Update app.py

app.py

@@ -1,29 +1,30 @@
 import os
-# [중요] OOM 방지를 위한 메모리 파편화 설정 (토치 로드 전에 설정해야 함)
-os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
-os.environ['SPCONV_ALGO'] = 'native'
-
-import gradio as gr
-from gradio_litmodel3d import LitModel3D
-
+import gc
 import shutil
-from typing import *
 import torch
+import torch.nn as nn
 import numpy as np
 import imageio
-import gc  # 가비지 컬렉션 추가
+from typing import *
 from easydict import EasyDict as edict
 from PIL import Image
+
+# [핵심 1] 메모리 파편화 방지 및 4090/L40S 등 최신 GPU 호환성 설정
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+os.environ['SPCONV_ALGO'] = 'native'
+
+import gradio as gr
+from gradio_litmodel3d import LitModel3D
 from trellis.pipelines import TrellisVGGTTo3DPipeline
 from trellis.representations import Gaussian, MeshExtractResult
 from trellis.utils import render_utils, postprocessing_utils
 
-# [중요] 모델 분산을 위한 accelerate 라이브러리 체크
+# [핵심 2] Accelerate 라이브러리 필수 로드 (없으면 에러 발생)
 try:
-    from accelerate import dispatch_model
-    ACCELERATE_AVAILABLE = True
+    from accelerate import dispatch_model, infer_auto_device_map
+    from accelerate.utils import get_balanced_memory
 except ImportError:
-    ACCELERATE_AVAILABLE = False
+    raise ImportError("Accelerate library is missing. Please run `pip install accelerate`.")
 
 MAX_SEED = np.iinfo(np.int32).max
 TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
@@ -33,46 +34,19 @@ def start_session(req: gr.Request):
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     os.makedirs(user_dir, exist_ok=True)
     
-    
 def end_session(req: gr.Request):
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     if os.path.exists(user_dir):
         shutil.rmtree(user_dir)
-    # 세션 종료 시 메모리 정리
     gc.collect()
     torch.cuda.empty_cache()
 
 def preprocess_image(image: Image.Image) -> Image.Image:
-    """
-    Preprocess the input image for 3D generation.
-    
-    This function is called when a user uploads an image or selects an example.
-    It applies background removal and other preprocessing steps necessary for
-    optimal 3D model generation.
-
-    Args:
-        image (Image.Image): The input image from the user
-
-    Returns:
-        Image.Image: The preprocessed image ready for 3D generation
-    """
+    # 파이프라인이 분산되어 있어도 전처리는 CPU/GPU0에서 수행됨
     processed_image = pipeline.preprocess_image(image)
     return processed_image
 
 def preprocess_videos(video: str) -> List[Tuple[Image.Image, str]]:
-    """
-    Preprocess the input video for multi-image 3D generation.
-    
-    This function is called when a user uploads a video.
-    It extracts frames from the video and processes each frame to prepare them
-    for the multi-image 3D generation pipeline.
-    
-    Args:
-        video (str): The path to the input video file
-        
-    Returns:
-        List[Tuple[Image.Image, str]]: The list of preprocessed images ready for 3D generation
-    """
     vid = imageio.get_reader(video, 'ffmpeg')
     fps = vid.get_meta_data()['fps']
     images = []
@@ -87,23 +61,10 @@ def preprocess_videos(video: str) -> List[Tuple[Image.Image, str]]:
     return processed_images
 
 def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:
-    """
-    Preprocess a list of input images for multi-image 3D generation.
-    
-    This function is called when users upload multiple images in the gallery.
-    It processes each image to prepare them for the multi-image 3D generation pipeline.
-    
-    Args:
-        images (List[Tuple[Image.Image, str]]): The input images from the gallery
-        
-    Returns:
-        List[Image.Image]: The preprocessed images ready for 3D generation
-    """
     images = [image[0] for image in images]
     processed_images = [pipeline.preprocess_image(image) for image in images]
     return processed_images
 
-
 def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
     return {
         'gaussian': {
@@ -119,11 +80,10 @@ def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
             'faces': mesh.faces.cpu().numpy(),
         },
     }
-    
-    
+
 def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
-    # 언팩 시 바로 CUDA로 올리면 메모리 튈 수 있으므로 상황에 맞게 device 설정
-    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+    # 결과물 렌더링 시에는 GPU 하나(보통 0번)나 CPU로 모아야 함
+    device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
     
     gs = Gaussian(
         aabb=state['gaussian']['aabb'],
@@ -143,27 +103,11 @@ def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
         vertices=torch.tensor(state['mesh']['vertices'], device=device),
         faces=torch.tensor(state['mesh']['faces'], device=device),
     )
-    
     return gs, mesh
 
