fal 1.41.1__tar.gz → 1.43.0__tar.gz

{fal-1.41.1/fal.egg-info → fal-1.43.0}/PKG-INFO

@@ -1,12 +1,12 @@
 Metadata-Version: 2.4
 Name: fal
-Version: 1.41.1
+Version: 1.43.0
 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]<0.21.0,>=0.18.0
-Requires-Dist: isolate-proto<0.18.0,>=0.17.0
+Requires-Dist: isolate-proto<0.19.0,>=0.18.0
 Requires-Dist: grpcio<2,>=1.64.0
 Requires-Dist: dill==0.3.7
 Requires-Dist: cloudpickle==3.0.0
@@ -25,6 +25,7 @@ Requires-Dist: pydantic!=2.0.*,!=2.1.*,!=2.2.*,!=2.3.*,!=2.4.*,<2.11
 Requires-Dist: fastapi<1,>=0.99.1
 Requires-Dist: starlette-exporter>=0.21.0
 Requires-Dist: httpx>=0.15.4
+Requires-Dist: httpx-sse
 Requires-Dist: attrs>=21.3.0
 Requires-Dist: python-dateutil<3,>=2.8.0
 Requires-Dist: types-python-dateutil<3,>=2.8.0

{fal-1.41.1 → fal-1.43.0/fal.egg-info}/PKG-INFO

@@ -1,12 +1,12 @@
 Metadata-Version: 2.4
 Name: fal
-Version: 1.41.1
+Version: 1.43.0
 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]<0.21.0,>=0.18.0
-Requires-Dist: isolate-proto<0.18.0,>=0.17.0
+Requires-Dist: isolate-proto<0.19.0,>=0.18.0
 Requires-Dist: grpcio<2,>=1.64.0
 Requires-Dist: dill==0.3.7
 Requires-Dist: cloudpickle==3.0.0
@@ -25,6 +25,7 @@ Requires-Dist: pydantic!=2.0.*,!=2.1.*,!=2.2.*,!=2.3.*,!=2.4.*,<2.11
 Requires-Dist: fastapi<1,>=0.99.1
 Requires-Dist: starlette-exporter>=0.21.0
 Requires-Dist: httpx>=0.15.4
+Requires-Dist: httpx-sse
 Requires-Dist: attrs>=21.3.0
 Requires-Dist: python-dateutil<3,>=2.8.0
 Requires-Dist: types-python-dateutil<3,>=2.8.0

{fal-1.41.1 → fal-1.43.0}/fal.egg-info/requires.txt

@@ -1,5 +1,5 @@
 isolate[build]<0.21.0,>=0.18.0
-isolate-proto<0.18.0,>=0.17.0
+isolate-proto<0.19.0,>=0.18.0
 grpcio<2,>=1.64.0
 dill==0.3.7
 cloudpickle==3.0.0
@@ -18,6 +18,7 @@ pydantic!=2.0.*,!=2.1.*,!=2.2.*,!=2.3.*,!=2.4.*,<2.11
 fastapi<1,>=0.99.1
 starlette-exporter>=0.21.0
 httpx>=0.15.4
+httpx-sse
 attrs>=21.3.0
 python-dateutil<3,>=2.8.0
 types-python-dateutil<3,>=2.8.0

{fal-1.41.1 → fal-1.43.0}/pyproject.toml

@@ -23,7 +23,7 @@ readme = "README.md"
 requires-python = ">=3.8"
 dependencies = [
     "isolate[build]>=0.18.0,<0.21.0",
-    "isolate-proto>=0.17.0,<0.18.0",
+    "isolate-proto>=0.18.0,<0.19.0",
     "grpcio>=1.64.0,<2",
     "dill==0.3.7",
     "cloudpickle==3.0.0",
@@ -44,6 +44,7 @@ dependencies = [
     "starlette-exporter>=0.21.0",
     # rest-api-client dependencies
     "httpx>=0.15.4",
+    "httpx-sse",
     "attrs>=21.3.0",
     "python-dateutil>=2.8.0,<3",
     "types-python-dateutil>=2.8.0,<3",

{fal-1.41.1 → fal-1.43.0}/src/fal/_fal_version.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '1.41.1'
-__version_tuple__ = version_tuple = (1, 41, 1)
+__version__ = version = '1.43.0'
+__version_tuple__ = version_tuple = (1, 43, 0)
 
