Support unaligned M

* speedup scalars in cuda

* comment

.circleci/config.yml

@@ -7,15 +7,6 @@ parameters:
   nightly_build:
     type: boolean
     default: false
-  weekly_build:
-    type: boolean
-    default: false
-  test_release:
-    type: boolean
-    default: false
-  linux_release:
-    type: boolean
-    default: false
 
 jobs:
   build_documentation:
@@ -38,7 +29,7 @@ jobs:
             pip install --upgrade pip
             pip install --upgrade cmake
             pip install -r docs/requirements.txt
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install . -v
+            pip install . -v
       - when:
           condition:
             not: << parameters.upload-docs >>
@@ -70,9 +61,9 @@ jobs:
                  git push -f origin gh-pages
 
   linux_build_and_test:
-    docker:
-      - image: cimg/python:3.9
-
+    machine:
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
@@ -84,37 +75,34 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
             sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get upgrade -y
+            pip install --upgrade cmake
+            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
             sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
       - run:
           name: Install Python package
           command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py develop
+            pip install -e ".[dev]"
       - run:
           name: Generate package stubs
           command: |
             echo "stubs"
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
-            python3 -m unittest discover python/tests -v
+            python -m unittest discover python/tests -v
             mpirun --bind-to none -host localhost:8 -np 8 python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build CPP only
           command: |
-            mkdir -p build && cd build 
+            mkdir -p build && cd build
             cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
             make -j `nproc`
       - run:
@@ -154,15 +142,14 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-            CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
               pip install -e . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
@@ -170,7 +157,8 @@ jobs:
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
             mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build example extension
           command: |
@@ -205,13 +193,34 @@ jobs:
           name: Run Python tests with JIT
           command: |
             source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
+            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
               pip install -e . -v
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
               python -m xmlrunner discover -v python/tests -o test-results/gpu_jit
 
+  cuda_build_and_test:
+    machine:
+      image: linux-cuda-12:2023.11.1
+      resource_class: gpu.nvidia.small.gen2
+    steps:
+      - checkout
+      - run:
+          name: Install Python package
+          command: |
+            sudo apt-get update
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            python3 -m venv env
+            source env/bin/activate
+            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              pip install -e ".[dev]"
+      - run:
+          name: Run Python tests
+          command: |
+            source env/bin/activate
+            LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
+            LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
+
   build_release:
     parameters:
       python_version:
@@ -252,21 +261,29 @@ jobs:
           command: |
             source env/bin/activate
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
             source env/bin/activate
-            << parameters.build_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              python -m build -w
+            python setup.py clean --all
+            << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
+      - when:
+          condition:
+            equal: ["3.9", << parameters.python_version >>]
+          steps:
+            - run:
+                name: Build common package
+                command: |
+                  source env/bin/activate
+                  python setup.py clean --all
+                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
           condition: << parameters.build_env >>
           steps:
@@ -283,52 +300,100 @@ jobs:
       python_version:
         type: string
         default: "3.9"
-      extra_env:
+      build_env:
         type: string
-        default: "DEV_RELEASE=1"
-    docker:
-      - image: ubuntu:20.04
+        default: ""
+    machine:
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
           name: Build wheel
           command: |
             PYTHON=python<< parameters.python_version >>
-            apt-get update
-            apt-get upgrade -y
-            DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get -y install tzdata
-            apt-get install -y apt-utils
-            apt-get install -y software-properties-common
-            add-apt-repository -y ppa:deadsnakes/ppa
-            apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
-            apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            apt-get install -y build-essential git
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
+            sudo apt-get update
+            sudo apt-get upgrade -y
+            TZ=Etc/UTC sudo apt-get -y install tzdata
+            sudo apt-get install -y apt-utils
+            sudo apt-get install -y software-properties-common
+            sudo add-apt-repository -y ppa:deadsnakes/ppa
+            sudo apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
+            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install -y build-essential git
             $PYTHON -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install --upgrade setuptools
-            pip install numpy
             pip install auditwheel
             pip install patchelf
             pip install build
             pip install twine
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              pip install . -v
+            << parameters.build_env >> pip install ".[dev]" -v
             pip install typing_extensions
-            python setup.py generate_stubs 
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python -m build --wheel
-            auditwheel show dist/*
-            auditwheel repair dist/* --plat manylinux_2_31_x86_64
+            python setup.py generate_stubs
+            python setup.py clean --all
+            MLX_BUILD_STAGE=1 << parameters.build_env >> python -m build -w
+            bash python/scripts/repair_linux.sh
+      - when:
+          condition:
+            equal: ["3.9", << parameters.python_version >>]
+          steps:
+            - run:
+                name: Build common package
+                command: |
+                  source env/bin/activate
+                  python setup.py clean --all
+                  << parameters.build_env >> MLX_BUILD_STAGE=2 \
+                    python -m build -w
+                  auditwheel repair dist/mlx_cpu*.whl --plat manylinux_2_35_x86_64
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload packages
+                command: |
+                  source env/bin/activate
+                  twine upload wheelhouse/*.whl
+      - store_artifacts:
+          path: wheelhouse/
+
+  build_cuda_release:
+    parameters:
+      build_env:
+        type: string
+        default: ""
+    machine:
+      image: linux-cuda-12:2024.11.1
+      resource_class: gpu.nvidia.small.gen2
+    steps:
+      - checkout
       - run:
-          name: Upload package
+          name: Build wheel
           command: |
+            sudo apt-get update
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install zip
+            python -m venv env
             source env/bin/activate
-            twine upload wheelhouse/*
+            pip install auditwheel
+            pip install patchelf
+            pip install build
+            pip install twine
+            << parameters.build_env >> MLX_BUILD_STAGE=2 \
+              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              python -m build -w
+            bash python/scripts/repair_cuda.sh
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload package
+                command: |
+                  source env/bin/activate
+                  twine upload wheelhouse/*.whl
       - store_artifacts:
           path: wheelhouse/
 
@@ -340,22 +405,19 @@ workflows:
             pattern: "^(?!pull/)[-\\w]+$"
             value: << pipeline.git.branch >>
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
-        - not: << pipeline.parameters.test_release >>
     jobs:
       - mac_build_and_test:
           matrix:
             parameters:
               macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test
+      - cuda_build_and_test 
       - build_documentation 
 
   build_pypi_release:
     when:
       and:
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
-        - not: << pipeline.parameters.test_release >>
     jobs:
       - build_release:
           filters:
@@ -437,6 +499,25 @@ workflows:
             branches:
               ignore: /.*/
           upload-docs: true
+      - build_linux_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              build_env: ["PYPI_RELEASE=1"]
+      - build_cuda_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              build_env: ["PYPI_RELEASE=1"]
 
   prb:
     when:
@@ -455,6 +536,8 @@ workflows:
               macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test:
           requires: [ hold ]
+      - cuda_build_and_test:
+          requires: [ hold ]
   nightly_build:
     when:
       and:
@@ -513,88 +596,8 @@ workflows:
               - macosx_deployment_target: "15.0"
                 xcode_version: "15.0.0"
                 python_version: "3.13"
-  weekly_build:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.weekly_build >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-  linux_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.linux_release >>
-    jobs:
       - build_linux_release:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]
+      - build_cuda_release

ACKNOWLEDGMENTS.md

@@ -19,6 +19,7 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />

CMakeLists.txt

@@ -22,7 +22,7 @@ project(
 
 # ----------------------------- Setup -----------------------------
 set(CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_INSTALL_MESSAGE NEVER)
@@ -64,10 +64,8 @@ if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
       message(WARNING "Building for x86_64 arch is not officially supported.")
     endif()
   endif()
-
 else()
   set(MLX_BUILD_METAL OFF)
-  message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
 endif()
 
 # ----------------------------- Lib -----------------------------

benchmarks/cpp/irregular_strides.cpp

@@ -1,5 +1,6 @@
 // Copyright © 2023 Apple Inc.
 
+#include <cstring>
 #include <iostream>
 #include <sstream>
 

benchmarks/cpp/single_ops.cpp

@@ -192,6 +192,22 @@ void time_reductions() {
 
   auto argmin_along_1 = [&a]() { return mx::argmin(a, 1, false); };
   TIME(argmin_along_1);
+
+  auto indices = mx::array({1});
+  auto updates = mx::reshape(mx::array({NAN}), {1, 1, 1});
+  std::vector<int> axes{0};
+  auto b = scatter(a, {indices}, updates, axes);
+  mx::eval(b);
+
+  auto max_along_0 = [&b]() { return mx::max(b, 0, false); };
+  TIME(max_along_0);
+  auto max_along_1 = [&b]() { return mx::max(b, 1, false); };
+  TIME(max_along_1);
+
+  auto min_along_0 = [&b]() { return mx::min(b, 0, false); };
+  TIME(min_along_0);
+  auto min_along_1 = [&b]() { return mx::min(b, 1, false); };
+  TIME(min_along_1);
 }
 
 void time_gather_scatter() {

benchmarks/python/comparative/bench_torch.py

@@ -5,6 +5,7 @@ import os
 import time
 
 import torch
+import torch.cuda
 import torch.mps
 
 
@@ -44,8 +45,10 @@ def bench(f, *args):
 
 
 def sync_if_needed(x):
-    if x.device != torch.device("cpu"):
+    if x.device == torch.device("mps"):
         torch.mps.synchronize()
+    elif x.device == torch.device("cuda"):
+        torch.cuda.synchronize()
 
 
 @torch.no_grad()
@@ -99,6 +102,14 @@ def reduction(op, axis, x):
     sync_if_needed(x)
 
 
+@torch.no_grad()
+def sum_and_add(axis, x, y):
+    z = x.sum(axis=axis, keepdims=True)
+    for i in range(50):
+        z = (z + y).sum(axis=axis, keepdims=True)
+    sync_if_needed(x)
+
+
 @torch.no_grad()
 def softmax(axis, x):
     ys = []
@@ -340,7 +351,11 @@ if __name__ == "__main__":
         args.axis.pop(0)
 
     torch.set_num_threads(1)
-    device = "cpu" if args.cpu else "mps"
+    device = "mps"
+    if torch.cuda.is_available():
+        device = "cuda"
+    if args.cpu:
+        device = "cpu"
 
     types = args.dtype
     if not types:
@@ -460,5 +475,8 @@ if __name__ == "__main__":
     elif args.benchmark == "selu":
         print(bench(selu, x))
 
+    elif args.benchmark == "sum_and_add":
+        print(bench(sum_and_add, axis, *xs))
+
     else:
         raise ValueError(f"Unknown benchmark `{args.benchmark}`.")

benchmarks/python/conv_unaligned_bench.py

@@ -0,0 +1,107 @@
+import math
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+N_warmup = 10
+N_iter_bench = 100
+N_iter_func = 5
+
+
+def bench(f, a, b):
+    for i in range(N_warmup):
+        f(a, b)
+    torch.mps.synchronize()
+
+    s = time.perf_counter_ns()
+    for i in range(N_iter_bench):
+        f(a, b)
+    e = time.perf_counter_ns()
+    return (e - s) * 1e-9
+
+
+def make_mx_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    def mx_conv_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = mx.conv2d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_2D
+
+
+def make_pt_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    @torch.no_grad()
+    def pt_conv_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = torch.conv2d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        torch.mps.synchronize()
+        return ys
+
+    return pt_conv_2D
+
+
+def bench_shape(N, H, W, C, kH, kW, O, strides, padding, groups, np_dtype):
+    scale = 1.0 / math.sqrt(kH * kH * C)
+    a_np = np.random.uniform(0, 0.5, (N, H, W, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, kH, kW, int(C / groups))).astype(
+        np_dtype
+    )
+
+    a_mx = mx.array(a_np)
+    b_mx = mx.array(b_np)
+
+    a_pt = torch.from_numpy(a_np.transpose((0, 3, 1, 2))).to("mps")
+    b_pt = torch.from_numpy(b_np.transpose((0, 3, 1, 2))).to("mps")
+
+    torch.mps.synchronize()
+
+    f_mx = make_mx_conv_2D(strides, padding, groups)
+    f_pt = make_pt_conv_2D(strides, padding, groups)
+
+    time_torch = bench(f_pt, a_pt, b_pt)
+    time_mlx = bench(f_mx, a_mx, b_mx)
+
+    out_mx = mx.conv2d(a_mx, b_mx, stride=strides, padding=padding, groups=groups)
+    out_pt = torch.conv2d(
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.permute(out_pt, (0, 2, 3, 1))
+    out_pt = out_pt.numpy(force=True)
+
+    atol = 2e-5 if np_dtype == np.float32 else 1e-4
+
+    if not np.allclose(out_pt, out_mx, atol=atol):
+        print(
+            f"Failed at {(N, H, W, C)}, {(O, kH, kW, C)} [strides = {strides}, padding = {padding}, groups = {groups}] with max(|a - b|) = {np.max(np.abs(out_pt - out_mx))}"
+        )
+
+    return time_mlx, time_torch
+
+
+if __name__ == "__main__":
+    dtype = "float32"
+    shapes = (
+        (4, 32, 32, 21, 3, 3, 128),
+        (4, 32, 32, 21, 3, 3, 37),
+        (4, 32, 32, 370, 3, 3, 370),
+        (4, 32, 32, 370, 7, 7, 128),
+        (2, 320, 640, 21, 7, 7, 21),
+    )
+    for N, H, W, C, kh, kw, O in shapes:
+        time_mlx, time_torch = bench_shape(
+            N, H, W, C, kh, kw, O, (1, 1), (0, 0), 1, dtype
+        )
+        diff = time_torch / time_mlx - 1.0
+
+        print(
+            f"({N}, {H:3d}, {W:3d}, {C:3d}), ({O:3d}, {kh:2d}, {kw:2d}, {C:3d}), {dtype}, {100. * diff:+5.2f}%"
+        )
+        if time_mlx >= 2.0 * time_torch:
+            print("ATTENTION ^^^^^^^")

benchmarks/python/layer_norm_bench.py

@@ -1,5 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.
 
+from functools import partial
+
 import mlx.core as mx
 import mlx.nn as nn
 from time_utils import time_fn
@@ -18,51 +20,63 @@ def layer_norm(x, w, b, eps):
     return y
 
 
-def time_layer_norm():
+def time_layer_norm(N, dt):
+    L = 1024
     f1 = lambda x, w, b, y: (layer_norm(x, w, b, 1e-5) * y).sum()
     f2 = lambda x, w, b, y: (mx.fast.layer_norm(x, w, b, 1e-5) * y).sum()
     g1 = mx.grad(f1, argnums=(0, 1, 2))
     g2 = mx.grad(f2, argnums=(0, 1, 2))
 
-    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
-    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
-    b = mx.random.uniform(shape=(4096,)).astype(mx.float16)
-    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    x = mx.random.uniform(shape=(8, L, N)).astype(dt)
+    w = mx.random.uniform(shape=(N,)).astype(dt)
+    b = mx.random.uniform(shape=(N,)).astype(dt)
+    y = mx.random.uniform(shape=(8, L, N)).astype(dt)
     mx.eval(x, w, b, y)
 
-    def layer_norm_loop(g, x, w, b):
+    def layer_norm_loop(f, x, w, b):
+        for _ in range(32):
+            x = f(x, w, b)
+        return x
+
+    time_fn(layer_norm_loop, partial(layer_norm, eps=1e-5), x, w, b)
+    time_fn(layer_norm_loop, partial(mx.fast.layer_norm, eps=1e-5), x, w, b)
+
+    def layer_norm_grad_loop(g, x, w, b):
         gx, gw, gb = x, w, b
         for _ in range(32):
             gx, gw, gb = g(gx, gw, gb, y)
         return gx, gw, gb
 
-    time_fn(layer_norm_loop, g1, x, w, b)
-    time_fn(layer_norm_loop, g2, x, w, b)
-    time_fn(layer_norm_loop, mx.compile(g1), x, w, b)
-    time_fn(layer_norm_loop, mx.compile(g2), x, w, b)
+    time_fn(layer_norm_grad_loop, g1, x, w, b)
+    time_fn(layer_norm_grad_loop, g2, x, w, b)
+    time_fn(layer_norm_grad_loop, mx.compile(g1), x, w, b)
+    time_fn(layer_norm_grad_loop, mx.compile(g2), x, w, b)
 
     f1 = lambda x, y: (layer_norm(x, None, None, 1e-5) * y).sum()
     f2 = lambda x, y: (mx.fast.layer_norm(x, None, None, 1e-5) * y).sum()
     g1 = mx.grad(f1, argnums=(0,))
     g2 = mx.grad(f2, argnums=(0,))
 
-    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
-    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
-    b = mx.random.uniform(shape=(4096,)).astype(mx.float16)
-    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    x = mx.random.uniform(shape=(8, L, N)).astype(dt)
+    w = mx.random.uniform(shape=(N,)).astype(dt)
+    b = mx.random.uniform(shape=(N,)).astype(dt)
+    y = mx.random.uniform(shape=(8, L, N)).astype(dt)
     mx.eval(x, w, b, y)
 
-    def layer_norm_loop(g, x):
+    def layer_norm_grad_x_loop(g, x):
         gx = x
         for _ in range(32):
             gx = g(gx, y)
         return gx
 
-    time_fn(layer_norm_loop, g1, x)
-    time_fn(layer_norm_loop, g2, x)
-    time_fn(layer_norm_loop, mx.compile(g1), x)
-    time_fn(layer_norm_loop, mx.compile(g2), x)
+    time_fn(layer_norm_grad_x_loop, g1, x)
+    time_fn(layer_norm_grad_x_loop, g2, x)
+    time_fn(layer_norm_grad_x_loop, mx.compile(g1), x)
+    time_fn(layer_norm_grad_x_loop, mx.compile(g2), x)
 
 
 if __name__ == "__main__":
-    time_layer_norm()
+    for dt in [mx.float32, mx.float16, mx.bfloat16]:
+        for n in [1024, 2048, 4096, 8192, 8192 + 1024]:
+            print(dt, n)
+            time_layer_norm(n, dt)

benchmarks/python/single_ops.py

@@ -51,6 +51,20 @@ def time_maximum():
     time_fn(mx.maximum, a, b)
 
 
+def time_max():
+    a = mx.random.uniform(shape=(32, 1024, 1024))
+    a[1, 1] = mx.nan
+    mx.eval(a)
+    time_fn(mx.max, a, 0)
+
+
+def time_min():
+    a = mx.random.uniform(shape=(32, 1024, 1024))
+    a[1, 1] = mx.nan
+    mx.eval(a)
+    time_fn(mx.min, a, 0)
+
+
 def time_negative():
     a = mx.random.uniform(shape=(10000, 1000))
     mx.eval(a)
@@ -108,6 +122,8 @@ if __name__ == "__main__":
 
     time_add()
     time_matmul()
+    time_min()
+    time_max()
     time_maximum()
     time_exp()
     time_negative()

docs/src/dev/custom_metal_kernels.rst

@@ -8,23 +8,26 @@ MLX supports writing custom Metal kernels through the Python and C++ APIs.
 Simple Example
 --------------
 
+.. currentmodule:: mlx.core
+
 Let's write a custom kernel that computes ``exp`` elementwise:
 
 .. code-block:: python
 
-  def exp_elementwise(a: mx.array):
-      source = """
-          uint elem = thread_position_in_grid.x;
-          T tmp = inp[elem];
-          out[elem] = metal::exp(tmp);
-      """
+  source = """
+      uint elem = thread_position_in_grid.x;
+      T tmp = inp[elem];
+      out[elem] = metal::exp(tmp);
+  """
 
-      kernel = mx.fast.metal_kernel(
-          name="myexp",
-          input_names=["inp"],
-          output_names=["out"],
-          source=source,
-      )
+  kernel = mx.fast.metal_kernel(
+      name="myexp",
+      input_names=["inp"],
+      output_names=["out"],
+      source=source,
+  )
+
+  def exp_elementwise(a: mx.array):
       outputs = kernel(
           inputs=[a],
           template=[("T", mx.float32)],
@@ -39,8 +42,13 @@ Let's write a custom kernel that computes ``exp`` elementwise:
   b = exp_elementwise(a)
   assert mx.allclose(b, mx.exp(a))
 
+Every time you make a kernel, a new Metal library is created and possibly
+JIT compiled. To reduce the overhead from that, build the kernel once with
+:func:`fast.metal_kernel` and then use it many times.
+
 .. note::
-    We are only required to pass the body of the Metal kernel in ``source``.
+   Only pass the body of the Metal kernel in ``source``. The function
+   signature is generated automatically.
 
 The full function signature will be generated using:
 
@@ -78,44 +86,51 @@ Putting this all together, the generated function signature for ``myexp`` is as
 
   template [[host_name("custom_kernel_myexp_float")]] [[kernel]] decltype(custom_kernel_myexp_float<float>) custom_kernel_myexp_float<float>;
 
-Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads <https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_ function.
-This means we will launch ``mx.prod(grid)`` threads, subdivided into ``threadgroup`` size threadgroups.
-For optimal performance, each thread group dimension should be less than or equal to the corresponding grid dimension.
+Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads
+<https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_
+function. This means we will launch ``mx.prod(grid)`` threads, subdivided into
+``threadgroup`` size threadgroups.  For optimal performance, each thread group
+dimension should be less than or equal to the corresponding grid dimension.
 
-Passing ``verbose=True`` to ``mx.fast.metal_kernel.__call__`` will print the generated code for debugging purposes.
+Passing ``verbose=True`` to :func:`ast.metal_kernel.__call__` will print the
+generated code for debugging purposes.
 
 Using Shape/Strides
 -------------------
 
-``mx.fast.metal_kernel`` supports an argument ``ensure_row_contiguous`` which is ``True`` by default.
-This will copy the ``mx.array`` inputs if needed before the kernel is launched to ensure that the memory layout is row contiguous.
-Generally this makes writing the kernel easier, since we don't have to worry about gaps or the ordering of the dims
-when indexing.
+:func:`fast.metal_kernel` supports an argument ``ensure_row_contiguous`` which
+is ``True`` by default. This will copy the array inputs if needed
+before the kernel is launched to ensure that the memory layout is row
+contiguous.  Generally this makes writing the kernel easier, since we don't
+have to worry about gaps or the ordering of the dims when indexing.
 
-If we want to avoid this copy, ``metal_kernel`` automatically passes ``a_shape``, ``a_strides`` and ``a_ndim`` for each
-input array ``a`` if any are present in ``source``.
-We can then use MLX's built in indexing utils to fetch the right elements for each thread.
+If we want to avoid this copy, :func:`fast.metal_kernel` automatically passes
+``a_shape``, ``a_strides`` and ``a_ndim`` for each input array ``a`` if any are
+present in ``source``. We can then use MLX's built in indexing utils to fetch
+the right elements for each thread.
 
-Let's convert ``myexp`` above to support arbitrarily strided arrays without relying on a copy from ``ensure_row_contiguous``:
+Let's convert ``myexp`` above to support arbitrarily strided arrays without
+relying on a copy from ``ensure_row_contiguous``:
 
 .. code-block:: python
+   
+  source = """
+      uint elem = thread_position_in_grid.x;
+      // Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
+      uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
+      T tmp = inp[loc];
+      // Output arrays are always row contiguous
+      out[elem] = metal::exp(tmp);
+  """
+
+  kernel = mx.fast.metal_kernel(
+      name="myexp_strided",
+      input_names=["inp"],
+      output_names=["out"],
+      source=source
+  )
 
   def exp_elementwise(a: mx.array):
-      source = """
-          uint elem = thread_position_in_grid.x;
-          // Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
-          uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
-          T tmp = inp[loc];
-          // Output arrays are always row contiguous
-          out[elem] = metal::exp(tmp);
-      """
-
-      kernel = mx.fast.metal_kernel(
-          name="myexp_strided",
-          input_names=["inp"],
-          output_names=["out"],
-          source=source
-      )
       outputs = kernel(
           inputs=[a],
           template=[("T", mx.float32)],
@@ -142,137 +157,139 @@ We'll start with the following MLX implementation using standard ops:
 
 .. code-block:: python
 
-    def grid_sample_ref(x, grid):
-        N, H_in, W_in, _ = x.shape
-        ix = ((grid[..., 0] + 1) * W_in - 1) / 2
-        iy = ((grid[..., 1] + 1) * H_in - 1) / 2
+  def grid_sample_ref(x, grid):
+      N, H_in, W_in, _ = x.shape
+      ix = ((grid[..., 0] + 1) * W_in - 1) / 2
+      iy = ((grid[..., 1] + 1) * H_in - 1) / 2
 
-        ix_nw = mx.floor(ix).astype(mx.int32)
-        iy_nw = mx.floor(iy).astype(mx.int32)
+      ix_nw = mx.floor(ix).astype(mx.int32)
+      iy_nw = mx.floor(iy).astype(mx.int32)
 
-        ix_ne = ix_nw + 1
-        iy_ne = iy_nw
+      ix_ne = ix_nw + 1
+      iy_ne = iy_nw
 
-        ix_sw = ix_nw
-        iy_sw = iy_nw + 1
+      ix_sw = ix_nw
+      iy_sw = iy_nw + 1
 
-        ix_se = ix_nw + 1
-        iy_se = iy_nw + 1
+      ix_se = ix_nw + 1
+      iy_se = iy_nw + 1
 
-        nw = (ix_se - ix)    * (iy_se - iy)
-        ne = (ix    - ix_sw) * (iy_sw - iy)
-        sw = (ix_ne - ix)    * (iy    - iy_ne)
-        se = (ix    - ix_nw) * (iy    - iy_nw)
+      nw = (ix_se - ix)    * (iy_se - iy)
+      ne = (ix    - ix_sw) * (iy_sw - iy)
+      sw = (ix_ne - ix)    * (iy    - iy_ne)
+      se = (ix    - ix_nw) * (iy    - iy_nw)
 
-        I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
-        I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
-        I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
-        I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
+      I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
+      I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
+      I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
+      I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
 
-        mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
-        mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
-        mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
-        mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
+      mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
+      mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
+      mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
+      mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
 
-        I_nw *= mask_nw[..., None]
-        I_ne *= mask_ne[..., None]
-        I_sw *= mask_sw[..., None]
-        I_se *= mask_se[..., None]
+      I_nw *= mask_nw[..., None]
+      I_ne *= mask_ne[..., None]
+      I_sw *= mask_sw[..., None]
+      I_se *= mask_se[..., None]
 
-        output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
+      output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
 
-        return output
+      return output
 
-Now let's use ``mx.custom_function`` together with ``mx.fast.metal_kernel``
+Now let's use :func:`custom_function` together with :func:`fast.metal_kernel`
 to write a fast GPU kernel for both the forward and backward passes.
 
 First we'll implement the forward pass as a fused kernel:
 
 .. code-block:: python
 
-    @mx.custom_function
-    def grid_sample(x, grid):
+  source = """
+      uint elem = thread_position_in_grid.x;
+      int H = x_shape[1];
+      int W = x_shape[2];
+      int C = x_shape[3];
+      int gH = grid_shape[1];
+      int gW = grid_shape[2];
 
-        assert x.ndim == 4, "`x` must be 4D."
-        assert grid.ndim == 4, "`grid` must be 4D."
+      int w_stride = C;
+      int h_stride = W * w_stride;
+      int b_stride = H * h_stride;
 
-        B, _, _, C = x.shape
-        _, gN, gM, D = grid.shape
-        out_shape = (B, gN, gM, C)
+      uint grid_idx = elem / C * 2;
+      float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+      float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
 
-        assert D == 2, "Last dim of `grid` must be size 2."
+      int ix_nw = floor(ix);
+      int iy_nw = floor(iy);
 
-        source = """
-            uint elem = thread_position_in_grid.x;
-            int H = x_shape[1];
-            int W = x_shape[2];
-            int C = x_shape[3];
-            int gH = grid_shape[1];
-            int gW = grid_shape[2];
+      int ix_ne = ix_nw + 1;
+      int iy_ne = iy_nw;
 
-            int w_stride = C;
-            int h_stride = W * w_stride;
-            int b_stride = H * h_stride;
+      int ix_sw = ix_nw;
+      int iy_sw = iy_nw + 1;
 
-            uint grid_idx = elem / C * 2;
-            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
-            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+      int ix_se = ix_nw + 1;
+      int iy_se = iy_nw + 1;
 
-            int ix_nw = floor(ix);
-            int iy_nw = floor(iy);
+      T nw = (ix_se - ix)    * (iy_se - iy);
+      T ne = (ix    - ix_sw) * (iy_sw - iy);
+      T sw = (ix_ne - ix)    * (iy    - iy_ne);
+      T se = (ix    - ix_nw) * (iy    - iy_nw);
 
-            int ix_ne = ix_nw + 1;
-            int iy_ne = iy_nw;
+      int batch_idx = elem / C / gH / gW * b_stride;
+      int channel_idx = elem % C;
+      int base_idx = batch_idx + channel_idx;
 
-            int ix_sw = ix_nw;
-            int iy_sw = iy_nw + 1;
+      T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
+      T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
+      T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
+      T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
 
-            int ix_se = ix_nw + 1;
-            int iy_se = iy_nw + 1;
+      I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
+      I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
+      I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
+      I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
 
-            T nw = (ix_se - ix)    * (iy_se - iy);
-            T ne = (ix    - ix_sw) * (iy_sw - iy);
-            T sw = (ix_ne - ix)    * (iy    - iy_ne);
-            T se = (ix    - ix_nw) * (iy    - iy_nw);
+      out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
+  """
 
-            int batch_idx = elem / C / gH / gW * b_stride;
-            int channel_idx = elem % C;
-            int base_idx = batch_idx + channel_idx;
+  kernel = mx.fast.metal_kernel(
+      name="grid_sample",
+      input_names=["x", "grid"],
+      output_names=["out"],
+      source=source,
+  )
 
-            T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
-            T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
-            T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
-            T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
+  @mx.custom_function
+  def grid_sample(x, grid):
 
-            I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
-            I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
-            I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
-            I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
+      assert x.ndim == 4, "`x` must be 4D."
+      assert grid.ndim == 4, "`grid` must be 4D."
 
-            out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
-        """
-        kernel = mx.fast.metal_kernel(
-            name="grid_sample",
-            input_names=["x", "grid"],
-            output_names=["out"],
-            source=source,
-        )
-        outputs = kernel(
-            inputs=[x, grid],
-            template=[("T", x.dtype)],
-            output_shapes=[out_shape],
-            output_dtypes=[x.dtype],
-            grid=(np.prod(out_shape), 1, 1),
-            threadgroup=(256, 1, 1),
-        )
-        return outputs[0]
+      B, _, _, C = x.shape
+      _, gN, gM, D = grid.shape
+      out_shape = (B, gN, gM, C)
+
+      assert D == 2, "Last dim of `grid` must be size 2."
+
+      outputs = kernel(
+          inputs=[x, grid],
+          template=[("T", x.dtype)],
+          output_shapes=[out_shape],
+          output_dtypes=[x.dtype],
+          grid=(np.prod(out_shape), 1, 1),
+          threadgroup=(256, 1, 1),
+      )
+      return outputs[0]
 
 For a reasonably sized input such as:
 
 .. code-block:: python
 
-    x.shape = (8, 1024, 1024, 64)
-    grid.shape = (8, 256, 256, 2)
+  x.shape = (8, 1024, 1024, 64)
+  grid.shape = (8, 256, 256, 2)
 
 On an M1 Max, we see a big performance improvement:
 
@@ -281,11 +298,11 @@ On an M1 Max, we see a big performance improvement:
 Grid Sample VJP
 ---------------
 
-Since we decorated ``grid_sample`` with ``mx.custom_function``, we can now define
-its custom vjp transform so MLX can differentiate it.
+Since we decorated ``grid_sample`` with :func:`custom_function`, we can now
+define its custom vjp transform so MLX can differentiate it.
 
