[CUDA] --compress-mode requires CUDA 12.8 (#2407)

mlx/backend/cuda/CMakeLists.txt

@@ -88,7 +88,9 @@ endif()
 target_compile_options(
   mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--Wno-deprecated-gpu-targets>")
 
-if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.4.0)
+# Use stronger binaries compression. This feature was introduced in CUDA 12.8
+# and requires drivers released after CUDA 12.4.
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
   target_compile_options(
     mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--compress-mode=size>")
 endif()

mlx/primitives.cpp

@@ -1271,19 +1271,6 @@ std::vector<array> Convolution::vjp(
         has_neg_padding |= (pd < 0);
       }
 
-      auto padding_lo_ = std::vector<int>(padding_lo);
-      auto padding_hi_ = std::vector<int>(padding_hi);
-
-      // Use negative padding on the gradient output
-      if (has_neg_padding) {
-        for (auto& p : padding_lo_) {
-          p = std::max(0, p);
-        }
-        for (auto& p : padding_hi_) {
-          p = std::max(0, p);
-        }
-      }
-
       auto wt_trans = group_transpose(wt, 0, 1, -1);
       auto grad = conv_general(
           /* const array& input = */ cotan,
@@ -1305,12 +1292,9 @@ std::vector<array> Convolution::vjp(
         for (int i = 0; i < grad.ndim() - 2; i++) {
           if (padding_lo[i] < 0) {
             starts[i + 1] -= padding_lo[i];
-            padding_lo[i] = 0;
           }
-
           if (padding_hi[i] < 0) {
             stops[i + 1] += padding_hi[i];
-            padding_hi[i] = 0;
           }
         }
 

python/src/ops.cpp

@@ -4258,7 +4258,7 @@ void init_ops(nb::module_& m) {
 
         .. math::
 
-          w_i = s \hat{w_i} - \beta
+          w_i = s \hat{w_i} + \beta
 
         Args:
           w (array): Matrix to be quantized