diff --git a/hasktorch.cabal b/hasktorch.cabal
--- a/hasktorch.cabal
+++ b/hasktorch.cabal
@@ -1,5 +1,5 @@
 name:                hasktorch
-version:             0.2.0.1
+version:             0.2.1.0
 synopsis:            Haskell bindings to libtorch, supporting both typed and untyped tensors.
 description:         Hasktorch is a library for tensors and neural networks in Haskell. It is an independent open source community project which leverages the core C++ libraries shared by PyTorch.
 homepage:            https://github.com/hasktorch/hasktorch#readme
@@ -102,8 +102,8 @@
  ghc-options:         -fplugin GHC.TypeLits.Normalise -fplugin GHC.TypeLits.KnownNat.Solver -fplugin GHC.TypeLits.Extra.Solver -fconstraint-solver-iterations=0 -fplugin GHC.NotExport.Plugin
  build-depends:       async >= 2.2.5 && < 2.3
                     , base >= 4.7 && < 5
-                    , libtorch-ffi == 2.0.*
-                    , libtorch-ffi-helper == 2.0.*
+                    , libtorch-ffi == 2.0.1.*
+                    , libtorch-ffi-helper == 2.0.0.*
                     , finite-typelits >= 0.1 && < 0.3
                     , ghc-typelits-extra >= 0.4.6 && < 0.5
                     , ghc-typelits-knownnat >= 0.7.9 && < 0.8
diff --git a/src/Torch/Backend.hs b/src/Torch/Backend.hs
--- a/src/Torch/Backend.hs
+++ b/src/Torch/Backend.hs
@@ -7,12 +7,13 @@
 import qualified Torch.Internal.Const as ATen
 import qualified Torch.Internal.Type as ATen
 
-data Backend = CPU | CUDA | HIP | SparseCPU | SparseCUDA | XLA
+data Backend = CPU | CUDA | HIP | SparseCPU | SparseCUDA | XLA | MPS
   deriving (Eq, Show)
 
 instance Castable Backend ATen.Backend where
   cast CPU f = f ATen.bCPU
   cast CUDA f = f ATen.bCUDA
+  cast MPS f = f ATen.bMPS
   cast HIP f = f ATen.bHIP
   cast SparseCPU f = f ATen.bSparseCPU
   cast SparseCUDA f = f ATen.bSparseCUDA
@@ -21,6 +22,7 @@
   uncast x f
     | x == ATen.bCPU = f CPU
     | x == ATen.bCUDA = f CUDA
+    | x == ATen.bMPS = f MPS
     | x == ATen.bHIP = f HIP
     | x == ATen.bSparseCPU = f SparseCPU
     | x == ATen.bSparseCUDA = f SparseCUDA
diff --git a/src/Torch/Device.hs b/src/Torch/Device.hs
--- a/src/Torch/Device.hs
+++ b/src/Torch/Device.hs
@@ -8,7 +8,7 @@
 import qualified Torch.Internal.Const as ATen
 import qualified Torch.Internal.Type as ATen
 
-data DeviceType = CPU | CUDA
+data DeviceType = CPU | CUDA | MPS
   deriving (Eq, Ord, Show)
 
 data Device = Device {deviceType :: DeviceType, deviceIndex :: I.Int16}
@@ -17,7 +17,9 @@
 instance Castable DeviceType ATen.DeviceType where
   cast CPU f = f ATen.kCPU
   cast CUDA f = f ATen.kCUDA
+  cast MPS f = f ATen.kMPS
 
   uncast x f
     | x == ATen.kCPU = f CPU
     | x == ATen.kCUDA = f CUDA
+    | x == ATen.kMPS = f MPS
diff --git a/src/Torch/Random.hs b/src/Torch/Random.hs
--- a/src/Torch/Random.hs
+++ b/src/Torch/Random.hs
@@ -54,6 +54,11 @@
       ATen.generator_set_current_seed genPtr seed
       genenerator <- newTVarIO (Right genPtr)
       return $ UnsafeGenerator genenerator
+    Device MPS _ -> do
+      genPtr <- ATen.newMPSGenerator
+      ATen.generator_set_current_seed genPtr seed
+      genenerator <- newTVarIO (Right genPtr)
+      return $ UnsafeGenerator genenerator
 
 type RandomGenFunc = ForeignPtr ATen.IntArray -> ForeignPtr ATen.Generator -> ForeignPtr ATen.TensorOptions -> IO (ForeignPtr ATen.Tensor)
 
