fal 1.9.4__tar.gz → 1.11.0__tar.gz

{fal-1.9.4/fal.egg-info → fal-1.11.0}/PKG-INFO

@@ -1,12 +1,12 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: fal
-Version: 1.9.4
+Version: 1.11.0
 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]<0.17.0,>=0.16.1
-Requires-Dist: isolate-proto<0.7.0,>=0.6.7
+Requires-Dist: isolate-proto<0.8.0,>=0.7.0
 Requires-Dist: grpcio==1.64.0
 Requires-Dist: dill==0.3.7
 Requires-Dist: cloudpickle==3.0.0

{fal-1.9.4 → fal-1.11.0/fal.egg-info}/PKG-INFO

@@ -1,12 +1,12 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: fal
-Version: 1.9.4
+Version: 1.11.0
 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]<0.17.0,>=0.16.1
-Requires-Dist: isolate-proto<0.7.0,>=0.6.7
+Requires-Dist: isolate-proto<0.8.0,>=0.7.0
 Requires-Dist: grpcio==1.64.0
 Requires-Dist: dill==0.3.7
 Requires-Dist: cloudpickle==3.0.0

{fal-1.9.4 → fal-1.11.0}/fal.egg-info/SOURCES.txt

@@ -99,8 +99,10 @@ src/fal/auth/auth0.py
 src/fal/auth/local.py
 src/fal/cli/__init__.py
 src/fal/cli/_utils.py
+src/fal/cli/api.py
 src/fal/cli/apps.py
 src/fal/cli/auth.py
+src/fal/cli/cli_nested_json.py
 src/fal/cli/create.py
 src/fal/cli/debug.py
 src/fal/cli/deploy.py

{fal-1.9.4 → fal-1.11.0}/fal.egg-info/requires.txt

@@ -1,5 +1,5 @@
 isolate[build]<0.17.0,>=0.16.1
-isolate-proto<0.7.0,>=0.6.7
+isolate-proto<0.8.0,>=0.7.0
 grpcio==1.64.0
 dill==0.3.7
 cloudpickle==3.0.0

{fal-1.9.4 → fal-1.11.0}/pyproject.toml

@@ -23,7 +23,7 @@ readme = "README.md"
 requires-python = ">=3.8"
 dependencies = [
     "isolate[build]>=0.16.1,<0.17.0",
-    "isolate-proto>=0.6.7,<0.7.0",
+    "isolate-proto>=0.7.0,<0.8.0",
     "grpcio==1.64.0",
     "dill==0.3.7",
     "cloudpickle==3.0.0",

{fal-1.9.4 → fal-1.11.0}/src/fal/_fal_version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '1.9.4'
-__version_tuple__ = version_tuple = (1, 9, 4)
+__version__ = version = '1.11.0'
+__version_tuple__ = version_tuple = (1, 11, 0)

{fal-1.9.4 → fal-1.11.0}/src/fal/api.py

@@ -53,6 +53,7 @@ from fal.exceptions import (
 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_KEEP_ALIVE,
     FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
@@ -400,6 +401,7 @@ class FalServerlessHost(Host):
             "keep_alive",
             "max_concurrency",
             "min_concurrency",
+            "concurrency_buffer",
             "max_multiplexing",
             "setup_function",
             "metadata",
@@ -451,6 +453,7 @@ class FalServerlessHost(Host):
         scheduler_options = options.host.get("_scheduler_options", None)
         max_concurrency = options.host.get("max_concurrency")
         min_concurrency = options.host.get("min_concurrency")
+        concurrency_buffer = options.host.get("concurrency_buffer")
         max_multiplexing = options.host.get("max_multiplexing")
         exposed_port = options.get_exposed_port()
         request_timeout = options.host.get("request_timeout")
@@ -466,6 +469,7 @@ class FalServerlessHost(Host):
             max_multiplexing=max_multiplexing,
             max_concurrency=max_concurrency,
             min_concurrency=min_concurrency,
+            concurrency_buffer=concurrency_buffer,
             request_timeout=request_timeout,
             startup_timeout=startup_timeout,
         )
@@ -523,6 +527,7 @@ class FalServerlessHost(Host):
         keep_alive = options.host.get("keep_alive", FAL_SERVERLESS_DEFAULT_KEEP_ALIVE)
         max_concurrency = options.host.get("max_concurrency")
         min_concurrency = options.host.get("min_concurrency")
+        concurrency_buffer = options.host.get("concurrency_buffer")
         max_multiplexing = options.host.get("max_multiplexing")
         base_image = options.host.get("_base_image", None)
         scheduler = options.host.get("_scheduler", None)
@@ -542,6 +547,7 @@ class FalServerlessHost(Host):
             max_multiplexing=max_multiplexing,
             max_concurrency=max_concurrency,
             min_concurrency=min_concurrency,
+            concurrency_buffer=concurrency_buffer,
             request_timeout=request_timeout,
             startup_timeout=startup_timeout,
         )
@@ -709,6 +715,7 @@ def function(
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
+    concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -737,6 +744,7 @@ def function(
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
+    concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -815,6 +823,7 @@ def function(
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
+    concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -848,6 +857,7 @@ def function(
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
+    concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -875,6 +885,7 @@ def function(
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
+    concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,
@@ -902,6 +913,7 @@ def function(
     keep_alive: int = FAL_SERVERLESS_DEFAULT_KEEP_ALIVE,
     max_multiplexing: int = FAL_SERVERLESS_DEFAULT_MAX_MULTIPLEXING,
     min_concurrency: int = FAL_SERVERLESS_DEFAULT_MIN_CONCURRENCY,
+    concurrency_buffer: int = FAL_SERVERLESS_DEFAULT_CONCURRENCY_BUFFER,
     request_timeout: int | None = None,
     startup_timeout: int | None = None,
     setup_function: Callable[..., None] | None = None,

