Add a test

Co-authored-by: Kanishk <kanishk.chores@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

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

@@ -20,7 +20,7 @@ runs:
       run: |
         pip install auditwheel "build<=1.4.2" patchelf setuptools
         python setup.py clean --all
-        MLX_DISABLE_SM90A_KERNELS=1 MLX_BUILD_STAGE=2 python -m build -w
+        MLX_BUILD_STAGE=2 python -m build -w
 
         auditwheel repair dist/mlx_cuda*.whl \
           --plat manylinux_2_35_${{ inputs.arch }} \

.github/actions/setup-linux/action.yml

@@ -14,6 +14,9 @@ inputs:
     description: 'Whether to enable ccache'
     required: false
     default: 'true'
+  ccache-key:
+    required: false
+    default: 'ccache'
 
 runs:
   using: "composite"
@@ -33,7 +36,7 @@ runs:
       if: ${{ inputs.use-ccache == 'true' }}
       uses: hendrikmuhs/ccache-action@v1.2
       with:
-        key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ inputs.toolkit }}
+        key: ${{ inputs.ccache-key }}-${{ runner.os }}-${{ runner.arch }}-${{ inputs.toolkit }}
         max-size: 1GB
         # ccache-action bug: running "apt-get update" fails on large arm runner.
         update-package-index: false

.github/workflows/nightly.yml

@@ -85,20 +85,24 @@ jobs:
 
   build_cuda_release:
     if: github.repository == 'ml-explore/mlx'
-    runs-on: ubuntu-22-large
+    strategy:
+      matrix:
+        arch: ['x86_64', 'aarch64']
+        toolkit: ['cuda-12.9', 'cuda-13.0']
+    runs-on: ${{ matrix.arch == 'x86_64' && 'ubuntu-22-large' || 'ubuntu-22-large-arm' }}
     steps:
       - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
         with:
-          toolkit: 'cuda-12.9'
+          toolkit: ${{ matrix.toolkit }}
+          ccache-key: 'ccache-release'
       - name: Build Python package
         uses: ./.github/actions/build-cuda-release
         with:
-          toolkit: 'cuda-12.9'
-          arch: 'x86_64'
+          arch: ${{ matrix.arch }}
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
-          name: mlx-cuda
+          name: mlx-${{ matrix.toolkit }}-${{ matrix.arch }}
           path: wheelhouse/mlx_cuda_*.whl
           retention-days: 7

.github/workflows/release.yml

@@ -146,7 +146,7 @@ jobs:
       - uses: ./.github/actions/setup-linux
         with:
           toolkit: ${{ matrix.toolkit }}
-          use-ccache: false
+          ccache-key: 'ccache-release'
       - name: Build Python package
         uses: ./.github/actions/build-cuda-release
         with:

docs/src/python/linalg.rst

@@ -14,6 +14,7 @@ Linear Algebra
     cholesky
     cholesky_inv
     cross
+    det
     qr
     svd
     eigvals
@@ -23,5 +24,6 @@ Linear Algebra
     lu
     lu_factor
     pinv
+    slogdet
     solve
     solve_triangular

mlx/backend/cpu/distributed.cpp

@@ -98,15 +98,6 @@ void Recv::eval_cpu(
 void ReduceScatter::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
-  assert(inputs.size() == 1);
-  assert(outputs.size() == 1);
-
-  auto [in, copied] = ensure_row_contiguous(inputs[0], stream());
-  outputs[0].set_data(allocator::malloc(outputs[0].nbytes()));
-  distributed::detail::sum_scatter(group(), in, outputs[0], stream());
-  if (copied) {
-    auto& enc = cpu::get_command_encoder(stream());
-    enc.add_temporary(in);
-  }
+  throw std::runtime_error("[ReduceScatter] Not implemented yet.");
 }
 } // namespace mlx::core::distributed

mlx/backend/cpu/luf.cpp

@@ -67,11 +67,10 @@ void luf_impl(
               /* ipiv */ reinterpret_cast<int*>(pivots_ptr),
               /* info */ &info);
 
-          if (info != 0) {
+          if (info < 0) {
             std::stringstream ss;
             ss << "[LUF::eval_cpu] sgetrf_ failed with code " << info
-               << ((info > 0) ? " because matrix is singular"
-                              : " because argument had an illegal value");
+               << " because argument had an illegal value";
             throw std::runtime_error(ss.str());
           }
 

mlx/backend/cuda/CMakeLists.txt

@@ -168,9 +168,8 @@ set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
                                      "${MLX_CUDA_ARCHITECTURES}")
 
 # Skip Hopper-only kernels when not building for sm90a.
-if(NOT DEFINED ENV{MLX_DISABLE_SM90A_KERNELS}
-   AND (("90a" IN_LIST MLX_CUDA_ARCHITECTURES) OR ("90a-real" IN_LIST
-                                                   MLX_CUDA_ARCHITECTURES)))
+if(("90a" IN_LIST MLX_CUDA_ARCHITECTURES) OR ("90a-real" IN_LIST
+                                              MLX_CUDA_ARCHITECTURES))
   target_compile_definitions(mlx PRIVATE MLX_CUDA_SM90A_ENABLED)
 endif()
 

mlx/backend/cuda/gemms/gather_gemm.cu

@@ -43,6 +43,12 @@ class GatherGemm {
   using ElementD = typename CollectiveEpilogue::ElementD;
   using StrideD = typename CollectiveEpilogue::StrideD;
 
+  static_assert(
+      cute::is_same_v<
+          ElementAccumulator,
+          typename CollectiveEpilogue::ElementAccumulator>,
+      "Mainloop and epilogue do not agree on accumulator value type.");
+
   static constexpr int SharedStorageSize = static_cast<int>(cute::max(
       sizeof(typename CollectiveMainloop::SharedStorage),
       sizeof(typename CollectiveEpilogue::SharedStorage)));
@@ -98,7 +104,9 @@ class GatherGemm {
 
   CUTLASS_DEVICE void operator()(const Params& params, char* smem_buf) {
     int thread_idx = int(threadIdx.x);
-    auto [m_coord, n_coord, l_coord] = uint3(blockIdx);
+    int m_coord = int(blockIdx.x);
+    int n_coord = int(blockIdx.y);
+    int l_coord = int(blockIdx.z);
 
     auto shape_MNKL = append<4>(params.problem_shape, Int<1>{});
     auto cta_tile = TileShape{};
@@ -220,7 +228,7 @@ void gather_mm(
       using TileShape = Shape<_128, _128, _8>;
       using DispatchPolicy = cutlass::gemm::MainloopSm70TwoStage;
       using TiledMma = TiledMMA<
-          MMA_Atom<UniversalFMA<Accumulator, Element, Element, Element>>,
+          MMA_Atom<UniversalFMA<Accumulator, Element, Element, Accumulator>>,
           Layout<Shape<_16, _16, _1>>>;
 
       using CopyTraitsA = SimtCopyTraits<Element, k_major_a.value>;
@@ -296,9 +304,6 @@ void cutlass_gather_mm(
   int n = out.shape(-1);
   int k = a.shape(-1);
   int l = out.size() / (m * n);
-  if (m < 16 || n < 16) {
-    throw std::invalid_argument("[gather_mm] M/N is too small.");
-  }
 
   encoder.set_input_array(a);
   encoder.set_input_array(b);

mlx/backend/cuda/gemms/grouped_gemm_unaligned.cu

@@ -245,7 +245,7 @@ void grouped_gemm_v2(
           LayoutB,
           cutlass::ComplexTransform::kNone,
           GemmConfiguration::kAlignmentAB,
-          typename GemmConfiguration::Element,
+          typename GemmConfiguration::Accumulator,
           cutlass::layout::RowMajor,
           typename GemmConfiguration::Accumulator,
           typename GemmConfiguration::OpClass,

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

@@ -52,7 +52,7 @@ bool supports_qmm_sm90(
   if (!biases) {
     return false;
   }
-  if (!x.flags().row_contiguous || !is_last_2_dims_row_contiguous(w) ||
+  if (!is_last_2_dims_row_contiguous(w) ||
       !is_last_2_dims_row_contiguous(scales) ||
       !is_last_2_dims_row_contiguous(*biases)) {
     return false;
@@ -139,7 +139,7 @@ bool supports_qmm_sm80(
   if ((n % 128 != 0) || (k % std::max(64, group_size) != 0)) {
     return false;
   }
-  if (!x.flags().row_contiguous || !is_last_2_dims_row_contiguous(w) ||
+  if (!is_last_2_dims_row_contiguous(w) ||
       !is_last_2_dims_row_contiguous(scales)) {
     return false;
   }
@@ -224,7 +224,7 @@ bool supports_qmm_naive(
   if (transpose && (k % std::max(64, group_size) != 0)) {
     return false;
   }
-  if (!x.flags().row_contiguous || !is_last_2_dims_row_contiguous(w) ||
+  if (!is_last_2_dims_row_contiguous(w) ||
       !is_last_2_dims_row_contiguous(scales)) {
     return false;
   }
@@ -265,7 +265,7 @@ void qmm_naive(
     if constexpr (k_major.value) {
       if (has_k_residue) {
         throw std::invalid_argument(
-            "[quantized_matmul] K must be multiples of group_size.");
+            "[quantized_matmul] K must be multiples of max(64, group_size).");
       }
       f.template operator()<false>();
     } else {
@@ -276,7 +276,8 @@ void qmm_naive(
   };
   int m = out.ndim() > 1 ? out.shape(-2) : 1;
   int k = x.shape(-1);
-  bool has_k_residue = k % group_size != 0;
+  int tile_k = std::max(64, group_size);
+  bool has_k_residue = k % tile_k != 0;
   bool sm80 = encoder.device().compute_capability_major() >= 8;
   dispatch_bool(transpose, [&](auto k_major) {
     dispatch_k(k_major, has_k_residue, [&]<bool HasKResidue>() {
@@ -342,7 +343,7 @@ bool supports_qmv(
   if (k % 8 != 0) {
     return false;
   }
-  if (!x.flags().row_contiguous || !is_last_2_dims_row_contiguous(w) ||
+  if (!is_last_2_dims_row_contiguous(w) ||
       !is_last_2_dims_row_contiguous(scales)) {
     return false;
   }

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

@@ -60,7 +60,9 @@ __global__ void qmm_naive_kernel(
   CUTE_STATIC_ASSERT_V(congruent(select<0,1,3>(shape_MNKL), dC));
 
   int thread_idx = int(threadIdx.x);
-  auto [m_coord, n_coord, l_coord] = static_cast<uint3>(blockIdx);
+  int m_coord = int(blockIdx.x);
+  int n_coord = int(blockIdx.y);
+  int l_coord = int(blockIdx.z);
 
   auto m_max_coord = size<0>(shape_MNKL) - size<0>(cta_tiler) * m_coord; // M - BLK_M * m_coord
   auto n_max_coord = size<1>(shape_MNKL) - size<1>(cta_tiler) * n_coord; // N - BLK_N * n_coord

