win: fix cuda build (#3532)

Co-authored-by: Cheng <git@zcbenz.com>

Author: dhiltgen

mlx/backend/cuda/CMakeLists.txt

@@ -122,7 +122,8 @@ target_compile_options(
 
 # Ignore warnings from CUTLASS.
 target_compile_options(
-  mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe="--diag_suppress=2908,2361">)
+  mlx
+  PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe="--diag_suppress=186,2908,2361">)
 
 if(NOT MSVC)
   # Required for generating optimized CUTLASS code.

mlx/backend/cuda/gemms/gather_gemm.cu

@@ -340,7 +340,7 @@ void gather_mm(
   dim3 num_blocks{uint32_t(ceil_div(m, size<0>(cta_tiler))),
                   uint32_t(ceil_div(n, size<1>(cta_tiler))),
                   uint32_t(l)};
-  dim3 block_dims{num_threads};
+  dim3 block_dims{uint32_t(num_threads)};
   void* args[] = {
       &shape_MNKL,
       &cta_tiler,

mlx/backend/cuda/quantized/qmm/qmm_naive.cu

@@ -148,7 +148,7 @@ void qmm_naive(
   dim3 num_blocks{uint32_t(ceil_div(m, size<0>(cta_tiler))),
                   uint32_t(ceil_div(n, size<1>(cta_tiler))),
                   uint32_t(l)};
-  dim3 block_dims{num_threads};
+  dim3 block_dims{uint32_t(num_threads)};
   void* args[] = {
       &shape_MNKL,
       &cta_tiler,

mlx/backend/cuda/quantized/qmm/qmm_sm80.cu

@@ -138,7 +138,7 @@ void qmm_sm80(
   dim3 num_blocks{uint32_t(ceil_div(m, size<0>(cta_tiler))),
                   uint32_t(ceil_div(n, size<1>(cta_tiler))),
                   uint32_t(l)};
-  dim3 block_dims{num_threads};
+  dim3 block_dims{uint32_t(num_threads)};
   void* args[] = {
       &shape_MNKL, &cta_tiler,
       &A, &dA,