Merge remote-tracking branch 'origin/main' into freider/cbor2-function-calling

Author: freider

CHANGELOG.md

@@ -6,20 +6,14 @@ This changelog documents user-facing updates (features, enhancements, fixes, and
 
 <!-- NEW CONTENT GENERATED BELOW. PLEASE PRESERVE THIS COMMENT. -->
 
-#### 1.1.4.dev11 (2025-08-22)
-
-Forbid the use of `encrypted_ports`, `h2_ports`, and `unencrypted_ports` in Sandbox creation when `block_network` is `True`.
-
-
-#### 1.1.4.dev7 (2025-08-22)
-
-The type returned by `modal.experimental.get_cluster_info()` now also includes the cluster ID - shared across the set of tasks that spin up in tandem when using the `@clustered` decorator.
-
-
-#### 1.1.4.dev5 (2025-08-21)
-
-- Added an `idle_timeout` param to `Sandbox.create()` which, when provided, will have the sandbox terminate after `idle_timeout` seconds of idleness.
+### 1.1.4 (2025-09-03)
 
+- Added a `startup_timeout` parameter to the `@app.function()` and `@app.cls()` decorators. When used, this configures the timeout applied to each container's startup period separately from the input `timeout`. For backwards compatibility, `timeout` still applies to the startup phase when `startup_timeout` is unset.
+- Added an optional `idle_timeout` parameter to `modal.Sandbox.create()`. When provided, Sandboxes will terminate after `idle_timeout` seconds of idleness.
+- The dataclass returned by `modal.experimental.get_cluster_info()` now includes a `cluster_id` field to identify the clustered set of containers.
+- When `block_network=True` is set in `modal.Sandbox.create()`, we now raise an error if any of `encrypted_ports`, `h2_ports`, or `unencrypted_ports` are also set.
+- Functions decorated with `@modal.asgi_app()` now return an HTTP 408 (request timeout) error code instead of a 502 (gateway timeout) in rare cases when an input fails to arrive at the container, e.g. due to cancellation.
+- `modal.Sandbox.create()` now warns when an invalid `name=` is passed, applying the same rules as other Modal object names: names must be alphanumeric and not longer than 64 characters. This will become an error in the future.
 
 ### 1.1.3 (2025-08-19)
 
@@ -94,7 +88,7 @@ This release also includes a number of other new features and bug fixes:
 - Added a `build_args` parameter to `modal.Image.from_dockerfile` for passing arguments through to `ARG` instructions in the Dockerfile.
 - It's now possible to use `@modal.experimental.clustered` and `i6pn` networking with `modal.Cls`.
 - Fixed a bug where `Cls.with_options` would fail when provided with a `modal.Secret` object that was already hydrated.
-- Fixed a bug where the timeout specified in `modal.Sandbox.exec()` was not respected by `modal.Sandbox.wait()` or `modal.Sandbox.poll()`.
+- Fixed a bug where the timeout specified in `modal.Sandbox.exec()` was not respected by `ContainerProcess.wait()` or `ContainerProcess.poll()`.
 - Fixed retry handling when using `modal run --detach` directly against a remote Function.
 
 Finally, this release introduces a small number of deprecations and potentially-breaking changes:

modal/_functions.py

@@ -674,6 +674,7 @@ def from_local(
         proxy: Optional[_Proxy] = None,
         retries: Optional[Union[int, Retries]] = None,
         timeout: int = 300,
+        startup_timeout: Optional[int] = None,
         min_containers: Optional[int] = None,
         max_containers: Optional[int] = None,
         buffer_containers: Optional[int] = None,
@@ -966,6 +967,7 @@ async def _load(self: _Function, resolver: Resolver, existing_object_id: Optiona
                     proxy_id=(proxy.object_id if proxy else None),
                     retry_policy=retry_policy,
                     timeout_secs=timeout_secs or 0,
+                    startup_timeout_secs=startup_timeout or timeout_secs,
                     pty_info=pty_info,
                     cloud_provider_str=cloud if cloud else "",
                     runtime=config.get("function_runtime"),
@@ -1019,6 +1021,7 @@ async def _load(self: _Function, resolver: Resolver, existing_object_id: Optiona
                         autoscaler_settings=function_definition.autoscaler_settings,
                         worker_id=function_definition.worker_id,
                         timeout_secs=function_definition.timeout_secs,
+                        startup_timeout_secs=function_definition.startup_timeout_secs,
                         web_url=function_definition.web_url,
                         web_url_info=function_definition.web_url_info,
                         webhook_config=function_definition.webhook_config,
@@ -1471,6 +1474,7 @@ def _initialize_from_empty(self):
         self._info = None
         self._serve_mounts = frozenset()
         self._metadata = None
+        self._experimental_flash_urls = None
 
