feat(logger): 重构日志系统配置

- 添加RotatingFileHandler支持日志轮转
- 配置多个日志处理器分别输出到控制台、全部日志文件和错误日志文件
- 设置不同日志级别的输出过滤
- 确保日志文件目录存在并正确初始化日志系统
- 移除原有的基础日志配置方式

.env.example

@@ -32,11 +32,17 @@ TEMP_DIR=tmp/
 #UPLOAD_TIMEOUT=600                   # 上传超时（秒）
 
 # ===================
-# 硬件加速
+# 硬件加速与多显卡
 # ===================
-# 可选值: none, qsv, cuda
+# 硬件加速类型: none, qsv, cuda
 HW_ACCEL=none
 
+# GPU 设备列表（逗号分隔的设备索引）
+# 不配置时：自动检测所有设备
+# 单设备示例：GPU_DEVICES=0
+# 多设备示例：GPU_DEVICES=0,1,2
+#GPU_DEVICES=0,1
+
 # ===================
 # 素材缓存
 # ===================

domain/config.py

@@ -5,12 +5,15 @@ Worker 配置模型
 定义 Worker 运行时的配置参数。
 """
 
+import logging
 import os
 from dataclasses import dataclass, field
 from typing import List, Optional
 
 from constant import HW_ACCEL_NONE, HW_ACCEL_QSV, HW_ACCEL_CUDA, HW_ACCEL_TYPES
 
+logger = logging.getLogger(__name__)
+
 
 # 默认支持的任务类型
 DEFAULT_CAPABILITIES = [
@@ -59,6 +62,9 @@ class WorkerConfig:
     # 硬件加速配置
     hw_accel: str = HW_ACCEL_NONE  # 硬件加速类型: none, qsv, cuda
 
+    # GPU 设备配置（多显卡调度）
+    gpu_devices: List[int] = field(default_factory=list)  # 空列表表示使用默认设备
+
     # 素材缓存配置
     cache_enabled: bool = True  # 是否启用素材缓存
     cache_dir: str = ""  # 缓存目录，默认为 temp_dir/cache
@@ -113,6 +119,16 @@ class WorkerConfig:
         if hw_accel not in HW_ACCEL_TYPES:
             hw_accel = HW_ACCEL_NONE
 
+        # GPU 设备列表（用于多显卡调度）
+        gpu_devices_str = os.getenv('GPU_DEVICES', '')
+        gpu_devices: List[int] = []
+        if gpu_devices_str:
+            try:
+                gpu_devices = [int(d.strip()) for d in gpu_devices_str.split(',') if d.strip()]
+            except ValueError:
+                logger.warning(f"Invalid GPU_DEVICES value: {gpu_devices_str}, using auto-detect")
+                gpu_devices = []
+
         # 素材缓存配置
         cache_enabled = os.getenv('CACHE_ENABLED', 'true').lower() in ('true', '1', 'yes')
         cache_dir = os.getenv('CACHE_DIR', '')  # 空字符串表示使用默认路径
@@ -132,6 +148,7 @@ class WorkerConfig:
             download_timeout=download_timeout,
             upload_timeout=upload_timeout,
             hw_accel=hw_accel,
+            gpu_devices=gpu_devices,
             cache_enabled=cache_enabled,
             cache_dir=cache_dir if cache_dir else os.path.join(temp_dir, 'cache'),
             cache_max_size_gb=cache_max_size_gb
@@ -156,3 +173,11 @@ class WorkerConfig:
     def is_cuda(self) -> bool:
         """是否使用 CUDA 硬件加速"""
         return self.hw_accel == HW_ACCEL_CUDA
+
+    def has_multi_gpu(self) -> bool:
+        """是否配置了多 GPU"""
+        return len(self.gpu_devices) > 1
+
+    def get_gpu_devices(self) -> List[int]:
+        """获取 GPU 设备列表"""
+        return self.gpu_devices.copy()

domain/gpu.py

@@ -0,0 +1,31 @@
+# -*- coding: utf-8 -*-
+"""
+GPU 设备模型
+
+定义 GPU 设备的数据结构。
+"""
+
+from dataclasses import dataclass
+from typing import Optional
+
+
+@dataclass
+class GPUDevice:
+    """
+    GPU 设备信息
+
+    Attributes:
+        index: 设备索引（对应 nvidia-smi 中的 GPU ID）
+        name: 设备名称（如 "NVIDIA GeForce RTX 3090"）
+        memory_total: 显存总量（MB），可选
+        available: 设备是否可用
+    """
+    index: int
+    name: str
+    memory_total: Optional[int] = None
+    available: bool = True
+
+    def __str__(self) -> str:
+        status = "available" if self.available else "unavailable"
+        mem_info = f", {self.memory_total}MB" if self.memory_total else ""
+        return f"GPU[{self.index}]: {self.name}{mem_info} ({status})"

handlers/base.py

@@ -11,6 +11,7 @@ import logging
 import shutil
 import tempfile
 import subprocess
+import threading
 from abc import ABC
 from typing import Optional, List, Dict, Any, Tuple, TYPE_CHECKING
 
@@ -75,23 +76,33 @@ def get_video_encode_args(hw_accel: str = HW_ACCEL_NONE) -> List[str]:
         ]
 
 
-def get_hwaccel_decode_args(hw_accel: str = HW_ACCEL_NONE) -> List[str]:
+def get_hwaccel_decode_args(hw_accel: str = HW_ACCEL_NONE, device_index: Optional[int] = None) -> List[str]:
     """
     获取硬件加速解码参数（输入文件之前使用）
 
     Args:
         hw_accel: 硬件加速类型 (none, qsv, cuda)
+        device_index: GPU 设备索引，用于多显卡调度
 
     Returns:
         FFmpeg 硬件加速解码参数列表
     """
     if hw_accel == HW_ACCEL_CUDA:
         # CUDA 硬件加速解码
-        # 注意：使用 cuda 作为 hwaccel，但输出到系统内存以便 CPU 滤镜处理
-        return ['-hwaccel', 'cuda', '-hwaccel_output_format', 'cuda']
+        args = ['-hwaccel', 'cuda']
+        # 多显卡模式下指定设备
+        if device_index is not None:
+            args.extend(['-hwaccel_device', str(device_index)])
+        args.extend(['-hwaccel_output_format', 'cuda'])
+        return args
     elif hw_accel == HW_ACCEL_QSV:
         # QSV 硬件加速解码
-        return ['-hwaccel', 'qsv', '-hwaccel_output_format', 'qsv']
+        args = ['-hwaccel', 'qsv']
+        # QSV 在 Windows 上使用 -qsv_device
+        if device_index is not None:
+            args.extend(['-qsv_device', str(device_index)])
+        args.extend(['-hwaccel_output_format', 'qsv'])
+        return args
     else:
         return []
 
@@ -103,6 +114,10 @@ def get_hwaccel_filter_prefix(hw_accel: str = HW_ACCEL_NONE) -> str:
     注意：由于大多数复杂滤镜（如 lut3d, overlay, crop 等）不支持硬件表面，
     我们需要在滤镜链开始时将硬件表面下载到系统内存。
 
+    CUDA/QSV hwdownload 只支持 nv12 格式输出，因此需要两步转换：
+    1. hwdownload,format=nv12 - 从 GPU 下载到 CPU
+    2. format=yuv420p - 转换为标准格式（确保与 RGBA/YUVA overlay 混合时颜色正确）
+
     Args:
         hw_accel: 硬件加速类型
 
@@ -110,9 +125,9 @@ def get_hwaccel_filter_prefix(hw_accel: str = HW_ACCEL_NONE) -> str:
         需要添加到滤镜链开头的 hwdownload 滤镜字符串
     """
     if hw_accel == HW_ACCEL_CUDA:
-        return 'hwdownload,format=nv12,'
+        return 'hwdownload,format=nv12,format=yuv420p,'
     elif hw_accel == HW_ACCEL_QSV:
-        return 'hwdownload,format=nv12,'
+        return 'hwdownload,format=nv12,format=yuv420p,'
     else:
         return ''
 
@@ -128,6 +143,8 @@ AUDIO_ENCODE_ARGS = [
     '-ac', '2',
 ]
 
+FFMPEG_LOGLEVEL = 'error'
+
 
 def subprocess_args(include_stdout: bool = True) -> Dict[str, Any]:
     """
@@ -248,9 +265,13 @@ class BaseHandler(TaskHandler, ABC):
     - 临时目录管理
     - 文件下载/上传
     - FFmpeg 命令执行
+    - GPU 设备管理（多显卡调度）
     - 日志记录
     """
 
