modal 1.5.3.dev5__tar.gz → 1.5.3.dev6__tar.gz

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: modal
-Version: 1.5.3.dev5
+Version: 1.5.3.dev6
 Summary: Python client library for Modal
 Author-email: Modal Labs <support@modal.com>
 License-Expression: Apache-2.0

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/modal/_functions.py

@@ -2073,11 +2073,13 @@ class _FunctionCall(typing.Generic[ReturnType], _Object, type_prefix="fc"):
 
     @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.
+        """Fetch information about the graph of Inputs this FunctionCall is part of.
 
-        See [`modal.call_graph`](https://modal.com/docs/sdk/py/latest/modal.call_graph) reference page
-        for documentation on the structure of the returned `InputInfo` items.
+        Note: the call graph data is not populated in real-time, and its capture is best-effort.
+        We do not recommend relying on this method for critical use cases.
+
+        See the [`modal.types`](/docs/sdk/py/latest/modal.types) reference for information
+        on the return values.
 
         Returns:
             A list of `InputInfo` nodes describing the call graph.
@@ -2088,19 +2090,13 @@ class _FunctionCall(typing.Generic[ReturnType], _Object, type_prefix="fc"):
         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`](https://modal.com/docs/sdk/py/latest/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.
+    async def cancel(self, terminate_containers: bool = False):
+        """Cancel the FunctionCall and terminate its inputs without retrying.
 
         Args:
-            terminate_containers: If True, forcibly terminate workers running cancelled inputs.
+            terminate_containers: If True, terminate the containers running the cancelled
+                inputs. Any other inputs running concurrently on those containers will be
+                rescheduled.
         """
         request = api_pb2.FunctionCallCancelRequest(
             function_call_id=self.object_id, terminate_containers=terminate_containers

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/modal/_utils/task_command_router_client.py

@@ -1,6 +1,7 @@
 # Copyright Modal Labs 2025
 import asyncio
 import base64
+import io
 import json
 import socket
 import ssl
@@ -8,16 +9,16 @@ import time
 import typing
 import urllib.parse
 import weakref
-from collections.abc import AsyncGenerator, Callable
+from collections.abc import AsyncGenerator, AsyncIterable, Callable
 from contextlib import suppress
-from typing import TypeVar
+from typing import BinaryIO, TypeVar
 
 import grpclib.client
 from grpclib import GRPCError, Status
 from grpclib.exceptions import StreamTerminatedError
 
 from modal.config import logger
-from modal.exception import ExecTimeoutError, TimeoutError as ModalTimeoutError
+from modal.exception import AuthError, ExecTimeoutError, InternalError, ServiceError, TimeoutError as ModalTimeoutError
 from modal_proto import api_pb2, task_command_router_pb2 as sr_pb2
 from modal_proto.task_command_router_grpc import TaskCommandRouterStub
 
@@ -26,6 +27,17 @@ from .._utils.grpc_utils import ModalChannel, create_channel_config
 from .async_utils import aclosing, retry
 from .grpc_utils import RETRYABLE_GRPC_STATUS_CODES
 
+STREAMING_STDIN_CHUNK_SIZE = 256 * 1024
+
+_STREAMING_STDIN_RESUMABLE_EXCEPTIONS = (
+    AuthError,
+    InternalError,
+    ServiceError,
+    StreamTerminatedError,
+    OSError,
+    asyncio.TimeoutError,
+)
+
 
 @retry(n_attempts=34, base_delay=1, max_delay=10, attempt_timeout=10, total_timeout=310)
 async def _connect_channel(channel: grpclib.client.Channel):
@@ -468,6 +480,118 @@ class TaskCommandRouterClient:
                 lambda: self._call_with_auth_retry(self._stub.TaskExecStdinWrite, request)
             )
 
