modal 1.5.3.dev1__tar.gz → 1.5.3.dev2__tar.gz

{modal-1.5.3.dev1 → modal-1.5.3.dev2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: modal
-Version: 1.5.3.dev1
+Version: 1.5.3.dev2
 Summary: Python client library for Modal
 Author-email: Modal Labs <support@modal.com>
 License-Expression: Apache-2.0

{modal-1.5.3.dev1 → modal-1.5.3.dev2}/modal/_functions.py

@@ -671,6 +671,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
         # Experimental: Clustered functions
         cluster_size: int | None = None,
         rdma: bool | None = None,
+        fabric_size: int | None = None,
         single_use_containers: bool = False,
         ephemeral_disk: int | None = None,
         include_source: bool = True,
@@ -1009,6 +1010,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                     mount_client_dependencies=mount_client_dependencies,
                     # ---
                     _experimental_group_size=cluster_size or 0,  # Experimental: Clustered functions
+                    _experimental_fabric_size=fabric_size or 0,  # Experimental: Clustered functions
                     _experimental_concurrent_cancellations=True,
                     # --- These are deprecated in favor of autoscaler_settings
                     warm_pool_size=min_containers or 0,
@@ -1051,6 +1053,7 @@ class _Function(typing.Generic[P, ReturnType, OriginalReturnType], _Object, type
                         method_definitions_set=function_definition.method_definitions_set,
                         experimental_options=experimental_options or {},
                         _experimental_group_size=function_definition._experimental_group_size,
+                        _experimental_fabric_size=function_definition._experimental_fabric_size,
                         _experimental_buffer_containers=function_definition._experimental_buffer_containers,
                         _experimental_custom_scaling=function_definition._experimental_custom_scaling,
                         snapshot_debug=function_definition.snapshot_debug,

{modal-1.5.3.dev1 → modal-1.5.3.dev2}/modal/_partial_function.py

@@ -74,6 +74,7 @@ class _PartialFunctionParams:
     target_concurrent_inputs: int | None = None
     build_timeout: int | None = None
     rdma: bool | None = None
+    fabric_size: int | None = None
     http_config: api_pb2.HTTPConfig | None = None
 
     def update(self, other: "_PartialFunctionParams") -> None:
@@ -777,7 +778,7 @@ def _concurrent(
 
 # NOTE: clustered is currently exposed through modal.experimental, not the top-level namespace
 def _clustered(
-    size: int, broadcast: bool = True, rdma: bool = False
+    size: int, broadcast: bool = True, rdma: bool = False, fabric_size: int | None = None
 ) -> Callable[
     [Callable[P, ReturnType] | _PartialFunction[P, ReturnType, ReturnType]],
     _PartialFunction[P, ReturnType, ReturnType],
@@ -791,6 +792,11 @@ def _clustered(
         If True, inputs will be sent simultaneously to each container. Otherwise,
         inputs will be sent only to the rank-0 container, which is responsible for
         delegating to the workers.
+    fabric_size: int | None = None
+        Experimental: constrains placement across GPU-memory fabrics. The cluster
+        is placed in co-fabric blocks of `fabric_size` containers, so every block
+        of `fabric_size` consecutive ranks shares a fabric (the scale-up domain)
+        and communicates over cross-node NVLink. Must evenly divide `size`.
     """
 
     assert broadcast, "broadcast=False has not been implemented yet!"
@@ -798,8 +804,14 @@ def _clustered(
     if size <= 0:
         raise ValueError("cluster size must be greater than 0")
 
+    if fabric_size is not None:
+        if not isinstance(fabric_size, int) or fabric_size <= 0:
+            raise ValueError("fabric_size must be a positive integer")
+        if size % fabric_size != 0:
+            raise ValueError(f"fabric_size must evenly divide the cluster size ({size} % {fabric_size} != 0)")
+
     flags = _PartialFunctionFlags.CLUSTERED
-    params = _PartialFunctionParams(cluster_size=size, rdma=rdma)
+    params = _PartialFunctionParams(cluster_size=size, rdma=rdma, fabric_size=fabric_size)
 
     def wrapper(
         obj: _PartialFunction[P, ReturnType, ReturnType] | Callable[P, ReturnType],