 __commit_id__ = commit_id = None

{fal-1.41.1 → fal-1.43.0}/src/fal/api.py

@@ -21,6 +21,7 @@ from typing import (
     Iterator,
     Literal,
     NamedTuple,
+    Optional,
     TypeVar,
     cast,
     overload,
@@ -53,10 +54,13 @@ from fal.exceptions._cuda import _is_cuda_oom_exception
 from fal.logging.isolate import IsolateLogPrinter
 from fal.sdk import (
     FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
+    FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER_PERC,
     FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
+    AuthModeLiteral,
     Credentials,
+    DeploymentStrategyLiteral,
     FalServerlessClient,
     FalServerlessConnection,
     HostedRunState,
@@ -424,6 +428,7 @@ class FalServerlessHost(Host):
             "max_concurrency",
             "min_concurrency",
             "concurrency_buffer",
+            "concurrency_buffer_perc",
             "max_multiplexing",
             "setup_function",
             "metadata",
@@ -465,12 +470,13 @@ class FalServerlessHost(Host):
         self,
         func: Callable[ArgsT, ReturnT],
         options: Options,
-        application_name: str | None = None,
-        application_auth_mode: Literal["public", "shared", "private"] | None = None,
-        metadata: dict[str, Any] | None = None,
-        deployment_strategy: Literal["recreate", "rolling"] = "recreate",
+        *,
+        application_name: Optional[str] = None,
+        application_auth_mode: Optional[AuthModeLiteral] = None,
+        metadata: Optional[dict[str, Any]] = None,
+        deployment_strategy: DeploymentStrategyLiteral,
         scale: bool = True,
-    ) -> str | None:
+    ) -> Optional[str]:
         from isolate.backends.common import active_python
 
         environment_options = options.environment.copy()
@@ -487,6 +493,7 @@ class FalServerlessHost(Host):
         max_concurrency = options.host.get("max_concurrency")
         min_concurrency = options.host.get("min_concurrency")
         concurrency_buffer = options.host.get("concurrency_buffer")
+        concurrency_buffer_perc = options.host.get("concurrency_buffer_perc")
         max_multiplexing = options.host.get("max_multiplexing")
         exposed_port = options.get_exposed_port()
         request_timeout = options.host.get("request_timeout")
@@ -503,6 +510,7 @@ class FalServerlessHost(Host):
             max_concurrency=max_concurrency,
             min_concurrency=min_concurrency,
             concurrency_buffer=concurrency_buffer,
+            concurrency_buffer_perc=concurrency_buffer_perc,
             request_timeout=request_timeout,
             startup_timeout=startup_timeout,
         )
@@ -560,6 +568,7 @@ class FalServerlessHost(Host):
         max_concurrency = options.host.get("max_concurrency")
         min_concurrency = options.host.get("min_concurrency")
         concurrency_buffer = options.host.get("concurrency_buffer")
+        concurrency_buffer_perc = options.host.get("concurrency_buffer_perc")
         max_multiplexing = options.host.get("max_multiplexing")
         base_image = options.host.get("_base_image", None)
         scheduler = options.host.get("_scheduler", None)
@@ -580,6 +589,7 @@ class FalServerlessHost(Host):
             max_concurrency=max_concurrency,
             min_concurrency=min_concurrency,
             concurrency_buffer=concurrency_buffer,
+            concurrency_buffer_perc=concurrency_buffer_perc,
             request_timeout=request_timeout,
             startup_timeout=startup_timeout,
         )
@@ -770,6 +780,7 @@ def function(
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
+    concurrency_buffer_perc: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER_PERC,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -800,6 +811,7 @@ def function(
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
+    concurrency_buffer_perc: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER_PERC,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -882,6 +894,7 @@ def function(
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
+    concurrency_buffer_perc: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER_PERC,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -917,6 +930,7 @@ def function(
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
+    concurrency_buffer_perc: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER_PERC,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -946,6 +960,7 @@ def function(
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
+    concurrency_buffer_perc: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER_PERC,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -975,6 +990,7 @@ def function(
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
     concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
+    concurrency_buffer_perc: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER_PERC,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,

{fal-1.41.1 → fal-1.43.0}/src/fal/app.py

@@ -11,7 +11,7 @@ import time
 import typing
 from contextlib import asynccontextmanager, contextmanager
 from dataclasses import dataclass
-from typing import Any, Callable, ClassVar, Literal, TypeVar
+from typing import Any, Callable, ClassVar, Optional, TypeVar
 
 import fastapi
 import grpc.aio as async_grpc
@@ -29,6 +29,7 @@ from fal.api import (
 )
 from fal.exceptions import FalServerlessException, RequestCancelledException
 from fal.logging import get_logger
+from fal.sdk import AuthModeLiteral
 from fal.toolkit.file import request_lifecycle_preference
 from fal.toolkit.file.providers.fal import LIFECYCLE_PREFERENCE
 
