modal 1.0.3.dev10__py3-none-any.whl → 1.2.3.dev7__py3-none-any.whl

modal/_functions.py

@@ -6,21 +6,21 @@ import textwrap
 import time
 import typing
 import warnings
-from collections.abc import AsyncGenerator, Sequence, Sized
+from collections.abc import AsyncGenerator, Collection, Sequence, Sized
 from dataclasses import dataclass
 from pathlib import PurePosixPath
-from typing import TYPE_CHECKING, Any, Callable, Optional, Union
+from typing import TYPE_CHECKING, Any, AsyncIterator, Callable, Optional, Union
 
 import typing_extensions
 from google.protobuf.message import Message
 from grpclib import GRPCError, Status
 from synchronicity.combined_types import MethodWithAio
-from synchronicity.exceptions import UserCodeException
 
 from modal_proto import api_pb2
 from modal_proto.modal_api_grpc import ModalClientModal
 
-from ._object import _get_environment_name, _Object, live_method, live_method_gen
+from ._load_context import LoadContext
+from ._object import _Object, live_method, live_method_gen
 from ._pty import get_pty_info
 from ._resolver import Resolver
 from ._resources import convert_fn_config_to_resources_config
@@ -41,21 +41,20 @@ from ._utils.async_utils import (
     synchronizer,
     warn_if_generator_is_not_consumed,
 )
-from ._utils.deprecation import deprecation_warning
+from ._utils.blob_utils import MAX_OBJECT_SIZE_BYTES
+from ._utils.deprecation import deprecation_warning, warn_if_passing_namespace
 from ._utils.function_utils import (
     ATTEMPT_TIMEOUT_GRACE_PERIOD,
     OUTPUTS_TIMEOUT,
     FunctionCreationStatus,
     FunctionInfo,
-    IncludeSourceMode,
     _create_input,
     _process_result,
     _stream_function_call_data,
     get_function_type,
-    get_include_source_mode,
     is_async,
 )
-from ._utils.grpc_utils import RetryWarningMessage, retry_transient_errors
+from ._utils.grpc_utils import Retry, RetryWarningMessage
 from ._utils.mount_utils import validate_network_file_systems, validate_volumes
 from .call_graph import InputInfo, _reconstruct_call_graph
 from .client import _Client
@@ -63,8 +62,6 @@ from .cloud_bucket_mount import _CloudBucketMount, cloud_bucket_mounts_to_proto
 from .config import config
 from .exception import (
     ExecutionError,
-    FunctionTimeoutError,
-    InternalFailure,
     InvalidError,
     NotFoundError,
     OutputExpiredError,
@@ -75,12 +72,16 @@ from .mount import _get_client_mount, _Mount
 from .network_file_system import _NetworkFileSystem, network_file_system_mount_protos
 from .output import _get_output_manager
 from .parallel_map import (
+    _experimental_spawn_map_async,
+    _experimental_spawn_map_sync,
     _for_each_async,
     _for_each_sync,
     _map_async,
     _map_invocation,
+    _map_invocation_inputplane,
     _map_sync,
     _spawn_map_async,
+    _spawn_map_invocation,
     _spawn_map_sync,
     _starmap_async,
     _starmap_sync,
@@ -94,12 +95,14 @@ from .secret import _Secret
 from .volume import _Volume
 
 if TYPE_CHECKING:
-    import modal._partial_function
     import modal.app
     import modal.cls
-    import modal.partial_function
 
 MAX_INTERNAL_FAILURE_COUNT = 8
+TERMINAL_STATUSES = (
+    api_pb2.GenericResult.GENERIC_STATUS_SUCCESS,
+    api_pb2.GenericResult.GENERIC_STATUS_TERMINATED,
+)
 
 
 @dataclasses.dataclass
@@ -144,7 +147,13 @@ class _Invocation:
         stub = client.stub
 
         function_id = function.object_id
-        item = await _create_input(args, kwargs, stub, method_name=function._use_method_name)
+        item = await _create_input(
+            args,
+            kwargs,
+            stub,
+            function=function,
+            function_call_invocation_type=function_call_invocation_type,
+        )
 
         request = api_pb2.FunctionMapRequest(
             function_id=function_id,
@@ -156,21 +165,22 @@ class _Invocation:
 
         if from_spawn_map:
             request.from_spawn_map = True
-            response = await retry_transient_errors(
-                client.stub.FunctionMap,
+            response = await client.stub.FunctionMap(
                 request,
-                max_retries=None,
-                max_delay=30.0,
-                retry_warning_message=RetryWarningMessage(
-                    message="Warning: `.spawn_map(...)` for function `{self._function_name}` is waiting to create"
-                    "more function calls. This may be due to hitting rate limits or function backlog limits.",
-                    warning_interval=10,
-                    errors_to_warn_for=[Status.RESOURCE_EXHAUSTED],
+                retry=Retry(
+                    max_retries=None,
+                    max_delay=30.0,
+                    warning_message=RetryWarningMessage(
+                        message="Warning: `.spawn_map(...)` for function `{self._function_name}` is waiting to create"
+                        "more function calls. This may be due to hitting rate limits or function backlog limits.",
+                        warning_interval=10,
+                        errors_to_warn_for=[Status.RESOURCE_EXHAUSTED],
+                    ),
+                    additional_status_codes=[Status.RESOURCE_EXHAUSTED],
                 ),
-                additional_status_codes=[Status.RESOURCE_EXHAUSTED],
             )
         else:
-            response = await retry_transient_errors(client.stub.FunctionMap, request)
+            response = await client.stub.FunctionMap(request)
 
         function_call_id = response.function_call_id
         if response.pipelined_inputs:
@@ -190,10 +200,7 @@ class _Invocation:
         request_put = api_pb2.FunctionPutInputsRequest(
             function_id=function_id, inputs=[item], function_call_id=function_call_id
         )
-        inputs_response: api_pb2.FunctionPutInputsResponse = await retry_transient_errors(
-            client.stub.FunctionPutInputs,
-            request_put,
-        )
+        inputs_response: api_pb2.FunctionPutInputsResponse = await client.stub.FunctionPutInputs(request_put)
         processed_inputs = inputs_response.inputs
         if not processed_inputs:
             raise Exception("Could not create function call - the input queue seems to be full")
@@ -210,7 +217,11 @@ class _Invocation:
         return _Invocation(stub, function_call_id, client, retry_context)
 
     async def pop_function_call_outputs(
-        self, timeout: Optional[float], clear_on_success: bool, input_jwts: Optional[list[str]] = None
+        self,
+        index: int = 0,
+        timeout: Optional[float] = None,
+        clear_on_success: bool = False,
+        input_jwts: Optional[list[str]] = None,
     ) -> api_pb2.FunctionGetOutputsResponse:
         t0 = time.time()
         if timeout is None:
@@ -228,11 +239,12 @@ class _Invocation:
                 clear_on_success=clear_on_success,
                 requested_at=time.time(),
                 input_jwts=input_jwts,
+                start_idx=index,
+                end_idx=index,
             )
-            response: api_pb2.FunctionGetOutputsResponse = await retry_transient_errors(
-                self.stub.FunctionGetOutputs,
+            response: api_pb2.FunctionGetOutputsResponse = await self.stub.FunctionGetOutputs(
                 request,
-                attempt_timeout=backend_timeout + ATTEMPT_TIMEOUT_GRACE_PERIOD,
+                retry=Retry(attempt_timeout=backend_timeout + ATTEMPT_TIMEOUT_GRACE_PERIOD),
             )
 
