Windows 11 + Blackwell (sm_120) + Conda 環境構築ガイド: FramePack + PyTorch Nightly (CUDA 12.8) - Qiita (2025)

nvcc --version# 出力例: release 12.8, V12.8.xnvidia-smi# 出力例: Driver Version: 552.xx 以上, CUDA Version: 12.8

conda init powershell

Set-ExecutionPolicy -Scope CurrentUser RemoteSigned -Force

conda config --env --add channels conda-forgeconda config --env --add channels pytorch-nightlyconda config --env --add channels nvidia

conda install -y cuda-version=12.8 cuda-nvcc_win-64=12.8.*

# (framepack) 環境で実行nvcc --version# 出力例: release 12.8, V12.8.x (Conda 環境内のものが優先されるはず)５. PyTorch Nightly (CUDA 12.8 版) のインストールBlackwell (sm 120) サポートが含まれる PyTorch Nightly ビルドをインストールします。

python -m pip install --upgrade pippython -m pip install --pre torch torchvision torchaudio --index-url [https://download.pytorch.org/whl/nightly/cu128]

import torchprint(torch.version.cuda)# 出力: 12.8print(torch.cuda.is_available())# 出力: Trueprint(torch.cuda.get_device_capability())# 出力: (12, 0) <- sm_120 の意味print(torch.cuda.get_arch_list())# 出力例: ['sm_120'] や ['compute_120', 'sm_120']

# 任意の作業ディレクトリに移動 (例: C:\Users\YourUser\)git clone [https://github.com/lllyasviel/FramePack.git](https://github.com/lllyasviel/FramePack.git)cd FramePackpython -m pip install -r requirements.txt

# 管理者権限で PowerShell を起動して実行reg add "HKLM\System\CurrentControlSet\Control\FileSystem" /v LongPathsEnabled /t REG_DWORD /d 1 /fgit config --global core.longpaths true

# (framepack) 環境で実行conda activate framepackpip install ninja

# FramePack ディレクトリ内に xformers をクローンする場合# cd C:\Users\YourUser\FramePack # FramePack をクローンした場所に適宜変更# 現在地が FramePack ディレクトリであると仮定git clone [https://github.com/facebookresearch/xformers.git](https://github.com/facebookresearch/xformers.git)cd xformersgit submodule update --init --recursive

# 現在の PowerShell セッションでのみ有効set TORCH_CUDA_ARCH_LIST=12.0# システム環境変数として恒久的に設定する場合 (要再起動 or 新しいターミナル)# setx TORCH_CUDA_ARCH_LIST "12.0;12.0+PTX"

# cd C:\Users\YourUser\FramePack\xformers # 必要に応じて xformers ディレクトリに移動Get-ChildItem -Path . -Recurse -Filter "*arch*.cmake" | Select-Object FullName

cmake/select_compute_arch.cmake (古いバージョン)cmake/xformers_cuda.cmake (新しいバージョン)

# 例: cmake\xformers_cuda.cmake が対象ファイルだった場合$path = "cmake\xformers_cuda.cmake" # 見つかったパスに置き換える(Get-Content $path -Raw) -replace "'70;75;80;86;90'", "'70;75;80;86;90;120'" | Set-Content $path -NoNewline

set XFORMERS_DISABLE_FLASH_ATTN=1

# cd C:\Users\YourUser\FramePack\xformers # xformers ディレクトリにいることを確認set TORCH_CUDA_ARCH_LIST=12.0 # 念のため再設定pip install -r requirements.txtpip install .

cd ..

# 現在地が FramePack ディレクトリであると仮定git clone [https://github.com/Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention)cd flash-attentionset TORCH_CUDA_ARCH_LIST=12.0 # ここでも必要python setup.py bdist_wheel# dist ディレクトリに生成された .whl ファイルをインストールpip install dist\flash_attn*.whlcd ..

