diff --git a/.github/workflows/build-release-conda.yml b/.github/workflows/build-release-conda.yml index 9f7f1ff..e7fa697 100644 --- a/.github/workflows/build-release-conda.yml +++ b/.github/workflows/build-release-conda.yml @@ -212,6 +212,11 @@ jobs: python -c "import ctranslate2; print(f'CTranslate2 version: {ctranslate2.__version__}')" echo "Note: CUDA availability check skipped (no GPU on GitHub runners)" + - name: Install Modal dependencies + run: | + pip install modal questionary + python -c "import modal; print(f'Modal version: {modal.__version__}')" + - name: Check cached models run: | echo "Checking for cached models..." @@ -377,6 +382,7 @@ jobs: run: | cd dist\faster_whisper_transwithai_chickenrice infer.exe --help + modal_infer.exe --help - name: Upload artifact uses: actions/upload-artifact@v4 @@ -647,4 +653,4 @@ jobs: repository: ${{ github.repository }} tag_name: ${{ github.ref }} files: faster_whisper_transwithai_windows_cu128-chickenrice.zip - token: ${{ secrets.GITHUB_TOKEN }} \ No newline at end of file + token: ${{ secrets.GITHUB_TOKEN }} diff --git a/.gitignore b/.gitignore index 0d6661a..8ff3ddd 100644 --- a/.gitignore +++ b/.gitignore @@ -10,6 +10,7 @@ __pycache__/ # C extensions *.so +node_modules/ # Distribution / packaging .Python build/ diff --git a/environment-cuda118.yml b/environment-cuda118.yml index 021b529..32d2dc0 100644 --- a/environment-cuda118.yml +++ b/environment-cuda118.yml @@ -37,7 +37,8 @@ dependencies: - backports.functools-lru-cache # Fix for PyInstaller ModuleNotFoundError # Build tools - - pyinstaller>=6.0.0 + # PyInstaller 6.17.0 has a conda hook regression (KeyError: 'depends') in some conda environments + - pyinstaller==6.16.0 - setuptools>=65.0.0 - wheel>=0.38.0 - build>=0.10.0 diff --git a/environment-cuda122.yml b/environment-cuda122.yml index f4ac70a..c338836 100644 --- a/environment-cuda122.yml +++ b/environment-cuda122.yml @@ -34,7 +34,8 @@ dependencies: - backports.functools-lru-cache # Fix for PyInstaller ModuleNotFoundError # Build tools - - pyinstaller>=6.0.0 + # PyInstaller 6.17.0 has a conda hook regression (KeyError: 'depends') in some conda environments + - pyinstaller==6.16.0 - setuptools>=65.0.0 - wheel>=0.38.0 - build>=0.10.0 diff --git a/environment-cuda128.yml b/environment-cuda128.yml index 6a1d637..b9f2ca5 100644 --- a/environment-cuda128.yml +++ b/environment-cuda128.yml @@ -34,7 +34,8 @@ dependencies: - backports.functools-lru-cache # Fix for PyInstaller ModuleNotFoundError # Build tools - - pyinstaller>=6.0.0 + # PyInstaller 6.17.0 has a conda hook regression (KeyError: 'depends') in some conda environments + - pyinstaller==6.16.0 - setuptools>=65.0.0 - wheel>=0.38.0 - build>=0.10.0 diff --git a/environment-modal.yml b/environment-modal.yml new file mode 100644 index 0000000..2bef4a6 --- /dev/null +++ b/environment-modal.yml @@ -0,0 +1,21 @@ +# Conda environment for Modal inference (local client only) +# This environment is for running modal_infer.py locally to submit jobs to Modal +name: faster-whisper-modal +channels: + - conda-forge + - defaults + +dependencies: + # Python version + - python=3.10 + + # Core dependencies + - pip + + # Pip dependencies + - pip: + # Modal client for submitting jobs + - modal + + # Interactive CLI prompts + - questionary diff --git a/modal_infer.py b/modal_infer.py new file mode 100644 index 0000000..d6abd78 --- /dev/null +++ b/modal_infer.py @@ -0,0 +1,805 @@ +"""feature-modal: 交互式 CLI，完成 Modal