@@ -311,7 +312,7 @@ class App(BaseServable):
         "keep_alive": 60,
     }
     app_name: ClassVar[str]
-    app_auth: ClassVar[Literal["private", "public", "shared", None]] = None
+    app_auth: ClassVar[Optional[AuthModeLiteral]] = None
     request_timeout: ClassVar[int | None] = None
     startup_timeout: ClassVar[int | None] = None
 

{fal-1.41.1 → fal-1.43.0}/src/fal/cli/_utils.py

@@ -1,10 +1,13 @@
 from __future__ import annotations
 
+from typing import Any, Optional
+
 from fal.project import find_project_root, find_pyproject_toml, parse_pyproject_toml
+from fal.sdk import AuthModeLiteral, DeploymentStrategyLiteral
 
 
 def get_client(host: str, team: str | None = None):
-    from fal.sdk import FalServerlessClient, get_default_credentials
+    from fal.sdk import FalServerlessClient, get_default_credentials  # noqa: PLC0415
 
     credentials = get_default_credentials(team=team)
     return FalServerlessClient(host, credentials)
@@ -17,7 +20,9 @@ def is_app_name(app_ref: tuple[str, str | None]) -> bool:
     return is_single_file and not is_python_file
 
 
-def get_app_data_from_toml(app_name):
+def get_app_data_from_toml(
+    app_name,
+) -> tuple[str, Optional[AuthModeLiteral], Optional[DeploymentStrategyLiteral], bool]:
     toml_path = find_pyproject_toml()
 
     if toml_path is None:
@@ -27,12 +32,12 @@ def get_app_data_from_toml(app_name):
     apps = fal_data.get("apps", {})
 
     try:
-        app_data = apps[app_name]
+        app_data: dict[str, Any] = apps[app_name]
     except KeyError:
         raise ValueError(f"App {app_name} not found in pyproject.toml")
 
     try:
-        app_ref = app_data.pop("ref")
+        app_ref: str = app_data.pop("ref")
     except KeyError:
         raise ValueError(f"App {app_name} does not have a ref key in pyproject.toml")
 
@@ -40,12 +45,15 @@ def get_app_data_from_toml(app_name):
     project_root, _ = find_project_root(None)
     app_ref = str(project_root / app_ref)
 
-    app_auth = app_data.pop("auth", "private")
-    app_deployment_strategy = app_data.pop("deployment_strategy", "recreate")
+    app_auth: Optional[AuthModeLiteral] = app_data.pop("auth", None)
+    app_deployment_strategy: Optional[DeploymentStrategyLiteral] = app_data.pop(
+        "deployment_strategy", None
+    )
 
+    app_reset_scale: bool
     if "no_scale" in app_data:
         # Deprecated
-        app_no_scale = app_data.pop("no_scale")
+        app_no_scale: bool = app_data.pop("no_scale")
         print("[WARNING] no_scale is deprecated, use app_scale_settings instead")
         app_reset_scale = not app_no_scale
     else:

{fal-1.41.1 → fal-1.43.0}/src/fal/cli/apps.py

@@ -33,13 +33,20 @@ def _apps_table(apps: list[AliasInfo]):
     table.add_column("Regions")
 
     for app in apps:
+        if app.concurrency_buffer_perc > 0:
+            concurrency_buffer_str = (
+                f"{app.concurrency_buffer_perc}%, min {app.concurrency_buffer}"
+            )
+        else:
+            concurrency_buffer_str = str(app.concurrency_buffer)
+
         table.add_row(
             app.alias,
             app.revision,
             app.auth_mode,
             str(app.min_concurrency),
             str(app.max_concurrency),
-            str(app.concurrency_buffer),
+            concurrency_buffer_str,
             str(app.max_multiplexing),
             str(app.keep_alive),
             str(app.request_timeout),
@@ -165,6 +172,7 @@ def _scale(args):
             and args.max_concurrency is None
             and args.min_concurrency is None
             and args.concurrency_buffer is None
+            and args.concurrency_buffer_perc is None
             and args.request_timeout is None
             and args.startup_timeout is None
             and args.machine_types is None
@@ -180,6 +188,7 @@ def _scale(args):
             max_concurrency=args.max_concurrency,
             min_concurrency=args.min_concurrency,
             concurrency_buffer=args.concurrency_buffer,
+            concurrency_buffer_perc=args.concurrency_buffer_perc,
             request_timeout=args.request_timeout,
             startup_timeout=args.startup_timeout,
             machine_types=args.machine_types,
@@ -225,7 +234,12 @@ def _add_scale_parser(subparsers, parents):
     parser.add_argument(
         "--concurrency-buffer",
         type=int,
-        help="Concurrency buffer",
+        help="Concurrency buffer (min)",
+    )
+    parser.add_argument(
+        "--concurrency-buffer-perc",
+        type=int,
+        help="Concurrency buffer %",
     )
     parser.add_argument(
         "--request-timeout",