# 現在地が FramePack ディレクトリであると仮定python - <<'PY'import torchimport xformersimport xformers.ops as xopsprint(f"Torch CUDA: {torch.version.cuda}")print(f"GPU arch: {torch.cuda.get_device_capability()}")print(f"xformers: {xformers.__version__}")try: # Create dummy tensors on GPU with float16 for attention check q = torch.randn(1, 1, 8, 32, device='cuda', dtype=torch.float16) k = torch.randn(1, 1, 8, 32, device='cuda', dtype=torch.float16) v = torch.randn(1, 1, 8, 32, device='cuda', dtype=torch.float16) out = xops.memory_efficient_attention(q, k, v) print("xformers memory efficient attention: OK")except Exception as e: print(f"xformers memory efficient attention: FAILED ({e})")

Torch CUDA: 12.8GPU arch: (12, 0)xformers: 0.0.26+xxxxxxxx # バージョンは変動しますxformers memory efficient attention: OK

<details><summary><code>framepack_worker.py</code> コード (クリックで展開)</summary>

# framepack_worker.pyimport osimport sysimport torchimport randomimport numpy as npfrom PIL import Imagefrom tqdm import tqdmfrom einops import rearrangefrom pathlib import Pathimport mathimport timeimport traceback# HF_HOME 環境変数を設定 (キャッシュディレクトリを指定したい場合)# os.environ['HF_HOME'] = 'huggingface_cache'# --- 必要ライブラリのインポート ---try: from pipelines.pipeline_framepack import FramePackPipeline from models.unet_3d_condition import UNet3DConditionModel from models.transformer import TransformerWrapper from models.pipeline_utils import get_quant_config, get_transformer_map from diffusers import AutoencoderKL from transformers import CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection, CLIPImageProcessorexcept ImportError as e: print(f"Error importing necessary libraries: {e}") print("Please ensure you have installed all requirements from requirements.txt in the FramePack directory.") sys.exit(1)# --- モデル初期化関数 ---def init_models(model_dir="checkpoints/stable-video-diffusion-img2vid-xt-1-1", device="cuda", dtype=torch.float16, low_vram_mode=False): # low_vram_mode は VRAM 24GB未満の場合 True を推奨 print(f"Initializing models from {model_dir}...") if not Path(model_dir).exists(): print(f"[ERROR] Model directory not found: {model_dir}") print("Please download the SVD checkpoints and place them in the specified directory.") sys.exit(1) transformer_map = get_transformer_map() # この変数は現状使われていないが将来のため残す transformer_model_path = os.path.join(model_dir, 'transformer') unet_model_path = os.path.join(model_dir, 'unet') vae_model_path = os.path.join(model_dir, 'vae') image_encoder_path = os.path.join(model_dir, 'image_encoder') text_encoder_path = os.path.join(model_dir, 'text_encoder') text_encoder_2_path = os.path.join(model_dir, 'text_encoder_2') pipeline = None # エラー時に None を返すための初期化 try: # Load models print("Loading VAE...") vae = AutoencoderKL.from_pretrained(vae_model_path, torch_dtype=dtype) print("Loading Image Encoder...") feature_extractor = CLIPImageProcessor.from_pretrained(image_encoder_path) image_encoder = CLIPVisionModelWithProjection.from_pretrained(image_encoder_path, torch_dtype=dtype) print("Loading Text Encoder 1...") tokenizer = CLIPTokenizer.from_pretrained(text_encoder_path) text_encoder = CLIPTextModel.from_pretrained(text_encoder_path, torch_dtype=dtype) print("Loading Text Encoder 2...") tokenizer_2 = CLIPTokenizer.from_pretrained(text_encoder_2_path) text_encoder_2 = CLIPTextModel.from_pretrained(text_encoder_2_path, torch_dtype=dtype) print("Loading UNet...") unet = UNet3DConditionModel.from_pretrained(unet_model_path, torch_dtype=dtype) print("Loading