fal 1.29.0__tar.gz → 1.30.0__tar.gz

{fal-1.29.0/fal.egg-info → fal-1.30.0}/PKG-INFO

@@ -1,12 +1,12 @@
 Metadata-Version: 2.4
 Name: fal
-Version: 1.29.0
+Version: 1.30.0
 Summary: fal is an easy-to-use Serverless Python Framework
 Author: Features & Labels <support@fal.ai>
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 Requires-Dist: isolate[build]<0.19.0,>=0.18.0
-Requires-Dist: isolate-proto<0.11.0,>=0.10.2
+Requires-Dist: isolate-proto<0.13.0,>=0.12.0
 Requires-Dist: grpcio<2,>=1.64.0
 Requires-Dist: dill==0.3.7
 Requires-Dist: cloudpickle==3.0.0
@@ -31,7 +31,7 @@ Requires-Dist: python-dateutil<3,>=2.8.0
 Requires-Dist: types-python-dateutil<3,>=2.8.0
 Requires-Dist: importlib-metadata>=4.4; python_version < "3.10"
 Requires-Dist: msgpack<2,>=1.0.7
-Requires-Dist: websockets<13,>=12.0
+Requires-Dist: websockets<16,>=12.0
 Requires-Dist: pillow<11,>=10.2.0
 Requires-Dist: pyjwt[crypto]<3,>=2.8.0
 Requires-Dist: uvicorn<1,>=0.29.0

{fal-1.29.0 → fal-1.30.0/fal.egg-info}/PKG-INFO

@@ -1,12 +1,12 @@
 Metadata-Version: 2.4
 Name: fal
-Version: 1.29.0
+Version: 1.30.0
 Summary: fal is an easy-to-use Serverless Python Framework
 Author: Features & Labels <support@fal.ai>
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 Requires-Dist: isolate[build]<0.19.0,>=0.18.0
-Requires-Dist: isolate-proto<0.11.0,>=0.10.2
+Requires-Dist: isolate-proto<0.13.0,>=0.12.0
 Requires-Dist: grpcio<2,>=1.64.0
 Requires-Dist: dill==0.3.7
 Requires-Dist: cloudpickle==3.0.0
@@ -31,7 +31,7 @@ Requires-Dist: python-dateutil<3,>=2.8.0
 Requires-Dist: types-python-dateutil<3,>=2.8.0
 Requires-Dist: importlib-metadata>=4.4; python_version < "3.10"
 Requires-Dist: msgpack<2,>=1.0.7
-Requires-Dist: websockets<13,>=12.0
+Requires-Dist: websockets<16,>=12.0
 Requires-Dist: pillow<11,>=10.2.0
 Requires-Dist: pyjwt[crypto]<3,>=2.8.0
 Requires-Dist: uvicorn<1,>=0.29.0

{fal-1.29.0 → fal-1.30.0}/fal.egg-info/SOURCES.txt

@@ -155,6 +155,7 @@ src/fal/toolkit/utils/__init__.py
 src/fal/toolkit/utils/download_utils.py
 src/fal/toolkit/utils/endpoint.py
 src/fal/toolkit/utils/retry.py
+src/fal/toolkit/utils/setup_utils.py
 src/fal/toolkit/video/__init__.py
 src/fal/toolkit/video/video.py
 tests/__init__.py

{fal-1.29.0 → fal-1.30.0}/fal.egg-info/requires.txt

@@ -1,5 +1,5 @@
 isolate[build]<0.19.0,>=0.18.0
-isolate-proto<0.11.0,>=0.10.2
+isolate-proto<0.13.0,>=0.12.0
 grpcio<2,>=1.64.0
 dill==0.3.7
 cloudpickle==3.0.0
@@ -23,7 +23,7 @@ attrs>=21.3.0
 python-dateutil<3,>=2.8.0
 types-python-dateutil<3,>=2.8.0
 msgpack<2,>=1.0.7
-websockets<13,>=12.0
+websockets<16,>=12.0
 pillow<11,>=10.2.0
 pyjwt[crypto]<3,>=2.8.0
 uvicorn<1,>=0.29.0

