perf(render_video): 优化视频编码的关键帧间隔策略

- 根据视频总帧数动态计算 GOP 大小，避免关键帧过多
- 短视频使用全部帧数作为 GOP，确保只在开头有关键帧
- 正常视频保持 2 秒一个关键帧的策略
- 添加强制第一帧为关键帧的设置
- 优化关键帧最小间隔参数，提升编码效率

handlers/base.py

@@ -114,6 +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
 
@@ -121,9 +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 ''
 
@@ -460,6 +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}")

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

@@ -123,9 +123,13 @@ 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")
@@ -271,12 +275,24 @@ 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
         ])
@@ -332,7 +348,7 @@ class RenderSegmentVideoHandler(BaseHandler):
             render_spec=render_spec,
             output_spec=output_spec,
             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
         )
@@ -353,16 +369,28 @@ 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')
 
@@ -376,7 +404,7 @@ class RenderSegmentVideoHandler(BaseHandler):
         render_spec: RenderSpec,
         output_spec: OutputSpec,
         lut_file: Optional[str] = None,
-        has_overlay: bool = False,
+        overlay_file: Optional[str] = None,
         overlap_head_ms: int = 0,
         overlap_tail_ms: int = 0
     ) -> str:
@@ -387,7 +415,7 @@ class RenderSegmentVideoHandler(BaseHandler):
             render_spec: 渲染规格
             output_spec: 输出规格
             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）
 
@@ -399,6 +427,10 @@ 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)
@@ -458,7 +490,10 @@ 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)
@@ -483,7 +518,13 @@ 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 ''
 
@@ -492,7 +533,10 @@ 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
@@ -506,7 +550,10 @@ 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）
@@ -531,15 +578,16 @@ 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
@@ -547,24 +595,21 @@ 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
@@ -587,7 +632,15 @@ 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）

services/cache.py

@@ -5,6 +5,7 @@
 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
@@ -14,6 +15,8 @@ import uuid
 from typing import Optional, Tuple
 from urllib.parse import urlparse, unquote
 
+import psutil
+
 from services import storage
 
 logger = logging.getLogger(__name__)
@@ -62,6 +65,7 @@ class MaterialCache:
 
     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):
         """
@@ -100,6 +104,96 @@ class MaterialCache:
         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:
@@ -111,9 +205,14 @@ class MaterialCache:
         while True:
             try:
                 fd = os.open(lock_path, os.O_CREAT | os.O_EXCL | os.O_WRONLY)
-                os.close(fd)
+                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
@@ -126,12 +225,7 @@ class MaterialCache:
         """释放缓存锁"""
         if not lock_path:
             return
-        try:
-            os.remove(lock_path)
-        except FileNotFoundError:
-            return
-        except Exception as e:
-            logger.warning(f"Cache lock release error: {e}")
+        self._remove_lock_file(lock_path)
 
     def is_cached(self, url: str) -> Tuple[bool, str]:
         """
@@ -251,6 +345,66 @@ class MaterialCache:
         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:
         """
         检查并清理缓存（LRU 策略）
@@ -292,6 +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']

services/storage.py

@@ -9,12 +9,36 @@ 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
@@ -62,6 +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
@@ -69,6 +94,8 @@ 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:
@@ -79,7 +106,7 @@ def upload_file(url: str, file_path: str, max_retries: int = 5, timeout: int = 6
                     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}")
@@ -119,6 +146,7 @@ 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

services/task_executor.py

@@ -12,6 +12,12 @@ 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
@@ -179,9 +185,10 @@ class TaskExecutor:
         )
         lease_service.start()
 
-        # 获取 GPU 设备
+        # 获取 GPU 设备（仅对需要 GPU 的任务类型）
         device_index = None
-        if self.gpu_scheduler.enabled:
+        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}")
@@ -227,8 +234,8 @@ class TaskExecutor:
             if handler:
                 handler.clear_gpu_device()
 
-            # 释放 GPU 设备
-            if self.gpu_scheduler.enabled:
+            # 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