+    # 线程本地存储：用于存储当前线程的 GPU 设备索引
+    _thread_local = threading.local()
+
     def __init__(self, config: WorkerConfig, api_client: 'APIClientV2'):
         """
         初始化处理器
@@ -267,6 +288,39 @@ class BaseHandler(TaskHandler, ABC):
             max_size_gb=config.cache_max_size_gb
         )
 
+    # ========== GPU 设备管理 ==========
+
+    def set_gpu_device(self, device_index: int) -> None:
+        """
+        设置当前线程的 GPU 设备索引
+
+        由 TaskExecutor 在任务执行前调用。
+
+        Args:
+            device_index: GPU 设备索引
+        """
+        self._thread_local.gpu_device = device_index
+
+    def get_gpu_device(self) -> Optional[int]:
+        """
+        获取当前线程的 GPU 设备索引
+
+        Returns:
+            GPU 设备索引，未设置则返回 None
+        """
+        return getattr(self._thread_local, 'gpu_device', None)
+
+    def clear_gpu_device(self) -> None:
+        """
+        清除当前线程的 GPU 设备索引
+
+        由 TaskExecutor 在任务执行后调用。
+        """
+        if hasattr(self._thread_local, 'gpu_device'):
+            del self._thread_local.gpu_device
+
+    # ========== FFmpeg 参数生成 ==========
+
     def get_video_encode_args(self) -> List[str]:
         """
         获取当前配置的视频编码参数
@@ -278,12 +332,13 @@ class BaseHandler(TaskHandler, ABC):
 
     def get_hwaccel_decode_args(self) -> List[str]:
         """
-        获取硬件加速解码参数（在输入文件之前使用）
+        获取硬件加速解码参数（支持设备指定）
 
         Returns:
             FFmpeg 硬件加速解码参数列表
         """
-        return get_hwaccel_decode_args(self.config.hw_accel)
+        device_index = self.get_gpu_device()
+        return get_hwaccel_decode_args(self.config.hw_accel, device_index)
 
     def get_hwaccel_filter_prefix(self) -> str:
         """
@@ -409,6 +464,11 @@ class BaseHandler(TaskHandler, ABC):
             if result:
                 file_size = os.path.getsize(file_path)
                 logger.info(f"[task:{task_id}] Uploaded: {file_path} ({file_size} bytes)")
+
+                # 将上传成功的文件加入缓存
+                if access_url:
+                    self.material_cache.add_to_cache(access_url, file_path)
+
                 return access_url
             else:
                 logger.error(f"[task:{task_id}] Upload failed: {file_path}")
@@ -437,22 +497,28 @@ class BaseHandler(TaskHandler, ABC):
         if timeout is None:
             timeout = self.config.ffmpeg_timeout
 
+        cmd_to_run = list(cmd)
+        if cmd_to_run and cmd_to_run[0] == 'ffmpeg' and '-loglevel' not in cmd_to_run:
+            cmd_to_run[1:1] = ['-loglevel', FFMPEG_LOGLEVEL]
+
         # 日志记录命令（限制长度）
-        cmd_str = ' '.join(cmd)
+        cmd_str = ' '.join(cmd_to_run)
         if len(cmd_str) > 500:
             cmd_str = cmd_str[:500] + '...'
         logger.info(f"[task:{task_id}] FFmpeg: {cmd_str}")
 
         try:
+            run_args = subprocess_args(False)
+            run_args['stdout'] = subprocess.DEVNULL
+            run_args['stderr'] = subprocess.PIPE
             result = subprocess.run(
-                cmd,
-                capture_output=True,
+                cmd_to_run,
                 timeout=timeout,
-                **subprocess_args(False)
+                **run_args
             )
 
             if result.returncode != 0:
-                stderr = result.stderr.decode('utf-8', errors='replace')[:1000]
+                stderr = (result.stderr or b'').decode('utf-8', errors='replace')[:1000]
                 logger.error(f"[task:{task_id}] FFmpeg failed (code={result.returncode}): {stderr}")
                 return False
 

handlers/finalize_mp4.py

@@ -127,9 +127,9 @@ class FinalizeMp4Handler(BaseHandler):
         concat_file = os.path.join(work_dir, 'concat.txt')
         with open(concat_file, 'w', encoding='utf-8') as f:
             for ts_file in ts_files:
-                # 路径中的反斜杠需要转义或使用正斜杠
-                ts_path = ts_file.replace('\\', '/')
-                f.write(f"file '{ts_path}'\n")
+                # FFmpeg concat 路径相对于 concat.txt 所在目录，只需写文件名
+                ts_filename = os.path.basename(ts_file)
+                f.write(f"file '{ts_filename}'\n")
 
         # 3. 构建合并命令（remux，不重编码）
         cmd = [

handlers/render_video.py

@@ -123,21 +123,48 @@ class RenderSegmentVideoHandler(BaseHandler):
             overlay_file = None
             if render_spec.overlay_url:
                 # 根据 URL 后缀确定文件扩展名
-                ext = '.png'
-                if render_spec.overlay_url.lower().endswith('.jpg') or render_spec.overlay_url.lower().endswith('.jpeg'):
+                url_lower = render_spec.overlay_url.lower()
+                if url_lower.endswith('.jpg') or url_lower.endswith('.jpeg'):
                     ext = '.jpg'
+                elif url_lower.endswith('.mov'):
+                    ext = '.mov'
+                else:
+                    ext = '.png'  # 默认
                 overlay_file = os.path.join(work_dir, f'overlay{ext}')
                 if not self.download_file(render_spec.overlay_url, overlay_file):
                     logger.warning(f"[task:{task.task_id}] Failed to download overlay, continuing without it")
                     overlay_file = None
 
-            # 6. 计算 overlap 时长（用于转场帧冻结）
+            # 6. 探测源视频时长（仅对视频素材）
+            # 用于检测时长不足并通过冻结最后一帧补足
+            source_duration_sec = None
+            if not is_image:
+                source_duration = self.probe_duration(input_file)
+                if source_duration:
+                    source_duration_sec = source_duration
+                    speed = float(render_spec.speed) if render_spec.speed else 1.0
+                    if speed > 0:
+                        # 计算变速后的有效时长
+                        effective_duration_sec = source_duration_sec / speed
+                        required_duration_sec = duration_ms / 1000.0
+
+                        # 如果源视频时长不足，记录日志
+                        if effective_duration_sec < required_duration_sec:
+                            shortage_sec = required_duration_sec - effective_duration_sec
+                            logger.warning(
+                                f"[task:{task.task_id}] Source video duration insufficient: "
+                                f"effective={effective_duration_sec:.2f}s (speed={speed}), "
+                                f"required={required_duration_sec:.2f}s, "
+                                f"will freeze last frame for {shortage_sec:.2f}s"
+                            )
+
+            # 7. 计算 overlap 时长（用于转场帧冻结）
             # 头部 overlap: 来自前一片段的出场转场
             overlap_head_ms = task.get_overlap_head_ms()
             # 尾部 overlap: 当前片段的出场转场
             overlap_tail_ms = task.get_overlap_tail_ms_v2()
 
-            # 7. 构建 FFmpeg 命令
+            # 8. 构建 FFmpeg 命令
             output_file = os.path.join(work_dir, 'output.mp4')
             cmd = self._build_command(
                 input_file=input_file,
@@ -148,28 +175,29 @@ class RenderSegmentVideoHandler(BaseHandler):
                 lut_file=lut_file,
                 overlay_file=overlay_file,
                 overlap_head_ms=overlap_head_ms,
-                overlap_tail_ms=overlap_tail_ms
+                overlap_tail_ms=overlap_tail_ms,
+                source_duration_sec=source_duration_sec
             )
 
-            # 8. 执行 FFmpeg
+            # 9. 执行 FFmpeg
             if not self.run_ffmpeg(cmd, task.task_id):
                 return TaskResult.fail(
                     ErrorCode.E_FFMPEG_FAILED,
                     "FFmpeg rendering failed"
                 )
 
-            # 9. 验证输出文件
+            # 10. 验证输出文件
             if not self.ensure_file_exists(output_file, min_size=4096):
                 return TaskResult.fail(
                     ErrorCode.E_FFMPEG_FAILED,
                     "Output file is missing or too small"
                 )
 
-            # 10. 获取实际时长
+            # 11. 获取实际时长
             actual_duration = self.probe_duration(output_file)
             actual_duration_ms = int(actual_duration * 1000) if actual_duration else duration_ms
 