App 构建、音频上传、推理执行与结果回传。""" + +from __future__ import annotations + +import argparse +import logging +import sys +from dataclasses import dataclass +from datetime import datetime +from pathlib import Path +import subprocess +from typing import Dict, List, Optional, Sequence, Tuple +from uuid import uuid4 + +try: + import questionary # type: ignore + from questionary import Choice # type: ignore +except ImportError: # pragma: no cover + questionary = None # type: ignore + Choice = None # type: ignore + +try: + import modal +except ImportError as exc: # pragma: no cover + print("未检测到 modal 包，请先运行 `python -m pip install modal questionary`。") + raise + +APP_NAME = "Faster-Whisper-TransWithAI-ChickenRice" +REPO_URL = "https://github.com/TransWithAI/Faster-Whisper-TransWithAI-ChickenRice" +VOLUME_NAME = "Faster_Whisper" +VOLUME_ROOT = Path("/Faster_Whisper") +REMOTE_MOUNT = VOLUME_ROOT +APP_ROOT_REL = Path(APP_NAME) +SESSION_SUBDIR = Path("sessions") +REPO_VOLUME_DIR = VOLUME_ROOT / "repo" +SUB_FORMATS = "srt,vtt,lrc" +SUB_SUFFIXES = {".srt", ".vtt", ".lrc"} +AUDIO_SUFFIXES = { + ".mp3", + ".wav", + ".flac", + ".m4a", + ".aac", + ".ogg", + ".wma", + ".mkv", + ".avi", + ".mov", + ".webm", + ".flv", + ".wmv", +} +VIDEO_NEED_CONVERT = {".mp4"} # 需要用户手动转换的格式 +DEFAULT_GPU_CHOICES = [ + "T4", + "L4", + "L40S", + "A10G", + "A100-40GB", + "A100-80GB", + "H100", + "H200", + "B200", +] + + +@dataclass +class ModelProfile: + key: str + label: str + hf_repo: Optional[str] + target_dir: Optional[str] + description: str + + +@dataclass +class UserSelection: + run_mode: str # once or persistent + gpu_choice: str + input_path: Path + model_profile: ModelProfile + custom_repo: Optional[str] + custom_target_dir: Optional[str] + enable_batching: bool + batch_size: Optional[int] + max_batch_size: int + timeout_minutes: int + + +@dataclass +class UploadManifest: + session_id: str + source_type: str # file or directory + local_source: Path + remote_inputs_rel: List[Path] + remote_output_rel: Path + local_output_dir: Path + remote_logs_rel: Path + original_filename: Optional[str] = None # 原始文件名（用于恢复空格） + + +@dataclass +class ScanResult: + audio_files: List[Path] + mp4_files: List[Path] + + +class NoAudioFilesError(Exception): + pass + + +def rel_to_volume_path(path: Path) -> str: + posix = path.as_posix() + if not posix.startswith("/"): + posix = "/" + posix + return posix + + +def rel_to_container_path(path: Path) -> str: + return str((REMOTE_MOUNT / path).as_posix()) + + +def volume_path_to_relative(path: str) -> Path: + return Path(path.lstrip("/")) + + +def container_to_volume_path(container_path: str) -> str: + prefix = str(REMOTE_MOUNT) + if not container_path.startswith(prefix): + raise ValueError(f"路径 {container_path} 不在挂载点 {prefix} 下") + rel = container_path[len(prefix) :] + if not rel.startswith("/"): + rel = "/" + rel + return rel + + +MODEL_PRESETS: Dict[str, ModelProfile] = { + "chickenrice": ModelProfile( + key="chickenrice", + label="海南鸡（日文转中文优化）", + hf_repo="chickenrice0721/whisper-large-v2-translate-zh-v0.2-st-ct2", + target_dir="chickenrice-v2", + description="默认高精度模型，建议 L40S 以上", + ), + "base": ModelProfile( + key="base", + label="基础版（whisper-base）", + hf_repo="openai/whisper-base", + target_dir="whisper-base", + description="适合低显存/快速校验", + ), + "custom": ModelProfile( + key="custom", + label="自定义 HuggingFace 模型", + hf_repo=None, + target_dir=None, + description="手动输入 HF repo 与目标目录", + ), +} + + +def