+    async def exec_stdin_write_stream(
+        self,
+        task_id: str,
+        exec_id: str,
+        source: BinaryIO,
+        chunk_size: int = STREAMING_STDIN_CHUNK_SIZE,
+        max_resume_attempts: int = 9,
+    ) -> int:
+        """Stream `source` into the exec's stdin, with bounded resume on transient failures.
+
+        Streams the full contents of `source` in one client-streaming RPC and
+        closes stdin (EOF) on success. On a resumable error, queries
+        `exec_stdin_status` for the server's canonical offset, seeks `source`
+        to that point, and reopens the stream.
+
+        Args:
+            task_id: The task ID of the task running the exec'd command.
+            exec_id: The execution ID of the command to write to.
+            source: A seekable byte source (file, BytesIO, etc.).
+            chunk_size: Bytes per outbound message.
+            max_resume_attempts: Upper bound on resume retries before giving up.
+
+        Returns:
+            The total number of bytes streamed into the exec's stdin.
+        """
+        offset = 0
+        attempt = 0
+        while True:
+
+            async def _chunks():
+                while True:
+                    chunk = source.read(chunk_size)
+                    if not chunk:
+                        return
+                    yield chunk
+
+            try:
+                source.seek(offset)
+                await self._exec_stdin_write_stream(task_id, exec_id, offset, _chunks())
+                return source.tell()
+            except (*_STREAMING_STDIN_RESUMABLE_EXCEPTIONS, AttributeError) as e:
+                # StreamTerminatedError surfaces as AttributeError in grpclib<=0.4.7;
+                # see call_with_retries_on_transient_errors.
+                if isinstance(e, AttributeError) and "_write_appdata" not in str(e):
+                    raise
+                attempt += 1
+                if attempt > max_resume_attempts:
+                    raise
+                status = await self.exec_stdin_status(task_id=task_id, exec_id=exec_id)
+                if status.closed:
+                    pos = source.tell()
+                    if status.num_bytes_written == pos and source.seek(0, io.SEEK_END) == pos:
+                        logger.debug(f"exec_stdin_write_stream completed but response was lost: {e}")
+                        return pos
+                    raise
+                offset = status.num_bytes_written
+                logger.debug(f"exec_stdin_write_stream resuming from offset {offset} after error: {e}")
+
+    async def _exec_stdin_write_stream(
+        self,
+        task_id: str,
+        exec_id: str,
+        start_offset: int,
+        chunks: AsyncIterable[bytes],
+    ) -> sr_pb2.TaskExecStdinWriteStreamResponse | None:
+        """Single client-streaming attempt: Start, Data chunks, then End (EOF).
+
+        Does not retry; `exec_stdin_write_stream` owns resume.
+        """
+        with grpc_error_converter():
+            stream = self._stub.TaskExecStdinWriteStream.open(metadata=self._get_metadata())
+            async with stream as s:
+                start = sr_pb2.TaskExecStdinWriteStreamRequest(
+                    start=sr_pb2.TaskExecStdinWriteStreamStart(
+                        task_id=task_id,
+                        exec_id=exec_id,
+                        offset=start_offset,
+                    ),
+                )
+                await s.send_message(start)
+                async for data in chunks:
+                    if not data:
+                        continue
+                    await s.send_message(sr_pb2.TaskExecStdinWriteStreamRequest(data=data))
+                # The server closes stdin only on this explicit End message. A
+                # stream that breaks before it leaves stdin open for resume.
+                await s.send_message(
+                    sr_pb2.TaskExecStdinWriteStreamRequest(end=sr_pb2.TaskExecStdinWriteStreamEnd()),
+                    end=True,
+                )
+                return await s.recv_message()
+
+    async def exec_stdin_status(self, task_id: str, exec_id: str) -> sr_pb2.TaskExecStdinStatusResponse:
+        """Read the current stdin write status for an exec'd command, to support retries from the right offset.
+
+        Used by streaming clients to find the resume offset after a stream
+        failure. Evicts any in-flight stdin stream for the exec, so it is not
+        safe to call for read-only observability.
+
+        Args:
+            task_id: The task ID of the task running the exec'd command.
+            exec_id: The execution ID of the command to query.
+        Raises:
+            Other errors: If retries are exhausted on transient errors or if there's an error
+              from the RPC itself.
+        """
+        request = sr_pb2.TaskExecStdinStatusRequest(task_id=task_id, exec_id=exec_id)
+        with grpc_error_converter():
+            return await call_with_retries_on_transient_errors(
+                lambda: self._call_with_auth_retry(self._stub.TaskExecStdinStatus, request)
+            )
+
     async def sandbox_stdin_write_v2(
         self, task_id: str, offset: int, data: bytes, eof: bool
     ) -> sr_pb2.SandboxStdinWriteV2Response:

