds, image_grid_thw = multi_apply(self.get_semantic_features,
+                                                   pixel_values, resize=False)
+        image_embeds = [x[0] for x in image_embeds]    # a list of embeds
+        image_grid_thw = torch.cat(image_grid_thw, dim=0)   # b 3
+
+        return image_embeds, image_grid_thw
+
+    @torch.no_grad()
+    def get_semantic_features(self, pixel_values, resize=True):
+        # pixel_values: [-1, 1]
+        pixel_values = (pixel_values + 1.0) / 2     # [0, 1]
+        pixel_values = pixel_values - self.vit_mean.view(1, 3, 1, 1)
+        pixel_values = pixel_values / self.vit_std.view(1, 3, 1, 1)
+
+        if resize:
+           pixel_values = F.interpolate(pixel_values, size=(self.vit_input_size, self.vit_input_size),
+                                        mode='bilinear')
+        b, c, h, w = pixel_values.shape
+
+        patch_size = self.lmm.config.vision_config.patch_size
+        spatial_merge_size = self.lmm.config.vision_config.spatial_merge_size
+        temporal_patch_size = self.lmm.config.vision_config.temporal_patch_size
+
+        pixel_values = pixel_values[:, None].expand(b, temporal_patch_size, c, h, w)
+
+        grid_t = 1
+        grid_h, grid_w = h // patch_size, w // patch_size
+
+        pixel_values = pixel_values.view(
+            b,
+            grid_t,
+            temporal_patch_size,
+            c,
+            grid_h // spatial_merge_size,
+            spatial_merge_size,
+            patch_size,
+            grid_w // spatial_merge_size,
+            spatial_merge_size,
+            patch_size,
+        )
+
+        pixel_values = rearrange(
+            pixel_values, 'b t tp c h m p w n q -> (b t h w m n) (c tp p q)')
+
+        image_grid_thw = torch.tensor([(grid_t, grid_h, grid_w)] * b).to(self.device).long()
+
+        image_embeds = self.lmm.visual(pixel_values, grid_thw=image_grid_thw)
+        image_embeds = rearrange(image_embeds, '(b l) d -> b l d', b=b)
+
+        return image_embeds, image_grid_thw
+
+    @torch.no_grad()
+    def prepare_text2image_prompts(self, texts):
+        texts = [self.prompt_template['GENERATION'].format(input=text) for text in texts]
+        texts = [self.prompt_template['INSTRUCTION'].format(input=text) for text in texts]
+
+        return self.tokenizer(
+            texts, add_special_tokens=True, return_tensors='pt', padding=True, padding_side='left').to(self.device)
+
+    @torch.no_grad()
+    def prepare_image2image_prompts(self, texts, num_refs, ref_lens):
+        prompts = []
+        cnt = 0
+        for text, num_ref in zip(texts, num_refs):
+            image_tokens = ''
+            for _ in range(num_ref):
+                image_tokens +=  self.prompt_template['IMG_START_TOKEN'] + \
+                                 self.prompt_template['IMG_CONTEXT_TOKEN'] * ref_lens[cnt] + \
+                                 self.prompt_template['IMG_END_TOKEN']
+                cnt += 1
+
+            prompts.append(self.prompt_template['INSTRUCTION'].format(input=f'{image_tokens}\n{text}'))
+
+        return self.tokenizer(
+            prompts, add_special_tokens=True, return_tensors='pt', padding=True, padding_side='left').to(self.device)
+
+
+    def text2image_loss(self, data_dict):
+        # obtain image latents
+        if 'image_latents' in data_dict:
+            image_latents = data_dict['image_latents']   # .to(dtype=self.dtype, device=self.device)
+            image_latents = [x.to(dtype=self.dtype, device=self.device) for x in image_latents]
+        else:
+            pixel_values = [p.to(dtype=self.dtype, device=self.device) for p in data_dict['pixel_values']]
+            image_latents = [self.pixels_to_latents(p[None])[0] for p in pixel_values]
+
+        b = len(image_latents)
+
+        texts = ['' if random.uniform(0, 1) < self.unconditional else text
+                 for text in data_dict['texts']]
+
+        text_inputs = self.prepare_text2image_prompts(texts)
+        hidden_states = self.meta_queries[None].expand(b, self.num_queries, -1)
+
+        inputs = self.prepare_forward_input(query_embeds=hidden_states, **text_inputs)
+
+        max_length = self.max_length + self.num_queries
+        inputs_embeds = inputs['inputs_embeds'][:, -max_length:]
+        attention_mask = inputs['attention_mask'][:, -max_length:]
+        position_ids = inputs['position_ids'][..., -max_length:]
+
+        output = self.llm(inputs_embeds=inputs_embeds,
+                          attention_mask=attention_mask,
+                          position_ids=position_ids,
+                          return_dict=True)
+
+        # hidden_states = output.last_hidden_state[:, -self.num_queries:]#query only
+        hidden_states = output.last_hidden_state
+        pooled_out, seq_out = self.llm2dit(hidden_states)
+
+        loss_diff = self.diff_loss(model_input=image_latents,
+                                   pooled_prompt_embeds=pooled_out,
+                                   prompt_embeds=seq_out)
+
+        return loss_diff
+
+
+    def image2image_loss(self, data_dict):
+        pixel_values_src = data_dict['pixel_values_src']
+
+        num_refs = [len(ref_images) for ref_images in pixel_values_src]
+
+        pixel_values_src = [[img.to(dtype=self.dtype, device=self.device) for img in ref_images]
+                            for ref_images in pixel_values_src]
+        image_latents_src = [[self.pixels_to_latents(img[None])[0] for img in ref_images]
+                             for ref_images in pixel_values_src]
+        image_embeds, image_grid_thw = self.get_semantic_features_dynamic(
+            [img for ref_images in pixel_values_src for img in ref_images])
+
+        ref_lens = [len(x) for x in image_embeds]
+
+        pixel_values = [p.to(dtype=self.dtype, device=self.device) for p in data_dict['pixel_values']]
+        image_latents = [self.pixels_to_latents(p[None])[0] for p in pixel_values]
+
+        b = len(image_latents)
+        text_inputs = self.prepare_image2image_prompts(data_dict['texts'], num_refs=num_refs, ref_lens=ref_lens)
+
+        hidden_states = self.meta_queries[None].expand(b, self.num_queries, -1)
+
+        inputs = self.prepare_forward_input(query_embeds=hidden_states,
+                                            image_embeds=torch.cat(image_embeds),
+                                            image_grid_thw=image_grid_thw,
+                                            **text_inputs)
+
+
+        max_length = self.max_length + max(num_refs) * max(ref_lens) + self.num_queries
+        inputs_embeds = inputs['inputs_embeds'][:, -max_length:]
+        attention_mask = inputs['attention_mask'][:, -max_length:]
+        position_ids = inputs['position_ids'][..., -max_length:]
+
+        output = self.llm(inputs_embeds=inputs_embeds,
+                          attention_mask=attention_mask,
+                          position_ids=position_ids,
+                          return_dict=True)
+
+        # hidden_states = output.last_hidden_state[:, -self.num_queries:] #query only
+        hidden_states = output.last_hidden_state
+
+        # if res_vit:
+        #     image_embeds=torch.cat(image_embeds)
+        #     hidden_states[input_ids == self.image_token_id] = 0.5 * (hidden_states[input_ids == self.image_token_id]) + 0.5 * (image_embeds.contiguous().view(-1, self.llm.config.hidden_size))
+
+
+        pooled_out, seq_out = self.llm2dit(hidden_states)
+
+
+        loss_diff = self.diff_loss(model_input=image_latents,
+                                   pooled_prompt_embeds=pooled_out,
+                                   prompt_embeds=seq_out,
+                                   cond_intput=image_latents_src)
+
+        return loss_diff
+
+
+
+    @torch.no_grad()
+    def generate(self,
+                 prompt,
+                 cfg_prompt,
+                 pixel_values_src=None,
+                 cfg_scale=4.5,
+                 num_steps=50,
+                 generator=None,
+                 height=512,
+                 width=512,
+                 progress_bar=True):
+        assert len(prompt) == len(cfg_prompt)
+        b = len(prompt)
+
+        if pixel_values_src is not None:
+            num_refs = [len(ref_images) for ref_images in pixel_values_src]
+            pixel_values_src = [[img.to(dtype=self.dtype, device=self.device) for img in ref_imgs]
+                                for ref_imgs in pixel_values_src]
+            image_embeds, image_grid_thw = self.get_semantic_features_dynamic(
+                [img for ref_images in pixel_values_src for img in ref_images])
+            ref_lens = [len(x) for x in image_embeds]
+
+            text_inputs = self.prepare_image2image_prompts(prompt + cfg_prompt, num_refs=num_refs*2, ref_lens=ref_lens*2)
+            text_inputs.update(image_embeds=torch.cat(image_embeds*2),
+                               image_grid_thw=torch.cat([image_grid_thw]*2),)
+            cond_latents = [[self.pixels_to_latents(img[None])[0] for img in ref_imgs]
+                            for ref_imgs in pixel_values_src]
+            cond_latents = cond_latents * 2
+        else:
+            text_inputs = self.prepare_text2image_prompts(prompt + cfg_prompt)
+            cond_latents = None
+
+        hidden_states = self.meta_queries[None].expand(2*b, self.num_queries, -1)
+        inputs = self.prepare_forward_input(query_embeds=hidden_states, **text_inputs)
+
+        output = self.llm(**inputs, return_dict=True)
+
+        # hidden_states = output.last_hidden_state[:, -self.num_queries:] #query only
+        hidden_states = output.last_hidden_state 
+        pooled_out, seq_out = self.llm2dit(hidden_states)
+
+        pipeline = StableDiffusion3Pipeline(
+            transformer=self.transformer,
+            scheduler=self.test_scheduler,
+            vae=self.vae,
+            text_encoder=None,
+            tokenizer=None,
+            text_encoder_2=None,
+            tokenizer_2=None,
+            text_encoder_3=None,
+            tokenizer_3=None,
+        )
+
+        pipeline.set_progress_bar_config(disable=not progress_bar)
+
+        samples = pipeline(
+            height=height,
+            width=width,
+            guidance_scale=cfg_scale,
+            num_inference_steps=num_steps,
+            prompt_embeds=seq_out[:b],
+            pooled_prompt_embeds=pooled_out[:b],
+            negative_prompt_embeds=seq_out[b:],
+            negative_pooled_prompt_embeds=pooled_out[b:],
+            generator=generator,
+            output_type='latent',
+            cond_latents=cond_latents
+        ).images.to(self.dtype)
+
+        return self.latents_to_pixels(samples)
+
+    def diff_loss(self, model_input, pooled_prompt_embeds, prompt_embeds, cond_intput=None):
+        # Sample noise that we'll add to the latents
+        # import pdb; pdb.set_trace()
+        noise = [torch.randn_like(x) for x in model_input]
+        bsz = len(model_input)
+
+        u = compute_density_for_timestep_sampling(
+            weighting_scheme=self.weighting_scheme,
+            batch_size=bsz,
+            logit_mean=self.logit_mean,
+            logit_std=self.logit_std,
+        )
+
+        if self.train_scheduler.use_dynamic_shifting:
+            assert self.weighting_scheme == 'logit_normal'
+            # follow flux
+            # import pdb; pdb.set_trace()
+            image_seq_lens = [math.prod(x.shape[-2:]) // self.transformer.patch_size ** 2 for x in model_input]
+            mu = calculate_shift(
+                torch.tensor(image_seq_lens, dtype=self.dtype, device=self.device),
+                self.train_scheduler.config.get("base_image_seq_len", 256),
+                self.train_scheduler.config.get("max_image_seq_len", 4096),
+                self.train_scheduler.config.get("base_shift", 0.5),
+                self.train_scheduler.config.get("max_shift", 1.15)
+            )
+
+            if self.train_scheduler.config.time_shift_type == "exponential":
+                shift = torch.exp(mu)
+            elif self.train_scheduler.config.time_shift_type == "linear":
+                shift = mu
+            else:
+                raise NotImplementedError
+
+            sigmas = u.to(dtype=self.dtype, device=self.device)
+            sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)
+            timesteps = sigmas * self.train_scheduler.num_train_timesteps
+            sigmas = sigmas.view(-1, 1, 1, 1)
+
+        else:
+            # Sample a random timestep for each image
+            # for weighting schemes where we sample timesteps non-uniformly
+            indices = (u * self.train_scheduler.config.num_train_timesteps).long()
+            timesteps = self.train_scheduler.timesteps[indices].to(device=self.device)
+
+            # Add noise according to flow matching.