 The backwards pass requires atomically updating ``x_grad``/``grid_grad`` and so
-requires a few extra ``mx.fast.metal_kernel`` features:
+requires a few extra :func:`fast.metal_kernel` features:
 
 * ``init_value=0``
     Initialize all of the kernel's outputs to this value before it runs. This allows us to update only part of the output arrays with the kernel.
@@ -299,128 +316,129 @@ We can then implement the backwards pass as follows:
 
 .. code-block:: python
 
-    @grid_sample.vjp
-    def grid_sample_vjp(primals, cotangent, _):
-        x, grid = primals
-        B, _, _, C = x.shape
-        _, gN, gM, D = grid.shape
+  source = """
+      uint elem = thread_position_in_grid.x;
+      int H = x_shape[1];
+      int W = x_shape[2];
+      int C = x_shape[3];
+      // Pad C to the nearest larger simdgroup size multiple
+      int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
 
-        assert D == 2, "Last dim of `grid` must be size 2."
+      int gH = grid_shape[1];
+      int gW = grid_shape[2];
 
-        source = """
-            uint elem = thread_position_in_grid.x;
-            int H = x_shape[1];
-            int W = x_shape[2];
-            int C = x_shape[3];
-            // Pad C to the nearest larger simdgroup size multiple
-            int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
+      int w_stride = C;
+      int h_stride = W * w_stride;
+      int b_stride = H * h_stride;
 
-            int gH = grid_shape[1];
-            int gW = grid_shape[2];
+      uint grid_idx = elem / C_padded * 2;
+      float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+      float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
 
-            int w_stride = C;
-            int h_stride = W * w_stride;
-            int b_stride = H * h_stride;
+      int ix_nw = floor(ix);
+      int iy_nw = floor(iy);
 
-            uint grid_idx = elem / C_padded * 2;
-            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
-            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+      int ix_ne = ix_nw + 1;
+      int iy_ne = iy_nw;
 
-            int ix_nw = floor(ix);
-            int iy_nw = floor(iy);
+      int ix_sw = ix_nw;
+      int iy_sw = iy_nw + 1;
 
-            int ix_ne = ix_nw + 1;
-            int iy_ne = iy_nw;
+      int ix_se = ix_nw + 1;
+      int iy_se = iy_nw + 1;
 
-            int ix_sw = ix_nw;
-            int iy_sw = iy_nw + 1;
+      T nw = (ix_se - ix)    * (iy_se - iy);
+      T ne = (ix    - ix_sw) * (iy_sw - iy);
+      T sw = (ix_ne - ix)    * (iy    - iy_ne);
+      T se = (ix    - ix_nw) * (iy    - iy_nw);
 
-            int ix_se = ix_nw + 1;
-            int iy_se = iy_nw + 1;
+      int batch_idx = elem / C_padded / gH / gW * b_stride;
+      int channel_idx = elem % C_padded;
+      int base_idx = batch_idx + channel_idx;
 
-            T nw = (ix_se - ix)    * (iy_se - iy);
-            T ne = (ix    - ix_sw) * (iy_sw - iy);
-            T sw = (ix_ne - ix)    * (iy    - iy_ne);
-            T se = (ix    - ix_nw) * (iy    - iy_nw);
+      T gix = T(0);
+      T giy = T(0);
+      if (channel_idx < C) {
+          int cot_index = elem / C_padded * C + channel_idx;
+          T cot = cotangent[cot_index];
+          if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
+              int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
 
-            int batch_idx = elem / C_padded / gH / gW * b_stride;
-            int channel_idx = elem % C_padded;
-            int base_idx = batch_idx + channel_idx;
+              T I_nw = x[offset];
+              gix -= I_nw * (iy_se - iy) * cot;
+              giy -= I_nw * (ix_se - ix) * cot;
+          }
+          if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
+              int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
 
-            T gix = T(0);
-            T giy = T(0);
-            if (channel_idx < C) {
-                int cot_index = elem / C_padded * C + channel_idx;
-                T cot = cotangent[cot_index];
-                if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
-                    int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
+              T I_ne = x[offset];
+              gix += I_ne * (iy_sw - iy) * cot;
+              giy -= I_ne * (ix - ix_sw) * cot;
+          }
+          if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
+              int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
 
-                    T I_nw = x[offset];
-                    gix -= I_nw * (iy_se - iy) * cot;
-                    giy -= I_nw * (ix_se - ix) * cot;
-                }
-                if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
-                    int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
+              T I_sw = x[offset];
+              gix -= I_sw * (iy - iy_ne) * cot;
+              giy += I_sw * (ix_ne - ix) * cot;
+          }
+          if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
+              int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
 
-                    T I_ne = x[offset];
-                    gix += I_ne * (iy_sw - iy) * cot;
-                    giy -= I_ne * (ix - ix_sw) * cot;
-                }
-                if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
-                    int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
+              T I_se = x[offset];
+              gix += I_se * (iy - iy_nw) * cot;
+              giy += I_se * (ix - ix_nw) * cot;
+          }
+      }
 
-                    T I_sw = x[offset];
-                    gix -= I_sw * (iy - iy_ne) * cot;
-                    giy += I_sw * (ix_ne - ix) * cot;
-                }
-                if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
-                    int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
+      T gix_mult = W / 2;
+      T giy_mult = H / 2;
 
-                    T I_se = x[offset];
-                    gix += I_se * (iy - iy_nw) * cot;
-                    giy += I_se * (ix - ix_nw) * cot;
-                }
-            }
+      // Reduce across each simdgroup first.
+      // This is much faster than relying purely on atomics.
+      gix = simd_sum(gix);
+      giy = simd_sum(giy);
 
-            T gix_mult = W / 2;
-            T giy_mult = H / 2;
+      if (thread_index_in_simdgroup == 0) {
+          atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
+          atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
+      }
+  """
+  kernel = mx.fast.metal_kernel(
+      name="grid_sample_grad",
+      input_names=["x", "grid", "cotangent"],
+      output_names=["x_grad", "grid_grad"],
+      source=source,
+      atomic_outputs=True,
+  )
 
-            // Reduce across each simdgroup first.
-            // This is much faster than relying purely on atomics.
-            gix = simd_sum(gix);
-            giy = simd_sum(giy);
+  @grid_sample.vjp
+  def grid_sample_vjp(primals, cotangent, _):
+      x, grid = primals
+      B, _, _, C = x.shape
+      _, gN, gM, D = grid.shape
 
-            if (thread_index_in_simdgroup == 0) {
-                atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
-                atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
-            }
-        """
-        kernel = mx.fast.metal_kernel(
-            name="grid_sample_grad",
-            input_names=["x", "grid", "cotangent"],
-            output_names=["x_grad", "grid_grad"],
-            source=source,
-            atomic_outputs=True,
-        )
-        # pad the output channels to simd group size
-        # so that our `simd_sum`s don't overlap.
-        simdgroup_size = 32
-        C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
-        grid_size = B * gN * gM * C_padded
-        outputs = kernel(
-            inputs=[x, grid, cotangent],
-            template=[("T", x.dtype)],
-            output_shapes=[x.shape, grid.shape],
-            output_dtypes=[x.dtype, x.dtype],
-            grid=(grid_size, 1, 1),
-            threadgroup=(256, 1, 1),
-            init_value=0,
-        )
-        return outputs[0], outputs[1]
+      assert D == 2, "Last dim of `grid` must be size 2."
+
+      # pad the output channels to simd group size
+      # so that our `simd_sum`s don't overlap.
+      simdgroup_size = 32
+      C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
+      grid_size = B * gN * gM * C_padded
+      outputs = kernel(
+          inputs=[x, grid, cotangent],
+          template=[("T", x.dtype)],
+          output_shapes=[x.shape, grid.shape],
+          output_dtypes=[x.dtype, x.dtype],
+          grid=(grid_size, 1, 1),
+          threadgroup=(256, 1, 1),
+          init_value=0,
+      )
+      return outputs[0], outputs[1]
 
 There's an even larger speed up for the vjp:
 

docs/src/dev/extensions.rst

@@ -138,13 +138,13 @@ more concrete:
         * representing the vectorized computation and the axis which
         * corresponds to the output vectorized dimension.
         */
-        virtual std::pair<std::vector<array>, std::vector<int>> vmap(
+        std::pair<std::vector<array>, std::vector<int>> vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) override;
 
-        /** Print the primitive. */
-        void print(std::ostream& os) override {
-            os << "Axpby";
+        /** The name of primitive. */
+        const char* name() const override {
+          return "Axpby";
         }
 
         /** Equivalence check **/
@@ -397,11 +397,11 @@ below.
         std::ostringstream kname;
         kname << "axpby_" << "general_" << type_to_name(out);
 
-        // Make sure the metal library is available
-        d.register_library("mlx_ext");
+        // Load the metal library
+        auto lib = d.get_library("mlx_ext");
 
         // Make a kernel from this metal library
-        auto kernel = d.get_kernel(kname.str(), "mlx_ext");
+        auto kernel = d.get_kernel(kname.str(), lib);
 
         // Prepare to encode kernel
         auto& compute_encoder = d.get_command_encoder(s.index);

docs/src/install.rst

@@ -23,13 +23,24 @@ To install from PyPI you must meet the following requirements:
     MLX is only available on devices running macOS >= 13.5
     It is highly recommended to use macOS 14 (Sonoma)
 
+CUDA
+^^^^
 
-MLX is also available on conda-forge. To install MLX with conda do:
+MLX has a CUDA backend which you can use on any Linux platform with CUDA 12
+and SM 7.0 (Volta) and up. To install MLX with CUDA support, run:
 
 .. code-block:: shell
 
-   conda install conda-forge::mlx
+    pip install "mlx[cuda]"
 
+CPU-only (Linux)
+^^^^^^^^^^^^^^^^
+
+For a CPU-only version of MLX that runs on Linux use:
+
+.. code-block:: shell
+
+    pip install "mlx[cpu]"
 
 Troubleshooting
 ^^^^^^^^^^^^^^^
@@ -65,6 +76,8 @@ Build Requirements
 Python API
 ^^^^^^^^^^
 
+.. _python install:
+
 To build and install the MLX python library from source, first, clone MLX from
 `its GitHub repo <https://github.com/ml-explore/mlx>`_:
 
@@ -76,20 +89,20 @@ Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
+  pip install .
 
 For developing, install the package with development dependencies, and use an
 editable install:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
+  pip install -e ".[dev]"
 
 Once the development dependencies are installed, you can build faster with:
 
 .. code-block:: shell
 
- CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
+ python setup.py build_ext --inplace
 
 Run the tests with:
 
@@ -107,6 +120,8 @@ IDE:
 C++ API
 ^^^^^^^
 
+.. _cpp install:
+
 Currently, MLX must be built and installed from source.
 
 Similarly to the python library, to build and install the MLX C++ library start
@@ -185,6 +200,7 @@ should point to the path to the built metal library.
 
       xcrun -sdk macosx --show-sdk-version
 
+
 Binary Size Minimization
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -213,6 +229,50 @@ be anwywhere from a few hundred millisecond to a few seconds depending on the
 application. Once a kernel is compiled, it will be cached by the system. The
 Metal kernel cache persists across reboots.
 
+Linux
+^^^^^
+
+To build from source on Linux (CPU only), install the BLAS and LAPACK headers.
+For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   apt-get update -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+
+From here follow the instructions to install either the :ref:`Python <python
+install>` or :ref:`C++ <cpp install>` APIs.
+
+CUDA
+^^^^
+
+To build from source on Linux with CUDA, install the BLAS and LAPACK headers
+and the CUDA toolkit. For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+   dpkg -i cuda-keyring_1.1-1_all.deb
+   apt-get update -y
+   apt-get -y install cuda-toolkit-12-9
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+
+
+When building either the Python or C++ APIs make sure to pass the cmake flag
+``MLX_BUILD_CUDA=ON``. For example, to build the Python API run:
+
+.. code-block:: shell
+
+  CMAKE_ARGS="-DMLX_BUILD_CUDA=ON" pip install -e ".[dev]"
+
+To build the C++ package run:
+
+.. code-block:: shell
+
+   mkdir -p build && cd build
+   cmake .. -DMLX_BUILD_CUDA=ON && make -j
+
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 

docs/src/python/optimizers/common_optimizers.rst

@@ -19,3 +19,4 @@ Common Optimizers
    Adamax
    Lion
    MultiOptimizer
+   Muon

docs/src/usage/indexing.rst

@@ -107,6 +107,16 @@ same array:
   >>> a
   array([1, 2, 0], dtype=int32)
 
+
+Note, unlike NumPy, updates to the same location are nondeterministic:
+
+.. code-block:: shell
+
+  >>> a = mx.array([1, 2, 3])
+  >>> a[[0, 0]] = mx.array([4, 5])
+
+The first element of ``a`` could be ``4`` or ``5``.
+
 Transformations of functions which use in-place updates are allowed and work as
 expected. For example:
 

examples/extensions/axpby/axpby.cpp

@@ -172,11 +172,11 @@ void Axpby::eval_gpu(
   kname << (contiguous_kernel ? "contiguous_" : "general_");
   kname << type_to_name(out);
 
-  // Make sure the metal library is available
-  d.register_library("mlx_ext");
+  // Load the metal library
+  auto lib = d.get_library("mlx_ext");
 
   // Make a kernel from this metal library
-  auto kernel = d.get_kernel(kname.str(), "mlx_ext");
+  auto kernel = d.get_kernel(kname.str(), lib);
 
   // Prepare to encode kernel
   auto& compute_encoder = d.get_command_encoder(s.index);

examples/extensions/axpby/axpby.h

@@ -74,9 +74,9 @@ class Axpby : public mx::Primitive {
       const std::vector<mx::array>& inputs,
       const std::vector<int>& axes) override;
 
-  /** Print the primitive. */
-  void print(std::ostream& os) override {
-    os << "Axpby";
+  /** The name of primitive. */
+  const char* name() const override {
+    return "Axpby";
   }
 
   /** Equivalence check **/

mlx/CMakeLists.txt

@@ -55,6 +55,9 @@ endif()
 
 if(MLX_BUILD_CUDA)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda)
+else()
+  target_sources(mlx
+                 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda/no_cuda.cpp)
 endif()
 
 if(MLX_BUILD_METAL OR MLX_BUILD_CUDA)

mlx/backend/common/compiled.cpp

@@ -14,6 +14,8 @@ void print_constant(std::ostream& os, const array& x) {
       return print_float_constant<float16_t>(os, x);
     case bfloat16:
       return print_float_constant<bfloat16_t>(os, x);
+    case float64:
+      return print_float_constant<double>(os, x);
     case complex64:
       return print_complex_constant<complex64_t>(os, x);
     case int8:
@@ -50,6 +52,8 @@ std::string get_type_string(Dtype d) {
       return "float16_t";
     case bfloat16:
       return "bfloat16_t";
+    case float64:
+      return "double";
     case complex64:
       return "complex64_t";
     case bool_:

mlx/backend/common/compiled.h

@@ -18,8 +18,12 @@ std::string get_type_string(Dtype d);
 template <typename T>
 void print_float_constant(std::ostream& os, const array& x) {
   auto old_precision = os.precision();
-  os << std::setprecision(std::numeric_limits<float>::digits10 + 1)
-     << x.item<T>() << std::setprecision(old_precision);
+  if constexpr (std::is_same_v<T, double>) {
+    os << std::setprecision(std::numeric_limits<double>::digits10 + 1);
+  } else {
+    os << std::setprecision(std::numeric_limits<float>::digits10 + 1);
+  }
+  os << x.item<T>() << std::setprecision(old_precision);
 }
 
 template <typename T>

mlx/backend/common/copy.h

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "mlx/array.h"
+#include "mlx/backend/common/utils.h"
 
 namespace mlx::core {
 
@@ -26,7 +26,7 @@ inline bool set_copy_output_data(const array& in, array& out, CopyType ctype) {
   if (ctype == CopyType::Vector) {
     // If the input is donateable, we are doing a vector copy and the types
     // have the same size, then the input buffer can hold the output.
-    if (in.is_donatable() && in.itemsize() == out.itemsize()) {
+    if (is_donatable(in, out)) {
       out.copy_shared_buffer(in);
       return true;
     } else {

mlx/backend/common/matmul.h

@@ -0,0 +1,67 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/utils.h"
+
+#include <sstream>
+
+namespace mlx::core {
+
+inline std::tuple<Shape, Strides, Strides> collapse_batches(
+    const array& a,
+    const array& b) {
+  if (a.ndim() == 2) {
+    return {{1}, {0}, {0}};
+  }
+
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
+  Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
+  Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
+
+  auto [batch_shape, batch_strides] =
+      collapse_contiguous_dims(A_bshape, std::vector{A_bstride, B_bstride});
+
+  auto a_batch_strides = batch_strides[0];
+  auto b_batch_strides = batch_strides[1];
+
+  if (batch_shape.empty()) {
+    batch_shape.push_back(1);
+    a_batch_strides.push_back(0);
+    b_batch_strides.push_back(0);
+  }
+
+  return std::make_tuple(batch_shape, a_batch_strides, b_batch_strides);
+}
+
+inline std::tuple<Shape, Strides, Strides, Strides>
+collapse_batches(const array& a, const array& b, const array& c) {
+  if (a.ndim() == 2) {
+    return {{1}, {0}, {0}, {0}};
+  }
+
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
+  Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
+  Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
+  Strides C_bstride{c.strides().begin(), c.strides().end() - 2};
+
+  auto [batch_shape, batch_strides] = collapse_contiguous_dims(
+      A_bshape, std::vector{A_bstride, B_bstride, C_bstride});
+
+  auto A_batch_stride = batch_strides[0];
+  auto B_batch_stride = batch_strides[1];
+  auto C_batch_stride = batch_strides[2];
+
+  if (batch_shape.empty()) {
+    batch_shape.push_back(1);
+    A_batch_stride.push_back(0);
+    B_batch_stride.push_back(0);
+    C_batch_stride.push_back(0);
+  }
+
+  return std::make_tuple(
+      batch_shape, A_batch_stride, B_batch_stride, C_batch_stride);
+}
+
+} // namespace mlx::core

mlx/backend/common/reduce.cpp

@@ -5,11 +5,9 @@
 namespace mlx::core {
 
 std::pair<Shape, Strides> shapes_without_reduction_axes(
-    const array& x,
+    Shape shape,
+    Strides strides,
     const std::vector<int>& axes) {
-  auto shape = x.shape();
-  auto strides = x.strides();
-
   for (int i = axes.size() - 1; i >= 0; i--) {
     int a = axes[i];
     shape.erase(shape.begin() + a);
@@ -19,6 +17,15 @@ std::pair<Shape, Strides> shapes_without_reduction_axes(
   return std::make_pair(shape, strides);
 }
 
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    const array& x,
+    const std::vector<int>& axes) {
+  auto shape = x.shape();
+  auto strides = x.strides();
+  return shapes_without_reduction_axes(
+      std::move(shape), std::move(strides), axes);
+}
+
 ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
   // The data is all there and we are reducing over everything
   if (x.size() == x.data_size() && axes.size() == x.ndim() &&

mlx/backend/common/reduce.h

@@ -51,5 +51,9 @@ ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes);
 std::pair<Shape, Strides> shapes_without_reduction_axes(
     const array& x,
     const std::vector<int>& axes);
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    Shape shape,
+    Strides strides,
+    const std::vector<int>& axes);
 
 } // namespace mlx::core

mlx/backend/common/unary.h

@@ -0,0 +1,26 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/allocator.h"
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+inline void set_unary_output_data(const array& in, array& out) {
+  if (in.flags().contiguous) {
+    if (is_donatable(in, out)) {
+      out.copy_shared_buffer(in);
+    } else {
+      out.set_data(
+          allocator::malloc(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+    }
+  } else {
+    out.set_data(allocator::malloc(out.nbytes()));
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/utils.cpp

@@ -1,14 +1,20 @@
 // Copyright © 2023-2024 Apple Inc.
 
+#include <dlfcn.h>
+
 #include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive) {
-  std::ostringstream op_t;
-  primitive->print(op_t);
-  return op_t.str();
+std::filesystem::path current_binary_dir() {
+  static std::filesystem::path binary_dir = []() {
+    Dl_info info;
+    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+      throw std::runtime_error("Unable to get current binary dir.");
+    }
+    return std::filesystem::path(info.dli_fname).parent_path();
+  }();
+  return binary_dir;
 }
 
 std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
@@ -199,14 +205,27 @@ Dims get_2d_grid_dims_common(
       }
     }
   }
-  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX || divisor > 1) {
+  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX) {
     throw std::runtime_error("Unable to safely factor shape.");
   }
   if (grid_y > grid_x) {
     std::swap(grid_x, grid_y);
   }
+  if (divisor > 1) {
+    grid_x = ((grid_x + divisor - 1) / divisor) * divisor;
+  }
   return std::make_tuple(
       static_cast<uint32_t>(grid_x), static_cast<uint32_t>(grid_y), 1);
 }
 
+std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
+  auto [bx, by, bz] = get_block_dims_common(dim0, dim1, dim2);
+  auto gx = (dim0 + bx - 1) / bx;
+  auto gy = (dim1 + by - 1) / by;
+  auto gz = (dim2 + bz - 1) / bz;
+
+  return std::make_pair(
+      std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
+}
+
 } // namespace mlx::core

mlx/backend/common/utils.h

@@ -2,6 +2,7 @@
 
 #pragma once
 
+#include <filesystem>
 #include <tuple>
 #include <vector>
 
@@ -9,7 +10,8 @@
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive);
+// Return the directory that contains current shared library.
+std::filesystem::path current_binary_dir();
 
 inline int64_t
 elem_to_loc(int elem, const Shape& shape, const Strides& strides) {
@@ -95,6 +97,9 @@ Dims get_2d_grid_dims_common(
     const Strides& strides,
     size_t divisor);
 
+// Get both the block and a grid of blocks that covers dim0, dim1 and dim2.
+std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2);
+
 struct ContiguousIterator {
   inline void step() {
     int dims = shape_.size();

mlx/backend/cpu/cholesky.cpp

@@ -20,7 +20,7 @@ void cholesky_impl(const array& a, array& factor, bool upper, Stream stream) {
 
   // The decomposition is computed in place, so just copy the input to the
   // output.
-  copy(
+  copy_cpu(
       a,
       factor,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/compiled.cpp

@@ -231,7 +231,7 @@ inline void build_kernel(
       os << "static_cast<" << get_type_string(x.dtype()) << ">(tmp_"
          << namer.get_name(x.inputs()[0]) << ");" << std::endl;
     } else {
-      x.primitive().print(os);
+      os << x.primitive().name();
       os << "()(";
       for (int i = 0; i < x.inputs().size() - 1; i++) {
         os << "tmp_" << namer.get_name(x.inputs()[i]) << ", ";

mlx/backend/cpu/conv.cpp

@@ -883,7 +883,7 @@ void explicit_gemm_conv_1D_cpu(
   // Fill with zeros
   std::vector<array> temps;
   temps.push_back(array(0, conv_dtype));
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = padding_lo[0] * in_padded.strides()[1];
@@ -895,7 +895,7 @@ void explicit_gemm_conv_1D_cpu(
       in_padded_slice.size(),
       data_offset);
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
   temps.push_back(in_padded_slice);
 
   // Make strided view
@@ -920,7 +920,7 @@ void explicit_gemm_conv_1D_cpu(
   // Materialize strided view
   Shape strided_reshape = {N * oH, wH * C};
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -938,13 +938,13 @@ void explicit_gemm_conv_1D_cpu(
         wt.size(),
         0);
     gemm_wt = array(wt_transpose.shape(), float32, nullptr, {});
-    copy(wt_transpose, gemm_wt, CopyType::General, stream);
+    copy_cpu(wt_transpose, gemm_wt, CopyType::General, stream);
     temps.push_back(gemm_wt);
   } else if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
@@ -991,7 +991,7 @@ void explicit_gemm_conv_1D_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }
@@ -1029,7 +1029,7 @@ void explicit_gemm_conv_2D_cpu(
   // Fill with zeros
   std::vector<array> temps;
   temps.push_back(array(0, conv_dtype));
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = padding_lo[0] * in_padded.strides()[1] +
@@ -1044,7 +1044,7 @@ void explicit_gemm_conv_2D_cpu(
   temps.push_back(in_padded_slice);
 
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
 
   // Make strided view
   Shape strided_shape = {N, oH, oW, wH, wW, C};
@@ -1065,7 +1065,7 @@ void explicit_gemm_conv_2D_cpu(
   // Materialize strided view
   Shape strided_reshape = {N * oH * oW, wH * wW * C};
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -1076,7 +1076,7 @@ void explicit_gemm_conv_2D_cpu(
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
@@ -1116,7 +1116,7 @@ void explicit_gemm_conv_2D_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }
@@ -1156,7 +1156,7 @@ void explicit_gemm_conv_ND_cpu(
 
   // Fill with zeros
   std::vector<array> temps = {array(0, conv_dtype)};
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = 0;
@@ -1173,7 +1173,7 @@ void explicit_gemm_conv_ND_cpu(
       data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
   temps.push_back(in_padded_slice);
 
   // Make strided view
@@ -1212,7 +1212,7 @@ void explicit_gemm_conv_ND_cpu(
   }
 
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -1223,13 +1223,13 @@ void explicit_gemm_conv_ND_cpu(
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
   if (flip) {
     auto gemm_wt_ = array(gemm_wt.shape(), float32, nullptr, {});
-    copy(gemm_wt, gemm_wt_, CopyType::Vector, stream);
+    copy_cpu(gemm_wt, gemm_wt_, CopyType::Vector, stream);
     temps.push_back(gemm_wt_);
 
     // Calculate the total size of the spatial dimensions
@@ -1284,7 +1284,7 @@ void explicit_gemm_conv_ND_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }

mlx/backend/cpu/copy.cpp

@@ -295,7 +295,11 @@ inline void copy_inplace_dispatch(
 
 } // namespace
 
-void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy_cpu_inplace(
+    const array& src,
+    array& dst,
+    CopyType ctype,
+    Stream stream) {
   auto& encoder = cpu::get_command_encoder(stream);
   encoder.set_input_array(src);
   encoder.set_output_array(dst);
@@ -305,7 +309,7 @@ void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
        ctype]() mutable { copy_inplace_dispatch(src, dst, ctype); });
 }
 
-void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy_cpu(const array& src, array& dst, CopyType ctype, Stream stream) {
   bool donated = set_copy_output_data(src, dst, ctype);
   if (donated && src.dtype() == dst.dtype()) {
     // If the output has the same type as the input then there is nothing to
@@ -315,10 +319,10 @@ void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
   if (ctype == CopyType::GeneralGeneral) {
     ctype = CopyType::General;
   }
-  copy_inplace(src, dst, ctype, stream);
+  copy_cpu_inplace(src, dst, ctype, stream);
 }
 
-void copy_inplace(
+void copy_cpu_inplace(
     const array& src,
     array& dst,
     const Shape& data_shape,

mlx/backend/cpu/copy.h

@@ -10,10 +10,14 @@
 
 namespace mlx::core {
 
-void copy(const array& src, array& dst, CopyType ctype, Stream stream);
-void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_cpu(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_cpu_inplace(
+    const array& src,
+    array& dst,
+    CopyType ctype,
+    Stream stream);
 
-void copy_inplace(
+void copy_cpu_inplace(
     const array& src,
     array& dst,
     const Shape& data_shape,

mlx/backend/cpu/distributed.cpp

@@ -14,7 +14,7 @@ std::pair<array, bool> ensure_row_contiguous(const array& arr, Stream stream) {
     return {arr, false};
   } else {
     array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-    copy(arr, arr_copy, CopyType::General, stream);
+    copy_cpu(arr, arr_copy, CopyType::General, stream);
     return {arr_copy, true};
   }
 };
@@ -35,7 +35,7 @@ void AllReduce::eval_cpu(
       return in;
     } else {
       array arr_copy(in.shape(), in.dtype(), nullptr, {});
-      copy(in, arr_copy, CopyType::General, s);
+      copy_cpu(in, arr_copy, CopyType::General, s);
       out.copy_shared_buffer(arr_copy);
       return arr_copy;
     }

mlx/backend/cpu/eig.cpp

@@ -135,7 +135,7 @@ void Eig::eval_cpu(
       : array(a.shape(), complex64, nullptr, {});
 
   auto a_copy = array(a.shape(), a.dtype(), nullptr, {});
-  copy(
+  copy_cpu(
       a,
       a_copy,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/eigh.cpp

@@ -196,7 +196,7 @@ void Eigh::eval_cpu(
 
   values.set_data(allocator::malloc(values.nbytes()));
 
-  copy(
+  copy_cpu(
       a,
       vectors,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/hadamard.cpp

@@ -96,7 +96,7 @@ void Hadamard::eval_cpu(const std::vector<array>& inputs, array& out) {
   if (in.flags().row_contiguous && in.is_donatable()) {
     out.copy_shared_buffer(in);
   } else {
-    copy(
+    copy_cpu(
         in,
         out,
         in.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/indexing.cpp

@@ -517,7 +517,7 @@ void Scatter::eval_cpu(const std::vector<array>& inputs, array& out) {
   // Copy src into out (copy allocates memory for out)
   auto ctype =
       src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-  copy(src, out, ctype, stream());
+  copy_cpu(src, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   std::vector<array> inds;
@@ -686,7 +686,7 @@ void ScatterAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
   // Copy src into out (copy allocates memory for out)
   auto ctype =
       src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-  copy(src, out, ctype, stream());
+  copy_cpu(src, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_input_array(idx);

mlx/backend/cpu/inverse.cpp

@@ -115,7 +115,7 @@ void inverse_impl(
   //   (A⁻¹)ᵀ = (Aᵀ)⁻¹
 
   // The inverse is computed in place, so just copy the input to the output.
-  copy(
+  copy_cpu(
       a,
       inv,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/logsumexp.cpp

@@ -88,7 +88,7 @@ void LogSumExp::eval_cpu(const std::vector<array>& inputs, array& out) {
       return x;
     } else {
       auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General, s);
+      copy_cpu(x, x_copy, CopyType::General, s);
       encoder.add_temporary(x_copy);
       return x_copy;
     }

mlx/backend/cpu/luf.cpp

@@ -31,7 +31,7 @@ void luf_impl(
   strides[ndim - 1] = M;
   strides[ndim - 2] = 1;
   lu.set_data(allocator::malloc(lu.nbytes()), lu.nbytes(), strides, flags);
-  copy_inplace(
+  copy_cpu_inplace(
       a,
       lu,
       a.shape(),

mlx/backend/cpu/masked_mm.cpp

@@ -6,6 +6,7 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/gemm.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
@@ -52,6 +53,58 @@ inline void mask_matrix(
   }
 }
 
+template <typename T>
+inline void segmented_mm(
+    const T* a,
+    const T* b,
+    const uint32_t* segments,
+    T* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides,
+    size_t num_segments,
+    const Shape& segments_shape,
+    const Strides& segments_strides) {
+  int ndim = a_shape.size();
+  Shape a_copy = a_shape;
+  Shape b_copy = b_shape;
+  int32_t M = a_copy[ndim - 2];
+  int32_t N = b_copy[ndim - 1];
+  for (int i = 0; i < num_segments; i++) {
+    uint32_t k_start =
+        segments[elem_to_loc(2 * i, segments_shape, segments_strides)];
+    uint32_t k_end =
+        segments[elem_to_loc(2 * i + 1, segments_shape, segments_strides)];
+    if (k_end <= k_start) {
+      std::fill_n(out + i * M * N, M * N, T(0));
+      continue;
+    }
+    a_copy[ndim - 1] = k_end - k_start;
+    b_copy[ndim - 2] = k_end - k_start;
+    matmul<T>(
+        a + k_start * a_strides[ndim - 1],
+        b + k_start * b_strides[ndim - 2],
+        out + i * M * N,
+        a_transposed,
+        b_transposed,
+        lda,
+        ldb,
+        N,
+        1.0,
+        0.0,
+        1,
+        a_copy,
+        a_strides,
+        b_copy,
+        b_strides);
+  }
+}
+
 } // namespace
 
 void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -71,20 +124,20 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
         if (!expand_all && stx == arr.shape(-1) && sty == 1) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(false, stx, arr_copy, true);
           }
           return std::make_tuple(false, stx, arr, false);
         } else if (!expand_all && stx == 1 && sty == arr.shape(-2)) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(true, sty, arr_copy, true);
           }
           return std::make_tuple(true, sty, arr, false);
         } else {
           array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-          copy(arr, arr_copy, CopyType::General, s);
+          copy_cpu(arr, arr_copy, CopyType::General, s);
           int64_t stx = arr.shape(-1);
           return std::make_tuple(false, stx, arr_copy, true);
         }
@@ -333,7 +386,7 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       int64_t stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -437,4 +490,121 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   encoder.add_temporaries(std::move(temps));
 }
 
+void SegmentedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& s = stream();
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto check_transpose = [&s, &encoder](const array& x) {
+    auto stx = x.strides()[x.ndim() - 2];
+    auto sty = x.strides()[x.ndim() - 1];
+    if (stx == x.shape(-1) && sty == 1) {
+      return std::make_tuple(false, stx, x);
+    } else if (stx == 1 && sty == x.shape(-2)) {
+      return std::make_tuple(true, sty, x);
+    } else {
+      array xc(x.shape(), x.dtype(), nullptr, {});
+      copy_cpu(x, xc, CopyType::General, s);
+      encoder.add_temporary(xc);
+      int64_t stx = x.shape(-1);
+      return std::make_tuple(false, stx, xc);
+    }
+  };
+
+  auto [a_transposed, lda, a] = check_transpose(inputs[0]);
+  auto [b_transposed, ldb, b] = check_transpose(inputs[1]);
+  auto& segments = inputs[2];
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(segments);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    segments = array::unsafe_weak_copy(segments),
+                    out_ptr = out.data<void>(),
+                    a_transposed = a_transposed,
+                    b_transposed = b_transposed,
+                    lda = lda,
+                    ldb = ldb]() {
+    switch (a.dtype()) {
+      case float64:
+        segmented_mm<double>(
+            a.data<double>(),
+            b.data<double>(),
+            segments.data<uint32_t>(),
+            static_cast<double*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float32:
+        segmented_mm<float>(
+            a.data<float>(),
+            b.data<float>(),
+            segments.data<uint32_t>(),
+            static_cast<float*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float16:
+        segmented_mm<float16_t>(
+            a.data<float16_t>(),
+            b.data<float16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<float16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case bfloat16:
+        segmented_mm<bfloat16_t>(
+            a.data<bfloat16_t>(),
+            b.data<bfloat16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<bfloat16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      default:
+        throw std::invalid_argument(
+            "Segmented mm supports only real float types.");
+    }
+  });
+}
+
 } // namespace mlx::core

mlx/backend/cpu/matmul.cpp

@@ -81,7 +81,7 @@ void matmul_general(
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, stream);
+      copy_cpu(arr, temps.back(), CopyType::General, stream);
       stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -142,7 +142,7 @@ void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   CopyType ctype = c.data_size() == 1
       ? CopyType::Scalar
       : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-  copy(c, out, ctype, stream());
+  copy_cpu(c, out, ctype, stream());
   if (inputs[0].shape(-1) == 0) {
     return;
   }

mlx/backend/cpu/primitives.cpp

@@ -22,7 +22,7 @@ void reshape(const array& in, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
   if (copy_necessary) {
     out.set_data(allocator::malloc(out.nbytes()));
-    copy_inplace(in, out, CopyType::General, out.primitive().stream());
+    copy_cpu_inplace(in, out, CopyType::General, out.primitive().stream());
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
@@ -175,7 +175,7 @@ void AsType::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -198,7 +198,7 @@ void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
     size_t data_offset = strides[axis_] * sizes[i];
     out_slice.copy_shared_buffer(
         out, strides, flags, out_slice.size(), data_offset);
-    copy_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
+    copy_cpu_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
   }
 }
 
@@ -211,7 +211,7 @@ void Contiguous::eval_cpu(const std::vector<array>& inputs, array& out) {
        (allow_col_major_ && in.flags().col_contiguous))) {
     out.copy_shared_buffer(in);
   } else {
-    copy(in, out, CopyType::General, stream());
+    copy_cpu(in, out, CopyType::General, stream());
   }
 }
 
@@ -235,7 +235,7 @@ void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
   } else {
     ctype = CopyType::General;
   }
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -251,7 +251,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(val.dtype() == in.dtype() && in.dtype() == out.dtype());
 
   // Fill output with val
-  copy(val, out, CopyType::Scalar, stream());
+  copy_cpu(val, out, CopyType::Scalar, stream());
 
   // Find offset for start of input values
   size_t data_offset = 0;
@@ -266,7 +266,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
       out, out.strides(), out.flags(), out_slice.size(), data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
+  copy_cpu_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
 }
 
 void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -340,7 +340,7 @@ void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
   out.set_data(allocator::malloc(out.nbytes()));
   auto [in_offset, donated] =
       compute_dynamic_offset(inputs[1], in.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ in,
       /* array& dst = */ out,
       /* const Shape& data_shape = */ out.shape(),
@@ -372,11 +372,11 @@ void DynamicSliceUpdate::eval_cpu(
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   auto [out_offset, donated] =
       compute_dynamic_offset(inputs[2], out.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -412,14 +412,14 @@ void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   // Calculate out strides, initial offset and if copy needs to be made
   auto [data_offset, out_strides] =
       prepare_slice(out, start_indices_, strides_);
 
