diff --git a/mlx/backend/cuda/CMakeLists.txt b/mlx/backend/cuda/CMakeLists.txt
index 1d089710..00f639c5 100644
--- a/mlx/backend/cuda/CMakeLists.txt
+++ b/mlx/backend/cuda/CMakeLists.txt
@@ -31,6 +31,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/gemms/block_mask.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/gemms/gemv.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/gather_gemm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/gemms/grouped_gemm_unaligned.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/jit_module.cpp
@@ -119,11 +120,11 @@ target_compile_options(mlx
 target_compile_options(
   mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>")
 
-if(MSVC)
-  # Ignore warnings from CUTLASS.
-  target_compile_options(
-    mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe="--diag_suppress=2908">)
-else()
+# Ignore warnings from CUTLASS.
+target_compile_options(
+  mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe="--diag_suppress=2908,2361">)
+
+if(NOT MSVC)
   # Required for generating optimized CUTLASS code.
   target_compile_options(
     mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-fno-strict-aliasing>")
diff --git a/mlx/backend/cuda/gemms/gather_gemm.cu b/mlx/backend/cuda/gemms/gather_gemm.cu
new file mode 100644
index 00000000..2c971be4
--- /dev/null
+++ b/mlx/backend/cuda/gemms/gather_gemm.cu
@@ -0,0 +1,334 @@
+// Copyright © 2026 Apple Inc.
+
+#include "mlx/backend/cuda/cutlass_utils.cuh"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+
+#include <cutlass/epilogue/collective/collective_epilogue.hpp>
+#include <cutlass/epilogue/thread/linear_combination.h>
+#include <cutlass/gemm/collective/collective_mma.hpp>
+#include <cutlass/gemm/device/gemm_universal_adapter.h>
+#include <cutlass/gemm/dispatch_policy.hpp>
+#include <cutlass/gemm/kernel/gemm_universal.hpp>
+
+// We can't put kernel code in mlx::core due to name conflicts of "Shape".
+namespace cutlass_gemm {
+
+using namespace cute;
+
+// Modified from cutlass/include/cutlass/gemm/kernel/sm70_gemm.hpp to fuse
+// gather into GEMM.
+template <
+    class ProblemShape_,
+    class CollectiveMainloop_,
+    class CollectiveEpilogue_>
+class GatherGemm {
+ public:
+  using ProblemShape = ProblemShape_;
+  using CollectiveMainloop = CollectiveMainloop_;
+  using TileShape = typename CollectiveMainloop::TileShape;
+  using TiledMma = typename CollectiveMainloop::TiledMma;
+  using ArchTag = typename CollectiveMainloop::ArchTag;
+  using ElementA = typename CollectiveMainloop::ElementA;
+  using StrideA = typename CollectiveMainloop::StrideA;
+  using ElementB = typename CollectiveMainloop::ElementB;
+  using StrideB = typename CollectiveMainloop::StrideB;
+  using DispatchPolicy = typename CollectiveMainloop::DispatchPolicy;
+  using ElementAccumulator = typename CollectiveMainloop::ElementAccumulator;
+
+  using CollectiveEpilogue = CollectiveEpilogue_;
+  using ElementC = typename CollectiveEpilogue::ElementC;
+  using StrideC = typename CollectiveEpilogue::StrideC;
+  using ElementD = typename CollectiveEpilogue::ElementD;
+  using StrideD = typename CollectiveEpilogue::StrideD;
+
+  static constexpr int SharedStorageSize = static_cast<int>(cute::max(
+      sizeof(typename CollectiveMainloop::SharedStorage),
+      sizeof(typename CollectiveEpilogue::SharedStorage)));
+  static constexpr uint32_t MaxThreadsPerBlock =
+      CUTE_STATIC_V(size(TiledMma{}));
+  static constexpr uint32_t MinBlocksPerMultiprocessor = 1;
+
+  struct Arguments {
+    ProblemShape problem_shape;
+    const uint32_t* lhs_indices;
+    const uint32_t* rhs_indices;
+    typename CollectiveMainloop::Arguments mainloop;
+    typename CollectiveEpilogue::Arguments epilogue;
+  };
+
+  struct Params {
+    ProblemShape problem_shape;
+    const uint32_t* lhs_indices;
+    const uint32_t* rhs_indices;
+    typename CollectiveMainloop::Params mainloop;
+    typename CollectiveEpilogue::Params epilogue;
+  };
+
+  static Params to_underlying_arguments(
+      const Arguments& args,
+      void* workspace) {
+    return {
+        args.problem_shape,
+        args.lhs_indices,
+        args.rhs_indices,
+        CollectiveMainloop::to_underlying_arguments(
+            args.problem_shape, args.mainloop, workspace),
+        CollectiveEpilogue::to_underlying_arguments(
