diff --git a/docs/source/devguide/mixins.rst b/docs/source/devguide/mixins.rst
index 1a548b26..695f177a 100644
--- a/docs/source/devguide/mixins.rst
+++ b/docs/source/devguide/mixins.rst
@@ -79,7 +79,7 @@ convenience.
    class CrossValidationMixin(object):
        def cross_validate(modelName, Xname, Yname, *args, **kwargs):
             # get the cross validation task
-            task = self.runtime.task('custom.tasks.cross_validate')
+            task = self.task('custom.tasks.cross_validate')
             return task.delay(modelName, Xname, Yname, *args, **kwargs)
             
    
diff --git a/omegaml/backends/keras.py b/omegaml/backends/keras.py
index 9127e62a..f235da01 100644
--- a/omegaml/backends/keras.py
+++ b/omegaml/backends/keras.py
@@ -1,6 +1,4 @@
 import os
-from keras import Sequential, Model
-from keras.engine.saving import load_model
 from mongoengine import GridFSProxy
 
 from omegaml.backends import BaseModelBackend
@@ -11,6 +9,7 @@ class KerasBackend(BaseModelBackend):
 
     @classmethod
     def supports(self, obj, name, **kwargs):
+        from keras import Sequential, Model
         return isinstance(obj, (Sequential, Model))
 
     def put_model(self, obj, name, attributes=None):
@@ -31,6 +30,8 @@ def put_model(self, obj, name, attributes=None):
             gridfile=gridfile).save()
 
     def get_model(self, name, version=-1):
+        from keras.engine.saving import load_model
+
         filename = self.model_store._get_obj_store_key(name, 'h5')
         packagefname = os.path.join(self.model_store.tmppath, name)
         dirname = os.path.dirname(packagefname)
@@ -61,7 +62,7 @@ def _fit_tpu(self, modelname, Xname, Yname=None, tpu_specs=None, **kwargs):
         import tensorflow as tf
         # adopted from https://www.dlology.com/blog/how-to-train-keras-model-x20-times-faster-with-tpu-for-free/
         # This address identifies the TPU we'll use when configuring TensorFlow.
-        TPU_WORKER = 'grpc://' + os.environ.get('COLAB_TPU_ADDR','')
+        TPU_WORKER = 'grpc://' + os.environ.get('COLAB_TPU_ADDR', '')
         tf.logging.set_verbosity(tf.logging.INFO)
         model = self.get_model(modelname)
         tpu_model = tf.contrib.tpu.keras_to_tpu_model(
diff --git a/omegaml/defaults.py b/omegaml/defaults.py
index 8ca649b1..36b91aa3 100644
--- a/omegaml/defaults.py
+++ b/omegaml/defaults.py
@@ -44,6 +44,7 @@
 #: storage backends
 OMEGA_STORE_BACKENDS = {
     'sklearn.joblib': 'omegaml.backends.ScikitLearnBackend',
+    'keras.h5': 'omegaml.backends.keras.KerasBackend',
 }
 #: storage mixins
 OMEGA_STORE_MIXINS = [
diff --git a/omegaml/runtimes/daskruntime.py b/omegaml/runtimes/daskruntime.py
index d8959321..85a74345 100644
--- a/omegaml/runtimes/daskruntime.py
+++ b/omegaml/runtimes/daskruntime.py
@@ -12,7 +12,7 @@ class DaskTask(object):
     A dask remote function wrapper mimicking a Celery task
     """
 
-    def __init__(self, fn, client, pure=True):
+    def __init__(self, fn, client, pure=True, **kwargs):
         """
         :param fn: (function) the function to be called
         :param client: (dask client) the dask client to use
@@ -22,12 +22,15 @@ def __init__(self, fn, client, pure=True):
         self.client = client
         self.fn = fn
         self.pure = pure
+        self.kwargs = kwargs
 
     def delay(self, *args, **kwargs):
         """
         submit the function and execute on cluster.  
         """
         kwargs['pure'] = kwargs.get('pure', self.pure)
+        if self.kwargs:
+            kwargs.update(self.kwargs)
         return DaskAsyncResult(self.client.submit(self.fn, *args, **kwargs))
 
 
@@ -90,7 +93,7 @@ def job(self, jobname):
         """
         return OmegaJobProxy(jobname, runtime=self)
 
-    def task(self, name):
+    def task(self, name, **kwargs):
         """
         retrieve the task function from the task module
 