+            # zt = (1 - texp) * x + texp * z1
+            sigmas = self.get_sigmas(timesteps, n_dim=model_input[0].ndim + 1)
+
+        noisy_model_input = [(1.0 - x) * y + x * z  for x, y, z in zip(sigmas, model_input, noise)]
+
+        # Predict the noise residual
+        model_pred = self.transformer(
+            hidden_states=noisy_model_input,
+            cond_hidden_states=cond_intput,
+            encoder_hidden_states=prompt_embeds,
+            pooled_projections=pooled_prompt_embeds,
+            timestep=timesteps,
+            return_dict=False,
+        )[0]
+
+
+        # these weighting schemes use a uniform timestep sampling
+        # and instead post-weight the loss
+        weighting = compute_loss_weighting_for_sd3(weighting_scheme=self.weighting_scheme, sigmas=sigmas)
+
+        # flow matching loss
+        # target = noise - model_input
+        target = [x - y for x, y in zip(noise, model_input)]
+
+        loss = [(x.float() * (y.float() - z.float()) ** 2).mean() for x, y, z in zip(weighting, model_pred, target)]
+        loss = sum(loss) / len(loss)
+
+        return loss
+
+    def get_sigmas(self, timesteps, n_dim=4):
+        sigmas = self.train_scheduler.sigmas.to(device=self.device, dtype=self.dtype)
+        schedule_timesteps = self.train_scheduler.timesteps.to(self.device)
+        timesteps = timesteps.to(self.device)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+        sigma = sigmas[step_indices].flatten()
+        while len(sigma.shape) < n_dim:
+            sigma = sigma.unsqueeze(-1)
+        return sigma
+
+
+def resize_image(x, image_size, unit_image_size=32):
+    w, h = x.size
+    if w >= h and w >= image_size:
+        target_w = image_size
+        target_h = h * (target_w / w)
+        target_h = math.ceil(target_h / unit_image_size) * unit_image_size
+
+    elif h >= w and h >= image_size:
+        target_h = image_size
+        target_w = w * (target_h / h)
+        target_w = math.ceil(target_w / unit_image_size) * unit_image_size
+
+    else:
+        target_h = math.ceil(h / unit_image_size) * unit_image_size
+        target_w = math.ceil(w / unit_image_size) * unit_image_size
+
+    x = x.resize(size=(target_w, target_h))
+
+    return x
+
+
+if __name__ == "__main__":
+    import os
+    import argparse
+    from glob import glob
+    from mmengine.config import Config
+    from PIL import Image
+    import numpy as np
+
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('config', help='log file path.')
+    parser.add_argument("--checkpoint", type=str, default=None)
+    parser.add_argument("--image", type=str, default=None)
+    parser.add_argument("--prompt", type=str, default='a dog on the left and a cat on the right')
+    parser.add_argument("--cfg_prompt", type=str, default='')
+    parser.add_argument("--cfg_scale", type=float, default=4.0)
+    parser.add_argument("--num_steps", type=int, default=50)
+    parser.add_argument("--height", type=int, default=512)
+    parser.add_argument("--width", type=int, default=512)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--grid_size", type=int, default=2)
+    parser.add_argument('--output', type=str, default='output.jpg')
+
+    args = parser.parse_args()
+    config = Config.fromfile(args.config)
+    model = BUILDER.build(config.model).cuda().bfloat16().eval()
+
+    if args.checkpoint is not None:
+        print(f"Load checkpoint: {args.checkpoint}", flush=True)
+        checkpoint = guess_load_checkpoint(args.checkpoint)
+        info = model.load_state_dict(checkpoint, strict=False)
+
+    generator = torch.Generator(device=model.device).manual_seed(args.seed)
+    # repeat
+    bsz = args.grid_size ** 2
+
+    prompt = [args.prompt] * bsz
+    cfg_prompt = [args.cfg_prompt] * bsz
+
+    if args.image is not None:
+
+        if os.path.isdir(args.image):
+            ref_images = glob(f"{args.image}/*")
+            ref_images = [Image.open(path) for path in ref_images]
+        else:
+            ref_images = [Image.open(args.image)]
+
+        ref_images = [resize_image(img, max(args.width, args.height), 32) for img in ref_images]
+
+        if len(ref_images) == 1:
+            width, height = ref_images[0].size
+        else:
+            width, height = args.width, args.height
+
+        pixel_values_src = [torch.from_numpy(np.array(img)).to(dtype=model.dtype, device=model.device)
+                            for img in ref_images]
+        pixel_values_src = [rearrange(img, 'h w c -> c h w') for img in pixel_values_src]
+        pixel_values_src = [2 * (img / 255) - 1 for img in pixel_values_src]
+
+        pixel_values_src = [pixel_values_src, ] * bsz
+    else:
+        width, height = args.width, args.height
+        pixel_values_src = None
+
+    samples = model.generate(prompt=prompt, cfg_prompt=cfg_prompt, pixel_values_src=pixel_values_src,
+                             cfg_scale=args.cfg_scale, num_steps=args.num_steps,
+                             generator=generator, height=height, width=width)
+
+
+    samples = rearrange(samples, '(m n) c h w -> (m h) (n w) c', m=args.grid_size, n=args.grid_size)
+    samples = torch.clamp(
+        127.5 * samples + 128.0, 0, 255).to("cpu", dtype=torch.uint8).numpy()
+
+    Image.fromarray(samples).save(args.output)

src/models/sd3_kontext/qwen2_5_vl_sd3_hf_dynamic_fusion.py

@@ -0,0 +1,824 @@
+import random
+import torch
+import math
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.modules.module import T
+import torch.distributed as dist
+from mmengine.logging import print_log
+from src.models.connector import ConnectorConfig, ConnectorEncoder
+from xtuner.model.utils import guess_load_checkpoint
+from xtuner.registry import BUILDER
+from diffusers.training_utils import compute_density_for_timestep_sampling, compute_loss_weighting_for_sd3
+from peft import LoraConfig
+from src.models.sd3_kontext.pipeline_stable_diffusion_3_dynamic import StableDiffusion3Pipeline, calculate_shift
+from mmengine.model import BaseModel
+from functools import partial
+from six.moves import map, zip
+from copy import deepcopy
+from einops import rearrange
+
+
+IMAGE_MEAN = (0.48145466, 0.4578275, 0.40821073)
+IMAGE_STD = (0.26862954, 0.26130258, 0.27577711)
+
+
+def multi_apply(func, *args, **kwargs):
+    pfunc = partial(func, **kwargs) if kwargs else func
+    map_results = map(pfunc, *args)
+    return tuple(map(list, zip(*map_results)))
+
+
+def find_target_linear_names(model, num_lora_modules=-1, lora_namespan_exclude=[], verbose=True):
+    linear_cls = torch.nn.modules.Linear
+    embedding_cls = torch.nn.modules.Embedding
+    lora_module_names = []
+
+    for name, module in model.named_modules():
+        if any(ex_keyword in name for ex_keyword in lora_namespan_exclude):
+            continue
+        if isinstance(module, (linear_cls, embedding_cls)):
+            lora_module_names.append(name)
+    
+    if num_lora_modules > 0:
+        lora_module_names = lora_module_names[-num_lora_modules:]
+    if verbose:
+        print(f"Found {len(lora_module_names)} lora modules: {lora_module_names}")
+    return lora_module_names
+
+
+class Qwen2p5VLStableDiffusion3HF(BaseModel):
+    def __init__(self,
+                 transformer,
+                 train_scheduler,
+                 test_scheduler,
+                 vae,
+                 lmm,
+                 tokenizer,
+                 prompt_template,
+                 connector,
+                 num_queries=64,
+                 vit_input_size=448,
+                 max_length=1024,
+                 freeze_lmm=True,
+                 freeze_mq=False,
+                 res_vit=False,
+                 pretrained_pth=None,
+                 use_activation_checkpointing=False,
+                 lora_modules='auto',  
+                 lora_rank=64,
+                 lora_alpha=128,
+                 freeze_transformer=True,
+                 unconditional=0.1,
+                 ema_cfg=None,
+                 weighting_scheme='none',
+                 logit_mean=0.0,
+                 logit_std=1.0,
+                 ):
+        super().__init__()
+
+        self.lmm = BUILDER.build(lmm)
+        if freeze_lmm:
+            self.lmm.requires_grad_(False)
+        self.freeze_lmm = freeze_lmm
+
+        self.transformer = BUILDER.build(transformer)
+        if freeze_transformer:
+            self.transformer.requires_grad_(False)
+        self.freeze_transformer = freeze_transformer
+        self.res_vit = res_vit
+
+
+
+        self.weighting_scheme = weighting_scheme
+        self.logit_mean = logit_mean
+        self.logit_std = logit_std
+
+        self.vae = BUILDER.build(vae)
+        self.vae.requires_grad_(False)
+
+        self.use_activation_checkpointing = use_activation_checkpointing
+        self.tokenizer = BUILDER.build(tokenizer)
+
+        self.prompt_template = prompt_template
+        self.vit_input_size = vit_input_size
+        self.max_length = max_length
+        self.image_token_id = self.tokenizer.convert_tokens_to_ids(prompt_template['IMG_CONTEXT_TOKEN'])
+        self.register_buffer('vit_mean', torch.tensor(IMAGE_MEAN), persistent=False)
+        self.register_buffer('vit_std', torch.tensor(IMAGE_STD), persistent=False)
+
+        self.num_queries = num_queries
+        self.connector = ConnectorEncoder(ConnectorConfig(**connector))
+
+
+        self.projector_1 = nn.Linear(self.llm.config.hidden_size*6, self.connector.config.hidden_size)
+        self.projector_2 = nn.Linear(self.connector.config.hidden_size, self.transformer.config.pooled_projection_dim)
+        self.projector_3 = nn.Linear(self.connector.config.hidden_size, self.transformer.config.joint_attention_dim)
+
+        # zero out
+        nn.init.zeros_(self.projector_2.weight)
+        nn.init.zeros_(self.projector_3.weight)
+        nn.init.zeros_(self.projector_2.bias)
+        nn.init.zeros_(self.projector_3.bias)
+
+        self.meta_queries = nn.Parameter(
+            torch.zeros(num_queries, self.llm.config.hidden_size))
+        nn.init.normal_(self.meta_queries, std=1 / math.sqrt(self.llm.config.hidden_size))
+
+        if freeze_mq:
+            self.projector_1.requires_grad_(False)
+            self.projector_2.requires_grad_(False)
+            self.projector_3.requires_grad_(False)
+            self.connector.requires_grad_(False)
+            self.meta_queries.requires_grad_(False)
+        self.freeze_mq = freeze_mq
+
+        self.unconditional = unconditional
+
+        self.train_scheduler = BUILDER.build(train_scheduler)
+        self.test_scheduler = BUILDER.build(test_scheduler)
+
+        if use_activation_checkpointing:
+            self.gradient_checkpointing_enable()
+
+        if lora_modules is not None:
+            assert self.freeze_lmm
+            self.llm.config.tie_word_embeddings = False
+            if lora_modules == 'auto':
+                lora_modules = find_target_linear_names(self.lmm)
+            # now we will add new LoRA weights the transformer layers
+            transformer_lora_config = LoraConfig(
+                r=lora_rank,
+                lora_alpha=lora_alpha,
+                init_lora_weights="gaussian",
+                target_modules=lora_modules,
+                lora_dropout=0.05,
+            )
+            self.lmm.add_adapter(transformer_lora_config)
+
+        if pretrained_pth is not None:
+            pretrained_state_dict = guess_load_checkpoint(pretrained_pth)
+            info = self.load_state_dict(pretrained_state_dict, strict=False)
+            print_log(f'Load pretrained weight from {pretrained_pth}')
+
+
+        self.ema_cfg = ema_cfg
+        if ema_cfg is not None:
+            self.ema = nn.ModuleDict()
+            self.ema.steps = 0
+            if not self.freeze_transformer:
+                self.ema.update(dict(transformer=deepcopy(self.transformer)))
+
+            if not self.freeze_mq:
+                self.ema.update(dict(projector_1=deepcopy(self.projector_1),
+                                     projector_2=deepcopy(self.projector_2),
+                                     projector_3=deepcopy(self.projector_3),
+                                     connector=deepcopy(self.connector)
+                                     )
+                                )
+                self.ema.register_buffer('meta_queries', deepcopy(self.meta_queries.data))
+
