fix: fallback windows asr to cpu when gpu runtime is missing

CHANGELOG.md

@@ -4,6 +4,11 @@
 
 ## 2026-04-06
 
+### Windows ASR GPU 失败时自动回退 CPU
+
+- 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。
+
 ### Windows ASR 默认改成 GPU 优先与自动语言识别
 
 - Windows `ASR HTTP` 现在默认不再强锁 `zh + cpu + int8`，而是改成：

deploy/storyforge-windows-asr-http/app.py

@@ -40,6 +40,9 @@ def describe_compute_mode() -> str:
 
 
 def build_runtime_profiles() -> list[tuple[str, str]]:
+    fallback_profile = getattr(app.state, "runtime_fallback_profile", None)
+    if fallback_profile:
+        return [fallback_profile]
     device = describe_device_mode()
     compute = describe_compute_mode()
     if device != "auto":
@@ -49,6 +52,20 @@ def build_runtime_profiles() -> list[tuple[str, str]]:
     return [("cuda", "int8_float16"), ("cpu", "int8")]
 
 
+def should_retry_on_cpu(exc: Exception) -> bool:
+    if describe_device_mode() != "auto":
+        return False
+    message = str(exc).lower()
+    return any(token in message for token in ("cublas", "cudnn", "cuda"))
+
+
+def activate_cpu_fallback() -> None:
+    app.state.runtime_fallback_profile = ("cpu", "int8")
+    app.state.runtime_device = "cpu"
+    app.state.runtime_compute_type = "int8"
+    get_model.cache_clear()
+
+
 @lru_cache(maxsize=1)
 def get_model():
     from faster_whisper import WhisperModel
@@ -93,15 +110,19 @@ def root() -> dict[str, str]:
     return {"service": "storyforge-windows-asr", "docs": "/docs"}
 
 
-@app.post("/transcribe", response_model=None)
-async def transcribe(wav: UploadFile = File(...)):
-    started = time.perf_counter()
-    suffix = Path(wav.filename or "segment.wav").suffix or ".wav"
-    with tempfile.NamedTemporaryFile(prefix="storyforge-asr-", suffix=suffix, delete=False) as handle:
-        temp_path = Path(handle.name)
-        handle.write(await wav.read())
-
+def transcribe_file(temp_path: Path, started: float) -> dict[str, object]:
+    model = get_model()
     try:
+        segments, info = model.transcribe(
+            str(temp_path),
+            language=resolve_language(),
+            beam_size=max(1, BEAM_SIZE),
+            vad_filter=VAD_FILTER,
+        )
+    except Exception as exc:
+        if not should_retry_on_cpu(exc):
+            raise
+        activate_cpu_fallback()
         model = get_model()
         segments, info = model.transcribe(
             str(temp_path),
@@ -119,6 +140,18 @@ async def transcribe(wav: UploadFile = File(...)):
         "detected_language_probability": getattr(info, "language_probability", None),
         "error_message": None if text else "empty transcription",
     }
+
+
+@app.post("/transcribe", response_model=None)
+async def transcribe(wav: UploadFile = File(...)):
+    started = time.perf_counter()
+    suffix = Path(wav.filename or "segment.wav").suffix or ".wav"
+    with tempfile.NamedTemporaryFile(prefix="storyforge-asr-", suffix=suffix, delete=False) as handle:
+        temp_path = Path(handle.name)
+        handle.write(await wav.read())
+
+    try:
+        return transcribe_file(temp_path, started)
     except Exception as exc:
         return JSONResponse(
             status_code=500,

tests/test_windows_asr_http.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import importlib.util
 import os
+import tempfile
 import unittest
 from pathlib import Path
 
@@ -69,3 +70,45 @@ class WindowsAsrHttpTests(unittest.TestCase):
         self.assertEqual(module.resolve_language(), "zh")
         self.assertEqual(module.describe_language_mode(), "zh")
         self.assertEqual(module.build_runtime_profiles(), [("cpu", "int8")])
+
+    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)
+        os.environ.pop("WHISPER_COMPUTE_TYPE", None)
+
+        module = load_windows_asr_app()
+
+        class FakeInfo:
+            language = "zh"
+            language_probability = 0.88
+
+        class FakeSegment:
+            text = "中英混合 hello world"
+
+        class BrokenGpuModel:
+            def transcribe(self, *_args, **_kwargs):
+                raise RuntimeError("Library cublas64_12.dll is not found or cannot be loaded")
+
+        class CpuModel:
+            def transcribe(self, *_args, **_kwargs):
+                return iter([FakeSegment()]), FakeInfo()
+
+        models = [BrokenGpuModel(), CpuModel()]
+        cache_cleared = []
+
+        def fake_get_model():
+            return models.pop(0)
+
+        fake_get_model.cache_clear = lambda: cache_cleared.append(True)
+        module.get_model = fake_get_model
+        module.app.state.runtime_device = "cuda"
+        module.app.state.runtime_compute_type = "int8_float16"
+
+        with tempfile.NamedTemporaryFile(suffix=".wav") as handle:
+            payload = module.transcribe_file(Path(handle.name), started=0.0)
+
+        self.assertTrue(payload["success"])
+        self.assertEqual(payload["text"], "中英混合 hello world")
+        self.assertEqual(payload["detected_language"], "zh")
+        self.assertEqual(module.app.state.runtime_fallback_profile, ("cpu", "int8"))
+        self.assertTrue(cache_cleared)