   // Do copy
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -456,9 +456,9 @@ void View::eval_cpu(const std::vector<array>& inputs, array& out) {
     if (in.dtype() == bool_) {
       auto in_tmp = array(in.shape(), uint8, nullptr, {});
       in_tmp.copy_shared_buffer(in);
-      copy_inplace(in_tmp, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in_tmp, tmp, CopyType::General, stream());
     } else {
-      copy_inplace(in, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in, tmp, CopyType::General, stream());
     }
 
     auto flags = out.flags();

mlx/backend/cpu/qrf.cpp

@@ -26,7 +26,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
   strides[in.ndim() - 2] = 1;
   strides[in.ndim() - 1] = M;
   in.set_data(allocator::malloc(in.nbytes()), in.nbytes(), strides, flags);
-  copy_inplace(a, in, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(a, in, CopyType::GeneralGeneral, stream);
   auto& encoder = cpu::get_command_encoder(stream);
   q.set_data(allocator::malloc(q.nbytes()));
   r.set_data(allocator::malloc(r.nbytes()));

mlx/backend/cpu/quantized.cpp

@@ -529,7 +529,7 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
       return arr;
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       return temps.back();
     }
   };
@@ -579,7 +579,7 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return arr;
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       return temps.back();
     }
   };
@@ -713,7 +713,7 @@ void fast::AffineQuantize::eval_cpu(
       return std::make_pair(arr, false);
     } else {
       array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General, s);
+      copy_cpu(arr, arr_copy, CopyType::General, s);
       return std::make_pair(arr_copy, true);
     }
   };

mlx/backend/cpu/reduce.cpp

@@ -325,7 +325,15 @@ struct MaxReduce {
   };
 
   template <int N, typename T>
-  T operator()(simd::Simd<T, N> x) {
+  std::enable_if_t<std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    return simd::max(x);
+  };
+
+  template <int N, typename T>
+  std::enable_if_t<!std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    if (simd::any(x != x)) {
+      return static_cast<T>(NAN);
+    }
     return simd::max(x);
   };
 };
@@ -342,7 +350,15 @@ struct MinReduce {
   };
 
   template <int N, typename T>
-  T operator()(simd::Simd<T, N> x) {
+  std::enable_if_t<std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    return simd::min(x);
+  };
+
+  template <int N, typename T>
+  std::enable_if_t<!std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    if (simd::any(x != x)) {
+      return static_cast<T>(NAN);
+    }
     return simd::min(x);
   };
 };
@@ -527,10 +543,10 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
             reduce_dispatch_min_max<uint64_t>(in, out, reduce_type_, axes_);
             break;
           case int8:
-            reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
+            reduce_dispatch_min_max<int8_t>(in, out, reduce_type_, axes_);
             break;
           case int16:
-            reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
+            reduce_dispatch_min_max<int16_t>(in, out, reduce_type_, axes_);
             break;
           case int32:
             reduce_dispatch_min_max<int32_t>(in, out, reduce_type_, axes_);

mlx/backend/cpu/scan.cpp

@@ -251,7 +251,7 @@ void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto in = inputs[0];
   if (!in.flags().row_contiguous) {
     array arr_copy(in.shape(), in.dtype(), nullptr, {});
-    copy(in, arr_copy, CopyType::General, stream());
+    copy_cpu(in, arr_copy, CopyType::General, stream());
     in = arr_copy;
     encoder.add_temporary(arr_copy);
   }

mlx/backend/cpu/softmax.cpp

@@ -132,7 +132,7 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
       return x;
     } else {
       array x_copy(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General, s);
+      copy_cpu(x, x_copy, CopyType::General, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }

mlx/backend/cpu/sort.cpp

@@ -334,8 +334,10 @@ void Sort::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  CopyType ctype = (in.flags().contiguous && in.strides()[axis_] != 0)
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_cpu(in, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_output_array(out);
@@ -426,8 +428,10 @@ void Partition::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  CopyType ctype = (in.flags().contiguous && in.strides()[axis_] != 0)
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_cpu(in, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_output_array(out);

mlx/backend/cpu/svd.cpp

@@ -31,7 +31,7 @@ void svd_impl(
 
   // lapack clobbers the input, so we have to make a copy.
   array in(a.shape(), a.dtype(), nullptr, {});
-  copy(
+  copy_cpu(
       a,
       in,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/unary.h

@@ -2,32 +2,13 @@
 
 #pragma once
 
-#include "mlx/allocator.h"
-#include "mlx/array.h"
-#include "mlx/backend/common/utils.h"
+#include "mlx/backend/common/unary.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
 #include "mlx/utils.h"
 
 namespace mlx::core {
 
-void set_unary_output_data(const array& in, array& out) {
-  if (in.flags().contiguous) {
-    if (is_donatable(in, out)) {
-      out.copy_shared_buffer(in);
-    } else {
-      auto size = in.data_size();
-      out.set_data(
-          allocator::malloc(size * out.itemsize()),
-          size,
-          in.strides(),
-          in.flags());
-    }
-  } else {
-    out.set_data(allocator::malloc(out.nbytes()));
-  }
-}
-
 template <typename T, typename U = T, typename Op>
 void unary_op(const T* a, U* out, size_t shape, size_t stride) {
   for (size_t i = 0; i < shape; i += 1) {

mlx/backend/cuda/CMakeLists.txt

@@ -1,32 +1,98 @@
 # Filename rules in cuda backend:
 #
 # * Use .cu/.cuh if code contains device code, and .cpp/.h if not.
-# * Device-only kernel code should be put in kernels/ subdir.
-# * Files in kernels/ subdir should not include files outside.
+# * Device-only code should be put in device/ subdir.
+# * Files in device/ subdir should not include files outside.
 target_sources(
   mlx
   PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary_two.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_contiguous.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_dynamic.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_input.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/jit_module.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/kernel_utils.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/layer_norm.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/logsumexp.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/random.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/all_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/col_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/init_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/sort.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/qmm.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
 target_compile_definitions(mlx PRIVATE MLX_USE_CUDA)
 
+# Embed kernel sources in binary for JIT compilation.
+file(
+  GLOB MLX_JIT_SOURCES
+  RELATIVE ${CMAKE_CURRENT_SOURCE_DIR}
+  "${CMAKE_CURRENT_SOURCE_DIR}/device/*.h"
+  "${CMAKE_CURRENT_SOURCE_DIR}/device/*.cuh")
+string(JOIN ":" MLX_JIT_SOURCES_ARG ${MLX_JIT_SOURCES})
+add_custom_command(
+  OUTPUT gen/cuda_jit_sources.h
+  COMMAND
+    ${CMAKE_COMMAND} -DMLX_SOURCE_ROOT=${CMAKE_CURRENT_SOURCE_DIR}
+    -DMLX_JIT_SOURCES=${MLX_JIT_SOURCES_ARG} -P
+    "${CMAKE_CURRENT_SOURCE_DIR}/bin2h.cmake"
+  DEPENDS bin2h.cmake ${MLX_JIT_SOURCES})
+add_custom_target(cuda_jit_sources DEPENDS gen/cuda_jit_sources.h)
+add_dependencies(mlx cuda_jit_sources)
+target_include_directories(mlx PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/gen")
+
 # Enable defining device lambda functions.
 target_compile_options(mlx
                        PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>")
 
+# Enable calling host constexpr functions from device. This is needed because
+# the constexpr version of isnan is host only.
+target_compile_options(
+  mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>")
+
+# CUDA 12.8 emits warning #20280-D for copy kernels which is a false positive.
+# Explicitly pass this flag to suppress the warning, it is safe to set it to
+# true but the warning wouldn't be suppressed.
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
+  target_compile_options(
+    mlx
+    PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--static-global-template-stub=false>")
+endif()
+
+# Suppress warning when building for compute capability 7 used by V100.
+target_compile_options(
+  mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--Wno-deprecated-gpu-targets>")
+
 # Compute capability 7 is required for synchronization between CPU/GPU with
 # managed memory. TODO: Add more architectures for potential performance gain.
 set(MLX_CUDA_ARCHITECTURES
-    "70;80"
+    "80"
     CACHE STRING "CUDA architectures")
 message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
@@ -53,6 +119,25 @@ target_link_libraries(mlx PUBLIC $<BUILD_INTERFACE:nvtx3-cpp>)
 find_package(CUDAToolkit REQUIRED)
 target_include_directories(mlx PRIVATE ${CUDAToolkit_INCLUDE_DIRS})
 
+# Use cublasLt.
+target_link_libraries(mlx PRIVATE CUDA::cublasLt)
+
+# Use NVRTC and driver APIs.
+target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
+
 # Suppress nvcc warnings on MLX headers.
 target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
                                    --diag_suppress=997>)
+
+# Install CCCL headers for JIT.
+install(DIRECTORY ${cccl_SOURCE_DIR}/include/cuda
+        DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/cccl)
+
+# Make Thunderkittens available
+FetchContent_Declare(
+  kittens
+  GIT_REPOSITORY https://github.com/HazyResearch/ThunderKittens.git
+  GIT_TAG aaab847f430ed313ed466e64b25b9177babd1db8
+  GIT_SHALLOW TRUE)
+FetchContent_MakeAvailable(kittens)
+target_include_directories(mlx BEFORE PRIVATE "${kittens_SOURCE_DIR}/include")

mlx/backend/cuda/allocator.cpp

@@ -3,6 +3,7 @@
 #include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/utils.h"
 #include "mlx/backend/cuda/worker.h"
+#include "mlx/utils.h"
 
 #include <cuda_runtime.h>
 #include <fmt/format.h>
@@ -14,9 +15,57 @@ namespace mlx::core {
 
 namespace cu {
 
+constexpr int page_size = 16384;
+
+// Any allocations smaller than this will try to use the small pool
+constexpr int small_block_size = 8;
+
+// The small pool size in bytes. This should be a multiple of the host page
+// size and small_block_size.
+constexpr int small_pool_size = 4 * page_size;
+
+SmallSizePool::SmallSizePool() {
+  CHECK_CUDA_ERROR(cudaMallocManaged(&buffer_, small_pool_size));
+  end_ = reinterpret_cast<void*>(
+      reinterpret_cast<char*>(buffer_) + small_pool_size);
+  next_free_ = reinterpret_cast<Block*>(buffer_);
+
+  auto num_blocks = small_pool_size / small_block_size;
+  auto curr = next_free_;
+  for (size_t i = 0; i < num_blocks - 1; ++i) {
+    curr->next = reinterpret_cast<Block*>(
+        reinterpret_cast<char*>(buffer_) + (i + 1) * small_block_size);
+    curr = curr->next;
+  }
+  curr->next = nullptr;
+}
+
+SmallSizePool::~SmallSizePool() {
+  CHECK_CUDA_ERROR(cudaFree(buffer_));
+}
+
+void* SmallSizePool::malloc() {
+  if (next_free_ == nullptr) {
+    return nullptr;
+  }
+  Block* b = next_free_;
+  next_free_ = next_free_->next;
+  return static_cast<void*>(b);
+}
+
+void SmallSizePool::free(void* p) {
+  auto b = static_cast<Block*>(p);
+  b->next = next_free_;
+  next_free_ = b;
+}
+
+bool SmallSizePool::in_pool(void* p) {
+  return (p >= buffer_) && (p < end_);
+}
+
 CudaAllocator::CudaAllocator()
     : buffer_cache_(
-          getpagesize(),
+          page_size,
           [](CudaBuffer* buf) { return buf->size; },
           [this](CudaBuffer* buf) {
             cuda_free(buf->data);
@@ -31,7 +80,16 @@ CudaAllocator::CudaAllocator()
 
 Buffer CudaAllocator::malloc(size_t size) {
   // Find available buffer from cache.
+  auto orig_size = size;
   std::unique_lock lock(mutex_);
+  if (size <= small_block_size) {
+    size = 8;
+  } else if (size < page_size) {
+    size = next_power_of_2(size);
+  } else {
+    size = page_size * ((size + page_size - 1) / page_size);
+  }
+
   CudaBuffer* buf = buffer_cache_.reuse_from_cache(size);
   if (!buf) {
     // If we have a lot of memory pressure or are over the maximum cache size,
@@ -43,11 +101,19 @@ Buffer CudaAllocator::malloc(size_t size) {
 
     lock.unlock();
     buf = new CudaBuffer{nullptr, size};
-    cudaError_t err = cudaMallocManaged(&buf->data, size);
-    if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
-      throw std::runtime_error(fmt::format(
-          "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+
+    // Try the scalar pool first
+    if (size <= small_block_size) {
+      buf->data = scalar_pool_.malloc();
     }
+    if (!buf->data) {
+      cudaError_t err = cudaMallocManaged(&buf->data, size);
+      if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
+        throw std::runtime_error(fmt::format(
+            "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+      }
+    }
+
     lock.lock();
   }
   active_memory_ += size;
@@ -106,8 +172,11 @@ void CudaAllocator::cuda_free(void* buf) {
       return;
     }
   }
-
-  cudaFree(buf);
+  if (scalar_pool_.in_pool(buf)) {
+    scalar_pool_.free(buf);
+  } else {
+    cudaFree(buf);
+  }
 }
 
 size_t CudaAllocator::get_active_memory() const {

mlx/backend/cuda/allocator.h

@@ -22,6 +22,28 @@ struct CudaBuffer {
   size_t size;
 };
 
+class SmallSizePool {
+ private:
+  struct Block {
+    Block* next;
+  };
+
+  void* buffer_{nullptr};
+  Block* next_free_{nullptr};
+  void* end_{nullptr};
+
+ public:
+  SmallSizePool();
+  ~SmallSizePool();
+
+  SmallSizePool(const SmallSizePool&) = delete;
+  SmallSizePool& operator=(const SmallSizePool&) = delete;
+
+  void* malloc();
+  void free(void* p);
+  bool in_pool(void* p);
+};
+
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
@@ -60,6 +82,7 @@ class CudaAllocator : public allocator::Allocator {
   BufferCache<CudaBuffer> buffer_cache_;
   size_t active_memory_{0};
   size_t peak_memory_{0};
+  SmallSizePool scalar_pool_;
 };
 
 CudaAllocator& allocator();

mlx/backend/cuda/arg_reduce.cu

@@ -0,0 +1,183 @@
+// Copyright © 2025 Apple Inc.
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/fp16_math.cuh"
+#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <nvtx3/nvtx3.hpp>
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_reduce.cuh>
+
+#include <cassert>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T>
+struct IndexValPair {
+  uint32_t index;
+  T val;
+};
+
+template <typename T>
+struct ArgMin {
+  constexpr __device__ T init() {
+    return Limits<T>::max();
+  }
+
+  __device__ IndexValPair<T> operator()(
+      const IndexValPair<T>& best,
+      const IndexValPair<T>& current) {
+    if (best.val > current.val ||
+        (best.val == current.val && best.index > current.index)) {
+      return current;
+    } else {
+      return best;
+    }
+  }
+
+  template <int N>
+  __device__ IndexValPair<T>
+  reduce_many(IndexValPair<T> best, T (&vals)[N], uint32_t offset) {
+    for (int i = 0; i < N; i++) {
+      if (vals[i] < best.val) {
+        best.val = vals[i];
+        best.index = offset + i;
+      }
+    }
+    return best;
+  }
+};
+
+template <typename T>
+struct ArgMax {
+  constexpr __device__ T init() {
+    return Limits<T>::min();
+  }
+
+  __device__ IndexValPair<T> operator()(
+      const IndexValPair<T>& best,
+      const IndexValPair<T>& current) {
+    if (best.val < current.val ||
+        (best.val == current.val && best.index > current.index)) {
+      return current;
+    } else {
+      return best;
+    }
+  }
+
+  template <int N>
+  __device__ IndexValPair<T>
+  reduce_many(IndexValPair<T> best, T (&vals)[N], uint32_t offset) {
+    for (int i = 0; i < N; i++) {
+      if (vals[i] > best.val) {
+        best.val = vals[i];
+        best.index = offset + i;
+      }
+    }
+    return best;
+  }
+};
+
+template <typename T, typename Op, int BLOCK_DIM, int N_READS = 4>
+__global__ void arg_reduce_general(
+    const T* in,
+    uint32_t* out,
+    size_t size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides in_strides,
+    const __grid_constant__ Strides out_strides,
+    int32_t ndim,
+    int64_t axis_stride,
+    int32_t axis_size) {
+  auto block = cg::this_thread_block();
+
+  int64_t index = cg::this_grid().block_rank();
+  if (index >= size) {
+    return;
+  }
+
+  int64_t in_idx = elem_to_loc(index, shape.data(), in_strides.data(), ndim);
+  int64_t out_idx = elem_to_loc(index, shape.data(), out_strides.data(), ndim);
+
+  Op op;
+  T init = op.init();
+  IndexValPair<T> best{0, init};
+
+  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    T vals[N_READS];
+    auto tid = r * BLOCK_DIM + block.thread_index().x;
+    cub::LoadDirectBlocked(
+        tid, strided_iterator(in + in_idx, axis_stride), vals, axis_size, init);
+    best = op.reduce_many(best, vals, tid * N_READS);
+  }
+
+  typedef cub::BlockReduce<IndexValPair<T>, BLOCK_DIM> BlockReduceT;
+  __shared__ typename BlockReduceT::TempStorage temp;
+
+  best = BlockReduceT(temp).Reduce(best, op);
+
+  if (block.thread_rank() == 0) {
+    out[out_idx] = best.index;
+  }
+}
+
+} // namespace cu
+
+void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("ArgReduce::eval_gpu");
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  out.set_data(allocator::malloc(out.nbytes()));
+  auto& s = stream();
+
+  // Prepare the shapes, strides and axis arguments.
+  Shape shape = remove_index(in.shape(), axis_);
+  Strides in_strides = remove_index(in.strides(), axis_);
+  Strides out_strides = out.ndim() == in.ndim()
+      ? remove_index(out.strides(), axis_)
+      : out.strides();
+  int64_t axis_stride = in.strides()[axis_];
+  int32_t axis_size = in.shape()[axis_];
+  int32_t ndim = shape.size();
+
+  // ArgReduce.
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  dispatch_real_types(in.dtype(), "ArgReduce", [&](auto type_tag) {
+    using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    constexpr uint32_t N_READS = 4;
+    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      dim3 num_blocks = get_2d_grid_dims(out.shape(), out.strides());
+      auto kernel =
+          cu::arg_reduce_general<T, cu::ArgMax<T>, block_dim(), N_READS>;
+      if (reduce_type_ == ArgReduce::ArgMin) {
+        kernel = cu::arg_reduce_general<T, cu::ArgMin<T>, block_dim(), N_READS>;
+      }
+      encoder.add_kernel_node(
+          kernel,
+          num_blocks,
+          block_dim(),
+          0,
+          in.data<T>(),
+          out.data<uint32_t>(),
+          out.size(),
+          const_param(shape),
+          const_param(in_strides),
+          const_param(out_strides),
+          ndim,
+          axis_stride,
+          axis_size);
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/bin2h.cmake

@@ -0,0 +1,150 @@
+# Based on: https://github.com/sivachandran/cmake-bin2h
+#
+# Copyright 2020 Sivachandran Paramasivam
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+include(CMakeParseArguments)
+
+# Function to wrap a given string into multiple lines at the given column
+# position.
+#
+# Parameters:
+#
+# * VARIABLE - The name of the CMake variable holding the string.
+# * AT_COLUMN - The column position at which string will be wrapped.
+function(WRAP_STRING)
+  set(oneValueArgs VARIABLE AT_COLUMN)
+  cmake_parse_arguments(WRAP_STRING "${options}" "${oneValueArgs}" "" ${ARGN})
+
+  string(LENGTH ${${WRAP_STRING_VARIABLE}} stringLength)
+  math(EXPR offset "0")
+
+  while(stringLength GREATER 0)
+    if(stringLength GREATER ${WRAP_STRING_AT_COLUMN})
+      math(EXPR length "${WRAP_STRING_AT_COLUMN}")
+    else()
+      math(EXPR length "${stringLength}")
+    endif()
+
+    string(SUBSTRING ${${WRAP_STRING_VARIABLE}} ${offset} ${length} line)
+    set(lines "${lines}\n ${line}")
+
+    math(EXPR stringLength "${stringLength} - ${length}")
+    math(EXPR offset "${offset} + ${length}")
+  endwhile()
+
+  set(${WRAP_STRING_VARIABLE}
+      "${lines}"
+      PARENT_SCOPE)
+endfunction()
+
+# Function to embed contents of a file as byte array in C/C++ header file(.h).
+# The header file will contain a byte array and integer variable holding the
+# size of the array.
+#
+# Parameters:
+#
+# * SOURCE_FILES - The paths of source files whose contents will be embedded in
+#   the header file.
+# * VARIABLE_NAME - The name of the variable for the byte array. The string
+#   "_SIZE" will be append to this name and will be used a variable name for
+#   size variable.
+# * HEADER_FILE - The path of header file.
+# * APPEND - If specified appends to the header file instead of overwriting it
+# * HEADER_NAMESPACE - The namespace, where the array should be located in.
+# * NULL_TERMINATE - If specified a null byte(zero) will be append to the byte
+#   array.
+#
+# Usage:
+#
+# bin2h(SOURCE_FILE "Logo.png" HEADER_FILE "Logo.h" VARIABLE_NAME "LOGO_PNG")
+function(BIN2H)
+  set(options APPEND NULL_TERMINATE)
+  set(oneValueArgs VARIABLE_NAME HEADER_FILE HEADER_NAMESPACE)
+  set(multiValueArgs SOURCE_FILES)
+  cmake_parse_arguments(BIN2H "${options}" "${oneValueArgs}"
+                        "${multiValueArgs}" ${ARGN})
+
+  set(arrayDefinition "")
+  foreach(SOURCE_FILE IN LISTS BIN2H_SOURCE_FILES)
+    # get filename without extension
+    get_filename_component(FILE_NAME_WE ${SOURCE_FILE} NAME_WE)
+    # convert the filename to a valid C identifier
+    string(MAKE_C_IDENTIFIER "${FILE_NAME_WE}" VALID_FILE_NAME)
+
+    # reads source file contents as hex string
+    file(READ ${SOURCE_FILE} hexString HEX)
+
+    # append null
+    if(BIN2H_NULL_TERMINATE)
+      string(APPEND hexString "00")
+    endif()
+
+    # wraps the hex string into multiple lines
+    wrap_string(VARIABLE hexString AT_COLUMN 24)
+
+    # strip the © in source code
+    string(REGEX REPLACE "c2a9" "2020" arrayValues ${hexString})
+
+    string(REGEX REPLACE "([0-9a-f][0-9a-f])" " 0x\\1," arrayValues
+                         ${arrayValues})
+
+    # make a full variable name for the array
+    set(FULL_VARIABLE_NAME "${BIN2H_VARIABLE_NAME}_${VALID_FILE_NAME}")
+
+    # declares byte array and the length variables
+    string(APPEND arrayDefinition
+           "constexpr char ${FULL_VARIABLE_NAME}[] = {${arrayValues}\n};\n\n")
+  endforeach()
+
+  # add namespace wrapper if defined
+  if(DEFINED BIN2H_HEADER_NAMESPACE)
+    set(namespaceStart "namespace ${BIN2H_HEADER_NAMESPACE} {")
+    set(namespaceEnd "} // namespace ${BIN2H_HEADER_NAMESPACE}")
+    set(declarations "${namespaceStart}\n\n${arrayDefinition}${namespaceEnd}\n")
+  endif()
+
+  set(arrayIncludes "#pragma once")
+  string(PREPEND declarations "${arrayIncludes}\n\n")
+
+  if(BIN2H_APPEND)
+    file(APPEND ${BIN2H_HEADER_FILE} "${declarations}")
+  else()
+    file(WRITE ${BIN2H_HEADER_FILE} "${declarations}")
+  endif()
+endfunction()
+
+# ----------------------------- CLI args -----------------------------
+
+string(REPLACE ":" ";" MLX_JIT_SOURCES_LIST ${MLX_JIT_SOURCES})
+foreach(source ${MLX_JIT_SOURCES_LIST})
+  list(APPEND MLX_JIT_SOURCES_ABS "${MLX_SOURCE_ROOT}/${source}")
+endforeach()
+
+bin2h(
+  SOURCE_FILES
+  ${MLX_JIT_SOURCES_ABS}
+  NULL_TERMINATE
+  VARIABLE_NAME
+  "jit_source"
+  HEADER_NAMESPACE
+  "mlx::core"
+  HEADER_FILE
+  "${CMAKE_CURRENT_BINARY_DIR}/gen/cuda_jit_sources.h")

mlx/backend/cuda/binary.cu

@@ -0,0 +1,362 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void binary_ss(const In* a, const In* b, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (int i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[0], b[0]);
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a[0], b[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void binary_sv(const In* a, const In* b, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[0], b[i]);
+    }
+  } else {
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a[0], b_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void binary_vs(const In* a, const In* b, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[0]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a_vec.val[i], b[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[i]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a_vec.val[i], b_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
+__global__ void binary_g_nd(
+    const In* a,
+    const In* b,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
+        index, shape.data(), a_strides.data(), b_strides.data());
+    out[index] = Op{}(a[a_idx], b[b_idx]);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void binary_g(
+    const In* a,
+    const In* b,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides a_strides,
+    const __grid_constant__ Strides b_strides,
+    int ndim) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [a_idx, b_idx] = elem_to_loc_4d(
+        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
+    out[index] = Op{}(a[a_idx], b[b_idx]);
+  }
+}
+
+template <typename Op, typename In, typename Out>
+constexpr bool supports_binary_op() {
+  if (std::is_same_v<Op, Add> || std::is_same_v<Op, Divide> ||
+      std::is_same_v<Op, Maximum> || std::is_same_v<Op, Minimum> ||
+      std::is_same_v<Op, Multiply> || std::is_same_v<Op, Subtract> ||
+      std::is_same_v<Op, Power> || std::is_same_v<Op, Remainder>) {
+    return std::is_same_v<In, Out>;
+  }
+  if (std::is_same_v<Op, Equal> || std::is_same_v<Op, Greater> ||
+      std::is_same_v<Op, GreaterEqual> || std::is_same_v<Op, Less> ||
+      std::is_same_v<Op, LessEqual> || std::is_same_v<Op, NotEqual>) {
+    return std::is_same_v<Out, bool>;
+  }
+  if (std::is_same_v<Op, LogicalAnd> || std::is_same_v<Op, LogicalOr>) {
+    return std::is_same_v<Out, bool> && std::is_same_v<In, bool>;
+  }
+  if (std::is_same_v<Op, NaNEqual>) {
+    return std::is_same_v<Out, bool> && is_inexact_v<In>;
+  }
+  if (std::is_same_v<Op, LogAddExp>) {
+    return std::is_same_v<In, Out> && is_inexact_v<In>;
+  }
+  if (std::is_same_v<Op, ArcTan2>) {
+    return std::is_same_v<In, Out> && is_floating_v<In>;
+  }
+  if (std::is_same_v<Op, BitwiseAnd> || std::is_same_v<Op, BitwiseOr> ||
+      std::is_same_v<Op, BitwiseXor>) {
+    return std::is_same_v<In, Out> && std::is_integral_v<In>;
+  }
+  if (std::is_same_v<Op, LeftShift> || std::is_same_v<Op, RightShift>) {
+    return std::is_same_v<In, Out> && std::is_integral_v<In> &&
+        !std::is_same_v<In, bool>;
+  }
+  return false;
+}
+
+} // namespace cu
+
+template <typename Op>
+void binary_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    array& out,
+    const char* op,
+    const Stream& s) {
+  assert(inputs.size() > 1);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  if (out.size() == 0) {
+    return;
+  }
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  dispatch_all_types(a.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
+      using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
+      if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+        using InType = cuda_type_t<CTYPE_IN>;
+        using OutType = cuda_type_t<CTYPE_OUT>;
+        auto bopt = get_binary_op_type(a, b);
+        if (bopt == BinaryOpType::General) {
+          dispatch_bool(
+              a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
+                  out.data_size() > INT32_MAX,
+              [&](auto large) {
+                using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+                Shape shape;
+                std::vector<Strides> strides;
+                std::tie(shape, strides) = collapse_contiguous_dims(a, b, out);
+                auto& a_strides = strides[0];
+                auto& b_strides = strides[1];
+                int ndim = shape.size();
+                if (ndim <= 3) {
+                  dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                    auto kernel = cu::
+                        binary_g_nd<Op, InType, OutType, IdxT, dims_constant()>;
+                    auto [num_blocks, block_dims] =
+                        get_launch_args(kernel, out, large());
+                    encoder.add_kernel_node(
+                        kernel,
+                        num_blocks,
+                        block_dims,
+                        0,
+                        a.data<InType>(),
+                        b.data<InType>(),
+                        out.data<OutType>(),
+                        out.size(),
+                        const_param<dims_constant()>(shape),
+                        const_param<dims_constant()>(a_strides),
+                        const_param<dims_constant()>(b_strides));
+                  });
+                } else {
+                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
+                  auto [num_blocks, block_dims] =
+                      get_launch_args(kernel, out, large());
+                  encoder.add_kernel_node(
+                      kernel,
+                      num_blocks,
+                      block_dims,
+                      0,
+                      a.data<InType>(),
+                      b.data<InType>(),
+                      out.data<OutType>(),
+                      out.size(),
+                      const_param(shape),
+                      const_param(a_strides),
+                      const_param(b_strides),
+                      ndim);
+                }
+              });
+        } else {
+          dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
+            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
+            auto kernel = cu::binary_ss<Op, InType, OutType, IdxT, N_READS>;
+            if (bopt == BinaryOpType::ScalarVector) {
+              kernel = cu::binary_sv<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorScalar) {
+              kernel = cu::binary_vs<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorVector) {
+              kernel = cu::binary_vv<Op, InType, OutType, IdxT, N_READS>;
+            }
+            auto [num_blocks, block_dims] = get_launch_args(
+                kernel,
+                out.data_size(),
+                out.shape(),
+                out.strides(),
+                large(),
+                N_READS);
+            encoder.add_kernel_node(
+                kernel,
+                num_blocks,
+                block_dims,
+                0,
+                a.data<InType>(),
+                b.data<InType>(),
+                out.data<OutType>(),
+                out.data_size());
+          });
+        }
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do binary op {} on inputs of {} with result of {}.",
+            op,
+            dtype_to_string(a.dtype()),
+            dtype_to_string(out.dtype())));
+      }
+    });
+  });
+}
+
+template <typename Op>
+void binary_op_gpu(
+    const std::vector<array>& inputs,
+    array& out,
+    const char* op,
+    const Stream& s) {
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+  binary_op_gpu_inplace<Op>(inputs, out, op, s);
+}
+
+#define BINARY_GPU(func)                                              \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    nvtx3::scoped_range r(#func "::eval_gpu");                        \
+    auto& s = out.primitive().stream();                               \
+    binary_op_gpu<cu::func>(inputs, out, name(), s);                  \
+  }
+
+BINARY_GPU(Add)
+BINARY_GPU(ArcTan2)
+BINARY_GPU(Divide)
+BINARY_GPU(Remainder)
+BINARY_GPU(Greater)
+BINARY_GPU(GreaterEqual)
+BINARY_GPU(Less)
+BINARY_GPU(LessEqual)
+BINARY_GPU(LogicalAnd)
+BINARY_GPU(LogicalOr)
+BINARY_GPU(LogAddExp)
+BINARY_GPU(Maximum)
+BINARY_GPU(Minimum)
+BINARY_GPU(Multiply)
+BINARY_GPU(NotEqual)
+BINARY_GPU(Power)
+BINARY_GPU(Subtract)
+
+void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Equal::eval_gpu");
+  auto& s = out.primitive().stream();
+  if (equal_nan_) {
+    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
+  } else {
+    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
+  }
+}
+
+void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
+  auto& s = out.primitive().stream();
+  switch (op_) {
+    case BitwiseBinary::And:
+      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Or:
+      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Xor:
+      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::LeftShift:
+      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::RightShift:
+      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
+      break;
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cuda/binary_two.cu

@@ -0,0 +1,337 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_ss(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[0], b[0]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a[0], b[0]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_sv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[0], b[i]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a[0], b_vec.val[i]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_vs(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[i], b[0]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a_vec.val[i], b[0]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[i], b[i]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a_vec.val[i], b_vec.val[i]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
+__global__ void binary_two_g_nd(
+    const In* a,
+    const In* b,
+    Out* out_a,
+    Out* out_b,
+    IdxT size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
+        index, shape.data(), a_strides.data(), b_strides.data());
+    auto out = Op{}(a[a_idx], b[b_idx]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void binary_two_g(
+    const In* a,
+    const In* b,
+    Out* out_a,
+    Out* out_b,
+    IdxT size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides a_strides,
+    const __grid_constant__ Strides b_strides,
+    int ndim) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [a_idx, b_idx] = elem_to_loc_4d(
+        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
+    auto out = Op{}(a[a_idx], b[b_idx]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out>
+constexpr bool supports_binary_two_op() {
+  if (std::is_same_v<Op, DivMod>) {
+    return std::is_same_v<In, Out> &&
+        (std::is_integral_v<Out> || is_floating_v<Out>);
+  }
+  return false;
+}
+
+} // namespace cu
+
+template <typename Op>
+void binary_two_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const char* op,
+    const Stream& s) {
+  assert(inputs.size() > 1);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  auto& out_a = outputs[0];
+  auto& out_b = outputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out_a, bopt);
+  set_binary_op_output_data(a, b, out_b, bopt);
+
+  if (out_a.size() == 0) {
+    return;
+  }
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out_a);
+  encoder.set_output_array(out_b);
+  dispatch_all_types(a.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out_a.dtype(), [&](auto out_type_tag) {
+      using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
+      using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
+      if constexpr (cu::supports_binary_two_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+        using InType = cuda_type_t<CTYPE_IN>;
+        using OutType = cuda_type_t<CTYPE_OUT>;
+
+        auto bopt = get_binary_op_type(a, b);
+        if (bopt == BinaryOpType::General) {
+          dispatch_bool(
+              a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
+                  out_a.data_size() > INT32_MAX,
+              [&](auto large) {
+                using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+                Shape shape;
+                std::vector<Strides> strides;
+                std::tie(shape, strides) =
+                    collapse_contiguous_dims(a, b, out_a);
+                auto& a_strides = strides[0];
+                auto& b_strides = strides[1];
+                int ndim = shape.size();
+                if (ndim <= 3) {
+                  dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                    auto kernel = cu::binary_two_g_nd<
+                        Op,
+                        InType,
+                        OutType,
+                        IdxT,
+                        dims_constant()>;
+                    auto [num_blocks, block_dims] =
+                        get_launch_args(kernel, out_a, large());
+                    encoder.add_kernel_node(
+                        kernel,
+                        num_blocks,
+                        block_dims,
+                        0,
+                        a.data<InType>(),
+                        b.data<InType>(),
+                        out_a.data<OutType>(),
+                        out_b.data<OutType>(),
+                        out_a.size(),
+                        const_param<dims_constant()>(shape),
+                        const_param<dims_constant()>(a_strides),
+                        const_param<dims_constant()>(b_strides));
+                  });
+                } else {
+                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT>;
+                  auto [num_blocks, block_dims] =
+                      get_launch_args(kernel, out_a, large());
+                  encoder.add_kernel_node(
+                      kernel,
+                      num_blocks,
+                      block_dims,
+                      0,
+                      a.data<InType>(),
+                      b.data<InType>(),
+                      out_a.data<OutType>(),
+                      out_b.data<OutType>(),
+                      out_a.size(),
+                      const_param(shape),
+                      const_param(a_strides),
+                      const_param(b_strides),
+                      ndim);
+                }
+              });
+        } else {
+          dispatch_bool(out_a.data_size() > UINT32_MAX, [&](auto large) {
+            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
+            auto kernel = cu::binary_two_ss<Op, InType, OutType, IdxT, N_READS>;
+            if (bopt == BinaryOpType::ScalarVector) {
+              kernel = cu::binary_two_sv<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorScalar) {
+              kernel = cu::binary_two_vs<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorVector) {
+              kernel = cu::binary_two_vv<Op, InType, OutType, IdxT, N_READS>;
+            }
+            auto [num_blocks, block_dims] = get_launch_args(
+                kernel,
+                out_a.data_size(),
+                out_a.shape(),
+                out_a.strides(),
+                large(),
+                N_READS);
+            encoder.add_kernel_node(
+                kernel,
+                num_blocks,
+                block_dims,
+                0,
+                a.data<InType>(),
+                b.data<InType>(),
+                out_a.data<OutType>(),
+                out_b.data<OutType>(),
+                out_a.data_size());
+          });
+        }
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do binary op {} on inputs of {} with result of {}.",
+            op,
+            dtype_to_string(a.dtype()),
+            dtype_to_string(out_a.dtype())));
+      }
+    });
+  });
+}
+
+template <typename Op>
+void binary_two_op_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const char* op,
+    const Stream& s) {
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, outputs[0], bopt);
+  set_binary_op_output_data(a, b, outputs[1], bopt);
+  binary_two_op_gpu_inplace<Op>(inputs, outputs, op, s);
+}
+
+void DivMod::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("DivMod::eval_gpu");
+  auto& s = outputs[0].primitive().stream();
+  binary_two_op_gpu<cu::DivMod>(inputs, outputs, name(), s);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/compiled.cpp