@@ -67,7 +72,10 @@
           let device =
                 if generatorIsCuda v'
                   then Device {deviceType = CUDA, deviceIndex = fromIntegral $ generatorDevice v'}
-                  else Device {deviceType = CPU, deviceIndex = 0}
+                  else
+                    if generatorIsMps v'
+                    then Device {deviceType = MPS, deviceIndex = 0}
+                    else Device {deviceType = CPU, deviceIndex = 0}
               seed = generatorSeed v'
           writeTVar generator $ seed `seq` deviceType device `seq` deviceIndex device `seq` Left (seed, device)
           return $ Right v'
@@ -80,6 +88,10 @@
           gen <- ATen.newCUDAGenerator (fromIntegral idx)
           ATen.generator_set_current_seed gen seed
           return gen
+        Device MPS _ -> do
+          gen <- ATen.newMPSGenerator
+          ATen.generator_set_current_seed gen seed
+          return gen
     tensor <- cast3 func size genPtr options
     nextGenenerator <- newTVarIO (Right genPtr)
     return (tensor, UnsafeGenerator nextGenenerator)
@@ -89,6 +101,9 @@
 
     generatorIsCuda :: ForeignPtr ATen.Generator -> Bool
     generatorIsCuda gen = unsafePerformIO $ cast1 ATen.generator_is_cuda gen
+
+    generatorIsMps :: ForeignPtr ATen.Generator -> Bool
+    generatorIsMps gen = unsafePerformIO $ cast1 ATen.generator_is_mps gen
 
     generatorDevice :: ForeignPtr ATen.Generator -> Int
     generatorDevice gen = unsafePerformIO $ cast1 ATen.generator_get_device gen
diff --git a/src/Torch/Tensor.hs b/src/Torch/Tensor.hs
--- a/src/Torch/Tensor.hs
+++ b/src/Torch/Tensor.hs
@@ -134,11 +134,19 @@
     then do
       isCUDA <- cast1 ATen.tensor_is_cuda t :: IO Bool
       if isCUDA then cuda <$> cast1 ATen.tensor_get_device t else pure cpu
-    else pure cpu
+    else do
+      hasMPS <- cast0 ATen.hasMPS :: IO Bool
+      if hasMPS
+        then do
+        isMPS <- cast1 ATen.tensor_is_mps t :: IO Bool
+        if isMPS then pure mps else pure cpu
+      else
+        pure cpu
   where
     cpu = Device {deviceType = CPU, deviceIndex = 0}
     cuda :: Int -> Device
     cuda di = Device {deviceType = CUDA, deviceIndex = fromIntegral di}
+    mps = Device {deviceType = MPS, deviceIndex = 0}
 
 -- | Returns the data type of the input tensor
 dtype ::
@@ -221,7 +229,10 @@
   -- | output
   Tensor
 _toDevice device' t = unsafePerformIO $ do
-  hasCUDA <- cast0 ATen.hasCUDA :: IO Bool
+  hasDevice <- case deviceType device' of
+    CPU -> pure True
+    CUDA -> cast0 ATen.hasCUDA
+    MPS -> cast0 ATen.hasMPS
   let device = Torch.Tensor.device t
   t' <-
     toDevice'
@@ -229,7 +240,7 @@
       (deviceType device')
       (deviceIndex device)
       (deviceIndex device')
-      hasCUDA
+      hasDevice
   check
     (deviceType device')
     (deviceType $ Torch.Tensor.device t')
@@ -241,6 +252,8 @@
     toDevice' CUDA CUDA di di' True | di /= di' = getOpts t >>= withDeviceIndex di' >>= to t -- copy from di to di'
     toDevice' CPU CUDA 0 di' True | di' >= 0 = getOpts t >>= withDeviceIndex di' >>= to t -- copy from cpu:0 to cuda:di'
     toDevice' CUDA CPU di 0 True | di >= 0 = getOpts t >>= withDeviceType CPU >>= to t -- copy from cuda:di to cpu:0
+    toDevice' CPU MPS 0 0 True = getOpts t >>= withDeviceType MPS >>= to t -- copy from cpu:0 to mps:0'
+    toDevice' MPS CPU 0 0 True = getOpts t >>= withDeviceType CPU >>= to t -- copy from mps:0 to cpu:0
     toDevice' dt dt' di di' _ =
       error $
         "cannot move tensor from \""
@@ -364,6 +377,9 @@
 
 toCUDA :: Tensor -> Tensor
 toCUDA t = unsafePerformIO $ (cast1 ATen.tensor_cuda) t
+
+toMPS :: Tensor -> Tensor
+toMPS t = unsafePerformIO $ (cast1 ATen.tensor_mps) t
 