-            # 11. 上传产物
+            # 12. 上传产物
             video_url = self.upload_file(task.task_id, 'video', output_file)
             if not video_url:
                 return TaskResult.fail(
@@ -177,7 +205,7 @@ class RenderSegmentVideoHandler(BaseHandler):
                     "Failed to upload video"
                 )
 
-            # 12. 构建结果（包含 overlap 信息）
+            # 13. 构建结果（包含 overlap 信息）
             result_data = {
                 'videoUrl': video_url,
                 'actualDurationMs': actual_duration_ms,
@@ -271,12 +299,24 @@ class RenderSegmentVideoHandler(BaseHandler):
 
         cmd.extend(['-vf', ','.join(filters)])
 
+        # 计算总帧数，动态调整 GOP
+        total_frames = int(actual_duration_sec * fps)
+        if total_frames <= 1:
+            gop_size = 1
+        elif total_frames < fps:
+            gop_size = total_frames
+        else:
+            gop_size = fps * 2  # 正常情况，2 秒一个关键帧
+
         # 编码参数
         cmd.extend([
             '-c:v', 'libx264',
             '-preset', 'fast',
             '-crf', '18',
             '-r', str(fps),
+            '-g', str(gop_size),
+            '-keyint_min', str(min(gop_size, fps // 2 or 1)),
+            '-force_key_frames', 'expr:eq(n,0)',
             '-an',  # 无音频
             output_file
         ])
@@ -294,7 +334,8 @@ class RenderSegmentVideoHandler(BaseHandler):
         lut_file: Optional[str] = None,
         overlay_file: Optional[str] = None,
         overlap_head_ms: int = 0,
-        overlap_tail_ms: int = 0
+        overlap_tail_ms: int = 0,
+        source_duration_sec: Optional[float] = None
     ) -> List[str]:
         """
         构建 FFmpeg 渲染命令
@@ -309,6 +350,7 @@ class RenderSegmentVideoHandler(BaseHandler):
             overlay_file: 叠加层文件路径（可选）
             overlap_head_ms: 头部 overlap 时长（毫秒）
             overlap_tail_ms: 尾部 overlap 时长（毫秒）
+            source_duration_sec: 源视频实际时长（秒），用于检测时长不足
 
         Returns:
             FFmpeg 命令参数列表
@@ -331,10 +373,12 @@ class RenderSegmentVideoHandler(BaseHandler):
         filters = self._build_video_filters(
             render_spec=render_spec,
             output_spec=output_spec,
+            duration_ms=duration_ms,
             lut_file=lut_file,
-            has_overlay=overlay_file is not None,
+            overlay_file=overlay_file,
             overlap_head_ms=overlap_head_ms,
-            overlap_tail_ms=overlap_tail_ms
+            overlap_tail_ms=overlap_tail_ms,
+            source_duration_sec=source_duration_sec
         )
 
         # 应用滤镜
@@ -353,16 +397,28 @@ class RenderSegmentVideoHandler(BaseHandler):
         fps = output_spec.fps
         cmd.extend(['-r', str(fps)])
 
-        # GOP 大小（关键帧间隔）
-        gop_size = fps * 2  # 2秒一个关键帧
-        cmd.extend(['-g', str(gop_size)])
-        cmd.extend(['-keyint_min', str(gop_size)])
-
         # 时长（包含 overlap 区域）
         total_duration_ms = duration_ms + overlap_head_ms + overlap_tail_ms
         duration_sec = total_duration_ms / 1000.0
         cmd.extend(['-t', str(duration_sec)])
 
+        # 动态调整 GOP 大小：对于短视频，GOP 不能大于总帧数
+        total_frames = int(duration_sec * fps)
+        if total_frames <= 1:
+            gop_size = 1
+        elif total_frames < fps:
+            # 短于 1 秒的视频，使用全部帧数作为 GOP（整个视频只有开头一个关键帧）
+            gop_size = total_frames
+        else:
+            # 正常情况，2 秒一个关键帧
+            gop_size = fps * 2
+
+        cmd.extend(['-g', str(gop_size)])
+        cmd.extend(['-keyint_min', str(min(gop_size, fps // 2 or 1))])
+
+        # 强制第一帧为关键帧
+        cmd.extend(['-force_key_frames', 'expr:eq(n,0)'])
+
         # 无音频（视频片段不包含音频）
         cmd.append('-an')
 
@@ -375,10 +431,12 @@ class RenderSegmentVideoHandler(BaseHandler):
         self,
         render_spec: RenderSpec,
         output_spec: OutputSpec,
+        duration_ms: int,
         lut_file: Optional[str] = None,
-        has_overlay: bool = False,
+        overlay_file: Optional[str] = None,
         overlap_head_ms: int = 0,
-        overlap_tail_ms: int = 0
+        overlap_tail_ms: int = 0,
+        source_duration_sec: Optional[float] = None
     ) -> str:
         """
         构建视频滤镜链
@@ -386,10 +444,12 @@ class RenderSegmentVideoHandler(BaseHandler):
         Args:
             render_spec: 渲染规格
             output_spec: 输出规格
+            duration_ms: 目标时长（毫秒）
             lut_file: LUT 文件路径
-            has_overlay: 是否有叠加层
+            overlay_file: 叠加层文件路径（支持图片 png/jpg 和视频 mov）
             overlap_head_ms: 头部 overlap 时长（毫秒）
             overlap_tail_ms: 尾部 overlap 时长（毫秒）
+            source_duration_sec: 源视频实际时长（秒），用于检测时长不足
 
         Returns:
             滤镜字符串
@@ -399,6 +459,10 @@ class RenderSegmentVideoHandler(BaseHandler):
         height = output_spec.height
         fps = output_spec.fps
 
+        # 判断 overlay 类型
+        has_overlay = overlay_file is not None
+        is_video_overlay = has_overlay and overlay_file.lower().endswith('.mov')
+
         # 解析 effects
         effects = render_spec.get_effects()
         has_camera_shot = any(e.effect_type == 'cameraShot' for e in effects)
@@ -458,23 +522,45 @@ class RenderSegmentVideoHandler(BaseHandler):
                 base_filters=filters,
                 effects=effects,
                 fps=fps,
+                width=width,
+                height=height,
                 has_overlay=has_overlay,
+                is_video_overlay=is_video_overlay,
                 overlap_head_ms=overlap_head_ms,
                 overlap_tail_ms=overlap_tail_ms,
-                use_hwdownload=bool(hwaccel_prefix)
+                use_hwdownload=bool(hwaccel_prefix),
+                duration_ms=duration_ms,
+                render_spec=render_spec,
+                source_duration_sec=source_duration_sec
             )
 
-        # 6. 帧冻结（tpad）- 用于转场 overlap 区域
+        # 6. 帧冻结（tpad）- 用于转场 overlap 区域和时长不足补足
         # 注意：tpad 必须在缩放之后应用
         tpad_parts = []
+
+        # 计算是否需要额外的尾部冻结（源视频时长不足）
+        extra_tail_freeze_sec = 0.0
+        if source_duration_sec is not None:
+            speed = float(render_spec.speed) if render_spec.speed else 1.0
+            if speed > 0:
+                # 计算变速后的有效时长
+                effective_duration_sec = source_duration_sec / speed
+                required_duration_sec = duration_ms / 1000.0
+
+                # 如果源视频时长不足，需要冻结最后一帧来补足
+                if effective_duration_sec < required_duration_sec:
+                    extra_tail_freeze_sec = required_duration_sec - effective_duration_sec
+
         if overlap_head_ms > 0:
             # 头部冻结：将第一帧冻结指定时长
             head_duration_sec = overlap_head_ms / 1000.0
             tpad_parts.append(f"start_mode=clone:start_duration={head_duration_sec}")
-        if overlap_tail_ms > 0:
-            # 尾部冻结：将最后一帧冻结指定时长
-            tail_duration_sec = overlap_tail_ms / 1000.0
-            tpad_parts.append(f"stop_mode=clone:stop_duration={tail_duration_sec}")
+
+        # 尾部冻结：合并 overlap 和时长不足的冻结
+        total_tail_freeze_sec = (overlap_tail_ms / 1000.0) + extra_tail_freeze_sec
+        if total_tail_freeze_sec > 0:
+            # 将最后一帧冻结指定时长
+            tpad_parts.append(f"stop_mode=clone:stop_duration={total_tail_freeze_sec}")
 
         if tpad_parts:
             filters.append(f"tpad={':'.join(tpad_parts)}")
@@ -483,7 +569,13 @@ class RenderSegmentVideoHandler(BaseHandler):
         if has_overlay:
             # 使用 filter_complex 格式
             base_filters = ','.join(filters) if filters else 'copy'
-            return f"[0:v]{base_filters}[base];[base][1:v]overlay=0:0"
+            overlay_scale = f"scale={width}:{height}"
+            # 视频 overlay 使用 eof_action=pass（结束后消失），图片 overlay 使用默认行为（保持显示）
+            overlay_params = 'eof_action=pass' if is_video_overlay else ''
+            overlay_filter = f"overlay=0:0:{overlay_params}" if overlay_params else 'overlay=0:0'
+            # 视频 overlay 需要在末尾统一颜色范围，避免 overlay 结束后 range 从 tv 变为 pc
+            range_fix = ',format=yuv420p,setrange=tv' if is_video_overlay else ''
+            return f"[0:v]{base_filters}[base];[1:v]{overlay_scale}[overlay];[base][overlay]{overlay_filter}{range_fix}"
         else:
             return ','.join(filters) if filters else ''
 