@@ -107,7 +110,7 @@ def task(self, name):
         func = getattr(mod, funcname)
         # we pass pure=False to force dask to reevaluate the task
         # http://distributed.readthedocs.io/en/latest/client.html?highlight=pure#pure-functions-by-default
-        return DaskTask(func, self.client, pure=False)
+        return DaskTask(func, self.client, pure=False, **kwargs)
 
     def settings(self):
         """
diff --git a/omegaml/runtimes/mixins/gridsearch.py b/omegaml/runtimes/mixins/gridsearch.py
index 5c7f3492..0d7092a0 100644
--- a/omegaml/runtimes/mixins/gridsearch.py
+++ b/omegaml/runtimes/mixins/gridsearch.py
@@ -4,7 +4,7 @@ def _common_kwargs(self):
         return dict(pure_python=self.pure_python)
 
     def gridsearch(self, Xname, Yname, parameters=None, pure_python=False, **kwargs):
-        gs_task = self.runtime.task('omegaml.tasks.omega_gridsearch')
+        gs_task = self.task('omegaml.tasks.omega_gridsearch')
         Xname = self._ensure_data_is_stored(Xname, prefix='_fitX')
         if Yname is not None:
             Yname = self._ensure_data_is_stored(Yname, prefix='_fitY')
diff --git a/omegaml/runtimes/mixins/modelmixin.py b/omegaml/runtimes/mixins/modelmixin.py
index dec969c7..24d7588b 100644
--- a/omegaml/runtimes/mixins/modelmixin.py
+++ b/omegaml/runtimes/mixins/modelmixin.py
@@ -31,7 +31,7 @@ def fit(self, Xname, Yname=None, **kwargs):
         :param Yname: name of Y dataset or data
         :return: the model (self) or the string representation (python clients)
         """
-        omega_fit = self.runtime.task('omegaml.tasks.omega_fit')
+        omega_fit = self.task('omegaml.tasks.omega_fit')
         Xname = self._ensure_data_is_stored(Xname, prefix='_fitX')
         if Yname is not None:
             Yname = self._ensure_data_is_stored(Yname, prefix='_fitY')
@@ -54,7 +54,7 @@ def partial_fit(self, Xname, Yname=None, **kwargs):
         :param Yname: name of Y dataset or data
         :return: the model (self) or the string representation (python clients) 
         """
-        omega_fit = self.runtime.task('omegaml.tasks.omega_partial_fit')
+        omega_fit = self.task('omegaml.tasks.omega_partial_fit')
         Xname = self._ensure_data_is_stored(Xname, prefix='_fitX')
         if Yname is not None:
             Yname = self._ensure_data_is_stored(Yname, prefix='_fitY')
@@ -73,7 +73,7 @@ def transform(self, Xname, rName=None, **kwargs):
         :return: the data returned by .transform, or the metadata of the rName
             dataset if rName was given
         """
-        omega_transform = self.runtime.task('omegaml.tasks.omega_transform')
+        omega_transform = self.task('omegaml.tasks.omega_transform')
         Xname = self._ensure_data_is_stored(Xname)
         return omega_transform.delay(self.modelname, Xname,
                                      rName=rName,
@@ -93,7 +93,7 @@ def fit_transform(self, Xname, Yname=None, rName=None, **kwargs):
            dataset if rName was given
         """
 