@@ -0,0 +1,301 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/compiled.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/jit_module.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/graph_utils.h"
+#include "mlx/primitives.h"
+
+#include <fmt/format.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+struct FusedKernelBuilder {
+  std::string os;
+  const std::string& kernel_name;
+  const std::vector<array>& inputs;
+  const std::vector<array>& outputs;
+  const std::vector<array>& tape;
+  const std::function<bool(size_t)>& is_constant;
+
+  void build(const char* name, bool contiguous) {
+    NodeNamer namer;
+
+    // Function parameters.
+    std::vector<std::string> params;
+    for (size_t i = 0; i < inputs.size(); ++i) {
+      if (is_constant(i)) {
+        continue;
+      }
+      const auto& x = inputs[i];
+      const std::string& xname = namer.get_name(x);
+      params.push_back(
+          fmt::format("const {}* {}", dtype_to_cuda_type(x.dtype()), xname));
+      if (!is_scalar(x) && !contiguous) {
+        params.push_back(fmt::format(
+            "const __grid_constant__ cuda::std::array<int64_t, NDIM> {}_strides",
+            xname));
+      }
+    }
+    for (const auto& x : outputs) {
+      params.push_back(fmt::format(
+          "{}* {}", dtype_to_cuda_type(x.dtype()), namer.get_name(x)));
+    }
+    if (!contiguous) {
+      params.push_back(
+          "const __grid_constant__ cuda::std::array<int32_t, NDIM> shape");
+    }
+    params.push_back("IdxT size");
+
+    // Build function signature.
+    if (contiguous) {
+      os += "template <typename IdxT = uint32_t, int work_per_thread = 1>\n";
+    } else {
+      os +=
+          "template <int NDIM, typename IdxT = uint32_t, int work_per_thread = 1>\n";
+    }
+    os += fmt::format("__global__ void {}(\n", kernel_name + name);
+    for (size_t i = 0; i < params.size(); ++i) {
+      os += "    ";
+      os += params[i];
+      if (i != params.size() - 1) {
+        os += ",\n";
+      }
+    }
+    os += ") {\n";
+
+    // Index. For non contiguous kernels we create a separate index
+    // variable per variable otherwise everyone uses `index`.
+    os +=
+        "  IdxT index = cg::this_grid().thread_rank() * work_per_thread;\n"
+        "  if (index >= size) {\n"
+        "    return;\n"
+        "  }\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "  IdxT " + xname + "_idx = 0;\n";
+      }
+      os += "  {\n";
+      os += "    IdxT loc = index;\n";
+      os +=
+          "    #pragma unroll\n"
+          "    for (int i = NDIM - 1; i >= 0; i--) {\n";
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "      " + xname + "_idx += (loc \% shape[i]) * IdxT(" + xname +
+            "_strides[i]);\n";
+      }
+      os +=
+          "      loc /= shape[i];\n"
+          "    }\n"
+          "  }\n";
+    }
+
+    // Work loop
+    os +=
+        "\n"
+        "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
+
+    // Read inputs.
+    for (size_t i = 0; i < inputs.size(); ++i) {
+      const auto& x = inputs[i];
+      const std::string& xname = namer.get_name(x);
+      std::string type = dtype_to_cuda_type(x.dtype());
+      std::string value;
+      if (is_constant(i)) {
+        std::ostringstream ss;
+        print_constant(ss, x);
+        value = fmt::format("static_cast<{}>({})", type, ss.str());
+      } else if (is_scalar(x)) {
+        value = fmt::format("{}[0]", xname);
+      } else if (contiguous) {
+        value = fmt::format("{}[index]", xname);
+      } else {
+        value = fmt::format("{}[{}_idx]", xname, xname);
+      }
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
+    }
+
+    // Write tape.
+    for (const auto& x : tape) {
+      const std::string& xname = namer.get_name(x);
+      std::string type = dtype_to_cuda_type(x.dtype());
+      std::string value;
+      if (is_static_cast(x.primitive())) {
+        value = fmt::format(
+            "static_cast<{}>(tmp_{})", type, namer.get_name(x.inputs()[0]));
+      } else {
+        value = x.primitive().name();
+        value += "{}(";
+        for (size_t i = 0; i < x.inputs().size() - 1; ++i) {
+          value += fmt::format("tmp_{}, ", namer.get_name(x.inputs()[i]));
+        }
+        value += fmt::format("tmp_{})", namer.get_name(x.inputs().back()));
+      }
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
+    }
+
+    // Write output.
+    for (const auto& x : outputs) {
+      os += fmt::format("    {0}[index] = tmp_{0};\n", namer.get_name(x));
+    }
+
+    // End of work loop
+    os +=
+        "\n"
+        "    index++;\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "    " + xname + "_idx += " + xname + "_strides[NDIM - 1];\n";
+      }
+    }
+    os += "  }\n";
+
+    os += "}\n";
+  }
+};
+
+} // namespace cu
+
+constexpr const char* g_jit_includes = R"(
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/device/ternary_ops.cuh"
+#include "mlx/backend/cuda/device/unary_ops.cuh"
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cooperative_groups.h>
+
+#define inf cuda::std::numeric_limits<float>::infinity()
+)";
+
+void Compiled::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("Compiled::eval_gpu");
+  auto& s = stream();
+
+  cu::JitModule& mod = cu::get_jit_module(s.device, lib_name(), [&]() {
+    // Build source code.
+    cu::FusedKernelBuilder builder{
+        g_jit_includes, lib_name(), inputs_, outputs_, tape_, is_constant_};
+    builder.os +=
+        "namespace mlx::core::cu {\n\n"
+        "namespace cg = cooperative_groups;\n\n";
+    builder.build("_contiguous", true);
+    builder.os += "\n";
+    builder.build("_strided", false);
+    builder.os += "\n} // namespace mlx::core::cu\n";
+    // Build kernel names.
+    std::vector<std::string> kernel_names;
+    for (auto work_per_thread : std::array<int, 2>{1, 4}) {
+      kernel_names.push_back(fmt::format(
+          "mlx::core::cu::{}_contiguous<uint32_t, {}>",
+          lib_name(),
+          work_per_thread));
+      kernel_names.push_back(fmt::format(
+          "mlx::core::cu::{}_contiguous<int64_t, {}>",
+          lib_name(),
+          work_per_thread));
+      for (int i = 1; i <= MAX_NDIM; ++i) {
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, uint32_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, int64_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
+      }
+    }
+    return std::make_pair(std::move(builder.os), std::move(kernel_names));
+  });
+
+  // Collapse contiguous dims to route to a faster kernel if possible. Also
+  // handle all broadcasting.
+  auto [contiguous, shape, strides_vec] =
+      compiled_collapse_contiguous_dims(inputs, outputs[0], is_constant_);
+
+  // Whether to use large index.
+  bool large = compiled_use_large_index(inputs, outputs, contiguous);
+
+  cu::KernelArgs args;
+  // Put inputs.
+  int strides_index = 1;
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    if (is_constant_(i)) {
+      continue;
+    }
+    const auto& x = inputs[i];
+    args.append(x);
+    if (!contiguous && !is_scalar(x)) {
+      args.append_ptr(strides_vec[strides_index++].data());
+    }
+  }
+
+  // Put outputs.
+  compiled_allocate_outputs(inputs, outputs, is_constant_, contiguous);
+  for (auto& x : outputs) {
+    args.append(x);
+  }
+
+  // Put shape and size.
+  if (!contiguous) {
+    args.append_ptr(shape.data());
+  }
+  if (large) {
+    args.append<int64_t>(outputs[0].data_size());
+  } else {
+    args.append<uint32_t>(outputs[0].data_size());
+  }
+
+  // Choose work per thread
+  int work_per_thread = 4;
+  if (!contiguous && shape.back() % work_per_thread != 0) {
+    work_per_thread = 1;
+  }
+
+  // Launch kernel.
+  const char* index_type = large ? "int64_t" : "uint32_t";
+  std::string kernel_name = fmt::format("mlx::core::cu::{}", lib_name());
+  if (contiguous) {
+    kernel_name +=
+        fmt::format("_contiguous<{}, {}>", index_type, work_per_thread);
+  } else {
+    kernel_name += fmt::format(
+        "_strided<{}, {}, {}>", shape.size(), index_type, work_per_thread);
+  }
+  auto& encoder = cu::get_command_encoder(s);
+  for (const auto& in : inputs) {
+    encoder.set_input_array(in);
+  }
+  for (const auto& out : outputs) {
+    encoder.set_output_array(out);
+  }
+
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] =
+      get_launch_args(kernel, outputs[0], large, work_per_thread);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+}
+
+} // namespace mlx::core

mlx/backend/cuda/copy.cpp

@@ -1,26 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/gpu/copy.h"
-
-namespace mlx::core {
-
-void copy_gpu_inplace(
-    const array& in,
-    array& out,
-    const Shape& data_shape,
-    const Strides& strides_in_pre,
-    const Strides& strides_out_pre,
-    int64_t inp_offset,
-    int64_t out_offset,
-    CopyType ctype,
-    const Stream& s,
-    const std::optional<array>& dynamic_i_offset /* = std::nullopt */,
-    const std::optional<array>& dynamic_o_offset /* = std::nullopt */) {
-  throw std::runtime_error("copy_gpu_inplace not implemented in CUDA backend.");
-}
-
-void fill_gpu(const array& val, array& out, const Stream& s) {
-  throw std::runtime_error("fill_gpu not implemented in CUDA backend.");
-}
-
-} // namespace mlx::core

mlx/backend/cuda/copy.cu

@@ -0,0 +1,87 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cuda/copy/copy.cuh"
+
+namespace mlx::core {
+
+void copy_gpu_inplace(
+    const array& in,
+    array& out,
+    const Shape& shape,
+    const Strides& strides_in,
+    const Strides& strides_out,
+    int64_t offset_in,
+    int64_t offset_out,
+    CopyType ctype,
+    const Stream& s,
+    const std::optional<array>& dynamic_offset_in,
+    const std::optional<array>& dynamic_offset_out) {
+  if (out.size() == 0) {
+    return;
+  }
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  if (ctype == CopyType::Scalar || ctype == CopyType::Vector) {
+    copy_contiguous(encoder, ctype, in, out, offset_in, offset_out);
+    return;
+  }
+
+  if (ctype == CopyType::General || ctype == CopyType::GeneralGeneral) {
+    auto [shape_collapsed, strides_vec] = collapse_contiguous_dims(
+        shape, std::vector{strides_in, strides_out}, INT32_MAX);
+    if (ctype == CopyType::General) {
+      copy_general_input(
+          encoder,
+          ctype,
+          in,
+          out,
+          offset_in,
+          offset_out,
+          shape_collapsed,
+          strides_vec[0]);
+    } else {
+      if (dynamic_offset_in || dynamic_offset_out) {
+        copy_general_dynamic(
+            encoder,
+            ctype,
+            in,
+            out,
+            offset_in,
+            offset_out,
+            shape_collapsed,
+            strides_vec[0],
+            strides_vec[1],
+            dynamic_offset_in ? *dynamic_offset_in : array(0, int64),
+            dynamic_offset_out ? *dynamic_offset_out : array(0, int64));
+      } else {
+        copy_general(
+            encoder,
+            ctype,
+            in,
+            out,
+            offset_in,
+            offset_out,
+            shape_collapsed,
+            strides_vec[0],
+            strides_vec[1]);
+      }
+    }
+    return;
+  }
+}
+
+void fill_gpu(const array& in, array& out, const Stream& s) {
+  if (out.size() == 0) {
+    return;
+  }
+  out.set_data(allocator::malloc(out.nbytes()));
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  copy_contiguous(encoder, CopyType::Scalar, in, out, 0, 0);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/copy/copy.cuh

@@ -0,0 +1,55 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/cast_op.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+
+namespace mlx::core {
+
+void copy_contiguous(
+    cu::CommandEncoder& encoder,
+    CopyType ctype,
+    const array& in,
+    array& out,
+    int64_t offset_in,
+    int64_t offset_out);
+
+void copy_general(
+    cu::CommandEncoder& encoder,
+    CopyType ctype,
+    const array& in,
+    array& out,
+    int64_t offset_in,
+    int64_t offset_out,
+    const Shape& shape,
+    const Strides& strides_in,
+    const Strides& strides_out);
+
+void copy_general_dynamic(
+    cu::CommandEncoder& encoder,
+    CopyType ctype,
+    const array& in,
+    array& out,
+    int64_t offset_in,
+    int64_t offset_out,
+    const Shape& shape,
+    const Strides& strides_in,
+    const Strides& strides_out,
+    const array& dynamic_offset_in,
+    const array& dynamic_offset_out);
+
+void copy_general_input(
+    cu::CommandEncoder& encoder,
+    CopyType ctype,
+    const array& in,
+    array& out,
+    int64_t offset_in,
+    int64_t offset_out,
+    const Shape& shape,
+    const Strides& strides_in);
+
+} // namespace mlx::core

mlx/backend/cuda/copy/copy_contiguous.cu

@@ -0,0 +1,94 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/copy/copy.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename In, typename Out, typename IdxT, int N_READS>
+__global__ void copy_s(const In* in, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = cast_to<Out>(in[0]);
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = cast_to<Out>(in[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename In, typename Out, typename IdxT, int N_READS>
+__global__ void copy_v(const In* in, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = cast_to<Out>(in[i]);
+    }
+  } else {
+    auto in_vec = load_vector<N_READS>(in, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = cast_to<Out>(in_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+} // namespace cu
+
+void copy_contiguous(
+    cu::CommandEncoder& encoder,
+    CopyType ctype,
+    const array& in,
+    array& out,
+    int64_t in_offset,
+    int64_t out_offset) {
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
+        using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
+        using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
+        using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+        // TODO: Choose optimized value based on type size.
+        constexpr int N_READS = 4;
+        auto kernel = cu::copy_s<InType, OutType, IdxT, N_READS>;
+        if (ctype == CopyType::Vector) {
+          kernel = cu::copy_v<InType, OutType, IdxT, N_READS>;
+        }
+        auto [num_blocks, block_dims] = get_launch_args(
+            kernel,
+            out.data_size(),
+            out.shape(),
+            out.strides(),
+            large(),
+            N_READS);
+        encoder.add_kernel_node(
+            kernel,
+            num_blocks,
+            block_dims,
+            0,
+            in.data<InType>() + in_offset,
+            out.data<OutType>() + out_offset,
+            out.data_size());
+      });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/copy/copy_general.cu

@@ -0,0 +1,112 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/copy/copy.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename In, typename Out, typename IdxT, int NDIM>
+__global__ void copy_gg_nd(
+    const In* in,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_out) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [idx_in, idx_out] = elem_to_loc_nd<NDIM>(
+        index, shape.data(), strides_in.data(), strides_out.data());
+    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
+  }
+}
+
+template <typename In, typename Out, typename IdxT>
+__global__ void copy_gg(
+    const In* in,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides strides_in,
+    const __grid_constant__ Strides strides_out,
+    int ndim) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [idx_in, idx_out] = elem_to_loc_4d(
+        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
+    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
+  }
+}
+
+} // namespace cu
+
+void copy_general(
+    cu::CommandEncoder& encoder,
+    CopyType ctype,
+    const array& in,
+    array& out,
+    int64_t offset_in,
+    int64_t offset_out,
+    const Shape& shape,
+    const Strides& strides_in,
+    const Strides& strides_out) {
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
+            using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            int ndim = shape.size();
+            size_t data_size = 1;
+            for (auto& s : shape)
+              data_size *= s;
+            if (ndim <= 3) {
+              dispatch_1_2_3(ndim, [&](auto ndim_constant) {
+                auto kernel =
+                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>;
+                auto [num_blocks, block_dims] = get_launch_args(
+                    kernel, data_size, shape, out.strides(), large());
+                encoder.add_kernel_node(
+                    kernel,
+                    num_blocks,
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    data_size,
+                    const_param<ndim_constant()>(shape),
+                    const_param<ndim_constant()>(strides_in),
+                    const_param<ndim_constant()>(strides_out));
+              });
+            } else { // ndim >= 4
+              auto kernel = cu::copy_gg<InType, OutType, IdxT>;
+              auto [num_blocks, block_dims] = get_launch_args(
+                  kernel, data_size, shape, out.strides(), large());
+              encoder.add_kernel_node(
+                  kernel,
+                  num_blocks,
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  data_size,
+                  const_param(shape),
+                  const_param(strides_in),
+                  const_param(strides_out),
+                  ndim);
+            }
+          });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/copy/copy_general_dynamic.cu

@@ -0,0 +1,119 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/copy/copy.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename In, typename Out, typename IdxT, int NDIM>
+__global__ void copy_gg_dynamic_nd(
+    const In* in,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_out,
+    const int64_t* offset_in,
+    const int64_t* offset_out) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [idx_in, idx_out] = elem_to_loc_nd<NDIM>(
+        index, shape.data(), strides_in.data(), strides_out.data());
+    out[idx_out + *offset_out] = CastOp<In, Out>{}(in[idx_in + *offset_in]);
+  }
+}
+
+template <typename In, typename Out, typename IdxT>
+__global__ void copy_gg_dynamic(
+    const In* in,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides strides_in,
+    const __grid_constant__ Strides strides_out,
+    int ndim,
+    const int64_t* offset_in,
+    const int64_t* offset_out) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [idx_in, idx_out] = elem_to_loc_4d(
+        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
+    out[idx_out + *offset_out] = CastOp<In, Out>{}(in[idx_in + *offset_in]);
+  }
+}
+
+} // namespace cu
+
+void copy_general_dynamic(
+    cu::CommandEncoder& encoder,
+    CopyType ctype,
+    const array& in,
+    array& out,
+    int64_t offset_in,
+    int64_t offset_out,
+    const Shape& shape,
+    const Strides& strides_in,
+    const Strides& strides_out,
+    const array& dynamic_offset_in,
+    const array& dynamic_offset_out) {
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
+            using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            int ndim = shape.size();
+            if (ndim <= 3) {
+              dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                auto kernel = cu::
+                    copy_gg_dynamic_nd<InType, OutType, IdxT, dims_constant()>;
+                auto [num_blocks, block_dims] =
+                    get_launch_args(kernel, out, large());
+                encoder.add_kernel_node(
+                    kernel,
+                    num_blocks,
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    out.size(),
+                    const_param<dims_constant()>(shape),
+                    const_param<dims_constant()>(strides_in),
+                    const_param<dims_constant()>(strides_out),
+                    dynamic_offset_in.data<int64_t>(),
+                    dynamic_offset_out.data<int64_t>());
+              });
+            } else { // ndim >= 4
+              auto kernel = cu::copy_gg_dynamic<InType, OutType, IdxT>;
+              auto [num_blocks, block_dims] =
+                  get_launch_args(kernel, out, large());
+              encoder.add_kernel_node(
+                  kernel,
+                  num_blocks,
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  out.size(),
+                  const_param(shape),
+                  const_param(strides_in),
+                  const_param(strides_out),
+                  ndim,
+                  dynamic_offset_in.data<int64_t>(),
+                  dynamic_offset_out.data<int64_t>());
+            }
+          });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/copy/copy_general_input.cu

@@ -0,0 +1,102 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/copy/copy.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename In, typename Out, typename IdxT, int NDIM>
+__global__ void copy_g_nd(
+    const In* in,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    IdxT idx_in = elem_to_loc_nd<NDIM>(index, shape.data(), strides_in.data());
+    out[index] = CastOp<In, Out>{}(in[idx_in]);
+  }
+}
+
+template <typename In, typename Out, typename IdxT>
+__global__ void copy_g(
+    const In* in,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides strides_in,
+    int ndim) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    IdxT idx_in = elem_to_loc_4d(index, shape.data(), strides_in.data(), ndim);
+    out[index] = CastOp<In, Out>{}(in[idx_in]);
+  }
+}
+
+} // namespace cu
+
+void copy_general_input(
+    cu::CommandEncoder& encoder,
+    CopyType ctype,
+    const array& in,
+    array& out,
+    int64_t offset_in,
+    int64_t offset_out,
+    const Shape& shape,
+    const Strides& strides_in) {
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
+            using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            int ndim = shape.size();
+            if (ndim <= 3) {
+              dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                auto kernel =
+                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>;
+                auto [num_blocks, block_dims] =
+                    get_launch_args(kernel, out, large());
+                encoder.add_kernel_node(
+                    kernel,
+                    num_blocks,
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    out.size(),
+                    const_param<dims_constant()>(shape),
+                    const_param<dims_constant()>(strides_in));
+              });
+            } else { // ndim >= 4
+              auto kernel = cu::copy_g<InType, OutType, IdxT>;
+              auto [num_blocks, block_dims] =
+                  get_launch_args(kernel, out, large());
+              encoder.add_kernel_node(
+                  kernel,
+                  num_blocks,
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  out.size(),
+                  const_param(shape),
+                  const_param(strides_in),
+                  ndim);
+            }
+          });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/cuda.cpp

@@ -0,0 +1,11 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/cuda.h"
+
+namespace mlx::core::cu {
+
+bool is_available() {
+  return true;
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/cuda.h

@@ -0,0 +1,10 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+namespace mlx::core::cu {
+
+/* Check if the CUDA backend is available. */
+bool is_available();
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device.cpp

@@ -2,46 +2,49 @@
 
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/worker.h"
-#include "mlx/backend/metal/metal.h"
+#include "mlx/utils.h"
 
 #include <fmt/format.h>
 #include <nvtx3/nvtx3.hpp>
+#include <future>
+#include <unordered_set>
 
 namespace mlx::core {
 
+// Can be tuned with MLX_MAX_OPS_PER_BUFFER
+// This should be less than 255
+constexpr int default_max_nodes_per_graph = 20;
+
+int cuda_graph_cache_size() {
+  static int cache_size = []() {
+    return env::get_var("MLX_CUDA_GRAPH_CACHE_SIZE", 100);
+  }();
+  return cache_size;
+}
+
 namespace cu {
 
-DeviceStream::DeviceStream(Device& device) : device_(device), stream_(device) {}
-
-void DeviceStream::synchronize() {
-  cudaStreamSynchronize(stream_);
-}
-
-cudaStream_t DeviceStream::schedule_cuda_stream() {
-  // TODO: Return a stream that maximizes parallelism.
-  return stream_;
-}
-
-cudaStream_t DeviceStream::last_cuda_stream() {
-  return stream_;
-}
-
-CommandEncoder& DeviceStream::get_encoder() {
-  if (!encoder_) {
-    encoder_ = std::make_unique<CommandEncoder>(*this);
-  }
-  return *encoder_;
-}
-
 Device::Device(int device) : device_(device) {
+  CHECK_CUDA_ERROR(cudaDeviceGetAttribute(
+      &compute_capability_major_, cudaDevAttrComputeCapabilityMajor, device_));
+  CHECK_CUDA_ERROR(cudaDeviceGetAttribute(
+      &compute_capability_minor_, cudaDevAttrComputeCapabilityMinor, device_));
   // Validate the requirements of device.
   int attr = 0;
-  cudaDeviceGetAttribute(&attr, cudaDevAttrConcurrentManagedAccess, device_);
+  CHECK_CUDA_ERROR(cudaDeviceGetAttribute(
+      &attr, cudaDevAttrConcurrentManagedAccess, device_));
   if (attr != 1) {
     throw std::runtime_error(fmt::format(
         "Device {} does not support synchronization in managed memory.",
         device_));
   }
+  // The cublasLt handle is used by matmul.
+  make_current();
+  cublasLtCreate(&lt_);
+}
+
+Device::~Device() {
+  cublasLtDestroy(lt_);
 }
 
 void Device::make_current() {
@@ -54,45 +57,271 @@ void Device::make_current() {
   }
 }
 
-DeviceStream& Device::get_stream(Stream s) {
-  auto it = streams_.find(s.index);
-  if (it == streams_.end()) {
-    it = streams_.try_emplace(s.index, *this).first;
+CommandEncoder& Device::get_command_encoder(Stream s) {
+  auto it = encoders_.find(s.index);
+  if (it == encoders_.end()) {
+    it = encoders_.try_emplace(s.index, *this).first;
   }
   return it->second;
 }
 
-CommandEncoder::CommandEncoder(DeviceStream& s)
-    : device_(s.device()), stream_(s) {}
+CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
+  CHECK_CUDA_ERROR(
+      cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
+}
+
+CommandEncoder::CaptureContext::~CaptureContext() {
+  CHECK_CUDA_ERROR(cudaStreamEndCapture(enc.stream(), &graph));
+  size_t num_nodes;
+  CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, NULL, &num_nodes));
+  if (num_nodes == 1) {
+    cudaGraphNode_t captured_node;
+    CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, &captured_node, &num_nodes));
+    CUDA_KERNEL_NODE_PARAMS params;
+    CHECK_CUDA_ERROR(cuGraphKernelNodeGetParams(captured_node, &params));
+    cudaGraphNode_t node;
+    CHECK_CUDA_ERROR(cuGraphAddKernelNode(&node, enc.graph_, NULL, 0, &params));
+    enc.insert_graph_dependencies(GraphNode{node, 'K'});
+  } else {
+    cudaGraphNode_t node;
+    CHECK_CUDA_ERROR(
+        cudaGraphAddChildGraphNode(&node, enc.graph_, NULL, 0, graph));
+    enc.insert_graph_dependencies(GraphNode{node, 'G'});
+  }
+  CHECK_CUDA_ERROR(cudaGraphDestroy(graph));
+}
+
+CommandEncoder::ConcurrentContext::ConcurrentContext(CommandEncoder& enc)
+    : enc(enc) {
+  enc.in_concurrent_ = true;
+}
+
+CommandEncoder::ConcurrentContext::~ConcurrentContext() {
+  enc.in_concurrent_ = false;
+
+  // Use an empty graph node for synchronization
+  CommandEncoder::GraphNode empty{NULL, 'E', std::to_string(enc.node_count_++)};
+  enc.empty_node_count_++;
+  CHECK_CUDA_ERROR(cudaGraphAddEmptyNode(&empty.node, enc.graph_, NULL, 0));
+
+  // Insert the concurrent -> empty node dependencies
+  for (auto& from : enc.concurrent_nodes_) {
+    enc.from_nodes_.push_back(from.node);
+    enc.to_nodes_.push_back(empty.node);
+    enc.graph_key_ += from.id;
+    enc.graph_key_ += from.node_type;
+    enc.graph_key_ += empty.id;
+    enc.graph_key_ += empty.node_type;
+  }
+
+  // Insert the input -> concurrent node dependencies without updating output
+  // nodes
+  auto outputs = std::move(enc.active_outputs_);
+  enc.insert_graph_dependencies(std::move(enc.concurrent_nodes_));
+
+  // Update output node to be the empty node
+  for (auto o : outputs) {
+    enc.node_map_.emplace(o, empty).first->second = empty;
+  }
+}
+
+void CommandEncoder::insert_graph_dependencies(GraphNode node) {
+  if (node.node_type == 'G') {
+    graph_node_count_++;
+  }
+  node.id = std::to_string(node_count_++);
+  if (in_concurrent_) {
+    concurrent_nodes_.push_back(std::move(node));
+  } else {
+    std::vector<GraphNode> nodes;
+    nodes.push_back(std::move(node));
+    insert_graph_dependencies(std::move(nodes));
+  }
+}
+
+void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
+  std::vector<GraphNode> deps;
+  {
+    // Dependencies must be added in the same order to produce a consistent
+    // topology
+    std::unordered_set<cudaGraphNode_t> set_deps;
+    for (auto d : active_deps_) {
+      if (auto it = node_map_.find(d); it != node_map_.end()) {
+        auto [_, inserted] = set_deps.insert(it->second.node);
+        if (inserted) {
+          deps.push_back(it->second);
+        }
+      }
+    }
+  }
+  active_deps_.clear();
+
+  for (auto o : active_outputs_) {
+    for (auto& node : nodes) {
+      node_map_.emplace(o, node).first->second = node;
+    }
+  }
+  active_outputs_.clear();
+
+  for (auto& from : deps) {
+    for (auto& to : nodes) {
+      from_nodes_.push_back(from.node);
+      to_nodes_.push_back(to.node);
+      graph_key_ += from.id;
+      graph_key_ += from.node_type;
+      graph_key_ += to.id;
+      graph_key_ += to.node_type;
+    }
+  }
+}
+
+CommandEncoder::CommandEncoder(Device& d) : device_(d), stream_(d) {
+  CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
+}
+
+void clear_graphs(std::unordered_map<std::string, cudaGraphExec_t>& graphs) {
+  for (auto& [_, graph_exec] : graphs) {
+    CHECK_CUDA_ERROR(cudaGraphExecDestroy(graph_exec));
+  }
+  graphs.clear();
+}
+
+CommandEncoder::~CommandEncoder() {
+  clear_graphs(graph_cache_);
+}
 
 void CommandEncoder::add_completed_handler(std::function<void()> task) {
   worker_.add_task(std::move(task));
 }
 
-void CommandEncoder::end_encoding() {
-  if (!temporaries_.empty()) {
-    add_completed_handler([temporaries = std::move(temporaries_)]() {});
-  }
+void CommandEncoder::set_input_array(const array& arr) {
+  auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
+  active_deps_.push_back(id);
+}
 
-  // There is no kernel running, run completion handlers immediately.
-  if (!has_gpu_work_) {
-    worker_.consume_in_this_thread();
-    return;
-  }
-  has_gpu_work_ = false;
+void CommandEncoder::set_output_array(const array& arr) {
+  auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
+  active_deps_.push_back(id);
+  active_outputs_.push_back(id);
+}
 
-  // Put completion handlers in a batch.
-  worker_.end_batch();
-
-  // Signaling kernel completion is expensive, delay until enough batches.
-  // TODO: This number is arbitrarily picked, profile for a better stragety.
-  if (worker_.uncommited_batches() > 8) {
+void CommandEncoder::maybe_commit() {
+  if (node_count_ >= env::max_ops_per_buffer(default_max_nodes_per_graph)) {
     commit();
   }
 }
 
