From 2d7550a69bb0e8fc40326b94630fec57d9a6c7e5 Mon Sep 17 00:00:00 2001
From: Dhruv Kapur <37783178+dkapur17@users.noreply.github.com>
Date: Thu, 19 Feb 2026 11:28:19 -0500
Subject: [PATCH] Change reduce_value type from double to float

Fixes data type mismatch between the Python binding and CUDA code (which expects a float), potentially leading to nan gradients.
---
 minitorch/cuda_kernel_ops.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/minitorch/cuda_kernel_ops.py b/minitorch/cuda_kernel_ops.py
index a3acfad..a8a0281 100644
--- a/minitorch/cuda_kernel_ops.py
+++ b/minitorch/cuda_kernel_ops.py
@@ -173,7 +173,7 @@ def ret(a: Tensor, dim: int) -> Tensor:
                 np.ctypeslib.ndpointer(dtype=np.int32, ndim=1, flags='C_CONTIGUOUS'),    # in_shape
                 np.ctypeslib.ndpointer(dtype=np.int32, ndim=1, flags='C_CONTIGUOUS'),    # in_strides
                 ctypes.c_int,                                                            # reduce_dim
-                ctypes.c_double,                                                         # reduce_value
+                ctypes.c_float,                                                         # reduce_value
                 ctypes.c_int,                                                            # shape_len
                 ctypes.c_int,                                                            # fn_id
             ]
@@ -375,4 +375,4 @@ def matrix_multiply(a: Tensor, b: Tensor) -> Tensor:
         if more_3d:
             out = out.view(*ls)
             # print(f"Debug in matmul: output shape {out.shape}")
-        return out
\ No newline at end of file
+        return out