fal-1.11.0/src/fal/cli/api.py

@@ -0,0 +1,72 @@
+import re
+
+import rich
+
+import fal.apps
+
+# = or := only
+KV_SPLIT_RE = re.compile(r"(=|:=)")
+
+
+def _api(args):
+    """Handle the api command execution."""
+    from rich.console import Group
+    from rich.live import Live
+    from rich.panel import Panel
+    from rich.text import Text
+
+    from . import cli_nested_json
+
+    params = [KV_SPLIT_RE.split(param) for param in args.params]
+    params = cli_nested_json.interpret_nested_json(  # type: ignore
+        [(key, value) for key, _, value in params]
+    )
+
+    handle = fal.apps.submit(args.model_id, params)  # type: ignore
+    logs = []  # type: ignore
+
+    with Live(auto_refresh=False) as live:
+        for event in handle.iter_events(logs=True):
+            if isinstance(event, fal.apps.Queued):
+                status = Text(f"⏳ Queued (position: {event.position})", style="yellow")
+            elif isinstance(event, fal.apps.InProgress):
+                status = Text("🔄 In Progress", style="blue")
+                if event.logs:
+                    logs.extend(log.get("message", str(log)) for log in event.logs)
+                    logs = logs[-10:]  # Keep only last 10 logs
+            else:
+                status = Text("✅ Done", style="green")
+
+            status_panel = Panel(status, title="Status")
+            logs_panel = Panel("\n".join(logs), title="Logs")
+
+            live.update(Group(status_panel, logs_panel))
+            live.refresh()
+
+        # Show final result
+        result = handle.get()
+        live.update(rich.pretty.Pretty(result))
+
+
+def add_parser(main_subparsers, parents):
+    """Add the api command to the main parser."""
+    api_help = "Call a fal API endpoint directly"
+    parser = main_subparsers.add_parser(
+        "api",
+        description=api_help,
+        help=api_help,
+        parents=parents,
+    )
+
+    parser.add_argument(
+        "model_id",
+        help="Name of the Model ID to call",
+    )
+
+    parser.add_argument(
+        "params",
+        nargs="*",
+        help="Key-value pairs (e.g. key=value or nested[a][b]=value)",
+    )
+
+    parser.set_defaults(func=_api)

{fal-1.9.4 → fal-1.11.0}/src/fal/cli/apps.py

@@ -17,6 +17,7 @@ def _apps_table(apps: list[AliasInfo]):
     table.add_column("Auth")
     table.add_column("Min Concurrency")
     table.add_column("Max Concurrency")
+    table.add_column("Concurrency Buffer")
     table.add_column("Max Multiplexing")
     table.add_column("Keep Alive")
     table.add_column("Request Timeout")
@@ -32,6 +33,7 @@ def _apps_table(apps: list[AliasInfo]):
             app.auth_mode,
             str(app.min_concurrency),
             str(app.max_concurrency),
+            str(app.concurrency_buffer),
             str(app.max_multiplexing),
             str(app.keep_alive),
             str(app.request_timeout),
@@ -50,6 +52,15 @@ def _list(args):
     client = FalServerlessClient(args.host)
     with client.connect() as connection:
         apps = connection.list_aliases()
+
+        if args.filter:
+            apps = [app for app in apps if args.filter in app.alias]
+
+        if args.sort_by_runners:
+            apps.sort(key=lambda x: x.active_runners)
+        else:
+            apps.sort(key=lambda x: x.alias)
+
         table = _apps_table(apps)
 
     args.console.print(table)
@@ -63,6 +74,16 @@ def _add_list_parser(subparsers, parents):
         help=list_help,
         parents=parents,
     )
+    parser.add_argument(
+        "--sort-by-runners",
+        action="store_true",
+        help="Sort by number of runners ascending",
+    )
+    parser.add_argument(
+        "--filter",
+        type=str,
+        help="Filter applications by alias contents",
+    )
     parser.set_defaults(func=_list)
 
 
@@ -130,6 +151,7 @@ def _scale(args):
             and args.max_multiplexing is None
             and args.max_concurrency is None
             and args.min_concurrency is None
+            and args.concurrency_buffer is None
             and args.request_timeout is None
             and args.startup_timeout is None
             and args.machine_types is None
@@ -144,6 +166,7 @@ def _scale(args):
             max_multiplexing=args.max_multiplexing,
             max_concurrency=args.max_concurrency,
             min_concurrency=args.min_concurrency,
+            concurrency_buffer=args.concurrency_buffer,
             request_timeout=args.request_timeout,
             startup_timeout=args.startup_timeout,
             machine_types=args.machine_types,
@@ -186,6 +209,11 @@ def _add_scale_parser(subparsers, parents):
         type=int,
         help="Minimum concurrency",
     )
+    parser.add_argument(
+        "--concurrency-buffer",
+        type=int,
+        help="Concurrency buffer",
+    )
     parser.add_argument(
         "--request-timeout",
         type=int,
@@ -258,13 +286,23 @@ def _runners(args):
     table = Table()
     table.add_column("Runner ID")
     table.add_column("In Flight Requests")
-    table.add_column("Expires in")
+    table.add_column("Missing Leases")
+    table.add_column("Expires In")
     table.add_column("Uptime")
 
     for runner in runners:
+        num_leases_with_request = len(
+            [
+                lease
+                for lease in runner.external_metadata.get("leases", [])
+                if lease.get("request_id") is not None
+            ]
+        )
+
         table.add_row(
             runner.runner_id,
             str(runner.in_flight_requests),
+            str(runner.in_flight_requests - num_leases_with_request),
             (
                 "N/A (active)"
                 if runner.expiration_countdown is None