+            args.problem_shape, args.epilogue, workspace)};
+  }
+
+  static cutlass::Status
+  initialize_workspace(const Arguments&, void*, cudaStream_t, void*) {
+    return cutlass::Status::kSuccess;
+  }
+
+  static dim3 get_grid_shape(const Params& params) {
+    auto [m, n, k, l] = params.problem_shape;
+    return dim3{
+        uint32_t(ceil_div(m, shape<0>(TileShape{}))),
+        uint32_t(ceil_div(n, shape<1>(TileShape{}))),
+        uint32_t(l)};
+  }
+
+  static dim3 get_block_shape() {
+    return dim3{MaxThreadsPerBlock, 1, 1};
+  }
+
+  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);
+
+    auto shape_MNKL = append<4>(params.problem_shape, Int<1>{});
+    auto cta_tile = TileShape{};
+    auto cta_coord = make_coord(m_coord, n_coord, _, l_coord);
+
+    // Represent the full tensors.
+    Tensor mA_mkl = make_tensor(
+        make_gmem_ptr(params.mainloop.ptr_A),
+        select<0, 2, 3>(shape_MNKL),
+        params.mainloop.dA);
+    Tensor mB_nkl = make_tensor(
+        make_gmem_ptr(params.mainloop.ptr_B),
+        select<1, 2, 3>(shape_MNKL),
+        params.mainloop.dB);
+
+    // Get batch slice.
+    Tensor mA_mk = mA_mkl(_, _, params.lhs_indices[l_coord]);
+    Tensor mB_nk = mB_nkl(_, _, params.rhs_indices[l_coord]);
+
+    // Slice to get the tiles this thread block is responsible for.
+    Tensor gA =
+        local_tile(mA_mk, cta_tile, take<0, 3>(cta_coord), Step<_1, X, _1>{});
+    Tensor gB =
+        local_tile(mB_nk, cta_tile, take<0, 3>(cta_coord), Step<X, _1, _1>{});
+
+    // Compute tile residues for predication.
+    auto m_max_coord = size<0>(shape_MNKL) - size<0>(gA) * get<0>(cta_coord);
+    auto n_max_coord = size<1>(shape_MNKL) - size<0>(gB) * get<1>(cta_coord);
+    auto k_residue = size<2>(shape_MNKL) - size<1>(gA) * size<2>(gA);
+    auto residue_mnk = make_tuple(m_max_coord, n_max_coord, k_residue);
+
+    // Allocate the tiled_mma and the accumulators for the (M,N) cta_tile.
+    TiledMma tiled_mma;
+    Tensor accum = partition_fragment_C(tiled_mma, take<0, 2>(cta_tile));
+    clear(accum);
+
+    auto k_tile_iter = make_coord_iterator(shape<2>(gA));
+    int k_tile_count = size<2>(gA);
+
+    // Perform the collective scoped MMA.
+    CollectiveMainloop collective_mma;
+    collective_mma(
+        accum,
+        gA,
+        gB,
+        accum,
+        k_tile_iter,
+        k_tile_count,
+        residue_mnk,
+        thread_idx,
+        smem_buf);
+
+    // Epilogue and write to out.
+    CollectiveEpilogue epilogue(params.epilogue);
+    epilogue(
+        shape_MNKL,
+        cta_tile,
+        cta_coord,
+        accum,
+        tiled_mma,
+        residue_mnk,
+        thread_idx,
+        smem_buf);
+  }
+};
+
+template <typename Element, bool KMajor>
+struct SimtCopyTraits {};
+
+template <typename Element>
+struct SimtCopyTraits<Element, true> {
+  using GmemTiledCopy = decltype(make_tiled_copy(
+      Copy_Atom<UniversalCopy<Element>, Element>{},
+      Layout<Shape<_32, _8>, Stride<_8, _1>>{},
+      Layout<Shape<_1, _1>>{}));
+  using SmemLayout = Layout<Shape<_128, _8>, Stride<_1, Int<128 + 1>>>;
+  using SmemCopyAtom = Copy_Atom<DefaultCopy, Element>;
+};
+
+template <typename Element>
+struct SimtCopyTraits<Element, false> {
+  using GmemTiledCopy = decltype(make_tiled_copy(
+      Copy_Atom<UniversalCopy<Element>, Element>{},
+      Layout<Shape<_32, _8>, Stride<_1, _32>>{},
+      Layout<Shape<_1, _1>>{}));
+  using SmemLayout = Layout<Shape<_128, _8>, Stride<_1, _128>>;
+  using SmemCopyAtom = Copy_Atom<DefaultCopy, Element>;
+};
+
+template <typename F>
+void dispatch_stride(bool k_major, int m, int k, F&& f) {
+  if (k_major) {
+    f(make_stride(k, Int<1>{}, m * k), std::true_type{});
+  } else {
+    f(make_stride(Int<1>{}, m, m * k), std::false_type{});
+  }
+}
+
+template <typename Element, typename F>
+void gather_mm(
+    int m,
+    int n,
+    int k,
+    int l,
+    bool a_transposed,
+    bool b_transposed,
+    const Element* A,
+    const Element* B,
+    const uint32_t* lhs_indices,
+    const uint32_t* rhs_indices,
+    Element* C,
+    F&& launch_kernel) {
+  auto problem_shape = make_shape(m, n, k, l);
+  auto dC = make_stride(m, Int<1>{}, m * n);
+  dispatch_stride(!a_transposed, m, k, [&](auto dA, auto k_major_a) {