             if len(response.outputs) > 0:
@@ -252,21 +264,19 @@ class _Invocation:
 
         item = api_pb2.FunctionRetryInputsItem(input_jwt=ctx.input_jwt, input=ctx.item.input)
         request = api_pb2.FunctionRetryInputsRequest(function_call_jwt=ctx.function_call_jwt, inputs=[item])
-        await retry_transient_errors(
-            self.stub.FunctionRetryInputs,
-            request,
-        )
+        await self.stub.FunctionRetryInputs(request)
 
-    async def _get_single_output(self, expected_jwt: Optional[str] = None) -> Any:
+    async def _get_single_output(self, expected_jwt: Optional[str] = None) -> api_pb2.FunctionGetOutputsItem:
         # waits indefinitely for a single result for the function, and clear the outputs buffer after
         item: api_pb2.FunctionGetOutputsItem = (
             await self.pop_function_call_outputs(
+                index=0,
                 timeout=None,
                 clear_on_success=True,
                 input_jwts=[expected_jwt] if expected_jwt else None,
             )
         ).outputs[0]
-        return await _process_result(item.result, item.data_format, self.stub, self.client)
+        return item
 
     async def run_function(self) -> Any:
         # Use retry logic only if retry policy is specified and
@@ -278,33 +288,38 @@ class _Invocation:
             or ctx.function_call_invocation_type != api_pb2.FUNCTION_CALL_INVOCATION_TYPE_SYNC
             or not ctx.sync_client_retries_enabled
         ):
-            return await self._get_single_output()
+            item = await self._get_single_output()
+            return await _process_result(item.result, item.data_format, self.stub, self.client)
 
         # User errors including timeouts are managed by the user specified retry policy.
         user_retry_manager = RetryManager(ctx.retry_policy)
 
         while True:
-            try:
-                return await self._get_single_output(ctx.input_jwt)
-            except (UserCodeException, FunctionTimeoutError) as exc:
+            item = await self._get_single_output(ctx.input_jwt)
+            if item.result.status in TERMINAL_STATUSES:
+                return await _process_result(item.result, item.data_format, self.stub, self.client)
+
+            if item.result.status != api_pb2.GenericResult.GENERIC_STATUS_INTERNAL_FAILURE:
+                # non-internal failures get a delay before retrying
                 delay_ms = user_retry_manager.get_delay_ms()
                 if delay_ms is None:
-                    raise exc
+                    # no more retries, this should raise an error when the non-success status is converted
+                    # to an exception:
+                    return await _process_result(item.result, item.data_format, self.stub, self.client)
                 await asyncio.sleep(delay_ms / 1000)
-            except InternalFailure:
-                # For system failures on the server, we retry immediately,
-                # and the failure does not count towards the retry policy.
-                pass
+
             await self._retry_input()
 
-    async def poll_function(self, timeout: Optional[float] = None):
+    async def poll_function(self, timeout: Optional[float] = None, *, index: int = 0):
         """Waits up to timeout for a result from a function.
 
         If timeout is `None`, waits indefinitely. This function is not
         cancellation-safe.
         """
         response: api_pb2.FunctionGetOutputsResponse = await self.pop_function_call_outputs(
-            timeout=timeout, clear_on_success=False
+            index=index,
+            timeout=timeout,
+            clear_on_success=False,
         )
         if len(response.outputs) == 0 and response.num_unfinished_inputs == 0:
             # if no unfinished inputs and no outputs, then function expired
@@ -322,7 +337,7 @@ class _Invocation:
         items_total: Union[int, None] = None
         async with aclosing(
             async_merge(
-                _stream_function_call_data(self.client, self.function_call_id, variant="data_out"),
+                _stream_function_call_data(self.client, None, self.function_call_id, variant="data_out"),
                 callable_to_agen(self.run_function),
             )
         ) as streamer:
@@ -337,11 +352,45 @@ class _Invocation:
                 if items_total is not None and items_received >= items_total:
                     break
 
+    async def enumerate(self, start_index: int, end_index: int):
+        """Iterate over the results of the function call in the range [start_index, end_index)."""
+        limit = 49
+        current_index = start_index
+        while current_index < end_index:
+            # batch_end_indx is inclusive, so we subtract 1 to get the last index in the batch.
+            batch_end_index = min(current_index + limit, end_index) - 1
+            request = api_pb2.FunctionGetOutputsRequest(
+                function_call_id=self.function_call_id,
+                timeout=0,
+                last_entry_id="0-0",
+                clear_on_success=False,
+                requested_at=time.time(),
+                start_idx=current_index,
+                end_idx=batch_end_index,
+            )
+            response: api_pb2.FunctionGetOutputsResponse = await self.stub.FunctionGetOutputs(
+                request, retry=Retry(attempt_timeout=ATTEMPT_TIMEOUT_GRACE_PERIOD)
+            )
+
+            outputs = list(response.outputs)
+            outputs.sort(key=lambda x: x.idx)
+            for output in outputs:
+                if output.idx != current_index:
+                    break
+                result = await _process_result(output.result, output.data_format, self.stub, self.client)
+                yield output.idx, result
+                current_index += 1
+
+            # We're missing current_index, so we need to poll the function for the next result
+            if len(outputs) < (batch_end_index - current_index + 1):
+                result = await self.poll_function(index=current_index)
+                yield current_index, result
+                current_index += 1
+
 
 class _InputPlaneInvocation:
     """Internal client representation of a single-input call to a Modal Function using the input
-    plane server API. As of 4/22/2025, this class is experimental and not used in production.
-    It is OK to make breaking changes to this class."""
+    plane server API."""
 
     stub: ModalClientModal
 
@@ -352,12 +401,16 @@ class _InputPlaneInvocation:
         client: _Client,
         input_item: api_pb2.FunctionPutInputsItem,
         function_id: str,
+        retry_policy: api_pb2.FunctionRetryPolicy,
+        input_plane_region: str,
     ):
         self.stub = stub
         self.client = client  # Used by the deserializer.
         self.attempt_token = attempt_token
         self.input_item = input_item
         self.function_id = function_id
+        self.retry_policy = retry_policy
+        self.input_plane_region = input_plane_region
 
