modal 1.0.6.dev9__py3-none-any.whl → 1.0.6.dev15__py3-none-any.whl

modal/_container_entrypoint.py

@@ -15,7 +15,6 @@ if telemetry_socket:
     instrument_imports(telemetry_socket)
 
 import asyncio
-import concurrent.futures
 import inspect
 import queue
 import signal
@@ -49,7 +48,6 @@ from ._runtime.container_io_manager import (
     ContainerIOManager,
     IOContext,
     UserException,
-    _ContainerIOManager,
 )
 
 if TYPE_CHECKING:
@@ -198,21 +196,16 @@ def call_function(
 
                 # Send up to this many outputs at a time.
                 generator_queue: asyncio.Queue[Any] = await container_io_manager._queue_create.aio(1024)
-                generator_output_task = asyncio.create_task(
-                    container_io_manager.generator_output_task.aio(
-                        function_call_ids[0],
-                        io_context.finalized_function.data_format,
-                        generator_queue,
-                    )
-                )
-
-                item_count = 0
-                async for value in res:
-                    await container_io_manager._queue_put.aio(generator_queue, value)
-                    item_count += 1
+                async with container_io_manager.generator_output_sender(
+                    function_call_ids[0],
+                    io_context.finalized_function.data_format,
+                    generator_queue,
+                ):
+                    item_count = 0
+                    async for value in res:
+                        await container_io_manager._queue_put.aio(generator_queue, value)
+                        item_count += 1
 
-                await container_io_manager._queue_put.aio(generator_queue, _ContainerIOManager._GENERATOR_STOP_SENTINEL)
-                await generator_output_task  # Wait to finish sending generator outputs.
                 message = api_pb2.GeneratorDone(items_total=item_count)
                 await container_io_manager.push_outputs.aio(
                     io_context,
@@ -249,20 +242,17 @@ def call_function(
 
                 # Send up to this many outputs at a time.
                 generator_queue: asyncio.Queue[Any] = container_io_manager._queue_create(1024)
-                generator_output_task: concurrent.futures.Future = container_io_manager.generator_output_task(  # type: ignore
+
+                with container_io_manager.generator_output_sender(
                     function_call_ids[0],
                     io_context.finalized_function.data_format,
                     generator_queue,
-                    _future=True,  # type: ignore  # Synchronicity magic to return a future.
-                )
-
-                item_count = 0
-                for value in res:
-                    container_io_manager._queue_put(generator_queue, value)
-                    item_count += 1
+                ):
+                    item_count = 0
+                    for value in res:
+                        container_io_manager._queue_put(generator_queue, value)
+                        item_count += 1
 
-                container_io_manager._queue_put(generator_queue, _ContainerIOManager._GENERATOR_STOP_SENTINEL)
-                generator_output_task.result()  # Wait to finish sending generator outputs.
                 message = api_pb2.GeneratorDone(items_total=item_count)
                 container_io_manager.push_outputs(io_context, started_at, message, api_pb2.DATA_FORMAT_GENERATOR_DONE)
             else:

modal/_functions.py

@@ -40,6 +40,7 @@ from ._utils.async_utils import (
     synchronizer,
     warn_if_generator_is_not_consumed,
 )
+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,
@@ -145,6 +146,7 @@ class _Invocation:
             args,
             kwargs,
             stub,
+            max_object_size_bytes=function._max_object_size_bytes,
             method_name=function._use_method_name,
             function_call_invocation_type=function_call_invocation_type,
         )
@@ -386,7 +388,13 @@ class _InputPlaneInvocation:
         function_id = function.object_id
         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, method_name=function._use_method_name)
+        input_item = await _create_input(
+            args,
+            kwargs,
+            control_plane_stub,
+            max_object_size_bytes=function._max_object_size_bytes,
+            method_name=function._use_method_name,
+        )
 
         request = api_pb2.AttemptStartRequest(
             function_id=function_id,
@@ -443,8 +451,10 @@ class _InputPlaneInvocation:
                         self.attempt_token = retry_response.attempt_token
                         continue
 