     def _hydrate_metadata(self, metadata: Optional[Message]):
         # Overridden concrete implementation of base class method
@@ -1498,6 +1502,7 @@ def _hydrate_metadata(self, metadata: Optional[Message]):
         self._max_object_size_bytes = (
             metadata.max_object_size_bytes if metadata.HasField("max_object_size_bytes") else MAX_OBJECT_SIZE_BYTES
         )
+        self._experimental_flash_urls = metadata._experimental_flash_urls
 
     def _get_metadata(self):
         # Overridden concrete implementation of base class method
@@ -1515,6 +1520,7 @@ def _get_metadata(self):
             input_plane_url=self._input_plane_url,
             input_plane_region=self._input_plane_region,
             max_object_size_bytes=self._max_object_size_bytes,
+            _experimental_flash_urls=self._experimental_flash_urls,
         )
 
     def _check_no_web_url(self, fn_name: str):
@@ -1545,6 +1551,11 @@ async def get_web_url(self) -> Optional[str]:
         """URL of a Function running as a web endpoint."""
         return self._web_url
 
+    @live_method
+    async def _experimental_get_flash_urls(self) -> Optional[list[str]]:
+        """URL of the flash service for the function."""
+        return list(self._experimental_flash_urls) if self._experimental_flash_urls else None
+
     @property
     async def is_generator(self) -> bool:
         """mdmd:hidden"""

modal/_runtime/asgi.py

@@ -120,7 +120,7 @@ async def disconnect_app():
 
         async def handle_first_input_timeout():
             if scope["type"] == "http":
-                await messages_from_app.put({"type": "http.response.start", "status": 502})
+                await messages_from_app.put({"type": "http.response.start", "status": 408})
                 await messages_from_app.put(
                     {
                         "type": "http.response.body",

modal/app.py

@@ -641,7 +641,8 @@ def function(
         scaledown_window: Optional[int] = None,  # Max time (in seconds) a container can remain idle while scaling down.
         proxy: Optional[_Proxy] = None,  # Reference to a Modal Proxy to use in front of this function.
         retries: Optional[Union[int, Retries]] = None,  # Number of times to retry each input in case of failure.
-        timeout: int = 300,  # Maximum execution time in seconds.
+        timeout: int = 300,  # Maximum execution time for inputs and startup time in seconds.
+        startup_timeout: Optional[int] = None,  # Maximum startup time in seconds with higher precedence than `timeout`.
         name: Optional[str] = None,  # Sets the Modal name of the function within the app
         is_generator: Optional[
             bool
@@ -816,6 +817,7 @@ def f(self, x):
                 batch_max_size=batch_max_size,
                 batch_wait_ms=batch_wait_ms,
                 timeout=timeout,
+                startup_timeout=startup_timeout or timeout,
                 cloud=cloud,
                 webhook_config=webhook_config,
                 enable_memory_snapshot=enable_memory_snapshot,
@@ -869,7 +871,8 @@ def cls(
         scaledown_window: Optional[int] = None,  # Max time (in seconds) a container can remain idle while scaling down.
         proxy: Optional[_Proxy] = None,  # Reference to a Modal Proxy to use in front of this function.
         retries: Optional[Union[int, Retries]] = None,  # Number of times to retry each input in case of failure.
-        timeout: int = 300,  # Maximum execution time in seconds; applies independently to startup and each input.
+        timeout: int = 300,  # Maximum execution time for inputs and startup time in seconds.
+        startup_timeout: Optional[int] = None,  # Maximum startup time in seconds with higher precedence than `timeout`.
         cloud: Optional[str] = None,  # Cloud provider to run the function on. Possible values are aws, gcp, oci, auto.
         region: Optional[Union[str, Sequence[str]]] = None,  # Region or regions to run the function on.
         enable_memory_snapshot: bool = False,  # Enable memory checkpointing for faster cold starts.
@@ -1002,6 +1005,7 @@ def wrapper(wrapped_cls: Union[CLS_T, _PartialFunction]) -> CLS_T:
                 batch_max_size=batch_max_size,
                 batch_wait_ms=batch_wait_ms,
                 timeout=timeout,
+                startup_timeout=startup_timeout or timeout,
                 cloud=cloud,
                 enable_memory_snapshot=enable_memory_snapshot,
                 block_network=block_network,