     @staticmethod
     async def create(
@@ -367,25 +420,39 @@ class _InputPlaneInvocation:
         *,
         client: _Client,
         input_plane_url: str,
+        input_plane_region: str,
     ) -> "_InputPlaneInvocation":
         stub = await client.get_stub(input_plane_url)
 
         function_id = function.object_id
-        input_item = await _create_input(args, kwargs, stub, method_name=function._use_method_name)
+        control_plane_stub = client.stub
+        # Note: Blob upload is done on the control plane stub, not the input plane stub!
+        input_item = await _create_input(
+            args,
+            kwargs,
+            control_plane_stub,
+            function=function,
+        )
 
         request = api_pb2.AttemptStartRequest(
             function_id=function_id,
             parent_input_id=current_input_id() or "",
             input=input_item,
         )
-        response = await retry_transient_errors(stub.AttemptStart, request)
+
+        metadata = await client.get_input_plane_metadata(input_plane_region)
+        response = await stub.AttemptStart(request, metadata=metadata)
         attempt_token = response.attempt_token
 
-        return _InputPlaneInvocation(stub, attempt_token, client, input_item, function_id)
+        return _InputPlaneInvocation(
+            stub, attempt_token, client, input_item, function_id, response.retry_policy, input_plane_region
+        )
 
     async def run_function(self) -> Any:
+        # User errors including timeouts are managed by the user-specified retry policy.
+        user_retry_manager = RetryManager(self.retry_policy)
+
         # This will retry when the server returns GENERIC_STATUS_INTERNAL_FAILURE, i.e. lost inputs or worker preemption
-        # TODO(ryan): add logic to retry for user defined retry policy
         internal_failure_count = 0
         while True:
             await_request = api_pb2.AttemptAwaitRequest(
@@ -393,33 +460,85 @@ class _InputPlaneInvocation:
                 timeout_secs=OUTPUTS_TIMEOUT,
                 requested_at=time.time(),
             )
-            await_response: api_pb2.AttemptAwaitResponse = await retry_transient_errors(
-                self.stub.AttemptAwait,
+            metadata = await self.client.get_input_plane_metadata(self.input_plane_region)
+            await_response: api_pb2.AttemptAwaitResponse = await self.stub.AttemptAwait(
                 await_request,
-                attempt_timeout=OUTPUTS_TIMEOUT + ATTEMPT_TIMEOUT_GRACE_PERIOD,
+                retry=Retry(attempt_timeout=OUTPUTS_TIMEOUT + ATTEMPT_TIMEOUT_GRACE_PERIOD),
+                metadata=metadata,
             )
 
-            try:
-                if await_response.HasField("output"):
-                    return await _process_result(
-                        await_response.output.result, await_response.output.data_format, self.stub, self.client
-                    )
-            except InternalFailure as e:
+            # Keep awaiting until we get an output.
+            if not await_response.HasField("output"):
+                continue
+
+            # If we have a final output, return.
+            if await_response.output.result.status in TERMINAL_STATUSES:
+                return await _process_result(
+                    await_response.output.result, await_response.output.data_format, self.client.stub, self.client
+                )
+
+            # We have a failure (internal or application), so see if there are any retries left, and if so, retry.
+            if await_response.output.result.status == api_pb2.GenericResult.GENERIC_STATUS_INTERNAL_FAILURE:
                 internal_failure_count += 1
-                # Limit the number of times we retry
-                if internal_failure_count >= MAX_INTERNAL_FAILURE_COUNT:
-                    raise e
-                # For system failures on the server, we retry immediately,
-                # and the failure does not count towards the retry policy.
-                retry_request = api_pb2.AttemptRetryRequest(
-                    function_id=self.function_id,
-                    parent_input_id=current_input_id() or "",
-                    input=self.input_item,
+                # Limit the number of times we retry internal failures.
+                if internal_failure_count < MAX_INTERNAL_FAILURE_COUNT:
+                    # We immediately retry internal failures and the failure doesn't count towards the retry policy.
+                    self.attempt_token = await self._retry_input(metadata)
+                    continue
+            elif (delay_ms := user_retry_manager.get_delay_ms()) is not None:
+                # We still have user retries left, so sleep and retry.
+                await asyncio.sleep(delay_ms / 1000)
+                self.attempt_token = await self._retry_input(metadata)
+                continue
+
+            # No more retries left.
+            return await _process_result(
+                await_response.output.result, await_response.output.data_format, self.client.stub, self.client
+            )
+
+    async def _retry_input(self, metadata: list[tuple[str, str]]) -> str:
+        retry_request = api_pb2.AttemptRetryRequest(
+            function_id=self.function_id,
+            parent_input_id=current_input_id() or "",
+            input=self.input_item,
+            attempt_token=self.attempt_token,
+        )
+        retry_response = await self.stub.AttemptRetry(retry_request, metadata=metadata)
+        return retry_response.attempt_token
+
+    async def run_generator(self):
+        items_received = 0
+        # populated when self.run_function() completes
+        items_total: Union[int, None] = None
+        async with aclosing(
+            async_merge(
+                _stream_function_call_data(
+                    self.client,
+                    self.stub,
+                    function_call_id=None,
+                    variant="data_out",
                     attempt_token=self.attempt_token,
-                )
-                # TODO(ryan): Add exponential backoff?
-                retry_response = await retry_transient_errors(self.stub.AttemptRetry, retry_request)
-                self.attempt_token = retry_response.attempt_token
+                ),
+                callable_to_agen(self.run_function),
+            )
+        ) as streamer:
+            async for item in streamer:
+                if isinstance(item, api_pb2.GeneratorDone):
+                    items_total = item.items_total
+                else:
+                    yield item
+                    items_received += 1
+                # The comparison avoids infinite loops if a non-deterministic generator is retried
+                # and produces less data in the second run than what was already sent.
+                if items_total is not None and items_received >= items_total:
+                    break
+
+    @staticmethod
+    async def _get_metadata(input_plane_region: str, client: _Client) -> list[tuple[str, str]]:
+        if not input_plane_region:
+            return []
+        token = await client._auth_token_manager.get_token()
+        return [("x-modal-input-plane-region", input_plane_region), ("x-modal-auth-token", token)]
 
 
 # Wrapper type for api_pb2.FunctionStats
@@ -461,7 +580,7 @@ class _FunctionSpec:
 
     image: Optional[_Image]
     mounts: Sequence[_Mount]
-    secrets: Sequence[_Secret]
+    secrets: Collection[_Secret]
     network_file_systems: dict[Union[str, PurePosixPath], _NetworkFileSystem]
     volumes: dict[Union[str, PurePosixPath], Union[_Volume, _CloudBucketMount]]
     # TODO(irfansharif): Somehow assert that it's the first kind, in sandboxes
@@ -474,6 +593,21 @@ class _FunctionSpec:
     proxy: Optional[_Proxy]
 