Transformer...") transformer = TransformerWrapper.from_pretrained(transformer_model_path, unet_config=unet.config, low_vram_mode=low_vram_mode, torch_dtype=dtype) # Move to device vae = vae.to(device).eval() image_encoder = image_encoder.to(device).eval() text_encoder = text_encoder.to(device).eval() text_encoder_2 = text_encoder_2.to(device).eval() unet.to(device).eval() transformer.to(device).eval() # Check VRAM and determine high_vram flag high_vram = False if device == "cuda" and torch.cuda.is_available(): try: gpu_properties = torch.cuda.get_device_properties(torch.cuda.current_device()) total_memory_gb = gpu_properties.total_memory / (1024 ** 3) print(f"Detected GPU: {gpu_properties.name} with {total_memory_gb:.2f} GB VRAM") if total_memory_gb >= 20: # 20GB以上を目安とする high_vram = True except Exception as e: print(f"Could not detect GPU VRAM: {e}") elif device != "cuda": print("Running on CPU or other device, high_vram set to False.") else: print("CUDA not available, high_vram set to False.") print(f"High VRAM mode determined: {high_vram}") # Create pipeline (do this *after* moving models to device) pipeline = FramePackPipeline( vae=vae, text_encoder=text_encoder, text_encoder_2=text_encoder_2, tokenizer=tokenizer, tokenizer_2=tokenizer_2, unet=unet, transformer=transformer, feature_extractor=feature_extractor, image_encoder=image_encoder ) # pipeline.enable_model_cpu_offload() # 必要に応じて有効化 print("Models initialized successfully.") return high_vram, text_encoder, text_encoder_2, tokenizer, tokenizer_2, vae, feature_extractor, image_encoder, transformer, pipeline except Exception as e: print(f"[ERROR] Failed to initialize models: {e}") traceback.print_exc() # Return None for all models if initialization fails return False, None, None, None, None, None, None, None, None, None# --- 動画生成ワーカー関数 ---def worker(input_image: Image.Image, prompt: str, pipeline: FramePackPipeline, # Pass the initialized pipeline transformer: TransformerWrapper, # Needed for TeaCache control n_prompt: str = "", seed: int = 42, total_second_length: float = 6.0, target_fps: int = 30, latent_window_size: int = 32, steps: int = 25, cfg: float = 1.0, gs: float = 7.0, rs: float = 0.0, gpu_memory_preservation: float = 6.0, #GB use_teacache: bool = True, chunk_size: int = 16, # Smaller chunk for lower VRAM outputs_folder: str = "outputs", high_vram: bool = False): if pipeline is None or transformer is None: print("[ERROR] Pipeline or Transformer not initialized. Cannot generate video.") return None try: start_time = time.time() output_dir = Path(outputs_folder) output_dir.mkdir(parents=True, exist_ok=True) # Determine frames based on seconds and FPS total_frames = math.ceil(total_second_length * target_fps) print(f"Target: {total_second_length} seconds, {target_fps} FPS => {total_frames} frames") # Adjust chunk size based on VRAM if needed if not high_vram and chunk_size > 8: print(f"Low VRAM detected, reducing decode_chunk_size from {chunk_size} to 8") decode_chunk_size = 8 else: decode_chunk_size = chunk_size # --- Run Pipeline --- generator = torch.Generator(device=pipeline.device).manual_seed(seed) # Enable TeaCache if requested and available teacache_enabled_here = False if use_teacache and hasattr(transformer, 'enable_teacache'): print("Enabling TeaCache...") transformer.enable_teacache(chunk_size=latent_window_size) # chunk_size for teacache might relate to latent_window teacache_enabled_here = True elif use_teacache: print("[WARN] TeaCache requested but 'enable_teacache' not found on transformer.") # Resize input image - SVD expects 1024x576 img_w, img_h = input_image.size if img_w != 1024 or img_h != 576: print(f"Resizing input image from {img_w}x{img_h} to 1024x576") input_image = input_image.resize((1024, 576)) output_video_path = pipeline( prompt=prompt, negative_prompt=n_prompt, image=input_image, num_inference_steps=steps, guidance_scale=gs, num_frames=total_frames, latent_window_size=latent_window_size, cfg_scale=cfg, revision_scale=rs, generator=generator, output_type="filepath", # Get filepath directly output_folder=str(output_dir), fps=target_fps, memory_limit_gb=gpu_memory_preservation, decode_chunk_size=decode_chunk_size # Use potentially adjusted chunk size ) # Disable TeaCache after use if it was enabled if teacache_enabled_here and hasattr(transformer, 'disable_teacache'): print("Disabling TeaCache...") transformer.disable_teacache() end_time = time.time() elapsed = end_time - start_time print(f"Video generation took {elapsed:.2f} seconds.") if output_video_path and Path(output_video_path).exists(): print(f"Saved video to: {output_video_path}") return output_video_path else: print("[ERROR] Pipeline finished but output video path not found or invalid.") return None except Exception as e: print(f"[ERROR] Error during video generation: {e}") traceback.print_exc() # Ensure teacache is disabled in case of error if teacache_enabled_here and hasattr(transformer, 'disable_teacache'): try: transformer.disable_teacache() except Exception as te: print(f"Error disabling teacache after exception: {te}") return None# --- Self-test block ---if __name__ == '__main__': print("--- Testing framepack_worker module ---") # --- Initialize --- # Set low_vram_mode=True if you have < 24GB VRAM for testing high_vram, text_encoder, text_encoder_2, tokenizer, tokenizer_2, vae, \ feature_extractor, image_encoder, transformer, pipeline = init_models(low_vram_mode=False) if pipeline is None: print("[FAIL] Model initialization failed. Aborting test.") else: print("[OK] Model initialization successful.") # --- Prepare Inputs --- test_img_path = Path("test_input_worker.png") test_output_folder = "worker_test_output" # Create a dummy image if none exists if not test_img_path.exists(): print(f"Creating dummy input image: {test_img_path}") try: Image.new('RGB', (1024, 576), color = 'blue').save(test_img_path) except Exception as e: print(f"[FAIL] Could not create dummy image: {e}") sys.exit(1) try: input_img = Image.open(test_img_path).convert("RGB") print(f"Loaded test image: {test_img_path}") except Exception as e: print(f"[FAIL] Could not load test image {test_img_path}: {e}") sys.exit(1) test_prompt = "a cinematic shot of a fluffy cat wearing a tiny wizard hat, exploring a mystical library" # --- Run Worker --- print("Running worker function...") mp4_path = worker( input_image=input_img, prompt=test_prompt, pipeline=pipeline, # Pass initialized pipeline transformer=transformer, # Pass transformer for TeaCache seed=12345, total_second_length=2.0, # Short video for testing target_fps=15, # Lower FPS for faster test steps=10, # Fewer steps for faster test outputs_folder=test_output_folder, high_vram=high_vram, # Use detected VRAM mode use_teacache=True # Test with TeaCache ) if mp4_path and Path(mp4_path).exists(): print(f"[PASS] Worker test successful, video saved to {mp4_path}") else: print("[FAIL] Worker test failed.")&lt;/details>