+void CommandEncoder::add_kernel_node(
+    void* func,
+    dim3 grid_dim,
+    dim3 block_dim,
+    uint32_t smem_bytes,
+    void** params) {
+  cudaKernelNodeParams kernel_params = {0};
+  kernel_params.func = func;
+  kernel_params.gridDim = grid_dim;
+  kernel_params.blockDim = block_dim;
+  kernel_params.kernelParams = params;
+  kernel_params.sharedMemBytes = smem_bytes;
+  cudaGraphNode_t node;
+  CHECK_CUDA_ERROR(
+      cudaGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
+  insert_graph_dependencies(GraphNode{node, 'K'});
+}
+
+void CommandEncoder::add_kernel_node(
+    CUfunction func,
+    dim3 grid_dim,
+    dim3 block_dim,
+    uint32_t smem_bytes,
+    void** params) {
+  CUDA_KERNEL_NODE_PARAMS kernel_params = {0};
+  kernel_params.func = func;
+  kernel_params.gridDimX = grid_dim.x;
+  kernel_params.gridDimY = grid_dim.y;
+  kernel_params.gridDimZ = grid_dim.z;
+  kernel_params.blockDimX = block_dim.x;
+  kernel_params.blockDimY = block_dim.y;
+  kernel_params.blockDimZ = block_dim.z;
+  kernel_params.kernelParams = params;
+  kernel_params.sharedMemBytes = smem_bytes;
+  CUgraphNode node;
+  CHECK_CUDA_ERROR(
+      cuGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
+  insert_graph_dependencies(GraphNode{node, 'K'});
+}
+
 void CommandEncoder::commit() {
-  worker_.commit(stream_.last_cuda_stream());
+  if (!temporaries_.empty()) {
+    add_completed_handler([temporaries = std::move(temporaries_)]() {});
+  }
+  if (node_count_ > 0) {
+    if (!from_nodes_.empty()) {
+      CHECK_CUDA_ERROR(cudaGraphAddDependencies(
+          graph_, from_nodes_.data(), to_nodes_.data(), from_nodes_.size()));
+    }
+
+    graph_key_ += ".";
+    graph_key_ += std::to_string(node_count_);
+    graph_key_ += ".";
+    graph_key_ += std::to_string(graph_node_count_);
+    graph_key_ += ".";
+    graph_key_ += std::to_string(empty_node_count_);
+
+    cudaGraphExec_t& graph_exec = graph_cache_[graph_key_];
+
+    if (graph_exec != nullptr) {
+      cudaGraphExecUpdateResult update_result;
+#if CUDART_VERSION >= 12000
+      cudaGraphExecUpdateResultInfo info;
+      cudaGraphExecUpdate(graph_exec, graph_, &info);
+      update_result = info.result;
+#else
+      cudaGraphNode_t error_node;
+      cudaGraphExecUpdate(graph_exec, graph_, &error_node, &update_result);
+#endif // CUDART_VERSION >= 12000
+      if (update_result != cudaGraphExecUpdateSuccess) {
+        cudaGetLastError(); // reset error
+        CHECK_CUDA_ERROR(cudaGraphExecDestroy(graph_exec));
+        graph_exec = nullptr;
+      }
+    }
+    if (graph_exec == nullptr) {
+      CHECK_CUDA_ERROR(
+          cudaGraphInstantiate(&graph_exec, graph_, NULL, NULL, 0));
+    }
+    device_.make_current();
+    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
+
+    // TODO smarter cache policy
+    if (graph_cache_.size() > cuda_graph_cache_size()) {
+      clear_graphs(graph_cache_);
+    }
+
+    // Reset state
+    node_count_ = 0;
+    graph_node_count_ = 0;
+    from_nodes_.clear();
+    to_nodes_.clear();
+    graph_key_.clear();
+    node_map_.clear();
+    CHECK_CUDA_ERROR(cudaGraphDestroy(graph_));
+    CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
+  }
+
+  // Put completion handlers in a batch.
+  worker_.end_batch();
+  worker_.commit(stream_);
+}
+
+void CommandEncoder::synchronize() {
+  cudaStreamSynchronize(stream_);
+  auto p = std::make_shared<std::promise<void>>();
+  std::future<void> f = p->get_future();
+  add_completed_handler([p = std::move(p)]() { p->set_value(); });
+  worker_.end_batch();
+  commit();
+  f.wait();
 }
 
 Device& device(mlx::core::Device device) {
@@ -104,12 +333,8 @@ Device& device(mlx::core::Device device) {
   return it->second;
 }
 
-DeviceStream& get_stream(Stream s) {
-  return device(s.device).get_stream(s);
-}
-
 CommandEncoder& get_command_encoder(Stream s) {
-  return get_stream(s).get_encoder();
+  return device(s.device).get_command_encoder(s);
 }
 
 } // namespace cu

mlx/backend/cuda/device.h

@@ -6,46 +6,125 @@
 #include "mlx/backend/cuda/worker.h"
 #include "mlx/stream.h"
 
+#include <cublasLt.h>
+#include <cuda.h>
 #include <thrust/execution_policy.h>
 
 #include <unordered_map>
 
 namespace mlx::core::cu {
 
-class Device;
-class CommandEncoder;
-
-class DeviceStream {
+class CommandEncoder {
  public:
-  explicit DeviceStream(Device& device);
+  struct CaptureContext {
+    CaptureContext(CommandEncoder& enc);
+    ~CaptureContext();
+    cudaGraph_t graph;
+    CommandEncoder& enc;
+  };
+  struct ConcurrentContext {
+    ConcurrentContext(CommandEncoder& enc);
+    ~ConcurrentContext();
+    CommandEncoder& enc;
+  };
 
-  DeviceStream(const DeviceStream&) = delete;
-  DeviceStream& operator=(const DeviceStream&) = delete;
+  explicit CommandEncoder(Device& d);
+  ~CommandEncoder();
 
-  // Wait until kernels in the stream complete.
-  void synchronize();
+  CommandEncoder(const CommandEncoder&) = delete;
+  CommandEncoder& operator=(const CommandEncoder&) = delete;
 
-  // Return a cuda stream for launching kernels.
-  cudaStream_t schedule_cuda_stream();
-
-  // Return the last cuda stream used.
-  cudaStream_t last_cuda_stream();
-
-  CommandEncoder& get_encoder();
-
-  Device& device() {
-    return device_;
+  CaptureContext capture_context() {
+    return CaptureContext{*this};
+  }
+  ConcurrentContext concurrent_context() {
+    return ConcurrentContext{*this};
   }
 
+  void set_input_array(const array& arr);
+  void set_output_array(const array& arr);
+
+  template <typename F, typename... Params>
+  void add_kernel_node(
+      F* func,
+      dim3 grid_dim,
+      dim3 block_dim,
+      uint32_t smem_bytes,
+      Params&&... params) {
+    constexpr size_t num = sizeof...(Params);
+    void* ptrs[num];
+    size_t i = 0;
+    ([&](auto&& p) { ptrs[i++] = static_cast<void*>(&p); }(
+         std::forward<Params>(params)),
+     ...);
+    add_kernel_node((void*)func, grid_dim, block_dim, smem_bytes, ptrs);
+  }
+
+  void add_kernel_node(
+      CUfunction func,
+      dim3 grid_dim,
+      dim3 block_dim,
+      uint32_t smem_bytes,
+      void** params);
+
+  void add_kernel_node(
+      void* func,
+      dim3 grid_dim,
+      dim3 block_dim,
+      uint32_t smem_bytes,
+      void** params);
+
+  void add_temporary(const array& arr) {
+    temporaries_.push_back(arr.data_shared_ptr());
+  }
+
+  void add_completed_handler(std::function<void()> task);
+  void maybe_commit();
+  void commit();
+
+  CudaStream& stream() {
+    return stream_;
+  }
+
+  // Wait until kernels and completion handlers are finished
+  void synchronize();
+
  private:
+  struct GraphNode {
+    cudaGraphNode_t node;
+    // K = kernel
+    // E = empty
+    // G = subgraph
+    char node_type;
+    std::string id;
+  };
+
+  void insert_graph_dependencies(GraphNode node);
+  void insert_graph_dependencies(std::vector<GraphNode> nodes);
+
   Device& device_;
   CudaStream stream_;
-  std::unique_ptr<CommandEncoder> encoder_;
+  cudaGraph_t graph_;
+  Worker worker_;
+  char node_count_{0};
+  char graph_node_count_{0};
+  char empty_node_count_{0};
+  bool in_concurrent_{false};
+  std::vector<cudaGraphNode_t> from_nodes_;
+  std::vector<cudaGraphNode_t> to_nodes_;
+  std::string graph_key_;
+  std::vector<GraphNode> concurrent_nodes_;
+  std::vector<std::shared_ptr<array::Data>> temporaries_;
+  std::unordered_map<std::string, cudaGraphExec_t> graph_cache_;
+  std::vector<std::uintptr_t> active_deps_;
+  std::vector<std::uintptr_t> active_outputs_;
+  std::unordered_map<std::uintptr_t, GraphNode> node_map_;
 };
 
 class Device {
  public:
   explicit Device(int device);
+  ~Device();
 
   Device(const Device&) = delete;
   Device& operator=(const Device&) = delete;
@@ -53,72 +132,30 @@ class Device {
   // Make this device the current cuda device, required by some cuda calls.
   void make_current();
 
-  DeviceStream& get_stream(Stream s);
+  CommandEncoder& get_command_encoder(Stream s);
 
   int cuda_device() const {
     return device_;
   }
+  int compute_capability_major() const {
+    return compute_capability_major_;
+  }
+  int compute_capability_minor() const {
+    return compute_capability_minor_;
+  }
+  cublasLtHandle_t lt_handle() const {
+    return lt_;
+  }
 
  private:
   int device_;
-  std::unordered_map<int, DeviceStream> streams_;
-};
-
-class CommandEncoder {
- public:
-  explicit CommandEncoder(DeviceStream& stream);
-
-  CommandEncoder(const CommandEncoder&) = delete;
-  CommandEncoder& operator=(const CommandEncoder&) = delete;
-
-  void set_input_array(const array& arr) {}
-  void set_output_array(const array& arr) {}
-
-  void add_temporary(const array& arr) {
-    temporaries_.push_back(arr.data_shared_ptr());
-  }
-
-  void add_completed_handler(std::function<void()> task);
-  void end_encoding();
-  void commit();
-
-  // Schedule a cuda stream for |fun| to launch kernels, and check error
-  // afterwards.
-  template <typename F>
-  void launch_kernel(F&& fun) {
-    launch_kernel(stream_.schedule_cuda_stream(), std::forward<F>(fun));
-  }
-
-  template <typename F>
-  void launch_kernel(cudaStream_t stream, F&& fun) {
-    device_.make_current();
-    fun(stream);
-    check_cuda_error("kernel launch", cudaGetLastError());
-    has_gpu_work_ = true;
-  }
-
-  Device& device() {
-    return device_;
-  }
-
-  DeviceStream& stream() {
-    return stream_;
-  }
-
-  bool has_gpu_work() const {
-    return has_gpu_work_;
-  }
-
- private:
-  Device& device_;
-  DeviceStream& stream_;
-  Worker worker_;
-  bool has_gpu_work_{false};
-  std::vector<std::shared_ptr<array::Data>> temporaries_;
+  int compute_capability_major_;
+  int compute_capability_minor_;
+  cublasLtHandle_t lt_;
+  std::unordered_map<int, CommandEncoder> encoders_;
 };
 
 Device& device(mlx::core::Device device);
-DeviceStream& get_stream(Stream s);
 CommandEncoder& get_command_encoder(Stream s);
 
 // Return an execution policy that does not sync for result.

mlx/backend/cuda/device/atomic_ops.cuh

@@ -0,0 +1,67 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device/complex.cuh"
+#include "mlx/backend/cuda/device/fp16_math.cuh"
+
+#include <cuda/atomic>
+
+namespace mlx::core::cu {
+
+template <typename T>
+inline __device__ void atomic_add(T* out, T val) {
+  cuda::atomic_ref<T, cuda::thread_scope_device> ref(*out);
+  ref += val;
+}
+
+template <typename T>
+inline __device__ void atomic_prod(T* out, T val) {
+  cuda::atomic_ref<T, cuda::thread_scope_device> ref(*out);
+  T old = ref.load();
+  while (!ref.compare_exchange_strong(old, old * val)) {
+  }
+}
+
+template <typename T>
+inline __device__ void atomic_max(T* out, T val) {
+  cuda::atomic_ref<T, cuda::thread_scope_device> ref(*out);
+  ref.fetch_max(val);
+}
+
+template <typename T>
+inline __device__ void atomic_min(T* out, T val) {
+  cuda::atomic_ref<T, cuda::thread_scope_device> ref(*out);
+  ref.fetch_min(val);
+}
+
+// Somehow cuda::atomic_ref does not provide atomic add for following types.
+template <typename T>
+inline __device__ void atomic_add_general(T* out, T val) {
+  cuda::atomic_ref<T, cuda::thread_scope_device> ref(*out);
+  T old = ref.load();
+  while (!ref.compare_exchange_strong(old, old + val)) {
+  }
+}
+
+inline __device__ void atomic_add(__half* out, __half val) {
+  atomicAdd(out, val);
+}
+
+inline __device__ void atomic_add(complex64_t* out, complex64_t val) {
+#if __CUDA_ARCH__ < 900
+  atomic_add_general(out, val);
+#else
+  atomicAdd(out, val);
+#endif
+}
+
+inline __device__ void atomic_add(__nv_bfloat16* out, __nv_bfloat16 val) {
+#if __CUDA_ARCH__ < 800
+  atomic_add_general(out, val);
+#else
+  atomicAdd(out, val);
+#endif
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/binary_ops.cuh

@@ -0,0 +1,293 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device/unary_ops.cuh"
+
+#include <cuda/std/array>
+
+namespace mlx::core::cu {
+
+struct Add {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x + y;
+  }
+};
+
+struct FloorDivide {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    if constexpr (cuda::std::is_integral_v<T>) {
+      return x / y;
+    } else {
+      return truncf(x / y);
+    }
+  }
+};
+
+struct Divide {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x / y;
+  }
+};
+
+struct Remainder {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    if constexpr (cuda::std::is_integral_v<T>) {
+      if constexpr (cuda::std::is_signed_v<T>) {
+        auto r = x % y;
+        if (r != 0 && (r < 0 != y < 0)) {
+          r += y;
+        }
+        return r;
+      } else {
+        return x % y;
+      }
+    } else if constexpr (is_complex_v<T>) {
+      return x % y;
+    } else {
+      T r = fmod(x, y);
+      if (r != 0 && (r < 0 != y < 0)) {
+        r = r + y;
+      }
+      return r;
+    }
+  }
+};
+
+struct Equal {
+  template <typename T>
+  __device__ bool operator()(T x, T y) {
+    return x == y;
+  }
+};
+
+struct NaNEqual {
+  template <typename T>
+  __device__ bool operator()(T x, T y) {
+    if constexpr (is_complex_v<T>) {
+      return x == y ||
+          (isnan(x.real()) && isnan(y.real()) && isnan(x.imag()) &&
+           isnan(y.imag())) ||
+          (x.real() == y.real() && isnan(x.imag()) && isnan(y.imag())) ||
+          (isnan(x.real()) && isnan(y.real()) && x.imag() == y.imag());
+    } else {
+      return x == y || (isnan(x) && isnan(y));
+    }
+  }
+};
+
+struct Greater {
+  template <typename T>
+  __device__ bool operator()(T x, T y) {
+    return x > y;
+  }
+};
+
+struct GreaterEqual {
+  template <typename T>
+  __device__ bool operator()(T x, T y) {
+    return x >= y;
+  }
+};
+
+struct Less {
+  template <typename T>
+  __device__ bool operator()(T x, T y) {
+    return x < y;
+  }
+};
+
+struct LessEqual {
+  template <typename T>
+  __device__ bool operator()(T x, T y) {
+    return x <= y;
+  }
+};
+
+struct LogAddExp {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag()) || isnan(y.real()) ||
+          isnan(y.imag())) {
+        return {
+            cuda::std::numeric_limits<float>::quiet_NaN(),
+            cuda::std::numeric_limits<float>::quiet_NaN()};
+      }
+      auto max = x.real() > y.real() ? x : y;
+      auto min = x.real() < y.real() ? x : y;
+      auto min_real = min.real();
+      auto max_real = max.real();
+      if (!isfinite(min_real) && (min_real == max_real)) {
+        if (min_real < 0) {
+          return min;
+        } else {
+          return Log{}(Exp{}(min) + Exp{}(max));
+        }
+      } else {
+        return Log1p{}(Exp{}(min - max)) + max;
+      }
+    } else {
+      if (isnan(x) || isnan(y)) {
+        return cuda::std::numeric_limits<T>::quiet_NaN();
+      }
+      T maxval = max(x, y);
+      T minval = min(x, y);
+      return (minval == -cuda::std::numeric_limits<T>::infinity() ||
+              maxval == cuda::std::numeric_limits<T>::infinity())
+          ? maxval
+          : T(float(maxval) + log1p(expf(minval - maxval)));
+    }
+  };
+};
+
+struct Maximum {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    if constexpr (cuda::std::is_integral_v<T>) {
+      return max(x, y);
+    } else if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag())) {
+        return x;
+      }
+      return x > y ? x : y;
+    } else {
+      if (isnan(x)) {
+        return x;
+      }
+      return x > y ? x : y;
+    }
+  }
+};
+
+struct Minimum {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    if constexpr (cuda::std::is_integral_v<T>) {
+      return min(x, y);
+    } else if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag())) {
+        return x;
+      }
+      return x < y ? x : y;
+    } else {
+      if (isnan(x)) {
+        return x;
+      }
+      return x < y ? x : y;
+    }
+  }
+};
+
+struct Multiply {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x * y;
+  }
+};
+
+struct NotEqual {
+  template <typename T>
+  __device__ bool operator()(T x, T y) {
+    if constexpr (is_complex_v<T>) {
+      return x.real() != y.real() || x.imag() != y.imag();
+    } else {
+      return x != y;
+    }
+  }
+};
+
+struct Power {
+  template <typename T>
+  __device__ T operator()(T base, T exp) {
+    if constexpr (cuda::std::is_integral_v<T>) {
+      T res = 1;
+      while (exp) {
+        if (exp & 1) {
+          res *= base;
+        }
+        exp >>= 1;
+        base *= base;
+      }
+      return res;
+    } else if constexpr (is_complex_v<T>) {
+      return pow(base, exp);
+    } else {
+      return powf(base, exp);
+    }
+  }
+};
+
+struct Subtract {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x - y;
+  }
+};
+
+struct LogicalAnd {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x && y;
+  };
+};
+
+struct LogicalOr {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x || y;
+  };
+};
+
+struct BitwiseAnd {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x & y;
+  };
+};
+
+struct BitwiseOr {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x | y;
+  };
+};
+
+struct BitwiseXor {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x ^ y;
+  };
+};
+
+struct LeftShift {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x << y;
+  };
+};
+
+struct RightShift {
+  template <typename T>
+  __device__ T operator()(T x, T y) {
+    return x >> y;
+  };
+};
+
+struct ArcTan2 {
+  template <typename T>
+  __device__ T operator()(T y, T x) {
+    return atan2f(y, x);
+  }
+};
+
+struct DivMod {
+  template <typename T>
+  __device__ cuda::std::array<T, 2> operator()(T x, T y) {
+    return {FloorDivide{}(x, y), Remainder{}(x, y)};
+  };
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/cast_op.cuh

@@ -0,0 +1,130 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device/complex.cuh"
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <thrust/iterator/transform_iterator.h>
+
+namespace mlx::core::cu {
+
+// An op that does static_cast, with custom conversions for some types.
+template <typename SrcT, typename DstT, typename = void>
+struct CastOp {
+  static constexpr bool is_castable = cuda::std::is_convertible_v<SrcT, DstT>;
+
+  __device__ DstT operator()(SrcT x) {
+    return static_cast<DstT>(x);
+  }
+};
+
+// Castings between complex and boolean.
+template <typename T>
+struct CastOp<complex_t<T>, bool> {
+  static constexpr bool is_castable = true;
+
+  __device__ bool operator()(complex_t<T> x) {
+    return x.real() != 0 && x.imag() != 0;
+  }
+};
+
+template <typename T>
+struct CastOp<bool, complex_t<T>> {
+  static constexpr bool is_castable = true;
+
+  __device__ complex_t<T> operator()(bool x) {
+    return x ? complex_t<T>{1, 1} : complex_t<T>{0, 0};
+  }
+};
+
+// Converting a complex number to real number discards the imaginary part.
+template <typename T, typename DstT>
+struct CastOp<complex_t<T>, DstT, cuda::std::enable_if_t<!is_complex_v<DstT>>> {
+  static constexpr bool is_castable = cuda::std::is_convertible_v<T, DstT>;
+
+  __device__ DstT operator()(complex_t<T> x) {
+    static_assert(!is_complex_v<DstT>);
+    return static_cast<DstT>(x.real());
+  }
+};
+
+// Allow converting a real number to complex number.
+template <typename SrcT, typename T>
+struct CastOp<SrcT, complex_t<T>, cuda::std::enable_if_t<!is_complex_v<SrcT>>> {
+  static constexpr bool is_castable = cuda::std::is_convertible_v<SrcT, T>;
+
+  __device__ complex_t<T> operator()(SrcT x) {
+    static_assert(!is_complex_v<SrcT>);
+    return complex_t<T>{static_cast<T>(x), 0};
+  }
+};
+
+// Do nothing when no casting is needed.
+template <typename SrcT, typename DstT>
+struct CastOp<
+    SrcT,
+    DstT,
+    cuda::std::enable_if_t<cuda::std::is_same_v<SrcT, DstT>>> {
+  static constexpr bool is_castable = true;
+
+  __device__ SrcT operator()(SrcT x) {
+    return x;
+  }
+};
+
+// In CUDA 11 the half types do not define conversions between some types,
+// provide fallbacks here.
+#if CUDART_VERSION < 12000
+template <typename SrcT, typename DstT>
+struct CastOp<
+    SrcT,
+    DstT,
+    cuda::std::enable_if_t<
+        !cuda::std::is_convertible_v<SrcT, DstT> && !is_complex_v<SrcT> &&
+        (cuda::std::is_same_v<DstT, __half> ||
+         cuda::std::is_same_v<DstT, __nv_bfloat16>)>> {
+  static constexpr bool is_castable = true;
+
+  __device__ DstT operator()(SrcT x) {
+    return DstT(static_cast<float>(x));
+  }
+};
+
+template <typename SrcT, typename DstT>
+struct CastOp<
+    SrcT,
+    DstT,
+    cuda::std::enable_if_t<
+        !cuda::std::is_convertible_v<SrcT, DstT> && !is_complex_v<SrcT> &&
+        !cuda::std::is_same_v<DstT, __half> &&
+        !cuda::std::is_same_v<DstT, __nv_bfloat16> &&
+        (cuda::std::is_same_v<SrcT, __half> ||
+         cuda::std::is_same_v<SrcT, __nv_bfloat16>)>> {
+  static constexpr bool is_castable = true;
+
+  __device__ DstT operator()(SrcT x) {
+    return DstT(static_cast<float>(x));
+  }
+};
+#endif // CUDART_VERSION < 12000
+
+// Helper to deduce the SrcT.
+template <typename DstT, typename SrcT>
+inline __host__ __device__ auto cast_to(SrcT x) {
+  return CastOp<SrcT, DstT>{}(x);
+}
+
+// Return an iterator that cast the value to DstT using CastOp.
+template <typename DstT, typename Iterator>
+inline __host__ __device__ auto make_cast_iterator(Iterator it) {
+  using SrcT = typename cuda::std::iterator_traits<Iterator>::value_type;
+  if constexpr (std::is_same_v<SrcT, DstT>) {
+    return it;
+  } else {
+    return thrust::make_transform_iterator(it, CastOp<SrcT, DstT>{});
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/complex.cuh

@@ -0,0 +1,60 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+// Make multiplication and division faster.
+#define LIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS
+
+#include <cuda/std/complex>
+#include <cuda/std/type_traits>
+
+namespace mlx::core::cu {
+
+// TODO: Consider using a faster implementation as cuda::std::complex has to
+// conform to C++ standard.
+template <typename T>
+using complex_t = cuda::std::complex<T>;
+
+using complex64_t = complex_t<float>;
+using complex128_t = complex_t<double>;
+
+template <typename T>
+struct is_complex : cuda::std::false_type {};
+
+template <typename T>
+struct is_complex<cuda::std::complex<T>> : cuda::std::true_type {};
+
+template <typename T>
+inline constexpr bool is_complex_v = is_complex<T>::value;
+
+// cuda::std::complex is missing some operators.
+template <typename T>
+inline __host__ __device__ complex_t<T> operator%(
+    complex_t<T> a,
+    complex_t<T> b) {
+  T r = a.real() - floor(a.real() / b.real()) * b.real();
+  T i = a.imag() - floor(a.imag() / b.imag()) * b.imag();
+  return complex_t<T>{r, i};
+}
+
+template <typename T>
+inline __host__ __device__ bool operator>(complex_t<T> a, complex_t<T> b) {
+  return (a.real() > b.real()) || (a.real() == b.real() && a.imag() > b.imag());
+}
+
+template <typename T>
+inline __host__ __device__ bool operator<(complex_t<T> a, complex_t<T> b) {
+  return operator>(b, a);
+}
+
+template <typename T>
+inline __host__ __device__ bool operator<=(complex_t<T> a, complex_t<T> b) {
+  return !(a > b);
+}
+
+template <typename T>
+inline __host__ __device__ bool operator>=(complex_t<T> a, complex_t<T> b) {
+  return !(a < b);
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/config.h

@@ -0,0 +1,12 @@
+// Copyright © 2025 Apple Inc.
+
+// This file is used by both CUDA kernel code and host-only C++ code.
+
+#pragma once
+
+// The maximum dimensions of shape/strides passed as kernel parameters.
+#define MAX_NDIM 10
+
+// All existing NVIDIA hardware has a fixed 32 warp size. Though a built-in
+// warpSize variable exists, using it would prevent compile-time optimizations.
+#define WARP_SIZE 32

mlx/backend/cuda/device/fp16_math.cuh

@@ -0,0 +1,194 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <cuda/std/limits>
+#include <cuda/std/type_traits>
+
+namespace mlx::core::cu {
+
+///////////////////////////////////////////////////////////////////////////////
+// Unary ops for half types.
+///////////////////////////////////////////////////////////////////////////////
+
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
+#define MLX_DEFINE_UNARY_OP(NAME, HALF_OP)           \
+  template <typename T>                              \
+  __forceinline__ __device__ auto NAME(T x) {        \
+    if constexpr (cuda::std::is_same_v<T, __half>) { \
+      return HALF_OP(x);                             \
+    } else {                                         \
+      return ::NAME(x);                              \
+    }                                                \
+  }
+#else
+#define MLX_DEFINE_UNARY_OP(NAME, HALF_OP)                         \
+  template <typename T>                                            \
+  __forceinline__ __device__ auto NAME(T x) {                      \
+    if constexpr (cuda::std::is_same_v<T, __half>) {               \
+      return HALF_OP(x);                                           \
+    } else if constexpr (cuda::std::is_same_v<T, __nv_bfloat16>) { \
+      return HALF_OP(x);                                           \
+    } else {                                                       \
+      return ::NAME(x);                                            \
+    }                                                              \
+  }
+#endif
+
+#define MLX_DEFINE_UNARY_OP_FALLBCK(NAME)                          \
+  template <typename T>                                            \
+  __forceinline__ __device__ auto NAME(T x) {                      \
+    if constexpr (cuda::std::is_same_v<T, __half>) {               \
+      return ::NAME(__half2float(x));                              \
+    } else if constexpr (cuda::std::is_same_v<T, __nv_bfloat16>) { \
+      return ::NAME(__bfloat162float(x));                          \
+    } else {                                                       \
+      return ::NAME(x);                                            \
+    }                                                              \
+  }
+
+MLX_DEFINE_UNARY_OP(abs, __habs)
+MLX_DEFINE_UNARY_OP(ceil, hceil)
+MLX_DEFINE_UNARY_OP(cos, hcos)
+MLX_DEFINE_UNARY_OP(exp, hexp)
+MLX_DEFINE_UNARY_OP(floor, hfloor)
+MLX_DEFINE_UNARY_OP(isnan, __hisnan)
+MLX_DEFINE_UNARY_OP(log, hlog)
+MLX_DEFINE_UNARY_OP(log2, hlog2)
+MLX_DEFINE_UNARY_OP(log10, hlog10)
+MLX_DEFINE_UNARY_OP(rint, hrint)
+MLX_DEFINE_UNARY_OP(rsqrt, hrsqrt)
+MLX_DEFINE_UNARY_OP(sin, hsin)
+MLX_DEFINE_UNARY_OP(sqrt, hsqrt)
+MLX_DEFINE_UNARY_OP_FALLBCK(acos)
+MLX_DEFINE_UNARY_OP_FALLBCK(acosh)
+MLX_DEFINE_UNARY_OP_FALLBCK(asin)
+MLX_DEFINE_UNARY_OP_FALLBCK(asinh)
+MLX_DEFINE_UNARY_OP_FALLBCK(atan)
+MLX_DEFINE_UNARY_OP_FALLBCK(atanh)
+MLX_DEFINE_UNARY_OP_FALLBCK(cosh)
+MLX_DEFINE_UNARY_OP_FALLBCK(log1p)
+MLX_DEFINE_UNARY_OP_FALLBCK(sinh)
+MLX_DEFINE_UNARY_OP_FALLBCK(tan)
+#if __CUDA_ARCH__ >= 1280
+MLX_DEFINE_UNARY_OP(tanh, htanh)
+#else
+MLX_DEFINE_UNARY_OP_FALLBCK(tanh)
+#endif
+
+#undef MLX_DEFINE_UNARY_OP
+#undef MLX_DEFINE_UNARY_OP_FALLBCK
+
+///////////////////////////////////////////////////////////////////////////////
+// Binary ops for half types.
+///////////////////////////////////////////////////////////////////////////////
+
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
+#define MLX_DEFINE_BINARY_OP(NAME, HALF_OP)          \
+  template <typename T>                              \
+  __forceinline__ __device__ auto NAME(T x, T y) {   \
+    if constexpr (cuda::std::is_same_v<T, __half>) { \
+      return HALF_OP(x, y);                          \
+    } else {                                         \
+      return ::NAME(x, y);                           \
+    }                                                \
+  }
+#else
+#define MLX_DEFINE_BINARY_OP(NAME, HALF_OP)                        \
+  template <typename T>                                            \
+  __forceinline__ __device__ auto NAME(T x, T y) {                 \
+    if constexpr (cuda::std::is_same_v<T, __half>) {               \
+      return HALF_OP(x, y);                                        \
+    } else if constexpr (cuda::std::is_same_v<T, __nv_bfloat16>) { \
+      return HALF_OP(x, y);                                        \
+    } else {                                                       \
+      return ::NAME(x, y);                                         \
+    }                                                              \
+  }
+#endif
+
+MLX_DEFINE_BINARY_OP(max, __hmax)
+MLX_DEFINE_BINARY_OP(min, __hmin)
+
+#undef MLX_DEFINE_BINARY_OP
+
+template <typename T>
+__forceinline__ __device__ T fmod(T x, T y) {
+  if constexpr (cuda::std::is_same_v<T, __half>) {
+    return __float2half(::fmod(__half2float(x), __half2float(y)));
+#if CUDART_VERSION >= 12000 || __CUDA_ARCH__ >= 800
+  } else if constexpr (cuda::std::is_same_v<T, __nv_bfloat16>) {
+    return __float2bfloat16(::fmod(__bfloat162float(x), __bfloat162float(y)));
+#endif
+  } else {
+    return ::fmod(x, y);
+  }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// Additional C++ operator overrides between half types and native types.
+///////////////////////////////////////////////////////////////////////////////
+
+template <typename T, typename U>
+constexpr bool is_integral_except =
+    cuda::std::is_integral_v<T> && !cuda::std::is_same_v<T, U>;
+
+template <typename T, typename U>
+constexpr bool is_arithmetic_except =
+    cuda::std::is_arithmetic_v<T> && !cuda::std::is_same_v<T, U>;
+
+#define MLX_DEFINE_HALF_OP(HALF, HALF2FLOAT, FLOAT2HALF, OP)          \
+  template <                                                          \
+      typename T,                                                     \
+      typename = cuda::std::enable_if_t<is_integral_except<T, HALF>>> \
+  __forceinline__ __device__ HALF operator OP(HALF x, T y) {          \
+    return FLOAT2HALF(HALF2FLOAT(x) OP static_cast<float>(y));        \
+  }                                                                   \
+  template <                                                          \
+      typename T,                                                     \
+      typename = cuda::std::enable_if_t<is_integral_except<T, HALF>>> \
+  __forceinline__ __device__ HALF operator OP(T x, HALF y) {          \
+    return FLOAT2HALF(static_cast<float>(x) OP HALF2FLOAT(y));        \
+  }
+
+#define MLX_DEFINE_HALF_CMP(HALF, HALF2FLOAT, OP)                       \
+  template <                                                            \
+      typename T,                                                       \
+      typename = cuda::std::enable_if_t<is_arithmetic_except<T, HALF>>> \
+  __forceinline__ __device__ bool operator OP(HALF x, T y) {            \
+    return HALF2FLOAT(x) OP static_cast<float>(y);                      \
+  }                                                                     \
+  template <                                                            \
+      typename T,                                                       \
+      typename = cuda::std::enable_if_t<is_arithmetic_except<T, HALF>>> \
+  __forceinline__ __device__ bool operator OP(T x, HALF y) {            \
+    return static_cast<float>(y) OP HALF2FLOAT(x);                      \
+  }
+
+MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, +)
+MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, -)
+MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, *)
+MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, /)
+MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, +)
+MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, -)
+MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, *)
+MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, /)
+MLX_DEFINE_HALF_CMP(__half, __half2float, <)
+MLX_DEFINE_HALF_CMP(__half, __half2float, >)
+MLX_DEFINE_HALF_CMP(__half, __half2float, <=)
+MLX_DEFINE_HALF_CMP(__half, __half2float, >=)
+MLX_DEFINE_HALF_CMP(__half, __half2float, ==)
+MLX_DEFINE_HALF_CMP(__half, __half2float, !=)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, <)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, >)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, <=)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, >=)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, ==)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, !=)
+
+#undef MLX_DEFINE_HALF_OP
+#undef MLX_DEFINE_HALF_CMP
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/gather.cuh

@@ -0,0 +1,53 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device/indexing.cuh"
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core::cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T, typename IdxT, int NIDX, int IDX_NDIM, typename LocT>
+__global__ void gather(
+    const T* src,
+    T* out,
+    LocT size,
+    const __grid_constant__ Shape src_shape,
+    const __grid_constant__ Strides src_strides,
+    int32_t src_ndim,
+    const __grid_constant__ Shape slice_sizes,
+    uint32_t slice_size,
+    const __grid_constant__ cuda::std::array<int32_t, NIDX> axes,
+    const __grid_constant__ cuda::std::array<IdxT*, NIDX> indices,
+    const __grid_constant__ cuda::std::array<int32_t, NIDX * IDX_NDIM>
+        indices_shape,
+    const __grid_constant__ cuda::std::array<int64_t, NIDX * IDX_NDIM>
+        indices_strides) {
+  LocT out_idx = cg::this_grid().thread_rank();
+  if (out_idx >= size) {
+    return;
+  }
+
+  LocT src_elem = out_idx % slice_size;
+  LocT idx_elem = out_idx / slice_size;
+
+  LocT src_loc =
+      elem_to_loc(src_elem, slice_sizes.data(), src_strides.data(), src_ndim);
+
+#pragma unroll
+  for (int i = 0; i < NIDX; ++i) {
+    LocT idx_loc = elem_to_loc_nd<IDX_NDIM>(
+        idx_elem,
+        indices_shape.data() + i * IDX_NDIM,
+        indices_strides.data() + i * IDX_NDIM);
+    int32_t axis = axes[i];
+    LocT idx_val = absolute_index(indices[i][idx_loc], src_shape[axis]);
+    src_loc += idx_val * src_strides[axis];
+  }
+
+  out[out_idx] = src[src_loc];
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/gather_axis.cuh