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

@@ -48,7 +48,9 @@ __global__ void qmm_sm80_kernel(
   CUTE_STATIC_ASSERT_V(congruent(select<0,1,3>(shape_MNKL), dC));
 
   int thread_idx = int(threadIdx.x);
-  auto [m_coord, n_coord, l_coord] = static_cast<uint3>(blockIdx);
+  int m_coord = int(blockIdx.x);
+  int n_coord = int(blockIdx.y);
+  int l_coord = int(blockIdx.z);
 
   // For gather, use index lookup for input batch slicing.
   uint32_t a_batch = lhs_indices ? lhs_indices[l_coord] : l_coord;

mlx/backend/cuda/quantized/quantized.cpp

@@ -17,7 +17,7 @@ void QuantizedMatmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
 
-  const array& x = inputs[0];
+  array x = ensure_row_contiguous(inputs[0], encoder, s);
   const array& w = inputs[1];
   const array& scales = inputs[2];
   std::optional<array> biases;
@@ -146,15 +146,17 @@ void GatherQMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
 
-  const array& x = inputs[0];
+  array x = ensure_row_contiguous(inputs[0], encoder, s);
   const array& w = inputs[1];
   const array& scales = inputs[2];
   std::optional<array> biases;
   if (inputs.size() == 6) {
     biases = inputs[3];
   }
-  array lhs_indices = ensure_contiguous(inputs[inputs.size() - 2], encoder, s);
-  array rhs_indices = ensure_contiguous(inputs[inputs.size() - 1], encoder, s);
+  array lhs_indices =
+      ensure_row_contiguous(inputs[inputs.size() - 2], encoder, s);
+  array rhs_indices =
+      ensure_row_contiguous(inputs[inputs.size() - 1], encoder, s);
 
   int M = out.ndim() > 1 ? out.shape(-2) : 1;
   int N = out.shape(-1);

mlx/backend/metal/device_info.cpp

@@ -13,7 +13,12 @@ bool is_available() {
 }
 
 int device_count() {
-  return 1;
+  try {
+    metal::device(Device::gpu);
+    return 1;
+  } catch (...) {
+    return 0;
+  }
 }
 
 const std::unordered_map<std::string, std::variant<std::string, size_t>>&

mlx/backend/metal/kernels/fp_quantized.h

@@ -533,6 +533,7 @@ METAL_FUNC void fp_qvm_impl(
     device T* y,
     const int in_vec_size,
     const int out_vec_size,
+    const int in_vec_stride,
     uint3 tid [[threadgroup_position_in_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
@@ -563,7 +564,7 @@ METAL_FUNC void fp_qvm_impl(
   int out_col = pack_factor * tn * (tid.y * num_simdgroups + simd_gid);
   ws += out_col * bytes_per_pack / pack_factor + simd_lid * out_vec_size_w;
   scales += out_col / group_size + simd_lid * out_vec_size_g;
-  x += tid.x * in_vec_size + simd_lid;
+  x += tid.x * in_vec_stride + simd_lid;
   y += tid.x * out_vec_size + out_col;
 
   if (out_col >= out_vec_size) {
@@ -1122,7 +1123,16 @@ template <typename T, const int group_size, int bits, bool batched>
         tid);
   }
   fp_qvm_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
+      w,
+      scales,
+      x,
+      y,
+      in_vec_size,
+      out_vec_size,
+      in_vec_size,
+      tid,
+      simd_gid,
+      simd_lid);
 }
 
 template <typename T, const int group_size, int bits, int split_k = 32>
@@ -1164,8 +1174,20 @@ template <typename T, const int group_size, int bits, int split_k = 32>
   int in_vec_size_adj =
       tid.z % split_k == split_k - 1 ? final_block_size : in_vec_size;
 
+  // The in_vec_stride is the full K dimension, not the partition size
+  int in_vec_stride = (split_k - 1) * in_vec_size + final_block_size;
+
   fp_qvm_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size_adj, out_vec_size, tid, simd_gid, simd_lid);
+      w,
+      scales,
+      x,
+      y,
+      in_vec_size_adj,
+      out_vec_size,
+      in_vec_stride,
+      tid,
+      simd_gid,
+      simd_lid);
 }
 
 template <
@@ -1423,7 +1445,16 @@ template <typename T, int group_size, int bits>
       s_strides,
       tid);
   fp_qvm_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
+      w,
+      scales,
+      x,
+      y,
+      in_vec_size,
+      out_vec_size,
+      in_vec_size,
+      tid,
+      simd_gid,
+      simd_lid);
 }
 
 template <

mlx/backend/metal/kernels/indexing/scatter.h

@@ -80,7 +80,9 @@ template <
     uint3 gsize [[threads_per_grid]]) {
   Op op;
 
-  IdxT idx = IdxT(gid.z) * gsize.y + gid.y * gsize.x + gid.x * NWORK;
+  IdxT idx =
+      (IdxT(gid.z) * IdxT(gsize.y) + IdxT(gid.y)) * IdxT(gsize.x) * NWORK +
+      IdxT(gid.x) * NWORK;
   IdxT out_idx;
   IdxT update_idx;
 

mlx/backend/metal/kernels/quantized.h

@@ -983,6 +983,7 @@ METAL_FUNC void qvm_impl(
     device T* y,
     const int in_vec_size,
     const int out_vec_size,
+    const int in_vec_stride,
     uint3 tid [[threadgroup_position_in_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
@@ -1016,7 +1017,7 @@ METAL_FUNC void qvm_impl(
   ws += out_col * bytes_per_pack / pack_factor + simd_lid * out_vec_size_w;
   scales += out_col / group_size + simd_lid * out_vec_size_g;
   biases += out_col / group_size + simd_lid * out_vec_size_g;
-  x += tid.x * in_vec_size + simd_lid;
+  x += tid.x * in_vec_stride + simd_lid;
   y += tid.x * out_vec_size + out_col;
 
   if (out_col >= out_vec_size) {
@@ -1643,6 +1644,7 @@ template <typename T, const int group_size, const int bits, bool batched>
       y,
       in_vec_size,
       out_vec_size,
+      in_vec_size,
       tid,
       simd_gid,
       simd_lid);
@@ -1691,6 +1693,9 @@ template <typename T, const int group_size, const int bits, int split_k = 32>
   int in_vec_size_adj =
       tid.z % split_k == split_k - 1 ? final_block_size : in_vec_size;
 
+  // The in_vec_stride is the full K dimension, not the partition size
+  int in_vec_stride = (split_k - 1) * in_vec_size + final_block_size;
+
   qvm_impl<T, group_size, bits>(
       w,
       scales,
@@ -1699,6 +1704,7 @@ template <typename T, const int group_size, const int bits, int split_k = 32>
       y,
       in_vec_size_adj,
       out_vec_size,
+      in_vec_stride,
       tid,
       simd_gid,
       simd_lid);
@@ -2077,6 +2083,7 @@ template <typename T, int group_size, int bits>
       y,
       in_vec_size,
       out_vec_size,
+      in_vec_size,
       tid,
       simd_gid,
       simd_lid);

mlx/distributed/jaccl/jaccl.cpp

@@ -145,16 +145,7 @@ class JACCLGroup : public GroupImpl {
   }
 
   void sum_scatter(const array& input, array& output, Stream stream) override {
-    auto in_ptr = input.data<char>();
-    auto out_ptr = output.data<char>();
-    size_t n_bytes = input.nbytes();
-    int dtype = dtype_to_jaccl_dtype(output.dtype());
-    auto& encoder = cpu::get_command_encoder(stream);
-    encoder.set_input_array(input);
-    encoder.set_output_array(output);
-    encoder.dispatch([in_ptr, out_ptr, n_bytes, dtype, this]() {
-      group_->sum_scatter(in_ptr, out_ptr, n_bytes, dtype);
-    });
+    throw std::runtime_error("[jaccl] sum_scatter not supported.");
   }
 
   std::shared_ptr<GroupImpl> split(int color, int key = -1) override {

mlx/distributed/jaccl/lib/README.md

@@ -29,6 +29,8 @@ in macOS 26.2.
 - **Point-to-Point Operations**:
   - `send`: Send data to a specific node
   - `recv`: Receive data from a specific node
+- **Synchronization**:
+  - `barrier`: Block until all nodes in the group reach this point
 - **Type Support**: Bool, Int8-64, UInt8-64, Float16, BFloat16, Float32,
   Float64, Complex64
 