{fal-1.41.1 → fal-1.43.0}/src/fal/cli/deploy.py

@@ -2,7 +2,9 @@ import argparse
 import json
 from collections import namedtuple
 from pathlib import Path
-from typing import Literal, Optional, Tuple, Union
+from typing import Optional, Tuple, Union, cast
+
+from fal.sdk import AuthModeLiteral, DeploymentStrategyLiteral
 
 from ._utils import get_app_data_from_toml, is_app_name
 from .parser import FalClientParser, RefAction, get_output_parser
@@ -67,8 +69,8 @@ def _deploy_from_reference(
     app_ref: Tuple[Optional[Union[Path, str]], ...],
     app_name: str,
     args,
-    auth: Optional[Literal["public", "shared", "private"]],
-    deployment_strategy: Optional[Literal["recreate", "rolling"]],
+    auth: Optional[AuthModeLiteral],
+    deployment_strategy: Optional[DeploymentStrategyLiteral],
     scale: bool,
 ):
     from fal.api import FalServerlessError, FalServerlessHost
@@ -99,7 +101,7 @@ def _deploy_from_reference(
     isolated_function = loaded.function
     app_name = app_name or loaded.app_name  # type: ignore
     app_auth = auth or loaded.app_auth
-    deployment_strategy = deployment_strategy or "recreate"
+    deployment_strategy = deployment_strategy or "rolling"
 
     app_id = host.register(
         func=isolated_function.func,
@@ -172,10 +174,12 @@ def _deploy(args):
     # path/to/myfile.py::MyApp
     else:
         file_path, func_name = args.app_ref
-        app_name = args.app_name
-        app_auth = args.auth
-        app_deployment_strategy = args.strategy
-        app_scale_settings = args.app_scale_settings
+        app_name = cast(str, args.app_name)
+        # default to be set in the backend
+        app_auth = cast(Optional[AuthModeLiteral], args.auth)
+        # default comes from the CLI
+        app_deployment_strategy = cast(DeploymentStrategyLiteral, args.strategy)
+        app_scale_settings = cast(bool, args.app_scale_settings)
 
     _deploy_from_reference(
         (file_path, func_name),
@@ -251,7 +255,7 @@ def add_parser(main_subparsers, parents):
         "--strategy",
         choices=["recreate", "rolling"],
         help="Deployment strategy.",
-        default="recreate",
+        default="rolling",
     )
     parser.add_argument(
         "--no-scale",

{fal-1.41.1 → fal-1.43.0}/src/fal/cli/parser.py

@@ -86,6 +86,37 @@ class SinceAction(argparse.Action):
 
         super().__init__(*args, **kwargs)
 
+        # If a default is provided as a string like "1h ago", parse it into a datetime
+        # so callers can rely on receiving a datetime even when the flag isn't passed.
+        default_value = getattr(self, "default", None)
+        if default_value is not None and default_value is not argparse.SUPPRESS:
+            if isinstance(default_value, str):
+                dt = self._parse_since(default_value)
+                if not dt:
+                    raise ValueError(
+                        f"Invalid 'default' value for SinceAction: {default_value!r}"
+                    )
+                if (
+                    self._limit
+                    and self._limit_dt is not None
+                    and dt < self._limit_dt - self.LIMIT_LEEWAY
+                ):
+                    raise ValueError(
+                        "Default since value is older than the allowed limit "
+                        f"{self._limit}."
+                    )
+                self.default = dt
+            elif isinstance(default_value, datetime):
+                if (
+                    self._limit
+                    and self._limit_dt is not None
+                    and default_value < self._limit_dt - self.LIMIT_LEEWAY
+                ):
+                    raise ValueError(
+                        "Default since value is older than the allowed limit "
+                        f"{self._limit}."
+                    )
+
     def __call__(self, parser, args, values, option_string=None):  # noqa: ARG002
         if values is None:
             setattr(args, self.dest, None)
@@ -102,7 +133,7 @@ class SinceAction(argparse.Action):
                 ),
             )
 
-        if self._limit_dt is not None:
+        if self._limit and self._limit_dt is not None:
             if dt < self._limit_dt - self.LIMIT_LEEWAY:
                 raise argparse.ArgumentError(
                     self,