 --------------------------------------------------------------------------------
 -- Indexing support
diff --git a/src/Torch/TensorOptions.hs b/src/Torch/TensorOptions.hs
--- a/src/Torch/TensorOptions.hs
+++ b/src/Torch/TensorOptions.hs
@@ -35,8 +35,14 @@
 
 withDevice :: Device -> TensorOptions -> TensorOptions
 withDevice Device {..} opts = unsafePerformIO $ do
-  hasCUDA <- cast0 ATen.hasCUDA
-  withDevice' deviceType deviceIndex hasCUDA opts
+  case deviceType of
+    CPU -> pure opts
+    CUDA -> do
+      hasCUDA <- cast0 ATen.hasCUDA
+      withDevice' deviceType deviceIndex hasCUDA opts
+    MPS -> do
+      hasMPS <- cast0 ATen.hasMPS
+      withDevice' deviceType deviceIndex hasMPS opts
   where
     withDeviceType :: DeviceType -> TensorOptions -> IO TensorOptions
     withDeviceType dt opts = cast2 ATen.tensorOptions_device_D opts dt
@@ -44,9 +50,9 @@
     withDeviceIndex di opts = cast2 ATen.tensorOptions_device_index_s opts di -- careful, this somehow implies deviceType = CUDA
     withDevice' ::
       DeviceType -> Int16 -> Bool -> TensorOptions -> IO TensorOptions
-    withDevice' CPU 0 False opts = pure opts
-    withDevice' CPU 0 True opts = pure opts >>= withDeviceType CPU
+    withDevice' CPU 0 _ opts = pure opts >>= withDeviceType CPU
     withDevice' CUDA di True opts | di >= 0 = pure opts >>= withDeviceIndex di
+    withDevice' MPS 0 True opts = pure opts >>= withDeviceType MPS
     withDevice' dt di _ _ =
       error $ "cannot move tensor to \"" <> show dt <> ":" <> show di <> "\""
 
diff --git a/src/Torch/Typed/Auxiliary.hs b/src/Torch/Typed/Auxiliary.hs
--- a/src/Torch/Typed/Auxiliary.hs
+++ b/src/Torch/Typed/Auxiliary.hs
@@ -347,6 +347,10 @@
     ( DTypeIsFloatingPoint '( 'D.CPU, 0) dtype,
       DTypeIsNotHalf '( 'D.CPU, 0) dtype
     )
+  StandardFloatingPointDTypeValidation '( 'D.MPS, 0) dtype =
+    ( DTypeIsFloatingPoint '( 'D.MPS, 0) dtype,
+      DTypeIsNotHalf '( 'D.MPS, 0) dtype
+    )
   StandardFloatingPointDTypeValidation '( 'D.CUDA, deviceIndex) dtype = DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype
   StandardFloatingPointDTypeValidation '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
diff --git a/src/Torch/Typed/Factories.hs b/src/Torch/Typed/Factories.hs
--- a/src/Torch/Typed/Factories.hs
+++ b/src/Torch/Typed/Factories.hs
@@ -90,6 +90,7 @@
       DTypeIsNotHalf '( 'D.CPU, 0) dtype
     )
   RandDTypeIsValid '( 'D.CUDA, _) dtype = ()
+  RandDTypeIsValid '( 'D.MPS, 0) dtype = ()
   RandDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 rand ::
diff --git a/src/Torch/Typed/Functional.hs b/src/Torch/Typed/Functional.hs
--- a/src/Torch/Typed/Functional.hs
+++ b/src/Torch/Typed/Functional.hs
@@ -157,6 +157,7 @@
 type family SumDTypeIsValid (device :: (D.DeviceType, Nat)) (dtype :: D.DType) :: Constraint where
   SumDTypeIsValid '( 'D.CPU, 0) dtype = DTypeIsNotHalf '( 'D.CPU, 0) dtype
   SumDTypeIsValid '( 'D.CUDA, _) dtype = ()
+  SumDTypeIsValid '( 'D.MPS, 0) dtype = ()
   SumDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | sumAll
@@ -249,6 +250,7 @@
 type family MinMaxDTypeIsValid (device :: (D.DeviceType, Nat)) (dtype :: D.DType) :: Constraint where
   MinMaxDTypeIsValid '( 'D.CPU, 0) dtype = DTypeIsNotHalf '( 'D.CPU, 0) dtype
   MinMaxDTypeIsValid '( 'D.CUDA, _) dtype = ()
+  MinMaxDTypeIsValid '( 'D.MPS, 0) dtype = ()
   MinMaxDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | min
@@ -939,6 +941,7 @@
 type family InverseShapeIsValid (device :: (D.DeviceType, Nat)) (shape :: [Nat]) :: Constraint where
   InverseShapeIsValid '( 'D.CPU, 0) _ = ()
   InverseShapeIsValid '( 'D.CUDA, _) shape = AllDimsPositive shape
+  InverseShapeIsValid '( 'D.MPS, 0) shape = AllDimsPositive shape
 