{fal-1.29.0 → fal-1.30.0}/pyproject.toml

@@ -23,7 +23,7 @@ readme = "README.md"
 requires-python = ">=3.8"
 dependencies = [
     "isolate[build]>=0.18.0,<0.19.0",
-    "isolate-proto>=0.10.2,<0.11.0",
+    "isolate-proto>=0.12.0,<0.13.0",
     "grpcio>=1.64.0,<2",
     "dill==0.3.7",
     "cloudpickle==3.0.0",
@@ -51,7 +51,7 @@ dependencies = [
     # For 3.9 and earlier, importlib-metadata's newer versions are included in the standard library.
     'importlib-metadata>=4.4; python_version < "3.10"',
     "msgpack>=1.0.7,<2",
-    "websockets>=12.0,<13",
+    "websockets>=12.0,<16",
     "pillow>=10.2.0,<11",
     "pyjwt[crypto]>=2.8.0,<3",
     "uvicorn>=0.29.0,<1",

{fal-1.29.0 → fal-1.30.0}/src/fal/_fal_version.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '1.29.0'
-__version_tuple__ = version_tuple = (1, 29, 0)
+__version__ = version = '1.30.0'
+__version_tuple__ = version_tuple = (1, 30, 0)
 
 __commit_id__ = commit_id = None

{fal-1.29.0 → fal-1.30.0}/src/fal/app.py

@@ -337,8 +337,10 @@ class App(BaseServable):
 
     def __init__(self, *, _allow_init: bool = False):
         if not _allow_init and not os.getenv("IS_ISOLATE_AGENT"):
+            cls_name = self.__class__.__name__
             raise NotImplementedError(
-                "Running apps through SDK is not implemented yet."
+                "Running apps through SDK is not implemented yet. "
+                f"Please use `fal run path/to/app.py::{cls_name}` to run your app."
             )
 
     @classmethod

{fal-1.29.0 → fal-1.30.0}/src/fal/cli/apps.py

@@ -5,6 +5,7 @@ from dataclasses import asdict
 from typing import TYPE_CHECKING
 
 import fal.cli.runners as runners
+from fal.sdk import RunnerState
 
 from ._utils import get_client
 from .parser import FalClientParser
@@ -298,7 +299,11 @@ def _runners(args):
         alias_runners = connection.list_alias_runners(alias=args.app_name)
 
     runners_table = runners.runners_table(alias_runners)
-    args.console.print(f"Runners: {len(alias_runners)}")
+    pending_runners = [
+        runner for runner in alias_runners if runner.state == RunnerState.PENDING
+    ]
+    args.console.print(f"Runners: {len(alias_runners) - len(pending_runners)}")
+    args.console.print(f"Pending Runners: {len(pending_runners)}")
     # Drop the alias column, which is the first column
     runners_table.columns.pop(0)
     args.console.print(runners_table)

{fal-1.29.0 → fal-1.30.0}/src/fal/cli/deploy.py

@@ -142,11 +142,16 @@ def _deploy_from_reference(
                 args.console.print(
                     f"\thttps://{playground_host}/models/{user.username}/{app_name}{endpoint}"
                 )
-            args.console.print("Endpoints:")
+            args.console.print("Synchronous Endpoints:")
             for endpoint in loaded.endpoints:
                 args.console.print(
                     f"\thttps://{endpoint_host}/{user.username}/{app_name}{endpoint}"
                 )
+            args.console.print("Asynchronous Endpoints (Recommended):")
+            for endpoint in loaded.endpoints:
+                args.console.print(
+                    f"\thttps://queue.{endpoint_host}/{user.username}/{app_name}{endpoint}"
+                )
 
 
 def _deploy(args):

{fal-1.29.0 → fal-1.30.0}/src/fal/cli/runners.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 from typing import List
 