-
 def get_seed(randomize_seed: bool, seed: int) -> int:
-    """
-    Get the random seed for generation.
-    
-    This function is called by the generate button to determine whether to use
-    a random seed or the user-specified seed value.
-    
-    Args:
-        randomize_seed (bool): Whether to generate a random seed
-        seed (int): The user-specified seed value
-        
-    Returns:
-        int: The seed to use for generation
-    """
     return np.random.randint(0, MAX_SEED) if randomize_seed else seed
 
-
 def generate_and_extract_glb(
     multiimages: List[Tuple[Image.Image, str]],
     seed: int,
@@ -176,61 +120,41 @@ def generate_and_extract_glb(
     texture_size: int,
     req: gr.Request,
 ) -> Tuple[dict, str, str, str]:
-    """
-    Convert an image to a 3D model and extract GLB file.
-
-    Args:
-        image (Image.Image): The input image.
-        multiimages (List[Tuple[Image.Image, str]]): The input images in multi-image mode.
-        is_multiimage (bool): Whether is in multi-image mode.
-        seed (int): The random seed.
-        ss_guidance_strength (float): The guidance strength for sparse structure generation.
-        ss_sampling_steps (int): The number of sampling steps for sparse structure generation.
-        slat_guidance_strength (float): The guidance strength for structured latent generation.
-        slat_sampling_steps (int): The number of sampling steps for structured latent generation.
-        multiimage_algo (Literal["multidiffusion", "stochastic"]): The algorithm for multi-image generation.
-        mesh_simplify (float): The mesh simplification factor.
-        texture_size (int): The texture resolution.
-
-    Returns:
-        dict: The information of the generated 3D model.
-        str: The path to the video of the 3D model.
-        str: The path to the extracted GLB file.
-        str: The path to the extracted GLB file (for download).
-    """
-    # [수정] 추론 시작 전 가비지 컬렉션 수행
+    
+    # [핵심 3] 실행 전 가비지 컬렉션으로 VRAM 확보
     gc.collect()
     torch.cuda.empty_cache()
     
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     image_files = [image[0] for image in multiimages]
 
-    # Generate 3D model
-    # [수정] torch.no_grad()를 사용하여 불필요한 그라디언트 메모리 사용 방지
-    with torch.no_grad():
-        outputs, _, _ = pipeline.run(
-            image=image_files,
-            seed=seed,
-            formats=["gaussian", "mesh"],
-            preprocess_image=False,
-            sparse_structure_sampler_params={
-                "steps": ss_sampling_steps,
-                "cfg_strength": ss_guidance_strength,
-            },
-            slat_sampler_params={
-                "steps": slat_sampling_steps,
-                "cfg_strength": slat_guidance_strength,
-            },
-            mode=multiimage_algo,
-        )
-
-    # Render video
-    # import uuid
-    # output_id = str(uuid.uuid4())
-    # os.makedirs(f"{TMP_DIR}/{output_id}", exist_ok=True)
-    # video_path = f"{TMP_DIR}/{output_id}/preview.mp4"
-    # glb_path = f"{TMP_DIR}/{output_id}/mesh.glb"
-
+    try:
+        # [핵심 4] 추론 모드 강제. Gradient 저장 안 함 (메모리 절약)
+        with torch.no_grad():
+            outputs, _, _ = pipeline.run(
+                image=image_files,
+                seed=seed,
+                formats=["gaussian", "mesh"],
+                preprocess_image=False,
+                sparse_structure_sampler_params={
+                    "steps": ss_sampling_steps,
+                    "cfg_strength": ss_guidance_strength,
+                },
+                slat_sampler_params={
+                    "steps": slat_sampling_steps,
+                    "cfg_strength": slat_guidance_strength,
+                },
+                mode=multiimage_algo,
+            )
+    except Exception as e:
+        # 에러 발생 시 메모리 비우고 에러 다시 던짐
+        torch.cuda.empty_cache()
+        raise e
+
+    # Render video (후처리는 단일 GPU 혹은 CPU에서 수행)
+    # 텐서들이 여러 GPU에 흩어져 있을 수 있으므로 .cpu() 혹은 단일 cuda로 모아야 함
+    # render_utils 내부에서 처리하겠지만 안전하게 진행
+    
     video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
     video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
     video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
@@ -244,46 +168,27 @@ def generate_and_extract_glb(
     glb_path = os.path.join(user_dir, 'sample.glb')
     glb.export(glb_path)
     
-    # Pack state for optional Gaussian extraction
+    # Pack state
     state = pack_state(gs, mesh)
     
