feat(gpu): 接入 P4000 NVENC 硬件编码加速

- packages/config/base.py: 新增 GPU_ENCODE_* 配置项（开关/endpoint/relay/secret/编码参数/fallback） - packages/shared/gpu_encoder.py: 新增 GpuEncoderClient，封装 health 探测 + mezzanine 上传 OSS + P4000 nvenc 编码 + relay 回传下载，失败抛 GpuEncodeError 触发 CPU 降级 - apps/api/app/api/routes/gpu_relay.py: 新增 internal PUT/GET/DELETE /api/v1/internal/gpu-relay/{key}（token 鉴权），P4000 PUT 编码结果，worker GET 下载 - apps/api/app/api/router.py: 注册 gpu_relay_router - apps/worker/video_processing/unified_render_service.py: _execute_ffmpeg 和 _render_pass_through 尝试 GPU 路径：CPU ultrafast mezzanine → P4000 nvenc → 输出到最终路径；任何失败自动回退到原 CPU libx264 路径 - apps/worker/worker_app/tasks/_startup.py: worker_ready 时探测 P4000 健康并打日志 - tests/unit/test_gpu_encoder.py: GpuEncoderClient 单测（health/sync 调用/失败/fallback/relay URL）架构： - 输入：CPU 输出 libx264 ultrafast mezzanine → 上传 OSS 临时前缀 → P4000 签名 URL 下载 - 输出：P4000 PUT → 宿主机 nginx（tailscale:80）→ API /api/ 反代 → gpu_relay 路由落盘到 generated/gpu_relay/ - 回传：worker 通过 docker 网络 http://xiaoxia-api-staging:8000 GET 下载最终 mp4 到 output_path - 降级：GPU 任何环节异常（health/上传/编码/回传/下载）→ 原 CPU 路径继续执行，不影响成片
2026-09-26 18:39:42 +08:00
parent a5075624f8
commit 105ab54059
7 changed files with 952 additions and 28 deletions
@@ -57,6 +57,7 @@ from video_processing.tts_engine import TtsEngine
 from video_processing.watermark_engine import WatermarkConfig, WatermarkEngine

 from packages.domain.render_layer_utils import LAYER_Z_INDEX as _IMPORTED_LAYER_Z_INDEX
+from packages.shared.gpu_encoder import GpuEncodeError, get_gpu_encoder
 from packages.domain.render_layer_utils import clip_adjusted_duration as _clip_adjusted_duration_pure
 from packages.domain.render_layer_utils import clip_effective_duration as _clip_effective_duration_pure
 from packages.domain.render_layer_utils import clip_playback_speed as _clip_playback_speed_pure
@@ -1621,20 +1622,27 @@ class UnifiedRenderService:
            effective_duration,
            has_audio,
        )
-        try:
-            run_ffmpeg(command)
-        except subprocess.CalledProcessError as e:
-            stderr_text = (e.stderr or "").strip()
-            stderr_tail = stderr_text[-1500:] if len(stderr_text) > 1500 else stderr_text
-            logger.error(
-                "直通渲染失败: plan_id=%s clip=%s exit_code=%d\nvf=%s\nstderr(last 1500):\n%s",
-                self.plan.id,
-                clip.clip_id,
-                e.returncode,
-                vf_str[:2000],
-                stderr_tail,
-            )
-            raise
+        # 尝试 GPU NVENC 加速
+        gpu_ok = False
+        if self._gpu_encode_available():
+            mezz_path = output_path.parent / f".{output_path.stem}.mezz{output_path.suffix}"
+            gpu_ok = self._ffmpeg_output_to_mezzanine(command, mezz_path, output_path)
+
+        if not gpu_ok:
+            try:
+                run_ffmpeg(command)
+            except subprocess.CalledProcessError as e:
+                stderr_text = (e.stderr or "").strip()
+                stderr_tail = stderr_text[-1500:] if len(stderr_text) > 1500 else stderr_text
+                logger.error(
+                    "直通渲染失败: plan_id=%s clip=%s exit_code=%d\nvf=%s\nstderr(last 1500):\n%s",
+                    self.plan.id,
+                    clip.clip_id,
+                    e.returncode,
+                    vf_str[:2000],
+                    stderr_tail,
+                )
+                raise

        return has_audio

@@ -2170,6 +2178,120 @@ class UnifiedRenderService:
        filter_complex = ";".join(filter_parts)
        return filter_complex, input_args