+                control_plane_stub = self.client.stub
+                # Note: Blob download is done on the control plane stub, not the input plane stub!
                 return await _process_result(
-                    await_response.output.result, await_response.output.data_format, self.stub, self.client
+                    await_response.output.result, await_response.output.data_format, control_plane_stub, self.client
                 )
 
 
@@ -1414,6 +1424,15 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         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
+        )
 
     def _get_metadata(self):
         # Overridden concrete implementation of base class method
@@ -1430,6 +1449,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
             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,
         )
 
     def _check_no_web_url(self, fn_name: str):

modal/_runtime/container_io_manager.py

@@ -290,7 +290,6 @@ class _ContainerIOManager:
 
     _client: _Client
 
-    _GENERATOR_STOP_SENTINEL: ClassVar[Sentinel] = Sentinel()
     _singleton: ClassVar[Optional["_ContainerIOManager"]] = None
 
     def _init(self, container_args: api_pb2.ContainerArguments, client: _Client):
@@ -508,33 +507,47 @@ class _ContainerIOManager:
         req = api_pb2.FunctionCallPutDataRequest(function_call_id=function_call_id, data_chunks=data_chunks)
         await retry_transient_errors(self._client.stub.FunctionCallPutDataOut, req)
 
-    async def generator_output_task(self, function_call_id: str, data_format: int, message_rx: asyncio.Queue) -> None:
-        """Task that feeds generator outputs into a function call's `data_out` stream."""
-        index = 1
-        received_sentinel = False
-        while not received_sentinel:
-            message = await message_rx.get()
-            if message is self._GENERATOR_STOP_SENTINEL:
-                break
-            # ASGI 'http.response.start' and 'http.response.body' msgs are observed to be separated by 1ms.
-            # If we don't sleep here for 1ms we end up with an extra call to .put_data_out().
-            if index == 1:
-                await asyncio.sleep(0.001)
-            serialized_messages = [serialize_data_format(message, data_format)]
-            total_size = len(serialized_messages[0]) + 512
-            while total_size < 16 * 1024 * 1024:  # 16 MiB, maximum size in a single message
-                try:
-                    message = message_rx.get_nowait()
-                except asyncio.QueueEmpty:
-                    break
-                if message is self._GENERATOR_STOP_SENTINEL:
-                    received_sentinel = True
+    @asynccontextmanager
+    async def generator_output_sender(
+        self, function_call_id: str, data_format: int, message_rx: asyncio.Queue
+    ) -> AsyncGenerator[None, None]:
+        """Runs background task that feeds generator outputs into a function call's `data_out` stream."""
+        GENERATOR_STOP_SENTINEL = Sentinel()
+
+        async def generator_output_task():
+            index = 1
+            received_sentinel = False
+            while not received_sentinel:
+                message = await message_rx.get()
+                if message is GENERATOR_STOP_SENTINEL:
                     break
-                else:
-                    serialized_messages.append(serialize_data_format(message, data_format))
-                    total_size += len(serialized_messages[-1]) + 512  # 512 bytes for estimated framing overhead
-            await self.put_data_out(function_call_id, index, data_format, serialized_messages)
-            index += len(serialized_messages)
+                # ASGI 'http.response.start' and 'http.response.body' msgs are observed to be separated by 1ms.
+                # If we don't sleep here for 1ms we end up with an extra call to .put_data_out().
+                if index == 1:
+                    await asyncio.sleep(0.001)
+                serialized_messages = [serialize_data_format(message, data_format)]
+                total_size = len(serialized_messages[0]) + 512
+                while total_size < 16 * 1024 * 1024:  # 16 MiB, maximum size in a single message
+                    try:
+                        message = message_rx.get_nowait()
+                    except asyncio.QueueEmpty:
+                        break
+                    if message is GENERATOR_STOP_SENTINEL:
+                        received_sentinel = True
+                        break
+                    else:
+                        serialized_messages.append(serialize_data_format(message, data_format))
+                        total_size += len(serialized_messages[-1]) + 512  # 512 bytes for estimated framing overhead
+                await self.put_data_out(function_call_id, index, data_format, serialized_messages)
+                index += len(serialized_messages)
+
+        task = asyncio.create_task(generator_output_task())
+        try:
+            yield
+        finally:
+            # gracefully stop the task after all current inputs have been sent
+            await message_rx.put(GENERATOR_STOP_SENTINEL)
+            await task
 
     async def _queue_create(self, size: int) -> asyncio.Queue:
         """Create a queue, on the synchronicity event loop (needed on Python 3.8 and 3.9)."""