 type family InverseDTypeIsValid (device :: (D.DeviceType, Nat)) (dtype :: D.DType) :: Constraint where
   InverseDTypeIsValid '( 'D.CPU, 0) dtype =
@@ -949,6 +952,10 @@
     ( DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype,
       DTypeIsNotHalf '( 'D.CUDA, deviceIndex) dtype
     )
+  InverseDTypeIsValid '( 'D.MPS, 0) dtype =
+    ( DTypeIsFloatingPoint '( 'D.MPS, 0) dtype,
+      DTypeIsNotHalf '( 'D.MPS, 0) dtype
+    )
   InverseDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | inverse
@@ -981,6 +988,10 @@
     ( DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype,
       DTypeIsNotHalf '( 'D.CUDA, deviceIndex) dtype
     )
+  SymeigDTypeIsValid '( 'D.MPS, 0) dtype =
+    ( DTypeIsFloatingPoint '( 'D.MPS, 0) dtype,
+      DTypeIsNotHalf '( 'D.MPS, 0) dtype
+    )
   SymeigDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | symeig
@@ -1086,6 +1097,10 @@
     ( DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype,
       DTypeIsNotHalf '( 'D.CUDA, deviceIndex) dtype
     )
+  EigDTypeIsValid '( 'D.MPS, 0) dtype =
+    ( DTypeIsFloatingPoint '( 'D.MPS, 0) dtype,
+      DTypeIsNotHalf '( 'D.MPS, 0) dtype
+    )
   EigDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 type family ToComplexNumber (dtype :: D.DType) :: D.DType where
@@ -1168,6 +1183,10 @@
     ( DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype,
       DTypeIsNotHalf '( 'D.CUDA, deviceIndex) dtype
     )
+  SVDDTypeIsValid '(D.MPS, 0) dtype =
+    ( DTypeIsFloatingPoint '(D.MPS, 0) dtype,
+      DTypeIsNotHalf '(D.MPS, 0) dtype
+    )
   SVDDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | Singular Value Decomposition
@@ -1233,6 +1252,10 @@
     ( DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype,
       DTypeIsNotHalf '( 'D.CUDA, deviceIndex) dtype
     )
+  CholeskyDTypeIsValid '(D.MPS, 0) dtype =
+    ( DTypeIsFloatingPoint '(D.MPS, 0) dtype,
+      DTypeIsNotHalf '(D.MPS, 0) dtype
+    )
   CholeskyDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | cholesky
@@ -1343,6 +1366,10 @@
     ( DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype,
       DTypeIsNotHalf '( 'D.CUDA, deviceIndex) dtype
     )
+  SolveDTypeIsValid '( 'D.MPS, 0) dtype =
+    ( DTypeIsFloatingPoint '( 'D.MPS, 0) dtype,
+      DTypeIsNotHalf '( 'D.MPS, 0) dtype
+    )
   SolveDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | solve
@@ -3030,6 +3057,7 @@
       DTypeIsNotHalf '(D.CPU, 0) dtype
     )
   DotDTypeIsValid '( 'D.CUDA, deviceIndex) dtype = DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype
+  DotDTypeIsValid '( 'D.MPS, 0) dtype = DTypeIsFloatingPoint '( 'D.MPS, 0) dtype
   DotDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | dot product
@@ -4168,6 +4196,10 @@
   GeluDTypeIsValid '( 'D.CUDA, deviceIndex) dtype =
     ( DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype,
       DTypeIsNotHalf '( 'D.CUDA, deviceIndex) dtype
+    )
+  GeluDTypeIsValid '( 'D.MPS, 0) dtype =
+    ( DTypeIsFloatingPoint '( 'D.MPS, 0) dtype,
+      DTypeIsNotHalf '( 'D.MPS, 0) dtype
     )
   GeluDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
diff --git a/src/Torch/Typed/Tensor.hs b/src/Torch/Typed/Tensor.hs
--- a/src/Torch/Typed/Tensor.hs
+++ b/src/Torch/Typed/Tensor.hs
@@ -140,6 +140,9 @@
 instance (KnownNat n) => KnownDevice '( 'D.CUDA, n) where
   deviceVal = D.Device D.CUDA (natValInt16 @n)
 