+    # ── GPU NVENC 加速 ────────────────────────────────────────────────────
+
+    def _gpu_encode_available(self) -> bool:
+        """GPU 编码客户端是否已配置且健康（缓存健康状态，单任务内只探测一次）。"""
+        if not getattr(self, "_gpu_health_ok", None):
+            client = get_gpu_encoder()
+            if client is None:
+                self._gpu_health_ok = False
+                return False
+            try:
+                health = client.check_health()
+                if health.ready:
+                    logger.info(
+                        "[gpu-encoder] healthy endpoint=%s gpu=%s",
+                        client.endpoint, health.gpu_name,
+                    )
+                    self._gpu_health_ok = True
+                else:
+                    logger.warning(
+                        "[gpu-encoder] not ready: %s (endpoint=%s)",
+                        health.error, client.endpoint,
+                    )
+                    self._gpu_health_ok = False
+            except Exception as e:  # noqa: BLE001
+                logger.warning("[gpu-encoder] health probe error (CPU fallback): %s", e)
+                self._gpu_health_ok = False
+        return self._gpu_health_ok
+
+    def _ffmpeg_output_to_mezzanine(
+        self,
+        base_command: list[str],
+        mezzanine_path: Path,
+        output_path: Path,
+    ) -> bool:
+        """用 CPU ultrafast 把滤镜链输出到 mezzanine_path，然后调 GPU 做最终编码。
+
+        base_command: 原本要执行的完整 ffmpeg 命令（含 -c:v libx264 -crf X -preset Y ... output_path）
+        我们把最后一个参数（output_path）替换成 mezzanine_path，并把编码参数改成 ultrafast，
+        成功后调用 gpu_encoder 做 nvenc 编码到 output_path。
+
+        任何失败返回 False，调用方走原始 CPU 路径。
+        """
+        client = get_gpu_encoder()
+        if client is None:
+            return False
+
+        # 构造 mezzanine 命令：替换编码参数和输出路径
+        mezz_cmd = list(base_command)
+        # 找到编码参数位置并替换
+        try:
+            i_crf = mezz_cmd.index("-crf")
+            mezz_cmd[i_crf + 1] = "20"
+            i_preset = mezz_cmd.index("-preset")
+            mezz_cmd[i_preset + 1] = "ultrafast"
+        except ValueError:
+            logger.warning("[gpu-encoder] could not find -crf/-preset in command, skip gpu")
+            return False
+
+        # 如果命令有音频编码 -c:a aac，我们保留音频让 GPU 侧不用单独处理
+        # （P4000 的 ffmpeg_args 可以直接 copy 音频？这里简单起见：把音频编码留在 mezzanine，
+        # 然后 GPU 侧直接 -c:a copy，避免重编码损失）
+        has_audio = "-c:a" in mezz_cmd
+
+        # 替换输出路径（最后一个参数）
+        mezz_cmd[-1] = str(mezzanine_path)
+
+        # 1) 跑 mezzanine
+        mezzanine_path.parent.mkdir(parents=True, exist_ok=True)
+        t0 = time.time()
+        try:
+            run_ffmpeg(mezz_cmd)
+        except subprocess.CalledProcessError as e:
+            logger.warning("[gpu-encoder] mezzanine encode failed (CPU fallback): %s", e)
+            return False
+        logger.info(
+            "[gpu-encoder] mezzanine ready: %s (%.1fs, %d bytes), dispatching to P4000 nvenc...",
+            mezzanine_path.name, time.time() - t0,
+            mezzanine_path.stat().st_size if mezzanine_path.exists() else 0,
+        )
+
+        # 2) GPU nvenc encode（含上传 mezzanine → OSS → P4000 下载+编码 → relay 回传）
+        try:
+            # GPU 侧：-i in.mp4 -c:v h264_nvenc ... 音频 copy（mezzanine 里音频已是 aac）
+            audio_args = ["-c:a", "copy"] if has_audio else None
+            client.encode_mezzanine_to_output(
+                mezzanine_path,
+                output_path,
+                audio_args=audio_args,
+            )
+            logger.info(
+                "[gpu-encoder] GPU nvenc encode done: %s (total %.1fs)",
+                output_path.name, time.time() - t0,
+            )
+            return True
+        except GpuEncodeError as e:
+            logger.warning("[gpu-encoder] GPU encode failed (CPU fallback): %s", e)
+            # 删除可能残留的不完整 output
+            try:
+                if output_path.exists():
+                    output_path.unlink()
+            except OSError:
+                pass
+            return False
+        except Exception as e:  # noqa: BLE001
+            logger.warning("[gpu-encoder] GPU encode unexpected error (CPU fallback): %s", e)
+            return False
+        finally:
+            # 清理 mezzanine
+            try:
+                if mezzanine_path.exists():
+                    mezzanine_path.unlink()
+            except OSError:
+                pass
+
    def _execute_ffmpeg(
        self,
        filter_complex: str,
@@ -2209,20 +2331,28 @@ class UnifiedRenderService:
            input_args.count("-i"),
            output_path,
        )
-        try:
-            run_ffmpeg(command)
-        except subprocess.CalledProcessError as e:
-            # 额外记录 filter_complex + stderr，方便排查滤镜链构建问题
-            stderr_text = (e.stderr or "").strip()
-            stderr_tail = stderr_text[-1500:] if len(stderr_text) > 1500 else stderr_text
-            logger.error(
-                "渲染失败: plan_id=%s exit_code=%d\nfilter_complex:\n%s\nstderr(last 1500):\n%s",
-                self.plan.id,
-                e.returncode,
-                filter_complex[:5000],
-                stderr_tail,
-            )
-            raise
+
+        # 尝试 GPU NVENC 加速：先出 ultrafast mezzanine，再交给 P4000 做最终编码
+        gpu_ok = False
+        if self._gpu_encode_available():
+            mezz_path = output_path.parent / f".{output_path.stem}.mezz{output_path.suffix}"
+            gpu_ok = self._ffmpeg_output_to_mezzanine(command, mezz_path, output_path)
+
+        if not gpu_ok:
+            try:
+                run_ffmpeg(command)
+            except subprocess.CalledProcessError as e:
+                # 额外记录 filter_complex + stderr，方便排查滤镜链构建问题
+                stderr_text = (e.stderr or "").strip()
+                stderr_tail = stderr_text[-1500:] if len(stderr_text) > 1500 else stderr_text
+                logger.error(
+                    "渲染失败: plan_id=%s exit_code=%d\nfilter_complex:\n%s\nstderr(last 1500):\n%s",
+                    self.plan.id,
+                    e.returncode,
+                    filter_complex[:5000],
+                    stderr_tail,
+                )
+                raise

    def _build_sticker_filters(self, input_label: str, output_label: str) -> tuple[str, list[str]]:
        """构建贴纸叠加滤镜链.