@@ -492,10 +584,16 @@ class RenderSegmentVideoHandler(BaseHandler):
         base_filters: List[str],
         effects: List[Effect],
         fps: int,
+        width: int,
+        height: int,
         has_overlay: bool = False,
+        is_video_overlay: bool = False,
         overlap_head_ms: int = 0,
         overlap_tail_ms: int = 0,
-        use_hwdownload: bool = False
+        use_hwdownload: bool = False,
+        duration_ms: int = 0,
+        render_spec: Optional[RenderSpec] = None,
+        source_duration_sec: Optional[float] = None
     ) -> str:
         """
         构建包含特效的 filter_complex 滤镜图
@@ -506,10 +604,16 @@ class RenderSegmentVideoHandler(BaseHandler):
             base_filters: 基础滤镜列表
             effects: 特效列表
             fps: 帧率
+            width: 输出宽度
+            height: 输出高度
             has_overlay: 是否有叠加层
+            is_video_overlay: 叠加层是否为视频格式（如 .mov）
             overlap_head_ms: 头部 overlap 时长
             overlap_tail_ms: 尾部 overlap 时长
             use_hwdownload: 是否使用了硬件加速解码（已在 base_filters 中包含 hwdownload）
+            duration_ms: 目标时长（毫秒）
+            render_spec: 渲染规格（用于获取变速参数）
+            source_duration_sec: 源视频实际时长（秒），用于检测时长不足
 
         Returns:
             filter_complex 格式的滤镜字符串
@@ -531,15 +635,16 @@ class RenderSegmentVideoHandler(BaseHandler):
                 if start_sec <= 0 or duration_sec <= 0:
                     continue
 
-                # cameraShot 实现：
+                # cameraShot 实现（定格效果）：
                 # 1. fps + split 分割
-                # 2. 第一路：trim(0, start+duration) + freezeframes
+                # 2. 第一路：trim(0, start) + tpad冻结duration秒
                 # 3. 第二路：trim(start, end)
                 # 4. concat 拼接
 
-                start_frame = start_sec * fps
                 split_out_a = f'[eff{effect_idx}_a]'
                 split_out_b = f'[eff{effect_idx}_b]'
+                frozen_out = f'[eff{effect_idx}_frozen]'
+                rest_out = f'[eff{effect_idx}_rest]'
                 effect_output = f'[v_eff{effect_idx}]'
 
                 # fps + split
@@ -547,37 +652,50 @@ class RenderSegmentVideoHandler(BaseHandler):
                     f"{current_output}fps=fps={fps},split{split_out_a}{split_out_b}"
                 )
 
-                # 第一路：trim + freezeframes（在 start 帧处冻结 duration 秒）
-                # freezeframes: 从 first 帧开始，用 replace 帧替换后续帧
-                # 这样实现定格效果：在 start_frame 位置冻结
+                # 第一路：trim(0, start) + tpad冻结
+                # tpad=stop_mode=clone 将最后一帧冻结指定时长
                 filter_parts.append(
-                    f"{split_out_a}trim=start=0:end={start_sec + duration_sec},"
-                    f"setpts=PTS-STARTPTS,"
-                    f"freezeframes=first={start_frame}:last={start_frame + duration_sec * fps - 1}:replace={start_frame}"
-                    f"{split_out_a}"
+                    f"{split_out_a}trim=start=0:end={start_sec},setpts=PTS-STARTPTS,"
+                    f"tpad=stop_mode=clone:stop_duration={duration_sec}{frozen_out}"
                 )
 
                 # 第二路：trim 从 start 开始
                 filter_parts.append(
-                    f"{split_out_b}trim=start={start_sec},setpts=PTS-STARTPTS{split_out_b}"
+                    f"{split_out_b}trim=start={start_sec},setpts=PTS-STARTPTS{rest_out}"
                 )
 
                 # concat 拼接
                 filter_parts.append(
-                    f"{split_out_a}{split_out_b}concat=n=2:v=1:a=0{effect_output}"
+                    f"{frozen_out}{rest_out}concat=n=2:v=1:a=0{effect_output}"
                 )
 
                 current_output = effect_output
                 effect_idx += 1
 
-        # 帧冻结（tpad）- 用于转场 overlap 区域
+        # 帧冻结（tpad）- 用于转场 overlap 区域和时长不足补足
         tpad_parts = []
+
+        # 计算是否需要额外的尾部冻结（源视频时长不足）
+        extra_tail_freeze_sec = 0.0
+        if source_duration_sec is not None and render_spec is not None and duration_ms > 0:
+            speed = float(render_spec.speed) if render_spec.speed else 1.0
+            if speed > 0:
+                # 计算变速后的有效时长
+                effective_duration_sec = source_duration_sec / speed
+                required_duration_sec = duration_ms / 1000.0
+
+                # 如果源视频时长不足，需要冻结最后一帧来补足
+                if effective_duration_sec < required_duration_sec:
+                    extra_tail_freeze_sec = required_duration_sec - effective_duration_sec
+
         if overlap_head_ms > 0:
             head_duration_sec = overlap_head_ms / 1000.0
             tpad_parts.append(f"start_mode=clone:start_duration={head_duration_sec}")
-        if overlap_tail_ms > 0:
-            tail_duration_sec = overlap_tail_ms / 1000.0
-            tpad_parts.append(f"stop_mode=clone:stop_duration={tail_duration_sec}")
+
+        # 尾部冻结：合并 overlap 和时长不足的冻结
+        total_tail_freeze_sec = (overlap_tail_ms / 1000.0) + extra_tail_freeze_sec
+        if total_tail_freeze_sec > 0:
+            tpad_parts.append(f"stop_mode=clone:stop_duration={total_tail_freeze_sec}")
 
         if tpad_parts:
             tpad_output = '[v_tpad]'
@@ -587,7 +705,15 @@ class RenderSegmentVideoHandler(BaseHandler):
         # 最终输出
         if has_overlay:
             # 叠加层处理
-            filter_parts.append(f"{current_output}[1:v]overlay=0:0")
+            # 视频 overlay 使用 eof_action=pass（结束后消失），图片 overlay 使用默认行为（保持显示）
+            overlay_params = 'eof_action=pass' if is_video_overlay else ''
+            overlay_filter = f"overlay=0:0:{overlay_params}" if overlay_params else 'overlay=0:0'
+            overlay_scale = f"scale={width}:{height}"
+            overlay_output = '[v_overlay]'
+            # 视频 overlay 需要在末尾统一颜色范围，避免 overlay 结束后 range 从 tv 变为 pc
+            range_fix = ',format=yuv420p,setrange=tv' if is_video_overlay else ''
+            filter_parts.append(f"[1:v]{overlay_scale}{overlay_output}")
+            filter_parts.append(f"{current_output}{overlay_output}{overlay_filter}{range_fix}")
         else:
             # 移除最后一个标签，直接输出
             # 将最后一个滤镜的输出标签替换为空（直接输出）

index.py

@@ -25,6 +25,8 @@ import sys
 import time
 import signal
 import logging
+import os
+from logging.handlers import RotatingFileHandler
 
 from dotenv import load_dotenv
 
@@ -34,11 +36,55 @@ from services.task_executor import TaskExecutor
 from constant import SOFTWARE_VERSION
 