setup_logger() -> Path: + logs_dir = Path("logs") + logs_dir.mkdir(exist_ok=True) + timestamp = datetime.now().strftime("%Y%m%d_%H%M%S") + log_path = logs_dir / f"modal_run_{timestamp}.log" + + logger = logging.getLogger() + logger.setLevel(logging.INFO) + formatter = logging.Formatter("%(asctime)s [%(levelname)s] %(message)s") + + fh = logging.FileHandler(log_path, encoding="utf-8") + fh.setFormatter(formatter) + logger.handlers.clear() + logger.addHandler(fh) + + sh = logging.StreamHandler(sys.stdout) + sh.setFormatter(formatter) + logger.addHandler(sh) + + logging.info("日志输出：%s", log_path) + return log_path + + +def ensure_questionary(): + if questionary is None or Choice is None: + raise RuntimeError("需要 questionary，请运行 `python -m pip install questionary`。") + + +def ask_selection() -> UserSelection: + ensure_questionary() + + gpu_choice = questionary.select( + "选择 GPU", + choices=DEFAULT_GPU_CHOICES, + ).ask() + if not gpu_choice: + raise KeyboardInterrupt + + model_key = questionary.select( + "选择模型：", + choices=[Choice(title=profile.label, value=key) for key, profile in MODEL_PRESETS.items()], + ).ask() + if not model_key: + raise KeyboardInterrupt + + model_profile = MODEL_PRESETS[model_key] + custom_repo = None + custom_target_dir = None + if model_key == "custom": + custom_repo = questionary.text("输入 HuggingFace repo（例如 user/repo）").ask() + if not custom_repo: + raise KeyboardInterrupt + custom_target_dir = questionary.text( + "输入 models 子目录名称（英文/数字）", default="custom-model" + ).ask() + if not custom_target_dir: + raise KeyboardInterrupt + + input_path_str = questionary.path( + "拖入或输入待处理的本地文件/文件夹路径：" + ).ask() + if not input_path_str: + raise KeyboardInterrupt + input_path = Path(input_path_str).expanduser().resolve() + if not input_path.exists(): + raise FileNotFoundError(f"路径不存在：{input_path}") + + enable_batching = questionary.confirm( + "启用批处理以加速（需要更高显存）？", default=False + ).ask() + if enable_batching is None: + raise KeyboardInterrupt + + batch_size = None + max_batch_size = 8 + if enable_batching: + batch_size_str = questionary.text( + "指定批次大小（留空自动探测）", default="" + ).ask() + if batch_size_str: + batch_size = int(batch_size_str) + max_batch_size_str = questionary.text( + "最大自动批次大小", default="8" + ).ask() + max_batch_size = int(max_batch_size_str or "8") + + timeout_minutes = int( + questionary.text("任务超时时间（分钟）", default="60").ask() or "60" + ) + + return UserSelection( + run_mode="once", + gpu_choice=gpu_choice, + input_path=input_path, + model_profile=model_profile, + custom_repo=custom_repo, + custom_target_dir=custom_target_dir, + enable_batching=bool(enable_batching), + batch_size=batch_size, + max_batch_size=max_batch_size, + timeout_minutes=timeout_minutes, + ) + + +def scan_audio_files(path: Path) -> ScanResult: + """扫描目录，返回音频文件和需要转换的 mp4 文件""" + audio_files: List[Path] = [] + mp4_files: List[Path] = [] + for file in path.rglob("*"): + if file.is_file(): + suffix = file.suffix.lower() + if suffix in AUDIO_SUFFIXES: + audio_files.append(file) + elif suffix in VIDEO_NEED_CONVERT: + mp4_files.append(file) + return ScanResult(audio_files=audio_files, mp4_files=mp4_files) + + +def validate_audio_path(path: Path) -> ScanResult: + """验证音频路径，返回扫描结果。