-from fal.sdk import RunnerInfo
+from fal.sdk import RunnerInfo, RunnerState
 
 from ._utils import get_client
 from .parser import FalClientParser
@@ -19,6 +19,7 @@ def runners_table(runners: List[RunnerInfo]):
     table.add_column("Expires In")
     table.add_column("Uptime")
     table.add_column("Revision")
+    table.add_column("State")
 
     for runner in runners:
         external_metadata = runner.external_metadata
@@ -45,6 +46,7 @@ def runners_table(runners: List[RunnerInfo]):
             ),
             f"{runner.uptime} ({runner.uptime.total_seconds()}s)",
             runner.revision,
+            runner.state.value,
         )
 
     return table
@@ -82,7 +84,11 @@ def _list(args):
     client = get_client(args.host, args.team)
     with client.connect() as connection:
         runners = connection.list_runners()
-        args.console.print(f"Runners: {len(runners)}")
+        pending_runners = [
+            runner for runner in runners if runner.state == RunnerState.PENDING
+        ]
+        args.console.print(f"Runners: {len(runners) - len(pending_runners)}")
+        args.console.print(f"Pending Runners: {len(pending_runners)}")
         args.console.print(runners_table(runners))
 
         requests_table = runners_requests_table(runners)

{fal-1.29.0 → fal-1.30.0}/src/fal/sdk.py

@@ -261,6 +261,21 @@ class AliasInfo:
     valid_regions: list[str]
 
 
+class RunnerState(Enum):
+    RUNNING = "running"
+    PENDING = "pending"
+    UNKNOWN = "unknown"
+
+    @staticmethod
+    def from_proto(proto: isolate_proto.RunnerInfo.State) -> RunnerState:
+        if proto is isolate_proto.RunnerInfo.State.RUNNING:
+            return RunnerState.RUNNING
+        elif proto is isolate_proto.RunnerInfo.State.PENDING:
+            return RunnerState.PENDING
+        else:
+            return RunnerState.UNKNOWN
+
+
 @dataclass
 class RunnerInfo:
     runner_id: str
@@ -270,6 +285,15 @@ class RunnerInfo:
     external_metadata: dict[str, Any]
     revision: str
     alias: str
+    state: RunnerState
+
+
+@dataclass
+class ServiceURLs:
+    playground: str
+    run: str
+    queue: str
+    ws: str
 
 
 @dataclass
@@ -279,12 +303,14 @@ class HostedRunResult(Generic[ResultT]):
     logs: list[Log] = field(default_factory=list)
     result: ResultT | None = None
     stream: Any = None
+    service_urls: ServiceURLs | None = None
 
 
 @dataclass
 class RegisterApplicationResult:
     result: RegisterApplicationResultType | None
     logs: list[Log] = field(default_factory=list)
+    service_urls: ServiceURLs | None = None
 
 
 @dataclass
@@ -421,6 +447,7 @@ def _from_grpc_runner_info(message: isolate_proto.RunnerInfo) -> RunnerInfo:
         external_metadata=external_metadata,
         revision=message.revision,
         alias=message.alias,
+        state=RunnerState.from_proto(message.state),
     )
 
 
@@ -435,6 +462,9 @@ def _from_grpc_register_application_result(
             if not message.HasField("result")
             else RegisterApplicationResultType(message.result.application_id)
         ),
+        service_urls=from_grpc(message.service_urls)
+        if message.HasField("service_urls")
+        else None,
     )
 
 
@@ -445,6 +475,13 @@ def _from_grpc_hosted_run_status(
     return HostedRunStatus(HostedRunState(message.state))
 
 
+@from_grpc.register(isolate_proto.ServiceURLs)
+def _from_grpc_service_urls(
+    message: isolate_proto.ServiceURLs,
+) -> ServiceURLs:
+    return ServiceURLs(message.playground, message.run, message.queue, message.ws)
+
+
 @from_grpc.register(isolate_proto.HostedRunResult)
 def _from_grpc_hosted_run_result(
     message: isolate_proto.HostedRunResult,
@@ -459,6 +496,9 @@ def _from_grpc_hosted_run_result(
         from_grpc(message.status),
         logs=[from_grpc(log) for log in message.logs],
         result=return_value,
+        service_urls=from_grpc(message.service_urls)
+        if message.HasField("service_urls")
+        else None,
     )
 
 
@@ -769,7 +809,7 @@ class FalServerlessConnection:
         return [from_grpc(alias) for alias in response.aliases]
 
     def list_alias_runners(self, alias: str) -> list[RunnerInfo]:
-        request = isolate_proto.ListAliasRunnersRequest(alias=alias)
+        request = isolate_proto.ListAliasRunnersRequest(alias=alias, list_pending=True)
         response = self.stub.ListAliasRunners(request)
         return [from_grpc(runner) for runner in response.runners]
 