+            self.ema.requires_grad_(False)   # parameters in ema are not learnable
+
+            if 'checkpoint' in ema_cfg:
+                ema_state_dict = guess_load_checkpoint(ema_cfg['checkpoint'])
+                info = self.ema.load_state_dict(ema_state_dict, strict=False)
+                print_log(f"Load ema weight from {ema_cfg['checkpoint']}")
+
+    @torch.no_grad()
+    def ema_step(self, ):
+        if self.ema_cfg is None:
+            return
+
+        steps = self.ema.steps
+        update_interval = self.ema_cfg.get('update_interval', 1)
+        save_interval = self.ema_cfg.get('save_interval', 1000)
+        momentum = self.ema_cfg.get('momentum', 0.99)
+
+        if steps % update_interval == 0 and steps > 0:
+            if not self.freeze_mq:
+                for ema_param, base_param in zip(self.ema.projector_1.parameters(), self.projector_1.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+                for ema_param, base_param in zip(self.ema.projector_2.parameters(), self.projector_2.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+                for ema_param, base_param in zip(self.ema.projector_3.parameters(), self.projector_3.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+                for ema_param, base_param in zip(self.ema.connector.parameters(), self.connector.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+                self.ema.meta_queries.data.lerp_(self.meta_queries.data.detach(), 1.0 - momentum)
+
+            if not self.freeze_transformer:
+                for ema_param, base_param in zip(self.ema.transformer.parameters(), self.transformer.parameters()):
+                    ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+
+        # print(f"steps: {steps}, rank: {dist.get_rank()},", flush=True)
+
+        if steps % save_interval == 0 and steps > 0:
+            is_ddp = dist.is_available() and dist.is_initialized()
+            is_primary_proc = (not is_ddp) or dist.get_rank() == 0
+            print(f"steps: {steps}, rank: {dist.get_rank()}, is_ddp:{is_ddp}, is_primary_proc: {is_primary_proc}.", flush=True)
+            if is_primary_proc:
+                save_path = self.ema_cfg.get('save_path')
+                torch.save(self.ema.state_dict(), save_path)
+            if is_ddp:
+                dist.barrier()
+
+        self.ema.steps = self.ema.steps + 1
+
+    def llm2dit(self, x):
+        x = self.connector(self.projector_1(x))
+        pooled_out = self.projector_2(x.mean(1))
+        seq_out = self.projector_3(x)
+
+        return pooled_out, seq_out
+
+    @property
+    def llm(self):
+        return self.lmm.language_model
+
+    def gradient_checkpointing_enable(self):
+        self.activation_checkpointing_enable()
+
+    def activation_checkpointing_enable(self):
+        self.llm.gradient_checkpointing_enable()
+        self.transformer.enable_gradient_checkpointing()
+        self.connector.gradient_checkpointing = True
+
+    def gradient_checkpointing_disable(self):
+        self.activation_checkpointing_disable()
+
+    def activation_checkpointing_disable(self):
+        self.llm.gradient_checkpointing_disable()
+        self.transformer.disable_gradient_checkpointing()
+        self.connector.gradient_checkpointing = False
+
+    @property
+    def device(self):
+        return self.llm.device
+
+    @property
+    def dtype(self):
+        return self.llm.dtype
+
+    def train(self: T, mode: bool = True) -> T:
+        super().train(mode=mode)
+        if self.vae is not None:
+            self.vae.train(mode=False)
+        if not mode:
+            self.gradient_checkpointing_disable()
+
+        return self
+
+    def state_dict(self, *args, **kwargs) -> dict:
+        state_dict = super().state_dict(*args, **kwargs)
+        state_dict = {k: v for k, v in state_dict.items()
+                      if 'vae.' not in k and 'lmm.' not in k and 'ema.' not in k}
+        return state_dict
+
+    @torch.no_grad()
+    def pixels_to_latents(self, x):
+        z = self.vae.encode(x).latent_dist.sample()
+        z = (z - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+        return z
+
+    @torch.no_grad()
+    def latents_to_pixels(self, z):
+        z = (z / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+        x_rec = self.vae.decode(z).sample
+        return x_rec
+
+    def forward(self, data, data_samples=None, mode='loss'):
+        if mode == 'loss':
+            self.ema_step()
+            return self.compute_loss(data_dict=data)
+        else:
+            raise NotImplementedError
+
+    def compute_loss(self, data_dict):
+        losses = {}
+        for data_type in ['text2image', 'image2image']:
+            if data_type in data_dict:
+                losses[f'loss_{data_type}'] = getattr(self, f'{data_type}_loss')(data_dict[data_type])
+        if len(losses) == 0:
+            if 'pixel_values_src' in data_dict:
+                losses[f'loss_image2image'] = self.image2image_loss(data_dict)
+            else:
+                losses[f'loss_text2image'] = self.text2image_loss(data_dict)
+
+        return losses
+
+
+    def prepare_forward_input(self,
+                              query_embeds,
+                              input_ids=None,
+                              image_embeds=None,
+                              image_grid_thw=None,
+                              attention_mask=None,
+                              past_key_values=None):
+        b, l, _ = query_embeds.shape
+        assert l > 0
+        attention_mask = attention_mask.to(device=self.device, dtype=torch.bool)
+
+        assert l == self.num_queries
+
+        input_ids = torch.cat([input_ids, input_ids.new_zeros(b, l)], dim=1)
+        attention_mask = torch.cat([attention_mask, attention_mask.new_ones(b, l)], dim=1)
+
+        position_ids, _ = self.lmm.model.get_rope_index(
+            input_ids=input_ids,
+            image_grid_thw=image_grid_thw,
+            video_grid_thw=None,
+            second_per_grid_ts=None,
+            attention_mask=attention_mask,
+        )
+
+        # prepare context
+        if past_key_values is not None:
+            inputs_embeds = query_embeds
+            position_ids = position_ids[..., -l:]
+        else:
+            input_ids = input_ids[:, :-l]   # context input_ids
+
+            if image_embeds is None:
+                inputs_embeds = self.llm.get_input_embeddings()(input_ids)
+            else:
+                inputs_embeds = torch.zeros(*input_ids.shape, self.llm.config.hidden_size,
+                                            device=self.device, dtype=self.dtype)
+                inputs_embeds[input_ids == self.image_token_id] = \
+                    image_embeds.contiguous().view(-1, self.llm.config.hidden_size)
+                inputs_embeds[input_ids != self.image_token_id] = self.llm.get_input_embeddings()(
+                    input_ids[input_ids != self.image_token_id]
+                )
+
+            inputs_embeds = torch.cat([inputs_embeds, query_embeds], dim=1)
+
+        inputs = dict(inputs_embeds=inputs_embeds,
+                      attention_mask=attention_mask,
+                      position_ids=position_ids,
+                      past_key_values=past_key_values)
+
+        return inputs
+
+
+    @torch.no_grad()
+    def get_semantic_features_dynamic(self, pixel_values):
+        # e.g., 512 -> 448
+        pixel_values = [F.interpolate(p[None], scale_factor=28 / 32, mode='bilinear') for p in pixel_values]
+        image_embeds, image_grid_thw = multi_apply(self.get_semantic_features,
+                                                   pixel_values, resize=False)
+        image_embeds = [x[0] for x in image_embeds]    # a list of embeds
+        image_grid_thw = torch.cat(image_grid_thw, dim=0)   # b 3
+
+        return image_embeds, image_grid_thw
+
+    @torch.no_grad()
+    def get_semantic_features(self, pixel_values, resize=True):
+        # pixel_values: [-1, 1]
+        pixel_values = (pixel_values + 1.0) / 2     # [0, 1]
+        pixel_values = pixel_values - self.vit_mean.view(1, 3, 1, 1)
+        pixel_values = pixel_values / self.vit_std.view(1, 3, 1, 1)
+
+        if resize:
+           pixel_values = F.interpolate(pixel_values, size=(self.vit_input_size, self.vit_input_size),
+                                        mode='bilinear')
+        b, c, h, w = pixel_values.shape
+
+        patch_size = self.lmm.config.vision_config.patch_size
+        spatial_merge_size = self.lmm.config.vision_config.spatial_merge_size
+        temporal_patch_size = self.lmm.config.vision_config.temporal_patch_size
+
+        pixel_values = pixel_values[:, None].expand(b, temporal_patch_size, c, h, w)
+
+        grid_t = 1
+        grid_h, grid_w = h // patch_size, w // patch_size
+
+        pixel_values = pixel_values.view(
+            b,
+            grid_t,
+            temporal_patch_size,
+            c,
+            grid_h // spatial_merge_size,
+            spatial_merge_size,
+            patch_size,
+            grid_w // spatial_merge_size,
+            spatial_merge_size,
+            patch_size,
+        )
+
+        pixel_values = rearrange(
+            pixel_values, 'b t tp c h m p w n q -> (b t h w m n) (c tp p q)')
+
+        image_grid_thw = torch.tensor([(grid_t, grid_h, grid_w)] * b).to(self.device).long()
+
+        image_embeds = self.lmm.visual(pixel_values, grid_thw=image_grid_thw)
+        image_embeds = rearrange(image_embeds, '(b l) d -> b l d', b=b)
+
+        return image_embeds, image_grid_thw
+
+    @torch.no_grad()
+    def prepare_text2image_prompts(self, texts):
+        texts = [self.prompt_template['GENERATION'].format(input=text) for text in texts]
+        texts = [self.prompt_template['INSTRUCTION'].format(input=text) for text in texts]
+
+        return self.tokenizer(
+            texts, add_special_tokens=True, return_tensors='pt', padding=True, padding_side='left').to(self.device)
+
+    @torch.no_grad()
+    def prepare_image2image_prompts(self, texts, num_refs, ref_lens):
+        prompts = []
+        cnt = 0
+        for text, num_ref in zip(texts, num_refs):
+            image_tokens = ''
+            for _ in range(num_ref):
+                image_tokens +=  self.prompt_template['IMG_START_TOKEN'] + \
+                                 self.prompt_template['IMG_CONTEXT_TOKEN'] * ref_lens[cnt] + \
+                                 self.prompt_template['IMG_END_TOKEN']
+                cnt += 1
+
+            prompts.append(self.prompt_template['INSTRUCTION'].format(input=f'{image_tokens}\n{text}'))
+
+        return self.tokenizer(
+            prompts, add_special_tokens=True, return_tensors='pt', padding=True, padding_side='left').to(self.device)
+
+
+    def text2image_loss(self, data_dict):
+        # obtain image latents
+        if 'image_latents' in data_dict:
+            image_latents = data_dict['image_latents']   # .to(dtype=self.dtype, device=self.device)
+            image_latents = [x.to(dtype=self.dtype, device=self.device) for x in image_latents]
+        else:
+            pixel_values = [p.to(dtype=self.dtype, device=self.device) for p in data_dict['pixel_values']]
+            image_latents = [self.pixels_to_latents(p[None])[0] for p in pixel_values]
+
+        b = len(image_latents)
+
+        texts = ['' if random.uniform(0, 1) < self.unconditional else text
+                 for text in data_dict['texts']]
+
+        text_inputs = self.prepare_text2image_prompts(texts)
+        hidden_states = self.meta_queries[None].expand(b, self.num_queries, -1)
+
+        inputs = self.prepare_forward_input(query_embeds=hidden_states, **text_inputs)
+
+        max_length = self.max_length + self.num_queries
+        inputs_embeds = inputs['inputs_embeds'][:, -max_length:]
+        attention_mask = inputs['attention_mask'][:, -max_length:]
+        position_ids = inputs['position_ids'][..., -max_length:]
+
+        output = self.llm(inputs_embeds=inputs_embeds,
+                          attention_mask=attention_mask,
+                          position_ids=position_ids,
+                          output_hidden_states=True,
+                          return_dict=True)
+
+        hidden_states = output.hidden_states
+        num_layers = len(hidden_states) - 1  # excpet embedding
+
+
+        selected_layers = list(range(num_layers - 1, 0, -6))
+        # [-2, -8, -14, ...]
