fal 1.13.2__tar.gz → 1.13.4__tar.gz

{fal-1.13.2/fal.egg-info → fal-1.13.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fal
-Version: 1.13.2
+Version: 1.13.4
 Summary: fal is an easy-to-use Serverless Python Framework
 Author: Features & Labels <support@fal.ai>
 Requires-Python: >=3.8

{fal-1.13.2 → fal-1.13.4/fal.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fal
-Version: 1.13.2
+Version: 1.13.4
 Summary: fal is an easy-to-use Serverless Python Framework
 Author: Features & Labels <support@fal.ai>
 Requires-Python: >=3.8

{fal-1.13.2 → fal-1.13.4}/src/fal/_fal_version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '1.13.2'
-__version_tuple__ = version_tuple = (1, 13, 2)
+__version__ = version = '1.13.4'
+__version_tuple__ = version_tuple = (1, 13, 4)

{fal-1.13.2 → fal-1.13.4}/src/fal/app.py

@@ -345,7 +345,25 @@ class App(BaseServable):
         async def provide_hints_headers(request, call_next):
             response = await call_next(request)
             try:
-                response.headers["X-Fal-Runner-Hints"] = ",".join(self.provide_hints())
+                # make sure the hints can be encoded in latin-1, so we don't crash
+                # when serving.
+                # https://github.com/encode/starlette/blob/a766a58d14007f07c0b5782fa78cdc370b892796/starlette/datastructures.py#L568
+                hints = []
+                for hint in self.provide_hints():
+                    try:
+                        _ = hint.encode("latin-1")
+                        hints.append(hint)
+                    except UnicodeEncodeError:
+                        from fastapi.logger import logger
+
+                        logger.warning(
+                            "Ignoring hint %s for %s because it can't be encoded in "
+                            "latin-1",
+                            hint,
+                            self.__class__.__name__,
+                        )
+
+                response.headers["X-Fal-Runner-Hints"] = ",".join(hints)
             except NotImplementedError:
                 # This lets us differentiate between apps that don't provide hints
                 # and apps that provide empty hints.

{fal-1.13.2 → fal-1.13.4}/src/fal/exceptions/_cuda.py

@@ -38,6 +38,9 @@ def _is_cuda_out_of_memory(exception: BaseException) -> bool:
     if "NVML_SUCCESS == r INTERNAL ASSERT FAILED" in msg:
         return True
 
+    if "CUDNN_STATUS_INTERNAL_ERROR" in msg:
+        return True
+
     return False
 
 

{fal-1.13.2 → fal-1.13.4}/tests/test_apps.py

@@ -223,6 +223,10 @@ class ExceptionApp(fal.App, keep_alive=300, max_concurrency=1):
         # https://github.com/pytorch/pytorch/issues/112377
         raise RuntimeError("NVML_SUCCESS == r INTERNAL ASSERT FAILED")
 
+    @fal.endpoint("/cuda-exception-3")
+    def cuda_exception_3(self) -> Output:
+        raise RuntimeError("cuDNN error: CUDNN_STATUS_INTERNAL_ERROR")
+
 
 class CancellableApp(fal.App, keep_alive=300, max_concurrency=1):
     task = None
@@ -910,6 +914,12 @@ def test_app_exceptions(test_exception_app: AppClient):
     assert cuda_exc.value.status_code == _CUDA_OOM_STATUS_CODE
     assert _CUDA_OOM_MESSAGE in cuda_exc.value.message
 
+    with pytest.raises(AppClientError) as cuda_exc:
+        test_exception_app.cuda_exception_3({})
+
+    assert cuda_exc.value.status_code == _CUDA_OOM_STATUS_CODE
+    assert _CUDA_OOM_MESSAGE in cuda_exc.value.message
+
 
 def test_kill_runner():
     app_alias = str(uuid.uuid4()) + "-sleep-alias"
@@ -956,3 +966,31 @@ def test_container_app_client(test_container_app: str):
 def test_container_build_args_app_client(test_container_build_args_app: str):
     response = apps.run(test_container_build_args_app, {})
     assert response == "built with build args"
+
+
+class HintsApp(fal.App, keep_alive=300, max_concurrency=1):
+    machine_type = "S"
+
+    def provide_hints(self) -> List[str]:
+        return ["é", "😀"]
+
+    @fal.endpoint("/add")
+    def add(self, input: Input) -> Output:
+        return Output(result=input.lhs + input.rhs)
+
+
+def test_hints_encoding():
+    """
+    Make sure that hints that can't be encoded in latin-1 don't crash the app
+    https://github.com/encode/starlette/blob/a766a58d14007f07c0b5782fa78cdc370b892796/starlette/datastructures.py#L568
+    """
+    with AppClient.connect(HintsApp) as client:
+        with httpx.Client() as httpx_client:
+            url = client.url + "/add"
+            resp = httpx_client.post(
+                url,
+                json={"lhs": 1, "rhs": 2},
+                timeout=30,
+            )
+            assert resp.is_success
+            assert resp.json()["result"] == 3