modal 1.1.4.dev33__tar.gz → 1.1.5.dev1__tar.gz

{modal-1.1.4.dev33 → modal-1.1.5.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: modal
-Version: 1.1.4.dev33
+Version: 1.1.5.dev1
 Summary: Python client library for Modal
 Author-email: Modal Labs <support@modal.com>
 License: Apache-2.0

{modal-1.1.4.dev33 → modal-1.1.5.dev1}/modal/client.pyi

@@ -33,7 +33,7 @@ class _Client:
         server_url: str,
         client_type: int,
         credentials: typing.Optional[tuple[str, str]],
-        version: str = "1.1.4.dev33",
+        version: str = "1.1.5.dev1",
     ):
         """mdmd:hidden
         The Modal client object is not intended to be instantiated directly by users.
@@ -164,7 +164,7 @@ class Client:
         server_url: str,
         client_type: int,
         credentials: typing.Optional[tuple[str, str]],
-        version: str = "1.1.4.dev33",
+        version: str = "1.1.5.dev1",
     ):
         """mdmd:hidden
         The Modal client object is not intended to be instantiated directly by users.

{modal-1.1.4.dev33 → modal-1.1.5.dev1}/modal/experimental/__init__.py

@@ -317,14 +317,11 @@ async def notebook_base_image(*, python_version: Optional[str] = None, force_bui
         # Install uv since it's faster than pip for installing packages.
         "pip install uv",
         # https://github.com/astral-sh/uv/issues/11480
-        "pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126",
+        "pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu129",
         f"uv pip install --system {shlex.join(sorted(environment_packages))}",
         f"uv pip install --system {shlex.join(sorted(kernelshim_packages))}",
     ]
 
-    # TODO: Also install the CUDA Toolkit, so `nvcc` is available.
-    # https://github.com/charlesfrye/cuda-modal/blob/7fef8db12402986cf42d9c8cca8c63d1da6d7700/cuda/use_cuda.py#L158-L188
-
     def build_dockerfile(version: ImageBuilderVersion) -> DockerfileSpec:
         return DockerfileSpec(
             commands=[

{modal-1.1.4.dev33 → modal-1.1.5.dev1}/modal/experimental/flash.py

@@ -337,25 +337,42 @@ class _FlashPrometheusAutoscaler:
         if not internal_metrics_list:
             return current_replicas
 
-        avg_internal_metric = sum(internal_metrics_list) / len(internal_metrics_list)
+        sum_metric = sum(internal_metrics_list)
+        containers_with_metrics = len(internal_metrics_list)
+        # n_containers_missing_metric is the number of unhealthy containers + number of cold starting containers
+        n_containers_missing_metric = current_replicas - containers_with_metrics
+        # n_containers_unhealthy is the number of live containers that are not emitting metrics i.e. unhealthy
+        n_containers_unhealthy = len(containers) - containers_with_metrics
+
+        # Scale up assuming that every unhealthy container is at 2x the target metric value.
+        scale_up_target_metric_value = (sum_metric + n_containers_unhealthy * self.target_metric_value) / (
+            (containers_with_metrics + n_containers_unhealthy) or 1
+        )
+
+        # Scale down assuming that every container (including cold starting containers) are at the target metric value.
+        scale_down_target_metric_value = (sum_metric + n_containers_missing_metric * self.target_metric_value) / (
+            current_replicas or 1
+        )
 
-        scale_factor = avg_internal_metric / self.target_metric_value
+        scale_up_ratio = scale_up_target_metric_value / self.target_metric_value
+        scale_down_ratio = scale_down_target_metric_value / self.target_metric_value
 
         desired_replicas = current_replicas
-        if scale_factor > 1 + self.scale_up_tolerance:
-            desired_replicas = math.ceil(current_replicas * scale_factor)
-        elif scale_factor < 1 - self.scale_down_tolerance:
-            desired_replicas = math.ceil(current_replicas * scale_factor)
+        if scale_up_ratio > 1 + self.scale_up_tolerance:
+            desired_replicas = math.ceil(current_replicas * scale_up_ratio)
+        elif scale_down_ratio < 1 - self.scale_down_tolerance:
+            desired_replicas = math.ceil(current_replicas * scale_down_ratio)
 
         logger.warning(
             f"[Modal Flash] Current replicas: {current_replicas}, "
-            f"avg internal metric `{self.target_metric}`: {avg_internal_metric}, "
+            f"sum internal metric `{self.target_metric}`: {sum_metric}, "
             f"target internal metric value: {self.target_metric_value}, "
-            f"scale factor: {scale_factor}, "
+            f"scale up ratio: {scale_up_ratio}, "
+            f"scale down ratio: {scale_down_ratio}, "
             f"desired replicas: {desired_replicas}"
         )
 
-        desired_replicas = max(1, min(desired_replicas, self.max_containers or 1000))
+        desired_replicas = max(1, min(desired_replicas, self.max_containers or 5000))
         return desired_replicas
 
     async def _compute_target_containers_prometheus(self, current_replicas: int) -> int:
@@ -405,9 +422,9 @@ class _FlashPrometheusAutoscaler:
         )
 
         # Scale down assuming that every container (including cold starting containers) are at the target metric value.
-        scale_down_target_metric_value = (
-            sum_metric + n_containers_missing_metric * target_metric_value
-        ) / current_replicas
+        scale_down_target_metric_value = (sum_metric + n_containers_missing_metric * target_metric_value) / (
+            current_replicas or 1
+        )
 
         scale_up_ratio = scale_up_target_metric_value / target_metric_value
         scale_down_ratio = scale_down_target_metric_value / target_metric_value

{modal-1.1.4.dev33 → modal-1.1.5.dev1}/modal.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: modal
-Version: 1.1.4.dev33
+Version: 1.1.5.dev1
 Summary: Python client library for Modal
 Author-email: Modal Labs <support@modal.com>
 License: Apache-2.0

{modal-1.1.4.dev33 → modal-1.1.5.dev1}/modal_version/__init__.py

@@ -1,4 +1,4 @@
 # Copyright Modal Labs 2025
 """Supplies the current version of the modal client library."""
 
-__version__ = "1.1.4.dev33"
+__version__ = "1.1.5.dev1"