+
+        
+        selected_hiddens = [hidden_states[i] for i in selected_layers]
+        merged_hidden = torch.cat(selected_hiddens, dim=-1)
+        pooled_out, seq_out = self.llm2dit(merged_hidden)
+
+        loss_diff = self.diff_loss(model_input=image_latents,
+                                   pooled_prompt_embeds=pooled_out,
+                                   prompt_embeds=seq_out)
+
+        return loss_diff
+
+
+    def image2image_loss(self, data_dict):
+        pixel_values_src = data_dict['pixel_values_src']
+
+        num_refs = [len(ref_images) for ref_images in pixel_values_src]
+
+        pixel_values_src = [[img.to(dtype=self.dtype, device=self.device) for img in ref_images]
+                            for ref_images in pixel_values_src]
+        image_latents_src = [[self.pixels_to_latents(img[None])[0] for img in ref_images]
+                             for ref_images in pixel_values_src]
+        image_embeds, image_grid_thw = self.get_semantic_features_dynamic(
+            [img for ref_images in pixel_values_src for img in ref_images])
+
+        ref_lens = [len(x) for x in image_embeds]
+
+        pixel_values = [p.to(dtype=self.dtype, device=self.device) for p in data_dict['pixel_values']]
+        image_latents = [self.pixels_to_latents(p[None])[0] for p in pixel_values]
+
+        b = len(image_latents)
+        text_inputs = self.prepare_image2image_prompts(data_dict['texts'], num_refs=num_refs, ref_lens=ref_lens)
+
+        hidden_states = self.meta_queries[None].expand(b, self.num_queries, -1)
+
+        inputs = self.prepare_forward_input(query_embeds=hidden_states,
+                                            image_embeds=torch.cat(image_embeds),
+                                            image_grid_thw=image_grid_thw,
+                                            **text_inputs)
+
+
+        max_length = self.max_length + max(num_refs) * max(ref_lens) + self.num_queries
+        inputs_embeds = inputs['inputs_embeds'][:, -max_length:]
+        attention_mask = inputs['attention_mask'][:, -max_length:]
+        position_ids = inputs['position_ids'][..., -max_length:]
+
+        output = self.llm(inputs_embeds=inputs_embeds,
+                          attention_mask=attention_mask,
+                          position_ids=position_ids,
+                          output_hidden_states = True,
+                          return_dict=True)
+
+        
+        hidden_states = output.hidden_states
+        num_layers = len(hidden_states) - 1  # except embedding
+
+        
+        selected_layers = list(range(num_layers - 1, 0, -6))
+        # [-2, -8, -14, ...]
+     
+
+        
+        selected_hiddens = [hidden_states[i] for i in selected_layers]
+        merged_hidden = torch.cat(selected_hiddens, dim=-1)
+        pooled_out, seq_out = self.llm2dit(merged_hidden)
+
+        # if res_vit:
+        #     image_embeds=torch.cat(image_embeds)
+        #     hidden_states[input_ids == self.image_token_id] = 0.5 * (hidden_states[input_ids == self.image_token_id]) + 0.5 * (image_embeds.contiguous().view(-1, self.llm.config.hidden_size))
+
+
+
+
+        loss_diff = self.diff_loss(model_input=image_latents,
+                                   pooled_prompt_embeds=pooled_out,
+                                   prompt_embeds=seq_out,
+                                   cond_intput=image_latents_src)
+
+        return loss_diff
+
+
+
+    @torch.no_grad()
+    def generate(self,
+                 prompt,
+                 cfg_prompt,
+                 pixel_values_src=None,
+                 cfg_scale=4.5,
+                 num_steps=50,
+                 generator=None,
+                 height=512,
+                 width=512,
+                 progress_bar=True):
+        assert len(prompt) == len(cfg_prompt)
+        b = len(prompt)
+
+        if pixel_values_src is not None:
+            num_refs = [len(ref_images) for ref_images in pixel_values_src]
+            pixel_values_src = [[img.to(dtype=self.dtype, device=self.device) for img in ref_imgs]
+                                for ref_imgs in pixel_values_src]
+            image_embeds, image_grid_thw = self.get_semantic_features_dynamic(
+                [img for ref_images in pixel_values_src for img in ref_images])
+            ref_lens = [len(x) for x in image_embeds]
+
+            text_inputs = self.prepare_image2image_prompts(prompt + cfg_prompt, num_refs=num_refs*2, ref_lens=ref_lens*2)
+            text_inputs.update(image_embeds=torch.cat(image_embeds*2),
+                               image_grid_thw=torch.cat([image_grid_thw]*2),)
+            cond_latents = [[self.pixels_to_latents(img[None])[0] for img in ref_imgs]
+                            for ref_imgs in pixel_values_src]
+            cond_latents = cond_latents * 2
+        else:
+            text_inputs = self.prepare_text2image_prompts(prompt + cfg_prompt)
+            cond_latents = None
+
+        hidden_states = self.meta_queries[None].expand(2*b, self.num_queries, -1)
+        inputs = self.prepare_forward_input(query_embeds=hidden_states, **text_inputs)
+
+        output = self.llm(**inputs, return_dict=True,output_hidden_states = True)
+
+         
+        hidden_states = output.hidden_states
+        num_layers = len(hidden_states) - 1  
+
+        
+        selected_layers = list(range(num_layers - 1, 0, -6))
+        
+     
+
+        
+        selected_hiddens = [hidden_states[i] for i in selected_layers]
+        merged_hidden = torch.cat(selected_hiddens, dim=-1)
+        pooled_out, seq_out = self.llm2dit(merged_hidden)
+    
+
+        pipeline = StableDiffusion3Pipeline(
+            transformer=self.transformer,
+            scheduler=self.test_scheduler,
+            vae=self.vae,
+            text_encoder=None,
+            tokenizer=None,
+            text_encoder_2=None,
+            tokenizer_2=None,
+            text_encoder_3=None,
+            tokenizer_3=None,
+        )
+
+        pipeline.set_progress_bar_config(disable=not progress_bar)
+
+        samples = pipeline(
+            height=height,
+            width=width,
+            guidance_scale=cfg_scale,
+            num_inference_steps=num_steps,
+            prompt_embeds=seq_out[:b],
+            pooled_prompt_embeds=pooled_out[:b],
+            negative_prompt_embeds=seq_out[b:],
+            negative_pooled_prompt_embeds=pooled_out[b:],
+            generator=generator,
+            output_type='latent',
+            cond_latents=cond_latents
+        ).images.to(self.dtype)
+
+        return self.latents_to_pixels(samples)
+
+    def diff_loss(self, model_input, pooled_prompt_embeds, prompt_embeds, cond_intput=None):
+        # Sample noise that we'll add to the latents
+        # import pdb; pdb.set_trace()
+        noise = [torch.randn_like(x) for x in model_input]
+        bsz = len(model_input)
+
+        u = compute_density_for_timestep_sampling(
+            weighting_scheme=self.weighting_scheme,
+            batch_size=bsz,
+            logit_mean=self.logit_mean,
+            logit_std=self.logit_std,
+        )
+
+        if self.train_scheduler.use_dynamic_shifting:
+            assert self.weighting_scheme == 'logit_normal'
+            # follow flux
+            # import pdb; pdb.set_trace()
+            image_seq_lens = [math.prod(x.shape[-2:]) // self.transformer.patch_size ** 2 for x in model_input]
+            mu = calculate_shift(
+                torch.tensor(image_seq_lens, dtype=self.dtype, device=self.device),
+                self.train_scheduler.config.get("base_image_seq_len", 256),
+                self.train_scheduler.config.get("max_image_seq_len", 4096),
+                self.train_scheduler.config.get("base_shift", 0.5),
+                self.train_scheduler.config.get("max_shift", 1.15)
+            )
+
+            if self.train_scheduler.config.time_shift_type == "exponential":
+                shift = torch.exp(mu)
+            elif self.train_scheduler.config.time_shift_type == "linear":
+                shift = mu
+            else:
+                raise NotImplementedError
+
+            sigmas = u.to(dtype=self.dtype, device=self.device)
+            sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)
+            timesteps = sigmas * self.train_scheduler.num_train_timesteps
+            sigmas = sigmas.view(-1, 1, 1, 1)
+
+        else:
+            # Sample a random timestep for each image
+            # for weighting schemes where we sample timesteps non-uniformly
+            indices = (u * self.train_scheduler.config.num_train_timesteps).long()
+            timesteps = self.train_scheduler.timesteps[indices].to(device=self.device)
+
+            # Add noise according to flow matching.
+            # zt = (1 - texp) * x + texp * z1
+            sigmas = self.get_sigmas(timesteps, n_dim=model_input[0].ndim + 1)
+
+        noisy_model_input = [(1.0 - x) * y + x * z  for x, y, z in zip(sigmas, model_input, noise)]
+
+        # Predict the noise residual
+        model_pred = self.transformer(
+            hidden_states=noisy_model_input,
+            cond_hidden_states=cond_intput,
+            encoder_hidden_states=prompt_embeds,
+            pooled_projections=pooled_prompt_embeds,
+            timestep=timesteps,
+            return_dict=False,
+        )[0]
+
+
+        # these weighting schemes use a uniform timestep sampling
+        # and instead post-weight the loss
+        weighting = compute_loss_weighting_for_sd3(weighting_scheme=self.weighting_scheme, sigmas=sigmas)
+
+        # flow matching loss
+        # target = noise - model_input
+        target = [x - y for x, y in zip(noise, model_input)]
+
+        loss = [(x.float() * (y.float() - z.float()) ** 2).mean() for x, y, z in zip(weighting, model_pred, target)]
+        loss = sum(loss) / len(loss)
+
+        return loss
+
+    def get_sigmas(self, timesteps, n_dim=4):
+        sigmas = self.train_scheduler.sigmas.to(device=self.device, dtype=self.dtype)
+        schedule_timesteps = self.train_scheduler.timesteps.to(self.device)
+        timesteps = timesteps.to(self.device)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+        sigma = sigmas[step_indices].flatten()
+        while len(sigma.shape) < n_dim:
+            sigma = sigma.unsqueeze(-1)
+        return sigma
+
+
+def resize_image(x, image_size, unit_image_size=32):
+    w, h = x.size
+    if w >= h and w >= image_size:
+        target_w = image_size
+        target_h = h * (target_w / w)
+        target_h = math.ceil(target_h / unit_image_size) * unit_image_size
+
+    elif h >= w and h >= image_size:
+        target_h = image_size
+        target_w = w * (target_h / h)
+        target_w = math.ceil(target_w / unit_image_size) * unit_image_size
+
+    else:
+        target_h = math.ceil(h / unit_image_size) * unit_image_size
+        target_w = math.ceil(w / unit_image_size) * unit_image_size
+
+    x = x.resize(size=(target_w, target_h))
+
+    return x
+
+
+if __name__ == "__main__":
+    import os
+    import argparse
+    from glob import glob
+    from mmengine.config import Config
+    from PIL import Image
+    import numpy as np
+
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('config', help='log file path.')