+    dispatch_stride(b_transposed, n, k, [&](auto dB, auto k_major_b) {
+      using Accumulator =
+          std::conditional_t<(sizeof(Element) < 4), float, Element>;
+      using TileShape = Shape<_128, _128, _8>;
+      using DispatchPolicy = cutlass::gemm::MainloopSm70TwoStage;
+      using TiledMma = TiledMMA<
+          MMA_Atom<UniversalFMA<Accumulator, Element, Element, Element>>,
+          Layout<Shape<_16, _16, _1>>>;
+
+      using CopyTraitsA = SimtCopyTraits<Element, k_major_a.value>;
+      using CopyTraitsB = SimtCopyTraits<Element, k_major_b.value>;
+
+      using CollectiveMainloop = cutlass::gemm::collective::CollectiveMma<
+          DispatchPolicy,
+          TileShape,
+          Element,
+          decltype(dA),
+          Element,
+          decltype(dB),
+          TiledMma,
+          typename CopyTraitsA::GmemTiledCopy,
+          typename CopyTraitsA::SmemLayout,
+          typename CopyTraitsA::SmemCopyAtom,
+          identity,
+          typename CopyTraitsB::GmemTiledCopy,
+          typename CopyTraitsB::SmemLayout,
+          typename CopyTraitsB::SmemCopyAtom,
+          identity>;
+
+      using CollectiveEpilogue = cutlass::epilogue::collective::DefaultEpilogue<
+          Element,
+          decltype(dC),
+          decltype(dC),
+          cutlass::epilogue::thread::
+              LinearCombination<Element, 1, Accumulator, Accumulator>,
+          cutlass::gemm::EpilogueDefault>;
+
+      using GemmKernel = GatherGemm<
+          decltype(problem_shape),
+          CollectiveMainloop,
+          CollectiveEpilogue>;
+      using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+
+      Gemm gemm;
+      typename Gemm::Arguments args{
+          problem_shape,
+          lhs_indices,
+          rhs_indices,
+          {A, dA, B, dB},
+          {{1.f, 0.f}, C, dC, C, dC}};
+
+      CHECK_CUTLASS_ERROR(gemm.initialize(args, nullptr));
+
+      auto* kernel = &cutlass::device_kernel<GemmKernel>;
+      void* kernel_params[] = {const_cast<Gemm::Params*>(&gemm.params())};
+      launch_kernel(
+          reinterpret_cast<void*>(kernel),
+          gemm.get_grid_shape(gemm.params()),
+          GemmKernel::get_block_shape(),
+          GemmKernel::SharedStorageSize,
+          kernel_params);
+    });
+  });
+}
+
+} // namespace cutlass_gemm
+
+namespace mlx::core {
+
+void cutlass_gather_mm(
+    bool a_transposed,
+    bool b_transposed,
+    const array& a,
+    const array& b,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    array& out,
+    cu::CommandEncoder& encoder) {
+  int m = out.shape(-2);
+  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);
+  encoder.set_input_array(lhs_indices);
+  encoder.set_input_array(rhs_indices);
+  encoder.set_output_array(out);
+
+  dispatch_float_types(out.dtype(), "gather_mm", [&](auto type_tag) {
+    using Element = cutlass_type_t<MLX_GET_TYPE(type_tag)>;
+    cutlass_gemm::gather_mm(
+        m,
+        n,
+        k,
+        l,
+        a_transposed,
+        b_transposed,
+        gpu_ptr<Element>(a),
+        gpu_ptr<Element>(b),
+        gpu_ptr<uint32_t>(lhs_indices),
+        gpu_ptr<uint32_t>(rhs_indices),
+        gpu_ptr<Element>(out),
+        [&](auto* kernel,
+            dim3 num_blocks,
+            dim3 block_dims,
+            uint32_t smem_bytes,
+            void** args) {
+          encoder.add_kernel_node_raw(
+              kernel, num_blocks, block_dims, {}, smem_bytes, args);
+        });
+  });
+}
+
+} // namespace mlx::core
diff --git a/mlx/backend/cuda/gemms/gather_gemm.h b/mlx/backend/cuda/gemms/gather_gemm.h
new file mode 100644
index 00000000..a2492a5f
--- /dev/null
+++ b/mlx/backend/cuda/gemms/gather_gemm.h
@@ -0,0 +1,23 @@
+// Copyright © 2026 Apple Inc.
+
+#pragma once
+
+namespace mlx::core {
+
+namespace cu {
+class CommandEncoder;
+}
+
+class array;
+
+void cutlass_gather_mm(
+    bool a_transposed,
+    bool b_transposed,
+    const array& a,
+    const array& b,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    array& out,
+    cu::CommandEncoder& encoder);
+
+} // namespace mlx::core
diff --git a/mlx/backend/cuda/gemms/gemv.cu b/mlx/backend/cuda/gemms/gemv.cu
index 7fdfc726..bb463bf7 100644
--- a/mlx/backend/cuda/gemms/gemv.cu
+++ b/mlx/backend/cuda/gemms/gemv.cu
@@ -167,7 +167,7 @@ __global__ void gemv_gather(
 }
 