 # 日志配置
-logging.basicConfig(
-    level=logging.INFO,
-    format='[%(asctime)s] [%(levelname)s] [%(name)s] %(message)s',
-    datefmt='%Y-%m-%d %H:%M:%S'
+def setup_logging():
+    """配置日志系统，输出到控制台和文件"""
+    # 日志格式
+    log_format = '[%(asctime)s] [%(levelname)s] [%(name)s] %(message)s'
+    date_format = '%Y-%m-%d %H:%M:%S'
+    formatter = logging.Formatter(log_format, date_format)
+
+    # 获取根logger
+    root_logger = logging.getLogger()
+    root_logger.setLevel(logging.INFO)
+
+    # 清除已有的handlers（避免重复）
+    root_logger.handlers.clear()
+
+    # 1. 控制台handler（只输出WARNING及以上级别）
+    console_handler = logging.StreamHandler()
+    console_handler.setLevel(logging.WARNING)
+    console_handler.setFormatter(formatter)
+    root_logger.addHandler(console_handler)
+
+    # 确保日志文件所在目录存在
+    log_dir = os.path.dirname(os.path.abspath(__file__))
+
+    # 2. 所有日志文件handler（all_log.log）
+    all_log_path = os.path.join(log_dir, 'all_log.log')
+    all_log_handler = RotatingFileHandler(
+        all_log_path,
+        maxBytes=10*1024*1024,  # 10MB
+        backupCount=5,
+        encoding='utf-8'
     )
+    all_log_handler.setLevel(logging.DEBUG)  # 记录所有级别
+    all_log_handler.setFormatter(formatter)
+    root_logger.addHandler(all_log_handler)
+
+    # 3. 错误日志文件handler（error.log）
+    error_log_path = os.path.join(log_dir, 'error.log')
+    error_log_handler = RotatingFileHandler(
+        error_log_path,
+        maxBytes=10*1024*1024,  # 10MB
+        backupCount=5,
+        encoding='utf-8'
+    )
+    error_log_handler.setLevel(logging.ERROR)  # 只记录ERROR及以上
+    error_log_handler.setFormatter(formatter)
+    root_logger.addHandler(error_log_handler)
+
+# 初始化日志系统
+setup_logging()
 logger = logging.getLogger('worker')
 
 

services/api_client.py

@@ -7,6 +7,7 @@ v2 API 客户端
 
 import logging
 import subprocess
+import time
 import requests
 from typing import Dict, List, Optional, Any
 
@@ -24,6 +25,8 @@ class APIClientV2:
     负责与渲染服务端的所有 HTTP 通信。
     """
 
+    SYSTEM_INFO_TTL_SECONDS = 30
+
     def __init__(self, config: WorkerConfig):
         """
         初始化 API 客户端
@@ -37,6 +40,15 @@ class APIClientV2:
         self.worker_id = config.worker_id
         self.session = requests.Session()
 
+        self._ffmpeg_version: Optional[str] = None
+        self._codec_info: Optional[str] = None
+        self._hw_accel_info: Optional[str] = None
+        self._gpu_info: Optional[str] = None
+        self._gpu_info_checked = False
+        self._static_system_info: Optional[Dict[str, Any]] = None
+        self._system_info_cache: Optional[Dict[str, Any]] = None
+        self._system_info_cache_ts = 0.0
+
         # 设置默认请求头
         self.session.headers.update({
             'Content-Type': 'application/json',
@@ -287,6 +299,8 @@ class APIClientV2:
 
     def _get_ffmpeg_version(self) -> str:
         """获取 FFmpeg 版本"""
+        if self._ffmpeg_version is not None:
+            return self._ffmpeg_version
         try:
             result = subprocess.run(
                 ['ffmpeg', '-version'],
@@ -299,13 +313,18 @@ class APIClientV2:
                 parts = first_line.split()
                 for i, part in enumerate(parts):
                     if part == 'version' and i + 1 < len(parts):
-                        return parts[i + 1]
-            return 'unknown'
+                        self._ffmpeg_version = parts[i + 1]
+                        return self._ffmpeg_version
+            self._ffmpeg_version = 'unknown'
+            return self._ffmpeg_version
         except Exception:
-            return 'unknown'
+            self._ffmpeg_version = 'unknown'
+            return self._ffmpeg_version
 
     def _get_codec_info(self) -> str:
         """获取支持的编解码器信息"""
+        if self._codec_info is not None:
+            return self._codec_info
         try:
             result = subprocess.run(
                 ['ffmpeg', '-codecs'],
@@ -324,37 +343,60 @@ class APIClientV2:
                 codecs.append('aac')
             if 'libfdk_aac' in output:
                 codecs.append('libfdk_aac')
-            return ', '.join(codecs) if codecs else 'unknown'
+            self._codec_info = ', '.join(codecs) if codecs else 'unknown'
+            return self._codec_info
         except Exception:
-            return 'unknown'
+            self._codec_info = 'unknown'
+            return self._codec_info
 
     def _get_system_info(self) -> Dict[str, Any]:
         """获取系统信息"""
         try:
+            now = time.monotonic()
+            if (
+                self._system_info_cache
+                and now - self._system_info_cache_ts < self.SYSTEM_INFO_TTL_SECONDS
+            ):
+                return self._system_info_cache
+
             import platform
             import psutil
 
-            info = {
+            if self._hw_accel_info is None:
+                self._hw_accel_info = get_hw_accel_info_str()
+
+            if self._static_system_info is None:
+                self._static_system_info = {
                     'os': platform.system(),
                     'cpu': f"{psutil.cpu_count()} cores",
                     'memory': f"{psutil.virtual_memory().total // (1024**3)}GB",
+                    'hwAccelConfig': self.config.hw_accel,  # 当前配置的硬件加速
+                    'hwAccelSupport': self._hw_accel_info,  # 系统支持的硬件加速
+                }
+
+            info = dict(self._static_system_info)
+            info.update({
                 'cpuUsage': f"{psutil.cpu_percent()}%",
                 'memoryAvailable': f"{psutil.virtual_memory().available // (1024**3)}GB",
-                'hwAccelConfig': self.config.hw_accel,  # 当前配置的硬件加速
-                'hwAccelSupport': get_hw_accel_info_str(),  # 系统支持的硬件加速
-            }
+            })
 
             # 尝试获取 GPU 信息
             gpu_info = self._get_gpu_info()
             if gpu_info:
                 info['gpu'] = gpu_info
 
+            self._system_info_cache = info
+            self._system_info_cache_ts = now
             return info
         except Exception:
             return {}
 
     def _get_gpu_info(self) -> Optional[str]:
         """获取 GPU 信息"""
+        if self._gpu_info_checked:
+            return self._gpu_info
+
+        self._gpu_info_checked = True
         try:
             result = subprocess.run(
                 ['nvidia-smi', '--query-gpu=name', '--format=csv,noheader'],
@@ -364,10 +406,11 @@ class APIClientV2:
             )
             if result.returncode == 0:
                 gpu_name = result.stdout.strip().split('\n')[0]
-                return gpu_name
+                self._gpu_info = gpu_name
         except Exception:
-            pass
-        return None
+            self._gpu_info = None
+
+        return self._gpu_info
 
     def close(self):
         """关闭会话"""

services/cache.py

@@ -5,14 +5,18 @@
 提供素材下载缓存功能，避免相同素材重复下载。
 """
 
+import json
 import os
 import hashlib
 import logging
 import shutil
 import time
+import uuid
 from typing import Optional, Tuple
 from urllib.parse import urlparse, unquote
 
+import psutil
+
 from services import storage
 
 logger = logging.getLogger(__name__)
@@ -59,6 +63,10 @@ class MaterialCache:
     负责素材文件的缓存存储和检索。
     """
 
+    LOCK_TIMEOUT_SEC = 30.0
+    LOCK_POLL_INTERVAL_SEC = 0.1
+    LOCK_STALE_SECONDS = 24 * 60 * 60
+
     def __init__(self, cache_dir: str, enabled: bool = True, max_size_gb: float = 0):
         """
         初始化缓存管理器
@@ -91,6 +99,134 @@ class MaterialCache:
         filename = f"{cache_key}{ext}"
         return os.path.join(self.cache_dir, filename)
 