 
+def _get_supported_input_output_formats(is_web_endpoint: bool, is_generator: bool, restrict_output: bool):
+    if is_web_endpoint:
+        supported_input_formats = [api_pb2.DATA_FORMAT_ASGI]
+        supported_output_formats = [api_pb2.DATA_FORMAT_ASGI, api_pb2.DATA_FORMAT_GENERATOR_DONE]
+    else:
+        supported_input_formats = [api_pb2.DATA_FORMAT_PICKLE, api_pb2.DATA_FORMAT_CBOR]
+        if restrict_output:
+            supported_output_formats = [api_pb2.DATA_FORMAT_CBOR]
+        else:
+            supported_output_formats = [api_pb2.DATA_FORMAT_PICKLE, api_pb2.DATA_FORMAT_CBOR]
+        if is_generator:
+            supported_output_formats.append(api_pb2.DATA_FORMAT_GENERATOR_DONE)
+    return supported_input_formats, supported_output_formats
+
+
 P = typing_extensions.ParamSpec("P")
 ReturnType = typing.TypeVar("ReturnType", covariant=True)
 OriginalReturnType = typing.TypeVar(
@@ -523,9 +657,10 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
     @staticmethod
     def from_local(
         info: FunctionInfo,
-        app,
+        app: Optional["modal.app._App"],  # App here should only be None in case of Image.run_function
         image: _Image,
-        secrets: Sequence[_Secret] = (),
+        env: Optional[dict[str, Optional[str]]] = None,
+        secrets: Optional[Collection[_Secret]] = None,
         schedule: Optional[Schedule] = None,
         is_generator: bool = False,
         gpu: Union[GPU_T, list[GPU_T]] = None,
@@ -537,7 +672,8 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         memory: Optional[Union[int, tuple[int, int]]] = None,
         proxy: Optional[_Proxy] = None,
         retries: Optional[Union[int, Retries]] = None,
-        timeout: Optional[int] = None,
+        timeout: int = 300,
+        startup_timeout: Optional[int] = None,
         min_containers: Optional[int] = None,
         max_containers: Optional[int] = None,
         buffer_containers: Optional[int] = None,
@@ -559,14 +695,16 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         rdma: Optional[bool] = None,
         max_inputs: Optional[int] = None,
         ephemeral_disk: Optional[int] = None,
-        # current default: first-party, future default: main-package
-        include_source: Optional[bool] = None,
+        include_source: bool = True,
         experimental_options: Optional[dict[str, str]] = None,
         _experimental_proxy_ip: Optional[str] = None,
         _experimental_custom_scaling_factor: Optional[float] = None,
-        _experimental_enable_gpu_snapshot: bool = False,
+        restrict_output: bool = False,
     ) -> "_Function":
-        """mdmd:hidden"""
+        """mdmd:hidden
+
+        Note: This is not intended to be public API.
+        """
         # Needed to avoid circular imports
         from ._partial_function import _find_partial_methods_for_user_cls, _PartialFunctionFlags
 
@@ -585,15 +723,10 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
             assert not webhook_config
             assert not schedule
 
-        include_source_mode = get_include_source_mode(include_source)
-        if include_source_mode != IncludeSourceMode.INCLUDE_NOTHING:
-            entrypoint_mounts = info.get_entrypoint_mount()
-        else:
-            entrypoint_mounts = {}
-
+        entrypoint_mount = info.get_entrypoint_mount() if include_source else {}
         all_mounts = [
             _get_client_mount(),
-            *entrypoint_mounts.values(),
+            *entrypoint_mount.values(),
         ]
 
         retry_policy = _parse_retries(
@@ -606,6 +739,13 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
             if is_generator:
                 raise InvalidError("Generator functions do not support retries.")
 
+        if timeout is None:  # type: ignore[unreachable]  # Help users who aren't using type checkers
+            raise InvalidError("The `timeout` parameter cannot be set to None: https://modal.com/docs/guide/timeouts")
+
+        secrets = secrets or []
+        if env:
+            secrets = [*secrets, _Secret.from_dict(env)]
+
         function_spec = _FunctionSpec(
             mounts=all_mounts,
             secrets=secrets,
@@ -621,34 +761,6 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
             proxy=proxy,
         )
 
-        if info.user_cls and not is_auto_snapshot:
-            build_functions = _find_partial_methods_for_user_cls(info.user_cls, _PartialFunctionFlags.BUILD).items()
-            for k, pf in build_functions:
-                build_function = pf.raw_f
-                snapshot_info = FunctionInfo(build_function, user_cls=info.user_cls)
-                snapshot_function = _Function.from_local(
-                    snapshot_info,
-                    app=None,
-                    image=image,
-                    secrets=secrets,
-                    gpu=gpu,
-                    network_file_systems=network_file_systems,
-                    volumes=volumes,
-                    memory=memory,
-                    timeout=pf.params.build_timeout,
-                    cpu=cpu,
-                    ephemeral_disk=ephemeral_disk,
-                    is_builder_function=True,
-                    is_auto_snapshot=True,
-                    scheduler_placement=scheduler_placement,
-                    include_source=include_source,
-                )
-                image = _Image._from_args(
-                    base_images={"base": image},
-                    build_function=snapshot_function,
-                    force_build=image.force_build or bool(pf.params.force_build),
-                )
-
         # Note that we also do these checks in FunctionCreate; could drop them here
         if min_containers is not None and not isinstance(min_containers, int):
             raise InvalidError(f"`min_containers` must be an int, not {type(min_containers).__name__}")
@@ -708,7 +820,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         validated_network_file_systems = validate_network_file_systems(network_file_systems)
 
         # Validate image
-        if image is not None and not isinstance(image, _Image):
+        if image is not None and not isinstance(image, _Image):  # type: ignore[unreachable]
             raise InvalidError(f"Expected modal.Image object. Got {type(image)}.")
 
         method_definitions: Optional[dict[str, api_pb2.MethodDefinition]] = None
@@ -721,17 +833,23 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
             for method_name, partial_function in interface_methods.items():
                 function_type = get_function_type(partial_function.params.is_generator)
                 function_name = f"{info.user_cls.__name__}.{method_name}"
+                is_web_endpoint = partial_function._is_web_endpoint()
                 method_schema = get_callable_schema(
                     partial_function._get_raw_f(),
-                    is_web_endpoint=partial_function._is_web_endpoint(),
+                    is_web_endpoint=is_web_endpoint,
                     ignore_first_argument=True,
                 )
+                method_input_formats, method_output_formats = _get_supported_input_output_formats(
+                    is_web_endpoint, partial_function.params.is_generator or False, restrict_output
+                )
 
                 method_definition = api_pb2.MethodDefinition(
                     webhook_config=partial_function.params.webhook_config,
                     function_type=function_type,
                     function_name=function_name,
                     function_schema=method_schema,
+                    supported_input_formats=method_input_formats,
+                    supported_output_formats=method_output_formats,
                 )
                 method_definitions[method_name] = method_definition
 
@@ -755,29 +873,43 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
 
             return deps
 
-        async def _preload(self: _Function, resolver: Resolver, existing_object_id: Optional[str]):
-            assert resolver.client and resolver.client.stub
+        if info.is_service_class():
+            # classes don't have data formats themselves - input/output formats are set per method above
+            supported_input_formats = []
+            supported_output_formats = []
+        else:
+            is_web_endpoint = webhook_config is not None and webhook_config.type != api_pb2.WEBHOOK_TYPE_UNSPECIFIED
+            supported_input_formats, supported_output_formats = _get_supported_input_output_formats(
+                is_web_endpoint, is_generator, restrict_output
+            )
 