modal/_runtime/container_io_manager.pyi

@@ -106,7 +106,6 @@ class _ContainerIOManager:
     _is_interactivity_enabled: bool
     _fetching_inputs: bool
     _client: modal.client._Client
-    _GENERATOR_STOP_SENTINEL: typing.ClassVar[Sentinel]
     _singleton: typing.ClassVar[typing.Optional[_ContainerIOManager]]
 
     def _init(self, container_args: modal_proto.api_pb2.ContainerArguments, client: modal.client._Client): ...
@@ -148,10 +147,10 @@ class _ContainerIOManager:
         """
         ...
 
-    async def generator_output_task(
+    def generator_output_sender(
         self, function_call_id: str, data_format: int, message_rx: asyncio.queues.Queue
-    ) -> None:
-        """Task that feeds generator outputs into a function call's `data_out` stream."""
+    ) -> typing.AsyncContextManager[None]:
+        """Runs background task that feeds generator outputs into a function call's `data_out` stream."""
         ...
 
     async def _queue_create(self, size: int) -> asyncio.queues.Queue:
@@ -268,7 +267,6 @@ class ContainerIOManager:
     _is_interactivity_enabled: bool
     _fetching_inputs: bool
     _client: modal.client.Client
-    _GENERATOR_STOP_SENTINEL: typing.ClassVar[Sentinel]
     _singleton: typing.ClassVar[typing.Optional[ContainerIOManager]]
 
     def __init__(self, /, *args, **kwargs):
@@ -367,16 +365,20 @@ class ContainerIOManager:
 
     put_data_out: __put_data_out_spec[typing_extensions.Self]
 
-    class __generator_output_task_spec(typing_extensions.Protocol[SUPERSELF]):
-        def __call__(self, /, function_call_id: str, data_format: int, message_rx: asyncio.queues.Queue) -> None:
-            """Task that feeds generator outputs into a function call's `data_out` stream."""
+    class __generator_output_sender_spec(typing_extensions.Protocol[SUPERSELF]):
+        def __call__(
+            self, /, function_call_id: str, data_format: int, message_rx: asyncio.queues.Queue
+        ) -> synchronicity.combined_types.AsyncAndBlockingContextManager[None]:
+            """Runs background task that feeds generator outputs into a function call's `data_out` stream."""
             ...
 
-        async def aio(self, /, function_call_id: str, data_format: int, message_rx: asyncio.queues.Queue) -> None:
-            """Task that feeds generator outputs into a function call's `data_out` stream."""
+        def aio(
+            self, /, function_call_id: str, data_format: int, message_rx: asyncio.queues.Queue
+        ) -> typing.AsyncContextManager[None]:
+            """Runs background task that feeds generator outputs into a function call's `data_out` stream."""
             ...
 
-    generator_output_task: __generator_output_task_spec[typing_extensions.Self]
+    generator_output_sender: __generator_output_sender_spec[typing_extensions.Self]
 
     class ___queue_create_spec(typing_extensions.Protocol[SUPERSELF]):
         def __call__(self, /, size: int) -> asyncio.queues.Queue:

modal/_utils/function_utils.py

@@ -34,7 +34,6 @@ from ..exception import (
 from ..mount import ROOT_DIR, _is_modal_path, _Mount
 from .blob_utils import (
     MAX_ASYNC_OBJECT_SIZE_BYTES,
-    MAX_OBJECT_SIZE_BYTES,
     blob_download,
     blob_upload_with_r2_failure_info,
 )
@@ -518,12 +517,13 @@ async def _process_result(result: api_pb2.GenericResult, data_format: int, stub,
 