+    def _get_lock_path(self, cache_key: str) -> str:
+        """获取缓存锁文件路径"""
+        assert self.cache_dir
+        return os.path.join(self.cache_dir, f"{cache_key}.lock")
+
+    def _write_lock_metadata(self, lock_fd: int, lock_path: str) -> bool:
+        """写入锁元数据，失败则清理锁文件"""
+        try:
+            try:
+                process_start_time = psutil.Process(os.getpid()).create_time()
+            except Exception as e:
+                process_start_time = None
+                logger.warning(f"Cache lock process start time error: {e}")
+            metadata = {
+                'pid': os.getpid(),
+                'process_start_time': process_start_time,
+                'created_at': time.time()
+            }
+            with os.fdopen(lock_fd, 'w', encoding='utf-8') as lock_file:
+                json.dump(metadata, lock_file)
+            return True
+        except Exception as e:
+            try:
+                os.close(lock_fd)
+            except Exception:
+                pass
+            self._remove_lock_file(lock_path, f"write metadata failed: {e}")
+            return False
+
+    def _read_lock_metadata(self, lock_path: str) -> Optional[dict]:
+        """读取锁元数据，失败返回 None（兼容历史空锁文件）"""
+        try:
+            with open(lock_path, 'r', encoding='utf-8') as lock_file:
+                data = json.load(lock_file)
+            return data if isinstance(data, dict) else None
+        except Exception:
+            return None
+
+    def _is_process_alive(self, pid: int, expected_start_time: Optional[float]) -> bool:
+        """判断进程是否存活并校验启动时间（防止 PID 复用）"""
+        try:
+            process = psutil.Process(pid)
+            if expected_start_time is None:
+                return process.is_running()
+            actual_start_time = process.create_time()
+            return abs(actual_start_time - expected_start_time) < 1.0
+        except psutil.NoSuchProcess:
+            return False
+        except Exception as e:
+            logger.warning(f"Cache lock process check error: {e}")
+            return True
+
+    def _is_lock_stale(self, lock_path: str) -> bool:
+        """判断锁是否过期（进程已退出或超过最大存活时长）"""
+        if not os.path.exists(lock_path):
+            return False
+        now = time.time()
+        metadata = self._read_lock_metadata(lock_path)
+        if metadata:
+            created_at = metadata.get('created_at')
+            if isinstance(created_at, (int, float)) and now - created_at > self.LOCK_STALE_SECONDS:
+                return True
+            pid = metadata.get('pid')
+            pid_value = int(pid) if isinstance(pid, int) or (isinstance(pid, str) and pid.isdigit()) else None
+            expected_start_time = metadata.get('process_start_time')
+            expected_start_time_value = (
+                expected_start_time if isinstance(expected_start_time, (int, float)) else None
+            )
+            if pid_value is not None and not self._is_process_alive(pid_value, expected_start_time_value):
+                return True
+            return self._is_lock_stale_by_mtime(lock_path, now)
+        return self._is_lock_stale_by_mtime(lock_path, now)
+
+    def _is_lock_stale_by_mtime(self, lock_path: str, now: float) -> bool:
+        """基于文件时间判断锁是否过期"""
+        try:
+            mtime = os.path.getmtime(lock_path)
+            return now - mtime > self.LOCK_STALE_SECONDS
+        except Exception as e:
+            logger.warning(f"Cache lock stat error: {e}")
+            return False
+
+    def _remove_lock_file(self, lock_path: str, reason: str = "") -> bool:
+        """删除锁文件"""
+        try:
+            os.remove(lock_path)
+            if reason:
+                logger.info(f"Cache lock removed: {lock_path} ({reason})")
+            return True
+        except FileNotFoundError:
+            return True
+        except Exception as e:
+            logger.warning(f"Cache lock remove error: {e}")
+            return False
+
+    def _acquire_lock(self, cache_key: str) -> Optional[str]:
+        """获取缓存锁（跨进程安全）"""
+        if not self.enabled:
+            return None
+
+        lock_path = self._get_lock_path(cache_key)
+        deadline = time.monotonic() + self.LOCK_TIMEOUT_SEC
+
+        while True:
+            try:
+                fd = os.open(lock_path, os.O_CREAT | os.O_EXCL | os.O_WRONLY)
+                if not self._write_lock_metadata(fd, lock_path):
+                    return None
+                return lock_path
+            except FileExistsError:
+                if self._is_lock_stale(lock_path):
+                    removed = self._remove_lock_file(lock_path, "stale lock")
+                    if removed:
+                        continue
+                if time.monotonic() >= deadline:
+                    logger.warning(f"Cache lock timeout: {lock_path}")
+                    return None
+                time.sleep(self.LOCK_POLL_INTERVAL_SEC)
+            except Exception as e:
+                logger.warning(f"Cache lock error: {e}")
+                return None
+
+    def _release_lock(self, lock_path: Optional[str]) -> None:
+        """释放缓存锁"""
+        if not lock_path:
+            return
+        self._remove_lock_file(lock_path)
+
     def is_cached(self, url: str) -> Tuple[bool, str]:
         """
         检查素材是否已缓存
@@ -136,8 +272,15 @@ class MaterialCache:
         if not self.enabled:
             return storage.download_file(url, dest, max_retries=max_retries, timeout=timeout)
 
-        # 检查缓存
-        cached, cache_path = self.is_cached(url)
+        cache_key = _extract_cache_key(url)
+        lock_path = self._acquire_lock(cache_key)
+        if not lock_path:
+            logger.warning(f"Cache lock unavailable, downloading without cache: {url[:80]}...")
+            return storage.download_file(url, dest, max_retries=max_retries, timeout=timeout)
+
+        try:
+            cache_path = self.get_cache_path(url)
+            cached = os.path.exists(cache_path) and os.path.getsize(cache_path) > 0
 
             if cached:
                 # 命中缓存，复制到目标路径
@@ -159,8 +302,11 @@ class MaterialCache:
             # 未命中缓存，下载到缓存目录
             logger.debug(f"Cache miss: {url[:80]}...")
 
-        # 先下载到临时文件
-        temp_cache_path = cache_path + '.downloading'
+            # 先下载到临时文件（唯一文件名，避免并发覆盖）
+            temp_cache_path = os.path.join(
+                self.cache_dir,
+                f"{cache_key}.{uuid.uuid4().hex}.downloading"
+            )
             try:
                 if not storage.download_file(url, temp_cache_path, max_retries=max_retries, timeout=timeout):
                     # 下载失败，清理临时文件
@@ -168,10 +314,13 @@ class MaterialCache:
                         os.remove(temp_cache_path)
                     return False
 
-            # 下载成功，移动到正式缓存路径
-            if os.path.exists(cache_path):
-                os.remove(cache_path)
-            os.rename(temp_cache_path, cache_path)
+                if not os.path.exists(temp_cache_path) or os.path.getsize(temp_cache_path) <= 0:
+                    if os.path.exists(temp_cache_path):
+                        os.remove(temp_cache_path)
+                    return False
+
+                # 下载成功，原子替换缓存文件
+                os.replace(temp_cache_path, cache_path)
 
                 # 复制到目标路径
                 shutil.copy2(cache_path, dest)
@@ -193,6 +342,68 @@ class MaterialCache:
                     except Exception:
                         pass
                 return False
+        finally:
+            self._release_lock(lock_path)
+
+    def add_to_cache(self, url: str, source_path: str) -> bool:
+        """
+        将本地文件添加到缓存
+
+        Args:
+            url: 对应的 URL（用于生成缓存键）
+            source_path: 本地文件路径
+
+        Returns:
+            是否成功
+        """
+        if not self.enabled:
+            return False
+
+        if not os.path.exists(source_path):
+            logger.warning(f"Source file not found for cache: {source_path}")
+            return False
+
+        cache_key = _extract_cache_key(url)
+        lock_path = self._acquire_lock(cache_key)
+        if not lock_path:
+            logger.warning(f"Cache lock unavailable for adding: {url[:80]}...")
+            return False
+
+        try:
+            cache_path = self.get_cache_path(url)
+
+            # 先复制到临时文件
+            temp_cache_path = os.path.join(
+                self.cache_dir,
+                f"{cache_key}.{uuid.uuid4().hex}.adding"
+            )
+
+            shutil.copy2(source_path, temp_cache_path)
+
+            # 原子替换
+            os.replace(temp_cache_path, cache_path)
+
+            # 更新访问时间
+            os.utime(cache_path, None)
+
+            logger.info(f"Added to cache: {url[:80]}... <- {source_path}")
+
+            # 检查清理
+            if self.max_size_bytes > 0:
+                self._cleanup_if_needed()
+
+            return True
+
+        except Exception as e:
+            logger.error(f"Failed to add to cache: {e}")
+            if 'temp_cache_path' in locals() and os.path.exists(temp_cache_path):
+                try:
+                    os.remove(temp_cache_path)
+                except Exception:
+                    pass
+            return False
+        finally:
+            self._release_lock(lock_path)
 
