depthtensor 2.4.2__tar.gz → 2.4.3__tar.gz

{depthtensor-2.4.2 → depthtensor-2.4.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: depthtensor
-Version: 2.4.2
+Version: 2.4.3
 Summary: A Hardware-Accelerated Tensor Computation and Autograd Engine
 Author-email: Le Hong Ha <lehahong1310@gmail.com>
 License: MIT License

{depthtensor-2.4.2 → depthtensor-2.4.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "depthtensor"
-version = "2.4.2"
+version = "2.4.3"
 description = "A Hardware-Accelerated Tensor Computation and Autograd Engine"
 readme = "README.md"
 authors = [

{depthtensor-2.4.2 → depthtensor-2.4.3}/src/DepthTensor/__init__.py

@@ -6,4 +6,4 @@ from ._core.ops import *
 from ._core.exceptions import *
 from ._core import random
 
-__version__ = "2.4.2"
+__version__ = "2.4.3"

{depthtensor-2.4.2 → depthtensor-2.4.3}/src/DepthTensor/_core/ops/elementwise.py

@@ -105,7 +105,7 @@ def create_2in_1out(op: BinaryOp, diff: BinaryDiff | None) -> BinaryFunc:
             if y.requires_grad:
                 if y.grad is None:
                     y.zeros_grad()
-                dx1, dx2 = diff(y, a1, a2, **kwds)
+                dx1, dx2 = diff(y, a1, a2, device=op_device, **kwds)
 
                 def backward() -> None:
                     if isinstance(x1, Tensor) and x1.requires_grad:
@@ -171,7 +171,7 @@ def create_1in_1out(op: UnaryOp, diff: UnaryDiff | None) -> UnaryFunc:
             if y.requires_grad:
                 if y.grad is None:
                     y.zeros_grad()
-                dx = diff(y, a, **kwds)
+                dx = diff(y, a, device=device_op, **kwds)
 
                 def backward() -> None:
                     if isinstance(x, Tensor) and x.requires_grad:

{depthtensor-2.4.2 → depthtensor-2.4.3}/src/DepthTensor/_core/ops/reduction.py

@@ -90,9 +90,12 @@ def max(
 
     arr = to_tensordata(a, device=device_op)
     if device_op == "cpu":
-        y = np.max(
-            arr, axis=axis, out=out, keepdims=keepdims, initial=initial, where=where
-        )
+        kwargs = {"axis": axis, "out": out, "keepdims": keepdims, "where": where}
+
+        if initial is not _NoValue:
+            kwargs["initial"] = initial
+
+        y = np.max(arr, **kwargs)
     else:
         if cp is None:
             raise CuPyNotFound(CUPY_NOT_FOUND_MSG)

{depthtensor-2.4.2 → depthtensor-2.4.3}/src/DepthTensor/autodiff.py

@@ -59,6 +59,13 @@ def differentiate(tensor: Tensor, grad: TensorData | None = None) -> list[Tensor
         if grad is not None:
             _grad_check(grad, tensor)
             tensor.grad += grad
+        else:
+            if tensor.device == "gpu":
+                if cp is None:
+                    raise CuPyNotFound(CUPY_NOT_FOUND_MSG)
+                tensor.grad = cp.ones(tensor.shape, tensor.dtype)
+            elif tensor.device == "cpu":
+                tensor.grad = np.ones(tensor.shape, tensor.dtype)
 
     for t in reversed(topo):
         if t.backward is None:

{depthtensor-2.4.2 → depthtensor-2.4.3}/src/DepthTensor/typing.py

@@ -65,7 +65,12 @@ class BinaryOp(Protocol):
 
 class BinaryDiff(Protocol):
     def __call__(
-        self, result: "Tensor", x1: TensorData, x2: TensorData, **kwds: Any
+        self,
+        result: "Tensor",
+        x1: TensorData,
+        x2: TensorData,
+        device: Device,
+        **kwds: Any,
     ) -> tuple[Callable[[], TensorData], Callable[[], TensorData]]: ...
 
 
@@ -91,5 +96,5 @@ class UnaryOp(Protocol):
 
 class UnaryDiff(Protocol):
     def __call__(
-        self, result: "Tensor", x: TensorData, **kwds: Any
+        self, result: "Tensor", x: TensorData, device: Device, **kwds: Any
     ) -> Callable[[], TensorData]: ...

{depthtensor-2.4.2 → depthtensor-2.4.3}/src/depthtensor.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: depthtensor
-Version: 2.4.2
+Version: 2.4.3
 Summary: A Hardware-Accelerated Tensor Computation and Autograd Engine
 Author-email: Le Hong Ha <lehahong1310@gmail.com>
 License: MIT License