 def should_upload(
     num_bytes: int,
+    max_object_size_bytes: int,
     function_call_invocation_type: Optional["api_pb2.FunctionCallInvocationType.ValueType"],
 ) -> bool:
     """
     Determine if the input should be uploaded to blob storage.
     """
-    return num_bytes > MAX_OBJECT_SIZE_BYTES or (
+    return num_bytes > max_object_size_bytes or (
         function_call_invocation_type == api_pb2.FUNCTION_CALL_INVOCATION_TYPE_ASYNC
         and num_bytes > MAX_ASYNC_OBJECT_SIZE_BYTES
     )
@@ -534,6 +534,7 @@ async def _create_input(
     kwargs,
     stub: ModalClientModal,
     *,
+    max_object_size_bytes: int,
     idx: Optional[int] = None,
     method_name: Optional[str] = None,
     function_call_invocation_type: Optional["api_pb2.FunctionCallInvocationType.ValueType"] = None,
@@ -548,7 +549,7 @@ async def _create_input(
 
     args_serialized = serialize((args, kwargs))
 
-    if should_upload(len(args_serialized), function_call_invocation_type):
+    if should_upload(len(args_serialized), max_object_size_bytes, function_call_invocation_type):
         args_blob_id, r2_failed, r2_latency_ms = await blob_upload_with_r2_failure_info(args_serialized, stub)
         return api_pb2.FunctionPutInputsItem(
             input=api_pb2.FunctionInput(

modal/client.pyi

@@ -31,7 +31,7 @@ class _Client:
         server_url: str,
         client_type: int,
         credentials: typing.Optional[tuple[str, str]],
-        version: str = "1.0.6.dev9",
+        version: str = "1.0.6.dev15",
     ):
         """mdmd:hidden
         The Modal client object is not intended to be instantiated directly by users.
@@ -160,7 +160,7 @@ class Client:
         server_url: str,
         client_type: int,
         credentials: typing.Optional[tuple[str, str]],
-        version: str = "1.0.6.dev9",
+        version: str = "1.0.6.dev15",
     ):
         """mdmd:hidden
         The Modal client object is not intended to be instantiated directly by users.

modal/functions.pyi

@@ -428,7 +428,7 @@ class Function(
 
     _call_generator: ___call_generator_spec[typing_extensions.Self]
 
-    class __remote_spec(typing_extensions.Protocol[P_INNER, ReturnType_INNER, SUPERSELF]):
+    class __remote_spec(typing_extensions.Protocol[ReturnType_INNER, P_INNER, SUPERSELF]):
         def __call__(self, /, *args: P_INNER.args, **kwargs: P_INNER.kwargs) -> ReturnType_INNER:
             """Calls the function remotely, executing it with the given arguments and returning the execution's result."""
             ...
@@ -437,7 +437,7 @@ class Function(
             """Calls the function remotely, executing it with the given arguments and returning the execution's result."""
             ...
 
-    remote: __remote_spec[modal._functions.P, modal._functions.ReturnType, typing_extensions.Self]
+    remote: __remote_spec[modal._functions.ReturnType, modal._functions.P, typing_extensions.Self]
 
     class __remote_gen_spec(typing_extensions.Protocol[SUPERSELF]):
         def __call__(self, /, *args, **kwargs) -> typing.Generator[typing.Any, None, None]:
@@ -464,7 +464,7 @@ class Function(
         """
         ...
 
-    class ___experimental_spawn_spec(typing_extensions.Protocol[P_INNER, ReturnType_INNER, SUPERSELF]):
+    class ___experimental_spawn_spec(typing_extensions.Protocol[ReturnType_INNER, P_INNER, SUPERSELF]):
         def __call__(self, /, *args: P_INNER.args, **kwargs: P_INNER.kwargs) -> FunctionCall[ReturnType_INNER]:
             """[Experimental] Calls the function with the given arguments, without waiting for the results.
 
@@ -488,7 +488,7 @@ class Function(
             ...
 