     def _cleanup_if_needed(self) -> None:
         """
@@ -209,7 +420,7 @@ class MaterialCache:
             total_size = 0
 
             for filename in os.listdir(self.cache_dir):
-                if filename.endswith('.downloading'):
+                if filename.endswith('.downloading') or filename.endswith('.lock'):
                     continue
                 file_path = os.path.join(self.cache_dir, filename)
                 if os.path.isfile(file_path):
@@ -235,6 +446,17 @@ class MaterialCache:
             for file_info in cache_files:
                 if total_size <= target_size:
                     break
+                # 从文件名提取 cache_key，检查是否有锁（说明正在被使用）
+                filename = os.path.basename(file_info['path'])
+                cache_key = os.path.splitext(filename)[0]
+                lock_path = self._get_lock_path(cache_key)
+                if os.path.exists(lock_path):
+                    if self._is_lock_stale(lock_path):
+                        self._remove_lock_file(lock_path, "cleanup stale lock")
+                    else:
+                        # 该文件正在被其他任务使用，跳过删除
+                        logger.debug(f"Cache cleanup: skipping locked file {filename}")
+                        continue
                 try:
                     os.remove(file_info['path'])
                     total_size -= file_info['size']
@@ -275,7 +497,7 @@ class MaterialCache:
         total_size = 0
 
         for filename in os.listdir(self.cache_dir):
-            if filename.endswith('.downloading'):
+            if filename.endswith('.downloading') or filename.endswith('.lock'):
                 continue
             file_path = os.path.join(self.cache_dir, filename)
             if os.path.isfile(file_path):

services/gpu_scheduler.py

@@ -0,0 +1,164 @@
+# -*- coding: utf-8 -*-
+"""
+GPU 调度器
+
+提供多 GPU 设备的轮询调度功能。
+"""
+
+import logging
+import threading
+from typing import List, Optional
+
+from domain.config import WorkerConfig
+from domain.gpu import GPUDevice
+from util.system import get_all_gpu_info, validate_gpu_device
+from constant import HW_ACCEL_CUDA, HW_ACCEL_QSV
+
+logger = logging.getLogger(__name__)
+
+
+class GPUScheduler:
+    """
+    GPU 调度器
+
+    实现多 GPU 设备的轮询（Round Robin）调度。
+    线程安全，支持并发任务执行。
+
+    使用方式：
+        scheduler = GPUScheduler(config)
+
+        # 在任务执行时
+        device_index = scheduler.acquire()
+        try:
+            # 执行任务
+            pass
+        finally:
+            scheduler.release(device_index)
+    """
+
+    def __init__(self, config: WorkerConfig):
+        """
+        初始化调度器
+
+        Args:
+            config: Worker 配置
+        """
+        self._config = config
+        self._devices: List[GPUDevice] = []
+        self._next_index: int = 0
+        self._lock = threading.Lock()
+        self._enabled = False
+
+        # 初始化设备列表
+        self._init_devices()
+
+    def _init_devices(self) -> None:
+        """初始化 GPU 设备列表"""
+        # 仅在启用硬件加速时才初始化
+        if self._config.hw_accel not in (HW_ACCEL_CUDA, HW_ACCEL_QSV):
+            logger.info("Hardware acceleration not enabled, GPU scheduler disabled")
+            return
+
+        configured_devices = self._config.gpu_devices
+
+        if configured_devices:
+            # 使用配置指定的设备
+            self._devices = self._validate_configured_devices(configured_devices)
+        else:
+            # 自动检测所有设备
+            self._devices = self._auto_detect_devices()
+
+        if self._devices:
+            self._enabled = True
+            device_info = ', '.join(str(d) for d in self._devices)
+            logger.info(f"GPU scheduler initialized with {len(self._devices)} device(s): {device_info}")
+        else:
+            logger.warning("No GPU devices available, scheduler disabled")
+
+    def _validate_configured_devices(self, indices: List[int]) -> List[GPUDevice]:
+        """
+        验证配置的设备列表
+
+        Args:
+            indices: 配置的设备索引列表
+
+        Returns:
+            验证通过的设备列表
+        """
+        devices = []
+        for index in indices:
+            if validate_gpu_device(index):
+                devices.append(GPUDevice(
+                    index=index,
+                    name=f"GPU-{index}",
+                    available=True
+                ))
+            else:
+                logger.warning(f"GPU device {index} is not available, skipping")
+        return devices
+
+    def _auto_detect_devices(self) -> List[GPUDevice]:
+        """
+        自动检测所有可用 GPU
+
+        Returns:
+            检测到的设备列表
+        """
+        all_devices = get_all_gpu_info()
+        # 过滤不可用设备
+        return [d for d in all_devices if d.available]
+
+    @property
+    def enabled(self) -> bool:
+        """调度器是否启用"""
+        return self._enabled
+
+    @property
+    def device_count(self) -> int:
+        """设备数量"""
+        return len(self._devices)
+
+    def acquire(self) -> Optional[int]:
+        """
+        获取下一个可用的 GPU 设备（轮询调度）
+
+        Returns:
+            GPU 设备索引，如果调度器未启用或无设备则返回 None
+        """
+        if not self._enabled or not self._devices:
+            return None
+
+        with self._lock:
+            device = self._devices[self._next_index]
+            self._next_index = (self._next_index + 1) % len(self._devices)
+            logger.debug(f"Acquired GPU device: {device.index}")
+            return device.index
+
+    def release(self, device_index: Optional[int]) -> None:
+        """
+        释放 GPU 设备
+
+        当前实现为无状态轮询，此方法仅用于日志记录。
+
+        Args:
+            device_index: 设备索引
+        """
+        if device_index is not None:
+            logger.debug(f"Released GPU device: {device_index}")
+
+    def get_status(self) -> dict:
+        """
+        获取调度器状态信息
+
+        Returns:
+            状态字典
+        """
+        return {
+            'enabled': self._enabled,
+            'device_count': len(self._devices),
+            'devices': [
+                {'index': d.index, 'name': d.name, 'available': d.available}
+                for d in self._devices
+            ],
+            'hw_accel': self._config.hw_accel,
+        }

services/storage.py

@@ -7,13 +7,38 @@
 
 import os
 import logging
+import subprocess
 from typing import Optional
+from urllib.parse import unquote
 
 import requests
 
 logger = logging.getLogger(__name__)
 
 
+# 文件扩展名到 Content-Type 的映射
+_CONTENT_TYPE_MAP = {
+    '.mp4': 'video/mp4',
+    '.aac': 'audio/aac',
+    '.ts': 'video/mp2t',
+    '.m4a': 'audio/mp4',
+}
+
+
+def _get_content_type(file_path: str) -> str:
+    """
+    根据文件扩展名获取 Content-Type
+
+    Args:
+        file_path: 文件路径
+
+    Returns:
+        Content-Type 字符串
+    """
+    ext = os.path.splitext(file_path)[1].lower()
+    return _CONTENT_TYPE_MAP.get(ext, 'application/octet-stream')
+
+
 def _apply_http_replace_map(url: str) -> str:
     """
     应用 HTTP_REPLACE_MAP 环境变量替换 URL
@@ -61,6 +86,7 @@ def upload_file(url: str, file_path: str, max_retries: int = 5, timeout: int = 6
 
     # 检查是否使用 rclone 上传
     if os.getenv("UPLOAD_METHOD") == "rclone":
+        logger.info(f"Uploading to: {url}")
         result = _upload_with_rclone(url, file_path)
         if result:
             return True
@@ -68,18 +94,20 @@ def upload_file(url: str, file_path: str, max_retries: int = 5, timeout: int = 6
 
     # 应用 HTTP_REPLACE_MAP 替换 URL
     http_url = _apply_http_replace_map(url)
+    content_type = _get_content_type(file_path)
+    logger.info(f"Uploading to: {http_url} (Content-Type: {content_type})")
 
     retries = 0
     while retries < max_retries:
         try:
             with open(file_path, 'rb') as f:
-                response = requests.put(
+                with requests.put(
                     http_url,
                     data=f,
                     stream=True,
                     timeout=timeout,
-                    headers={"Content-Type": "application/octet-stream"}
-                )
+                    headers={"Content-Type": content_type}
+                ) as response:
                     response.raise_for_status()
                     logger.info(f"Upload succeeded: {file_path}")
                     return True
@@ -111,7 +139,6 @@ def _upload_with_rclone(url: str, file_path: str) -> bool:
         return False
 
     config_file = os.getenv("RCLONE_CONFIG_FILE", "")
-    rclone_config = f"--config {config_file}" if config_file else ""
 