 bool can_use_gemv(int M, int N, int K, bool a_transposed, bool b_transposed) {
-  return K % 32 == 0 && ((M == 1 && b_transposed) || (N == 1 && !a_transposed));
+  return (M == 1 && b_transposed) || (N == 1 && !a_transposed);
 }
 
 template <typename F>
diff --git a/mlx/backend/cuda/gemms/grouped_gemm_unaligned.cu b/mlx/backend/cuda/gemms/grouped_gemm_unaligned.cu
index bf17f597..08b3f1c1 100644
--- a/mlx/backend/cuda/gemms/grouped_gemm_unaligned.cu
+++ b/mlx/backend/cuda/gemms/grouped_gemm_unaligned.cu
@@ -325,6 +325,11 @@ void cutlass_grouped_gemm_unaligned(
     const array& indices,
     array& out,
     cu::CommandEncoder& encoder) {
+  if (group_count > 1024) {
+    throw std::runtime_error(
+        "[gather_mm] Group count can not be larger than 1024.");
+  }
+
   int K = a.shape(-1);
   int N = b.shape(-1);
 
diff --git a/mlx/backend/cuda/matmul.cpp b/mlx/backend/cuda/matmul.cpp
index f06fed39..ddcfe577 100644
--- a/mlx/backend/cuda/matmul.cpp
+++ b/mlx/backend/cuda/matmul.cpp
@@ -4,6 +4,7 @@
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/gemms/block_mask.h"
 #include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/gemms/gather_gemm.h"
 #include "mlx/backend/cuda/gemms/gemv.h"
 #include "mlx/backend/cuda/gemms/grouped_gemm.h"
 #include "mlx/backend/gpu/copy.h"
@@ -137,40 +138,6 @@ void gemm_and_bias(
       encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides, alpha);
 }
 
-void gather_mm_rhs(
-    const array& a_,
-    const array& b_,
-    const array& indices_,
-    array& out,
-    cu::CommandEncoder& encoder,
-    Stream s) {
-  if (a_.size() / a_.shape(-2) / a_.shape(-1) != indices_.size()) {
-    throw std::runtime_error("[gather_mm] Broadcasting lhs is not supported.");
-  }
-
-  int group_count = b_.size() / b_.shape(-1) / b_.shape(-2);
-  if (group_count > 1024) {
-    throw std::runtime_error(
-        "[gather_mm] Group count can not be larger than 1024.");
-  }
-
-  auto [a_transposed, lda, a] = ensure_batch_contiguous(a_, encoder, s);
-  auto [b_transposed, ldb, b] = ensure_batch_contiguous(b_, encoder, s);
-  auto indices = ensure_row_contiguous(indices_, encoder, s);
-
-  cutlass_grouped_gemm_unaligned(
-      a_transposed,
-      lda,
-      b_transposed,
-      ldb,
-      group_count,
-      a,
-      b,
-      indices,
-      out,
-      encoder);
-}
-
 } // namespace
 