如果发现 mp4 文件会打印警告。""" + if path.is_file(): + suffix = path.suffix.lower() + if suffix in VIDEO_NEED_CONVERT: + raise ValueError( + f"文件 {path} 是 mp4 格式，请先使用 ffmpeg 转换为 mp3：\n" + f" ffmpeg -i \"{path}\" -vn -acodec libmp3lame \"{path.with_suffix('.mp3')}\"" + ) + if suffix not in AUDIO_SUFFIXES: + raise ValueError(f"文件 {path} 不属于支持的音/视频格式。") + return ScanResult(audio_files=[path], mp4_files=[]) + elif path.is_dir(): + scan_result = scan_audio_files(path) + if scan_result.mp4_files: + logging.warning("=" * 60) + logging.warning("发现 %d 个 mp4 文件，这些文件将被跳过：", len(scan_result.mp4_files)) + for mp4_file in scan_result.mp4_files: + logging.warning(" - %s", mp4_file) + logging.warning("请使用 ffmpeg 转换为 mp3 后再处理，例如：") + logging.warning(" ffmpeg -i \"input.mp4\" -vn -acodec libmp3lame \"output.mp3\"") + logging.warning("=" * 60) + if not scan_result.audio_files: + raise NoAudioFilesError(f"输入的文件夹内没有音频文件：{path}") + return scan_result + else: + raise ValueError(f"路径 {path} 既不是文件也不是文件夹。") + + +def upload_single_file( + volume: modal.Volume, + selection: UserSelection, + audio_file: Path, + base_dir: Optional[Path] = None, +) -> UploadManifest: + """上传单个音频文件到 Modal Volume。 + + Args: + volume: Modal Volume 实例 + selection: 用户选择配置 + audio_file: 要上传的音频文件路径 + base_dir: 基础目录（用于文件夹模式，输出到此目录） + """ + session_id = f"{datetime.now().strftime('%Y%m%d-%H%M%S')}-{uuid4().hex[:6]}" + remote_session_rel = SESSION_SUBDIR / session_id + remote_logs_rel = remote_session_rel / "logs" + + # 使用固定文件名避免全角字符等问题 + original_filename = audio_file.name + safe_filename = "todo" + audio_file.suffix.lower() + + with volume.batch_upload(force=True) as batch: + remote_rel = remote_session_rel / safe_filename + logging.info("上传文件 -> %s", rel_to_volume_path(remote_rel)) + batch.put_file(str(audio_file), rel_to_volume_path(remote_rel)) + + # 如果指定了 base_dir（文件夹模式），输出到 base_dir；否则输出到文件所在目录 + local_output_dir = base_dir if base_dir else audio_file.parent + + return UploadManifest( + session_id=session_id, + source_type="file", + local_source=audio_file, + remote_inputs_rel=[remote_rel], + remote_output_rel=remote_session_rel, + local_output_dir=local_output_dir, + remote_logs_rel=remote_logs_rel, + original_filename=original_filename, # 始终记录原始文件名 + ) + + +def build_job_payload(selection: UserSelection, manifest: UploadManifest) -> Dict: + model_profile = selection.model_profile + hf_repo = selection.custom_repo if model_profile.key == "custom" else model_profile.hf_repo + target_dir = ( + selection.custom_target_dir if model_profile.key == "custom" else model_profile.target_dir + ) or "custom-model" + + payload = { + "session_id": manifest.session_id, + "mount_root": str(REMOTE_MOUNT), + "repo_url": REPO_URL, + "remote_inputs": [rel_to_container_path(p) for p in manifest.remote_inputs_rel], + "remote_output_dir": rel_to_container_path(manifest.remote_output_rel), + "output_targets": [ + { + "remote_dir": rel_to_container_path(manifest.remote_output_rel), + "extensions": list(SUB_SUFFIXES), + } + ], + "input_mode": manifest.source_type, + "sub_formats": SUB_FORMATS, + "enable_batching": selection.enable_batching, + "batch_size": selection.batch_size, + "max_batch_size": selection.max_batch_size, + "timeout_seconds": selection.timeout_minutes * 60, + "model_profile": { + "label": model_profile.label, + "hf_repo": hf_repo, + "target_dir": target_dir, + }, + "remote_logs_dir": rel_to_container_path(manifest.remote_logs_rel), + "output_suffixes": list(SUB_SUFFIXES), + } + return payload + + +def build_modal_image() -> modal.Image: + return ( + modal.Image.micromamba(python_version="3.10") + .apt_install("git") + .micromamba_install( + spec_file="environment-cuda128.yml", + channels=["conda-forge", "defaults"], + ) + .pip_install("modal", "questionary") + ) + + +def