-        omega_fit_transform = self.runtime.task(
+        omega_fit_transform = self.task(
             'omegaml.tasks.omega_fit_transform')
         Xname = self._ensure_data_is_stored(Xname)
         if Yname is not None:
@@ -114,7 +114,7 @@ def predict(self, Xpath_or_data, rName=None, **kwargs):
         :return: the data returned by .predict, or the metadata of the rName
             dataset if rName was given
         """
-        omega_predict = self.runtime.task('omegaml.tasks.omega_predict')
+        omega_predict = self.task('omegaml.tasks.omega_predict')
         Xname = self._ensure_data_is_stored(Xpath_or_data)
         return omega_predict.delay(self.modelname, Xname, rName=rName,
                                    **self._common_kwargs, **kwargs)
@@ -131,7 +131,7 @@ def predict_proba(self, Xpath_or_data, rName=None, **kwargs):
         :return: the data returned by .predict_proba, or the metadata of the rName
            dataset if rName was given
         """
-        omega_predict_proba = self.runtime.task(
+        omega_predict_proba = self.task(
             'omegaml.tasks.omega_predict_proba')
         Xname = self._ensure_data_is_stored(Xpath_or_data)
         return omega_predict_proba.delay(self.modelname, Xname, rName=rName,
@@ -150,7 +150,7 @@ def score(self, Xname, yName, rName=None, **kwargs):
         :return: the data returned by .score, or the metadata of the rName
            dataset if rName was given
         """
-        omega_score = self.runtime.task('omegaml.tasks.omega_score')
+        omega_score = self.task('omegaml.tasks.omega_score')
         Xname = self._ensure_data_is_stored(Xname)
         yName = self._ensure_data_is_stored(yName)
         return omega_score.delay(self.modelname, Xname, yName, rName=rName,
@@ -168,7 +168,7 @@ def decision_function(self, Xname, rName=None, **kwargs):
         :return: the data returned by .score, or the metadata of the rName
            dataset if rName was given
         """
-        omega_decision_function = self.runtime.task('omegaml.tasks.omega_decision_function')
+        omega_decision_function = self.task('omegaml.tasks.omega_decision_function')
         Xname = self._ensure_data_is_stored(Xname)
         return omega_decision_function.delay(self.modelname, Xname, rName=rName,
                                              **self._common_kwargs, **kwargs)
diff --git a/omegaml/runtimes/modelproxy.py b/omegaml/runtimes/modelproxy.py
index ef71590d..11a5faa0 100644
--- a/omegaml/runtimes/modelproxy.py
+++ b/omegaml/runtimes/modelproxy.py
@@ -52,6 +52,7 @@ def __init__(self, modelname, runtime=None):
                                    False)
         self.pure_python = self.pure_python or self._client_is_pure_python()
         self.apply_mixins()
+        self.task_kwargs = {}
 
     def apply_mixins(self):
         """
@@ -61,6 +62,36 @@ def apply_mixins(self):
         for mixin in defaults.OMEGA_RUNTIME_MIXINS:
             extend_instance(self, mixin)
 
+    def task(self, name):
+        """
+        return the task from the runtime with requirements applied
+        """
+        kwargs = self.task_kwargs
+        return self.runtime.task(name, **kwargs)
+
+    def require(self, **kwargs):
+        """
+        specify requirements for the task execution
+
+        Use this to specify resource or routing requirements on the task call
+        sent to the runtime.
+
+        Args:
+            **kwargs: requirements specification that the runtime understands
+
+        Usage:
+            # celery runtime
+            om.runtime.model('foo').require(queue='gpu').fit(...)
+
+            # dask distributed runtime
+            om.runtime.model('foo').require(resources={...}).fit(...)
+
+        Returns:
+            self
+        """
+        self.task_kwargs.update(kwargs)
+        return self
+
     def _client_is_pure_python(self):
         try:
             import pandas as pd
diff --git a/omegaml/runtimes/runtime.py b/omegaml/runtimes/runtime.py
index 8c52d33e..6822fb65 100644
--- a/omegaml/runtimes/runtime.py
+++ b/omegaml/runtimes/runtime.py
@@ -6,6 +6,33 @@
 from omegaml.util import settings
 
 
+class CeleryTask(object):
+    """
+    A thin wrapper for a Celery.Task object
+
+    This is so that we can collect common delay arguments on the
+    .task() call
+    """
+
+    def __init__(self, task, **kwargs):
+        """
+
+        Args:
+            task (Celery.Task): the celery task object
+            **kwargs (dict): optional, the kwargs to pass to apply_async
+        """
+        self.task = task
+        self.kwargs = kwargs
+
+    def delay(self, *args, **kwargs):
+        """
+        submit the task with args and kwargs to pass on
+
+        This calls task.apply_async and passes on the self.kwargs.
+        """
+        return self.task.apply_async(args=args, kwargs=kwargs, **self.kwargs)
+
+
 class OmegaRuntime(object):
     """
     omegaml compute cluster gateway 
@@ -48,7 +75,7 @@ def job(self, jobname):
         """
         return OmegaJobProxy(jobname, runtime=self)
 
-    def task(self, name):
+    def task(self, name, **kwargs):
         """
         retrieve the task function from the celery instance
 
@@ -57,7 +84,7 @@ def task(self, name):
         import, which seems to confuse celery)
         """
         # import omegapkg.tasks
-        return self.celeryapp.tasks.get(name)
+        return CeleryTask(self.celeryapp.tasks.get(name), **kwargs)
 
     def settings(self):
         """