@@ -286,6 +288,9 @@ class Group {
   // Simple send/recv primitives.
   virtual void send(const void* input, size_t n_bytes, int dst) = 0;
   virtual void recv(void* output, size_t n_bytes, int src) = 0;
+
+  // Block until every rank reaches this point.
+  virtual void barrier() = 0;
 };
 ```
 

mlx/distributed/jaccl/lib/examples/CMakeLists.txt

@@ -35,6 +35,7 @@ endfunction()
 # Examples
 build_example(minimal_env.cpp)
 build_example(minimal_cfg.cpp)
+build_example(minimal_barrier.cpp)
 
 # Benchmarks
 build_example(allreduce_bench.cpp)

mlx/distributed/jaccl/lib/examples/minimal_barrier.cpp

@@ -0,0 +1,42 @@
+// Copyright © 2026 Apple Inc.
+//
+// Exercises Group::barrier(). Ranks arrive at the barrier at staggered times;
+// after the barrier returns we do a small all_sum to confirm the group is
+// healthy and that barrier() carried the correct fence semantics.
+
+#include <chrono>
+#include <iostream>
+#include <thread>
+
+#include <jaccl/jaccl.h>
+
+int main() {
+  auto group = jaccl::init();
+  if (!group) {
+    std::cerr << "Failed to initialize JACCL" << std::endl;
+    return 1;
+  }
+
+  int rank = group->rank();
+  int size = group->size();
+
+  std::this_thread::sleep_for(std::chrono::milliseconds(100 * rank));
+  std::cout << "rank " << rank << " entering barrier" << std::endl;
+
+  group->barrier();
+
+  std::cout << "rank " << rank << " exited barrier" << std::endl;
+
+  int in = rank + 1;
+  int out = 0;
+  group->all_sum(&in, &out, sizeof(in), jaccl::Int32);
+  int expected = size * (size + 1) / 2;
+  if (out != expected) {
+    std::cerr << "rank " << rank << ": post-barrier all_sum mismatch (got "
+              << out << ", expected " << expected << ")" << std::endl;
+    return 1;
+  }
+  std::cout << "rank " << rank << ": post-barrier all_sum OK (" << out << ")"
+            << std::endl;
+  return 0;
+}

mlx/distributed/jaccl/lib/jaccl/group.h

@@ -28,11 +28,9 @@ class Group {
 
   virtual void all_gather(const void* input, void* output, size_t n_bytes) = 0;
 
-  virtual void
-  sum_scatter(const void* input, void* output, size_t n_bytes, int dtype) = 0;
-
   virtual void send(const void* input, size_t n_bytes, int dst) = 0;
   virtual void recv(void* output, size_t n_bytes, int src) = 0;
+  virtual void barrier() = 0;
 };
 
 /**

mlx/distributed/jaccl/lib/jaccl/mesh.cpp

@@ -176,17 +176,6 @@ void MeshGroup::all_gather(const void* input, void* output, size_t n_bytes) {
       static_cast<const char*>(input), static_cast<char*>(output), n_bytes);
 }
 
-void MeshGroup::sum_scatter(
-    const void* input,
-    void* output,
-    size_t n_bytes,
-    int dtype) {
-  dispatch_all_types(dtype, [&](auto type_tag) {
-    using T = JACCL_GET_TYPE(type_tag);
-    reduce_scatter<T>(input, output, n_bytes, SumOp<T>{});
-  });
-}
-
 void MeshGroup::send(const void* input, size_t n_bytes, int dst) {
   mesh_.send(static_cast<const char*>(input), n_bytes, dst);
 }
@@ -195,16 +184,9 @@ void MeshGroup::recv(void* output, size_t n_bytes, int src) {
   mesh_.recv(static_cast<char*>(output), n_bytes, src);
 }
 
-template <typename T, typename ReduceOp>
-void MeshGroup::reduce_scatter(
-    const void* input,
-    void* output,
-    size_t n_bytes,
-    ReduceOp reduce_op) {
-  auto in_ptr = static_cast<const T*>(input);
-  auto out_ptr = static_cast<T*>(output);
-  int64_t count = n_bytes / sizeof(T);
-  mesh_.reduce_scatter(in_ptr, out_ptr, count, reduce_op);
+void MeshGroup::barrier() {
+  uint8_t b = 0;
+  all_sum(&b, &b, sizeof(b), Dtype::UInt8);
 }
 
 template <typename T, typename ReduceOp>

mlx/distributed/jaccl/lib/jaccl/mesh.h

@@ -44,20 +44,12 @@ class MeshGroup : public Group {
 
   void all_gather(const void* input, void* output, size_t n_bytes) override;
 
-  void sum_scatter(const void* input, void* output, size_t n_bytes, int dtype)
-      override;
-
   void send(const void* input, size_t n_bytes, int dst) override;
   void recv(void* output, size_t n_bytes, int src) override;
 
- private:
-  template <typename T, typename ReduceOp>
-  void reduce_scatter(
-      const void* input,
-      void* output,
-      size_t n_bytes,
-      ReduceOp reduce_op);
+  void barrier() override;
 
+ private:
   template <typename T, typename ReduceOp>
   void all_reduce(
       const void* input,

mlx/distributed/jaccl/lib/jaccl/mesh_impl.h

@@ -29,165 +29,6 @@ class MeshImpl {
 
   MeshImpl() : rank_(0), size_(1) {}
 
-  template <typename T, typename ReduceOp>
-  void reduce_scatter(const T* in, T* out, int64_t size, ReduceOp reduce_op) {
-    // Reduce-scatter for mesh topology.
-    //
-    // Each rank sends its entire input to all other ranks.
-    // Each rank reduces only its assigned chunk from all received inputs.
-
-    auto [sz, buffer_size] = buffer_size_from_message(size * sizeof(T));
-    int64_t N = buffer_size / sizeof(T);
-    constexpr int PIPELINE = 2;
-    constexpr int WC_NUM = PIPELINE * MESH_MAX_PEERS * 2;
-    int64_t total = static_cast<int64_t>(size);
-    int num_peers = size_ - 1;
-
-    // Calculate chunk for this rank
-    int64_t chunk_size = (total + size_ - 1) / size_;
-    int64_t my_chunk_start = rank_ * chunk_size;
-    int64_t my_chunk_size = std::min(chunk_size, total - my_chunk_start);
-
-    // Initialize output with our own chunk
-    if (my_chunk_size > 0) {
-      std::copy_n(in + my_chunk_start, my_chunk_size, out);
-    }
-
-    // A helper for convenient access to the staging buffer.
-    auto local_staging = [&](int buff) -> T* {
-      return reinterpret_cast<T*>(staging_mem_.get() + buff * MAX_BUFFER_SIZE);
-    };
-
-    // Counters to maintain the state of transfers
-    int in_flight = 0;
-    int64_t read_offset = 0;
-    int completed_send_count[PIPELINE] = {0};
-    int recv_end[MESH_MAX_PEERS] = {0};
-    int reduce_chunk = 0;
-    int reduce_rank = 0;
-
-    // Total number of chunks
-    int64_t total_chunks = (total + N - 1) / N;
-
-    // Prefill the pipeline
-    int buff = 0;
-    while (read_offset < total && buff < PIPELINE) {
-      post_recv_all(sz, buff);
-
-      // Copy the local data to send buffer and staging buffer
-      int64_t elems = std::min(N, total - read_offset);
-      std::copy(
-          in + read_offset, in + read_offset + elems, local_staging(buff));
-      std::copy(
-          in + read_offset,
-          in + read_offset + elems,
-          send_buffer(sz, buff).begin<T>());
-      recv_end[rank_]++;
-      post_send_all(sz, buff);
-
-      buff++;
-      in_flight += 2 * num_peers;
-      read_offset += N;
-    }
-
-    // Main loop
-    while (reduce_chunk < total_chunks) {
-      // Poll the hardware for completions.
-      ibv_wc wc[WC_NUM];
-      int n = poll(connections_, WC_NUM, wc);
-      for (int i = 0; i < n; i++) {
-        int work_type = wc[i].wr_id >> 16;
-        int buff = (wc[i].wr_id >> 8) & 0xff;
-        int rank = wc[i].wr_id & 0xff;
-
-        in_flight--;
-
-        if (work_type == SEND_WR && read_offset < total) {
-          completed_send_count[buff]++;
-          if (completed_send_count[buff] == num_peers) {
-            int64_t elems = std::min(N, total - read_offset);
-            std::copy(
-                in + read_offset,
-                in + read_offset + elems,
-                local_staging(buff));
-            std::copy(
-                in + read_offset,
-                in + read_offset + elems,
-                send_buffer(sz, buff).begin<T>());
-            recv_end[rank_]++;
-            post_send_all(sz, buff);
-
-            completed_send_count[buff] = 0;
-            in_flight += num_peers;
-            read_offset += N;
-          }
-        }
-
-        else if (work_type == RECV_WR) {
-          recv_end[rank]++;
-        }
-      }
-
-      // Process the received chunks in order, reducing only our chunk
-      while (reduce_chunk < total_chunks) {
-        int64_t w = static_cast<int64_t>(reduce_chunk) * N;
-        if (w >= read_offset) {
-          break;
-        }
-        if (recv_end[reduce_rank] <= reduce_chunk) {
-          break;
-        }
-        int b = reduce_chunk % PIPELINE;
-        int64_t elems = std::min(N, total - w);
-
-        // Check if this chunk overlaps with our output chunk
-        int64_t overlap_start = std::max(w, my_chunk_start);
-        int64_t overlap_end =
-            std::min(w + elems, my_chunk_start + my_chunk_size);
-
-        if (overlap_start < overlap_end) {
-          int64_t out_offset = overlap_start - my_chunk_start;
-          int64_t in_offset = overlap_start - w;
-          int64_t overlap_size = overlap_end - overlap_start;
-
-          // Data is read from the staging area for our own rank
-          if (reduce_rank == rank_) {
-            reduce_op(
-                local_staging(b) + in_offset, out + out_offset, overlap_size);
-          }
-          // Data is read from the recv buffers for other ranks
-          else {
-            reduce_op(
-                recv_buffer(sz, b, reduce_rank).begin<T>() + in_offset,
-                out + out_offset,
-                overlap_size);
-          }
-        }
-
-        // Check if we need to post another receive
-        int64_t next_chunk = static_cast<int64_t>(reduce_chunk) + PIPELINE;
-        if (next_chunk < total_chunks) {
-          recv_from(sz, reduce_rank, b);
-          in_flight++;
-        }
-
-        // Move to next rank's data for this chunk