-            assert resolver.app_id
+        async def _preload(
+            self: _Function, resolver: Resolver, load_context: LoadContext, existing_object_id: Optional[str]
+        ):
+            assert load_context.app_id
             req = api_pb2.FunctionPrecreateRequest(
-                app_id=resolver.app_id,
+                app_id=load_context.app_id,
                 function_name=info.function_name,
                 function_type=function_type,
                 existing_function_id=existing_object_id or "",
                 function_schema=get_callable_schema(info.raw_f, is_web_endpoint=bool(webhook_config))
                 if info.raw_f
                 else None,
+                supported_input_formats=supported_input_formats,
+                supported_output_formats=supported_output_formats,
             )
             if method_definitions:
                 for method_name, method_definition in method_definitions.items():
                     req.method_definitions[method_name].CopyFrom(method_definition)
             elif webhook_config:
                 req.webhook_config.CopyFrom(webhook_config)
-            response = await retry_transient_errors(resolver.client.stub.FunctionPrecreate, req)
-            self._hydrate(response.function_id, resolver.client, response.handle_metadata)
 
-        async def _load(self: _Function, resolver: Resolver, existing_object_id: Optional[str]):
-            assert resolver.client and resolver.client.stub
+            response = await load_context.client.stub.FunctionPrecreate(req)
+            self._hydrate(response.function_id, load_context.client, response.handle_metadata)
+
+        async def _load(
+            self: _Function, resolver: Resolver, load_context: LoadContext, existing_object_id: Optional[str]
+        ):
             with FunctionCreationStatus(resolver, tag) as function_creation_status:
                 timeout_secs = timeout
 
@@ -827,6 +959,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                         mount_path=path,
                         volume_id=volume.object_id,
                         allow_background_commits=True,
+                        read_only=volume._read_only,
                     )
                     for path, volume in validated_volumes_no_cloud_buckets
                 ]
@@ -843,6 +976,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                 function_schema = (
                     get_callable_schema(info.raw_f, is_web_endpoint=bool(webhook_config)) if info.raw_f else None
                 )
+
                 # Create function remotely
                 function_definition = api_pb2.Function(
                     module_name=info.module_name or "",
@@ -863,6 +997,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                     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"),
@@ -896,7 +1031,6 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                     _experimental_concurrent_cancellations=True,
                     _experimental_proxy_ip=_experimental_proxy_ip,
                     _experimental_custom_scaling=_experimental_custom_scaling_factor is not None,
-                    _experimental_enable_gpu_snapshot=_experimental_enable_gpu_snapshot,
                     # --- These are deprecated in favor of autoscaler_settings
                     warm_pool_size=min_containers or 0,
                     concurrency_limit=max_containers or 0,
@@ -904,6 +1038,8 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                     task_idle_timeout_secs=scaledown_window or 0,
                     # ---
                     function_schema=function_schema,
+                    supported_input_formats=supported_input_formats,
+                    supported_output_formats=supported_output_formats,
                 )
 
                 if isinstance(gpu, list):
@@ -917,6 +1053,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                         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,
@@ -933,12 +1070,13 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                         _experimental_group_size=function_definition._experimental_group_size,
                         _experimental_buffer_containers=function_definition._experimental_buffer_containers,
                         _experimental_custom_scaling=function_definition._experimental_custom_scaling,
-                        _experimental_enable_gpu_snapshot=_experimental_enable_gpu_snapshot,
                         _experimental_proxy_ip=function_definition._experimental_proxy_ip,
                         snapshot_debug=function_definition.snapshot_debug,
                         runtime_perf_record=function_definition.runtime_perf_record,
                         function_schema=function_schema,
                         untrusted=function_definition.untrusted,
+                        supported_input_formats=supported_input_formats,
+                        supported_output_formats=supported_output_formats,
                     )
 
                     ranked_functions = []
@@ -967,18 +1105,16 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                         ),
                     )
 
-                assert resolver.app_id
+                assert load_context.app_id
                 assert (function_definition is None) != (function_data is None)  # xor
                 request = api_pb2.FunctionCreateRequest(
-                    app_id=resolver.app_id,
+                    app_id=load_context.app_id,
                     function=function_definition,
                     function_data=function_data,
                     existing_function_id=existing_object_id or "",
                 )
                 try:
-                    response: api_pb2.FunctionCreateResponse = await retry_transient_errors(
-                        resolver.client.stub.FunctionCreate, request
-                    )
+                    response: api_pb2.FunctionCreateResponse = await load_context.client.stub.FunctionCreate(request)
                 except GRPCError as exc:
                     if exc.status == Status.INVALID_ARGUMENT:
                         raise InvalidError(exc.message)
@@ -993,10 +1129,14 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
             serve_mounts = {m for m in all_mounts if m.is_local()}
             serve_mounts |= image._serve_mounts
             obj._serve_mounts = frozenset(serve_mounts)
-            self._hydrate(response.function_id, resolver.client, response.handle_metadata)
+            self._hydrate(response.function_id, load_context.client, response.handle_metadata)
 
         rep = f"Function({tag})"
-        obj = _Function._from_loader(_load, rep, preload=_preload, deps=_deps)
+        # Pass a *reference* to the App's LoadContext - this is important since the App is
+        # the only way to infer a LoadContext for an `@app.function`, and the App doesn't
+        # get its client until *after* the Function is created.
+        load_context = app._root_load_context if app else LoadContext.empty()
+        obj = _Function._from_loader(_load, rep, preload=_preload, deps=_deps, load_context_overrides=load_context)
 
         obj._raw_f = info.raw_f
         obj._info = info
@@ -1038,7 +1178,12 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
 
         parent = self
 
-        async def _load(param_bound_func: _Function, resolver: Resolver, existing_object_id: Optional[str]):
+        async def _load(
+            param_bound_func: _Function,
+            resolver: Resolver,
+            load_context: LoadContext,
+            existing_object_id: Optional[str],
+        ):
             if not parent.is_hydrated:
                 # While the base Object.hydrate() method appears to be idempotent, it's not always safe
                 await parent.hydrate()
@@ -1071,7 +1216,6 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                 param_bound_func._hydrate_from_other(parent)
                 return
 
