From 32a16ef2f5261bbaa4d5d0e22631608c4a6aa5ad Mon Sep 17 00:00:00 2001
From: lvyufeng <lvyufeng@cqu.edu.cn>
Date: Sun, 21 Sep 2025 15:40:36 +0000
Subject: [PATCH] fix u-w class on GPU

---
 mindtorch/_apis/cpu.py     |  5 +++--
 mindtorch/_apis/gpu.py     | 13 ++++++++-----
 mindtorch/nn/functional.py |  3 +++
 mindtorch/ops/array.py     |  2 +-
 mindtorch/ops/other.py     |  3 +++
 5 files changed, 18 insertions(+), 8 deletions(-)

diff --git a/mindtorch/_apis/cpu.py b/mindtorch/_apis/cpu.py
index 85dc5e6d8..12fdda459 100644
--- a/mindtorch/_apis/cpu.py
+++ b/mindtorch/_apis/cpu.py
@@ -1089,7 +1089,7 @@ def leaky_relu(input, negative_slope):
     select_op = maximum
     if negative_slope > 1:
         select_op = minimum
-    return select_op(mul(negative_slope, input), input)
+    return select_op(mul(input, negative_slope), input)
 
 def ceil(input):
     return legacy.ceil(input)
@@ -1220,7 +1220,8 @@ def logsumexp(input, dim, keepdim=False):
     return add(input_logsumexp, input_max)
 
 def bernoulli(input, generator):
-    return legacy.bernoulli(input, seed, offset)
+    seed, offset = generator._step(12)  # pylint: disable=protected-access
+    return legacy.bernoulli(input, 0.5, seed.item(), offset.item())
 
 def right_shift(input, other):
     return legacy.right_shift(input, other)
diff --git a/mindtorch/_apis/gpu.py b/mindtorch/_apis/gpu.py
index 5affa4ca7..b59c8db14 100644
--- a/mindtorch/_apis/gpu.py
+++ b/mindtorch/_apis/gpu.py
@@ -746,9 +746,9 @@ def conv_transpose2d(input, weight, bias=None, stride=1, padding=0, output_paddi
     pad_mode = 'pad'
     pad = padding
     if isinstance(padding, tuple):
-        pad = (0, 0, padding[0], padding[0])
+        pad = (padding[0], padding[0], padding[1], padding[1])
     elif isinstance(padding, int):
-        pad = (0, 0) + (padding,) * 2
+        pad = (padding,) * 4
     if not isinstance(padding, (int, tuple)):
         pad_mode = padding
         pad = (0,) * 4
@@ -758,7 +758,6 @@ def conv_transpose2d(input, weight, bias=None, stride=1, padding=0, output_paddi
 
     in_channel, out_channels = weight.shape[0], weight.shape[1] * groups
     kernel_size = weight.shape[2:]
-
     n, _, h, w = input.shape
     h_add = _deconv_output_length(pad_mode, kernel_size[0], stride[0], dilation[0], pad[0] + pad[1])
     w_add = _deconv_output_length(pad_mode, kernel_size[1], stride[1], dilation[1], pad[2] + pad[3])
@@ -1004,7 +1003,7 @@ def leaky_relu(input, negative_slope):
     select_op = maximum
     if negative_slope > 1:
         select_op = minimum
-    return select_op(mul(negative_slope, input), input)
+    return select_op(mul(input, negative_slope), input)
 
 def ceil(input):
     return legacy.ceil(input)
@@ -1146,4 +1145,8 @@ def search_sorted(sorted_sequence, values, sorter, dtype, right):
     return legacy.search_sorted(sorted_sequence, values, sorter, dtype, right)
 
 def einsum(equation, operands):
-    return legacy.einsum(operands, equation)
\ No newline at end of file
+    return legacy.einsum(operands, equation)
+
+def unique2(input, sorted, return_inverse, return_counts):
+    outs = legacy.unique(input)
+    return outs + (None,)
diff --git a/mindtorch/nn/functional.py b/mindtorch/nn/functional.py
index 49dcac19d..ba128fd8a 100644
--- a/mindtorch/nn/functional.py
+++ b/mindtorch/nn/functional.py
@@ -47,6 +47,9 @@ def elu(input, alpha=1.0):
     return execute('elu', input, alpha)
 
 def glu(input, dim=-1):
+    if input.device.type == 'cuda':
+        x, y = input.chunk(2, dim)
+        return x * sigmoid(y)
     return execute('glu', input, dim)
 
 def softplus(input, beta=1, threshold=20):
diff --git a/mindtorch/ops/array.py b/mindtorch/ops/array.py
index 75fcd51de..fd2edd247 100644
--- a/mindtorch/ops/array.py
+++ b/mindtorch/ops/array.py
@@ -908,7 +908,7 @@ def range_(start, length):
         stacked_indices = indices[0]
     stacked_indices = stacked_indices.to(mindtorch.int32)
     stacked_indices = where(
-        stacked_indices < 0, stacked_indices + mindtorch.tensor(dim_sizes, device=stacked_indices.device), stacked_indices
+        stacked_indices < 0, stacked_indices + mindtorch.tensor(dim_sizes, dtype=stacked_indices.dtype, device=stacked_indices.device), stacked_indices
     )
     axis = dims[0]
     if len(dims) > 1:
diff --git a/mindtorch/ops/other.py b/mindtorch/ops/other.py
index f2098d395..a0a119bb7 100644
--- a/mindtorch/ops/other.py
+++ b/mindtorch/ops/other.py
@@ -27,6 +27,9 @@ def broadcast_tensors(*tensors):
 
 # broadcast_to
 def broadcast_to(input, shape):
+    if input.shape == shape:
+        return input
+
     new_shape = ()
     for s in shape:
         if not isinstance(s, int):