-        reduce_rank++;
-        if (reduce_rank >= size_) {
-          reduce_rank = 0;
-          reduce_chunk++;
-        }
-      }
-    }
-
-    // Drain remaining in-flight completions
-    while (in_flight > 0) {
-      ibv_wc wc[WC_NUM];
-      int n = poll(connections_, WC_NUM, wc);
-      in_flight -= n;
-    }
-  }
-
   template <typename T, typename ReduceOp>
   void all_reduce(const T* in, T* out, int64_t size, ReduceOp reduce_op) {
     // Fully connected all reduce with deterministic reduction order.

mlx/distributed/jaccl/lib/jaccl/ring.cpp

@@ -166,17 +166,6 @@ void RingGroup::all_gather(const void* input, void* output, size_t n_bytes) {
       n_conns_);
 }
 
-void RingGroup::sum_scatter(
-    const void* input,
-    void* output,
-    size_t n_bytes,
-    int dtype) {
-  dispatch_all_types(dtype, [&](auto type_tag) {
-    using T = JACCL_GET_TYPE(type_tag);
-    reduce_scatter<T>(input, output, n_bytes, SumOp<T>{});
-  });
-}
-
 void RingGroup::send(const void* input, size_t n_bytes, int dst) {
   int right = (rank_ + 1) % size_;
   int left = (rank_ + size_ - 1) % size_;
@@ -201,27 +190,9 @@ void RingGroup::recv(void* output, size_t n_bytes, int src) {
   ring_.recv(static_cast<char*>(output), n_bytes, src, n_conns_);
 }
 
-template <typename T, typename ReduceOp>
-void RingGroup::reduce_scatter(
-    const void* input,
-    void* output,
-    size_t n_bytes,
-    ReduceOp reduce_op) {
-  auto in_ptr = static_cast<const T*>(input);
-  auto out_ptr = static_cast<T*>(output);
-  int64_t count = n_bytes / sizeof(T);
-  if (count < size_ * 2 * n_conns_) {
-    ring_.reduce_scatter<1, T, ReduceOp>(in_ptr, out_ptr, count, 1, reduce_op);
-    return;
-  }
-
-  if (n_bytes <= 65536) {
-    ring_.reduce_scatter<2, T, ReduceOp>(in_ptr, out_ptr, count, 1, reduce_op);
-    return;
-  }
-
-  ring_.reduce_scatter<2, T, ReduceOp>(
-      in_ptr, out_ptr, count, n_conns_, reduce_op);
+void RingGroup::barrier() {
+  uint8_t b = 0;
+  all_sum(&b, &b, sizeof(b), Dtype::UInt8);
 }
 
 template <typename T, typename ReduceOp>

mlx/distributed/jaccl/lib/jaccl/ring.h

@@ -45,20 +45,12 @@ class RingGroup : public Group {
 
   void all_gather(const void* input, void* output, size_t n_bytes) override;
 
-  void sum_scatter(const void* input, void* output, size_t n_bytes, int dtype)
-      override;
-
   void send(const void* input, size_t n_bytes, int dst) override;
   void recv(void* output, size_t n_bytes, int src) override;
 
- private:
-  template <typename T, typename ReduceOp>
-  void reduce_scatter(
-      const void* input,
-      void* output,
-      size_t n_bytes,
-      ReduceOp reduce_op);
+  void barrier() override;
 
+ private:
   template <typename T, typename ReduceOp>
   void all_reduce(
       const void* input,

mlx/distributed/jaccl/lib/jaccl/ring_impl.h

@@ -45,174 +45,6 @@ class RingImpl {
 
   RingImpl() : rank_(0), size_(1), n_conns_(0) {}
 
-  template <int MAX_DIR, typename T, typename ReduceOp>
-  void reduce_scatter(
-      const T* in_ptr,
-      T* out_ptr,
-      int64_t size,
-      int n_wires,
-      ReduceOp reduce_op) {
-    constexpr int PIPELINE = 2;
-    constexpr int WC_NUM = PIPELINE * RING_MAX_CONNS * 2 * MAX_DIR;
-    int64_t chunk_size = size / size_;
-    int64_t size_per_wire =
-        (chunk_size + (MAX_DIR * n_wires) - 1) / (MAX_DIR * n_wires);
-    auto [sz, N] = buffer_size_from_message(size_per_wire * sizeof(T));
-    N /= sizeof(T);
-    int64_t n_steps = (size_per_wire + N - 1) / N;
-
-    // Counters to maintain the state of transfers
-    int in_flight = 0;
-    int64_t send_offset[MAX_DIR];
-    int64_t recv_offset[MAX_DIR];
-    int64_t send_limits[MAX_DIR];
-    int64_t recv_limits[MAX_DIR];
-    int send_count[MAX_DIR * RING_MAX_CONNS] = {0};
-    int recv_count[MAX_DIR * RING_MAX_CONNS] = {0};
-    send_offset[0] = ((rank_ + size_ - 1) % size_) * chunk_size;
-    recv_offset[0] = ((rank_ + size_ - 2) % size_) * chunk_size;
-    if constexpr (MAX_DIR == 2) {
-      send_offset[1] = ((rank_ + 1) % size_) * chunk_size;
-      recv_offset[1] = ((rank_ + 2) % size_) * chunk_size;
-      send_limits[0] = std::min(
-          n_wires * size_per_wire, std::max<int64_t>(0, size - send_offset[0]));
-      send_limits[1] =
-          std::min(chunk_size, std::max<int64_t>(0, size - send_offset[1]));
-      recv_limits[0] = std::min(
-          n_wires * size_per_wire, std::max<int64_t>(0, size - recv_offset[0]));
-      recv_limits[1] =
-          std::min(chunk_size, std::max<int64_t>(0, size - recv_offset[1]));
-    } else {
-      send_limits[0] =
-          std::min(chunk_size, std::max<int64_t>(0, size - send_offset[0]));
-      recv_limits[0] =
-          std::min(chunk_size, std::max<int64_t>(0, size - recv_offset[0]));
-    }
-
-    for (int k = 0; k < size_ - 1; k++) {
-      const T* read_ptr = (k == 0) ? in_ptr : out_ptr;
-      int64_t read_offset[MAX_DIR];
-      read_offset[0] = (k == 0) ? send_offset[0] : 0;
-      if constexpr (MAX_DIR == 2) {
-        read_offset[1] = (k == 0) ? send_offset[1] : 0;
-      }
-
-      // Prefill the pipeline
-      int buff = 0;
-      while (buff < n_steps && buff < PIPELINE) {
-        // Post receives
-        post_recv_all<MAX_DIR>(sz, buff, n_wires);
-        in_flight += MAX_DIR * n_wires;
-
-        // Copy to send buffers and also to the output buffer
-        for (int lr = 0; lr < MAX_DIR; lr++) {
-          for (int lw = 0; lw < n_wires; lw++) {
-            // send
-            int64_t offset = lw * N +
-                send_count[lr * RING_MAX_CONNS + lw] * n_wires * N +
-                lr * n_wires * size_per_wire;
-            std::copy(
-                read_ptr + read_offset[lr] + offset,
-                read_ptr + read_offset[lr] +
-                    std::max(offset, std::min(offset + N, send_limits[lr])),
-                send_buffer(sz, buff, lr, lw).begin<T>());
-            send_count[lr * RING_MAX_CONNS + lw]++;
-            in_flight++;
-            send_to(sz, buff, lr, lw);
-
-            // copy for in-place recv reduce
-            offset -= n_wires * N;
-            std::copy(
-                in_ptr + recv_offset[lr] + offset,
-                in_ptr + recv_offset[lr] +
-                    std::max(offset, std::min(offset + N, recv_limits[lr])),
-                out_ptr + offset);
-          }
-        }
-
-        buff++;
-      }
-
-      while (in_flight > 0) {
-        ibv_wc wc[WC_NUM];
-        int n = poll(left_, right_, WC_NUM, wc);
-        for (int i = 0; i < n; i++) {
-          int work_type = wc[i].wr_id >> 16;
-          int buff = (wc[i].wr_id >> 8) & 0xff;
-          int wire = wc[i].wr_id & 0xff;
-          int lr = wire / RING_MAX_CONNS;
-          int lw = wire % RING_MAX_CONNS;
-
-          in_flight--;
-
-          if (work_type == SEND_WR) {
-            int64_t offset = lw * N + send_count[wire] * n_wires * N +
-                lr * n_wires * size_per_wire;
-
-            // More stuff to send
-            if (send_count[wire] < n_steps) {
-              std::copy(
-                  read_ptr + read_offset[lr] + offset,
-                  read_ptr + read_offset[lr] +
-                      std::max(offset, std::min(offset + N, send_limits[lr])),
-                  send_buffer(sz, buff, lr, lw).begin<T>());
-              send_to(sz, buff, lr, lw);
-              in_flight++;
-              send_count[wire]++;
-            }
-
-            // Copy the input chunk into the output
-            offset -= n_wires * N;
-            std::copy(
-                in_ptr + recv_offset[lr] + offset,
-                in_ptr + recv_offset[lr] +
-                    std::max(offset, std::min(offset + N, recv_limits[lr])),
-                out_ptr + offset);
-          }
-
-          else if (work_type == RECV_WR) {
-            int64_t offset = lw * N + recv_count[wire] * n_wires * N +
-                lr * n_wires * size_per_wire;
-            reduce_op(
-                recv_buffer(sz, buff, lr, lw).begin<T>(),
-                out_ptr + recv_offset[lr] + offset,
-                std::max<int64_t>(0, std::min(N, recv_limits[lr] - offset)));
-            recv_count[wire]++;
-            if (recv_count[wire] + (PIPELINE - 1) < n_steps) {
-              recv_from(sz, buff, lr, lw);
-              in_flight++;
-            }
-          }
-        }
-      }
-
-      send_offset[0] = (send_offset[0] + size - chunk_size) % size;
-      recv_offset[0] = (recv_offset[0] + size - chunk_size) % size;
-      if constexpr (MAX_DIR == 2) {
-        send_offset[1] = (send_offset[1] + chunk_size) % size;
-        recv_offset[1] = (recv_offset[1] + chunk_size) % size;
-        send_limits[0] = std::min(
-            n_wires * size_per_wire,
-            std::max<int64_t>(0, size - send_offset[0]));
-        send_limits[1] =
-            std::min(chunk_size, std::max<int64_t>(0, size - send_offset[1]));
-        recv_limits[0] = std::min(
-            n_wires * size_per_wire,
-            std::max<int64_t>(0, size - recv_offset[0]));
-        recv_limits[1] =
-            std::min(chunk_size, std::max<int64_t>(0, size - recv_offset[1]));
-      } else {
-        send_limits[0] =
-            std::min(chunk_size, std::max<int64_t>(0, size - send_offset[0]));
-        recv_limits[0] =
-            std::min(chunk_size, std::max<int64_t>(0, size - recv_offset[0]));
-      }