run_remote_pipeline( + volume: modal.Volume, + selection: UserSelection, + manifest: UploadManifest, + payload: Dict, +) -> Dict: + logging.info("=== 开始构建 Modal 镜像 ===") + image = build_modal_image() + logging.info("✓ 镜像构建完成") + logging.info("使用 GPU：%s", selection.gpu_choice) + logging.info("超时时间：%d 分钟", selection.timeout_minutes) + app = modal.App(APP_NAME) + + @app.function( + image=image, + gpu=selection.gpu_choice, + timeout=selection.timeout_minutes * 60, + volumes={str(REMOTE_MOUNT): volume}, + serialized=True, + ) + def modal_pipeline(job_payload: Dict) -> Dict: + return _remote_pipeline(job_payload) + + logging.info("=== 开始远程执行 ===") + logging.info("正在启动 GPU 容器并执行推理任务...") + logging.info("（以下为远程容器输出）") + logging.info("-" * 60) + with app.run(): + result = modal_pipeline.remote(payload) + logging.info("-" * 60) + logging.info("✓ 远程执行完成") + return result # 直接返回 Dict，包含 created_files 和 log_content + + +def process_directory_files( + volume: modal.Volume, + selection: UserSelection, + audio_files: List[Path], +) -> Tuple[int, int]: + """处理文件夹中的所有音频文件，容器复用。 + + Args: + volume: Modal Volume 实例 + selection: 用户选择配置 + audio_files: 要处理的音频文件列表 + + Returns: + (成功数, 失败数) 元组 + """ + logging.info("=== 开始构建 Modal 镜像 ===") + image = build_modal_image() + logging.info("✓ 镜像构建完成") + logging.info("使用 GPU：%s", selection.gpu_choice) + logging.info("超时时间：%d 分钟", selection.timeout_minutes) + logging.info("待处理文件数：%d", len(audio_files)) + + app = modal.App(APP_NAME) + + @app.function( + image=image, + gpu=selection.gpu_choice, + timeout=selection.timeout_minutes * 60, + volumes={str(REMOTE_MOUNT): volume}, + serialized=True, + min_containers=1, # 保持容器预热，复用容器 + ) + def modal_pipeline(job_payload: Dict) -> Dict: + return _remote_pipeline(job_payload) + + success_count = 0 + fail_count = 0 + base_dir = selection.input_path # 文件夹模式下，输出到源文件夹 + + with app.run(): + for i, audio_file in enumerate(audio_files, 1): + logging.info("=" * 60) + logging.info("处理文件 [%d/%d]: %s", i, len(audio_files), audio_file.name) + logging.info("=" * 60) + try: + # 1. 上传单个文件 + manifest = upload_single_file(volume, selection, audio_file, base_dir) + + # 2. 构建 payload + payload = build_job_payload(selection, manifest) + + # 3. 执行推理（复用容器） + logging.info("正在执行推理...") + result = modal_pipeline.remote(payload) + + # 4. 写入结果文件到本地 + download_outputs(manifest, result) + + logging.info("✓ 文件 %s 处理完成", audio_file.name) + success_count += 1 + except Exception as e: + logging.error("✗ 文件 %s 处理失败: %s", audio_file.name, e) + fail_count += 1 + continue # 继续处理下一个文件 + + return success_count, fail_count + + +def download_outputs( + manifest: UploadManifest, + result: Dict, +) -> None: + """从远程结果中提取文件内容并写入本地""" + import base64 + + created_files = result.get("created_files", {}) + log_content = result.get("log_content") + + # 获取原始文件名的 stem（不含扩展名） + original_stem = Path(manifest.original_filename).stem if manifest.original_filename else "todo" + + for filename, content_b64 in created_files.items(): + content = base64.b64decode(content_b64) + # 将 todo.xxx 替换为原始文件名 + if filename.startswith("todo."): + suffix = Path(filename).suffix + new_filename = original_stem + suffix + else: + new_filename = filename + + local_path = manifest.local_output_dir / new_filename + local_path.parent.mkdir(parents=True, exist_ok=True) + local_path.write_bytes(content) + logging.info("写入文件: %s (%d bytes)", local_path, len(content)) + + # 