 void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -404,14 +371,12 @@ void GatherMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& encoder = cu::get_command_encoder(s);
 
   assert(inputs.size() == 4);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  auto& lhs_indices = inputs[2];
-  auto& rhs_indices = inputs[3];
+  auto& a_pre = inputs[0];
+  auto& b_pre = inputs[1];
 
   // Return 0s if either input is empty.
-  if (a.size() == 0 || b.size() == 0) {
-    array zero(0, a.dtype());
+  if (a_pre.size() == 0 || b_pre.size() == 0) {
+    array zero(0, a_pre.dtype());
     encoder.add_temporary(zero);
     fill_gpu(zero, out, s);
     return;
@@ -420,31 +385,44 @@ void GatherMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   // Extract shapes from inputs.
-  int M = a.shape(-2);
-  int N = b.shape(-1);
-  int K = a.shape(-1);
+  int M = a_pre.shape(-2);
+  int N = b_pre.shape(-1);
+  int K = a_pre.shape(-1);
+
+  auto [a_transposed, lda, a] = ensure_batch_contiguous(a_pre, encoder, s);
+  auto [b_transposed, ldb, b] = ensure_batch_contiguous(b_pre, encoder, s);
+  auto lhs_indices = ensure_row_contiguous(inputs[2], encoder, s);
+  auto rhs_indices = ensure_row_contiguous(inputs[3], encoder, s);
 
   // We are walking a in order and b is also in order so we can batch up the
   // matmuls and reuse reading a and b.
   if (M == 1 && right_sorted_ == true) {
-    gather_mm_rhs(a, b, rhs_indices, out, encoder, s);
+    cutlass_grouped_gemm_unaligned(
+        a_transposed,
+        lda,
+        b_transposed,
+        ldb,
+        b.size() / b.shape(-1) / b.shape(-2), // group_count
+        a,
+        b,
+        rhs_indices,
+        out,
+        encoder);
     return;
   }
 
-  auto [transposed_a, lda, a_] = check_transpose(encoder, s, a);
-  auto [transposed_b, ldb, b_] = check_transpose(encoder, s, b);
-  auto use_gemv = cu::can_use_gemv(M, N, K, transposed_a, transposed_b);
+  auto use_gemv = cu::can_use_gemv(M, N, K, a_transposed, b_transposed);
   if (M == 1 && use_gemv) {
-    gather_mv(b_, a_, rhs_indices, lhs_indices, out, N, K, encoder);
+    gather_mv(b, a, rhs_indices, lhs_indices, out, N, K, encoder);
     return;
   }
-
   if (N == 1 && use_gemv) {
-    gather_mv(a_, b_, lhs_indices, rhs_indices, out, M, K, encoder);
+    gather_mv(a, b, lhs_indices, rhs_indices, out, M, K, encoder);
     return;
   }
 
-  throw std::runtime_error("NYI");
+  cutlass_gather_mm(
+      a_transposed, b_transposed, a, b, lhs_indices, rhs_indices, out, encoder);
 }
 
 void SegmentedMM::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -473,14 +451,7 @@ void SegmentedMM::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
   auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
-  auto segments = [&] {
-    if (segments_pre.flags().row_contiguous) {
-      return segments_pre;
-    }
-    array copy = contiguous_copy_gpu(segments_pre, s);
-    encoder.add_temporary(copy);
-    return copy;
-  }();
+  auto segments = ensure_row_contiguous(segments_pre, encoder, s);
 
   cutlass_segmented_mm(
       a_transposed,
diff --git a/python/tests/cuda_skip.py b/python/tests/cuda_skip.py
index 8f91199b..2de3f3bc 100644
--- a/python/tests/cuda_skip.py
+++ b/python/tests/cuda_skip.py
@@ -1,8 +1,4 @@
 cuda_skip = {
-    # Gather matmul NYI
-    "TestBlas.test_gather_matmul",
-    "TestBlas.test_gather_matmul_grad",
-    "TestBlas.test_gather_mm_sorted_vjp",
     # Lapack ops NYI
     "TestLinalg.test_cholesky",
     "TestLinalg.test_cholesky_inv",