-      for (int i = 0; i < MAX_DIR * RING_MAX_CONNS; i++) {
-        send_count[i] = recv_count[i] = 0;
-      }
-    }
-  }
-
   template <int MAX_DIR, typename T, typename ReduceOp>
   void all_reduce(
       const T* in_ptr,

mlx/io/CMakeLists.txt

@@ -17,6 +17,11 @@ if(MLX_BUILD_GGUF)
                              PRIVATE $<BUILD_INTERFACE:${gguflib_SOURCE_DIR}>)
   add_library(gguflib STATIC ${gguflib_SOURCE_DIR}/fp16.c
                              ${gguflib_SOURCE_DIR}/gguflib.c)
+  # gguflib uses assert() to reject malformed tensor headers (e.g. ndim > 8).
+  # Those checks are otherwise compiled out by -DNDEBUG in release builds, which
+  # leaves out-of-bounds reads/writes unguarded when loading untrusted GGUF
+  # files. Force NDEBUG off for this target so the asserts stay live.
+  target_compile_options(gguflib PRIVATE -UNDEBUG)
   target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:gguflib>)
   target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gguf.cpp
                              ${CMAKE_CURRENT_SOURCE_DIR}/gguf_quants.cpp)

mlx/io/safetensors.cpp

@@ -28,6 +28,7 @@ using json = nlohmann::json;
 #define ST_U32 "U32"
 #define ST_U64 "U64"
 #define ST_F8_E4M3 "F8_E4M3"
+#define ST_F8_E8M0 "F8_E8M0"
 
 // Note: Complex numbers aren't in the spec yet so this could change -
 // https://github.com/huggingface/safetensors/issues/389
@@ -97,6 +98,8 @@ Dtype dtype_from_safetensor_str(std::string_view str) {
     return complex64;
   } else if (str == ST_F8_E4M3) {
     return uint8;
+  } else if (str == ST_F8_E8M0) {
+    return uint8;
   } else {
     std::ostringstream msg;
     msg << "[safetensor] unsupported dtype" << str;

mlx/linalg.cpp

@@ -705,4 +705,170 @@ array solve_triangular(
   return matmul(a_inv, b, s);
 }
 
-} // namespace mlx::core::linalg
+void validate_det(
+    const array& a,
+    const StreamOrDevice& stream,
+    const std::string& fname) {
+  check_cpu_stream(stream, fname);
+  if (issubdtype(a.dtype(), complexfloating)) {
+    throw std::invalid_argument(fname + " Complex inputs are not supported.");
+  }
+  if (a.ndim() < 2) {
+    std::ostringstream msg;
+    msg << fname
+        << " Arrays must have >= 2 dimensions. Received array "
+           "with "
+        << a.ndim() << " dimensions.";
+    throw std::invalid_argument(msg.str());
+  }
+
+  if (a.shape(-1) != a.shape(-2)) {
+    throw std::invalid_argument(fname + " Only defined for square matrices.");
+  }
+}
+
+array det_raw_small(const array& a, StreamOrDevice s) {
+  int n = a.shape(-1);
+
+  // Empty 0x0 matrix: determinant is the empty product = 1
+  if (n == 0) {
+    Shape out_shape(a.shape().begin(), a.shape().end() - 2);
+    return broadcast_to(array(1.0f, a.dtype()), std::move(out_shape), s);
+  }
+
+  // Helper to extract a[..., i, j] from the last two dims
+  auto elem = [&](int i, int j) {
+    auto starts = Shape(a.ndim(), 0);
+    auto stops = a.shape();
+    starts[a.ndim() - 2] = i;
+    stops[a.ndim() - 2] = i + 1;
+    starts[a.ndim() - 1] = j;
+    stops[a.ndim() - 1] = j + 1;
+    return squeeze(squeeze(slice(a, starts, stops, s), -1, s), -1, s);
+  };
+
+  if (n == 1) {
+    return elem(0, 0);
+  } else if (n == 2) {
+    return subtract(
+        multiply(elem(0, 0), elem(1, 1), s),
+        multiply(elem(0, 1), elem(1, 0), s),
+        s);
+  } else {
+    // 3x3: a00*(a11*a22 - a12*a21) - a01*(a10*a22 - a12*a20) + a02*(a10*a21 -
+    // a11*a20)
+    auto a00 = elem(0, 0), a01 = elem(0, 1), a02 = elem(0, 2);
+    auto a10 = elem(1, 0), a11 = elem(1, 1), a12 = elem(1, 2);
+    auto a20 = elem(2, 0), a21 = elem(2, 1), a22 = elem(2, 2);
+    return add(
+        subtract(
+            multiply(
+                a00,
+                subtract(multiply(a11, a22, s), multiply(a12, a21, s), s),
+                s),
+            multiply(
+                a01,
+                subtract(multiply(a10, a22, s), multiply(a12, a20, s), s),
+                s),
+            s),
+        multiply(
+            a02, subtract(multiply(a10, a21, s), multiply(a11, a20, s), s), s),
+        s);
+  }
+}
+
+std::pair<array, array> slogdet_impl(const array& input, StreamOrDevice s) {
+  int n = input.shape(-1);
+  auto dtype = input.dtype();
+
+  // Small-matrix fast path
+  if (n <= 3) {
+    auto raw = det_raw_small(input, s);
+    auto abs_raw = abs(raw, s);
+    auto sgn = sign(raw, s);
+    auto logabs = log(abs_raw, s);
+    return std::make_pair(sgn, logabs);
+  }
+
+  // General LU-based path
+  auto [LU, pivots] = lu_factor(input, s);
+
+  // Extract diagonal of U
+  auto diag = diagonal(LU, 0, -2, -1, s);
+
+  // Permutation parity: count positions where pivot[i] != i
+  int k = std::min(input.shape(-2), input.shape(-1));
+  auto iota = arange(0, k, uint32, s);
+  auto parity = astype(
+      sum(not_equal(pivots, iota, s),
+          /* axis = */ -1,
+          /* keepdims = */ false,
+          s),
+      int32,
+      s);
+
+  // Count negative diagonal elements
+  auto num_neg = astype(
+      sum(less(diag, array(0.0f, dtype), s),
+          /* axis = */ -1,
+          /* keepdims = */ false,
+          s),
+      int32,
+      s);
+
+  // sign = (-1)^(parity + num_neg)
+  auto total = add(parity, num_neg, s);
+  auto sign_val = astype(
+      subtract(
+          array(1, int32),
+          multiply(array(2, int32), remainder(total, array(2, int32), s), s),
+          s),
+      dtype,
+      s);
+
+  // logabsdet = sum(log(abs(diag)))
+  auto logabsdet =
+      sum(log(abs(diag, s), s), /* axis = */ -1, /* keepdims = */ false, s);
+
+  // Handle singular matrices: any zero on diagonal
+  auto is_zero =
+      any(equal(diag, array(0.0f, dtype), s),
+          /* axis = */ -1,
+          /* keepdims = */ false,
+          s);
+  sign_val = where(is_zero, array(0.0f, dtype), sign_val, s);
+  logabsdet = where(
+      is_zero,
+      array(-std::numeric_limits<float>::infinity(), dtype),
+      logabsdet,
+      s);
+
+  return std::make_pair(sign_val, logabsdet);
+}
+
+std::pair<array, array> slogdet(const array& a, StreamOrDevice s /* = {} */) {
+  validate_det(a, s, "[linalg::slogdet]");
+
+  auto dtype = at_least_float(a.dtype());
+  auto input = astype(a, dtype, s);
+  return slogdet_impl(input, s);
+}
+
+array det(const array& a, StreamOrDevice s /* = {} */) {
+  validate_det(a, s, "[linalg::det]");
+
+  auto dtype = at_least_float(a.dtype());
+  auto input = astype(a, dtype, s);
+  int n = input.shape(-1);
+
+  // Small-matrix fast path: compute directly, skip log/exp round-trip
+  if (n <= 3) {
+    return det_raw_small(input, s);
+  }
+
+  // General case: det = sign * exp(logabsdet)
+  auto [sign_val, logabsdet] = slogdet_impl(input, s);
+  return multiply(sign_val, exp(logabsdet, s), s);
+}
+
+} // namespace mlx::core::linalg

mlx/linalg.h

@@ -112,4 +112,8 @@ eigvalsh(const array& a, std::string UPLO = "L", StreamOrDevice s = {});
 MLX_API std::pair<array, array>
 eigh(const array& a, std::string UPLO = "L", StreamOrDevice s = {});
 
+MLX_API array det(const array& a, StreamOrDevice s = {});
+
+MLX_API std::pair<array, array> slogdet(const array& a, StreamOrDevice s = {});
+
 } // namespace mlx::core::linalg

mlx/version.h

@@ -5,8 +5,8 @@
 #include "mlx/api.h"
 
 #define MLX_VERSION_MAJOR 0
-#define MLX_VERSION_MINOR 31
-#define MLX_VERSION_PATCH 2
+#define MLX_VERSION_MINOR 32
+#define MLX_VERSION_PATCH 0
 #define MLX_VERSION_NUMERIC \
   (100000 * MLX_VERSION_MAJOR + 1000 * MLX_VERSION_MINOR + MLX_VERSION_PATCH)
 

python/src/convert.cpp

@@ -1,5 +1,8 @@
 // Copyright © 2024 Apple Inc.
 
+#include <limits>
+#include <sstream>
+
 #include <nanobind/stl/complex.h>
 
 #include "python/src/convert.h"
@@ -15,9 +18,15 @@ enum PyScalarT {
 };
 
 int check_shape_dim(int64_t dim) {
-  if (dim > std::numeric_limits<int>::max()) {
-    throw std::invalid_argument(
-        "Shape dimension falls outside supported `int` range.");
+  if (dim > std::numeric_limits<int>::max() ||
+      dim < std::numeric_limits<int>::min()) {
+    std::ostringstream msg;
+    msg << "Shape dimension " << dim << " is outside the supported range ["
+        << std::numeric_limits<int>::min() << ", "
+        << std::numeric_limits<int>::max()
+        << "]. MLX currently uses 32-bit integers for shape dimensions.";
+    PyErr_SetString(PyExc_OverflowError, msg.str().c_str());
+    nb::detail::raise_python_error();
   }
   return static_cast<int>(dim);
 }

python/src/convert.h

@@ -76,3 +76,7 @@ nb::object tolist(mx::array& a);
 mx::array create_array(nb::object v, std::optional<mx::Dtype> t);
 mx::array array_from_list(nb::list pl, std::optional<mx::Dtype> dtype);
 mx::array array_from_list(nb::tuple pl, std::optional<mx::Dtype> dtype);
+
+// Narrow a Python-side shape dimension (int64) to a C++ mx::ShapeElem (int32),
+// raising a clear error if the value would overflow.
+int check_shape_dim(int64_t dim);

python/src/linalg.cpp

@@ -660,4 +660,77 @@ void init_linalg(nb::module_& parent_module) {
         Returns:
             array: The unique solution to the system ``AX = B``.
       )pbdoc");
+
+  m.def(
+      "det",
+      &mx::linalg::det,
+      "a"_a,
+      nb::kw_only(),
+      "stream"_a = nb::none(),
+      nb::sig(
+          "def det(a: array, *, stream: Union[None, Stream, Device] = None) -> array"),
+      R"pbdoc(
+        Compute the determinant of a square matrix.