@@ -0,0 +1,65 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device/indexing.cuh"
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core::cu {
+
+namespace cg = cooperative_groups;
+
+template <
+    typename T,
+    typename IdxT,
+    int NDIM,
+    bool SrcC,
+    bool IdxC,
+    typename LocT>
+__global__ void gather_axis(
+    const T* src,
+    const IdxT* indices,
+    T* out,
+    LocT idx_size_pre,
+    LocT idx_size_axis,
+    LocT idx_size_post,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> src_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> idx_strides,
+    int32_t axis,
+    int32_t axis_size,
+    int64_t src_stride_axis,
+    int64_t idx_stride_axis) {
+  LocT index = cg::this_grid().thread_rank();
+  if (index >= idx_size_pre * idx_size_axis * idx_size_post) {
+    return;
+  }
+
+  auto [x, y, z] = index_to_dims(index, idx_size_axis, idx_size_pre);
+
+  LocT elem_idx = z * idx_size_post;
+
+  LocT idx_loc = y * idx_stride_axis;
+  if constexpr (IdxC) {
+    idx_loc += elem_idx * idx_size_axis + x;
+  } else {
+    idx_loc +=
+        elem_to_loc_nd<NDIM>(elem_idx + x, shape.data(), idx_strides.data());
+  }
+
+  auto idx_val = absolute_index(indices[idx_loc], axis_size);
+
+  LocT src_loc = idx_val * src_stride_axis;
+  if constexpr (SrcC) {
+    src_loc += elem_idx * axis_size + x;
+  } else {
+    src_loc +=
+        elem_to_loc_nd<NDIM>(elem_idx + x, shape.data(), src_strides.data());
+  }
+
+  LocT out_idx = y * idx_size_post + elem_idx * idx_size_axis + x;
+
+  out[out_idx] = src[src_loc];
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/indexing.cuh

@@ -0,0 +1,30 @@
+// Copyright © 2025 Apple Inc.
+
+#include <cuda/std/tuple>
+#include <cuda/std/type_traits>
+
+namespace mlx::core::cu {
+
+// Convert an absolute index to positions in a 3d grid, assuming the index is
+// calculated with:
+// index = x * dim1 * dim2 + y * dim2 + z
+template <typename T>
+inline __host__ __device__ cuda::std::tuple<T, T, T>
+index_to_dims(T index, T dim1, T dim2) {
+  T x = index / (dim1 * dim2);
+  T y = (index % (dim1 * dim2)) / dim2;
+  T z = index % dim2;
+  return cuda::std::make_tuple(x, y, z);
+}
+
+// Get absolute index from possible negative index.
+template <typename IdxT>
+inline __host__ __device__ auto absolute_index(IdxT idx, int32_t size) {
+  if constexpr (cuda::std::is_unsigned_v<IdxT>) {
+    return idx;
+  } else {
+    return static_cast<int32_t>(idx < 0 ? idx + size : idx);
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/scatter.cuh

@@ -0,0 +1,68 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device/indexing.cuh"
+#include "mlx/backend/cuda/device/scatter_ops.cuh"
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core::cu {
+
+namespace cg = cooperative_groups;
+
+template <
+    typename T,
+    typename IdxT,
+    typename Op,
+    int NIDX,
+    int IDX_NDIM,
+    typename LocT>
+__global__ void scatter(
+    const T* upd,
+    T* out,
+    LocT size,
+    const __grid_constant__ Shape upd_shape,
+    const __grid_constant__ Strides upd_strides,
+    int32_t upd_ndim,
+    LocT upd_post_idx_size,
+    const __grid_constant__ Shape out_shape,
+    const __grid_constant__ Strides out_strides,
+    int32_t out_ndim,
+    const __grid_constant__ cuda::std::array<int32_t, NIDX> axes,
+    const __grid_constant__ cuda::std::array<IdxT*, NIDX> indices,
+    const __grid_constant__ cuda::std::array<int32_t, NIDX * IDX_NDIM>
+        indices_shape,
+    const __grid_constant__ cuda::std::array<int64_t, NIDX * IDX_NDIM>
+        indices_strides) {
+  LocT upd_idx = cg::this_grid().thread_rank();
+  if (upd_idx >= size) {
+    return;
+  }
+
+  LocT out_elem = upd_idx % upd_post_idx_size;
+  LocT idx_elem = upd_idx / upd_post_idx_size;
+
+  LocT out_idx = elem_to_loc(
+      out_elem, upd_shape.data() + IDX_NDIM, out_strides.data(), out_ndim);
+
+#pragma unroll
+  for (int i = 0; i < NIDX; ++i) {
+    LocT idx_loc = elem_to_loc_nd<IDX_NDIM>(
+        idx_elem,
+        indices_shape.data() + i * IDX_NDIM,
+        indices_strides.data() + i * IDX_NDIM);
+    int32_t axis = axes[i];
+    LocT idx_val = absolute_index(indices[i][idx_loc], out_shape[axis]);
+    out_idx += idx_val * out_strides[axis];
+  }
+
+  LocT upd_loc = elem_to_loc(
+      out_elem + idx_elem * upd_post_idx_size,
+      upd_shape.data(),
+      upd_strides.data(),
+      upd_ndim);
+
+  Op{}(out + out_idx, upd[upd_loc]);
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/scatter_axis.cuh

@@ -0,0 +1,67 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device/indexing.cuh"
+#include "mlx/backend/cuda/device/scatter_ops.cuh"
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core::cu {
+
+namespace cg = cooperative_groups;
+
+template <
+    typename T,
+    typename IdxT,
+    typename Op,
+    int NDIM,
+    bool UpdC,
+    bool IdxC,
+    typename LocT>
+__global__ void scatter_axis(
+    const T* upd,
+    const IdxT* indices,
+    T* out,
+    LocT idx_size_pre,
+    LocT idx_size_axis,
+    LocT idx_size_post,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> upd_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> idx_strides,
+    int32_t axis,
+    int32_t axis_size,
+    int64_t upd_stride_axis,
+    int64_t idx_stride_axis) {
+  LocT index = cg::this_grid().thread_rank();
+  if (index >= idx_size_pre * idx_size_axis * idx_size_post) {
+    return;
+  }
+
+  auto [x, y, z] = index_to_dims(index, idx_size_axis, idx_size_pre);
+
+  LocT elem_idx = z * idx_size_post;
+
+  LocT idx_loc = y * idx_stride_axis;
+  if constexpr (IdxC) {
+    idx_loc += elem_idx * idx_size_axis + x;
+  } else {
+    idx_loc +=
+        elem_to_loc_nd<NDIM>(elem_idx + x, shape.data(), idx_strides.data());
+  }
+
+  auto idx_val = absolute_index(indices[idx_loc], axis_size);
+
+  LocT upd_loc = y * upd_stride_axis;
+  if constexpr (UpdC) {
+    upd_loc += elem_idx * idx_size_axis + x;
+  } else {
+    upd_loc +=
+        elem_to_loc_nd<NDIM>(elem_idx + x, shape.data(), upd_strides.data());
+  }
+
+  LocT out_idx = idx_val * idx_size_post + elem_idx * axis_size + x;
+
+  Op{}(out + out_idx, upd[upd_loc]);
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/scatter_ops.cuh

@@ -0,0 +1,44 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device/atomic_ops.cuh"
+
+namespace mlx::core::cu {
+
+struct ScatterAssign {
+  template <typename T>
+  __device__ void operator()(T* out, T val) const {
+    *out = val;
+  }
+};
+
+struct ScatterSum {
+  template <typename T>
+  __device__ void operator()(T* out, T val) const {
+    atomic_add(out, val);
+  }
+};
+
+struct ScatterProd {
+  template <typename T>
+  __device__ void operator()(T* out, T val) const {
+    atomic_prod(out, val);
+  }
+};
+
+struct ScatterMax {
+  template <typename T>
+  __device__ void operator()(T* out, T val) const {
+    atomic_max(out, val);
+  }
+};
+
+struct ScatterMin {
+  template <typename T>
+  __device__ void operator()(T* out, T val) const {
+    atomic_min(out, val);
+  }
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/ternary_ops.cuh

@@ -0,0 +1,13 @@
+// Copyright © 2025 Apple Inc.
+#pragma once
+
+namespace mlx::core::cu {
+
+struct Select {
+  template <typename T>
+  __device__ T operator()(bool condition, T x, T y) {
+    return condition ? x : y;
+  }
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/unary_ops.cuh

@@ -0,0 +1,337 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device/fp16_math.cuh"
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <math_constants.h>
+
+namespace mlx::core::cu {
+
+struct Abs {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (cuda::std::is_unsigned_v<T>) {
+      return x;
+    } else {
+      return abs(x);
+    }
+  }
+};
+
+struct ArcCos {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return acos(x);
+  }
+};
+
+struct ArcCosh {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return acosh(x);
+  }
+};
+
+struct ArcSin {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return asin(x);
+  }
+};
+
+struct ArcSinh {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return asinh(x);
+  }
+};
+
+struct ArcTan {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return atan(x);
+  }
+};
+
+struct ArcTanh {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return atanh(x);
+  }
+};
+
+struct BitwiseInvert {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return ~x;
+  }
+};
+
+struct Ceil {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (cuda::std::is_integral_v<T>) {
+      return x;
+    } else if constexpr (is_complex_v<T>) {
+      return T{ceil(x.real()), ceil(x.imag())};
+    } else {
+      return ceil(x);
+    }
+  }
+};
+
+struct Conjugate {
+  template <typename T>
+  __device__ complex_t<T> operator()(complex_t<T> x) {
+    return conj(x);
+  }
+};
+
+struct Cos {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return cos(x);
+  }
+};
+
+struct Cosh {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return cosh(x);
+  }
+};
+
+struct Erf {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (cuda::std::is_same_v<T, __half>) {
+      return erf(__half2float(x));
+    } else if constexpr (cuda::std::is_same_v<T, __nv_bfloat16>) {
+      return erf(__bfloat162float(x));
+    } else {
+      return erf(x);
+    }
+  }
+};
+
+struct ErfInv {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (cuda::std::is_same_v<T, __half>) {
+      return erfinv(__half2float(x));
+    } else if constexpr (cuda::std::is_same_v<T, __nv_bfloat16>) {
+      return erfinv(__bfloat162float(x));
+    } else {
+      return erfinv(x);
+    }
+  }
+};
+
+struct Exp {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return exp(x);
+  }
+};
+
+struct Expm1 {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (cuda::std::is_same_v<T, __half>) {
+      return expm1(__half2float(x));
+    } else if constexpr (cuda::std::is_same_v<T, __nv_bfloat16>) {
+      return expm1(__bfloat162float(x));
+    } else {
+      return expm1(x);
+    }
+  }
+};
+
+struct Floor {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (cuda::std::is_integral_v<T>) {
+      return x;
+    } else if constexpr (is_complex_v<T>) {
+      return T{floor(x.real()), floor(x.imag())};
+    } else {
+      return floor(x);
+    }
+  }
+};
+
+struct Imag {
+  template <typename T>
+  __device__ auto operator()(complex_t<T> x) {
+    return x.imag();
+  }
+};
+
+struct Log {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return log(x);
+  }
+};
+
+struct Log2 {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (is_complex_v<T>) {
+      auto y = Log{}(x);
+      return {y.real() / CUDART_LN2_F, y.imag() / CUDART_LN2_F};
+    } else {
+      return log2(x);
+    }
+  }
+};
+
+struct Log10 {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return log10(x);
+  }
+};
+
+struct Log1p {
+  template <typename T>
+  __device__ T operator()(T z) {
+    if constexpr (is_complex_v<T>) {
+      float x = z.real();
+      float y = z.imag();
+      float zabs = Abs{}(z).real();
+      float theta = atan2f(y, x + 1);
+      if (zabs < 0.5f) {
+        float r = x * (2 + x) + y * y;
+        if (r == 0) { // handle underflow
+          return {x, theta};
+        }
+        return {0.5f * log1pf(r), theta};
+      } else {
+        float z0 = hypotf(x + 1, y);
+        return {logf(z0), theta};
+      }
+    } else {
+      return log1p(z);
+    }
+  }
+};
+
+struct LogicalNot {
+  __device__ bool operator()(bool x) {
+    return !x;
+  }
+};
+
+struct Negative {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (is_complex_v<T>) {
+      return T{0, 0} - x;
+    } else {
+      return -x;
+    }
+  }
+};
+
+struct Real {
+  template <typename T>
+  __device__ auto operator()(complex_t<T> x) {
+    return x.real();
+  }
+};
+
+struct Round {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (is_complex_v<T>) {
+      return {rint(x.real()), rint(x.imag())};
+    } else {
+      return rint(x);
+    }
+  }
+};
+
+struct Sigmoid {
+  template <typename T>
+  __device__ T operator()(T x) {
+    T y = 1 / (1 + exp(-abs(x)));
+    return (x < 0) ? 1 - y : y;
+  }
+};
+
+struct Sign {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (cuda::std::is_unsigned_v<T>) {
+      return x != 0;
+    } else if constexpr (is_complex_v<T>) {
+      if (x.real() == 0 && x.imag() == 0) {
+        return x;
+      } else {
+        return x / Abs()(x);
+      }
+    } else if constexpr (cuda::std::is_same_v<T, __nv_bfloat16>) {
+      return static_cast<float>((x > T(0.f)) - (x < T(0.f)));
+    } else {
+      return (x > T(0)) - (x < T(0));
+    }
+  }
+};
+
+struct Sin {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return sin(x);
+  }
+};
+
+struct Sinh {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return sinh(x);
+  }
+};
+
+struct Square {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return x * x;
+  }
+};
+
+struct Sqrt {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return sqrt(x);
+  }
+};
+
+struct Rsqrt {
+  template <typename T>
+  __device__ T operator()(T x) {
+    if constexpr (is_complex_v<T>) {
+      return 1.0f / Sqrt{}(x);
+    } else {
+      return rsqrt(x);
+    }
+  }
+};
+
+struct Tan {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return tan(x);
+  }
+};
+
+struct Tanh {
+  template <typename T>
+  __device__ T operator()(T x) {
+    return tanh(x);
+  }
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/utils.cuh

@@ -0,0 +1,362 @@
+// Copyright © 2025 Apple Inc.
+
+// This file must not include any host-only code, utilies that work under both
+// host and device can be put here.
+//
+// See more about the requirements at:
+// https://docs.nvidia.com/cuda/nvrtc/#language
+
+#pragma once
+
+#include "mlx/backend/cuda/device/complex.cuh"
+#include "mlx/backend/cuda/device/config.h"
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <cuda/std/array>
+#include <cuda/std/limits>
+#include <cuda/std/tuple>
+
+namespace mlx::core::cu {
+
+///////////////////////////////////////////////////////////////////////////////
+// CUDA kernel utils
+///////////////////////////////////////////////////////////////////////////////
+
+// To pass shape/strides to kernels via constant memory, their size must be
+// known at compile time.
+using Shape = cuda::std::array<int32_t, MAX_NDIM>;
+using Strides = cuda::std::array<int64_t, MAX_NDIM>;
+
+// Vectorized load/store.
+template <typename T, int N>
+struct alignas(sizeof(T) * N) AlignedVector {
+  T val[N];
+};
+
+template <int N, typename T>
+inline __device__ AlignedVector<T, N> load_vector(
+    const T* ptr,
+    uint32_t offset) {
+  auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
+  return from[offset];
+}
+
+template <int N, typename T>
+inline __device__ void
+store_vector(T* ptr, uint32_t offset, const AlignedVector<T, N>& vec) {
+  auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
+  to[offset] = vec;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// Type limits utils
+///////////////////////////////////////////////////////////////////////////////
+
+template <typename T, typename = void>
+struct Limits {
+  static constexpr __host__ __device__ T max() {
+    return cuda::std::numeric_limits<T>::max();
+  }
+  static constexpr __host__ __device__ T min() {
+    return cuda::std::numeric_limits<T>::min();
+  }
+  static constexpr __host__ __device__ T finite_max() {
+    return cuda::std::numeric_limits<T>::max();
+  }
+  static constexpr __host__ __device__ T finite_min() {
+    return cuda::std::numeric_limits<T>::min();
+  }
+};
+
+template <typename T>
+struct Limits<
+    T,
+    cuda::std::enable_if_t<
+        cuda::std::is_same_v<T, float> || cuda::std::is_same_v<T, double>>> {
+  static constexpr __host__ __device__ T max() {
+    return cuda::std::numeric_limits<T>::infinity();
+  }
+  static constexpr __host__ __device__ T min() {
+    return -cuda::std::numeric_limits<T>::infinity();
+  }
+  static constexpr __host__ __device__ T finite_max() {
+    return cuda::std::numeric_limits<T>::max();
+  }
+  static constexpr __host__ __device__ T finite_min() {
+    return cuda::std::numeric_limits<T>::lowest();
+  }
+};
+
+// CUDA 11 does not have host side arithmatic operators for half types.
+template <typename T>
+struct Limits<
+    T,
+    cuda::std::enable_if_t<
+        cuda::std::is_same_v<T, __half> ||
+        cuda::std::is_same_v<T, __nv_bfloat16>>> {
+  static constexpr __host__ __device__ T max() {
+    return cuda::std::numeric_limits<T>::infinity();
+  }
+  static constexpr __host__ __device__ T min() {
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
+    return -cuda::std::numeric_limits<float>::infinity();
+#else
+    return -cuda::std::numeric_limits<T>::infinity();
+#endif
+  }
+  static constexpr __host__ __device__ T finite_max() {
+    return cuda::std::numeric_limits<T>::max();
+  }
+  static constexpr __host__ __device__ T finite_min() {
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
+    return cuda::std::numeric_limits<float>::lowest();
+#else
+    return cuda::std::numeric_limits<T>::lowest();
+#endif
+  }
+};
+
+template <>
+struct Limits<bool> {
+  static constexpr __host__ __device__ bool max() {
+    return true;
+  }
+  static constexpr __host__ __device__ bool min() {
+    return false;
+  }
+};
+
+template <typename T>
+struct Limits<complex_t<T>> {
+  static constexpr __host__ __device__ complex_t<T> max() {
+    return {Limits<T>::max(), Limits<T>::max()};
+  }
+  static constexpr __host__ __device__ complex_t<T> min() {
+    return {Limits<T>::min(), Limits<T>::min()};
+  }
+};
+
+///////////////////////////////////////////////////////////////////////////////
+// Indexing utils
+///////////////////////////////////////////////////////////////////////////////
+
+template <typename IdxT = int64_t>
+inline __host__ __device__ IdxT
+elem_to_loc(IdxT elem, const int* shape, const int64_t* strides, int ndim) {
+  IdxT loc = 0;
+  for (int i = ndim - 1; i >= 0 && elem > 0; --i) {
+    loc += (elem % shape[i]) * IdxT(strides[i]);
+    elem /= shape[i];
+  }
+  return loc;
+}
+
+// Optimize when the ndim is known at compile time.
+template <int NDIM, typename IdxT = int64_t>
+inline __host__ __device__ IdxT
+elem_to_loc_nd(IdxT elem, const int* shape, const int64_t* strides) {
+  IdxT loc = 0;
+#pragma unroll
+  for (int i = NDIM - 1; i >= 0; --i) {
+    loc += (elem % shape[i]) * IdxT(strides[i]);
+    elem /= shape[i];
+  }
+  return loc;
+}
+
+template <int NDIM, typename IdxT = int64_t>
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_nd(
+    IdxT elem,
+    const int* shape,
+    const int64_t* a_strides,
+    const int64_t* b_strides) {
+  IdxT a_loc = 0;
+  IdxT b_loc = 0;
+#pragma unroll
+  for (int i = NDIM - 1; i >= 0; --i) {
+    int dim_idx = elem % shape[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
+    b_loc += dim_idx * IdxT(b_strides[i]);
+    elem /= shape[i];
+  }
+  return cuda::std::make_tuple(a_loc, b_loc);
+}
+
+template <int NDIM, typename IdxT = int64_t>
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_nd(
+    IdxT elem,
+    const int* shape,
+    const int64_t* a_strides,
+    const int64_t* b_strides,
+    const int64_t* c_strides) {
+  IdxT a_loc = 0;
+  IdxT b_loc = 0;
+  IdxT c_loc = 0;
+#pragma unroll
+  for (int i = NDIM - 1; i >= 0; --i) {
+    int dim_idx = elem % shape[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
+    b_loc += dim_idx * IdxT(b_strides[i]);
+    c_loc += dim_idx * IdxT(c_strides[i]);
+    elem /= shape[i];
+  }
+  return cuda::std::make_tuple(a_loc, b_loc, c_loc);
+}
+
+// Optimized version when ndim is larger than 4.
+template <typename IdxT = int64_t>
+inline __host__ __device__ IdxT
+elem_to_loc_4d(IdxT elem, const int* shape, const int64_t* strides, int ndim) {
+  IdxT loc = 0;
+  for (int i = ndim - 1; i >= 0; --i) {
+    loc += (elem % shape[i]) * IdxT(strides[i]);
+    elem /= shape[i];
+  }
+  return loc;
+}
+
+template <typename IdxT = int64_t>
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_4d(
+    IdxT elem,
+    const int* shape,
+    const int64_t* a_strides,
+    const int64_t* b_strides,
+    int ndim) {
+  IdxT a_loc = 0;
+  IdxT b_loc = 0;
+  for (int i = ndim - 1; i >= 0; --i) {
+    int dim_idx = elem % shape[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
+    b_loc += dim_idx * IdxT(b_strides[i]);
+    elem /= shape[i];
+  }
+  return cuda::std::make_tuple(a_loc, b_loc);
+}
+
+template <typename IdxT = int64_t>
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_4d(
+    IdxT elem,
+    const int* shape,
+    const int64_t* a_strides,
+    const int64_t* b_strides,
+    const int64_t* c_strides,
+    int ndim) {
+  IdxT a_loc = 0;
+  IdxT b_loc = 0;
+  IdxT c_loc = 0;
+  for (int i = ndim - 1; i >= 0; --i) {
+    int dim_idx = elem % shape[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
+    b_loc += dim_idx * IdxT(b_strides[i]);
+    c_loc += dim_idx * IdxT(c_strides[i]);
+    elem /= shape[i];
+  }
+  return cuda::std::make_tuple(a_loc, b_loc, c_loc);
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// Elem to loc in a loop utils
+///////////////////////////////////////////////////////////////////////////////
+
+template <int DIM, bool General = true, typename OffsetT = size_t>
+struct LoopedElemToLoc {
+  int dim;
+  LoopedElemToLoc<DIM - 1, General, OffsetT> inner_looper;
+  OffsetT offset{0};
+  int index{0};
+
+  __device__ LoopedElemToLoc(int dim) : dim(dim), inner_looper(dim - 1) {}
+
+  __device__ void next(const int* shape, const int64_t* strides) {
+    if (dim == 0) {
+      return;
+    }
+    index++;
+    offset += OffsetT(strides[dim - 1]);
+    if (index >= shape[dim - 1]) {
+      index = 0;
+      inner_looper.next(shape, strides);
+      offset = inner_looper.offset;
+    }
+  }
+
+  __device__ void next(int n, const int* shape, const int64_t* strides) {
+    if (dim == 0) {
+      return;
+    }
+    index += n;
+    offset += n * OffsetT(strides[dim - 1]);
+
+    if (index >= shape[dim - 1]) {
+      int extra = index - shape[dim - 1];
+      if (extra >= shape[dim - 1]) {
+        inner_looper.next(1 + extra / shape[dim - 1], shape, strides);
+        extra = extra % shape[dim - 1];
+      } else {
+        inner_looper.next(shape, strides);
+      }
+      index = 0;
+      offset = inner_looper.offset;
+      if (extra > 0) {
+        next(extra, shape, strides);
+      }
+    }
+  }
+
+  __device__ OffsetT location() {
+    return offset;
+  }
+};
+
+template <typename OffsetT>
+struct LoopedElemToLoc<1, true, OffsetT> {
+  int dim;
+  OffsetT offset{0};
+  int index{0};
+
+  __device__ LoopedElemToLoc(int dim) : dim(dim) {}
+
+  __device__ void next(const int* shape, const int64_t* strides) {
+    index++;
+    if (dim > 1) {
+      offset = elem_to_loc<OffsetT>(index, shape, strides, dim);
+    } else {
+      offset += OffsetT(strides[0]);
+    }
+  }
+
+  __device__ void next(int n, const int* shape, const int64_t* strides) {
+    index += n;
+    if (dim > 1) {
+      offset = elem_to_loc<OffsetT>(index, shape, strides, dim);
+    } else {
+      offset = index * OffsetT(strides[0]);
+    }
+  }
+
+  __device__ OffsetT location() {
+    return offset;
+  }
+};
+
+template <typename OffsetT>
+struct LoopedElemToLoc<1, false, OffsetT> {
+  OffsetT offset{0};
+
+  __device__ LoopedElemToLoc(int) {}
+
+  __device__ void next(const int*, const int64_t* strides) {
+    offset += OffsetT(strides[0]);
+  }
+
+  __device__ void next(int n, const int*, const int64_t* strides) {
+    offset += n * OffsetT(strides[0]);
+  }
+
+  __device__ OffsetT location() {
+    return offset;
+  }
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/eval.cpp

@@ -37,22 +37,20 @@ void eval(array& arr) {
   }
 
   auto& encoder = cu::get_command_encoder(arr.primitive().stream());
-  if (encoder.has_gpu_work()) {
-    // Keep used buffers alive until kernel finishes running.
-    std::unordered_set<std::shared_ptr<array::Data>> buffers;
-    for (auto& in : arr.inputs()) {
-      buffers.insert(in.data_shared_ptr());
-    }
-    for (auto& s : arr.siblings()) {
-      buffers.insert(s.data_shared_ptr());
-    }
-    // Remove the output if it was donated to by an input.
-    if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
-      buffers.erase(it);
-    }
-    encoder.add_completed_handler([buffers = std::move(buffers)]() {});
+  // Keep used buffers alive until kernel finishes running.
+  std::unordered_set<std::shared_ptr<array::Data>> buffers;
+  for (auto& in : arr.inputs()) {
+    buffers.insert(in.data_shared_ptr());
   }
-  encoder.end_encoding();
+  for (auto& s : arr.siblings()) {
+    buffers.insert(s.data_shared_ptr());
+  }
+  // Remove the output if it was donated to by an input.
+  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
+    buffers.erase(it);
+  }
+  encoder.add_completed_handler([buffers = std::move(buffers)]() {});
+  encoder.maybe_commit();
 }
 
 void finalize(Stream s) {
@@ -62,7 +60,7 @@ void finalize(Stream s) {
 
 void synchronize(Stream s) {
   nvtx3::scoped_range r("gpu::synchronize");
-  cu::get_stream(s).synchronize();
+  cu::get_command_encoder(s).synchronize();
 }
 
 } // namespace mlx::core::gpu

mlx/backend/cuda/event.cu

@@ -61,7 +61,9 @@ void CudaEvent::wait(Stream s) {
   if (s.device == mlx::core::Device::cpu) {
     scheduler::enqueue(s, [*this]() mutable { wait(); });
   } else {
-    wait(cu::get_stream(s).last_cuda_stream());
+    auto& enc = cu::get_command_encoder(s);
+    enc.commit();
+    wait(enc.stream());
   }
 }
 
@@ -74,7 +76,9 @@ void CudaEvent::record(Stream s) {
   if (s.device == mlx::core::Device::cpu) {
     throw std::runtime_error("CudaEvent can not wait on cpu stream.");
   } else {
-    record(cu::get_stream(s).last_cuda_stream());
+    auto& enc = cu::get_command_encoder(s);
+    enc.commit();
+    record(enc.stream());
   }
 }
 
@@ -86,8 +90,6 @@ bool CudaEvent::completed() const {
 // SharedEvent implementations
 ///////////////////////////////////////////////////////////////////////////////
 
-namespace {
-
 __host__ __device__ void event_wait(SharedEvent::Atomic* ac, uint64_t value) {
   uint64_t current;
   while ((current = ac->load()) < value) {
@@ -108,8 +110,6 @@ __global__ void event_signal_kernel(SharedEvent::Atomic* ac, uint64_t value) {
   event_signal(ac, value);
 }
 
-} // namespace
-
 SharedEvent::SharedEvent() {
   // Allocate cuda::atomic on managed memory.
   Atomic* ac;
@@ -136,11 +136,9 @@ void SharedEvent::wait(Stream s, uint64_t value) {
     scheduler::enqueue(s, [*this, value]() mutable { wait(value); });
   } else {
     auto& encoder = get_command_encoder(s);
-    encoder.launch_kernel(
-        encoder.stream().last_cuda_stream(),
-        [this, value](cudaStream_t stream) { wait(stream, value); });
+    encoder.commit();
+    wait(encoder.stream(), value);
     encoder.add_completed_handler([ac = ac_]() {});
-    encoder.end_encoding();
   }
 }
 
@@ -162,11 +160,9 @@ void SharedEvent::signal(Stream s, uint64_t value) {
     scheduler::enqueue(s, [*this, value]() mutable { signal(stream, value); });
   } else {
     auto& encoder = get_command_encoder(s);
-    encoder.launch_kernel(
-        encoder.stream().last_cuda_stream(),
-        [this, value](cudaStream_t stream) { signal(stream, value); });
+    encoder.commit();
+    signal(encoder.stream(), value);
     encoder.add_completed_handler([ac = ac_]() {});
-    encoder.end_encoding();
   }
 }
 