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/modal/client.pyi

@@ -37,7 +37,7 @@ class _Client:
         server_url: str,
         client_type: int,
         credentials: typing.Optional[tuple[str, str]],
-        version: str = "1.5.3.dev5",
+        version: str = "1.5.3.dev6",
     ):
         """mdmd:hidden
         The Modal client object is not intended to be instantiated directly by users.
@@ -207,7 +207,7 @@ class Client:
         server_url: str,
         client_type: int,
         credentials: typing.Optional[tuple[str, str]],
-        version: str = "1.5.3.dev5",
+        version: str = "1.5.3.dev6",
     ):
         """mdmd:hidden
         The Modal client object is not intended to be instantiated directly by users.

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/modal/container_process.py

@@ -1,7 +1,7 @@
 # Copyright Modal Labs 2024
 import asyncio
 import platform
-from typing import Generic, TypeVar
+from typing import BinaryIO, Generic, TypeVar
 
 from modal_proto import api_pb2
 
@@ -117,6 +117,18 @@ class _ContainerProcess(Generic[T]):
             )
         return self._returncode
 
+    async def _stdin_write_stream(self, source: BinaryIO) -> int:
+        """mdmd:hidden
+        Stream `source` into the process's stdin and close it (EOF) on success.
+
+        Returns the total bytes streamed.
+        """
+        return await self._command_router_client.exec_stdin_write_stream(
+            task_id=self._task_id,
+            exec_id=self._process_id,
+            source=source,
+        )
+
     async def poll(self) -> int | None:
         """Check if the container process has finished running.
 

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/modal/container_process.pyi

@@ -53,6 +53,14 @@ class _ContainerProcess(typing.Generic[T]):
 
     @property
     def returncode(self) -> int: ...
+    async def _stdin_write_stream(self, source: typing.BinaryIO) -> int:
+        """mdmd:hidden
+        Stream `source` into the process's stdin and close it (EOF) on success.
+
+        Returns the total bytes streamed.
+        """
+        ...
+
     async def poll(self) -> typing.Optional[int]:
         """Check if the container process has finished running.
 
@@ -105,6 +113,25 @@ class ContainerProcess(typing.Generic[T]):
     @property
     def returncode(self) -> int: ...
 
+    class ___stdin_write_stream_spec(typing_extensions.Protocol):
+        def __call__(self, /, source: typing.BinaryIO) -> int:
+            """mdmd:hidden
+            Stream `source` into the process's stdin and close it (EOF) on success.
+
+            Returns the total bytes streamed.
+            """
+            ...
+
+        async def aio(self, /, source: typing.BinaryIO) -> int:
+            """mdmd:hidden
+            Stream `source` into the process's stdin and close it (EOF) on success.
+
+            Returns the total bytes streamed.
+            """
+            ...
+
+    _stdin_write_stream: ___stdin_write_stream_spec
+
     class __poll_spec(typing_extensions.Protocol):
         def __call__(self, /) -> typing.Optional[int]:
             """Check if the container process has finished running.