+
+        This function supports arrays with at least 2 dimensions. When the
+        input has more than two dimensions, the determinant is computed for
+        each matrix in the last two dimensions.
+
+        Args:
+            a (array): Input array.
+            stream (Stream, optional): Stream or device. Defaults to ``None``
+              in which case the default stream of the default device is used.
+
+        Returns:
+            array: The determinant(s) of the input matrix (matrices).
+
+        Example:
+            >>> A = mx.array([[1., 2.], [3., 4.]])
+            >>> mx.linalg.det(A, stream=mx.cpu)
+            array(-2, dtype=float32)
+      )pbdoc");
+
+  m.def(
+      "slogdet",
+      [](const mx::array& a, mx::StreamOrDevice s) {
+        auto result = mx::linalg::slogdet(a, s);
+        return nb::make_tuple(result.first, result.second);
+      },
+      "a"_a,
+      nb::kw_only(),
+      "stream"_a = nb::none(),
+      nb::sig(
+          "def slogdet(a: array, *, stream: Union[None, Stream, Device] = None) -> Tuple[array, array]"),
+      R"pbdoc(
+        Compute the sign and natural log of the absolute value of the
+        determinant of a square matrix.
+
+        This function supports arrays with at least 2 dimensions. When the
+        input has more than two dimensions, the sign and log-absolute-determinant
+        are computed for each matrix in the last two dimensions.
+
+        For a singular matrix, ``sign`` is 0 and ``logabsdet`` is ``-inf``.
+
+        The determinant can be reconstructed as ``det = sign * exp(logabsdet)``.
+        This is more numerically stable than computing the determinant directly
+        for matrices with large or small determinants.
+
+        Args:
+            a (array): Input array.
+            stream (Stream, optional): Stream or device. Defaults to ``None``
+              in which case the default stream of the default device is used.
+
+        Returns:
+            tuple(array, array): The ``sign`` and ``logabsdet`` of the
+              determinant. ``sign`` is -1, 0, or +1. ``logabsdet`` is the
+              natural log of the absolute value of the determinant.
+
+        Example:
+            >>> A = mx.array([[1., 2.], [3., 4.]])
+            >>> sign, logabsdet = mx.linalg.slogdet(A, stream=mx.cpu)
+            >>> sign
+            array(-1, dtype=float32)
+            >>> logabsdet
+            array(0.693147, dtype=float32)
+      )pbdoc");
 }

python/src/ops.cpp

@@ -15,6 +15,7 @@
 #include "mlx/einsum.h"
 #include "mlx/ops.h"
 #include "mlx/utils.h"
+#include "python/src/convert.h"
 #include "python/src/load.h"
 #include "python/src/small_vector.h"
 #include "python/src/utils.h"
@@ -45,6 +46,13 @@ double scalar_to_double(Scalar s) {
   }
 }
 
+mx::Shape to_shape(const nb::object& shape) {
+  if (nb::isinstance<nb::int_>(shape)) {
+    return {check_shape_dim(nb::cast<int64_t>(shape))};
+  }
+  return nb::cast<mx::Shape>(shape);
+}
+
 void init_ops(nb::module_& m) {
   m.def(
       "reshape",
@@ -1702,15 +1710,11 @@ void init_ops(nb::module_& m) {
       )pbdoc");
   m.def(
       "full",
-      [](const std::variant<int, mx::Shape>& shape,
+      [](const nb::object& shape,
          const ScalarOrArray& vals,
          std::optional<mx::Dtype> dtype,
          mx::StreamOrDevice s) {
-        if (auto pv = std::get_if<int>(&shape); pv) {
-          return mx::full({*pv}, to_array(vals, dtype), s);
-        } else {
-          return mx::full(std::get<mx::Shape>(shape), to_array(vals, dtype), s);
-        }
+        return mx::full(to_shape(shape), to_array(vals, dtype), s);
       },
       "shape"_a,
       "vals"_a,
@@ -1736,15 +1740,11 @@ void init_ops(nb::module_& m) {
       )pbdoc");
   m.def(
       "zeros",
-      [](const std::variant<int, mx::Shape>& shape,
+      [](const nb::object& shape,
          std::optional<mx::Dtype> dtype,
          mx::StreamOrDevice s) {
         auto t = dtype.value_or(mx::float32);
-        if (auto pv = std::get_if<int>(&shape); pv) {
-          return mx::zeros({*pv}, t, s);
-        } else {
-          return mx::zeros(std::get<mx::Shape>(shape), t, s);
-        }
+        return mx::zeros(to_shape(shape), t, s);
       },
       "shape"_a,
       "dtype"_a.none() = mx::float32,
@@ -1802,15 +1802,11 @@ void init_ops(nb::module_& m) {
       )pbdoc");
   m.def(
       "ones",
-      [](const std::variant<int, mx::Shape>& shape,
+      [](const nb::object& shape,
          std::optional<mx::Dtype> dtype,
          mx::StreamOrDevice s) {
         auto t = dtype.value_or(mx::float32);
-        if (auto pv = std::get_if<int>(&shape); pv) {
-          return mx::ones({*pv}, t, s);
-        } else {
-          return mx::ones(std::get<mx::Shape>(shape), t, s);
-        }
+        return mx::ones(to_shape(shape), t, s);
       },
       "shape"_a,
       "dtype"_a.none() = mx::float32,

python/src/small_vector.h

@@ -2,6 +2,11 @@
 
 #pragma once
 
+#include <cstdint>
+#include <limits>
+#include <sstream>
+#include <type_traits>
+
 #include "mlx/small_vector.h"
 
 #include <nanobind/stl/detail/nb_list.h>
@@ -14,11 +19,19 @@ struct type_caster<mlx::core::SmallVector<Type, Size, Alloc>> {
   using List = mlx::core::SmallVector<Type, Size, Alloc>;
   using Caster = make_caster<Type>;
 
+  // For narrow integer element types we fetch each element through a wider
+  // integer caster so we can emit a clean OverflowError on overflow instead of
+  // nanobind's generic "incompatible function arguments" TypeError.
+  static constexpr bool kNarrowInt = std::is_integral_v<Type> &&
+      !std::is_same_v<Type, bool> && (sizeof(Type) < sizeof(int64_t));
+
   NB_TYPE_CASTER(
       List,
       const_name("tuple[") + make_caster<Type>::Name + const_name(", ...]"))
 
-  bool from_python(handle src, uint8_t flags, cleanup_list* cleanup) noexcept {
+  // Not noexcept: on overflow of a narrow integer element we raise
+  // OverflowError so nanobind surfaces a clean error to the user.
+  bool from_python(handle src, uint8_t flags, cleanup_list* cleanup) {
     size_t size;
     PyObject* temp;
 
@@ -29,19 +42,39 @@ struct type_caster<mlx::core::SmallVector<Type, Size, Alloc>> {
     value.clear();
     value.reserve(size);
 
-    Caster caster;
     bool success = o != nullptr;
 
     flags = flags_for_local_caster<Type>(flags);
 
     for (size_t i = 0; i < size; ++i) {
-      if (!caster.from_python(o[i], flags, cleanup) ||
-          !caster.template can_cast<Type>()) {
-        success = false;
-        break;
+      if constexpr (kNarrowInt) {
+        make_caster<int64_t> wide;
+        if (!wide.from_python(o[i], flags, cleanup) ||
+            !wide.template can_cast<int64_t>()) {
+          success = false;
+          break;
+        }
+        int64_t v = wide.operator cast_t<int64_t>();
+        if (v > std::numeric_limits<Type>::max() ||
+            v < std::numeric_limits<Type>::min()) {
+          std::ostringstream msg;
+          msg << "Integer value " << v << " is outside the supported range ["
+              << static_cast<int64_t>(std::numeric_limits<Type>::min()) << ", "
+              << static_cast<int64_t>(std::numeric_limits<Type>::max()) << "].";
+          Py_XDECREF(temp);
+          PyErr_SetString(PyExc_OverflowError, msg.str().c_str());
+          raise_python_error();
+        }
+        value.push_back(static_cast<Type>(v));
+      } else {
+        Caster caster;
+        if (!caster.from_python(o[i], flags, cleanup) ||
+            !caster.template can_cast<Type>()) {
+          success = false;
+          break;
+        }
+        value.push_back(caster.operator cast_t<Type>());
       }
-
-      value.push_back(caster.operator cast_t<Type>());
     }
 
     Py_XDECREF(temp);

python/tests/cuda_skip.py

@@ -1,22 +1,5 @@
 cuda_skip = {
-    # Lapack ops NYI
-    "TestLinalg.test_cholesky",
-    "TestLinalg.test_cholesky_inv",
-    "TestLinalg.test_eig",
-    "TestLinalg.test_eigh",
-    "TestLinalg.test_inverse",
-    "TestVmap.test_vmap_inverse",
-    "TestLinalg.test_lu",
-    "TestLinalg.test_lu_factor",
-    "TestLinalg.test_pseudo_inverse",
-    "TestLinalg.test_qr_factorization",
-    "TestInit.test_orthogonal",
-    "TestLinalg.test_svd_decomposition",
-    "TestVmap.test_vmap_svd",
-    "TestLinalg.test_tri_inverse",
     # Quantization NYI
     "TestQuantized.test_gather_matmul_grad",
     "TestQuantized.test_gather_qmm",
-    "TestQuantized.test_gather_qmm_sorted",
-    "TestQuantized.test_gather_qmm_grad",
 }

python/tests/test_array.py

@@ -767,10 +767,13 @@ class TestArray(mlx_tests.MLXTestCase):
 
     def test_array_np_shape_dim_check(self):
         a_npy = np.empty(2**31, dtype=np.bool_)
-        with self.assertRaises(ValueError) as e:
+        with self.assertRaises(OverflowError) as e:
             mx.array(a_npy)
         self.assertEqual(
-            str(e.exception), "Shape dimension falls outside supported `int` range."
+            str(e.exception),
+            "Shape dimension 2147483648 is outside the supported range "
+            "[-2147483648, 2147483647]. MLX currently uses 32-bit integers "
+            "for shape dimensions.",
         )
 
     def test_dtype_promotion(self):
@@ -1547,6 +1550,30 @@ class TestArray(mlx_tests.MLXTestCase):
         a = a.at[:, :, :].add(update)
         self.assertEqualArray(a, update)
 