-            environment_name = _get_environment_name(None, resolver)
             assert parent is not None and parent.is_hydrated
 
             if options:
@@ -1080,6 +1224,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                         mount_path=path,
                         volume_id=volume.object_id,
                         allow_background_commits=True,
+                        read_only=volume._read_only,
                     )
                     for path, volume in options.validated_volumes
                 ]
@@ -1088,6 +1233,8 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                     replace_secret_ids=bool(options.secrets),
                     replace_volume_mounts=len(volume_mounts) > 0,
                     volume_mounts=volume_mounts,
+                    cloud_bucket_mounts=cloud_bucket_mounts_to_proto(options.cloud_bucket_mounts),
+                    replace_cloud_bucket_mounts=bool(options.cloud_bucket_mounts),
                     resources=options.resources,
                     retry_policy=options.retry_policy,
                     concurrency_limit=options.max_containers,
@@ -1098,6 +1245,8 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                     target_concurrent_inputs=options.target_concurrent_inputs,
                     batch_max_size=options.batch_max_size,
                     batch_linger_ms=options.batch_wait_ms,
+                    scheduler_placement=options.scheduler_placement,
+                    cloud_provider_str=options.cloud,
                 )
             else:
                 options_pb = None
@@ -1106,19 +1255,30 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                 function_id=parent.object_id,
                 serialized_params=serialized_params,
                 function_options=options_pb,
-                environment_name=environment_name
+                environment_name=load_context.environment_name
                 or "",  # TODO: investigate shouldn't environment name always be specified here?
             )
 
-            response = await retry_transient_errors(parent._client.stub.FunctionBindParams, req)
+            response = await parent._client.stub.FunctionBindParams(req)
             param_bound_func._hydrate(response.bound_function_id, parent._client, response.handle_metadata)
 
         def _deps():
             if options:
-                return [v for _, v in options.validated_volumes] + list(options.secrets)
+                all_deps = (
+                    [v for _, v in options.validated_volumes]
+                    + list(options.secrets)
+                    + [mount.secret for _, mount in options.cloud_bucket_mounts if mount.secret]
+                )
+                return [dep for dep in all_deps if not dep.is_hydrated]
             return []
 
-        fun: _Function = _Function._from_loader(_load, "Function(parametrized)", hydrate_lazily=True, deps=_deps)
+        fun: _Function = _Function._from_loader(
+            _load,
+            "Function(parametrized)",
+            hydrate_lazily=True,
+            deps=_deps,
+            load_context_overrides=self._load_context_overrides,
+        )
 
         fun._info = self._info
         fun._obj = obj
@@ -1169,7 +1329,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
             scaledown_window=scaledown_window,
         )
         request = api_pb2.FunctionUpdateSchedulingParamsRequest(function_id=self.object_id, settings=settings)
-        await retry_transient_errors(self.client.stub.FunctionUpdateSchedulingParams, request)
+        await self.client.stub.FunctionUpdateSchedulingParams(request)
 
         # One idea would be for FunctionUpdateScheduleParams to return the current (coalesced) settings
         # and then we could return them here (would need some ad hoc dataclass, which I don't love)
@@ -1212,34 +1372,47 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         await self.update_autoscaler(min_containers=warm_pool_size)
 
     @classmethod
-    def _from_name(cls, app_name: str, name: str, namespace, environment_name: Optional[str]):
+    def _from_name(
+        cls,
+        app_name: str,
+        name: str,
+        *,
+        load_context_overrides: LoadContext,
+    ):
         # internal function lookup implementation that allows lookup of class "service functions"
         # in addition to non-class functions
-        async def _load_remote(self: _Function, resolver: Resolver, existing_object_id: Optional[str]):
-            assert resolver.client and resolver.client.stub
+        async def _load_remote(
+            self: _Function, resolver: Resolver, load_context: LoadContext, existing_object_id: Optional[str]
+        ):
             request = api_pb2.FunctionGetRequest(
                 app_name=app_name,
                 object_tag=name,
-                namespace=namespace,
-                environment_name=_get_environment_name(environment_name, resolver) or "",
+                environment_name=load_context.environment_name,
             )
             try:
-                response = await retry_transient_errors(resolver.client.stub.FunctionGet, request)
-            except GRPCError as exc:
-                if exc.status == Status.NOT_FOUND:
-                    env_context = f" (in the '{environment_name}' environment)" if environment_name else ""
-                    raise NotFoundError(
-                        f"Lookup failed for Function '{name}' from the '{app_name}' app{env_context}: {exc.message}."
-                    )
-                else:
-                    raise
+                response = await load_context.client.stub.FunctionGet(request)
+            except NotFoundError as exc:
+                # refine the error message
+                env_context = (
+                    f" (in the '{load_context.environment_name}' environment)" if load_context.environment_name else ""
+                )
+                raise NotFoundError(
+                    f"Lookup failed for Function '{name}' from the '{app_name}' app{env_context}: {exc}."
+                ) from None
 
             print_server_warnings(response.server_warnings)
 
-            self._hydrate(response.function_id, resolver.client, response.handle_metadata)
+            self._hydrate(response.function_id, load_context.client, response.handle_metadata)
 
-        rep = f"Function.from_name({app_name}, {name})"
-        return cls._from_loader(_load_remote, rep, is_another_app=True, hydrate_lazily=True)
+        environment_rep = (
+            f", environment_name={load_context_overrides.environment_name!r}"
+            if load_context_overrides._environment_name  # slightly ugly - checking if _environment_name is overridden
+            else ""
+        )
+        rep = f"modal.Function.from_name('{app_name}', '{name}'{environment_rep})"
+        return cls._from_loader(
+            _load_remote, rep, is_another_app=True, hydrate_lazily=True, load_context_overrides=load_context_overrides
+        )
 
     @classmethod
     def from_name(
@@ -1247,14 +1420,15 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         app_name: str,
         name: str,
         *,
-        namespace=api_pb2.DEPLOYMENT_NAMESPACE_WORKSPACE,
+        namespace=None,  # mdmd:line-hidden
         environment_name: Optional[str] = None,
+        client: Optional[_Client] = None,
     ) -> "_Function":
         """Reference a Function from a deployed App by its name.
 
-        In contrast to `modal.Function.lookup`, this is a lazy method
-        that defers hydrating the local object with metadata from
-        Modal servers until the first time it is actually used.
+        This is a lazy method that defers hydrating the local
+        object with metadata from Modal servers until the first
+        time it is actually used.
 
         ```python
         f = modal.Function.from_name("other-app", "function")
@@ -1271,40 +1445,10 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                 f"instance.{method_name}.remote(...)\n",
             )
 
-        return cls._from_name(app_name, name, namespace, environment_name)
-
-    @staticmethod
-    async def lookup(
-        app_name: str,
-        name: str,
-        namespace=api_pb2.DEPLOYMENT_NAMESPACE_WORKSPACE,
-        client: Optional[_Client] = None,
-        environment_name: Optional[str] = None,
-    ) -> "_Function":
-        """mdmd:hidden
-        Lookup a Function from a deployed App by its name.
-
-        DEPRECATED: This method is deprecated in favor of `modal.Function.from_name`.
-
-        In contrast to `modal.Function.from_name`, this is an eager method
-        that will hydrate the local object with metadata from Modal servers.
-
-        ```python notest
-        f = modal.Function.lookup("other-app", "function")
-        ```
-        """
-        deprecation_warning(
-            (2025, 1, 27),
-            "`modal.Function.lookup` is deprecated and will be removed in a future release."
-            " It can be replaced with `modal.Function.from_name`."
-            "\n\nSee https://modal.com/docs/guide/modal-1-0-migration for more information.",
+        warn_if_passing_namespace(namespace, "modal.Function.from_name")
+        return cls._from_name(
+            app_name, name, load_context_overrides=LoadContext(environment_name=environment_name, client=client)
         )
-        obj = _Function.from_name(app_name, name, namespace=namespace, environment_name=environment_name)
-        if client is None:
-            client = await _Client.from_env()
-        resolver = Resolver(client=client)
-        await resolver.load(obj)
-        return obj
 