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/modal/functions.pyi

@@ -1044,11 +1044,13 @@ class FunctionCall(typing.Generic[modal._functions.ReturnType], modal.object.Obj
 
     class __get_call_graph_spec(typing_extensions.Protocol):
         def __call__(self, /) -> list[modal.types.InputInfo]:
-            """Returns a structure representing the call graph from a given root
-            call ID, along with the status of execution for each node.
+            """Fetch information about the graph of Inputs this FunctionCall is part of.
 
-            See [`modal.call_graph`](https://modal.com/docs/sdk/py/latest/modal.call_graph) reference page
-            for documentation on the structure of the returned `InputInfo` items.
+            Note: the call graph data is not populated in real-time, and its capture is best-effort.
+            We do not recommend relying on this method for critical use cases.
+
+            See the [`modal.types`](/docs/sdk/py/latest/modal.types) reference for information
+            on the return values.
 
             Returns:
                 A list of `InputInfo` nodes describing the call graph.
@@ -1056,11 +1058,13 @@ class FunctionCall(typing.Generic[modal._functions.ReturnType], modal.object.Obj
             ...
 
         async def aio(self, /) -> list[modal.types.InputInfo]:
-            """Returns a structure representing the call graph from a given root
-            call ID, along with the status of execution for each node.
+            """Fetch information about the graph of Inputs this FunctionCall is part of.
+
+            Note: the call graph data is not populated in real-time, and its capture is best-effort.
+            We do not recommend relying on this method for critical use cases.
 
-            See [`modal.call_graph`](https://modal.com/docs/sdk/py/latest/modal.call_graph) reference page
-            for documentation on the structure of the returned `InputInfo` items.
+            See the [`modal.types`](/docs/sdk/py/latest/modal.types) reference for information
+            on the return values.
 
             Returns:
                 A list of `InputInfo` nodes describing the call graph.
@@ -1071,26 +1075,22 @@ class FunctionCall(typing.Generic[modal._functions.ReturnType], modal.object.Obj
 
     class __cancel_spec(typing_extensions.Protocol):
         def __call__(self, /, terminate_containers: bool = False):
-            """Cancels the function call, which will stop its execution and mark its inputs as
-            [`TERMINATED`](https://modal.com/docs/sdk/py/latest/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.
+            """Cancel the FunctionCall and terminate its inputs without retrying.
 
             Args:
-                terminate_containers: If True, forcibly terminate workers running cancelled inputs.
+                terminate_containers: If True, terminate the containers running the cancelled
+                    inputs. Any other inputs running concurrently on those containers will be
+                    rescheduled.
             """
             ...
 
         async def aio(self, /, terminate_containers: bool = False):
-            """Cancels the function call, which will stop its execution and mark its inputs as
-            [`TERMINATED`](https://modal.com/docs/sdk/py/latest/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.
+            """Cancel the FunctionCall and terminate its inputs without retrying.
 
             Args:
-                terminate_containers: If True, forcibly terminate workers running cancelled inputs.
+                terminate_containers: If True, terminate the containers running the cancelled
+                    inputs. Any other inputs running concurrently on those containers will be
+                    rescheduled.
             """
             ...
 

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/modal/sandbox.py

@@ -332,6 +332,7 @@ class _Sandbox(_Object, type_prefix="sb"):
     _attached: bool
     _filesystem: _SandboxFilesystem | None
     _is_v2: bool = False
+    _app_id: str | None
 
