Remove unused variables (#706)

mlx/backend/accelerate/quantized.cpp

@@ -24,8 +24,6 @@ void _qmm_t_4_64(
   constexpr int bitmask = (1 << bits) - 1;
   constexpr int pack_factor = 32 / bits;
   constexpr int packs_in_group = group_size / pack_factor;
-  const int Kg = K / group_size;
-  const int Kw = K / pack_factor;
 
   for (int m = 0; m < M; m++) {
     const uint32_t* w_local = w;

mlx/backend/common/compiled.cpp

@@ -410,7 +410,6 @@ void Compiled::eval_cpu(
 
   // Get the kernel name from the lib
   int ndim = shape.size();
-  bool dynamic = ndim >= 8;
   auto kernel_name = kernel_lib_ + (contiguous ? "_contiguous" : "_strided_");
   if (!contiguous) {
     kernel_name += std::to_string(shape.size());

mlx/backend/metal/conv.cpp

@@ -182,7 +182,6 @@ void implicit_gemm_conv_2D_gpu(
 
   int implicit_M = conv_params.N * conv_params.oS[0] * conv_params.oS[1];
   int implicit_N = conv_params.O;
-  int implicit_K = conv_params.wS[0] * conv_params.wS[1] * conv_params.C;
 
   size_t grid_dim_x = (implicit_N + bn - 1) / bn;
   size_t grid_dim_y = (implicit_M + bm - 1) / bm;

mlx/backend/metal/indexing.cpp

@@ -167,10 +167,6 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto& upd = inputs.back();
   size_t nthreads = upd.size();
-  NS::UInteger thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
-  if (thread_group_size > nthreads) {
-    thread_group_size = nthreads;
-  }
 
   compute_encoder->setComputePipelineState(kernel);
 

mlx/backend/metal/primitives.cpp

@@ -691,7 +691,6 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   // organize into grid nkeys x elem_per_key
   MTL::Size grid_dims = MTL::Size(num_keys, half_size + odd, 1);
   NS::UInteger thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
-  auto nthreads = std::min(num_keys * (half_size + odd), thread_group_size);
   MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
   auto compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);

mlx/io/gguf_quants.cpp

@@ -114,7 +114,6 @@ void gguf_load_quantized(
         << "has incompatible last dim shape: " << shape[shape.size() - 1];
     throw std::runtime_error(msg.str());
   }
-  const uint64_t num_blocks = tensor.num_weights / weights_per_block;
 
   std::vector<int> weights_shape = shape;
   weights_shape.back() /= (weights_per_byte * 4);

mlx/primitives.cpp

@@ -628,7 +628,6 @@ std::vector<array> Convolution::vjp(
   auto& wt = primals[1];
   auto cotan = cotangents[0];
 
-  int N = in.shape(0);
   int O = wt.shape(0);
 
   // Resolve Padded input shapes and strides