+    def test_slice_update_contiguous_2d(self):
+        for shape in [(32, 32), (64, 64), (17, 33), (128, 128)]:
+            for upd_shape in [(16, 16), (8, 8), (4, 4)]:
+                if upd_shape[0] > shape[0] or upd_shape[1] > shape[1]:
+                    continue
+                y = mx.zeros(shape)
+                x = mx.random.normal(upd_shape)
+                z = y.at[: upd_shape[0], : upd_shape[1]].add(x)
+                diff = z[: upd_shape[0], : upd_shape[1]] - x
+                self.assertTrue(mx.allclose(diff, mx.zeros_like(diff)))
+
+        # Test non-zero offset slice update
+        y = mx.zeros((32, 32))
+        x = mx.random.normal((16, 16))
+        z = y.at[16:, 16:].add(x)
+        self.assertTrue(mx.allclose(z[16:, 16:] - x, mx.zeros_like(x)))
+
+        # Test with size divisible by 4, 2, and odd
+        for cols in [32, 18, 15]:
+            y = mx.zeros((32, cols))
+            x = mx.random.normal((16, cols))
+            z = y.at[:16, :].add(x)
+            self.assertTrue(mx.allclose(z[:16, :] - x, mx.zeros_like(x)))
+
     def test_slice_negative_step(self):
         a_np = np.arange(20)
         a_mx = mx.array(a_np)

python/tests/test_linalg.py

@@ -520,6 +520,19 @@ class TestLinalg(mlx_tests.MLXTestCase):
         P, L, U = mx.linalg.lu(a, stream=mx.cpu)
         self.assertTrue(mx.allclose(L[P, :] @ U, a))
 
+        # Test singular matrix (should not throw)
+        a = mx.array(
+            [
+                [1.0, 2.0, 3.0, 4.0],
+                [2.0, 4.0, 6.0, 8.0],
+                [0.0, 1.0, 1.0, 0.0],
+                [1.0, 0.0, 0.0, 1.0],
+            ]
+        )
+        P, L, U = mx.linalg.lu(a, stream=mx.cpu)
+        L_permuted = mx.take_along_axis(L, P[..., None], axis=-2)
+        self.assertTrue(mx.allclose(L_permuted @ U, a))
+
     def test_lu_factor(self):
         mx.random.seed(7)
 
@@ -616,6 +629,248 @@ class TestLinalg(mlx_tests.MLXTestCase):
         expected = np.linalg.solve(a, b)
         self.assertTrue(np.allclose(result, expected))
 
+    def test_det(self):
+        # 1x1 fast path
+        A = mx.array([[5.0]])
+        self.assertTrue(np.allclose(mx.linalg.det(A, stream=mx.cpu), 5.0))
+
+        # 2x2 fast path
+        A = mx.array([[1.0, 2.0], [3.0, 4.0]])
+        d = mx.linalg.det(A, stream=mx.cpu)
+        self.assertTrue(np.allclose(d, -2.0))
+
+        # 3x3 fast path
+        A = mx.array([[1.0, 2.0, 3.0], [0.0, 1.0, 4.0], [5.0, 6.0, 0.0]])
+        d = mx.linalg.det(A, stream=mx.cpu)
+        expected = np.linalg.det(np.array(A))
+        self.assertTrue(np.allclose(d, expected, atol=1e-5))
+
+        # 4x4 LU path: compare with numpy
+        np.random.seed(42)
+        A_np = np.random.randn(4, 4).astype(np.float32)
+        A_mx = mx.array(A_np)
+        d_mx = mx.linalg.det(A_mx, stream=mx.cpu)
+        d_np = np.linalg.det(A_np)
+        self.assertTrue(np.allclose(d_mx, d_np, atol=1e-4))
+
+        # 5x5 LU path
+        A_np = np.random.randn(5, 5).astype(np.float32)
+        A_mx = mx.array(A_np)
+        d_mx = mx.linalg.det(A_mx, stream=mx.cpu)
+        d_np = np.linalg.det(A_np)
+        self.assertTrue(np.allclose(d_mx, d_np, atol=1e-4))
+
+        # Identity matrix
+        A = mx.eye(5)
+        self.assertTrue(np.allclose(mx.linalg.det(A, stream=mx.cpu), 1.0))
+
+        # Batched: (3, 4, 4)
+        A_np = np.random.randn(3, 4, 4).astype(np.float32)
+        A_mx = mx.array(A_np)
+        d_mx = mx.linalg.det(A_mx, stream=mx.cpu)
+        d_np = np.linalg.det(A_np)
+        self.assertTrue(np.allclose(d_mx, d_np, atol=1e-4))
+
+        # Multi-batch: (2, 3, 3, 3)
+        A_np = np.random.randn(2, 3, 3, 3).astype(np.float32)
+        A_mx = mx.array(A_np)
+        d_mx = mx.linalg.det(A_mx, stream=mx.cpu)
+        d_np = np.linalg.det(A_np)
+        self.assertTrue(np.allclose(d_mx, d_np, atol=1e-4))
+
+        # Integer input auto-promotes to float
+        A = mx.array([[1, 2], [3, 4]])
+        d = mx.linalg.det(A, stream=mx.cpu)
+        self.assertTrue(np.allclose(d, -2.0))
+
+        # float64
+        A_np = np.random.randn(4, 4).astype(np.float64)
+        A_mx = mx.array(A_np)
+        d_mx = mx.linalg.det(A_mx, stream=mx.cpu)
+        d_np = np.linalg.det(A_np)
+        self.assertTrue(np.allclose(d_mx, d_np, atol=1e-10))
+
+        # Singular 4x4 matrix (LU path): det should be 0
+        A = mx.array(
+            [
+                [1.0, 2.0, 3.0, 4.0],
+                [2.0, 4.0, 6.0, 8.0],
+                [0.0, 1.0, 1.0, 0.0],
+                [1.0, 0.0, 0.0, 1.0],
+            ]
+        )
+        d = mx.linalg.det(A, stream=mx.cpu)
+        self.assertTrue(np.allclose(d, 0.0, atol=1e-5))
+
+        # Singular 5x5 matrix (LU path)
+        A_np = np.ones((5, 5), dtype=np.float32)
+        A_mx = mx.array(A_np)
+        d = mx.linalg.det(A_mx, stream=mx.cpu)
+        self.assertTrue(np.allclose(d, 0.0, atol=1e-5))
+
+        # Batched singular matrices (LU path)
+        A_np = np.array([np.diag([1.0, 2.0, 0.0, 3.0]), np.eye(4, dtype=np.float32)])
+        A_mx = mx.array(A_np)
+        d_mx = mx.linalg.det(A_mx, stream=mx.cpu)
+        d_np = np.linalg.det(A_np)
+        self.assertTrue(np.allclose(d_mx, d_np, atol=1e-5))
+
+        # Empty 0x0 matrix: det is the empty product = 1
+        d = mx.linalg.det(mx.zeros((0, 0)), stream=mx.cpu)
+        self.assertEqual(d.shape, ())
+        self.assertEqual(float(d), 1.0)
+
+        # Batched empty matrices: shape preserves batch dims
+        d = mx.linalg.det(mx.zeros((3, 0, 0)), stream=mx.cpu)
+        self.assertTrue(np.allclose(d, np.linalg.det(np.zeros((3, 0, 0)))))
+
+        # Error: non-square
+        with self.assertRaises(ValueError):
+            mx.linalg.det(mx.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]), stream=mx.cpu)
+
+        # Error: 1D
+        with self.assertRaises(ValueError):
+            mx.linalg.det(mx.array([1.0, 2.0]), stream=mx.cpu)
+
+        # Error: complex unsupported (small-matrix path)
+        with self.assertRaises(ValueError):
+            mx.linalg.det(mx.array([[1.0 + 1j, 2.0], [3.0, 4.0]]), stream=mx.cpu)
+
+        # Error: complex unsupported (LU path)
+        with self.assertRaises(ValueError):
+            mx.linalg.det(mx.eye(4).astype(mx.complex64), stream=mx.cpu)
+
+    def test_slogdet(self):
+        # 2x2: det = -2 => sign = -1, logabsdet = log(2)
+        A = mx.array([[1.0, 2.0], [3.0, 4.0]])
+        sign, logabsdet = mx.linalg.slogdet(A, stream=mx.cpu)
+        self.assertTrue(np.allclose(sign, -1.0))
+        self.assertTrue(np.allclose(logabsdet, np.log(2.0), atol=1e-5))
+
+        # Identity: sign = 1, logabsdet = 0
+        A = mx.eye(4)
+        sign, logabsdet = mx.linalg.slogdet(A, stream=mx.cpu)
+        self.assertTrue(np.allclose(sign, 1.0))
+        self.assertTrue(np.allclose(logabsdet, 0.0, atol=1e-6))
+
+        # Compare with numpy for random matrices
+        np.random.seed(42)
+        for n in [1, 2, 3, 4, 5]:
+            A_np = np.random.randn(n, n).astype(np.float32)
+            A_mx = mx.array(A_np)
+            sign_mx, logabs_mx = mx.linalg.slogdet(A_mx, stream=mx.cpu)
+            sign_np, logabs_np = np.linalg.slogdet(A_np)
+            with self.subTest(n=n):
+                self.assertTrue(np.allclose(sign_mx, sign_np, atol=1e-5))
+                self.assertTrue(np.allclose(logabs_mx, logabs_np, atol=1e-4))
+
+        # Singular matrix 2x2 (fast path): sign = 0, logabsdet = -inf
+        A = mx.array([[1.0, 2.0], [2.0, 4.0]])
+        sign, logabsdet = mx.linalg.slogdet(A, stream=mx.cpu)
+        self.assertEqual(float(sign), 0.0)
+        self.assertEqual(float(logabsdet), float("-inf"))
+
+        # Singular 4x4 matrix (LU path): sign = 0, logabsdet = -inf
+        A = mx.array(
+            [
+                [1.0, 2.0, 3.0, 4.0],
+                [2.0, 4.0, 6.0, 8.0],
+                [0.0, 1.0, 1.0, 0.0],
+                [1.0, 0.0, 0.0, 1.0],
+            ]
+        )
+        sign, logabsdet = mx.linalg.slogdet(A, stream=mx.cpu)
+        self.assertEqual(float(sign), 0.0)
+        self.assertEqual(float(logabsdet), float("-inf"))
+
+        # Singular 5x5 matrix (LU path): all-ones matrix
+        A = mx.array(np.ones((5, 5), dtype=np.float32))
+        sign, logabsdet = mx.linalg.slogdet(A, stream=mx.cpu)
+        self.assertEqual(float(sign), 0.0)
+        self.assertEqual(float(logabsdet), float("-inf"))
+
+        # Batched with mix of singular and non-singular (LU path)
+        A_np = np.array([np.diag([1.0, 2.0, 0.0, 3.0]), np.eye(4, dtype=np.float32)])
+        A_mx = mx.array(A_np)
+        sign_mx, logabs_mx = mx.linalg.slogdet(A_mx, stream=mx.cpu)
+        sign_np, logabs_np = np.linalg.slogdet(A_np)
+        self.assertTrue(np.allclose(sign_mx, sign_np, atol=1e-5))
+        # Check -inf for singular, 0.0 for identity
+        self.assertEqual(float(logabs_mx[0]), float("-inf"))
+        self.assertTrue(np.allclose(logabs_mx[1], 0.0, atol=1e-6))
+
+        # Batched
+        A_np = np.random.randn(3, 4, 4).astype(np.float32)
+        A_mx = mx.array(A_np)
+        sign_mx, logabs_mx = mx.linalg.slogdet(A_mx, stream=mx.cpu)
+        sign_np, logabs_np = np.linalg.slogdet(A_np)
+        self.assertTrue(np.allclose(sign_mx, sign_np, atol=1e-5))
+        self.assertTrue(np.allclose(logabs_mx, logabs_np, atol=1e-4))