     @staticmethod
     def _default_pty_info() -> api_pb2.PTYInfo:
@@ -500,7 +501,7 @@ class _Sandbox(_Object, type_prefix="sb"):
             create_resp = await load_context.client.stub.SandboxCreate(create_req)
             rpc_elapsed = time.monotonic() - rpc_start
             sandbox_id = create_resp.sandbox_id
-            self._hydrate(sandbox_id, load_context.client, None)
+            self._hydrate(sandbox_id, load_context.client, create_resp.metadata)
 
             if logger.isEnabledFor(logging.DEBUG):
                 total_elapsed = time.monotonic() - load_start
@@ -1017,7 +1018,7 @@ class _Sandbox(_Object, type_prefix="sb"):
             )
             rpc_elapsed = time.monotonic() - rpc_start
             sandbox_id = create_resp.sandbox_id
-            self._hydrate(sandbox_id, load_context.client, None)
+            self._hydrate(sandbox_id, load_context.client, create_resp.metadata)
             self._is_v2 = True
             self._task_id = create_resp.task_id
             self._hydrate_metadata_v2()
@@ -1070,7 +1071,22 @@ class _Sandbox(_Object, type_prefix="sb"):
             await resolver.load(obj, load_context)
         return obj
 
-    def _hydrate_metadata(self, handle_metadata: Message | None):
+    def _get_metadata(self) -> api_pb2.SandboxHandleMetadata:
+        metadata = api_pb2.SandboxHandleMetadata(app_id=self._app_id or "")
+        if hasattr(self, "_result") and self._result is not None:
+            metadata.result.CopyFrom(self._result)
+        return metadata
+
+    def _hydrate_metadata(self, handle_metadata: Message | None) -> None:
+        self._app_id = None
+        self._result = None
+        if handle_metadata is not None:
+            assert isinstance(handle_metadata, api_pb2.SandboxHandleMetadata), (
+                f"{type(handle_metadata)} is not SandboxHandleMetadata"
+            )
+            self._app_id = handle_metadata.app_id
+            if handle_metadata.HasField("result"):
+                self._result = handle_metadata.result
         self._stdout = StreamReader(
             _StreamReaderThroughServerParams(
                 file_descriptor=api_pb2.FILE_DESCRIPTOR_STDOUT,
@@ -1088,13 +1104,14 @@ class _Sandbox(_Object, type_prefix="sb"):
             by_line=True,
         )
         self._stdin = StreamWriter(_StreamWriterThroughServerParams(object_id=self.object_id, client=self._client))
-        self._result = None
         self._task_id = None
         self._tunnels = None
         self._enable_snapshot = False
         self._command_router_client = None
         self._filesystem = None
-        self._is_v2 = False
+        self._is_v2 = _is_v2_sandbox_id(self.object_id)
+        if self._is_v2:
+            self._hydrate_metadata_v2()
 
     def _hydrate_metadata_v2(self) -> None:
         """Wire up V2 stdio readers that read directly from the worker. Cheap
@@ -1133,11 +1150,13 @@ class _Sandbox(_Object, type_prefix="sb"):
         super()._initialize_from_other(other)
         self._attached = other._attached
         self._is_v2 = other._is_v2
+        self._app_id = other._app_id
 
     def _initialize_from_empty(self):
         super()._initialize_from_empty()
         self._attached = True
         self._is_v2 = False
+        self._app_id = None
 
     async def detach(self):
         """Disconnects your client from the sandbox and cleans up resources assoicated with the connection.
@@ -1200,7 +1219,7 @@ class _Sandbox(_Object, type_prefix="sb"):
 
         req = api_pb2.SandboxGetFromNameRequest(sandbox_name=name, app_name=app_name, environment_name=env_name)
         resp = await client.stub.SandboxGetFromName(req)
-        return _Sandbox._new_hydrated(resp.sandbox_id, client, None)
+        return _Sandbox._new_hydrated(resp.sandbox_id, client, resp.metadata)
 
     @staticmethod
     async def _experimental_from_name(
@@ -1235,7 +1254,7 @@ class _Sandbox(_Object, type_prefix="sb"):
         auth_token = await client._auth_token_manager.get_token()
         resp = await client.stub.SandboxGetFromNameV2(req, metadata=[("x-modal-auth-token", auth_token)])
 