     _experimental_spawn: ___experimental_spawn_spec[
-        modal._functions.P, modal._functions.ReturnType, typing_extensions.Self
+        modal._functions.ReturnType, modal._functions.P, typing_extensions.Self
     ]
 
     class ___spawn_map_inner_spec(typing_extensions.Protocol[P_INNER, SUPERSELF]):
@@ -497,7 +497,7 @@ class Function(
 
     _spawn_map_inner: ___spawn_map_inner_spec[modal._functions.P, typing_extensions.Self]
 
-    class __spawn_spec(typing_extensions.Protocol[P_INNER, ReturnType_INNER, SUPERSELF]):
+    class __spawn_spec(typing_extensions.Protocol[ReturnType_INNER, P_INNER, SUPERSELF]):
         def __call__(self, /, *args: P_INNER.args, **kwargs: P_INNER.kwargs) -> FunctionCall[ReturnType_INNER]:
             """Calls the function with the given arguments, without waiting for the results.
 
@@ -518,7 +518,7 @@ class Function(
             """
             ...
 
-    spawn: __spawn_spec[modal._functions.P, modal._functions.ReturnType, typing_extensions.Self]
+    spawn: __spawn_spec[modal._functions.ReturnType, modal._functions.P, typing_extensions.Self]
 
     def get_raw_f(self) -> collections.abc.Callable[..., typing.Any]:
         """Return the inner Python object wrapped by this Modal Function."""

modal/image.py

@@ -1222,6 +1222,114 @@ class _Image(_Object, type_prefix="im"):
             gpu_config=parse_gpu_config(gpu),
         )
 
+    def uv_pip_install(
+        self,
+        *packages: Union[str, list[str]],  # A list of Python packages, eg. ["numpy", "matplotlib>=3.5.0"]
+        requirements: Optional[list[str]] = None,  # Passes -r (--requirements) to uv pip install
+        find_links: Optional[str] = None,  # Passes -f (--find-links) to uv pip install
+        index_url: Optional[str] = None,  # Passes -i (--index-url) to uv pip install
+        extra_index_url: Optional[str] = None,  # Passes --extra-index-url to uv pip install
+        pre: bool = False,  # Allow pre-releases using uv pip install --prerelease allow
+        extra_options: str = "",  # Additional options to pass to pip install, e.g. "--no-build-isolation"
+        force_build: bool = False,  # Ignore cached builds, similar to 'docker build --no-cache'
+        uv_version: Optional[str] = None,  # uv version to use
+        secrets: Sequence[_Secret] = [],
+        gpu: GPU_T = None,
+    ) -> "_Image":
+        """Install a list of Python packages using uv pip install.
+
+        **Examples**
+
+        Simple installation:
+        ```python
+        image = modal.Image.debian_slim().uv_pip_install("torch==2.7.1", "numpy")
+        ```
+
+        This method assumes that:
+        - Python is on the `$PATH` and dependencies are installed with the first Python on the `$PATH`.
+        - Shell supports backticks for substitution
+        - `which` command is on the `$PATH`
+        """
+        pkgs = _flatten_str_args("uv_pip_install", "packages", packages)
+
+        if requirements is None or isinstance(requirements, list):
+            requirements = requirements or []
+        else:
+            raise InvalidError("requirements must be None or a list of strings")
+
+        if not pkgs and not requirements:
+            return self
+        elif not _validate_packages(pkgs):
+            raise InvalidError(
+                "Package list for `Image.uv_pip_install` cannot contain other arguments;"
+                " try the `extra_options` parameter instead."
+            )
+