+
+        # Multi-batch
+        A_np = np.random.randn(2, 3, 3, 3).astype(np.float32)
+        A_mx = mx.array(A_np)
+        sign_mx, logabs_mx = mx.linalg.slogdet(A_mx, stream=mx.cpu)
+        sign_np, logabs_np = np.linalg.slogdet(A_np)
+        self.assertTrue(np.allclose(sign_mx, sign_np, atol=1e-5))
+        self.assertTrue(np.allclose(logabs_mx, logabs_np, atol=1e-4))
+
+        # Numerical stability: large matrix where det overflows
+        # 0.1 * I_100 has det = 0.1^100 which underflows in float32
+        # but slogdet should give sign=1, logabsdet = 100*log(0.1)
+        n = 100
+        A = mx.array(0.1) * mx.eye(n)
+        sign, logabsdet = mx.linalg.slogdet(A, stream=mx.cpu)
+        self.assertTrue(np.allclose(sign, 1.0))
+        self.assertTrue(np.allclose(logabsdet, n * np.log(0.1), atol=1e-3))
+
+        # Verify det = sign * exp(logabsdet) for non-singular cases
+        A_np = np.random.randn(5, 5).astype(np.float32)
+        A_mx = mx.array(A_np)
+        sign_mx, logabs_mx = mx.linalg.slogdet(A_mx, stream=mx.cpu)
+        det_mx = mx.linalg.det(A_mx, stream=mx.cpu)
+        reconstructed = float(sign_mx) * np.exp(float(logabs_mx))
+        self.assertTrue(np.allclose(float(det_mx), reconstructed, rtol=1e-4))
+
+        # float64
+        A_np = np.random.randn(4, 4).astype(np.float64)
+        A_mx = mx.array(A_np)
+        sign_mx, logabs_mx = mx.linalg.slogdet(A_mx, stream=mx.cpu)
+        sign_np, logabs_np = np.linalg.slogdet(A_np)
+        self.assertTrue(np.allclose(sign_mx, sign_np))
+        self.assertTrue(np.allclose(logabs_mx, logabs_np, atol=1e-10))
+
+        # Empty 0x0 matrix: sign = 1, logabsdet = 0 (empty product)
+        sign, logabsdet = mx.linalg.slogdet(mx.zeros((0, 0)), stream=mx.cpu)
+        self.assertEqual(sign.shape, ())
+        self.assertEqual(logabsdet.shape, ())
+        self.assertEqual(float(sign), 1.0)
+        self.assertEqual(float(logabsdet), 0.0)
+
+        # Batched empty matrices
+        sign, logabsdet = mx.linalg.slogdet(mx.zeros((3, 0, 0)), stream=mx.cpu)
+        sign_np, logabs_np = np.linalg.slogdet(np.zeros((3, 0, 0)))
+        self.assertTrue(np.allclose(sign, sign_np))
+        self.assertTrue(np.allclose(logabsdet, logabs_np))
+
+        # Error: non-square
+        with self.assertRaises(ValueError):
+            mx.linalg.slogdet(
+                mx.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]), stream=mx.cpu
+            )
+
+        # Error: 1D
+        with self.assertRaises(ValueError):
+            mx.linalg.slogdet(mx.array([1.0, 2.0]), stream=mx.cpu)
+
+        # Error: complex unsupported (small-matrix path)
+        with self.assertRaises(ValueError):
+            mx.linalg.slogdet(mx.array([[1.0 + 1j, 2.0], [3.0, 4.0]]), stream=mx.cpu)
+
+        # Error: complex unsupported (LU path)
+        with self.assertRaises(ValueError):
+            mx.linalg.slogdet(mx.eye(4).astype(mx.complex64), stream=mx.cpu)
+
 
 if __name__ == "__main__":
     mlx_tests.MLXTestRunner()

python/tests/test_ops.py

@@ -92,6 +92,52 @@ class TestOps(mlx_tests.MLXTestCase):
             self.assertEqual(y.dtype, t)
             self.assertTrue(mx.array_equal(y, x))
 
+    def test_shape_overflow_error(self):
+        # Shape dimensions that don't fit in int32 should raise a clear
+        # OverflowError that names the offending value, rather than a generic
+        # "incompatible function arguments" TypeError. The overflow check
+        # lives in the mx::Shape type caster, so it applies to every op that
+        # takes a shape. See issue #2681.
+        too_big = 2**31
+
+        # Array creation ops — also exercise the scalar shape path.
+        for ctor in (mx.zeros, mx.ones):
+            with self.assertRaises(OverflowError) as cm:
+                ctor(too_big)
+            self.assertIn(str(too_big), str(cm.exception))
+            with self.assertRaises(OverflowError) as cm:
+                ctor([too_big])
+            self.assertIn(str(too_big), str(cm.exception))
+
+        with self.assertRaises(OverflowError) as cm:
+            mx.full(too_big, 0.0)
+        self.assertIn(str(too_big), str(cm.exception))
+        with self.assertRaises(OverflowError) as cm:
+            mx.full([too_big], 0.0)
+        self.assertIn(str(too_big), str(cm.exception))
+
+        # Other shape-taking ops should surface the same clean error.
+        a = mx.zeros(4)
+        with self.assertRaises(OverflowError) as cm:
+            mx.reshape(a, [too_big])
+        self.assertIn(str(too_big), str(cm.exception))
+
+        with self.assertRaises(OverflowError) as cm:
+            mx.broadcast_to(a, [too_big, 1])
+        self.assertIn(str(too_big), str(cm.exception))
+
+        # Negative overflow (< int32 min) is caught too.
+        too_negative = -(2**31) - 1
+        with self.assertRaises(OverflowError) as cm:
+            mx.zeros([too_negative])
+        self.assertIn(str(too_negative), str(cm.exception))
+
+        # Shapes that fit in int32 still go through unchanged.
+        self.assertEqual(mx.zeros(4).shape, (4,))
+        self.assertEqual(mx.zeros((2, 3)).shape, (2, 3))
+        self.assertEqual(mx.ones([2, 3]).shape, (2, 3))
+        self.assertEqual(mx.full((2, 3), 1.5).tolist(), [[1.5] * 3] * 2)
+
     def test_scalar_inputs(self):
         # Check combinations of python types
         a = mx.add(False, True)

python/tests/test_quantized.py

@@ -470,6 +470,30 @@ class TestQuantized(mlx_tests.MLXTestCase):
         self.assertEqual(y_q.shape, y_hat.shape)
         self.assertLess((y_q - y_hat).abs().max(), 2e-3)
 
+    def test_qvm_splitk_multi_row(self):
+        # Test qvm split_k with M > 1 to ensure the x row stride is correct
+        key = mx.random.key(0)
+        k1, k2 = mx.random.split(key)
+        tests = product(
+            [64, 32],  # group_size
+            [4, 8],  # bits
+            [128],  # out dim (N)
+            [2048, 4096],  # in dim (K) >= 1024 to trigger split_k
+            [2, 3],  # M (multiple rows)
+        )
+        for group_size, bits, N, K, M in tests:
+            with self.subTest(M=M, K=K, N=N, group_size=group_size, bits=bits):
+                x = 1e-1 * mx.random.normal(shape=(M, K), key=k1)
+                w = 1e-1 * mx.random.normal(shape=(K, N), key=k2)
+                w_q, scales, biases = mx.quantize(w, group_size, bits)
+                w_hat = mx.dequantize(w_q, scales, biases, group_size, bits)
+                y_q = mx.quantized_matmul(
+                    x, w_q, scales, biases, False, group_size, bits
+                )
+                y_hat = x @ w_hat
+                self.assertEqual(y_q.shape, y_hat.shape)
+                self.assertLess((y_q - y_hat).abs().max(), 2e-3)
+
     def test_fp_qvm(self):
         key = mx.random.key(0)
         k1, k2 = mx.random.split(key)
@@ -1046,7 +1070,7 @@ class TestQuantized(mlx_tests.MLXTestCase):
                 y3 = scatter_unsort(y3, inv_order, indices.shape)
                 y4 = scatter_unsort(y4, inv_order, indices.shape)
 
-                tol = 1.5e-5 if (dtype == mx.float32) else 2.5e-4
+                tol = 1.5e-5 if (dtype == mx.float32) else 1e-3
 
                 self.assertLess((y1 - y2).abs().max(), tol)
                 self.assertLess((y1 - y3).abs().max(), tol)

tests/linalg_tests.cpp

@@ -637,3 +637,68 @@ TEST_CASE("test solve_triangluar") {
   expected = array({-3., 2., 3.});
   CHECK(allclose(expected, result).item<bool>());
 }
+
+TEST_CASE("test det") {
+  // 1x1 fast path
+  {
+    array a = array({5.0f}, {1, 1});
+    auto d = det(a, Device::cpu);
+    CHECK_EQ(d.item<float>(), doctest::Approx(5.0f));
+  }
+
+  // 2x2 fast path: det([[1,2],[3,4]]) = -2
+  {
+    array a = array({1.0f, 2.0f, 3.0f, 4.0f}, {2, 2});
+    auto d = det(a, Device::cpu);
+    CHECK_EQ(d.item<float>(), doctest::Approx(-2.0f));
+  }
+
+  // 3x3 fast path: det([[1,2,3],[0,1,4],[5,6,0]]) = 1
+  {
+    array a =
+        array({1.0f, 2.0f, 3.0f, 0.0f, 1.0f, 4.0f, 5.0f, 6.0f, 0.0f}, {3, 3});
+    auto d = det(a, Device::cpu);
+    CHECK_EQ(d.item<float>(), doctest::Approx(1.0f));
+  }
+
+  // 4x4 LU path: identity matrix det = 1
+  {
+    array a = eye(4);
+    auto d = det(a, Device::cpu);
+    CHECK_EQ(d.item<float>(), doctest::Approx(1.0f));
+  }
+
+  // Non-square should throw
+  CHECK_THROWS(
+      det(array({1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f}, {2, 3}), Device::cpu));
+
+  // 1D should throw
+  CHECK_THROWS(det(array({1.0f, 2.0f}), Device::cpu));
+}
+
+TEST_CASE("test slogdet") {
+  // 2x2: det = -2, so sign = -1, logabsdet = log(2)
+  {
+    array a = array({1.0f, 2.0f, 3.0f, 4.0f}, {2, 2});
+    auto [s, logabs] = slogdet(a, Device::cpu);
+    CHECK_EQ(s.item<float>(), doctest::Approx(-1.0f));
+    CHECK_EQ(logabs.item<float>(), doctest::Approx(std::log(2.0f)));
+  }
+
+  // Identity: sign = 1, logabsdet = 0
+  {
+    array a = eye(4);
+    auto [s, logabs] = slogdet(a, Device::cpu);
+    CHECK_EQ(s.item<float>(), doctest::Approx(1.0f));
+    CHECK_EQ(logabs.item<float>(), doctest::Approx(0.0f));
+  }
+
+  // Singular: sign = 0, logabsdet = -inf
+  {
+    array a = array({1.0f, 2.0f, 2.0f, 4.0f}, {2, 2});
+    auto [s, logabs] = slogdet(a, Device::cpu);
+    CHECK_EQ(s.item<float>(), 0.0f);
+    CHECK(std::isinf(logabs.item<float>()));
+    CHECK(logabs.item<float>() < 0);
+  }
+}