     @property
     def tag(self) -> str:
@@ -1360,6 +1504,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         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
@@ -1377,6 +1522,17 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         self._method_handle_metadata = dict(metadata.method_handle_metadata)
         self._definition_id = metadata.definition_id
         self._input_plane_url = metadata.input_plane_url
+        self._input_plane_region = metadata.input_plane_region
+        # The server may pass back a larger max object size for some input plane users. This applies to input plane
+        # users only - anyone using the control plane will get the standard limit.
+        # There are some cases like FunctionPrecreate where this value is not set at all. We expect that this field
+        # will eventually be hydrated with the correct value, but just to be defensive, if the field is not set we use
+        # MAX_OBJECT_SIZE_BYTES, otherwise it would get set to 0. Accidentally using 0 would cause us to blob upload
+        # everything, so let's avoid that.
+        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
@@ -1392,6 +1548,11 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
             method_handle_metadata=self._method_handle_metadata,
             function_schema=self._metadata.function_schema if self._metadata else None,
             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,
+            supported_input_formats=self._metadata.supported_input_formats if self._metadata else [],
+            supported_output_formats=self._metadata.supported_output_formats if self._metadata else [],
         )
 
     def _check_no_web_url(self, fn_name: str):
@@ -1422,6 +1583,11 @@ Use the `Function.get_web_url()` method instead.
         """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"""
@@ -1438,7 +1604,11 @@ Use the `Function.get_web_url()` method instead.
 
     @live_method_gen
     async def _map(
-        self, input_queue: _SynchronizedQueue, order_outputs: bool, return_exceptions: bool
+        self,
+        input_queue: _SynchronizedQueue,
+        order_outputs: bool,
+        return_exceptions: bool,
+        wrap_returned_exceptions: bool,
     ) -> AsyncGenerator[Any, None]:
         """mdmd:hidden
 
@@ -1459,19 +1629,51 @@ Use the `Function.get_web_url()` method instead.
         else:
             count_update_callback = None
 
-        async with aclosing(
-            _map_invocation(
-                self,
-                input_queue,
-                self.client,
-                order_outputs,
-                return_exceptions,
-                count_update_callback,
-                api_pb2.FUNCTION_CALL_INVOCATION_TYPE_SYNC,
-            )
-        ) as stream:
-            async for item in stream:
-                yield item
+        if self._input_plane_url:
+            async with aclosing(
+                _map_invocation_inputplane(
+                    self,
+                    input_queue,
+                    self.client,
+                    order_outputs,
+                    return_exceptions,
+                    wrap_returned_exceptions,
+                    count_update_callback,
+                )
+            ) as stream:
+                async for item in stream:
+                    yield item
+        else:
+            async with aclosing(
+                _map_invocation(
+                    self,
+                    input_queue,
+                    self.client,
+                    order_outputs,
+                    return_exceptions,
+                    wrap_returned_exceptions,
+                    count_update_callback,
+                    api_pb2.FUNCTION_CALL_INVOCATION_TYPE_SYNC,
+                )
+            ) as stream:
+                async for item in stream:
+                    yield item
+
+    @live_method
+    async def _spawn_map(self, input_queue: _SynchronizedQueue) -> "_FunctionCall[ReturnType]":
+        self._check_no_web_url("spawn_map")
+        if self._is_generator:
+            raise InvalidError("A generator function cannot be called with `.spawn_map(...)`.")
+
+        assert self._function_name
+        function_call_id, num_inputs = await _spawn_map_invocation(
+            self,
+            input_queue,
+            self.client,
+        )
+        fc: _FunctionCall[ReturnType] = _FunctionCall._new_hydrated(function_call_id, self.client, None)
+        fc._num_inputs = num_inputs  # set the cached value of num_inputs
+        return fc
 
     async def _call_function(self, args, kwargs) -> ReturnType:
         invocation: Union[_Invocation, _InputPlaneInvocation]
@@ -1482,6 +1684,7 @@ Use the `Function.get_web_url()` method instead.
                 kwargs,
                 client=self.client,
                 input_plane_url=self._input_plane_url,
+                input_plane_region=self._input_plane_region,
             )
         else:
             invocation = await _Invocation.create(
@@ -1514,13 +1717,24 @@ Use the `Function.get_web_url()` method instead.
     @live_method_gen
     @synchronizer.no_input_translation
     async def _call_generator(self, args, kwargs):
-        invocation = await _Invocation.create(
-            self,
-            args,
-            kwargs,
-            client=self.client,
-            function_call_invocation_type=api_pb2.FUNCTION_CALL_INVOCATION_TYPE_SYNC_LEGACY,
-        )
+        invocation: Union[_Invocation, _InputPlaneInvocation]
+        if self._input_plane_url:
+            invocation = await _InputPlaneInvocation.create(
+                self,
+                args,
+                kwargs,
+                client=self.client,
+                input_plane_url=self._input_plane_url,
+                input_plane_region=self._input_plane_region,
+            )
+        else:
+            invocation = await _Invocation.create(
+                self,
+                args,
+                kwargs,
+                client=self.client,
+                function_call_invocation_type=api_pb2.FUNCTION_CALL_INVOCATION_TYPE_SYNC_LEGACY,
+            )
         async for res in invocation.run_generator():
             yield res
 
@@ -1584,8 +1798,9 @@ Use the `Function.get_web_url()` method instead.
         # "user code" to run on the synchronicity thread, which seems bad
         if not self._is_local():
             msg = (
-                "The definition for this function is missing here so it is not possible to invoke it locally. "
-                "If this function was retrieved via `Function.lookup` you need to use `.remote()`."
+                "The definition for this Function is missing, so it is not possible to invoke it locally. "
+                "If this function was retrieved via `Function.from_name`, "
+                "you need to use one of the remote invocation methods instead."
             )
             raise ExecutionError(msg)
 
@@ -1662,8 +1877,9 @@ Use the `Function.get_web_url()` method instead.
     async def spawn(self, *args: P.args, **kwargs: P.kwargs) -> "_FunctionCall[ReturnType]":
         """Calls the function with the given arguments, without waiting for the results.
 