-    # [수정] 사용 끝난 변수 명시적 삭제 및 메모리 정리
+    # 결과물 반환 후 메모리 정리
     del outputs, gs, mesh, glb
     gc.collect()
     torch.cuda.empty_cache()
     
     return state, video_path, glb_path, glb_path
 
-
 def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
-    """
-    Extract a Gaussian splatting file from the generated 3D model.
-    
-    This function is called when the user clicks "Extract Gaussian" button.
-    It converts the 3D model state into a .ply file format containing
-    Gaussian splatting data for advanced 3D applications.
-
-    Args:
-        state (dict): The state of the generated 3D model containing Gaussian data
-        req (gr.Request): Gradio request object for session management
-
-    Returns:
-        Tuple[str, str]: Paths to the extracted Gaussian file (for display and download)
-    """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     gs, _ = unpack_state(state)
     gaussian_path = os.path.join(user_dir, 'sample.ply')
     gs.save_ply(gaussian_path)
     
-    # [수정] 메모리 정리
     del gs
     torch.cuda.empty_cache()
-    
     return gaussian_path, gaussian_path
 
-
 def prepare_multi_example() -> List[Image.Image]:
-    # 에러 방지용 경로 체크
     if not os.path.exists("assets/example_multi_image"):
         return []
     multi_case = list(set([i.split('_')[0] for i in os.listdir("assets/example_multi_image")]))
@@ -300,21 +205,7 @@ def prepare_multi_example() -> List[Image.Image]:
             images.append(Image.fromarray(np.concatenate(_images, axis=1)))
     return images
 
-
 def split_image(image: Image.Image) -> List[Image.Image]:
-    """
-    Split a multi-view image into separate view images.
-    
-    This function is called when users select multi-image examples that contain
-    multiple views in a single concatenated image. It automatically splits them
-    based on alpha channel boundaries and preprocesses each view.
-    
-    Args:
-        image (Image.Image): A concatenated image containing multiple views
-        
-    Returns:
-        List[Image.Image]: List of individual preprocessed view images
-    """
     image = np.array(image)
     alpha = image[..., 3]
     alpha = np.any(alpha>0, axis=0)
@@ -338,20 +229,7 @@ demo = gr.Blocks(
 )
 with demo:
     gr.Markdown("""
-    # 💻 ReconViaGen
-    <p align="center">
-    <a title="Github" href="https://github.com/GAP-LAB-CUHK-SZ/ReconViaGen" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
-        <img src="https://img.shields.io/github/stars/GAP-LAB-CUHK-SZ/ReconViaGen?label=GitHub%20%E2%98%85&logo=github&color=C8C" alt="badge-github-stars">
-    </a>
-    <a title="Website" href="https://jiahao620.github.io/reconviagen/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
-        <img src="https://www.obukhov.ai/img/badges/badge-website.svg">
-    </a>
-    <a title="arXiv" href="https://jiahao620.github.io/reconviagen/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
-        <img src="https://www.obukhov.ai/img/badges/badge-pdf.svg">
-    </a>
-    </p>
-    
-    ✨This demo is partial. We will release the whole model later. Stay tuned!✨
+    # 💻 ReconViaGen (Multi-GPU Force Enabled)
     """)
     
     with gr.Row():
@@ -361,9 +239,6 @@ with demo:
                     input_video = gr.Video(label="Upload Video", interactive=True, height=300)
                     image_prompt = gr.Image(label="Image Prompt", format="png", visible=False, image_mode="RGBA", type="pil", height=300)
                     multiimage_prompt = gr.Gallery(label="Image Prompt", format="png", type="pil", height=300, columns=3)
-                    gr.Markdown("""
-                        Input different views of the object in separate images.
-                                """)
         
             with gr.Accordion(label="Generation Settings", open=False):
                 seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
@@ -384,9 +259,6 @@ with demo:
 
             generate_btn = gr.Button("Generate & Extract GLB", variant="primary")
             extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)
-            gr.Markdown("""
-                        *NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*
-                        """)
 
         with gr.Column():
             video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
@@ -398,7 +270,6 @@ with demo:
     
     output_buf = gr.State()
 
-    # Example images at the bottom of the page
     with gr.Row() as multiimage_example:
         examples_multi = gr.Examples(
             examples=prepare_multi_example(),
@@ -413,25 +284,12 @@ with demo:
     demo.load(start_session)
     demo.unload(end_session)
 
-    input_video.upload(
-        preprocess_videos,
-        inputs=[input_video],
-        outputs=[multiimage_prompt],
-    )
-    input_video.clear(
-        lambda: tuple([None, None]),
-        outputs=[input_video, multiimage_prompt],
-    )
-    multiimage_prompt.upload(
-        preprocess_images,
-        inputs=[multiimage_prompt],
-        outputs=[multiimage_prompt],
-    )
+    input_video.upload(preprocess_videos, inputs=[input_video], outputs=[multiimage_prompt])
+    input_video.clear(lambda: tuple([None, None]), outputs=[input_video, multiimage_prompt])
+    multiimage_prompt.upload(preprocess_images, inputs=[multiimage_prompt], outputs=[multiimage_prompt])
 
     generate_btn.click(
-        get_seed,
-        inputs=[randomize_seed, seed],
-        outputs=[seed],
+        get_seed, inputs=[randomize_seed, seed], outputs=[seed]
     ).then(
         generate_and_extract_glb,
         inputs=[multiimage_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo, mesh_simplify, texture_size],
@@ -446,46 +304,50 @@ with demo:
         outputs=[extract_gs_btn, download_glb, download_gs],
     )
     
-    extract_gs_btn.click(
-        extract_gaussian,
-        inputs=[output_buf],
-        outputs=[model_output, download_gs],
-    ).then(
-        lambda: gr.Button(interactive=True),
-        outputs=[download_gs],
+    extract_gs_btn.click(extract_gaussian, inputs=[output_buf], outputs=[model_output, download_gs]).then(
+        lambda: gr.Button(interactive=True), outputs=[download_gs]
     )
 
     model_output.clear(
         lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]),
         outputs=[download_glb, download_gs],
     )
-    
 
-# Launch the Gradio app
+# Launch Script
 if __name__ == "__main__":
-    print("Initializing Pipeline...")
+    print("🚀 Initializing Pipeline...")
+    # [핵심 5] 모델 로딩 시점에 device를 지정하지 않고, CPU에 둡니다.
+    # 이렇게 해야 accelerate가 '빈 모델' 상태 혹은 CPU 상태에서 가져가서 쪼갤 수 있습니다.
     pipeline = TrellisVGGTTo3DPipeline.from_pretrained("esther11/trellis-vggt-v0-2")
-    pipeline.cuda()
 
-    # [수정] 멀티 GPU 처리 로직 (Model Parallelism)
-    # 기존 DataParallel은 모델을 복제하여 VRAM을 2배로 쓰므로 OOM 발생함.
-    # 대신 accelerate의 dispatch_model을 사용하여 모델을 여러 GPU에 분할 적재해야 함.
-    if torch.cuda.device_count() > 1:
-        if ACCELERATE_AVAILABLE:
-            print(f"⚡ Accelerate detected: {torch.cuda.device_count()} GPUs found.")
-            print("Applying 'device_map=auto' to VGGT_model to split layers across GPUs (Memory Efficient).")
-            # VGGT_model이 가장 무거우므로 이를 여러 GPU에 쪼개서 올림
-            pipeline.VGGT_model = dispatch_model(pipeline.VGGT_model, device_map="auto")
-        else:
-            print("⚠️ 'accelerate' library not found. Cannot split model across GPUs.")
-            print("Installing 'accelerate' (`pip install accelerate`) is highly recommended for multi-GPU inference.")
-            # 라이브러리가 없으면 기본 동작(단일 GPU or 기존 상태) 유지
+    gpu_count = torch.cuda.device_count()
+    print(f"⚡ Detected {gpu_count} GPUs.")
+
+    if gpu_count > 1:
+        print("⚡ Multi-GPU Mode Activated: Distributing model across all available GPUs.")
+        
+        # [핵심 6] VGGT 모델 분산 (가장 큰 메모리 차지)
+        # device_map="balanced"는 가능한 한 GPU 메모리를 고르게 쓰려고 노력합니다.
+        # "auto"는 첫 번째 GPU부터 채우므로 OOM 위험이 있을 수 있어 balanced가 안전합니다.
+        pipeline.VGGT_model = dispatch_model(
+            pipeline.VGGT_model, 
+            device_map="balanced" 
+        )
+        
+        # [핵심 7] SLAT 모델 분산 (그 다음으로 큼)
+        pipeline.slat_model = dispatch_model(
+            pipeline.slat_model,
+            device_map="balanced"
+        )
+        
+        # [핵심 8] 나머지 가벼운 모델들은 GPU 0에 할당 (분산 오버헤드 방지)
+        pipeline.birefnet_model.to("cuda:0")
+        
+        print("✅ Model dispatched successfully via Accelerate.")
+        
     else:
-        print(f"Running on Single GPU: {torch.cuda.get_device_name(0)}")
+        # 단일 GPU이거나 GPU가 없는 경우 (사용자가 원치 않지만 안전망으로 둠)
+        print("⚠️ Warning: Only 1 GPU detected. Loading entire model to CUDA:0 (High OOM Risk).")
+        pipeline.cuda()
 
-    # 나머지 모델들도 CUDA로 이동 (accelerate 적용 안된 경우)
-    if not ACCELERATE_AVAILABLE or torch.cuda.device_count() <= 1:
-        pipeline.VGGT_model.cuda()
-    pipeline.birefnet_model.cuda()
-    
     demo.launch()