-        obj = _Sandbox._new_hydrated(resp.sandbox_id, client, None)
+        obj = _Sandbox._new_hydrated(resp.sandbox_id, client, resp.metadata)
         obj._is_v2 = True
         obj._hydrate_metadata_v2()
         return obj
@@ -1266,7 +1285,7 @@ class _Sandbox(_Object, type_prefix="sb"):
         else:
             resp = await client.stub.SandboxWait(req)
 
-        obj = _Sandbox._new_hydrated(sandbox_id, client, None)
+        obj = _Sandbox._new_hydrated(sandbox_id, client, resp.metadata)
         obj._is_v2 = is_v2
         if is_v2:
             obj._hydrate_metadata_v2()
@@ -2360,7 +2379,7 @@ class _Sandbox(_Object, type_prefix="sb"):
 
             for sandbox_info in resp.sandboxes:
                 sandbox_info: api_pb2.SandboxInfo
-                obj = _Sandbox._new_hydrated(sandbox_info.id, client, None)
+                obj = _Sandbox._new_hydrated(sandbox_info.id, client, sandbox_info.metadata)
                 obj._result = sandbox_info.task_info.result  # TODO: send SandboxInfo as metadata to _new_hydrated?
                 yield obj
 
@@ -2415,7 +2434,7 @@ class _Sandbox(_Object, type_prefix="sb"):
 
             for sandbox_info in resp.sandboxes:
                 sandbox_info: api_pb2.SandboxInfo
-                obj = _Sandbox._new_hydrated(sandbox_info.id, client, None)
+                obj = _Sandbox._new_hydrated(sandbox_info.id, client, sandbox_info.metadata)
                 # SandboxListV2 only returns V2 sandboxes; mark them as such so
                 # operations like wait/terminate/exec use the V2 RPCs and stdio.
                 obj._is_v2 = True

{modal-1.5.3.dev5 → modal-1.5.3.dev6}/modal/sandbox.pyi

@@ -150,6 +150,7 @@ class _Sandbox(modal._object._Object):
     _attached: bool
     _filesystem: typing.Optional[modal.sandbox_fs._SandboxFilesystem]
     _is_v2: bool
+    _app_id: typing.Optional[str]
 
     @staticmethod
     def _default_pty_info() -> modal_proto.api_pb2.PTYInfo: ...
@@ -424,7 +425,8 @@ class _Sandbox(modal._object._Object):
         """
         ...
 
-    def _hydrate_metadata(self, handle_metadata: typing.Optional[google.protobuf.message.Message]): ...
+    def _get_metadata(self) -> modal_proto.api_pb2.SandboxHandleMetadata: ...
+    def _hydrate_metadata(self, handle_metadata: typing.Optional[google.protobuf.message.Message]) -> None: ...
     def _hydrate_metadata_v2(self) -> None:
         """Wire up V2 stdio readers that read directly from the worker. Cheap
         to call eagerly: the router connection is opened lazily on first read.
@@ -1342,6 +1344,7 @@ class Sandbox(modal.object.Object):
     _attached: bool
     _filesystem: typing.Optional[modal.sandbox_fs.SandboxFilesystem]
     _is_v2: bool
+    _app_id: typing.Optional[str]
 
     def __init__(self, *args, **kwargs):
         """mdmd:hidden"""
@@ -1872,7 +1875,8 @@ class Sandbox(modal.object.Object):
 
     _experimental_create: typing.ClassVar[___experimental_create_spec]
 
-    def _hydrate_metadata(self, handle_metadata: typing.Optional[google.protobuf.message.Message]): ...
+    def _get_metadata(self) -> modal_proto.api_pb2.SandboxHandleMetadata: ...
+    def _hydrate_metadata(self, handle_metadata: typing.Optional[google.protobuf.message.Message]) -> None: ...
     def _hydrate_metadata_v2(self) -> None:
         """Wire up V2 stdio readers that read directly from the worker. Cheap
         to call eagerly: the router connection is opened lazily on first read.