JSON{ "DancePrompts": [ { "Prompt Number": 1, "Dance Prompt": "A girl starts hip-hop dance on a vibrant street stage, dynamic moves, energetic rhythm, urban background." }, { "Prompt Number": 2, "Dance Prompt": "A boy breakdancing on the street, power moves, spinning on the ground, graffiti wall background." }, { "Prompt Number": 3, "Dance Prompt": "Ballet dancer performing elegant pirouettes in a classical theater setting." } ]}

PlaintextFramePack/│ generate_dance_videos.py│ framepack_worker.py│ DancePrompts.json│ ... (other FramePack files)│└───dance_first_frame/│ 1.png│ 2.png│ 3.png│ ...│└───outputs/ (This will be created by the script)

# generate_dance_videos.pyimport jsonimport osfrom pathlib import Pathfrom PIL import Imageimport timeimport sys# --- スクリプトがあるディレクトリを基準にする ---SCRIPT_DIR = Path(__file__).parent.resolve()# --- 設定 ---# 相対パスで指定BASE_IMG_DIR = SCRIPT_DIR / "dance_first_frame" # 入力画像フォルダOUTPUT_DIR = SCRIPT_DIR / "outputs" # 出力動画フォルダJSON_PATH = SCRIPT_DIR / "DancePrompts.json" # JSONファイルパスMODEL_DIR = SCRIPT_DIR / "checkpoints/stable-video-diffusion-img2vid-xt-1-1" # モデルディレクトリTOTAL_SEC = 15.0 # 生成する動画の長さ（秒）TARGET_FPS = 30 # 目標フレームレートSTEPS = 25 # 推論ステップ数LAT_WIN = 32 # Latent Window SizeCFG = 1.0GS = 10.0 # Guidance ScaleRS = 0.0 # Revision ScaleGPU_MEM = 6.0 # 予約するGPUメモリ(GB) - これを超えるとCPUに移される可能性USE_TEACACHE = True # TeaCache を使うかCHUNK_SIZE = 16 # デコード/エンコードのチャンクサイズ (低VRAMなら小さく: 8 or 4)LOW_VRAM_MODE_INIT = False # モデル初期化時に低VRAMモードを使うか (例: < 24GBならTrue)# --- 設定ここまで ---# --- framepack_worker モジュールをインポート ---try: # スクリプトのディレクトリをPythonのパスに追加してインポートを試みる sys.path.insert(0, str(SCRIPT_DIR)) from framepack_worker import init_models, workerexcept ImportError: print("[ERROR] framepack_worker.py not found or cannot be imported.") print(f"Ensure framepack_worker.py is in the directory: {SCRIPT_DIR}") sys.exit(1)except Exception as e: print(f"[ERROR] Failed to import from framepack_worker: {e}") sys.exit(1)print("--- FramePack Batch Video Generation ---")print(f"Script directory: {SCRIPT_DIR}")print(f"Input image directory: {BASE_IMG_DIR}")print(f"Output directory: {OUTPUT_DIR}")print(f"JSON path: {JSON_PATH}")print(f"Model directory: {MODEL_DIR}")# 1. モデルの初期化 (一度だけ実行)print("\nInitializing models...")init_start_time = time.time()# モデルディレクトリを絶対パスに変換して渡すabsolute_model_dir = str(MODEL_DIR.resolve())high_vram, text_encoder, text_encoder_2, tokenizer, tokenizer_2, vae, \feature_extractor, image_encoder, transformer, pipeline = init_models( model_dir=absolute_model_dir, low_vram_mode=LOW_VRAM_MODE_INIT)# Check if initialization was successfulif pipeline is None: print("\n[FATAL] Model initialization failed. Please check logs and model paths. Exiting.") sys.exit(1)init_end_time = time.time()print(f"Model initialization took {init_end_time - init_start_time:.2f} seconds.")print(f"High VRAM mode detected by worker: {high_vram}") # Display the detected mode# 2. JSONデータの読み込みif not JSON_PATH.exists(): print(f"\n[ERROR] JSON file not found at: {JSON_PATH}") sys.exit(1)print(f"\nLoading prompts from {JSON_PATH.name}...")try: with open(JSON_PATH, encoding="utf-8") as f: table = json.load(f)["DancePrompts"] print(f"Found {len(table)} prompts.")except KeyError: print(f"[ERROR] JSON file must contain a key named 'DancePrompts' with a list of prompts.") sys.exit(1)except Exception as e: print(f"[ERROR] Failed to load or parse JSON: {e}") sys.exit(1)# 3. 出力ディレクトリ作成print(f"\nEnsuring output directory exists: {OUTPUT_DIR}")OUTPUT_DIR.mkdir(parents=True, exist_ok=True)# 4. 