+    parser.add_argument("--checkpoint", type=str, default=None)
+    parser.add_argument("--image", type=str, default=None)
+    parser.add_argument("--prompt", type=str, default='a dog on the left and a cat on the right')
+    parser.add_argument("--cfg_prompt", type=str, default='')
+    parser.add_argument("--cfg_scale", type=float, default=4.0)
+    parser.add_argument("--num_steps", type=int, default=50)
+    parser.add_argument("--height", type=int, default=512)
+    parser.add_argument("--width", type=int, default=512)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--grid_size", type=int, default=2)
+    parser.add_argument('--output', type=str, default='output.jpg')
+
+    args = parser.parse_args()
+    config = Config.fromfile(args.config)
+    model = BUILDER.build(config.model).cuda().bfloat16().eval()
+
+    if args.checkpoint is not None:
+        print(f"Load checkpoint: {args.checkpoint}", flush=True)
+        checkpoint = guess_load_checkpoint(args.checkpoint)
+        info = model.load_state_dict(checkpoint, strict=False)
+
+    generator = torch.Generator(device=model.device).manual_seed(args.seed)
+    # repeat
+    bsz = args.grid_size ** 2
+
+    prompt = [args.prompt] * bsz
+    cfg_prompt = [args.cfg_prompt] * bsz
+
+    if args.image is not None:
+
+        if os.path.isdir(args.image):
+            ref_images = glob(f"{args.image}/*")
+            ref_images = [Image.open(path) for path in ref_images]
+        else:
+            ref_images = [Image.open(args.image)]
+
+        ref_images = [resize_image(img, max(args.width, args.height), 32) for img in ref_images]
+
+        if len(ref_images) == 1:
+            width, height = ref_images[0].size
+        else:
+            width, height = args.width, args.height
+
+        pixel_values_src = [torch.from_numpy(np.array(img)).to(dtype=model.dtype, device=model.device)
+                            for img in ref_images]
+        pixel_values_src = [rearrange(img, 'h w c -> c h w') for img in pixel_values_src]
+        pixel_values_src = [2 * (img / 255) - 1 for img in pixel_values_src]
+
+        pixel_values_src = [pixel_values_src, ] * bsz
+    else:
+        width, height = args.width, args.height
+        pixel_values_src = None
+
+    samples = model.generate(prompt=prompt, cfg_prompt=cfg_prompt, pixel_values_src=pixel_values_src,
+                             cfg_scale=args.cfg_scale, num_steps=args.num_steps,
+                             generator=generator, height=height, width=width)
+
+
+    samples = rearrange(samples, '(m n) c h w -> (m h) (n w) c', m=args.grid_size, n=args.grid_size)
+    samples = torch.clamp(
+        127.5 * samples + 128.0, 0, 255).to("cpu", dtype=torch.uint8).numpy()
+
+    Image.fromarray(samples).save(args.output)

src/models/sd3_kontext/sd3_hf.py

@@ -0,0 +1,486 @@
+import random
+import torch
+import torch.nn as nn
+import torch.distributed as dist
+from copy import deepcopy
+from torch.nn.modules.module import T
+from xtuner.registry import BUILDER
+from mmengine.logging import print_log
+from xtuner.model.utils import guess_load_checkpoint
+from diffusers.training_utils import compute_density_for_timestep_sampling, compute_loss_weighting_for_sd3
+from peft import LoraConfig
+from src.models.sd3_kontext.pipeline_stable_diffusion_3 import StableDiffusion3Pipeline
+from mmengine.model import BaseModel
+
+
+
+class StableDiffusion3HF(BaseModel):
+    def __init__(self,
+                 text_encoder,
+                 text_encoder_2,
+                 text_encoder_3,
+                 transformer,
+                 train_scheduler,
+                 test_scheduler,
+                 vae,
+                 tokenizer,
+                 tokenizer_2,
+                 tokenizer_3,
+                 pretrained_pth=None,
+                 use_activation_checkpointing=True,
+                 lora_modules=None,  # ["to_k", "to_q", "to_v", "to_out.0"],
+                 lora_rank=8,
+                 lora_alpha=8,
+                 freeze_transformer=True,
+                 unconditional=0.1,
+                 weighting_scheme='none',
+                 logit_mean=0.0,
+                 logit_std=1.0,
+                 ema_cfg=None,
+                 ):
+        super().__init__()
+        self.use_activation_checkpointing = use_activation_checkpointing
+        self.text_encoder = BUILDER.build(text_encoder)
+        self.text_encoder_2 = BUILDER.build(text_encoder_2)
+        self.text_encoder_3 = BUILDER.build(text_encoder_3)
+        self.text_encoder.requires_grad_(False)
+        self.text_encoder_2.requires_grad_(False)
+        self.text_encoder_3.requires_grad_(False)
+
+        self.tokenizer = BUILDER.build(tokenizer)
+        self.tokenizer_2 = BUILDER.build(tokenizer_2)
+        self.tokenizer_3 = BUILDER.build(tokenizer_3)
+
+        self.unconditional = unconditional
+
+        self.train_scheduler = BUILDER.build(train_scheduler)
+        self.test_scheduler = BUILDER.build(test_scheduler)
+
+        self.transformer = BUILDER.build(transformer)
+        if freeze_transformer:
+            self.transformer.requires_grad_(False)
+        self.freeze_transformer = freeze_transformer
+
+        self.vae = BUILDER.build(vae)
+        self.vae.requires_grad_(False)
+
+        self.weighting_scheme = weighting_scheme
+        self.logit_mean = logit_mean
+        self.logit_std = logit_std
+
+        if use_activation_checkpointing:
+            self.gradient_checkpointing_enable()
+
+        if lora_modules is not None:
+            assert self.freeze_transformer
+            # now we will add new LoRA weights the transformer layers
+            transformer_lora_config = LoraConfig(
+                r=lora_rank,
+                lora_alpha=lora_alpha,
+                init_lora_weights="gaussian",
+                target_modules=lora_modules,
+            )
+            self.transformer.add_adapter(transformer_lora_config)
+
+        if pretrained_pth is not None:
+            pretrained_state_dict = guess_load_checkpoint(pretrained_pth)
+            info = self.load_state_dict(pretrained_state_dict, strict=False)
+            print_log(f'Load pretrained weight from {pretrained_pth}')
+
+
+        self.ema_cfg = ema_cfg
+        if ema_cfg is not None:
+            self.ema = nn.ModuleDict()
+            self.ema.steps = 0
+            assert not self.freeze_transformer
+            self.ema.update(dict(transformer=deepcopy(self.transformer)))
+            self.ema.requires_grad_(False)   # parameters in ema are not learnable
+
+            if 'checkpoint' in ema_cfg:
+                ema_state_dict = guess_load_checkpoint(ema_cfg['checkpoint'])
+                info = self.ema.load_state_dict(ema_state_dict, strict=False)
+                print_log(f"Load ema weight from {ema_cfg['checkpoint']}")
+
+    @torch.no_grad()
+    def ema_step(self, ):
+        if self.ema_cfg is None:
+            return
+
+        steps = self.ema.steps
+        update_interval = self.ema_cfg.get('update_interval', 1)
+        save_interval = self.ema_cfg.get('save_interval', 1)
+        momentum = self.ema_cfg.get('momentum', 0.99)
+
+        if steps % update_interval == 0 and steps > 0:
+            for ema_param, base_param in zip(self.ema.transformer.parameters(), self.transformer.parameters()):
+                ema_param.data.lerp_(base_param.data.detach(), 1.0 - momentum)
+
+        if steps % save_interval == 0 and steps > 0:
+            is_ddp = dist.is_available() and dist.is_initialized()
+            is_primary_proc = (not is_ddp) or dist.get_rank() == 0
+            print(f"steps: {steps}, rank: {dist.get_rank()}, is_ddp:{is_ddp}, is_primary_proc: {is_primary_proc}.", flush=True)
+            if is_primary_proc:
+                save_path = self.ema_cfg.get('save_path')
+                torch.save(self.ema.state_dict(), save_path)
+            if is_ddp:
+                dist.barrier()
+
+        self.ema.steps = self.ema.steps + 1
+
+
+    def gradient_checkpointing_enable(self):
+        self.activation_checkpointing_enable()
+
+    def activation_checkpointing_enable(self):
+        self.transformer.enable_gradient_checkpointing()
+
+    def gradient_checkpointing_disable(self):
+        self.activation_checkpointing_disable()
+
+    def activation_checkpointing_disable(self):
+        self.transformer.disable_gradient_checkpointing()
+
+    @property
+    def device(self):
+        return self.transformer.device
+
+    @property
+    def dtype(self):
+        return self.transformer.dtype
+
+    def train(self: T, mode: bool = True) -> T:
+        super().train(mode=mode)
+        self.vae.train(mode=False)
+        self.text_encoder.train(mode=False)
+        self.text_encoder_2.train(mode=False)
+        self.text_encoder_3.train(mode=False)
+        if not mode:
+            self.gradient_checkpointing_disable()
+
+        return self
+
+    def state_dict(self, *args, **kwargs) -> dict:
+        state_dict = super().state_dict(*args, **kwargs)
+        return {k: v for k, v in state_dict.items() if k.startswith('transformer.')}
+
+    @torch.no_grad()
+    def pixels_to_latents(self, x):
+        z = self.vae.encode(x).latent_dist.sample()
+        z = (z - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+        return z
+
+    @torch.no_grad()
+    def latents_to_pixels(self, z):
+        z = (z / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+        x_rec = self.vae.decode(z).sample
+        return x_rec
+
+    def forward(self, data, data_samples=None, mode='loss'):
+        if mode == 'loss':
+            self.ema_step()
+            return self.compute_loss(data_dict=data)
+        else:
+            raise NotImplementedError
+
+    def compute_loss(self, data_dict):
+        losses = {}
+        for data_type in ['text2image', 'image2image']:
+            if data_type in data_dict:
+                losses[f'loss_{data_type}'] = getattr(self, f'{data_type}_loss')(data_dict[data_type])
+        if len(losses) == 0:
+            if 'pixel_values_src' in data_dict:
+                losses[f'loss_image2image'] = self.image2image_loss(data_dict)
+            else:
+                losses[f'loss_text2image'] = self.text2image_loss(data_dict)
+
+        return losses
+
+    def text2image_loss(self, data_dict):
+
+        # obtain image latents
+        if 'image_latents' in data_dict:
+            image_latents = data_dict['image_latents'].to(dtype=self.dtype, device=self.device)
+        else:
+            pixel_values = data_dict['pixel_values'].to(dtype=self.dtype, device=self.device)
+            image_latents = self.pixels_to_latents(pixel_values)
+
+        texts = ['' if random.uniform(0, 1) < self.unconditional else text
+                 for text in data_dict['texts']]
+
+        pipeline = StableDiffusion3Pipeline(
+            transformer=None,
+            scheduler=None,
+            vae=None,
+            text_encoder=self.text_encoder,
+            tokenizer=self.tokenizer,
+            text_encoder_2=self.text_encoder_2,
+            tokenizer_2=self.tokenizer_2,
+            text_encoder_3=self.text_encoder_3,
+            tokenizer_3=self.tokenizer_3,
+        )
+
+        with torch.no_grad():
+            (
+                prompt_embeds,
+                _,
+                pooled_prompt_embeds,
+                _,
+            ) = pipeline.encode_prompt(
+                prompt=texts,
+                prompt_2=None,
+                prompt_3=None,
+                negative_prompt=None,
+                negative_prompt_2=None,
+                negative_prompt_3=None,
+                do_classifier_free_guidance=False,
+                device=self.device,
+                clip_skip=None,
+                num_images_per_prompt=1,
+                max_sequence_length=256,
+                lora_scale=None,
+            )
+
+        loss_diff = self.diff_loss(model_input=image_latents,
+                                   pooled_prompt_embeds=pooled_prompt_embeds,
+                                   prompt_embeds=prompt_embeds)
+
+        return loss_diff
+
+
+    def image2image_loss(self, data_dict):
+
+        pixel_values_src = data_dict['pixel_values_src'].to(dtype=self.dtype, device=self.device)
+        image_latents_src = self.pixels_to_latents(pixel_values_src)
+
+        pixel_values = data_dict['pixel_values'].to(dtype=self.dtype, device=self.device)
+        image_latents = self.pixels_to_latents(pixel_values)
+
+        pipeline = StableDiffusion3Pipeline(
+            transformer=None,
+            scheduler=None,
+            vae=None,
+            text_encoder=self.text_encoder,
+            tokenizer=self.tokenizer,
+            text_encoder_2=self.text_encoder_2,
+            tokenizer_2=self.tokenizer_2,
+            text_encoder_3=self.text_encoder_3,
+            tokenizer_3=self.tokenizer_3,
+        )
+
+        with torch.no_grad():
+            (
+                prompt_embeds,
+                _,
+                pooled_prompt_embeds,
+                _,
+            ) = pipeline.encode_prompt(
+                prompt=data_dict['texts'],
+                prompt_2=None,
+                prompt_3=None,
+                negative_prompt=None,
+                negative_prompt_2=None,
+                negative_prompt_3=None,
+                do_classifier_free_guidance=False,
+                device=self.device,
+                clip_skip=None,
+                num_images_per_prompt=1,
+                max_sequence_length=256,
+                lora_scale=None,
+            )
+
+
+        loss_diff = self.diff_loss(model_input=image_latents,
+                                   pooled_prompt_embeds=pooled_prompt_embeds,
+                                   prompt_embeds=prompt_embeds,
+                                   cond_intput=image_latents_src)
+
+        return loss_diff
+
+
+    @torch.no_grad()
+    def generate(self,
+                 prompt,
+                 cfg_prompt,
+                 pixel_values_src=None,
+                 cfg_scale=4.5,
+                 num_steps=50,
+                 generator=None,
+                 height=512,
+                 width=512,
+                 progress_bar=True):
+
+        pipeline = StableDiffusion3Pipeline(
+            transformer=self.transformer,
+            scheduler=self.test_scheduler,
+            vae=self.vae,
+            text_encoder=self.text_encoder,
+            tokenizer=self.tokenizer,
+            text_encoder_2=self.text_encoder_2,
+            tokenizer_2=self.tokenizer_2,
+            text_encoder_3=self.text_encoder_3,
+            tokenizer_3=self.tokenizer_3,
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = pipeline.encode_prompt(
+            prompt=prompt,
+            prompt_2=None,
+            prompt_3=None,
+            negative_prompt=cfg_prompt,
+            negative_prompt_2=None,
+            negative_prompt_3=None,
+            do_classifier_free_guidance=True,
+            device=self.device,
+            clip_skip=None,
+            num_images_per_prompt=1,
+            max_sequence_length=256,
+            lora_scale=None,
+        )
+
+        pipeline.set_progress_bar_config(disable=not progress_bar)
+
+        if pixel_values_src is not None:
+            cond_latents = self.pixels_to_latents(pixel_values_src)
+            cond_latents = torch.cat([cond_latents] * 2)
+        else:
+            cond_latents = None
+
+        samples = pipeline(
+            height=height,
+            width=width,
+            guidance_scale=cfg_scale,
+            num_inference_steps=num_steps,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            generator=generator,
+            output_type='latent',
+            cond_latents=cond_latents,
+        ).images.to(self.dtype)
+
+        return self.latents_to_pixels(samples)
+
+    def diff_loss(self, model_input, pooled_prompt_embeds, prompt_embeds, cond_intput=None):
+        # Sample noise that we'll add to the latents
+        noise = torch.randn_like(model_input)
+        bsz = model_input.shape[0]
+
+        # Sample a random timestep for each image
+        # for weighting schemes where we sample timesteps non-uniformly
+        u = compute_density_for_timestep_sampling(
+            weighting_scheme=self.weighting_scheme,
+            batch_size=bsz,
+            logit_mean=self.logit_mean,
+            logit_std=self.logit_std,
+        )
+        indices = (u * self.train_scheduler.config.num_train_timesteps).long()
+        timesteps = self.train_scheduler.timesteps[indices].to(device=model_input.device)
+
+        # Add noise according to flow matching.
+        # zt = (1 - texp) * x + texp * z1
+        sigmas = self.get_sigmas(timesteps, n_dim=model_input.ndim)
+        noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
+
+        # Predict the noise residual
+        model_pred = self.transformer(
+            hidden_states=noisy_model_input,
+            cond_hidden_states=cond_intput,
+            encoder_hidden_states=prompt_embeds,
+            pooled_projections=pooled_prompt_embeds,
+            timestep=timesteps,
+            return_dict=False,
+        )[0]
+
+        # these weighting schemes use a uniform timestep sampling
+        # and instead post-weight the loss
+        weighting = compute_loss_weighting_for_sd3(weighting_scheme=self.weighting_scheme, sigmas=sigmas)
+
+        # flow matching loss
+        target = noise - model_input
+
+        # Compute regular loss.
+        loss = torch.mean(
+            (weighting.float() * (model_pred.float() - target.float()) ** 2).reshape(target.shape[0], -1),
+            1,
+        )
+        loss = loss.mean()
+
+        return loss
+
+    def get_sigmas(self, timesteps, n_dim=4):
+        sigmas = self.train_scheduler.sigmas.to(device=self.device, dtype=self.dtype)
+        schedule_timesteps = self.train_scheduler.timesteps.to(self.device)
+        timesteps = timesteps.to(self.device)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+        sigma = sigmas[step_indices].flatten()
+        while len(sigma.shape) < n_dim:
+            sigma = sigma.unsqueeze(-1)
+        return sigma
+
+
+if __name__ == "__main__":
+    import argparse
+    from mmengine.config import Config
+    from einops import rearrange
+    from PIL import Image
+    import numpy as np
+    import torch.nn.functional as F
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('config', help='log file path.')
+    parser.add_argument("--checkpoint", type=str, default=None)
+    parser.add_argument("--image", type=str, default=None)
+    parser.add_argument("--prompt", type=str, default='a dog on the left and a cat on the right')
+    parser.add_argument("--cfg_prompt", type=str, default='')
+    parser.add_argument("--cfg_scale", type=float, default=3.5)
+    parser.add_argument("--num_steps", type=int, default=50)
+    parser.add_argument("--height", type=int, default=512)
+    parser.add_argument("--width", type=int, default=512)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--grid_size", type=int, default=2)
+    parser.add_argument('--output', type=str, default='output.jpg')
+
+    args = parser.parse_args()
+    config = Config.fromfile(args.config)
+    model = BUILDER.build(config.model).cuda().bfloat16().eval()
+
+    if args.checkpoint is not None:
+        print(f"Load checkpoint: {args.checkpoint}", flush=True)
+        checkpoint = guess_load_checkpoint(args.checkpoint)
+        info = model.load_state_dict(checkpoint, strict=False)
+
+    generator = torch.Generator(device=model.device).manual_seed(args.seed)
+    # repeat
+    bsz = args.grid_size ** 2
+
+    prompt = [args.prompt] * bsz
+    cfg_prompt = [args.cfg_prompt] * bsz
+
+    if args.image is not None:
+        image = Image.open(args.image)
+        img_w, img_h = image.size
+        image = image.resize(size=(args.height, args.width))
+        pixel_values_src = torch.from_numpy(np.array(image)).to(dtype=model.dtype, device=model.device)
+        pixel_values_src = rearrange(pixel_values_src, 'h w c -> c h w')[None]
+        pixel_values_src = 2 * (pixel_values_src / 255) - 1
+        pixel_values_src = pixel_values_src.expand(bsz, -1, -1, -1)
+    else:
+        pixel_values_src = None
+
+    samples = model.generate(prompt=prompt, cfg_prompt=cfg_prompt, pixel_values_src=pixel_values_src,
+                             cfg_scale=args.cfg_scale, num_steps=args.num_steps,
+                             generator=generator, height=args.height, width=args.width)
+
+    if pixel_values_src is not None:
+        samples = F.interpolate(samples, size=(img_h, img_w), mode='bilinear')
+
+    samples = rearrange(samples, '(m n) c h w -> (m h) (n w) c', m=args.grid_size, n=args.grid_size)
+    samples = torch.clamp(
+        127.5 * samples + 128.0, 0, 255).to("cpu", dtype=torch.uint8).numpy()
+
+    Image.fromarray(samples).save(args.output)

src/models/sd3_kontext/sd3_hf_dynamic.py

@@ -0,0 +1,353 @@
+import random
+import torch
+import math
+from src.models.sd3_kontext.pipeline_stable_diffusion_3_dynamic import StableDiffusion3Pipeline, calculate_shift
+from src.models.sd3_kontext.sd3_hf import StableDiffusion3HF
+from diffusers.training_utils import compute_density_for_timestep_sampling, compute_loss_weighting_for_sd3
+
+
+class StableDiffusion3HFDynamic(StableDiffusion3HF):
+
+    def text2image_loss(self, data_dict):
+
+        pixel_values = [img.to(dtype=self.dtype, device=self.device) for img in data_dict['pixel_values']]
+        image_latents = [self.pixels_to_latents(img[None])[0] for img in pixel_values]
+
+        texts = ['' if random.uniform(0, 1) < self.unconditional else text
+                 for text in data_dict['texts']]
+
+        pipeline = StableDiffusion3Pipeline(
+            transformer=None,
+            scheduler=None,
+            vae=None,
+            text_encoder=self.text_encoder,
+            tokenizer=self.tokenizer,
+            text_encoder_2=self.text_encoder_2,
+            tokenizer_2=self.tokenizer_2,
+            text_encoder_3=self.text_encoder_3,
+            tokenizer_3=self.tokenizer_3,
+        )
+
+        with torch.no_grad():
+            (
+                prompt_embeds,
+                _,
+                pooled_prompt_embeds,
+                _,
+            ) = pipeline.encode_prompt(
+                prompt=texts,
+                prompt_2=None,
+                prompt_3=None,
+                negative_prompt=None,
+                negative_prompt_2=None,
+                negative_prompt_3=None,
+                do_classifier_free_guidance=False,
+                device=self.device,
+                clip_skip=None,
+                num_images_per_prompt=1,
+                max_sequence_length=512,
+                lora_scale=None,
+            )
+
+        loss_diff = self.diff_loss(model_input=image_latents,
+                                   pooled_prompt_embeds=pooled_prompt_embeds,
+                                   prompt_embeds=prompt_embeds)
+
+        return loss_diff
+
+
+    def image2image_loss(self, data_dict):
+        pixel_values_src = [[img.to(dtype=self.dtype, device=self.device) for img in ref_imgs]
+                            for ref_imgs in data_dict['pixel_values_src']]
+        image_latents_src = [[self.pixels_to_latents(img[None])[0] for img in ref_imgs]
+                             for ref_imgs in pixel_values_src]
+
+        pixel_values = [img.to(dtype=self.dtype, device=self.device) for img in data_dict['pixel_values']]
+        image_latents = [self.pixels_to_latents(img[None])[0] for img in pixel_values]
+
+        pipeline = StableDiffusion3Pipeline(
+            transformer=None,
+            scheduler=None,
+            vae=None,
+            text_encoder=self.text_encoder,
+            tokenizer=self.tokenizer,
+            text_encoder_2=self.text_encoder_2,
+            tokenizer_2=self.tokenizer_2,
+            text_encoder_3=self.text_encoder_3,
+            tokenizer_3=self.tokenizer_3,
+        )
+
+        with torch.no_grad():
+            (
+                prompt_embeds,
+                _,
+                pooled_prompt_embeds,
+                _,
+            ) = pipeline.encode_prompt(
+                prompt=data_dict['texts'],
+                prompt_2=None,
+                prompt_3=None,
+                negative_prompt=None,
+                negative_prompt_2=None,
+                negative_prompt_3=None,
+                do_classifier_free_guidance=False,
+                device=self.device,
+                clip_skip=None,
+                num_images_per_prompt=1,
+                max_sequence_length=512,
+                lora_scale=None,
+            )
+
+
+        loss_diff = self.diff_loss(model_input=image_latents,
+                                   pooled_prompt_embeds=pooled_prompt_embeds,
+                                   prompt_embeds=prompt_embeds,
+                                   cond_intput=image_latents_src)
+
+        return loss_diff
+
+
+    def diff_loss(self, model_input, pooled_prompt_embeds, prompt_embeds, cond_intput=None):
+        # Sample noise that we'll add to the latents
+        # import pdb; pdb.set_trace()
+        noise = [torch.randn_like(x) for x in model_input]
+        bsz = len(model_input)
+
+        u = compute_density_for_timestep_sampling(
+            weighting_scheme=self.weighting_scheme,
+            batch_size=bsz,
+            logit_mean=self.logit_mean,
+            logit_std=self.logit_std,
+        )
+
+        if self.train_scheduler.use_dynamic_shifting:
+            assert self.weighting_scheme == 'logit_normal'
+            # follow flux
+            # import pdb; pdb.set_trace()
+            image_seq_lens = [math.prod(x.shape[-2:]) // self.transformer.patch_size ** 2 for x in model_input]
+            mu = calculate_shift(
+                torch.tensor(image_seq_lens, dtype=self.dtype, device=self.device),
+                self.train_scheduler.config.get("base_image_seq_len", 256),
+                self.train_scheduler.config.get("max_image_seq_len", 4096),
+                self.train_scheduler.config.get("base_shift", 0.5),
+                self.train_scheduler.config.get("max_shift", 1.15)
+            )
+
+            if self.train_scheduler.config.time_shift_type == "exponential":
+                shift = torch.exp(mu)
+            elif self.train_scheduler.config.time_shift_type == "linear":
+                shift = mu
+            else:
+                raise NotImplementedError
+
+            sigmas = u.to(dtype=self.dtype, device=self.device)
+            sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)
+            timesteps = sigmas * self.train_scheduler.num_train_timesteps
+            sigmas = sigmas.view(-1, 1, 1, 1)
+
+        else:
+            # Sample a random timestep for each image
+            # for weighting schemes where we sample timesteps non-uniformly
+            indices = (u * self.train_scheduler.config.num_train_timesteps).long()
+            timesteps = self.train_scheduler.timesteps[indices].to(device=self.device)
+
+            # Add noise according to flow matching.
+            # zt = (1 - texp) * x + texp * z1
+            sigmas = self.get_sigmas(timesteps, n_dim=model_input[0].ndim + 1)
+
+        noisy_model_input = [(1.0 - x) * y + x * z  for x, y, z in zip(sigmas, model_input, noise)]
+
+        # Predict the noise residual
+        model_pred = self.transformer(
+            hidden_states=noisy_model_input,
+            cond_hidden_states=cond_intput,
+            encoder_hidden_states=prompt_embeds,
+            pooled_projections=pooled_prompt_embeds,
+            timestep=timesteps,
+            return_dict=False,
+        )[0]
+
+
+        # these weighting schemes use a uniform timestep sampling
+        # and instead post-weight the loss
+        weighting = compute_loss_weighting_for_sd3(weighting_scheme=self.weighting_scheme, sigmas=sigmas)
+
+        # flow matching loss
+        # target = noise - model_input
+        target = [x - y for x, y in zip(noise, model_input)]
+
+        loss = [(x.float() * (y.float() - z.float()) ** 2).mean() for x, y, z in zip(weighting, model_pred, target)]
+        loss = sum(loss) / len(loss)
+
+        return loss
+
+
+
+    @torch.no_grad()
+    def generate(self,
+                 prompt,
+                 cfg_prompt,
+                 pixel_values_src=None,
+                 cfg_scale=4.5,
+                 num_steps=50,
+                 generator=None,
+                 height=512,
+                 width=512,
+                 progress_bar=True):
+
+        pipeline = StableDiffusion3Pipeline(
+            transformer=self.transformer,
+            scheduler=self.test_scheduler,
+            vae=self.vae,
+            text_encoder=self.text_encoder,
+            tokenizer=self.tokenizer,
+            text_encoder_2=self.text_encoder_2,
+            tokenizer_2=self.tokenizer_2,
+            text_encoder_3=self.text_encoder_3,
+            tokenizer_3=self.tokenizer_3,
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = pipeline.encode_prompt(
+            prompt=prompt,
+            prompt_2=None,
+            prompt_3=None,
+            negative_prompt=cfg_prompt,
+            negative_prompt_2=None,
+            negative_prompt_3=None,
+            do_classifier_free_guidance=True,
+            device=self.device,
+            clip_skip=None,
+            num_images_per_prompt=1,
+            max_sequence_length=512,
+            lora_scale=None,
+        )
+
+        pipeline.set_progress_bar_config(disable=not progress_bar)
+
+        if pixel_values_src is not None:
+            pixel_values_src = [[img.to(dtype=self.dtype, device=self.device) for img in ref_imgs]
+                                for ref_imgs in pixel_values_src]
+            cond_latents = [[self.pixels_to_latents(img[None])[0] for img in ref_imgs]
+                            for ref_imgs in pixel_values_src]
+            cond_latents = cond_latents * 2
+        else:
+            cond_latents = None
+
+        samples = pipeline(
+            height=height,
+            width=width,
+            guidance_scale=cfg_scale,
+            num_inference_steps=num_steps,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            generator=generator,
+            output_type='latent',
+            cond_latents=cond_latents,
+        ).images.to(self.dtype)
+
+        return self.latents_to_pixels(samples)
+
+
+def resize_image(x, image_size, unit_image_size=32):
+    w, h = x.size
+    if w >= h and w >= image_size:
+        target_w = image_size
+        target_h = h * (target_w / w)
+        target_h = math.ceil(target_h / unit_image_size) * unit_image_size
+
+    elif h >= w and h >= image_size:
+        target_h = image_size
+        target_w = w * (target_h / h)
+        target_w = math.ceil(target_w / unit_image_size) * unit_image_size
+
+    else:
+        target_h = math.ceil(h / unit_image_size) * unit_image_size
+        target_w = math.ceil(w / unit_image_size) * unit_image_size
+
+    x = x.resize(size=(target_w, target_h))
+
+    return x
+
+
+if __name__ == "__main__":
+    import os
+    import argparse
+    from glob import glob
+    from mmengine.config import Config
+    from einops import rearrange
+    from PIL import Image
+    import numpy as np
+    from xtuner.model.utils import guess_load_checkpoint
+    from xtuner.registry import BUILDER
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('config', help='log file path.')