写入 log 文件 + if log_content: + log_dir = Path("logs") + log_dir.mkdir(exist_ok=True) + log_path = log_dir / f"modal_run_{manifest.session_id}.log" + log_path.write_bytes(base64.b64decode(log_content)) + logging.info("写入日志: %s", log_path) + + +def summarize(manifest: UploadManifest, result: Dict) -> None: + logging.info("=== 运行完成 ===") + logging.info("Session: %s", manifest.session_id) + logging.info("源路径: %s", manifest.local_source) + logging.info("输出路径: %s", manifest.local_output_dir if manifest.source_type == "directory" else manifest.local_source.parent) + created_files = result.get("created_files", {}) + if created_files: + logging.info("新生成文件：") + for filename in created_files.keys(): + logging.info(" %s", filename) + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="feature-modal 运行脚本") + parser.add_argument( + "--non-interactive", + action="store_true", + help="跳过 questionary，全部通过参数指定（暂未实现）。", + ) + return parser.parse_args() + + +def prompt_exit(enabled: bool) -> None: + if not enabled: + return + try: + input("输入任意键退出...") + except EOFError: + pass + + +def main() -> int: + args = parse_args() + log_path = setup_logger() + exit_code = 0 + try: + selection = ask_selection() + volume = modal.Volume.from_name(VOLUME_NAME, create_if_missing=True) + + # 验证路径并获取扫描结果 + scan_result = validate_audio_path(selection.input_path) + + if selection.input_path.is_dir(): + # 文件夹模式：逐个处理文件，容器复用 + logging.info("检测到文件夹输入，将逐个处理 %d 个音频文件", len(scan_result.audio_files)) + success_count, fail_count = process_directory_files( + volume, selection, scan_result.audio_files + ) + logging.info("=" * 60) + logging.info("=== 批量处理完成 ===") + logging.info("成功: %d, 失败: %d", success_count, fail_count) + logging.info("输出路径: %s", selection.input_path) + logging.info("✅ 请在上方输出路径查看字幕结果。") + else: + # 单文件模式：保持原有逻辑 + manifest = upload_single_file(volume, selection, selection.input_path) + payload = build_job_payload(selection, manifest) + result = run_remote_pipeline(volume, selection, manifest, payload) + download_outputs(manifest, result) + summarize(manifest, result) + logging.info("✅ 请在上方输出路径查看字幕结果。") + except KeyboardInterrupt: + logging.warning("用户中断，未执行任何远程操作。") + exit_code = 1 + except Exception as exc: + if isinstance(exc, NoAudioFilesError): + logging.error("%s", exc) + else: + logging.exception("运行失败：%s", exc) + logging.error("日志见：%s", log_path) + exit_code = 1 + + + prompt_exit(not args.non_interactive) + return exit_code + +def _remote_pipeline(job: Dict) -> Dict: + import subprocess + from pathlib import Path + import os + + # 强制重新加载 Volume，确保看到最新上传的文件 + from modal import Volume + volume = Volume.from_name("Faster_Whisper") + volume.reload() + + def run(cmd: Sequence[str], cwd: Optional[str] = None, env: Optional[dict] = None) -> None: + print(" ".join(cmd), flush=True) + subprocess.run(cmd, check=True, cwd=cwd, env=env) + + mount_root = Path(job["mount_root"]) + repo_dir = REPO_VOLUME_DIR + + # log 文件放在 session 目录下，而不是 logs 子目录 + session_dir = Path(job["remote_output_dir"]) + session_dir.mkdir(parents=True, exist_ok=True) + log_file = session_dir / "modal_run.log" + + def log(msg: str) -> None: + line = f"[modal_run] {msg}" + print(line, flush=True) + with log_file.open("a", encoding="utf-8") as fh: + fh.write(line + "\n") + + if not (repo_dir / ".git").exists(): + log("开始克隆仓库...") + run(["git", "clone", REPO_URL, str(repo_dir)]) + else: + log("更新仓库...") + run(["git", "-C", str(repo_dir), "fetch", "origin", "main"]) + run(["git", "-C", str(repo_dir), "reset", "--hard", "origin/main"]) + + model_profile = job["model_profile"] + model_path = repo_dir / "models" + if model_profile.get("hf_repo"): + target_dir = model_profile["target_dir"] + model_path = repo_dir / "models" / target_dir + if not model_path.exists(): + log(f"模型 {model_profile['hf_repo']} 缺失，开始下载...") + cmd = [ + "python", + str(repo_dir / "download_models.py"), + "--hf-model", + model_profile["hf_repo"], + "--target-dir", + target_dir, + ] + env = os.environ.copy() + hf_token = env.get("HF_TOKEN") + if hf_token: + env["HF_TOKEN"] = hf_token + run(cmd, cwd=str(repo_dir), env=env) + else: + log("使用仓库默认模型目录。") + + def snapshot(path: str) -> set: + base = Path(path) + files = set() + if base.exists(): + for f in base.rglob("*"): + if f.is_file(): + files.add(str(f)) + return files + + before = { + target["remote_dir"]: snapshot(target["remote_dir"]) + for target in job["output_targets"] + } + + output_dir = Path(job["remote_output_dir"]) + output_dir.mkdir(parents=True, exist_ok=True) + + cmd = [ + "python", + str(repo_dir / "infer.py"), + "--device", + "cuda", + "--model_name_or_path", + str(model_path), + "--sub_formats", + job["sub_formats"], + "--log_level", + "INFO", + "--output_dir", + str(output_dir), + ] + if job["enable_batching"]: + cmd.append("--enable_batching") + if job["batch_size"]: + cmd.extend(["--batch_size", str(job["batch_size"])]) + cmd.extend(["--max_batch_size", str(job["max_batch_size"])]) + + cmd.extend(job["remote_inputs"]) + + # 在执行推理前，等待文件同步完成 + import time + log("等待文件同步...") + for input_path in job["remote_inputs"]: + input_file = Path(input_path) + session_dir = input_file.parent + max_wait = 180 # 最多等待 3 分钟 + waited = 0 + while not input_file.exists() and waited < max_wait: + time.sleep(1) + waited += 1 + if waited % 10 == 0: + # 打印当前 session 目录下的文件 + if session_dir.exists(): + files = list(session_dir.iterdir()) + file_names = [f.name for f in files] + log(f"等待文件出现: {input_path} ({waited}s)") + log(f" 当前 {session_dir} 下有 {len(files)} 个文件/文件夹: {file_names}") + else: + log(f"等待文件出现: {input_path} ({waited}s)") + log(f" 目录不存在: {session_dir}") + + if input_file.exists(): + log(f"文件已就绪: {input_path}") + else: + log(f"警告: 文件未出现: {input_path}") + + log(f"执行推理命令：{' '.join(cmd)}") + try: + run(cmd, cwd=str(repo_dir)) + except subprocess.CalledProcessError as e: + # 打印调试信息 + sessions_dir = mount_root / "sessions" + log(f"推理命令执行失败，错误码: {e.returncode}") + log(f"=== 调试信息 ===") + + # 统计 sessions 目录下的文件夹数量 + if sessions_dir.exists(): + session_folders = [d for d in sessions_dir.iterdir() if d.is_dir()] + log(f"sessions 目录下共有 {len(session_folders)} 个文件夹") + else: + log(f"sessions 目录不存在: {sessions_dir}") + + # 打印待处理文件所在目录的内容 + for input_path in job["remote_inputs"]: + input_dir = Path(input_path).parent + log(f"待处理文件目录: {input_dir}") + if input_dir.exists(): + log(f"目录内容:") + for item in input_dir.iterdir(): + item_type = "目录" if item.is_dir() else "文件" + log(f" [{item_type}] {item.name}") + else: + log(f"目录不存在: {input_dir}") + + log(f"=== 调试信息结束 ===") + raise + + def to_volume_path(path_str: str) -> str: + return container_to_volume_path(path_str) + + # 收集生成的文件内容（直接返回，避免 volume 同步问题） + import base64 + created_files = {} # {filename: base64_content} + for target in job["output_targets"]: + remote_dir = target["remote_dir"] + after = snapshot(remote_dir) + prev = before.get(remote_dir, set()) + new_files = sorted( + file + for file in after - prev + if Path(file).suffix.lower() in SUB_SUFFIXES + ) + for file_path in new_files: + file_path = Path(file_path) + if file_path.exists(): + content = file_path.read_bytes() + # 文件名使用相对于 remote_dir 的路径 + rel_name = file_path.name + created_files[rel_name] = base64.b64encode(content).decode() + log(f"读取生成文件: {file_path} ({len(content)} bytes)") + + # 读取 log 文件内容 + log_content = None + if log_file.exists(): + log_content = base64.b64encode(log_file.read_bytes()).decode() + + return { + "created_files": created_files, # {filename: base64_content} + "log_content": log_content, + } + + +if __name__ == "__main__": # pragma: no cover + sys.exit(main()) diff --git a/project.spec b/project.spec index b8078db..efcb59a 100644 --- a/project.spec +++ b/project.spec @@ -346,7 +346,7 @@ datas += [ ] a = Analysis( - ['infer.py'], + ['infer.py', 'modal_infer.py'], pathex=[], binaries=binaries, datas=datas, @@ -374,9 +374,9 @@ a = Analysis( pyz = PYZ(a.pure, a.zipped_data, cipher=block_cipher) -exe = EXE( +infer_exe = EXE( pyz, - a.scripts, + [a.scripts[0]], [], exclude_binaries=True, name='infer', @@ -393,8 +393,28 @@ exe = EXE( icon='transwithai.ico' if os.path.exists('transwithai.ico') else None, ) +modal_exe = EXE( + pyz, + [a.scripts[1]], + [], + exclude_binaries=True, + name='modal_infer', + debug=False, + bootloader_ignore_signals=False, + strip=False, + upx=False, + console=True, + disable_windowed_traceback=False, + argv_emulation=False, + target_arch=None, + codesign_identity=None, + entitlements_file=None, + icon='transwithai.ico' if os.path.exists('transwithai.ico') else None, +) + coll = COLLECT( - exe, + infer_exe, + modal_exe, a.binaries, a.zipfiles, a.datas, @@ -402,4 +422,4 @@ coll = COLLECT( upx=False, upx_exclude=[], name='faster_whisper_transwithai_chickenrice', -) \ No newline at end of file +) diff --git "a/\344\275\277\347\224\250\350\257\264\346\230\216.txt" "b/\344\275\277\347\224\250\350\257\264\346\230\216.txt" index 7afd611..a55207f 100644 --- "a/\344\275\277\347\224\250\350\257\264\346\230\216.txt" +++ "b/\344\275\277\347\224\250\350\257\264\346\230\216.txt" @@ -94,6 +94,83 @@ GPU模式（仅限NVIDIA显卡）: --- +=== Modal 云端推理模式 === + +如果本地没有GPU或显存不足，可以使用Modal云端GPU进行推理。 + +1. 环境配置： + +使用 Conda 创建轻量级环境（仅需 modal 和 questionary 库）： + +```bash +conda env create -f environment-modal.yml +conda activate faster-whisper-modal +``` + +或手动安装： +```bash +pip install modal questionary +``` + +2. Modal 账号设置： + +(1) 注册 Modal 账号： + 访问 https://modal.com/ 注册账号 + 新用户每月有 $30 免费额度，足够处理大量音频 + +(2) 配置 Modal Token： + 在终端运行： + ```bash + modal token new + ``` + 按提示完成浏览器授权 + +(3) 配置 HuggingFace Token（可选但推荐）： + + 为什么需要配置： + - Modal 需要从 HuggingFace 下载模型文件 + - 某些模型需要登录才能下载 + - 配置后可避免下载失败或速度慢的问题 + + 如何配置： + a. 获取 HF Token： + 访问 https://huggingface.co/settings/tokens + 创建一个 Read 权限的 token + + b. 在 Modal 中添加 Secret： + 访问 https://modal.com/secrets + 点击 "Create new secret" + 选择 "Hugging Face" 类型 + 输入你的 HF Token + 保存为 "huggingface-secret" + +3. 运行 Modal 推理： + +```bash +python modal_infer.py +``` + +程序会交互式询问： +- GPU 类型：T4（推荐）、A10G、A100、H100 等 +- 模型选择：基础版、海南鸡（日文转中文优化）、自定义模型 +- 输入文件：本地音频文件或文件夹路径 +- 批处理选项：是否启用批处理加速（需要更高显存） +- 超时时间：任务超时时间（分钟） + +处理流程： +1. 上传本地音频文件到 Modal Volume +2. 在云端 GPU 上运行推理 +3. 自动下载生成的字幕文件到本地 +4. 输出结果保存在源文件同目录 + +注意事项： +- 首次运行需要构建 Docker 镜像，可能需要几分钟 +- 模型首次下载会缓存在 Modal Volume，后续运行更快 +- 推荐使用 T4 GPU，性价比最高 +- 批处理模式可显著加速，但需要更多显存 + +--- + 相关项目： - Faster Whisper: https://github.com/SYSTRAN/faster-whisper