+        def build_dockerfile(version: ImageBuilderVersion) -> DockerfileSpec:
+            commands = ["FROM base"]
+            UV_ROOT = "/.uv"
+            if uv_version is None:
+                commands.append(f"COPY --from=ghcr.io/astral-sh/uv:latest /uv {UV_ROOT}/uv")
+            else:
+                commands.append(f"COPY --from=ghcr.io/astral-sh/uv:{uv_version} /uv {UV_ROOT}/uv")
+
+            # NOTE: Using `which python` assumes:
+            # - python is on the PATH and uv is installing into the first python in the PATH
+            # - the shell supports backticks for substitution
+            # - `which` command is on the PATH
+            uv_pip_args = ["--python `which python`", "--compile-bytecode"]
+            context_files = {}
+
+            if find_links:
+                uv_pip_args.append(f"--find-links {shlex.quote(find_links)}")
+            if index_url:
+                uv_pip_args.append(f"--index-url {shlex.quote(index_url)}")
+            if extra_index_url:
+                uv_pip_args.append(f"--extra-index-url {shlex.quote(extra_index_url)}")
+            if pre:
+                uv_pip_args.append("--prerelease allow")
+            if extra_options:
+                uv_pip_args.append(extra_options)
+
+            if requirements:
+
+                def _generate_paths(idx: int, req: str) -> dict:
+                    local_path = os.path.expanduser(req)
+                    basename = os.path.basename(req)
+
+                    # The requirement files can have the same name but in different directories:
+                    # requirements=["test/requirements.txt", "a/b/c/requirements.txt"]
+                    # To uniquely identify these files, we add a `idx` prefix to every file's basename
+                    # - `test/requirements.txt` -> `/.0_requirements.txt` in context -> `/.uv/0/requirements.txt` to uv
+                    # - `a/b/c/requirements.txt` -> `/.1_requirements.txt` in context -> `/.uv/1/requirements.txt` to uv
+                    return {
+                        "local_path": local_path,
+                        "context_path": f"/.{idx}_{basename}",
+                        "dest_path": f"{UV_ROOT}/{idx}/{basename}",
+                    }
+
+                requirement_paths = [_generate_paths(idx, req) for idx, req in enumerate(requirements)]
+                requirements_cli = " ".join(f"--requirements {req['dest_path']}" for req in requirement_paths)
+                uv_pip_args.append(requirements_cli)
+
+                commands.extend([f"COPY {req['context_path']} {req['dest_path']}" for req in requirement_paths])
+                context_files.update({req["context_path"]: req["local_path"] for req in requirement_paths})
+
+            uv_pip_args.extend(shlex.quote(p) for p in sorted(pkgs))
+            uv_pip_args_joined = " ".join(uv_pip_args)
+
+            commands.append(f"RUN {UV_ROOT}/uv pip install {uv_pip_args_joined}")
+
+            return DockerfileSpec(commands=commands, context_files=context_files)
+
+        return _Image._from_args(
+            base_images={"base": self},
+            dockerfile_function=build_dockerfile,
+            force_build=self.force_build or force_build,
+            gpu_config=parse_gpu_config(gpu),
+            secrets=secrets,
+        )
+
     def poetry_install_from_file(
         self,
         poetry_pyproject_toml: str,
@@ -1312,6 +1420,149 @@ class _Image(_Object, type_prefix="im"):
             gpu_config=parse_gpu_config(gpu),
         )
 