+    parser.add_argument("--checkpoint", type=str, default=None)
+    parser.add_argument("--image", type=str, default=None)
+    parser.add_argument("--prompt", type=str, default='a dog on the left and a cat on the right')
+    parser.add_argument("--cfg_prompt", type=str, default='')
+    parser.add_argument("--cfg_scale", type=float, default=3.5)
+    parser.add_argument("--num_steps", type=int, default=50)
+    parser.add_argument("--height", type=int, default=512)
+    parser.add_argument("--width", type=int, default=512)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--grid_size", type=int, default=2)
+    parser.add_argument('--output', type=str, default='output.jpg')
+
+    args = parser.parse_args()
+    config = Config.fromfile(args.config)
+    model = BUILDER.build(config.model).cuda().bfloat16().eval()
+
+    if args.checkpoint is not None:
+        print(f"Load checkpoint: {args.checkpoint}", flush=True)
+        checkpoint = guess_load_checkpoint(args.checkpoint)
+        info = model.load_state_dict(checkpoint, strict=False)
+
+    generator = torch.Generator(device=model.device).manual_seed(args.seed)
+    # repeat
+    bsz = args.grid_size ** 2
+
+    prompt = [args.prompt] * bsz
+    cfg_prompt = [args.cfg_prompt] * bsz
+
+    if args.image is not None:
+
+        if os.path.isdir(args.image):
+            ref_images = glob(f"{args.image}/*")
+            ref_images = [Image.open(path) for path in ref_images]
+        else:
+            ref_images = [Image.open(args.image)]
+
+        ref_images = [resize_image(img, max(args.width, args.height), 32) for img in ref_images]
+
+        if len(ref_images) == 1:
+            width, height = ref_images[0].size
+        else:
+            width, height = args.width, args.height
+
+        pixel_values_src = [torch.from_numpy(np.array(img)).to(dtype=model.dtype, device=model.device)
+                            for img in ref_images]
+        pixel_values_src = [rearrange(img, 'h w c -> c h w') for img in pixel_values_src]
+        pixel_values_src = [2 * (img / 255) - 1 for img in pixel_values_src]
+
+        pixel_values_src = [pixel_values_src, ] * bsz
+    else:
+        width, height = args.width, args.height
+        pixel_values_src = None
+
+    samples = model.generate(prompt=prompt, cfg_prompt=cfg_prompt, pixel_values_src=pixel_values_src,
+                             cfg_scale=args.cfg_scale, num_steps=args.num_steps,
+                             generator=generator, height=height, width=width)
+
+
+    samples = rearrange(samples, '(m n) c h w -> (m h) (n w) c', m=args.grid_size, n=args.grid_size)
+    samples = torch.clamp(
+        127.5 * samples + 128.0, 0, 255).to("cpu", dtype=torch.uint8).numpy()
+
+    Image.fromarray(samples).save(args.output)

src/models/sd3_kontext/transformer_sd3_dynamic.py

@@ -0,0 +1,639 @@
+# Copyright 2024 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, List, Optional, Tuple, Union
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from torch.nn.utils.rnn import pad_sequence
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin, SD3Transformer2DLoadersMixin
+from diffusers.models.attention import FeedForward, JointTransformerBlock, _chunked_feed_forward
+from diffusers.models.attention_processor import (
+    Attention,
+    AttentionProcessor,
+    FusedJointAttnProcessor2_0,
+    JointAttnProcessor2_0,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormContinuous, AdaLayerNormZero
+from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class CustomJointAttnProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("JointAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        batch_size = hidden_states.shape[0]
+
+        # `sample` projections.
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # `context` projections.
+        if encoder_hidden_states is not None:
+            ctx_len = encoder_hidden_states.shape[1]
+            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+
+            if attn.norm_added_q is not None:
+                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
+            if attn.norm_added_k is not None:
+                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
+
+            query = torch.cat([query, encoder_hidden_states_query_proj], dim=2)
+            key = torch.cat([key, encoder_hidden_states_key_proj], dim=2)
+            value = torch.cat([value, encoder_hidden_states_value_proj], dim=2)
+
+            if attention_mask is not None:
+                # import pdb; pdb.set_trace()
+                encoder_attention_mask = torch.ones(
+                    batch_size, ctx_len, dtype=torch.bool, device=hidden_states.device)
+                attention_mask = torch.cat([attention_mask, encoder_attention_mask], dim=1)
+
+
+        # import pdb; pdb.set_trace()
+        if attention_mask is not None:
+            attention_mask = attention_mask[:, None] * attention_mask[..., None]   # bsz, seqlen, seqlen
+            indices = range(attention_mask.shape[1])
+            attention_mask[:, indices, indices] = True
+            attention_mask = attention_mask[:, None]
+
+        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False,
+                                                       attn_mask=attention_mask)
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            # import pdb; pdb.set_trace()
+            # Split the attention outputs.
+            hidden_states, encoder_hidden_states = (
+                hidden_states[:, : residual.shape[1]],
+                hidden_states[:, residual.shape[1] :],
+            )
+            if not attn.context_pre_only:
+                encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if encoder_hidden_states is not None:
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
+
+
+class CustomJointTransformerBlock(JointTransformerBlock):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.attn.set_processor(CustomJointAttnProcessor2_0())
+        if self.attn2 is not None:
+            self.attn2.set_processor(CustomJointAttnProcessor2_0())
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor,
+        temb: torch.FloatTensor,
+        attention_mask: Optional[torch.BoolTensor] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ):
+        joint_attention_kwargs = joint_attention_kwargs or {}
+        if self.use_dual_attention:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp, norm_hidden_states2, gate_msa2 = self.norm1(
+                hidden_states, emb=temb
+            )
+        else:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+
+        if self.context_pre_only:
+            norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
+        else:
+            norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+                encoder_hidden_states, emb=temb
+            )
+
+        # Attention.
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            attention_mask=attention_mask,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            **joint_attention_kwargs,
+        )
+
+        # Process attention outputs for the `hidden_states`.
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = hidden_states + attn_output
+
+        if self.use_dual_attention:
+            attn_output2 = self.attn2(hidden_states=norm_hidden_states2, attention_mask=attention_mask,
+                                      **joint_attention_kwargs)
+            attn_output2 = gate_msa2.unsqueeze(1) * attn_output2
+            hidden_states = hidden_states + attn_output2
+
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = hidden_states + ff_output
+
+        # Process attention outputs for the `encoder_hidden_states`.
+        if self.context_pre_only:
+            encoder_hidden_states = None
+        else:
+            context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
+            encoder_hidden_states = encoder_hidden_states + context_attn_output
+
+            norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+            norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+            if self._chunk_size is not None:
+                # "feed_forward_chunk_size" can be used to save memory
+                context_ff_output = _chunked_feed_forward(
+                    self.ff_context, norm_encoder_hidden_states, self._chunk_dim, self._chunk_size
+                )
+            else:
+                context_ff_output = self.ff_context(norm_encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+        return encoder_hidden_states, hidden_states
+
+
+@maybe_allow_in_graph
+class SD3SingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+    ):
+        super().__init__()
+
+        self.norm1 = AdaLayerNormZero(dim)
+        self.attn = Attention(
+            query_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            bias=True,
+            processor=JointAttnProcessor2_0(),
+            eps=1e-6,
+        )
+
+        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
+        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+    def forward(self, hidden_states: torch.Tensor, temb: torch.Tensor):
+        # 1. Attention
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+        attn_output = self.attn(hidden_states=norm_hidden_states, encoder_hidden_states=None)
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = hidden_states + attn_output
+
+        # 2. Feed Forward
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp.unsqueeze(1)) + shift_mlp.unsqueeze(1)
+        ff_output = self.ff(norm_hidden_states)
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+        hidden_states = hidden_states + ff_output
+
+        return hidden_states
+
+
+class SD3Transformer2DModel(
+    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, SD3Transformer2DLoadersMixin
+):
+    """
+    The Transformer model introduced in [Stable Diffusion 3](https://huggingface.co/papers/2403.03206).
+
+    Parameters:
+        sample_size (`int`, defaults to `128`):
+            The width/height of the latents. This is fixed during training since it is used to learn a number of
+            position embeddings.
+        patch_size (`int`, defaults to `2`):
+            Patch size to turn the input data into small patches.
+        in_channels (`int`, defaults to `16`):
+            The number of latent channels in the input.
+        num_layers (`int`, defaults to `18`):
+            The number of layers of transformer blocks to use.
+        attention_head_dim (`int`, defaults to `64`):
+            The number of channels in each head.
+        num_attention_heads (`int`, defaults to `18`):
+            The number of heads to use for multi-head attention.
+        joint_attention_dim (`int`, defaults to `4096`):
+            The embedding dimension to use for joint text-image attention.
+        caption_projection_dim (`int`, defaults to `1152`):
+            The embedding dimension of caption embeddings.
+        pooled_projection_dim (`int`, defaults to `2048`):
+            The embedding dimension of pooled text projections.
+        out_channels (`int`, defaults to `16`):
+            The number of latent channels in the output.
+        pos_embed_max_size (`int`, defaults to `96`):
+            The maximum latent height/width of positional embeddings.
+        dual_attention_layers (`Tuple[int, ...]`, defaults to `()`):
+            The number of dual-stream transformer blocks to use.
+        qk_norm (`str`, *optional*, defaults to `None`):
+            The normalization to use for query and key in the attention layer. If `None`, no normalization is used.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["JointTransformerBlock", "CustomJointTransformerBlock"]
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: int = 128,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        num_layers: int = 18,
+        attention_head_dim: int = 64,
+        num_attention_heads: int = 18,
+        joint_attention_dim: int = 4096,
+        caption_projection_dim: int = 1152,
+        pooled_projection_dim: int = 2048,
+        out_channels: int = 16,
+        pos_embed_max_size: int = 96,
+        dual_attention_layers: Tuple[
+            int, ...
+        ] = (),  # () for sd3.0; (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12) for sd3.5
+        qk_norm: Optional[str] = None,
+    ):
+        super().__init__()
+        self.out_channels = out_channels if out_channels is not None else in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pos_embed = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=self.inner_dim,
+            pos_embed_max_size=pos_embed_max_size,  # hard-code for now.
+        )
+        self.time_text_embed = CombinedTimestepTextProjEmbeddings(
+            embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
+        )
+        self.context_embedder = nn.Linear(joint_attention_dim, caption_projection_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                CustomJointTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    context_pre_only=i == num_layers - 1,
+                    qk_norm=qk_norm,
+                    use_dual_attention=True if i in dual_attention_layers else False,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
+
+        self.gradient_checkpointing = False
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+    def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+        """
+        Sets the attention processor to use [feed forward
+        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+        Parameters:
+            chunk_size (`int`, *optional*):
+                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+                over each tensor of dim=`dim`.