mlx/backend/cuda/indexing.cpp

@@ -0,0 +1,428 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/compiled.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/jit_module.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include "cuda_jit_sources.h"
+
+#include <cuda.h>
+#include <fmt/format.h>
+#include <nvrtc.h>
+#include <nvtx3/nvtx3.hpp>
+
+#include <cassert>
+#include <numeric>
+
+namespace mlx::core {
+
+namespace {
+
+constexpr const char* g_scatter_ops[] = {"Max", "Min", "Sum", "Prod", "Assign"};
+
+void append_indices_arg(
+    cu::KernelArgs& args,
+    const std::vector<array>& inputs,
+    int nidx,
+    int idx_ndim) {
+  std::vector<const void*> indices(nidx);
+  for (int i = 0; i < nidx; ++i) {
+    indices[i] = inputs[i + 1].data<void>();
+  }
+  args.append(std::move(indices));
+  std::vector<int32_t> indices_shape(nidx * idx_ndim);
+  for (int i = 0; i < nidx; ++i) {
+    std::copy_n(
+        inputs[i + 1].shape().begin(),
+        idx_ndim,
+        indices_shape.data() + i * idx_ndim);
+  }
+  args.append(std::move(indices_shape));
+  std::vector<int64_t> indices_strides(nidx * idx_ndim);
+  for (int i = 0; i < nidx; ++i) {
+    std::copy_n(
+        inputs[i + 1].strides().begin(),
+        idx_ndim,
+        indices_strides.data() + i * idx_ndim);
+  }
+  args.append(std::move(indices_strides));
+}
+
+} // namespace
+
+void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Gather::eval_gpu");
+  assert(inputs.size() > 0);
+  const auto& src = inputs[0];
+
+  out.set_data(allocator::malloc(out.nbytes()));
+  if (out.size() == 0) {
+    return;
+  }
+
+  int nidx = inputs.size() - 1;
+  Dtype idx_dtype = nidx > 0 ? inputs[1].dtype() : int32;
+  int32_t idx_ndim = nidx > 0 ? inputs[1].ndim() : 0;
+
+  bool large = (nidx > 0 && inputs[1].size() > INT32_MAX) ||
+      (src.size() > INT32_MAX) || (out.size() > INT32_MAX);
+
+  uint32_t slice_size = std::accumulate(
+      slice_sizes_.begin(), slice_sizes_.end(), 1, std::multiplies<uint32_t>());
+
+  std::string module_name = fmt::format(
+      "gather_{}_{}_{}",
+      dtype_to_string(out.dtype()),
+      dtype_to_string(idx_dtype),
+      nidx);
+
+  auto& s = stream();
+  cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
+    std::vector<std::string> kernel_names;
+    for (int ndim = 0; ndim <= MAX_NDIM; ++ndim) {
+      for (int large = 0; large <= 1; ++large) {
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::gather<{}, {}, {}, {}, {}>",
+            dtype_to_cuda_type(out.dtype()),
+            dtype_to_cuda_type(idx_dtype),
+            nidx,
+            ndim,
+            large ? "int64_t" : "int32_t"));
+      }
+    }
+    return std::make_pair(jit_source_gather, std::move(kernel_names));
+  });
+
+  cu::KernelArgs args;
+  args.append(src);
+  args.append(out);
+  if (large) {
+    args.append<int64_t>(out.size());
+  } else {
+    args.append<int32_t>(out.size());
+  }
+  args.append_ndim(src.shape());
+  args.append_ndim(src.strides());
+  args.append<int32_t>(src.ndim());
+  args.append_ndim(slice_sizes_);
+  args.append(slice_size);
+  args.append(axes_);
+  append_indices_arg(args, inputs, nidx, idx_ndim);
+
+  std::string kernel_name = fmt::format(
+      "mlx::core::cu::gather<{}, {}, {}, {}, {}>",
+      dtype_to_cuda_type(out.dtype()),
+      dtype_to_cuda_type(idx_dtype),
+      nidx,
+      idx_ndim,
+      large ? "int64_t" : "int32_t");
+
+  auto& encoder = cu::get_command_encoder(s);
+  for (const auto& in : inputs) {
+    encoder.set_input_array(in);
+  }
+  encoder.set_output_array(out);
+
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+}
+
+void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Gather::eval_gpu");
+  assert(inputs.size() > 1);
+  auto& upd = inputs.back();
+
+  // Copy src into out.
+  CopyType copy_type;
+  if (inputs[0].data_size() == 1) {
+    copy_type = CopyType::Scalar;
+  } else if (inputs[0].flags().row_contiguous) {
+    copy_type = CopyType::Vector;
+  } else {
+    copy_type = CopyType::General;
+  }
+  copy_gpu(inputs[0], out, copy_type);
+
+  // Empty update.
+  if (upd.size() == 0) {
+    return;
+  }
+
+  int nidx = axes_.size();
+  Dtype idx_dtype = nidx > 0 ? inputs[1].dtype() : int32;
+  int32_t idx_ndim = nidx > 0 ? inputs[1].ndim() : 0;
+
+  bool large = (nidx > 0 && inputs[1].size() > INT32_MAX) ||
+      (upd.size() > INT32_MAX) || (out.size() > INT32_MAX);
+
+  int32_t upd_post_idx_size = std::accumulate(
+      upd.shape().begin() + idx_ndim,
+      upd.shape().end(),
+      1,
+      std::multiplies<int32_t>());
+
+  const char* op = g_scatter_ops[reduce_type_];
+  std::string module_name = fmt::format(
+      "scatter_{}_{}_{}_{}",
+      dtype_to_string(out.dtype()),
+      dtype_to_string(idx_dtype),
+      op,
+      nidx);
+
+  auto& s = stream();
+  cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
+    std::vector<std::string> kernel_names;
+    for (int ndim = 0; ndim <= MAX_NDIM; ++ndim) {
+      for (int large = 0; large <= 1; ++large) {
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::scatter<{}, {}, mlx::core::cu::Scatter{}, {}, {}, {}>",
+            dtype_to_cuda_type(out.dtype()),
+            dtype_to_cuda_type(idx_dtype),
+            op,
+            nidx,
+            ndim,
+            large ? "int64_t" : "int32_t"));
+      }
+    }
+    return std::make_pair(jit_source_scatter, std::move(kernel_names));
+  });
+
+  cu::KernelArgs args;
+  args.append(upd);
+  args.append(out);
+  if (large) {
+    args.append<int64_t>(upd.size());
+  } else {
+    args.append<int32_t>(upd.size());
+  }
+  args.append_ndim(upd.shape());
+  args.append_ndim(upd.strides());
+  args.append<int32_t>(upd.ndim());
+  if (large) {
+    args.append<int64_t>(upd_post_idx_size);
+  } else {
+    args.append<int32_t>(upd_post_idx_size);
+  }
+  args.append_ndim(out.shape());
+  args.append_ndim(out.strides());
+  args.append<int32_t>(out.ndim());
+  args.append(axes_);
+  append_indices_arg(args, inputs, nidx, idx_ndim);
+
+  std::string kernel_name = fmt::format(
+      "mlx::core::cu::scatter<{}, {}, mlx::core::cu::Scatter{}, {}, {}, {}>",
+      dtype_to_cuda_type(out.dtype()),
+      dtype_to_cuda_type(idx_dtype),
+      op,
+      nidx,
+      idx_ndim,
+      large ? "int64_t" : "int32_t");
+
+  auto& encoder = cu::get_command_encoder(s);
+  for (const auto& in : inputs) {
+    encoder.set_input_array(in);
+  }
+  encoder.set_output_array(out);
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] = get_launch_args(kernel, upd, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+}
+
+void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("GatherAxis::eval_gpu");
+  assert(inputs.size() > 1);
+  const auto& src = inputs[0];
+  const auto& idx = inputs[1];
+
+  out.set_data(allocator::malloc(out.nbytes()));
+  if (out.size() == 0) {
+    return;
+  }
+
+  bool large = idx.size() > INT32_MAX || src.size() > INT32_MAX;
+
+  std::string module_name = fmt::format(
+      "gather_axis_{}_{}",
+      dtype_to_string(out.dtype()),
+      dtype_to_string(idx.dtype()));
+
+  auto& s = stream();
+  cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
+    std::vector<std::string> kernel_names;
+    for (int ndim = 0; ndim <= MAX_NDIM; ++ndim) {
+      for (int contiguous = 0; contiguous < 4; ++contiguous) {
+        for (int large = 0; large <= 1; ++large) {
+          kernel_names.push_back(fmt::format(
+              "mlx::core::cu::gather_axis<{}, {}, {}, {}, {}, {}>",
+              dtype_to_cuda_type(out.dtype()),
+              dtype_to_cuda_type(idx.dtype()),
+              ndim,
+              contiguous & 1 ? true : false,
+              contiguous & 2 ? true : false,
+              large ? "int64_t" : "int32_t"));
+        }
+      }
+    }
+    return std::make_pair(jit_source_gather_axis, std::move(kernel_names));
+  });
+
+  size_t idx_size_pre = 1;
+  size_t idx_size_post = 1;
+  for (int i = 0; i < axis_; ++i) {
+    idx_size_pre *= idx.shape(i);
+  }
+  for (int i = axis_ + 1; i < idx.ndim(); ++i) {
+    idx_size_post *= idx.shape(i);
+  }
+  size_t idx_size_axis = idx.shape(axis_);
+
+  cu::KernelArgs args;
+  args.append(src);
+  args.append(idx);
+  args.append(out);
+  if (large) {
+    args.append<int64_t>(idx_size_pre);
+    args.append<int64_t>(idx_size_axis);
+    args.append<int64_t>(idx_size_post);
+  } else {
+    args.append<int32_t>(idx_size_pre);
+    args.append<int32_t>(idx_size_axis);
+    args.append<int32_t>(idx_size_post);
+  }
+  args.append(remove_index(idx.shape(), axis_));
+  args.append(remove_index(src.strides(), axis_));
+  args.append(remove_index(idx.strides(), axis_));
+  args.append<int32_t>(axis_);
+  args.append(src.shape(axis_));
+  args.append(src.strides(axis_));
+  args.append(idx.strides(axis_));
+
+  std::string kernel_name = fmt::format(
+      "mlx::core::cu::gather_axis<{}, {}, {}, {}, {}, {}>",
+      dtype_to_cuda_type(out.dtype()),
+      dtype_to_cuda_type(idx.dtype()),
+      src.ndim() - 1,
+      src.flags().row_contiguous,
+      idx.flags().row_contiguous,
+      large ? "int64_t" : "int32_t");
+
+  auto& encoder = cu::get_command_encoder(s);
+  for (const auto& in : inputs) {
+    encoder.set_input_array(in);
+  }
+  encoder.set_output_array(out);
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+}
+
+void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("ScatterAxis::eval_gpu");
+  assert(inputs.size() > 2);
+  const auto& src = inputs[0];
+  const auto& idx = inputs[1];
+  const auto& upd = inputs[2];
+
+  // Copy src into out.
+  CopyType copy_type;
+  if (src.data_size() == 1) {
+    copy_type = CopyType::Scalar;
+  } else if (src.flags().row_contiguous) {
+    copy_type = CopyType::Vector;
+  } else {
+    copy_type = CopyType::General;
+  }
+  copy_gpu(src, out, copy_type);
+
+  // Empty update.
+  if (upd.size() == 0) {
+    return;
+  }
+
+  bool large = idx.size() > INT32_MAX || src.size() > INT32_MAX;
+
+  const char* op = reduce_type_ == ScatterAxis::Sum ? "Sum" : "Assign";
+  std::string module_name = fmt::format(
+      "scatter_axis_{}_{}_{}",
+      dtype_to_string(out.dtype()),
+      dtype_to_string(idx.dtype()),
+      op);
+
+  auto& s = stream();
+  cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
+    std::vector<std::string> kernel_names;
+    for (int ndim = 0; ndim <= MAX_NDIM; ++ndim) {
+      for (int contiguous = 0; contiguous < 4; ++contiguous) {
+        for (int large = 0; large <= 1; ++large) {
+          kernel_names.push_back(fmt::format(
+              "mlx::core::cu::scatter_axis<{}, {}, mlx::core::cu::Scatter{}, {}, {}, {}, {}>",
+              dtype_to_cuda_type(out.dtype()),
+              dtype_to_cuda_type(idx.dtype()),
+              op,
+              ndim,
+              contiguous & 1 ? true : false,
+              contiguous & 2 ? true : false,
+              large ? "int64_t" : "int32_t"));
+        }
+      }
+    }
+    return std::make_pair(jit_source_scatter_axis, std::move(kernel_names));
+  });
+
+  size_t idx_size_pre = 1;
+  size_t idx_size_post = 1;
+  for (int i = 0; i < axis_; ++i) {
+    idx_size_pre *= idx.shape(i);
+  }
+  for (int i = axis_ + 1; i < idx.ndim(); ++i) {
+    idx_size_post *= idx.shape(i);
+  }
+  size_t idx_size_axis = idx.shape(axis_);
+
+  cu::KernelArgs args;
+  args.append(upd);
+  args.append(idx);
+  args.append(out);
+  if (large) {
+    args.append<int64_t>(idx_size_pre);
+    args.append<int64_t>(idx_size_axis);
+    args.append<int64_t>(idx_size_post);
+  } else {
+    args.append<int32_t>(idx_size_pre);
+    args.append<int32_t>(idx_size_axis);
+    args.append<int32_t>(idx_size_post);
+  }
+  args.append(remove_index(idx.shape(), axis_));
+  args.append(remove_index(upd.strides(), axis_));
+  args.append(remove_index(idx.strides(), axis_));
+  args.append<int32_t>(axis_);
+  args.append(out.shape(axis_));
+  args.append(upd.strides(axis_));
+  args.append(idx.strides(axis_));
+
+  std::string kernel_name = fmt::format(
+      "mlx::core::cu::scatter_axis<{}, {}, mlx::core::cu::Scatter{}, {}, {}, {}, {}>",
+      dtype_to_cuda_type(out.dtype()),
+      dtype_to_cuda_type(idx.dtype()),
+      op,
+      idx.ndim() - 1,
+      upd.flags().row_contiguous,
+      idx.flags().row_contiguous,
+      large ? "int64_t" : "int32_t");
+
+  auto& encoder = cu::get_command_encoder(s);
+  for (const auto& in : inputs) {
+    encoder.set_input_array(in);
+  }
+  encoder.set_output_array(out);
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+}
+
+} // namespace mlx::core

mlx/backend/cuda/iterators/general_iterator.cuh

@@ -0,0 +1,121 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <thrust/iterator/iterator_adaptor.h>
+#include <cuda/std/utility>
+
+#include "mlx/backend/cuda/kernel_utils.cuh"
+
+namespace mlx::core::cu {
+
+// Iterating non-contiguous array.
+template <typename Iterator, typename IdxT = int64_t>
+class general_iterator
+    : public thrust::
+          iterator_adaptor<general_iterator<Iterator, IdxT>, Iterator> {
+ public:
+  using super_t =
+      thrust::iterator_adaptor<general_iterator<Iterator, IdxT>, Iterator>;
+
+  using reference = typename super_t::reference;
+  using difference_type = typename super_t::difference_type;
+
+  __host__ __device__ general_iterator(
+      Iterator it,
+      IdxT index,
+      int ndim,
+      Shape shape,
+      Strides strides)
+      : super_t(it),
+        index_(index),
+        ndim_(ndim),
+        shape_(cuda::std::move(shape)),
+        strides_(cuda::std::move(strides)) {}
+
+  __host__ __device__ IdxT index() const {
+    return index_;
+  }
+
+  __host__ __device__ const Shape& shape() const {
+    return shape_;
+  }
+
+  __host__ __device__ const Strides& strides() const {
+    return strides_;
+  }
+
+ private:
+  friend class thrust::iterator_core_access;
+
+  __host__ __device__ bool equal(const general_iterator& other) const {
+    return this->base() == other.base() && this->index() == other.index();
+  }
+
+  __host__ __device__ void advance(difference_type n) {
+    this->index_ += n;
+  }
+
+  __host__ __device__ void increment() {
+    this->index_ += 1;
+  }
+
+  __host__ __device__ void decrement() {
+    this->index_ -= 1;
+  }
+
+  __host__ __device__ difference_type
+  distance_to(const general_iterator& other) const {
+    _CCCL_ASSERT(
+        this->base() == other.base(),
+        "Underlying iterator must point to same base iterator");
+    return other.index() - this->index();
+  }
+
+  // The dereference is device-only to avoid accidental running in host.
+  __device__ typename super_t::reference dereference() const {
+    IdxT offset = elem_to_loc(index_, shape_.data(), strides_.data(), ndim_);
+    return *(this->base() + offset);
+  }
+
+  IdxT index_;
+  int ndim_;
+  Shape shape_;
+  Strides strides_;
+};
+
+template <typename IdxT, typename Iterator>
+__host__ __device__ auto make_general_iterator(
+    Iterator it,
+    IdxT index,
+    int ndim,
+    Shape shape,
+    Strides strides) {
+  return general_iterator<Iterator, IdxT>(
+      it, index, ndim, cuda::std::move(shape), cuda::std::move(strides));
+}
+
+template <typename IdxT, typename Iterator>
+auto make_general_iterator(
+    Iterator it,
+    const std::vector<int32_t>& shape,
+    const std::vector<int64_t>& strides) {
+  return make_general_iterator<IdxT>(
+      it, 0, shape.size(), const_param(shape), const_param(strides));
+}
+
+template <typename IdxT, typename Iterator>
+auto make_general_iterators(
+    Iterator it,
+    IdxT size,
+    const std::vector<int32_t>& shape,
+    const std::vector<int64_t>& strides) {
+  auto ndim = shape.size();
+  auto shape_arg = const_param(shape);
+  auto strides_arg = const_param(strides);
+  return std::make_pair(
+      make_general_iterator<IdxT>(it, 0, ndim, shape_arg, strides_arg),
+      make_general_iterator<IdxT>(it, size, ndim, shape_arg, strides_arg));
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/iterators/strided_iterator.cuh

@@ -0,0 +1,60 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <thrust/iterator/iterator_adaptor.h>
+#include <thrust/iterator/iterator_facade.h>
+
+namespace mlx::core::cu {
+
+// RandomAccessIterator for strided access to array entries.
+template <typename Iterator, typename Stride = int64_t>
+class strided_iterator
+    : public thrust::
+          iterator_adaptor<strided_iterator<Iterator, Stride>, Iterator> {
+ public:
+  using super_t =
+      thrust::iterator_adaptor<strided_iterator<Iterator, Stride>, Iterator>;
+
+  using reference = typename super_t::reference;
+  using difference_type = typename super_t::difference_type;
+
+  __host__ __device__ strided_iterator(Iterator it, Stride stride)
+      : super_t(it), stride_(stride) {}
+
+  __host__ __device__ Stride stride() const {
+    return stride_;
+  }
+
+ private:
+  friend class thrust::iterator_core_access;
+
+  __host__ __device__ bool equal(const strided_iterator& other) const {
+    return this->base() == other.base();
+  }
+
+  __host__ __device__ void advance(difference_type n) {
+    this->base_reference() += n * stride_;
+  }
+
+  __host__ __device__ void increment() {
+    this->base_reference() += stride_;
+  }
+
+  __host__ __device__ void decrement() {
+    this->base_reference() -= stride_;
+  }
+
+  __host__ __device__ difference_type
+  distance_to(const strided_iterator& other) const {
+    const difference_type dist = other.base() - this->base();
+    _CCCL_ASSERT(
+        dist % stride() == 0,
+        "Underlying iterator difference must be divisible by the stride");
+    return dist / stride();
+  }
+
+  Stride stride_;
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/jit_module.cpp

@@ -0,0 +1,345 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/jit_module.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/version.h"
+
+#include "cuda_jit_sources.h"
+
+#include <cstdlib>
+#include <filesystem>
+#include <fstream>
+#include <unordered_map>
+
+#include <fmt/format.h>
+#include <nvrtc.h>
+#include <unistd.h>
+
+namespace mlx::core::cu {
+
+namespace {
+
+#define CHECK_NVRTC_ERROR(cmd) check_nvrtc_error(#cmd, (cmd))
+
+void check_nvrtc_error(const char* name, nvrtcResult err) {
+  if (err != NVRTC_SUCCESS) {
+    throw std::runtime_error(
+        fmt::format("{} failed: {}", name, nvrtcGetErrorString(err)));
+  }
+}
+
+// Return the location of the CUDA toolkit.
+const std::string& cuda_home() {
+  static std::string home = []() -> std::string {
+    const char* home = std::getenv("CUDA_HOME");
+    if (home) {
+      return home;
+    }
+    home = std::getenv("CUDA_PATH");
+    if (home) {
+      return home;
+    }
+#if defined(__linux__)
+    home = "/usr/local/cuda";
+    if (std::filesystem::exists(home)) {
+      return home;
+    }
+#endif
+    throw std::runtime_error(
+        "Environment variable CUDA_HOME or CUDA_PATH is not set.");
+  }();
+  return home;
+}
+
+// Return the location of CCCL headers shipped with the distribution.
+const std::string& cccl_dir() {
+  static std::string dir = []() {
+    std::filesystem::path path;
+#if defined(MLX_CCCL_DIR)
+    // First search the install dir if defined.
+    path = MLX_CCCL_DIR;
+    if (std::filesystem::exists(path)) {
+      return path.string();
+    }
+#endif
+    // Then search dynamically from the dir of libmlx.so file.
+    path = current_binary_dir().parent_path() / "include" / "cccl";
+    if (std::filesystem::exists(path)) {
+      return path.string();
+    }
+    // Finally check the environment variable.
+    path = std::getenv("MLX_CCCL_DIR");
+    if (!path.empty() && std::filesystem::exists(path)) {
+      return path.string();
+    }
+    return std::string();
+  }();
+  return dir;
+}
+
+// Get the cache directory for storing compiled results.
+const std::filesystem::path& ptx_cache_dir() {
+  static std::filesystem::path cache = []() -> std::filesystem::path {
+    std::filesystem::path cache;
+    if (auto c = std::getenv("MLX_PTX_CACHE_DIR"); c) {
+      cache = c;
+    } else {
+      cache =
+          std::filesystem::temp_directory_path() / "mlx" / version() / "ptx";
+    }
+    if (!std::filesystem::exists(cache)) {
+      std::error_code error;
+      if (!std::filesystem::create_directories(cache, error)) {
+        return std::filesystem::path();
+      }
+    }
+    return cache;
+  }();
+  return cache;
+}
+
+// Try to read the cached |ptx| and |ptx_kernels| from |cache_dir|.
+bool read_cached_ptx(
+    const std::filesystem::path& cache_dir,
+    const std::string& module_name,
+    std::vector<char>* ptx,
+    std::vector<std::pair<std::string, std::string>>* ptx_kernels) {
+  if (cache_dir.empty()) {
+    return false;
+  }
+
+  auto ptx_path = cache_dir / (module_name + ".ptx");
+  std::error_code error;
+  auto ptx_size = std::filesystem::file_size(ptx_path, error);
+  if (error) {
+    return false;
+  }
+  std::ifstream ptx_file(ptx_path, std::ios::binary);
+  if (!ptx_file.good()) {
+    return false;
+  }
+  ptx->resize(ptx_size);
+  ptx_file.read(ptx->data(), ptx_size);
+
+  std::ifstream txt_file(cache_dir / (module_name + ".txt"), std::ios::binary);
+  std::string line;
+  while (std::getline(txt_file, line)) {
+    auto tab = line.find('\t');
+    if (tab != std::string::npos) {
+      ptx_kernels->emplace_back(line.substr(0, tab), line.substr(tab + 1));
+    }
+  }
+  return true;
+}
+
+// Write the |ptx| and |ptx_kernels| to |cache_dir| with |name|.
+void write_cached_ptx(
+    const std::filesystem::path& cache_dir,
+    const std::string& module_name,
+    const std::vector<char>& ptx,
+    const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
+    const std::string& source_code) {
+  if (cache_dir.empty()) {
+    return;
+  }
+
+  std::ofstream ptx_file(cache_dir / (module_name + ".ptx"), std::ios::binary);
+  if (!ptx.empty()) {
+    ptx_file.write(&ptx.front(), ptx.size());
+  }
+  std::ofstream txt_file(cache_dir / (module_name + ".txt"), std::ios::binary);
+  for (const auto& [name, mangled] : ptx_kernels) {
+    txt_file << name << "\t" << mangled << std::endl;
+  }
+
+  std::ofstream source_file(cache_dir / (module_name + ".cu"));
+  source_file << source_code;
+}
+
+// Return if |device|'s version is not newer than |major|.|minor| version.
+inline bool version_lower_equal(Device& device, int major, int minor) {
+  if (device.compute_capability_major() < major) {
+    return true;
+  } else if (device.compute_capability_major() == major) {
+    return device.compute_capability_minor() <= minor;
+  } else {
+    return false;
+  }
+}
+
+// Return whether NVRTC supports compiling to |device|'s SASS code.
+bool compiler_supports_device_sass(Device& device) {
+  int nvrtc_major, nvrtc_minor;
+  CHECK_NVRTC_ERROR(nvrtcVersion(&nvrtc_major, &nvrtc_minor));
+  if (nvrtc_major < 9) {
+    return false;
+  } else if (nvrtc_major == 9) {
+    return version_lower_equal(device, 7, 2);
+  } else if (nvrtc_major == 10) {
+    return version_lower_equal(device, 7, 5);
+  } else if (nvrtc_major == 11 && nvrtc_minor == 0) {
+    return version_lower_equal(device, 8, 0);
+  } else if (nvrtc_major == 11 && nvrtc_minor < 8) {
+    return version_lower_equal(device, 8, 6);
+  } else {
+    return true;
+  }
+}
+
+#define INCLUDE_PREFIX "mlx/backend/cuda/device/"
+
+constexpr const char* g_include_names[] = {
+    INCLUDE_PREFIX "atomic_ops.cuh",
+    INCLUDE_PREFIX "binary_ops.cuh",
+    INCLUDE_PREFIX "cast_op.cuh",
+    INCLUDE_PREFIX "config.h",
+    INCLUDE_PREFIX "complex.cuh",
+    INCLUDE_PREFIX "fp16_math.cuh",
+    INCLUDE_PREFIX "indexing.cuh",
+    INCLUDE_PREFIX "scatter_ops.cuh",
+    INCLUDE_PREFIX "unary_ops.cuh",
+    INCLUDE_PREFIX "ternary_ops.cuh",
+    INCLUDE_PREFIX "utils.cuh",
+};
+
+#undef INCLUDE_PREFIX
+
+constexpr const char* g_headers[] = {
+    jit_source_atomic_ops,
+    jit_source_binary_ops,
+    jit_source_cast_op,
+    jit_source_config,
+    jit_source_complex,
+    jit_source_fp16_math,
+    jit_source_indexing,
+    jit_source_scatter_ops,
+    jit_source_unary_ops,
+    jit_source_ternary_ops,
+    jit_source_utils,
+};
+
+} // namespace
+
+JitModule::JitModule(
+    Device& device,
+    const std::string& module_name,
+    const KernelBuilder& builder) {
+  // Check cache.
+  std::vector<char> ptx;
+  std::vector<std::pair<std::string, std::string>> ptx_kernels;
+  if (!read_cached_ptx(ptx_cache_dir(), module_name, &ptx, &ptx_kernels)) {
+    // Create program.
+    auto [source_code, kernel_names] = builder();
+    nvrtcProgram prog;
+    CHECK_NVRTC_ERROR(nvrtcCreateProgram(
+        &prog,
+        source_code.c_str(),
+        (module_name + ".cu").c_str(),
+        std::size(g_headers),
+        g_headers,
+        g_include_names));
+    std::unique_ptr<nvrtcProgram, void (*)(nvrtcProgram*)> prog_freer(
+        &prog,
+        [](nvrtcProgram* p) { CHECK_NVRTC_ERROR(nvrtcDestroyProgram(p)); });
+    for (const auto& name : kernel_names) {
+      CHECK_NVRTC_ERROR(nvrtcAddNameExpression(prog, name.c_str()));
+    }
+
+    // Compile program.
+    std::vector<const char*> args;
+    bool use_sass = compiler_supports_device_sass(device);
+    std::string compute = fmt::format(
+        "--gpu-architecture={}_{}{}",
+        use_sass ? "sm" : "compute",
+        device.compute_capability_major(),
+        device.compute_capability_minor());
+    args.push_back(compute.c_str());
+    std::string cccl_include = cccl_dir();
+    if (!cccl_include.empty()) {
+      cccl_include = fmt::format("--include-path={}", cccl_include);
+      args.push_back(cccl_include.c_str());
+    }
+    std::string cuda_include =
+        fmt::format("--include-path={}/include", cuda_home());
+    args.push_back(cuda_include.c_str());
+    nvrtcResult compile_result =
+        nvrtcCompileProgram(prog, args.size(), args.data());
+    if (compile_result != NVRTC_SUCCESS) {
+      size_t log_size;
+      CHECK_NVRTC_ERROR(nvrtcGetProgramLogSize(prog, &log_size));
+      std::vector<char> log(log_size + 1, 0);
+      CHECK_NVRTC_ERROR(nvrtcGetProgramLog(prog, log.data()));
+      throw std::runtime_error(
+          fmt::format("Failed to compile kernel: {}.", log.data()));
+    }
+
+    // Get mangled names of kernel names.
+    for (const auto& name : kernel_names) {
+      const char* mangled;
+      CHECK_NVRTC_ERROR(nvrtcGetLoweredName(prog, name.c_str(), &mangled));
+      ptx_kernels.emplace_back(name, mangled);
+    }
+
+    // Get ptx data.
+    size_t ptx_size;
+    if (use_sass) {
+      CHECK_NVRTC_ERROR(nvrtcGetCUBINSize(prog, &ptx_size));
+    } else {
+      CHECK_NVRTC_ERROR(nvrtcGetPTXSize(prog, &ptx_size));
+    }
+    ptx.resize(ptx_size, 0);
+    if (use_sass) {
+      CHECK_NVRTC_ERROR(nvrtcGetCUBIN(prog, ptx.data()));
+    } else {
+      CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
+    }
+    write_cached_ptx(
+        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
+  }
+
+  // Load module.
+  char jit_log[4089] = {};
+  CUjit_option options[] = {
+      CU_JIT_ERROR_LOG_BUFFER, CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES};
+  void* values[] = {jit_log, reinterpret_cast<void*>(std::size(jit_log) - 1)};
+  CUresult jit_result = cuModuleLoadDataEx(
+      &module_, ptx.data(), std::size(options), options, values);
+  if (jit_result != CUDA_SUCCESS) {
+    throw std::runtime_error(fmt::format(
+        "Failed to load compiled {} kernel: {}.", module_name, jit_log));
+  }
+
+  // Load kernels.
+  for (const auto& [name, mangled] : ptx_kernels) {
+    CUfunction kernel;
+    CHECK_CUDA_ERROR(cuModuleGetFunction(&kernel, module_, mangled.c_str()));
+    kernels_[name] = kernel;
+  }
+}
+
+JitModule::~JitModule() {
+  CHECK_CUDA_ERROR(cuModuleUnload(module_));
+}
+
+CUfunction JitModule::get_kernel(const std::string& kernel_name) {
+  auto it = kernels_.find(kernel_name);
+  if (it == kernels_.end()) {
+    throw std::runtime_error(
+        fmt::format("There is no kernel named {}.", kernel_name));
+  }
+  return it->second;
+}
+
+JitModule& get_jit_module(
+    const mlx::core::Device& device,
+    const std::string& name,
+    const KernelBuilder& builder) {
+  static std::unordered_map<std::string, JitModule> map;
+  auto it = map.find(name);
+  if (it == map.end()) {
+    it = map.try_emplace(name, cu::device(device), name, builder).first;
+  }
+  return it->second;
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/jit_module.h

@@ -0,0 +1,107 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/config.h"
+
+#include <deque>
+#include <unordered_map>
+#include <utility>
+#include <variant>
+
+#include <cuda.h>
+#include <fmt/format.h>
+
+namespace mlx::core::cu {
+
+class Device;
+
+using KernelBuilderResult = std::pair<
+    /* source code */ std::string,
+    /* kernel names */ std::vector<std::string>>;
+using KernelBuilder = std::function<KernelBuilderResult()>;
+
+struct KernelArgs {
+  void** args() {
+    return args_.data();
+  }
+
+  void append(const array& a) {
+    append(reinterpret_cast<CUdeviceptr>(a.data<void>()));
+  }
+
+  template <typename T>
+  void append(T val) {
+    storage_.emplace_back(val);
+    append_ptr(&storage_.back());
+  }
+
+  template <typename T>
+  void append(std::vector<T> vec) {
+    if (vec.empty()) {
+      // The nullptr can not be used as arg, pass something not null.
+      append(std::monostate{});
+    } else {
+      append_ptr(vec.data());
+      storage_.emplace_back(std::move(vec));
+    }
+  }
+
+  // Make sure the arg is copied to an array with size of NDIM.
+  template <size_t NDIM = MAX_NDIM, typename T>
+  void append_ndim(std::vector<T> vec) {
+    if (vec.size() > NDIM) {
+      throw std::runtime_error(
+          fmt::format("ndim can not be larger than {}.", NDIM));
+    }
+    vec.resize(NDIM);
+    append(std::move(vec));
+  }
+
+  void append_ptr(const void* v) {
+    args_.push_back(const_cast<void*>(v));
+  }
+
+ private:
+  std::vector<void*> args_;
+
+  // The cuLaunchKernel API requires passing pointers to arguments so store
+  // temporary values untill kernel is launched.
+  using Arg = std::variant<
+      std::monostate,
+      CUdeviceptr,
+      int32_t,
+      uint32_t,
+      int64_t,
+      std::vector<const void*>,
+      std::vector<int32_t>,
+      std::vector<int64_t>>;
+  std::deque<Arg> storage_;
+};
+
+class JitModule {
+ public:
+  JitModule(
+      Device& device,
+      const std::string& module_name,
+      const KernelBuilder& builder);
+  ~JitModule();
+
+  JitModule(const JitModule&) = delete;
+  JitModule& operator=(const JitModule&) = delete;
+  CUfunction get_kernel(const std::string& kernel_name);
+
+ private:
+  CUmodule module_{nullptr};
+  std::unordered_map<std::string, CUfunction> kernels_;
+};
+
+JitModule& get_jit_module(
+    const mlx::core::Device& device,
+    const std::string& name,
+    const KernelBuilder& builder);
+
+} // namespace mlx::core::cu

mlx/backend/cuda/kernel_utils.cu

@@ -23,4 +23,11 @@ dim3 get_2d_grid_dims(
   return dim3(std::get<0>(dims), std::get<1>(dims), std::get<2>(dims));
 }
 
+std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2) {
+  auto [grid, block] = get_grid_and_block_common(dim0, dim1, dim2);
+  auto [gx, gy, gz] = grid;
+  auto [bx, by, bz] = block;
+  return std::make_pair(dim3(gx, gy, gz), dim3(bx, by, bz));
+}
+
 } // namespace mlx::core

mlx/backend/cuda/kernel_utils.cuh

@@ -1,19 +1,65 @@
 // Copyright © 2025 Apple Inc.
 
 // This file includes host-only utilies for writing CUDA kernels, the difference
-// from backend/cuda/kernels/utils.cuh is that the latter file only include
+// from backend/cuda/device/utils.cuh is that the latter file only include
 // device-only code.
 