+    def uv_sync(
+        self,
+        uv_project_dir: str = "./",  # Path to local uv managed project
+        *,
+        force_build: bool = False,  # Ignore cached builds, similar to 'docker build --no-cache'
+        groups: Optional[list[str]] = None,  # Dependency group to install using `uv sync --group`
+        extras: Optional[list[str]] = None,  # Optional dependencies to install using `uv sync --extra`
+        frozen: bool = True,  # If True, then we run `uv sync --frozen` when a uv.lock file is present
+        extra_options: str = "",  # Extra options to pass to `uv sync`
+        uv_version: Optional[str] = None,  # uv version to use
+        secrets: Sequence[_Secret] = [],
+        gpu: GPU_T = None,
+    ) -> "_Image":
+        """Creates a virtual environment with the dependencies in a uv managed project with `uv sync`.
+
+        **Examples**
+        ```python
+        image = modal.Image.debian_slim().uv_sync()
+        ```
+        """
+
+        def _normalize_items(items, name) -> list[str]:
+            if items is None:
+                return []
+            elif isinstance(items, list):
+                return items
+            else:
+                raise InvalidError(f"{name} must be None or a list of strings")
+
+        groups = _normalize_items(groups, "groups")
+        extras = _normalize_items(extras, "extras")
+
+        def _check_pyproject_toml(pyproject_toml: str, version: ImageBuilderVersion):
+            if not os.path.exists(pyproject_toml):
+                raise InvalidError(f"Expected {pyproject_toml} to exist")
+
+            import toml
+
+            with open(pyproject_toml) as f:
+                pyproject_toml_content = toml.load(f)
+
+            if (
+                "tool" in pyproject_toml_content
+                and "uv" in pyproject_toml_content["tool"]
+                and "workspace" in pyproject_toml_content["tool"]["uv"]
+            ):
+                raise InvalidError("uv workspaces are not supported")
+
+            if version > "2024.10":
+                # For builder version > 2024.10, modal is mounted at runtime and is not
+                # a requirement in `uv.lock`
+                return
+
+            dependencies = pyproject_toml_content["project"]["dependencies"]
+
+            for group in groups:
+                if (
+                    "dependency-groups" in pyproject_toml_content
+                    and group in pyproject_toml_content["dependency-groups"]
+                ):
+                    dependencies += pyproject_toml_content["dependency-groups"][group]
+
+            for extra in extras:
+                if (
+                    "project" in pyproject_toml_content
+                    and "optional-dependencies" in pyproject_toml_content["project"]
+                    and extra in pyproject_toml_content["project"]["optional-dependencies"]
+                ):
+                    dependencies += pyproject_toml_content["project"]["optional-dependencies"][extra]
+
+            PACKAGE_REGEX = re.compile(r"^[\w-]+")
+
+            def _extract_package(package) -> str:
+                m = PACKAGE_REGEX.match(package)
+                return m.group(0) if m else ""
+
+            if not any(_extract_package(dependency) == "modal" for dependency in dependencies):
+                raise InvalidError(
+                    "Image builder version <= 2024.10 requires modal to be specified in your pyproject.toml file"
+                )
+
+        def build_dockerfile(version: ImageBuilderVersion) -> DockerfileSpec:
+            uv_project_dir_ = os.path.expanduser(uv_project_dir)
+            pyproject_toml = os.path.join(uv_project_dir_, "pyproject.toml")
+
+            UV_ROOT = "/.uv"
+            uv_sync_args = [
+                f"--project={UV_ROOT}",
+                "--no-install-workspace",  # Do not install the root project or any "uv workspace"
+                "--compile-bytecode",
+            ]
+
+            for group in groups:
+                uv_sync_args.append(f"--group={group}")
+            for extra in extras:
+                uv_sync_args.append(f"--extra={extra}")
+            if extra_options:
+                uv_sync_args.append(extra_options)
+
+            commands = ["FROM base"]
+
+            if uv_version is None:
+                commands.append(f"COPY --from=ghcr.io/astral-sh/uv:latest /uv {UV_ROOT}/uv")
+            else:
+                commands.append(f"COPY --from=ghcr.io/astral-sh/uv:{uv_version} /uv {UV_ROOT}/uv")
+
+            context_files = {}
+
+            _check_pyproject_toml(pyproject_toml, version)
+
+            context_files["/.pyproject.toml"] = pyproject_toml
+            commands.append(f"COPY /.pyproject.toml {UV_ROOT}/pyproject.toml")
+
+            uv_lock = os.path.join(uv_project_dir_, "uv.lock")
+            if os.path.exists(uv_lock):
+                context_files["/.uv.lock"] = uv_lock
+                commands.append(f"COPY /.uv.lock {UV_ROOT}/uv.lock")
+
+                if frozen:
+                    # Do not update `uv.lock` when we have one when `frozen=True`. This it ehd efault because this
+                    # ensures that the runtime environment matches the local `uv.lock`.
+                    #
+                    # If `frozen=False`, then `uv sync` will update the the dependencies in the `uv.lock` file
+                    # during build time.
+                    uv_sync_args.append("--frozen")
+
+            uv_sync_args_joined = " ".join(uv_sync_args).strip()
+
+            commands += [
+                f"RUN {UV_ROOT}/uv sync {uv_sync_args_joined}",
+                f"ENV PATH={UV_ROOT}/.venv/bin:$PATH",
+            ]
+
+            return DockerfileSpec(commands=commands, context_files=context_files)
+
+        return _Image._from_args(
+            base_images={"base": self},
+            dockerfile_function=build_dockerfile,
+            force_build=self.force_build or force_build,
+            secrets=secrets,
+            gpu_config=parse_gpu_config(gpu),
+        )
+
     def dockerfile_commands(
         self,
         *dockerfile_commands: Union[str, list[str]],