+            dim (`int`, *optional*, defaults to `0`):
+                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+                or dim=1 (sequence length).
+        """
+        if dim not in [0, 1]:
+            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+        # By default chunk size is 1
+        chunk_size = chunk_size or 1
+
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, chunk_size, dim)
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.disable_forward_chunking
+    def disable_forward_chunking(self):
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, None, 0)
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedJointAttnProcessor2_0
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedJointAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        cond_hidden_states: torch.Tensor = None,
+        pooled_projections: torch.Tensor = None,
+        timestep: torch.LongTensor = None,
+        block_controlnet_hidden_states: List = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+        skip_layers: Optional[List[int]] = None,
+    ) -> Union[torch.Tensor, Transformer2DModelOutput]:
+        """
+        The [`SD3Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.Tensor` of shape `(batch size, channel, height, width)`):
+                Input `hidden_states`.
+            encoder_hidden_states (`torch.Tensor` of shape `(batch size, sequence_len, embed_dims)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            pooled_projections (`torch.Tensor` of shape `(batch_size, projection_dim)`):
+                Embeddings projected from the embeddings of input conditions.
+            timestep (`torch.LongTensor`):
+                Used to indicate denoising step.
+            block_controlnet_hidden_states (`list` of `torch.Tensor`):
+                A list of tensors that if specified are added to the residuals of transformer blocks.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+            skip_layers (`list` of `int`, *optional*):
+                A list of layer indices to skip during the forward pass.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        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."
+                )
+
+        latent_sizes = [hs.shape[-2:] for hs in hidden_states]
+        bsz = len(hidden_states)
+
+        hidden_states_list = []
+        for idx in range(bsz):
+            hidden_states_per_sample = self.pos_embed(hidden_states[idx][None])[0]
+            if cond_hidden_states is not None:
+                for ref in cond_hidden_states[idx]:
+                    hidden_states_per_sample = torch.cat(
+                        [hidden_states_per_sample, self.pos_embed(ref[None])[0]])
+
+            hidden_states_list.append(hidden_states_per_sample)
+
+        max_len = max([len(hs) for hs in hidden_states_list])
+        attention_mask = torch.zeros(bsz, max_len, dtype=torch.bool, device=self.device)
+        for i, hs in enumerate(hidden_states_list):
+            attention_mask[i, :len(hs)] = True  # right padding
+        # import pdb; pdb.set_trace()
+        hidden_states = pad_sequence(hidden_states_list, batch_first=True, padding_value=0.0, padding_side='right')
+
+        temb = self.time_text_embed(timestep, pooled_projections)
+        encoder_hidden_states = self.context_embedder(encoder_hidden_states)
+
+        if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs:
+            ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds")
+            ip_hidden_states, ip_temb = self.image_proj(ip_adapter_image_embeds, timestep)
+
+            joint_attention_kwargs.update(ip_hidden_states=ip_hidden_states, temb=ip_temb)
+
+        for index_block, block in enumerate(self.transformer_blocks):
+            # Skip specified layers
+            is_skip = True if skip_layers is not None and index_block in skip_layers else False
+
+            if torch.is_grad_enabled() and self.gradient_checkpointing and not is_skip:
+                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    attention_mask,
+                    joint_attention_kwargs,
+                )
+            elif not is_skip:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=temb,
+                    attention_mask=attention_mask,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+
+            # controlnet residual
+            if block_controlnet_hidden_states is not None and block.context_pre_only is False:
+                interval_control = len(self.transformer_blocks) / len(block_controlnet_hidden_states)
+                hidden_states = hidden_states + block_controlnet_hidden_states[int(index_block / interval_control)]
+
+        hidden_states = self.norm_out(hidden_states, temb)
+        hidden_states = self.proj_out(hidden_states)
+
+        patch_size = self.config.patch_size
+        latent_sizes = [(latent_size[0] // patch_size, latent_size[1] // patch_size)
+                        for latent_size in latent_sizes]
+
+        # import pdb; pdb.set_trace()
+        # unpatchify
+        output = [rearrange(hs[:math.prod(latent_size)], '(h w) (p q c) -> c (h p) (w q)',
+                            h=latent_size[0], w=latent_size[1], p=patch_size, q=patch_size)
+                  for hs, latent_size in zip(hidden_states, latent_sizes)]
+
+        try:
+            output = torch.stack(output)    # can be staked if all have the save shape
+        except:
+            # cannot be stacked
+            pass
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/optimisers/custom_adamw.py

@@ -0,0 +1,45 @@
+from torch.optim import AdamW
+
+
+class CustomAdamW(AdamW):
+    def __init__(self, params, weight_decay, *args, **kwargs):
+        # import pdb; pdb.set_trace()
+        if isinstance(params, dict):
+            params = [p for p in params.values() if p.requires_grad]
+        else:
+            params = [p for p in params if p.requires_grad]
+
+        # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.
+        # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.
+        decay_params = [p for p in params if p.dim() >= 2]
+        nodecay_params = [p for p in params if p.dim() < 2]
+        optim_groups = [
+            {'params': decay_params, 'weight_decay': weight_decay},
+            {'params': nodecay_params, 'weight_decay': 0.0}
+        ]
+        num_decay_params = sum(p.numel() for p in decay_params)
+        num_nodecay_params = sum(p.numel() for p in nodecay_params)
+        print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
+        print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
+        # Create AdamW optimizer and use the fused version if it is available
+        # fused_available = 'fused' in inspect.signature(AdamW).parameters
+        # extra_args = dict(fused=True) if fused_available else dict()
+        # print(f"using fused AdamW: {fused_available}")
+
+        # kwargs.update(extra_args)
+
+        super().__init__(params=optim_groups, *args, **kwargs)
+
+
+class ParamWiseAdamW(AdamW):
+    def __init__(self, params, *args, **kwargs):
+        assert isinstance(params, list)
+        for param in params:
+            assert isinstance(param, dict)
+            assert isinstance(param['params'], list)
+            assert len(param['params']) == 1
+
+            if param['params'][0].ndim == 1:
+                param['weight_decay'] = 0.0
+
+        super().__init__(params=params, *args, **kwargs)

src/runners/custom_runner.py

@@ -0,0 +1,177 @@
+import copy
+import logging
+import inspect
+
+from torch.utils.data import DataLoader
+from functools import partial
+from typing import Callable, Dict, List, Optional, Union
+
+from mmengine.logging import print_log
+from mmengine.dist import get_rank
+from mmengine.dataset import worker_init_fn as default_worker_init_fn
+from mmengine.utils import digit_version
+from mmengine.utils.dl_utils import TORCH_VERSION
+from mmengine.runner import FlexibleRunner
+from mmengine.registry import (
+    DATA_SAMPLERS,
+    DATASETS,
+    FUNCTIONS,
+)
+from xtuner.registry import BUILDER
+
+
+class CustomRunner(FlexibleRunner):
+    def __init__(
+        self,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+    @staticmethod
+    def build_dataloader(
+        dataloader: Union[DataLoader, Dict],
+        seed: Optional[int] = None,
+        diff_rank_seed: bool = False,
+    ) -> DataLoader:
+        """Build dataloader.
+
+        The method builds three components:
+
+        - Dataset
+        - Sampler
+        - Dataloader
+
+        An example of ``dataloader``::
+
+            dataloader = dict(
+                dataset=dict(type='ToyDataset'),
+                sampler=dict(type='DefaultSampler', shuffle=True),
+                batch_size=1,
+                num_workers=9
+            )
+
+        Args:
+            dataloader (DataLoader or dict): A Dataloader object or a dict to
+                build Dataloader object. If ``dataloader`` is a Dataloader
+                object, just returns itself.
+            seed (int, optional): Random seed. Defaults to None.
+            diff_rank_seed (bool): Whether or not set different seeds to
+                different ranks. If True, the seed passed to sampler is set
+                to None, in order to synchronize the seeds used in samplers
+                across different ranks. Defaults to False.
+
+        Returns:
+            Dataloader: DataLoader build from ``dataloader_cfg``.
+        """
+        if isinstance(dataloader, DataLoader):
+            return dataloader
+
+        dataloader_cfg = copy.deepcopy(dataloader)
+
+        # build dataset
+        dataset_cfg = dataloader_cfg.pop('dataset')
+        if isinstance(dataset_cfg, dict):
+            dataset = DATASETS.build(dataset_cfg)
+            if hasattr(dataset, 'full_init'):
+                dataset.full_init()
+        else:
+            # fallback to raise error in dataloader
+            # if `dataset_cfg` is not a valid type
+            dataset = dataset_cfg
+
+        # build sampler
+        sampler_cfg = dataloader_cfg.pop('sampler')
+        if isinstance(sampler_cfg, dict):
+            sampler_seed = None if diff_rank_seed else seed
+            sampler = DATA_SAMPLERS.build(
+                sampler_cfg,
+                default_args=dict(dataset=dataset, seed=sampler_seed))
+        else:
+            # fallback to raise error in dataloader
+            # if `sampler_cfg` is not a valid type
+            sampler = sampler_cfg
+
+        # build batch sampler
+        batch_sampler_cfg = dataloader_cfg.pop('batch_sampler', None)
+        if batch_sampler_cfg is None:
+            batch_sampler = None
+        elif isinstance(batch_sampler_cfg, dict):
+            batch_sampler = DATA_SAMPLERS.build(
+                batch_sampler_cfg,
+                default_args=dict(
+                    dataset=dataset,
+                    sampler=sampler,
+                    batch_size=dataloader_cfg.pop('batch_size')))
+        else:
+            # fallback to raise error in dataloader
+            # if `batch_sampler_cfg` is not a valid type
+            batch_sampler = batch_sampler_cfg
+
+        # build dataloader
+        init_fn: Optional[partial]
+        if 'worker_init_fn' in dataloader_cfg:
+            worker_init_fn_cfg = dataloader_cfg.pop('worker_init_fn')
+            worker_init_fn_type = worker_init_fn_cfg.pop('type')
+            worker_init_fn = FUNCTIONS.get(worker_init_fn_type)
+            assert callable(worker_init_fn)
+            init_fn = partial(worker_init_fn,
+                              **worker_init_fn_cfg)  # type: ignore
+        else:
+            if seed is not None:
+                disable_subprocess_warning = dataloader_cfg.pop(
+                    'disable_subprocess_warning', False)
+                assert isinstance(disable_subprocess_warning, bool), (
+                    'disable_subprocess_warning should be a bool, but got '
+                    f'{type(disable_subprocess_warning)}')
+                init_fn = partial(
+                    default_worker_init_fn,
+                    num_workers=dataloader_cfg.get('num_workers'),
+                    rank=get_rank(),
+                    seed=seed,
+                    disable_subprocess_warning=disable_subprocess_warning)
+            else:
+                init_fn = None
+
+        # `persistent_workers` requires pytorch version >= 1.7
+        if ('persistent_workers' in dataloader_cfg
+                and digit_version(TORCH_VERSION) < digit_version('1.7.0')):
+            print_log(
+                '`persistent_workers` is only available when '
+                'pytorch version >= 1.7',
+                logger='current',
+                level=logging.WARNING)
+            dataloader_cfg.pop('persistent_workers')
+
+        # The default behavior of `collat_fn` in dataloader is to
+        # merge a list of samples to form a mini-batch of Tensor(s).
+        # However, in mmengine, if `collate_fn` is not defined in
+        # dataloader_cfg, `pseudo_collate` will only convert the list of
+        # samples into a dict without stacking the batch tensor.
+        collate_fn_cfg = dataloader_cfg.pop('collate_fn',
+                                            dict(type='pseudo_collate'))
+        if isinstance(collate_fn_cfg, dict):
+            collate_fn_type = collate_fn_cfg.pop('type')
+            if isinstance(collate_fn_type, str):
+                collate_fn = FUNCTIONS.get(collate_fn_type)
+            elif inspect.isclass(collate_fn_type):
+                collate_fn_cfg['type'] = collate_fn_type
+                collate_fn = BUILDER.build(collate_fn_cfg)
+            else:
+                collate_fn = collate_fn_type
+            if not inspect.isclass(collate_fn_type):
+                collate_fn = partial(collate_fn, **collate_fn_cfg)  # type: ignore
+        elif callable(collate_fn_cfg):
+            collate_fn = collate_fn_cfg
+        else:
+            raise TypeError(
+                'collate_fn should be a dict or callable object, but got '
+                f'{collate_fn_cfg}')
+        data_loader = DataLoader(
+            dataset=dataset,
+            sampler=sampler if batch_sampler is None else None,
+            batch_sampler=batch_sampler,
+            collate_fn=collate_fn,
+            worker_init_fn=init_fn,
+            **dataloader_cfg)
+
+        return data_loader