{modal-1.5.3.dev1 → modal-1.5.3.dev2}/modal/app.py

@@ -920,6 +920,7 @@ class _App:
                 i6pn_enabled = i6pn or (f.flags & _PartialFunctionFlags.CLUSTERED)
                 cluster_size = f.params.cluster_size  # Experimental: Clustered functions
                 rdma = f.params.rdma
+                fabric_size = f.params.fabric_size
 
                 info = FunctionInfo(f.raw_f, serialized=serialized, name_override=name)
                 raw_f = f.raw_f
@@ -974,6 +975,7 @@ class _App:
 
                 cluster_size = None  # Experimental: Clustered functions
                 rdma = None
+                fabric_size = None
                 i6pn_enabled = i6pn
 
             if is_generator is None:
@@ -1016,6 +1018,7 @@ class _App:
                 i6pn_enabled=i6pn_enabled,
                 cluster_size=cluster_size,  # Experimental: Clustered functions
                 rdma=rdma,
+                fabric_size=fabric_size,  # Experimental: Clustered functions
                 include_source=include_source if include_source is not None else local_state.include_source_default,
                 experimental_options={k: str(v) for k, v in (experimental_options or {}).items()},
                 restrict_output=_experimental_restrict_output,
@@ -1151,16 +1154,19 @@ class _App:
                 if wrapped_cls.flags & _PartialFunctionFlags.CLUSTERED:
                     cluster_size = wrapped_cls.params.cluster_size
                     rdma = wrapped_cls.params.rdma
+                    fabric_size = wrapped_cls.params.fabric_size
 
                 else:
                     cluster_size = None
                     rdma = None
+                    fabric_size = None
             else:
                 user_cls = wrapped_cls
                 max_concurrent_inputs = None
                 target_concurrent_inputs = None
                 cluster_size = None
                 rdma = None
+                fabric_size = None
             if not inspect.isclass(user_cls):
                 raise TypeError("The @app.cls decorator must be used on a class.")
 
@@ -1246,6 +1252,7 @@ class _App:
                 i6pn_enabled=i6pn_enabled,
                 cluster_size=cluster_size,
                 rdma=rdma,
+                fabric_size=fabric_size,
                 include_source=include_source if include_source is not None else local_state.include_source_default,
                 experimental_options={k: str(v) for k, v in (experimental_options or {}).items()},
                 restrict_output=_experimental_restrict_output,
@@ -1403,6 +1410,7 @@ class _App:
             # Extract the underlying class if wrapped in a _PartialFunction (e.g., from @modal.clustered())
             cluster_size = None
             rdma = None
+            fabric_size = None
             user_cls = wrapped_user_cls
 
             if isinstance(wrapped_user_cls, _PartialFunction):
@@ -1410,6 +1418,7 @@ class _App:
                 if wrapped_user_cls.flags & _PartialFunctionFlags.CLUSTERED:
                     cluster_size = wrapped_user_cls.params.cluster_size
                     rdma = wrapped_user_cls.params.rdma
+                    fabric_size = wrapped_user_cls.params.fabric_size
 
             local_state = self._local_state
 
@@ -1449,6 +1458,7 @@ class _App:
                 i6pn_enabled=i6pn or (cluster_size is not None),
                 cluster_size=cluster_size,
                 rdma=rdma,
+                fabric_size=fabric_size,
                 include_source=include_source if include_source is not None else local_state.include_source_default,
                 experimental_options={k: str(v) for k, v in (experimental_options or {}).items()},
                 restrict_output=False,

{modal-1.5.3.dev1 → modal-1.5.3.dev2}/modal/client.pyi

@@ -35,7 +35,7 @@ class _Client:
         server_url: str,
         client_type: int,
         credentials: typing.Optional[tuple[str, str]],
-        version: str = "1.5.3.dev1",
+        version: str = "1.5.3.dev2",
     ):
         """mdmd:hidden
         The Modal client object is not intended to be instantiated directly by users.
@@ -205,7 +205,7 @@ class Client:
         server_url: str,
         client_type: int,
         credentials: typing.Optional[tuple[str, str]],
-        version: str = "1.5.3.dev1",
+        version: str = "1.5.3.dev2",
     ):
         """mdmd:hidden
         The Modal client object is not intended to be instantiated directly by users.

{modal-1.5.3.dev1 → modal-1.5.3.dev2}/modal/functions.pyi

@@ -109,6 +109,7 @@ class Function(
         i6pn_enabled: bool = False,
         cluster_size: typing.Optional[int] = None,
         rdma: typing.Optional[bool] = None,
+        fabric_size: typing.Optional[int] = None,
         single_use_containers: bool = False,
         ephemeral_disk: typing.Optional[int] = None,
         include_source: bool = True,

{modal-1.5.3.dev1 → modal-1.5.3.dev2}/modal.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: modal
-Version: 1.5.3.dev1
+Version: 1.5.3.dev2
 Summary: Python client library for Modal
 Author-email: Modal Labs <support@modal.com>
 License-Expression: Apache-2.0