@@ -797,6 +837,6 @@ class FalServerlessConnection:
         self.stub.KillRunner(request)
 
     def list_runners(self) -> list[RunnerInfo]:
-        request = isolate_proto.ListRunnersRequest()
+        request = isolate_proto.ListRunnersRequest(list_pending=True)
         response = self.stub.ListRunners(request)
         return [from_grpc(runner) for runner in response.runners]

{fal-1.29.0 → fal-1.30.0}/src/fal/toolkit/utils/__init__.py

@@ -1,3 +1,4 @@
 from __future__ import annotations
 
 from fal.toolkit.utils.download_utils import *  # noqa: F403
+from fal.toolkit.utils.setup_utils import patch_onnx_runtime  # noqa: F401

fal-1.30.0/src/fal/toolkit/utils/setup_utils.py

@@ -0,0 +1,34 @@
+import os
+
+
+def patch_onnx_runtime(
+    inter_op_num_threads: int = 16,
+    intra_op_num_threads: int = 16,
+    omp_num_threads: int = 16,
+):
+    """
+    Patch ONNX Runtime's defaults to set the number of threads for inter-op,
+    intra-op, and OpenMP.Trying to use an ONNX Runtime session within a fal app
+    without explicitly setting these parameters can lead to issues, for example,
+    it can cause several logs related to these parameters to be printed.
+    Please run this function before importing any ONNX Runtime modules
+    in your application.
+
+    Args:
+        inter_op_num_threads (int): Number of threads for inter-op parallelism.
+        intra_op_num_threads (int): Number of threads for intra-op parallelism.
+        omp_num_threads (int): Number of threads for OpenMP parallelism.
+
+    """
+    import onnxruntime as ort
+
+    os.environ["OMP_NUM_THREADS"] = str(omp_num_threads)
+
+    _default_session_options = ort.capi._pybind_state.get_default_session_options()
+
+    def get_default_session_options_new():
+        _default_session_options.inter_op_num_threads = inter_op_num_threads
+        _default_session_options.intra_op_num_threads = intra_op_num_threads
+        return _default_session_options
+
+    ort.capi._pybind_state.get_default_session_options = get_default_session_options_new

{fal-1.29.0 → fal-1.30.0}/tests/test_apps.py

@@ -727,7 +727,12 @@ def test_app_deploy_scale(host: api.FalServerlessHost):
     )
 
     options = replace(
-        addition_app.options, host={**addition_app.options.host, "max_multiplexing": 30}
+        addition_app.options,
+        host={
+            **addition_app.options.host,
+            "max_multiplexing": 3,
+            "max_concurrency": 2,
+        },
     )
     kwargs = dict(
         func=addition_app.func,
@@ -743,7 +748,10 @@ def test_app_deploy_scale(host: api.FalServerlessHost):
         found = next(filter(lambda alias: alias.alias == app_alias, res), None)
         assert found, f"Could not find app {app_alias} in {res}"
         assert found.revision == app_revision
-        assert found.max_multiplexing == 1
+        # multiplexing is revision-specific
+        assert found.max_multiplexing == 3, "Expected max_multiplexing to have changed"
+        # max_concurrency is alias-specific
+        assert found.max_concurrency == 1, "Expected max_concurrency to stay the same"
 
     app_revision = addition_app.host.register(**kwargs, scale=True)
 
@@ -752,7 +760,9 @@ def test_app_deploy_scale(host: api.FalServerlessHost):
         found = next(filter(lambda alias: alias.alias == app_alias, res), None)
         assert found, f"Could not find app {app_alias} in {res}"
         assert found.revision == app_revision
-        assert found.max_multiplexing == 30
+        # when scaling, all values are updated
+        assert found.max_multiplexing == 3
+        assert found.max_concurrency == 2
 
 
 # List aliases is taking long
@@ -1015,7 +1025,8 @@ def test_kill_runner(host: api.FalServerlessHost, test_sleep_app: str):
         client.kill_runner(runners[0].runner_id)
 
         runners = client.list_alias_runners(app_alias)
-        assert len(runners) == 0
+        num_runners = len([runner for runner in runners if runner.state == "running"])
+        assert num_runners == 0
 
 
 def test_container_app_client(test_container_app: str):