動画生成ループtotal_videos = len(table)generated_count = 0failed_count = 0total_gen_time = 0print("\n--- Starting Video Generation Loop ---")for i, row in enumerate(table): item_start_time = time.time() print("-" * 60) print(f"Processing item {i+1}/{total_videos}") try: idx = row["Prompt Number"] prompt = row["Dance Prompt"] print(f"Prompt Number: {idx}") print(f"Prompt: {prompt[:80]}...") # Display first 80 chars except KeyError as e: print(f"[WARN] Skipping row {i+1} due to missing key: {e}") failed_count += 1 continue # 画像パスを絶対パスに変換 img_path = (BASE_IMG_DIR / f"{idx}.png").resolve() print(f"Looking for input image: {img_path}") if not img_path.exists(): print(f"[!] Image not found, skipping.") failed_count += 1 continue elif not img_path.is_file(): print(f"[!] Path exists but is not a file, skipping: {img_path}") failed_count += 1 continue try: input_image = Image.open(img_path).convert("RGB") print(f"Loaded image: {img_path.name} ({input_image.width}x{input_image.height})") except Exception as e: print(f"[!] Failed to open image {img_path}, skipping: {e}") failed_count += 1 continue # --- worker 関数呼び出し --- print("Calling worker function...") gen_start_time = time.time() # 出力フォルダも絶対パスで渡す absolute_output_dir = str(OUTPUT_DIR.resolve()) # Pass all necessary arguments, including the initialized pipeline and transformer mp4_temp_path_str = worker( input_image=input_image, prompt=prompt, pipeline=pipeline, # Pass the initialized pipeline transformer=transformer, # Pass the initialized transformer n_prompt="", # Add negative prompt if needed seed=31337 + idx, # Use a unique seed per video total_second_length=TOTAL_SEC, target_fps=TARGET_FPS, latent_window_size=LAT_WIN, steps=STEPS, cfg=CFG, gs=GS, rs=RS, gpu_memory_preservation=GPU_MEM, use_teacache=USE_TEACACHE, chunk_size=CHUNK_SIZE, outputs_folder=absolute_output_dir, # Pass absolute path as string high_vram=high_vram # Pass the detected VRAM mode ) gen_end_time = time.time() loop_gen_time = gen_end_time - gen_start_time total_gen_time += loop_gen_time if mp4_temp_path_str: mp4_temp_path = Path(mp4_temp_path_str) # 分かりやすいファイル名に変更 (出力ディレクトリ基準で) new_name = OUTPUT_DIR / f"dance_{idx}.mp4" try: # Rename the file if it exists if mp4_temp_path.exists() and mp4_temp_path.is_file(): # If the returned path is already the final name, skip renaming if mp4_temp_path.resolve() != new_name.resolve(): mp4_temp_path.rename(new_name) print(f" → Renamed to {new_name.name}") else: print(f" → Saved as {new_name.name} (already correct name)") print(f" → Success! (Generation took {loop_gen_time:.2f} seconds)") generated_count += 1 else: print(f"[!] Worker returned path but file not found or invalid: {mp4_temp_path_str}") print(f" → Failed! (Generation attempt took {loop_gen_time:.2f} seconds)") failed_count += 1 except Exception as e: print(f"[!] Failed to rename {mp4_temp_path.name} to {new_name.name}: {e}") print(f" → Original file might be at: {mp4_temp_path_str}") # Show original path print(f" → Failed! (Generation attempt took {loop_gen_time:.2f} seconds)") failed_count += 1 else: print(f"[!] Video generation failed for prompt {idx}.") print(f" → Failed! (Generation attempt took {loop_gen_time:.2f} seconds)") failed_count += 1 item_end_time = time.time() print(f"Total time for item {idx}: {item_end_time - item_start_time:.2f} seconds")print("-" * 60)print("--- Generation Complete ---")print(f"Summary:")print(f" Total prompts processed: {total_videos}")print(f" Successfully generated: {generated_count}")print(f" Failed/Skipped: {failed_count}")if generated_count > 0: average_time = total_gen_time / generated_count print(f" Average generation time per successful video: {average_time:.2f} seconds.")print("-" * 60)

# (framepack) 環境で FramePack ディレクトリに移動していることを確認# cd C:\Path\To\Your\FramePack # 必要に応じて移動conda activate framepackpython demo_gradio.py --inbrowser

# 起動ログの例PyTorch compiled with CUDA: 12.8GPU capability: 12.0Using xformers: True # xFormersが有効になっているか[...]Running on local URL: [http://127.0.0.1:7860](http://127.0.0.1:7860)

# (framepack) 環境で FramePack ディレクトリに移動していることを確認# cd C:\Path\To\Your\FramePack # 必要に応じて移動conda activate framepackpython generate_dance_videos.py