+instance (KnownNat n) => KnownDevice '( 'D.MPS, n) where
+  deviceVal = D.Device D.MPS (natValInt16 @n)
+
 type Size = Type -> Type
 
 type Shape = [Type -> Type]
@@ -207,6 +210,8 @@
 
 type CUDATensor deviceIndex = Tensor '( 'D.CUDA, deviceIndex)
 
+type MPSTensor deviceIndex = Tensor '( 'D.MPS, 0)
+
 data UnknownShapeTensor device dtype = forall shape. UnknownShapeTensor (Tensor device dtype shape)
 
 type family ComputeHaskellType (dtype :: D.DType) :: Type where
@@ -310,6 +315,7 @@
   Just (SomeNat (Proxy :: Proxy n)) -> case deviceType of
     D.CPU -> SomeDevice $ Proxy @'( 'D.CPU, n)
     D.CUDA -> SomeDevice $ Proxy @'( 'D.CUDA, n)
+    D.MPS -> SomeDevice $ Proxy @'( 'D.MPS, n)
 
 withTensor ::
   D.Tensor ->
@@ -381,6 +387,7 @@
       DTypeIsNotHalf '( 'D.CPU, 0) dtype
     )
   BasicArithmeticDTypeIsValid '( 'D.CUDA, _) dtype = ()
+  BasicArithmeticDTypeIsValid '( 'D.MPS, 0) dtype = ()
   BasicArithmeticDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 add,
@@ -408,6 +415,7 @@
       DTypeIsNotHalf '( 'D.CPU, 0) dtype
     )
   ComparisonDTypeIsValid '( 'D.CUDA, _) dtype = ()
+  ComparisonDTypeIsValid '( 'D.MPS, 0) dtype = ()
   ComparisonDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 gt,
@@ -468,6 +476,7 @@
       DTypeIsNotHalf '( 'D.CPU, 0) dtype
     )
   MatMulDTypeIsValid '( 'D.CUDA, deviceIndex) dtype = DTypeIsFloatingPoint '( 'D.CUDA, deviceIndex) dtype
+  MatMulDTypeIsValid '( 'D.MPS, 0) dtype = DTypeIsFloatingPoint '( 'D.MPS, 0) dtype
   MatMulDTypeIsValid '(deviceType, _) dtype = UnsupportedDTypeForDevice deviceType dtype
 
 -- | matrix multiplication
@@ -635,6 +644,14 @@
   Tensor device dtype shape ->
   CUDATensor 0 dtype shape
 toCUDA t = UnsafeMkTensor $ D.toCUDA (toDynamic t)
+
+-- | move tensor to the first MPS device
+-- TODO: what if this fails?
+toMPS ::
+  forall device' device shape dtype.
+  Tensor device dtype shape ->
+  MPSTensor 0 dtype shape
+toMPS t = UnsafeMkTensor $ D.toMPS (toDynamic t)
 
 -- | move tensor to device
 -- TODO: what if this fails?