     # 替换 URL
     new_url = url
@@ -119,23 +146,35 @@ def _upload_with_rclone(url: str, file_path: str) -> bool:
     for src, dst in replace_list:
         new_url = new_url.replace(src, dst)
     new_url = new_url.split("?", 1)[0]  # 移除查询参数
+    new_url = unquote(new_url)  # 解码 URL 编码的字符（如 %2F -> /）
 
     if new_url == url:
         return False
 
-    cmd = (
-        f"rclone copyto --no-check-dest --ignore-existing "
-        f"--multi-thread-chunk-size 8M --multi-thread-streams 8 "
-        f"{rclone_config} {file_path} {new_url}"
-    )
-    logger.debug(f"rclone command: {cmd}")
+    cmd = [
+        "rclone",
+        "copyto",
+        "--no-check-dest",
+        "--ignore-existing",
+        "--multi-thread-chunk-size",
+        "8M",
+        "--multi-thread-streams",
+        "8",
+    ]
+    if config_file:
+        cmd.extend(["--config", config_file])
+    cmd.extend([file_path, new_url])
 
-    result = os.system(cmd)
-    if result == 0:
+    logger.debug(f"rclone command: {' '.join(cmd)}")
+
+    result = subprocess.run(cmd, capture_output=True, text=True)
+    if result.returncode == 0:
         logger.info(f"rclone upload succeeded: {file_path}")
         return True
 
-    logger.warning(f"rclone upload failed (code={result}): {file_path}")
+    stderr = (result.stderr or '').strip()
+    stderr = stderr[:500] if stderr else ""
+    logger.warning(f"rclone upload failed (code={result.returncode}): {file_path} {stderr}")
     return False
 
 
@@ -177,7 +216,7 @@ def download_file(
     retries = 0
     while retries < max_retries:
         try:
-            response = requests.get(http_url, timeout=timeout, stream=True)
+            with requests.get(http_url, timeout=timeout, stream=True) as response:
                 response.raise_for_status()
 
                 with open(file_path, 'wb') as f:

services/task_executor.py

@@ -12,9 +12,16 @@ from typing import Dict, Optional, TYPE_CHECKING
 
 from domain.task import Task, TaskType
 from domain.result import TaskResult, ErrorCode
+
+# 需要 GPU 加速的任务类型
+GPU_REQUIRED_TASK_TYPES = {
+    TaskType.RENDER_SEGMENT_VIDEO,
+    TaskType.COMPOSE_TRANSITION,
+}
 from domain.config import WorkerConfig
 from core.handler import TaskHandler
 from services.lease_service import LeaseService
+from services.gpu_scheduler import GPUScheduler
 
 if TYPE_CHECKING:
     from services.api_client import APIClientV2
@@ -60,6 +67,12 @@ class TaskExecutor:
         # 线程安全锁
         self.lock = threading.Lock()
 
+        # GPU 调度器（如果启用硬件加速）
+        self.gpu_scheduler = GPUScheduler(config)
+
+        if self.gpu_scheduler.enabled:
+            logger.info(f"GPU scheduler enabled with {self.gpu_scheduler.device_count} device(s)")
+
         # 注册处理器
         self._register_handlers()
 
@@ -130,6 +143,14 @@ class TaskExecutor:
                 logger.warning(f"[task:{task.task_id}] Task already running, skipping")
                 return False
 
+            # 检查并发上限
+            if len(self.current_tasks) >= self.config.max_concurrency:
+                logger.info(
+                    f"[task:{task.task_id}] Max concurrency reached "
+                    f"({self.config.max_concurrency}), rejecting task"
+                )
+                return False
+
             # 检查是否有对应的处理器
             if task.task_type not in self.handlers:
                 logger.error(f"[task:{task.task_id}] No handler for type: {task.task_type.value}")
@@ -164,15 +185,28 @@ class TaskExecutor:
         )
         lease_service.start()
 
+        # 获取 GPU 设备（仅对需要 GPU 的任务类型）
+        device_index = None
+        needs_gpu = task.task_type in GPU_REQUIRED_TASK_TYPES
+        if needs_gpu and self.gpu_scheduler.enabled:
+            device_index = self.gpu_scheduler.acquire()
+            if device_index is not None:
+                logger.info(f"[task:{task_id}] Assigned to GPU device {device_index}")
+
+        # 获取处理器（需要在设置 GPU 设备前获取）
+        handler = self.handlers.get(task.task_type)
+
         try:
             # 报告任务开始
             self.api_client.report_start(task_id)
 
-            # 获取处理器
-            handler = self.handlers.get(task.task_type)
             if not handler:
                 raise ValueError(f"No handler for task type: {task.task_type}")
 
+            # 设置 GPU 设备（线程本地存储）
+            if device_index is not None:
+                handler.set_gpu_device(device_index)
+
             # 执行前钩子
             handler.before_handle(task)
 
@@ -196,6 +230,14 @@ class TaskExecutor:
             self.api_client.report_fail(task_id, 'E_UNKNOWN', str(e))
 
         finally:
+            # 清除 GPU 设备设置
+            if handler:
+                handler.clear_gpu_device()
+
+            # 释放 GPU 设备（仅当实际分配了设备时）
+            if device_index is not None:
+                self.gpu_scheduler.release(device_index)
+
             # 停止租约续期
             lease_service.stop()
 

tests/unit/test_material_cache_lock.py

@@ -0,0 +1,15 @@
+# -*- coding: utf-8 -*-
+
+import os
+
+from services.cache import MaterialCache, _extract_cache_key
+
+
+def test_cache_lock_acquire_release(tmp_path):
+    cache = MaterialCache(cache_dir=str(tmp_path), enabled=True, max_size_gb=0)
+    cache_key = _extract_cache_key("https://example.com/path/file.mp4?token=abc")
+    lock_path = cache._acquire_lock(cache_key)
+    assert lock_path
+    assert os.path.exists(lock_path)
+    cache._release_lock(lock_path)
+    assert not os.path.exists(lock_path)

util/system.py

@@ -5,13 +5,17 @@
 提供系统信息采集功能。
 """
 
+import logging
 import os
 import platform
 import subprocess
-from typing import Optional, Dict, Any
+from typing import Optional, Dict, Any, List
 
 import psutil
 from constant import SOFTWARE_VERSION, DEFAULT_CAPABILITIES, HW_ACCEL_NONE, HW_ACCEL_QSV, HW_ACCEL_CUDA
+from domain.gpu import GPUDevice
+
+logger = logging.getLogger(__name__)
 
 
 def get_sys_info():
@@ -264,3 +268,78 @@ def get_hw_accel_info_str() -> str:
         return "No hardware acceleration available"
 
     return ', '.join(parts) + f" [recommended: {support['recommended']}]"
+
+
+def get_all_gpu_info() -> List[GPUDevice]:
+    """
+    获取所有 NVIDIA GPU 信息
+
+    使用 nvidia-smi 查询所有 GPU 设备。
+
+    Returns:
+        GPU 设备列表，失败返回空列表
+    """
+    try:
+        result = subprocess.run(
+            [
+                'nvidia-smi',
+                '--query-gpu=index,name,memory.total',
+                '--format=csv,noheader,nounits'
+            ],
+            capture_output=True,
+            text=True,
+            timeout=10
+        )
+
+        if result.returncode != 0:
+            return []
+
+        devices = []
+        for line in result.stdout.strip().split('\n'):
+            if not line.strip():
+                continue
+            parts = [p.strip() for p in line.split(',')]
+            if len(parts) >= 2:
+                index = int(parts[0])
+                name = parts[1]
+                memory = int(parts[2]) if len(parts) >= 3 else None
+                devices.append(GPUDevice(
+                    index=index,
+                    name=name,
+                    memory_total=memory,
+                    available=True
+                ))
+
+        return devices
+
+    except Exception as e:
+        logger.warning(f"Failed to detect GPUs: {e}")
+        return []
+
+
+def validate_gpu_device(index: int) -> bool:
+    """
+    验证指定索引的 GPU 设备是否可用
+
+    Args:
+        index: GPU 设备索引
+
+    Returns:
+        设备是否可用
+    """
+    try:
+        result = subprocess.run(
+            [
+                'nvidia-smi',
+                '-i', str(index),
+                '--query-gpu=name',
+                '--format=csv,noheader'
+            ],
+            capture_output=True,
+            text=True,
+            timeout=5
+        )
+        return result.returncode == 0 and bool(result.stdout.strip())
+    except Exception:
+        return False
+