diff --git a/setup.py b/setup.py
index e7de2ac56b..03ab976f9a 100644
--- a/setup.py
+++ b/setup.py
@@ -19,7 +19,7 @@ setup(name='tinygrad',
         "Programming Language :: Python :: 3",
         "License :: OSI Approved :: MIT License"
       ],
-      install_requires=["numpy", "requests", "pillow", "tqdm", "networkx", "pyopencl", "PyYAML",
+      install_requires=["numpy", "requests", "tqdm", "networkx", "pyopencl",
                         "pyobjc-framework-Metal; platform_system=='Darwin'",
                         "pyobjc-framework-Cocoa; platform_system=='Darwin'",
                         "pyobjc-framework-libdispatch; platform_system=='Darwin'"],
@@ -37,11 +37,11 @@ setup(name='tinygrad',
             "typing-extensions",
             "pre-commit",
             "ruff",
-            "types-PyYAML",
             "types-tqdm",
         ],
         'testing': [
             "torch",
+            "pillow",
             "pytest",
             "pytest-xdist",
             "onnx==1.14.1",
diff --git a/tinygrad/runtime/ops_torch.py b/tinygrad/runtime/ops_torch.py
index 7122eb54b2..3a183d0ddc 100644
--- a/tinygrad/runtime/ops_torch.py
+++ b/tinygrad/runtime/ops_torch.py
@@ -18,8 +18,8 @@ def as_strided(x, arg):
 
 torch_fxn_for_op: Dict[Op, Callable] = {**base_fxn_for_op, **{
   # TODO: torch.tensor should work here
-  #BufferOps.CONST: lambda val, dtype: torch.tensor(val, dtype=inverse_type_map[dtype]),
-  BufferOps.CONST: lambda val, dtype: torch.from_numpy(np.array(val, dtype=dtype.np)),
+  #BufferOps.CONST: lambda val, dtype: torch.tensor(val, device=device, dtype=inverse_type_map[dtype]),
+  BufferOps.CONST: lambda val, dtype: torch.from_numpy(np.array(val, dtype=dtype.np)).requires_grad_(False).to(device),
   UnaryOps.NOOP: lambda x: x.contiguous(), UnaryOps.SQRT: lambda x: x.sqrt(), UnaryOps.EXP2: lambda x: x.exp2(), UnaryOps.LOG2: lambda x: x.log2(), UnaryOps.SIN: torch.sin,
   UnaryOps.CAST: lambda x,y: (x.view if y[1] else x.type)(next(k for k,v in type_map.items() if v==y[0])),
   BinaryOps.MAX: torch.maximum, BinaryOps.CMPLT: lambda x,y: (x<y).type(torch.promote_types(x.dtype, y.dtype)), BinaryOps.SUB: lambda x,y: torch.logical_xor(x, y) if y.dtype is torch.bool else torch.sub(x, y),
@@ -32,7 +32,7 @@ torch_fxn_for_op: Dict[Op, Callable] = {**base_fxn_for_op, **{
 }}
 
 class RawTorchBuffer(RawBuffer):
-  def __init__(self, size:int, dtype:DType, buf:Optional[torch.Tensor]=None): super().__init__(size, dtype, buf if buf is not None else torch.empty([size], dtype=inverse_type_map[dtype]))
+  def __init__(self, size:int, dtype:DType, buf:Optional[torch.Tensor]=None): super().__init__(size, dtype, buf if buf is not None else torch.empty([size], device=device, dtype=inverse_type_map[dtype]))
   @classmethod
   def fromCPU(cls, x):
     buf = torch.from_numpy(x if all(s>=0 for s in x.strides) else x.copy()).requires_grad_(False).to(device)