diff --git a/CHANGELOG.md b/CHANGELOG.md
index f288fb8..4ad8ee4 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,6 +9,7 @@
 - Windows `ASR HTTP` 现在在 `auto` 模式下仍会优先尝试 `cuda + int8_float16`，但如果在真正推理阶段命中 `cublas/cudnn/cuda` 运行库缺失，会自动切回 `cpu + int8` 重试，不再把整次转写卡死在 GPU 路径。
 - 这让“默认优先用 GPU、但当前机器 CUDA 运行库不完整”的场景也能稳定返回结果，同时保留混合中英文自动识别。
 - `smoke_public_storyforge.sh` 与 `smoke_fnos_storyforge_lan.sh` 现在会覆盖 ASR 转写链路，公网 smoke 在遇到 `127.0.0.1` 这类服务器内网地址时会自动跳过真实转写，避免在开发机上误判。
+- Windows ASR 运行时现在会自动发现 venv 里的 `nvidia-cublas-cu12 / nvidia-cuda-runtime-cu12 / nvidia-cudnn-cu12` DLL 目录并注入搜索路径；实机验证后 `active_device` 已经恢复为 `cuda`，不再长期回退到 CPU。
 
 ### Windows ASR 默认改成 GPU 优先与自动语言识别
 
diff --git a/deploy/storyforge-windows-asr-http/app.py b/deploy/storyforge-windows-asr-http/app.py
index a3b4b28..7e32348 100644
--- a/deploy/storyforge-windows-asr-http/app.py
+++ b/deploy/storyforge-windows-asr-http/app.py
@@ -1,6 +1,8 @@
 from __future__ import annotations
 
 import os
+import sys
+import sysconfig
 import tempfile
 import time
 from functools import lru_cache
@@ -15,6 +17,7 @@ VAD_FILTER = os.getenv("WHISPER_VAD_FILTER", "1").strip().lower() not in {"0", "
 DOWNLOAD_ROOT = Path(os.getenv("WHISPER_DOWNLOAD_ROOT", str(Path(__file__).resolve().parent / "models-cache")))
 
 app = FastAPI(title="storyforge-windows-asr", version="1.0.0")
+_dll_handles: list[object] = []
 
 
 def describe_language_mode() -> str:
@@ -66,8 +69,44 @@ def activate_cpu_fallback() -> None:
     get_model.cache_clear()
 
 
+def find_windows_cuda_runtime_dirs(site_packages_root: Path | None = None) -> list[Path]:
+    root = site_packages_root or Path(sysconfig.get_paths()["purelib"])
+    dirs = []
+    for rel in (
+        "nvidia/cublas/bin",
+        "nvidia/cuda_runtime/bin",
+        "nvidia/cuda_nvrtc/bin",
+        "nvidia/cudnn/bin",
+    ):
+        path = root / rel
+        if path.exists():
+            dirs.append(path)
+    return dirs
+
+
+def configure_windows_cuda_runtime() -> None:
+    if sys.platform != "win32":
+        return
+    configured = getattr(app.state, "windows_cuda_runtime_dirs", None)
+    if configured is not None:
+        return
+    runtime_dirs = find_windows_cuda_runtime_dirs()
+    app.state.windows_cuda_runtime_dirs = [str(path) for path in runtime_dirs]
+    if not runtime_dirs:
+        return
+    path_parts = os.environ.get("PATH", "").split(os.pathsep)
+    for runtime_dir in runtime_dirs:
+        runtime_dir_str = str(runtime_dir)
+        if runtime_dir_str not in path_parts:
+            path_parts.insert(0, runtime_dir_str)
+        if hasattr(os, "add_dll_directory"):
+            _dll_handles.append(os.add_dll_directory(runtime_dir_str))
+    os.environ["PATH"] = os.pathsep.join(path_parts)
+
+
 @lru_cache(maxsize=1)
 def get_model():
+    configure_windows_cuda_runtime()
     from faster_whisper import WhisperModel
 
     DOWNLOAD_ROOT.mkdir(parents=True, exist_ok=True)
@@ -91,6 +130,7 @@ def get_model():
 
 @app.get("/health")
 def health() -> dict[str, object]:
+    configure_windows_cuda_runtime()
     return {
         "status": "ok",
         "service": "storyforge-windows-asr",
@@ -102,6 +142,7 @@ def health() -> dict[str, object]:
         "active_compute_type": getattr(app.state, "runtime_compute_type", ""),
         "download_root": str(DOWNLOAD_ROOT),
         "model_loaded": get_model.cache_info().currsize > 0,
+        "windows_cuda_runtime_dirs": getattr(app.state, "windows_cuda_runtime_dirs", []),
     }
 
 
diff --git a/deploy/storyforge-windows-asr-http/requirements.txt b/deploy/storyforge-windows-asr-http/requirements.txt
index 9ea4c4c..341b5ad 100644
--- a/deploy/storyforge-windows-asr-http/requirements.txt
+++ b/deploy/storyforge-windows-asr-http/requirements.txt
@@ -2,3 +2,6 @@ fastapi==0.116.1
 uvicorn[standard]==0.35.0
 python-multipart==0.0.20
 faster-whisper>=1.1,<2
+nvidia-cublas-cu12; platform_system == "Windows"
+nvidia-cuda-runtime-cu12; platform_system == "Windows"
+nvidia-cudnn-cu12; platform_system == "Windows"
diff --git a/tests/test_windows_asr_http.py b/tests/test_windows_asr_http.py
index be09519..b5e1c2c 100644
--- a/tests/test_windows_asr_http.py
+++ b/tests/test_windows_asr_http.py
@@ -71,6 +71,30 @@ class WindowsAsrHttpTests(unittest.TestCase):
         self.assertEqual(module.describe_language_mode(), "zh")
         self.assertEqual(module.build_runtime_profiles(), [("cpu", "int8")])
 
+    def test_windows_cuda_runtime_discovery_finds_nvidia_wheel_bins(self) -> None:
+        module = load_windows_asr_app()
+        with tempfile.TemporaryDirectory() as tempdir:
+            root = Path(tempdir)
+            for rel in [
+                "nvidia/cublas/bin",
+                "nvidia/cuda_runtime/bin",
+                "nvidia/cuda_nvrtc/bin",
+                "nvidia/cudnn/bin",
+            ]:
+                (root / rel).mkdir(parents=True, exist_ok=True)
+
+            dirs = module.find_windows_cuda_runtime_dirs(root)
+
+        self.assertEqual(
+            [path.as_posix() for path in dirs],
+            [
+                (root / "nvidia/cublas/bin").as_posix(),
+                (root / "nvidia/cuda_runtime/bin").as_posix(),
+                (root / "nvidia/cuda_nvrtc/bin").as_posix(),
+                (root / "nvidia/cudnn/bin").as_posix(),
+            ],
+        )
+
     def test_auto_runtime_falls_back_to_cpu_when_cuda_runtime_is_missing(self) -> None:
         os.environ.pop("WHISPER_LANGUAGE", None)
         os.environ.pop("WHISPER_DEVICE", None)