-        Returns a [`modal.FunctionCall`](/docs/reference/modal.FunctionCall) object, that can later be polled or
-        waited for using [`.get(timeout=...)`](/docs/reference/modal.FunctionCall#get).
+        Returns a [`modal.FunctionCall`](https://modal.com/docs/reference/modal.FunctionCall) object
+        that can later be polled or waited for using
+        [`.get(timeout=...)`](https://modal.com/docs/reference/modal.FunctionCall#get).
         Conceptually similar to `multiprocessing.pool.apply_async`, or a Future/Promise in other contexts.
         """
         self._check_no_web_url("spawn")
@@ -1685,10 +1901,9 @@ Use the `Function.get_web_url()` method instead.
     @live_method
     async def get_current_stats(self) -> FunctionStats:
         """Return a `FunctionStats` object describing the current function's queue and runner counts."""
-        resp = await retry_transient_errors(
-            self.client.stub.FunctionGetCurrentStats,
+        resp = await self.client.stub.FunctionGetCurrentStats(
             api_pb2.FunctionGetCurrentStatsRequest(function_id=self.object_id),
-            total_timeout=10.0,
+            retry=Retry(total_timeout=10.0),
         )
         return FunctionStats(backlog=resp.backlog, num_total_runners=resp.num_total_tasks)
 
@@ -1706,6 +1921,7 @@ Use the `Function.get_web_url()` method instead.
     starmap = MethodWithAio(_starmap_sync, _starmap_async, synchronizer)
     for_each = MethodWithAio(_for_each_sync, _for_each_async, synchronizer)
     spawn_map = MethodWithAio(_spawn_map_sync, _spawn_map_async, synchronizer)
+    experimental_spawn_map = MethodWithAio(_experimental_spawn_map_sync, _experimental_spawn_map_async, synchronizer)
 
 
 class _FunctionCall(typing.Generic[ReturnType], _Object, type_prefix="fc"):
@@ -1720,12 +1936,25 @@ class _FunctionCall(typing.Generic[ReturnType], _Object, type_prefix="fc"):
     """
 
     _is_generator: bool = False
+    _num_inputs: Optional[int] = None
 
     def _invocation(self):
         return _Invocation(self.client.stub, self.object_id, self.client)
 
-    async def get(self, timeout: Optional[float] = None) -> ReturnType:
-        """Get the result of the function call.
+    @live_method
+    async def num_inputs(self) -> int:
+        """Get the number of inputs in the function call."""
+        if self._num_inputs is None:
+            request = api_pb2.FunctionCallFromIdRequest(function_call_id=self.object_id)
+            resp = await self.client.stub.FunctionCallFromId(request)
+            self._num_inputs = resp.num_inputs  # cached
+        return self._num_inputs
+
+    @live_method
+    async def get(self, timeout: Optional[float] = None, *, index: int = 0) -> ReturnType:
+        """Get the result of the index-th input of the function call.
+        `.spawn()` calls have a single output, so only specifying `index=0` is valid.
+        A non-zero index is useful when your function has multiple outputs, like via `.spawn_map()`.
 
         This function waits indefinitely by default. It takes an optional
         `timeout` argument that specifies the maximum number of seconds to wait,
@@ -1733,27 +1962,59 @@ class _FunctionCall(typing.Generic[ReturnType], _Object, type_prefix="fc"):
 
         The returned coroutine is not cancellation-safe.
         """
-        return await self._invocation().poll_function(timeout=timeout)
+        return await self._invocation().poll_function(timeout=timeout, index=index)
+
+    @live_method_gen
+    async def iter(self, *, start: int = 0, end: Optional[int] = None) -> AsyncIterator[ReturnType]:
+        """Iterate in-order over the results of the function call.
+
+        Optionally, specify a range [start, end) to iterate over.
+
+        Example:
+        ```python
+        @app.function()
+        def my_func(a):
+            return a ** 2
+
+
+        @app.local_entrypoint()
+        def main():
+            fc = my_func.spawn_map([1, 2, 3, 4])
+            assert list(fc.iter()) == [1, 4, 9, 16]
+            assert list(fc.iter(start=1, end=3)) == [4, 9]
+        ```
+
+        If `end` is not provided, it will iterate over all results.
+        """
+        num_inputs = await self.num_inputs()
+        if end is None:
+            end = num_inputs
+        if start < 0 or end > num_inputs:
+            raise ValueError(f"Invalid index range: {start} to {end} for {num_inputs} inputs")
+        async for _, item in self._invocation().enumerate(start_index=start, end_index=end):
+            yield item
 
+    @live_method
     async def get_call_graph(self) -> list[InputInfo]:
         """Returns a structure representing the call graph from a given root
         call ID, along with the status of execution for each node.
 
-        See [`modal.call_graph`](/docs/reference/modal.call_graph) reference page
+        See [`modal.call_graph`](https://modal.com/docs/reference/modal.call_graph) reference page
         for documentation on the structure of the returned `InputInfo` items.
         """
         assert self._client and self._client.stub
         request = api_pb2.FunctionGetCallGraphRequest(function_call_id=self.object_id)
-        response = await retry_transient_errors(self._client.stub.FunctionGetCallGraph, request)
+        response = await self._client.stub.FunctionGetCallGraph(request)
         return _reconstruct_call_graph(response)
 
+    @live_method
     async def cancel(
         self,
         # if true, containers running the inputs are forcibly terminated
         terminate_containers: bool = False,
     ):
         """Cancels the function call, which will stop its execution and mark its inputs as
-        [`TERMINATED`](/docs/reference/modal.call_graph#modalcall_graphinputstatus).
+        [`TERMINATED`](https://modal.com/docs/reference/modal.call_graph#modalcall_graphinputstatus).
 
         If `terminate_containers=True` - the containers running the cancelled inputs are all terminated
         causing any non-cancelled inputs on those containers to be rescheduled in new containers.
@@ -1762,7 +2023,7 @@ class _FunctionCall(typing.Generic[ReturnType], _Object, type_prefix="fc"):
             function_call_id=self.object_id, terminate_containers=terminate_containers
         )
         assert self._client and self._client.stub
-        await retry_transient_errors(self._client.stub.FunctionCallCancel, request)
+        await self._client.stub.FunctionCallCancel(request)
 
     @staticmethod
     async def from_id(function_call_id: str, client: Optional[_Client] = None) -> "_FunctionCall[Any]":
@@ -1784,11 +2045,18 @@ class _FunctionCall(typing.Generic[ReturnType], _Object, type_prefix="fc"):
         if you no longer have access to the original object returned from `Function.spawn`.
 
         """
-        if client is None:
-            client = await _Client.from_env()
 
-        fc: _FunctionCall[Any] = _FunctionCall._new_hydrated(function_call_id, client, None)
-        return fc
+        async def _load(
+            self: _FunctionCall, resolver: Resolver, load_context: LoadContext, existing_object_id: Optional[str]
+        ):
+            # this loader doesn't do anything in practice, but it will get the client from the load_context
+            self._hydrate(function_call_id, load_context.client, None)
+
+        rep = f"FunctionCall.from_id({function_call_id!r})"
+
+        return _FunctionCall._from_loader(
+            _load, rep, hydrate_lazily=True, load_context_overrides=LoadContext(client=client)
+        )
 
     @staticmethod
     async def gather(*function_calls: "_FunctionCall[T]") -> typing.Sequence[T]: