patch bump for future version (#2804)

.github/actions/build-macos-release/action.yml

@@ -17,6 +17,8 @@ runs:
   steps:
     - name: Build Python package
       shell: bash -l {0}
+      env:
+        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
       run: |
         pip install build
         python setup.py clean --all
@@ -25,6 +27,8 @@ runs:
     - name: Build backend package
       if: ${{ inputs.build-backend }}
       shell: bash -l {0}
+      env:
+        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
       run: |
         python setup.py clean --all
         MLX_BUILD_STAGE=2 python -m build -w

.github/workflows/release.yml

@@ -23,7 +23,7 @@ jobs:
 
   build_documentation:
     if: github.repository == 'ml-explore/mlx'
-    runs-on: [self-hosted, macos]
+    runs-on: ubuntu-22.04
     steps:
       - uses: actions/checkout@v5
       - uses: ./.github/actions/build-docs
@@ -65,14 +65,14 @@ jobs:
         uses: actions/upload-artifact@v5
         with:
           overwrite: true
-          name: linux-wheels-${{ matrix.python_version }}
+          name: linux-wheels-${{ matrix.python_version }}-${{ matrix.arch }}
           path: wheelhouse/mlx-*.whl
       - name: Upload CPU artifacts
         if: matrix.python_version == '3.10'
         uses: actions/upload-artifact@v5
         with:
           overwrite: true
-          name: mlx-cpu
+          name: mlx-cpu-${{ matrix.arch }}
           path: wheelhouse/mlx_cpu-*.whl
   
   build_mac_release:
@@ -208,7 +208,8 @@ jobs:
     steps:
       - uses: actions/download-artifact@v6
         with:
-          name: mlx-cpu
+          pattern: mlx-cpu-*
+          merge-multiple: true
           path: dist
       - name: Display structure of downloaded files
         run: ls -R dist

mlx/backend/cuda/allocator.cpp

@@ -92,7 +92,7 @@ CudaAllocator::CudaAllocator()
           [this](CudaBuffer* buf) { cuda_free(buf); }) {
   size_t free, total;
   CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
-  memory_limit_ = total * 0.95;
+  memory_limit_ = total * 0.9;
   max_pool_size_ = memory_limit_;
 
   int device_count = 0;
@@ -176,7 +176,7 @@ CudaAllocator::malloc_async(size_t size, int device, cudaStream_t stream) {
     buffer_cache_.release_cached_buffers(get_cache_memory() - max_pool_size_);
   }
   // Copy to managed here if the buffer is not on the right device
-  if (buf->device != device) {
+  if (buf->device >= 0 && buf->device != device) {
     copy_to_managed(*buf);
   }
   return Buffer{buf};
@@ -219,9 +219,9 @@ void CudaAllocator::cuda_free(CudaBuffer* buf) {
     scalar_pool_.free(buf);
   } else {
     if (buf->device >= 0) {
-      cudaFreeAsync(buf->data, free_streams_[buf->device]);
+      CHECK_CUDA_ERROR(cudaFreeAsync(buf->data, free_streams_[buf->device]));
     } else {
-      cudaFree(buf->data);
+      CHECK_CUDA_ERROR(cudaFree(buf->data));
     }
     delete buf;
   }

mlx/backend/cuda/random.cu

@@ -139,10 +139,10 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   // keys has shape (N1, ..., NK, 2)
   // out has shape (N1, ..., NK, M1, M2, ...)
   auto& keys = inputs[0];
-  uint32_t num_keys = keys.size() / 2;
+  size_t num_keys = keys.size() / 2;
 
-  uint32_t elems_per_key = out.size() / num_keys;
-  uint32_t bytes_per_key = out.itemsize() * elems_per_key;
+  size_t elems_per_key = out.size() / num_keys;
+  size_t bytes_per_key = out.itemsize() * elems_per_key;
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
   out.set_data(cu::malloc_async(out.nbytes(), encoder));
@@ -150,19 +150,25 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
     return;
   }
 
-  uint32_t out_per_key = (bytes_per_key + 4 - 1) / 4;
-  uint32_t half_size = out_per_key / 2;
+  size_t out_per_key = (bytes_per_key + 4 - 1) / 4;
+  size_t half_size = out_per_key / 2;
+
   bool odd = out_per_key % 2;
+  if ((half_size + odd) >= UINT32_MAX || num_keys >= UINT32_MAX) {
+    throw std::runtime_error("[RandomBits::eval_gpu] Large size unsupported");
+  }
 
   encoder.set_input_array(keys);
   encoder.set_output_array(out);
-  dim3 grid_dims{num_keys, half_size + odd};
-  int64_t total = grid_dims.x * grid_dims.y;
-  int32_t threads_y = 1;
-  while ((total / threads_y) >= (1U << 31)) {
+  int64_t total = num_keys * (half_size + odd);
+  uint32_t threads_y = 1;
+  while ((total / threads_y) >= UINT_MAX) {
     threads_y *= 2;
   }
-  int32_t threads_x = cuda::ceil_div(total, threads_y);
+  uint32_t threads_x = cuda::ceil_div(total, threads_y);
+
+  dim3 grid_dims{
+      static_cast<uint32_t>(num_keys), static_cast<uint32_t>(half_size + odd)};
   auto [grid, block] = get_grid_and_block(threads_x, threads_y, 1);
   auto& stream = encoder.stream();
   if (keys.flags().row_contiguous) {

mlx/version.h

@@ -4,7 +4,7 @@
 
 #define MLX_VERSION_MAJOR 0
 #define MLX_VERSION_MINOR 30
-#define MLX_VERSION_PATCH 0
+#define MLX_VERSION_PATCH 1
 #define MLX_VERSION_NUMERIC \
   (100000 * MLX_VERSION_MAJOR + 1000 * MLX_VERSION_MINOR + MLX_VERSION_PATCH)
 

setup.py

@@ -24,8 +24,8 @@ def get_version():
             if "#define MLX_VERSION_PATCH" in l:
                 patch = l.split()[-1]
     version = f"{major}.{minor}.{patch}"
-    pypi_release = os.environ.get("PYPI_RELEASE", False)
-    dev_release = os.environ.get("DEV_RELEASE", False)
+    pypi_release = int(os.environ.get("PYPI_RELEASE", 0))
+    dev_release = int(os.environ.get("DEV_RELEASE", 0))
     if not pypi_release or dev_release:
         today = datetime.date.today()
         version = f"{version}.dev{today.year}{today.month:02d}{today.day:02d}"