modal/cls.py

@@ -12,7 +12,7 @@
 from modal_proto import api_pb2
 
 from ._functions import _Function, _parse_retries
-from ._object import _Object
+from ._object import _Object, live_method
 from ._partial_function import (
     _find_callables_for_obj,
     _find_partial_methods_for_user_cls,
@@ -510,6 +510,11 @@ def _get_method_names(self) -> Collection[str]:
         # returns method names for a *local* class only for now (used by cli)
         return self._method_partials.keys()
 
+    @live_method
+    async def _experimental_get_flash_urls(self) -> Optional[list[str]]:
+        """URL of the flash service for the class."""
+        return await self._get_class_service_function()._experimental_get_flash_urls()
+
     def _hydrate_metadata(self, metadata: Message):
         assert isinstance(metadata, api_pb2.ClassHandleMetadata)
         class_service_function = self._get_class_service_function()

modal/config.py

@@ -66,8 +66,8 @@
 * `traceback` (in the .toml file) / `MODAL_TRACEBACK` (as an env var).
   Defaults to False. Enables printing full tracebacks on unexpected CLI
   errors, which can be useful for debugging client issues.
-* `log_pattern` (in the .toml file) / MODAL_LOG_PATTERN` (as an env var).
-  Defaults to "[modal-client] %(asctime)s %(message)s"
+* `log_pattern` (in the .toml file) / `MODAL_LOG_PATTERN` (as an env var).
+  Defaults to `"[modal-client] %(asctime)s %(message)s"`
   The log formatting pattern that will be used by the modal client itself.
   See https://docs.python.org/3/library/logging.html#logrecord-attributes for available
   log attributes.

modal/container_process.py

@@ -10,6 +10,7 @@
 from ._utils.grpc_utils import retry_transient_errors
 from ._utils.shell_utils import stream_from_stdin, write_to_fd
 from .client import _Client
+from .config import logger
 from .exception import InteractiveTimeoutError, InvalidError
 from .io_streams import _StreamReader, _StreamWriter
 from .stream_type import StreamType
@@ -136,6 +137,7 @@ async def wait(self) -> int:
             self._returncode = await asyncio.wait_for(self._wait_for_completion(), timeout=timeout)
         except (asyncio.TimeoutError, TimeoutError):
             self._returncode = -1
+        logger.debug(f"ContainerProcess {self._process_id} wait completed with returncode {self._returncode}")
         return self._returncode
 
     async def attach(self):

modal/experimental/__init__.py

@@ -311,19 +311,17 @@ async def notebook_base_image(*, python_version: Optional[str] = None, force_bui
 
     commands: list[str] = [
         "apt-get update",
-        "apt-get install -y libpq-dev pkg-config cmake git curl wget unzip zip libsqlite3-dev openssh-server vim",
+        "apt-get install -y "
+        + "libpq-dev pkg-config cmake git curl wget unzip zip libsqlite3-dev openssh-server vim ffmpeg",
         _install_cuda_command(),
         # Install uv since it's faster than pip for installing packages.
         "pip install uv",
         # https://github.com/astral-sh/uv/issues/11480
-        "pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126",
+        "pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu129",
         f"uv pip install --system {shlex.join(sorted(environment_packages))}",
         f"uv pip install --system {shlex.join(sorted(kernelshim_packages))}",
     ]
 
-    # TODO: Also install the CUDA Toolkit, so `nvcc` is available.
-    # https://github.com/charlesfrye/cuda-modal/blob/7fef8db12402986cf42d9c8cca8c63d1da6d7700/cuda/use_cuda.py#L158-L188
-
     def build_dockerfile(version: ImageBuilderVersion) -> DockerfileSpec:
         return DockerfileSpec(
             commands=[