 #pragma once
 
-#include "mlx/array.h"
+#include <type_traits>
 
-#include <cuComplex.h>
+#include "mlx/array.h"
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cuda.h>
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
+#include <fmt/format.h>
+#include <cuda/cmath>
 
 namespace mlx::core {
 
+template <typename F>
+void dispatch_1_2_3(int n, F&& f) {
+  switch (n) {
+    case 1:
+      f(std::integral_constant<int, 1>{});
+      break;
+    case 2:
+      f(std::integral_constant<int, 2>{});
+      break;
+    case 3:
+      f(std::integral_constant<int, 3>{});
+      break;
+  }
+}
+
+template <typename F>
+void dispatch_bool(bool v, F&& f) {
+  if (v) {
+    f(std::true_type{});
+  } else {
+    f(std::false_type{});
+  }
+}
+
+template <typename F>
+void dispatch_block_dim(int threads, F&& f) {
+  if (threads <= WARP_SIZE) {
+    f(std::integral_constant<int, WARP_SIZE>{});
+  } else if (threads <= WARP_SIZE * 2) {
+    f(std::integral_constant<int, WARP_SIZE * 2>{});
+  } else if (threads <= WARP_SIZE * 4) {
+    f(std::integral_constant<int, WARP_SIZE * 4>{});
+  } else if (threads <= WARP_SIZE * 8) {
+    f(std::integral_constant<int, WARP_SIZE * 8>{});
+  } else if (threads <= WARP_SIZE * 16) {
+    f(std::integral_constant<int, WARP_SIZE * 16>{});
+  } else {
+    f(std::integral_constant<int, WARP_SIZE * 32>{});
+  }
+}
+
 // Maps CPU types to CUDA types.
 template <typename T>
 struct CTypeToCudaType {
@@ -32,12 +78,39 @@ struct CTypeToCudaType<bfloat16_t> {
 
 template <>
 struct CTypeToCudaType<complex64_t> {
-  using type = cuComplex;
+  using type = cu::complex64_t;
 };
 
 template <typename T>
 using cuda_type_t = typename CTypeToCudaType<T>::type;
 
+// Type traits for detecting floating numbers.
+template <typename T>
+inline constexpr bool is_floating_v =
+    cuda::std::is_same_v<T, float> || cuda::std::is_same_v<T, double> ||
+    cuda::std::is_same_v<T, float16_t> || cuda::std::is_same_v<T, bfloat16_t>;
+
+// Type traits for detecting complex numbers.
+template <typename T>
+inline constexpr bool is_complex_v = cuda::std::is_same_v<T, complex64_t> ||
+    cuda::std::is_same_v<T, complex128_t>;
+
+// Type traits for detecting complex or real floating point numbers.
+template <typename T>
+inline constexpr bool is_inexact_v = is_floating_v<T> || is_complex_v<T>;
+
+// Utility to copy data from vector to array in host.
+template <int NDIM = MAX_NDIM, typename T = int32_t>
+inline cuda::std::array<T, NDIM> const_param(const std::vector<T>& vec) {
+  if (vec.size() > NDIM) {
+    throw std::runtime_error(
+        fmt::format("ndim can not be larger than {}.", NDIM));
+  }
+  cuda::std::array<T, NDIM> result;
+  std::copy_n(vec.begin(), vec.size(), result.begin());
+  return result;
+}
+
 // Compute the grid and block dimensions, check backend/common/utils.h for docs.
 dim3 get_block_dims(int dim0, int dim1, int dim2, int pow2 = 10);
 dim3 get_2d_grid_dims(const Shape& shape, const Strides& strides);
@@ -45,5 +118,55 @@ dim3 get_2d_grid_dims(
     const Shape& shape,
     const Strides& strides,
     size_t divisor);
+std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2);
+
+// Return a block size that achieves maximum potential occupancy for kernel.
+template <typename T>
+inline uint max_occupancy_block_dim(T kernel) {
+  int _, block_dim;
+  if constexpr (std::is_same_v<T, CUfunction>) {
+    CHECK_CUDA_ERROR(
+        cuOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel, 0, 0, 0));
+  } else {
+    CHECK_CUDA_ERROR(
+        cudaOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel));
+  }
+  return block_dim;
+}
+
+// Get the num_blocks and block_dims that maximize occupancy for |kernel|,
+// assuming each thread handles |work_per_thread| elements of |arr|.
+template <typename T>
+inline std::tuple<dim3, uint> get_launch_args(
+    T kernel,
+    size_t size,
+    const Shape& shape,
+    const Strides& strides,
+    bool large,
+    int work_per_thread = 1) {
+  size_t nthreads = cuda::ceil_div(size, work_per_thread);
+  uint block_dim = max_occupancy_block_dim(kernel);
+  if (block_dim > nthreads) {
+    block_dim = nthreads;
+  }
+  dim3 num_blocks;
+  if (large) {
+    num_blocks = get_2d_grid_dims(shape, strides, work_per_thread);
+    num_blocks.x = cuda::ceil_div(num_blocks.x, block_dim);
+  } else {
+    num_blocks.x = cuda::ceil_div(nthreads, block_dim);
+  }
+  return std::make_tuple(num_blocks, block_dim);
+}
+
+template <typename T>
+inline std::tuple<dim3, uint> get_launch_args(
+    T kernel,
+    const array& arr,
+    bool large,
+    int work_per_thread = 1) {
+  return get_launch_args(
+      kernel, arr.size(), arr.shape(), arr.strides(), large, work_per_thread);
+}
 
 } // namespace mlx::core

mlx/backend/cuda/kernels/fp16_math.cuh

@@ -1,76 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-#include <cuda_bf16.h>
-#include <cuda_fp16.h>
-#include <cuda/std/limits>
-#include <cuda/std/type_traits>
-
-namespace mlx::core::cu {
-
-///////////////////////////////////////////////////////////////////////////////
-// Additional C++ operator overrides between half types and native types.
-///////////////////////////////////////////////////////////////////////////////
-
-template <typename T, typename U>
-constexpr bool is_integral_except =
-    cuda::std::is_integral_v<T> && !cuda::std::is_same_v<T, U>;
-
-template <typename T, typename U>
-constexpr bool is_arithmetic_except =
-    cuda::std::is_arithmetic_v<T> && !cuda::std::is_same_v<T, U>;
-
-#define MLX_DEFINE_HALF_OP(HALF, HALF2FLOAT, FLOAT2HALF, OP)          \
-  template <                                                          \
-      typename T,                                                     \
-      typename = cuda::std::enable_if_t<is_integral_except<T, HALF>>> \
-  __forceinline__ __device__ HALF operator OP(HALF x, T y) {          \
-    return FLOAT2HALF(HALF2FLOAT(x) OP static_cast<float>(y));        \
-  }                                                                   \
-  template <                                                          \
-      typename T,                                                     \
-      typename = cuda::std::enable_if_t<is_integral_except<T, HALF>>> \
-  __forceinline__ __device__ HALF operator OP(T x, HALF y) {          \
-    return FLOAT2HALF(static_cast<float>(x) OP HALF2FLOAT(y));        \
-  }
-
-#define MLX_DEFINE_HALF_CMP(HALF, HALF2FLOAT, OP)                       \
-  template <                                                            \
-      typename T,                                                       \
-      typename = cuda::std::enable_if_t<is_arithmetic_except<T, HALF>>> \
-  __forceinline__ __device__ bool operator OP(HALF x, T y) {            \
-    return HALF2FLOAT(x) OP static_cast<float>(y);                      \
-  }                                                                     \
-  template <                                                            \
-      typename T,                                                       \
-      typename = cuda::std::enable_if_t<is_arithmetic_except<T, HALF>>> \
-  __forceinline__ __device__ bool operator OP(T x, HALF y) {            \
-    return static_cast<float>(y) OP HALF2FLOAT(x);                      \
-  }
-
-MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, +)
-MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, -)
-MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, *)
-MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, /)
-MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, +)
-MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, -)
-MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, *)
-MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, /)
-MLX_DEFINE_HALF_CMP(__half, __half2float, <)
-MLX_DEFINE_HALF_CMP(__half, __half2float, >)
-MLX_DEFINE_HALF_CMP(__half, __half2float, <=)
-MLX_DEFINE_HALF_CMP(__half, __half2float, >=)
-MLX_DEFINE_HALF_CMP(__half, __half2float, ==)
-MLX_DEFINE_HALF_CMP(__half, __half2float, !=)
-MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, <)
-MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, >)
-MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, <=)
-MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, >=)
-MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, ==)
-MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, !=)
-
-#undef MLX_DEFINE_HALF_OP
-#undef MLX_DEFINE_HALF_CMP
-
-} // namespace mlx::core::cu

mlx/backend/cuda/layer_norm.cu

@@ -0,0 +1,404 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/reduce/reduce.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/fast_primitives.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#include <nvtx3/nvtx3.hpp>
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_reduce.cuh>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+inline __device__ float3 plus_f3(const float3& a, const float3& b) {
+  return {a.x + b.x, a.y + b.y, a.z + b.z};
+}
+
+// Similar to cub::BlockReduce, but result is broadcasted to every thread.
+template <typename T, int BLOCK_DIM>
+struct BlockBroadcastReduce {
+  static_assert(WARP_SIZE <= BLOCK_DIM && BLOCK_DIM <= WARP_SIZE * WARP_SIZE);
+  static_assert(BLOCK_DIM % WARP_SIZE == 0);
+  using TempStorage = T[BLOCK_DIM / WARP_SIZE];
+
+  cg::thread_block& block;
+  TempStorage& temp;
+
+  template <typename Op>
+  __device__ T Reduce(const T& input, const Op& op, const T& init_value) {
+    auto warp = cg::tiled_partition<WARP_SIZE>(block);
+    T x = cg::reduce(warp, input, op);
+    if (warp.thread_rank() == 0) {
+      temp[warp.meta_group_rank()] = x;
+    }
+    block.sync();
+    x = warp.thread_rank() < warp.meta_group_size() ? temp[warp.thread_rank()]
+                                                    : init_value;
+    return cg::reduce(warp, x, op);
+  }
+
+  __device__ T Sum(const T& input) {
+    return Reduce(input, cg::plus<T>{}, T{});
+  }
+};
+
+template <typename T, int BLOCK_DIM, int N_READS = 4>
+__global__ void layer_norm(
+    const T* x,
+    const T* w,
+    const T* b,
+    T* out,
+    float eps,
+    int32_t axis_size,
+    int64_t w_stride,
+    int64_t b_stride) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+
+  using BlockReduceT = BlockBroadcastReduce<float, BLOCK_DIM>;
+  __shared__ typename BlockReduceT::TempStorage temp;
+
+  x += grid.block_rank() * axis_size;
+  out += grid.block_rank() * axis_size;
+
+  // Sum.
+  float sum = 0;
+  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    auto index = r * BLOCK_DIM + block.thread_rank();
+    T xn[N_READS] = {};
+    cub::LoadDirectBlocked(index, x, xn, axis_size);
+    sum += static_cast<float>(cub::ThreadReduce(xn, cuda::std::plus<>{}));
+  }
+  sum = BlockReduceT{block, temp}.Sum(sum);
+
+  // Mean.
+  float mean = sum / axis_size;
+
+  // Normalizer.
+  float normalizer = 0;
+  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    auto index = r * BLOCK_DIM + block.thread_rank();
+    T xn[N_READS];
+    cub::LoadDirectBlocked(index, x, xn, axis_size, mean);
+    for (int i = 0; i < N_READS; ++i) {
+      float t = static_cast<float>(xn[i]) - mean;
+      normalizer += t * t;
+    }
+  }
+  normalizer = BlockReduceT{block, temp}.Sum(normalizer);
+  normalizer = rsqrt(normalizer / axis_size + eps);
+
+  // Outputs.
+  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    auto index = r * BLOCK_DIM + block.thread_rank();
+    T xn[N_READS];
+    T wn[N_READS];
+    T bn[N_READS];
+    cub::LoadDirectBlocked(index, x, xn, axis_size);
+    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    cub::LoadDirectBlocked(index, strided_iterator(b, b_stride), bn, axis_size);
+    for (int i = 0; i < N_READS; ++i) {
+      float norm = (static_cast<float>(xn[i]) - mean) * normalizer;
+      xn[i] = wn[i] * static_cast<T>(norm) + bn[i];
+    }
+    cub::StoreDirectBlocked(index, out, xn, axis_size);
+  }
+}
+
+template <typename T, bool HAS_W, int BLOCK_DIM, int N_READS = 4>
+__global__ void layer_norm_vjp(
+    const T* x,
+    const T* w,
+    const T* g,
+    T* gx,
+    T* gw,
+    float eps,
+    int32_t axis_size,
+    int64_t w_stride) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+
+  using BlockReduceF = BlockBroadcastReduce<float, BLOCK_DIM>;
+  using BlockReduceF3 = BlockBroadcastReduce<float3, BLOCK_DIM>;
+  __shared__ union {
+    typename BlockReduceF::TempStorage f;
+    typename BlockReduceF3::TempStorage f3;
+  } temp;
+
+  x += grid.block_rank() * axis_size;
+  g += grid.block_rank() * axis_size;
+  gx += grid.block_rank() * axis_size;
+  gw += grid.block_rank() * axis_size;
+
+  // Sum.
+  float sum = 0;
+  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    auto index = r * BLOCK_DIM + block.thread_rank();
+    T xn[N_READS] = {};
+    cub::LoadDirectBlocked(index, x, xn, axis_size);
+    sum += static_cast<float>(cub::ThreadReduce(xn, cuda::std::plus<>{}));
+  }
+  sum = BlockReduceF{block, temp.f}.Sum(sum);
+
+  // Mean.
+  float mean = sum / axis_size;
+
+  // Normalizer.
+  float3 factors = {};
+  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    T xn[N_READS];
+    T wn[N_READS] = {};
+    T gn[N_READS] = {};
+    auto index = r * BLOCK_DIM + block.thread_rank();
+    cub::LoadDirectBlocked(index, x, xn, axis_size, mean);
+    cub::LoadDirectBlocked(index, g, gn, axis_size);
+    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    for (int i = 0; i < N_READS; i++) {
+      float t = static_cast<float>(xn[i]) - mean;
+      float wi = wn[i];
+      float gi = gn[i];
+      float wg = wi * gi;
+      factors = plus_f3(factors, {wg, wg * t, t * t});
+    }
+  }
+  factors = BlockReduceF3{block, temp.f3}.Reduce(factors, plus_f3, {});
+  float meanwg = factors.x / axis_size;
+  float meanwgxc = factors.y / axis_size;
+  float normalizer2 = 1 / (factors.z / axis_size + eps);
+  float normalizer = sqrt(normalizer2);
+
+  // Outputs.
+  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    auto index = r * BLOCK_DIM + block.thread_rank();
+    T xn[N_READS];
+    T wn[N_READS];
+    T gn[N_READS];
+    cub::LoadDirectBlocked(index, x, xn, axis_size);
+    cub::LoadDirectBlocked(index, g, gn, axis_size);
+    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    for (int i = 0; i < N_READS; i++) {
+      float xi = (static_cast<float>(xn[i]) - mean) * normalizer;
+      float wi = wn[i];
+      float gi = gn[i];
+      xn[i] = normalizer * (wi * gi - meanwg) - xi * meanwgxc * normalizer2;
+      if constexpr (HAS_W) {
+        wn[i] = gi * xi;
+      }
+    }
+    cub::StoreDirectBlocked(index, gx, xn, axis_size);
+    if constexpr (HAS_W) {
+      cub::StoreDirectBlocked(index, gw, wn, axis_size);
+    }
+  }
+}
+
+} // namespace cu
+
+namespace fast {
+
+bool LayerNorm::use_fallback(Stream s) {
+  return s.device == Device::cpu;
+}
+
+// TODO: There are duplicate code with backend/metal/normalization.cpp
+void LayerNorm::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("LayerNorm::eval_gpu");
+  auto& s = stream();
+  auto& out = outputs[0];
+
+  // Make sure that the last dimension is contiguous.
+  auto set_output = [&s, &out](const array& x) {
+    bool no_copy = x.flags().contiguous && x.strides()[x.ndim() - 1] == 1;
+    if (no_copy && x.ndim() > 1) {
+      auto s = x.strides()[x.ndim() - 2];
+      no_copy &= (s == 0 || s == x.shape().back());
+    }
+    if (no_copy) {
+      if (x.is_donatable()) {
+        out.copy_shared_buffer(x);
+      } else {
+        out.set_data(
+            allocator::malloc(x.data_size() * x.itemsize()),
+            x.data_size(),
+            x.strides(),
+            x.flags());
+      }
+      return x;
+    } else {
+      array x_copy = contiguous_copy_gpu(x, s);
+      out.copy_shared_buffer(x_copy);
+      return x_copy;
+    }
+  };
+
+  const array x = set_output(inputs[0]);
+  const array& w = inputs[1];
+  const array& b = inputs[2];
+
+  int32_t axis_size = x.shape().back();
+  int32_t n_rows = x.data_size() / axis_size;
+  int64_t w_stride = (w.ndim() == 1) ? w.strides()[0] : 0;
+  int64_t b_stride = (b.ndim() == 1) ? b.strides()[0] : 0;
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(x);
+  encoder.set_input_array(w);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  dispatch_float_types(out.dtype(), "layernorm", [&](auto type_tag) {
+    constexpr uint32_t N_READS = 4;
+    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+      auto kernel = cu::layer_norm<DataType, block_dim(), N_READS>;
+      encoder.add_kernel_node(
+          kernel,
+          n_rows,
+          block_dim(),
+          0,
+          x.data<DataType>(),
+          w.data<DataType>(),
+          b.data<DataType>(),
+          out.data<DataType>(),
+          eps_,
+          axis_size,
+          w_stride,
+          b_stride);
+    });
+  });
+}
+
+void LayerNormVJP::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("LayerNormVJP::eval_gpu");
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  // Ensure row contiguity. We could relax this step by checking that the array
+  // is contiguous (no broadcasts or holes) and that the input strides are the
+  // same as the cotangent strides but for now this is simpler.
+  auto check_input = [&s](const array& x, bool& copied) {
+    if (x.flags().row_contiguous) {
+      copied = false;
+      return x;
+    }
+    copied = true;
+    return contiguous_copy_gpu(x, s);
+  };
+  bool donate_x = inputs[0].is_donatable();
+  bool donate_g = inputs[3].is_donatable();
+  bool copied;
+  auto x = check_input(inputs[0], copied);
+  donate_x |= copied;
+  const array& w = inputs[1];
+  const array& b = inputs[2];
+  bool g_copied;
+  auto g = check_input(inputs[3], g_copied);
+  donate_g |= g_copied;
+  array& gx = outputs[0];
+  array& gw = outputs[1];
+  array& gb = outputs[2];
+
+  // Check whether we had a weight.
+  bool has_w = w.ndim() != 0;
+
+  // Allocate space for the outputs.
+  bool g_in_gx = false;
+  if (donate_x) {
+    gx.copy_shared_buffer(x);
+  } else if (donate_g) {
+    gx.copy_shared_buffer(g);
+    g_in_gx = true;
+  } else {
+    gx.set_data(allocator::malloc(gx.nbytes()));
+  }
+  if (g_copied && !g_in_gx) {
+    encoder.add_temporary(g);
+  }
+
+  int32_t axis_size = x.shape().back();
+  int32_t n_rows = x.data_size() / axis_size;
+  int64_t w_stride = (w.ndim() == 1) ? w.strides()[0] : 0;
+
+  // Allocate a temporary to store the gradients for w and allocate the output
+  // gradient accumulators.
+  array gw_temp =
+      (has_w) ? array({n_rows, x.shape().back()}, gw.dtype(), nullptr, {}) : w;
+  bool g_in_gw = false;
+  if (has_w) {
+    if (!g_in_gx && donate_g) {
+      g_in_gw = true;
+      gw_temp.copy_shared_buffer(g);
+    } else {
+      gw_temp.set_data(allocator::malloc(gw_temp.nbytes()));
+      encoder.add_temporary(gw_temp);
+    }
+  }
+
+  // The gradient for b in case we had a b.
+  bool has_gb = (gb.ndim() == 1 && gb.size() == axis_size);
+  if (has_gb) {
+    ReductionPlan plan(
+        ReductionOpType::ContiguousStridedReduce, {n_rows}, {axis_size});
+    col_reduce(encoder, g, gb, Reduce::ReduceType::Sum, {0}, plan);
+  }
+
+  // Insert dependency if `g` was donated
+  if ((g_in_gx || g_in_gw) && has_gb) {
+    encoder.set_input_array(gb);
+  }
+  encoder.set_input_array(x);
+  encoder.set_input_array(w);
+  encoder.set_input_array(g);
+  encoder.set_output_array(gx);
+  encoder.set_output_array(gw_temp);
+  dispatch_float_types(gx.dtype(), "layernorm_vjp", [&](auto type_tag) {
+    dispatch_bool(has_w, [&](auto has_w_constant) {
+      constexpr int N_READS = 4;
+      dispatch_block_dim(
+          cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+            using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+            auto kernel = cu::layer_norm_vjp<
+                DataType,
+                has_w_constant.value,
+                block_dim(),
+                N_READS>;
+            encoder.add_kernel_node(
+                kernel,
+                n_rows,
+                block_dim(),
+                0,
+                x.data<DataType>(),
+                w.data<DataType>(),
+                g.data<DataType>(),
+                gx.data<DataType>(),
+                gw_temp.data<DataType>(),
+                eps_,
+                axis_size,
+                w_stride);
+          });
+    });
+  });
+
+  if (has_w) {
+    ReductionPlan plan(
+        ReductionOpType::ContiguousStridedReduce, {n_rows}, {axis_size});
+    col_reduce(encoder, gw_temp, gw, Reduce::ReduceType::Sum, {0}, plan);
+  }
+}
+
+} // namespace fast
+
+} // namespace mlx::core

mlx/backend/cuda/logsumexp.cu

@@ -0,0 +1,162 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/cast_op.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#include <nvtx3/nvtx3.hpp>
+#include <cub/block/block_load.cuh>
+
+#include <cassert>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T>
+inline __device__ T softmax_exp(T x) {
+  // Softmax doesn't need high precision exponential cause x is gonna be in
+  // (-oo, 0] anyway and subsequently it will be divided by sum(exp(x_i)).
+  return __expf(x);
+}
+
+template <typename T, typename AccT, int BLOCK_DIM, int N_READS = 4>
+__global__ void logsumexp(const T* in, T* out, int axis_size) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  in += grid.block_rank() * axis_size;
+
+  cg::greater<AccT> max_op;
+  cg::plus<AccT> plus_op;
+
+  // Thread reduce.
+  AccT prevmax;
+  AccT maxval = Limits<AccT>::finite_min();
+  AccT normalizer = 0;
+  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
+    AccT vals[N_READS];
+    cub::LoadDirectBlocked(
+        r * BLOCK_DIM + block.thread_rank(),
+        make_cast_iterator<AccT>(in),
+        vals,
+        axis_size,
+        Limits<AccT>::min());
+    prevmax = maxval;
+    maxval = max_op(maxval, cub::ThreadReduce(vals, max_op));
+    // Online normalizer calculation for softmax:
+    // https://github.com/NVIDIA/online-softmax
+    normalizer = normalizer * softmax_exp(prevmax - maxval);
+    for (int i = 0; i < N_READS; i++) {
+      normalizer = normalizer + softmax_exp(vals[i] - maxval);
+    }
+  }
+
+  // First warp reduce.
+  prevmax = maxval;
+  maxval = cg::reduce(warp, maxval, max_op);
+  normalizer = normalizer * softmax_exp(prevmax - maxval);
+  normalizer = cg::reduce(warp, normalizer, plus_op);
+
+  __shared__ AccT local_max[WARP_SIZE];
+  __shared__ AccT local_normalizer[WARP_SIZE];
+
+  // Write to shared memory and do second warp reduce.
+  prevmax = maxval;
+  if (warp.thread_rank() == 0) {
+    local_max[warp.meta_group_rank()] = maxval;
+  }
+  block.sync();
+  maxval = warp.thread_rank() < warp.meta_group_size()
+      ? local_max[warp.thread_rank()]
+      : Limits<AccT>::finite_min();
+  maxval = cg::reduce(warp, maxval, max_op);
+  normalizer = normalizer * softmax_exp(prevmax - maxval);
+  if (warp.thread_rank() == 0) {
+    local_normalizer[warp.meta_group_rank()] = normalizer;
+  }
+  block.sync();
+  normalizer = warp.thread_rank() < warp.meta_group_size()
+      ? local_normalizer[warp.thread_rank()]
+      : AccT{};
+  normalizer = cg::reduce(warp, normalizer, plus_op);
+
+  // Write output.
+  if (block.thread_rank() == 0) {
+    out[grid.block_rank()] = isinf(maxval) ? maxval : log(normalizer) + maxval;
+  }
+}
+
+} // namespace cu
+
+void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("LogSumExp::eval_gpu");
+  assert(inputs.size() == 1);
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  // Make sure that the last dimension is contiguous.
+  auto ensure_contiguous = [&s, &encoder](const array& x) {
+    if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
+      return x;
+    } else {
+      array x_copy = contiguous_copy_gpu(x, s);
+      encoder.add_temporary(x_copy);
+      return x_copy;
+    }
+  };
+
+  auto in = ensure_contiguous(inputs[0]);
+  if (in.flags().row_contiguous) {
+    out.set_data(allocator::malloc(out.nbytes()));
+  } else {
+    auto n = in.shape(-1);
+    auto flags = in.flags();
+    auto strides = in.strides();
+    for (auto& s : strides) {
+      s /= n;
+    }
+    bool col_contig = strides[0] == 1;
+    for (int i = 1; col_contig && i < strides.size(); ++i) {
+      col_contig &=
+          (out.shape(i) == 1 || strides[i - 1] == out.shape(i) * strides[i]);
+    }
+    flags.col_contiguous = col_contig;
+    out.set_data(
+        allocator::malloc(in.nbytes() / n),
+        in.data_size() / n,
+        std::move(strides),
+        flags);
+  }
+
+  int axis_size = in.shape().back();
+  int n_rows = in.data_size() / axis_size;
+
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  dispatch_float_types(out.dtype(), "logsumexp", [&](auto type_tag) {
+    constexpr int N_READS = 4;
+    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+      auto kernel = cu::logsumexp<DataType, float, block_dim(), N_READS>;
+      encoder.add_kernel_node(
+          kernel,
+          n_rows,
+          block_dim(),
+          0,
+          in.data<DataType>(),
+          out.data<DataType>(),
+          axis_size);
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/matmul.cpp

@@ -0,0 +1,516 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/matmul.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+#include "mlx/utils.h"
+
+#include <cublasLt.h>
+#include <fmt/format.h>
+#include <nvtx3/nvtx3.hpp>
+
+#include <numeric>
+
+namespace mlx::core {
+
+namespace cu {
+
+#define CHECK_CUBLAS_ERROR(cmd) check_cublas_error(#cmd, (cmd))
+
+void check_cublas_error(const char* name, cublasStatus_t err) {
+  if (err != CUBLAS_STATUS_SUCCESS) {
+    // TODO: Use cublasGetStatusString when it is widely available.
+    throw std::runtime_error(
+        fmt::format("{} failed with code: {}.", name, static_cast<int>(err)));
+  }
+}
+
+struct CublasPreference {
+  CublasPreference(Device& device) {
+    // The recommended cublas workspace size is 4 MiB for pre-Hopper and 32 MiB
+    // for Hopper+:
+    // https://docs.nvidia.com/cuda/cublas/#cublassetworkspace
+    uint64_t MiB = 1024 * 1024;
+    uint64_t workspace_size =
+        device.compute_capability_major() >= 9 ? 32 * MiB : 4 * MiB;
+
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceCreate(&pref_));
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceSetAttribute(
+        pref_,
+        CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
+        &workspace_size,
+        sizeof(uint64_t)));
+  }
+
+  ~CublasPreference() {
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceDestroy(pref_));
+  }
+
+  cublasLtMatmulPreference_t pref_{nullptr};
+};
+
+cublasLtMatmulPreference_t cublas_preference(Device& device) {
+  static CublasPreference pref(device);
+  return pref.pref_;
+}
+
+class MatMul {
+ public:
+  MatMul(
+      Device& device,
+      Dtype dtype,
+      bool a_transposed,
+      uint64_t a_rows,
+      uint64_t a_cols,
+      int64_t lda,
+      bool b_transposed,
+      uint64_t b_rows,
+      uint64_t b_cols,
+      int64_t ldb,
+      int32_t batch_count,
+      int64_t a_batch_stride,
+      int64_t b_batch_stride)
+      : handle_(device.lt_handle()), pref_(cublas_preference(device)) {
+    heuristic_.state = CUBLAS_STATUS_NOT_INITIALIZED;
+
+    auto scale_type = dtype_to_cuda_type(dtype);
+    if (dtype == bfloat16 || dtype == float16) {
+      scale_type = CUDA_R_32F;
+    }
+    CHECK_CUBLAS_ERROR(cublasLtMatmulDescCreate(
+        &matmul_desc_, dtype_to_compute_type(dtype), scale_type));
+    int32_t pointer_mode = CUBLASLT_POINTER_MODE_HOST;
+    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+        matmul_desc_,
+        CUBLASLT_MATMUL_DESC_POINTER_MODE,
+        &pointer_mode,
+        sizeof(int32_t)));
+    cublasOperation_t op = CUBLAS_OP_N;
+    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+        matmul_desc_,
+        CUBLASLT_MATMUL_DESC_TRANSA,
+        &op,
+        sizeof(cublasOperation_t)));
+    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+        matmul_desc_,
+        CUBLASLT_MATMUL_DESC_TRANSB,
+        &op,
+        sizeof(cublasOperation_t)));
+
+    auto type = dtype_to_cuda_type(dtype);
+    a_desc_ = create_matrix_layout(
+        type, a_rows, a_cols, a_transposed, lda, batch_count, a_batch_stride);
+    b_desc_ = create_matrix_layout(
+        type, b_rows, b_cols, b_transposed, ldb, batch_count, b_batch_stride);
+    out_desc_ = create_matrix_layout(
+        type, a_rows, b_cols, false, b_cols, batch_count, a_rows * b_cols);
+  }
+
+  MatMul(
+      Device& device,
+      Dtype dtype,
+      bool a_transposed,
+      uint64_t a_rows,
+      uint64_t a_cols,
+      int64_t lda,
+      bool b_transposed,
+      uint64_t b_rows,
+      uint64_t b_cols,
+      int64_t ldb,
+      int64_t ldc,
+      int32_t batch_count,
+      int64_t a_batch_stride,
+      int64_t b_batch_stride,
+      int64_t c_batch_stride)
+      : MatMul(
+            device,
+            dtype,
+            a_transposed,
+            a_rows,
+            a_cols,
+            lda,
+            b_transposed,
+            b_rows,
+            b_cols,
+            ldb,
+            batch_count,
+            a_batch_stride,
+            b_batch_stride) {
+    auto type = dtype_to_cuda_type(dtype);
+    c_desc_ = create_matrix_layout(
+        type, a_rows, b_cols, false, ldc, batch_count, c_batch_stride);
+  }
+
+  ~MatMul() {
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(a_desc_));
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(b_desc_));
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(c_desc_));
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(out_desc_));
+    CHECK_CUBLAS_ERROR(cublasLtMatmulDescDestroy(matmul_desc_));
+  }
+
+  void run(
+      cu::CommandEncoder& encoder,
+      void* out,
+      void* a,
+      void* b,
+      void* c = nullptr,
+      float alpha = 1,
+      float beta = 0) {
+    if (heuristic_.state != CUBLAS_STATUS_SUCCESS) {
+      int ret = 0;
+      CHECK_CUBLAS_ERROR(cublasLtMatmulAlgoGetHeuristic(
+          handle_,
+          matmul_desc_,
+          a_desc_,
+          b_desc_,
+          out_desc_,
+          out_desc_,
+          pref_,
+          1,
+          &heuristic_,
+          &ret));
+      if (ret == 0) {
+        throw std::runtime_error("Can not find algorithm for matmul.");
+      }
+    }
+
+    void* workspace_ptr = nullptr;
+    if (heuristic_.workspaceSize > 0) {
+      array workspace(
+          allocator::malloc(heuristic_.workspaceSize),
+          {static_cast<int>(heuristic_.workspaceSize)},
+          int8);
+      encoder.add_temporary(workspace);
+      workspace_ptr = workspace.data<void>();
+    }
+
+    auto capture = encoder.capture_context();
+    CHECK_CUBLAS_ERROR(cublasLtMatmul(
+        handle_,
+        matmul_desc_,
+        &alpha,
+        a,
+        a_desc_,
+        b,
+        b_desc_,
+        &beta,
+        c ? c : out,
+        c ? c_desc_ : out_desc_,
+        out,
+        out_desc_,
+        &heuristic_.algo,
+        workspace_ptr,
+        heuristic_.workspaceSize,
+        encoder.stream()));
+  }
+
+ private:
+  cublasComputeType_t dtype_to_compute_type(Dtype dtype) {
+    switch (dtype) {
+      case float16:
+        return CUBLAS_COMPUTE_32F;
+      case bfloat16:
+        return CUBLAS_COMPUTE_32F;
+      case float32:
+        return mlx::core::env::enable_tf32() ? CUBLAS_COMPUTE_32F_FAST_TF32
+                                             : CUBLAS_COMPUTE_32F;
+      case float64:
+      case complex64:
+        return CUBLAS_COMPUTE_64F;
+      default:
+        throw std::runtime_error(fmt::format(
+            "Unsupported dtype in MatMul: {}.", dtype_to_string(dtype)));
+    }
+  }
+
+  cudaDataType_t dtype_to_cuda_type(Dtype dtype) {
+    switch (dtype) {
+      case float16:
+        return CUDA_R_16F;
+      case bfloat16:
+        return CUDA_R_16BF;
+      case float32:
+        return CUDA_R_32F;
+      case float64:
+        return CUDA_R_64F;
+      case complex64:
+        return CUDA_C_32F;
+      default:
+        throw std::runtime_error(fmt::format(
+            "Unsupported dtype in MatMul: {}.", dtype_to_string(dtype)));
+    }
+  }
+
+  cublasLtMatrixLayout_t create_matrix_layout(
+      cudaDataType_t type,
+      uint64_t rows,
+      uint64_t cols,
+      bool transposed,
+      int64_t ld,
+      int32_t batch_count,
+      int64_t batch_stride) {
+    cublasLtMatrixLayout_t desc;
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutCreate(&desc, type, rows, cols, ld));
+    cublasLtOrder_t order =
+        transposed ? CUBLASLT_ORDER_COL : CUBLASLT_ORDER_ROW;
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+        desc, CUBLASLT_MATRIX_LAYOUT_ORDER, &order, sizeof(cublasLtOrder_t)));
+    if (batch_count > 1) {
+      CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+          desc,
+          CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT,
+          &batch_count,
+          sizeof(int32_t)));
+      CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+          desc,
+          CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET,
+          &batch_stride,
+          sizeof(int64_t)));
+    }
+    return desc;
+  }
+
+  cublasLtMatmulPreference_t pref_{nullptr};
+  cublasLtHandle_t handle_{nullptr};
+  cublasLtMatmulDesc_t matmul_desc_{nullptr};
+  cublasLtMatrixLayout_t a_desc_{nullptr};
+  cublasLtMatrixLayout_t b_desc_{nullptr};
+  cublasLtMatrixLayout_t c_desc_{nullptr};
+  cublasLtMatrixLayout_t out_desc_{nullptr};
+  cublasLtMatmulHeuristicResult_t heuristic_;
+};
+
+} // namespace cu
+
+namespace {
+
+std::tuple<bool, int64_t, array>
+check_transpose(cu::CommandEncoder& enc, const Stream& s, const array& arr) {
+  auto stx = arr.strides()[arr.ndim() - 2];
+  auto sty = arr.strides()[arr.ndim() - 1];
+  if (sty == 1 && stx == arr.shape(-1)) {
+    return std::make_tuple(false, stx, arr);
+  } else if (stx == 1 && sty == arr.shape(-2)) {
+    return std::make_tuple(true, sty, arr);
+  } else {
+    array arr_copy = contiguous_copy_gpu(arr, s);
+    enc.add_temporary(arr_copy);
+    return std::make_tuple(false, arr.shape(-1), arr_copy);
+  }
+}
+
+} // namespace
+
+void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Matmul::eval_gpu");
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  assert(inputs.size() == 2);
+  auto& a_pre = inputs[0];
+  auto& b_pre = inputs[1];
+  // Return 0s if either input is empty.
+  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;
+  }
+
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Init checks and prep
+
+  int M = a_pre.shape(-2);
+  int N = b_pre.shape(-1);
+  int K = a_pre.shape(-1);
+
+  // Keep a vector with copies to be cleared in the completed buffer to release
+  // the arrays
+  auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
+  auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Check and collapse batch dimensions
+
+  auto [batch_shape, a_batch_strides, b_batch_strides] = collapse_batches(a, b);
+
+  auto batch_count = out.size() / (M * N);
+
+  // Collapse batches into M if needed
+  if (batch_count > 1 && !a_transposed && batch_shape.size() == 1 &&
+      a.strides()[a.ndim() - 2] == K && a_batch_strides.back() == M * K &&
+      b_batch_strides.back() == 0) {
+    M *= batch_shape.back();
+    batch_count = 1;
+
+    a_batch_strides = {0};
+    b_batch_strides = {0};
+    batch_shape = {1};
+  }
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Invoke cublasLt
+
+  cu::MatMul matmul(
+      cu::device(s.device),
+      a.dtype(),
+      a_transposed,
+      M,
+      K,
+      lda,
+      b_transposed,
+      K,
+      N,
+      ldb,
+      batch_shape.back(),
+      a_batch_strides.back(),
+      b_batch_strides.back());
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  auto nbatch = batch_count / batch_shape.back();
+  if (nbatch == 1) {
+    matmul.run(encoder, out.data<int8_t>(), a.data<int8_t>(), b.data<int8_t>());
+    return;
+  }
+
+  ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
+  ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
+  auto concurrent = encoder.concurrent_context();
+  for (size_t i = 0; i < nbatch; ++i) {
+    matmul.run(
+        encoder,
+        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M * N,
+        a.data<int8_t>() + a.itemsize() * a_it.loc,
+        b.data<int8_t>() + b.itemsize() * b_it.loc);
+    a_it.step();
+    b_it.step();
+  }
+}
+
+void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("AddMM::eval_gpu");
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  assert(inputs.size() == 3);
+  auto& a_pre = inputs[0];
+  auto& b_pre = inputs[1];
+  auto c = inputs[2];
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Init checks and prep
+
+  int M = a_pre.shape(-2);
+  int N = b_pre.shape(-1);
+  int K = a_pre.shape(-1);
+
+  // Keep a vector with copies to be cleared in the completed buffer to release
+  // the arrays
+  auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
+  auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
+
+  int64_t ldc;
+  {
+    auto stx = c.strides()[c.ndim() - 2];
+    auto sty = c.strides()[c.ndim() - 1];
+    if (sty == 1 && stx == c.shape(-1)) {
+      ldc = stx;
+      out.set_data(allocator::malloc(out.nbytes()));
+    } else if (sty == 1 && stx == 0) {
+      ldc = 0;
+      out.set_data(allocator::malloc(out.nbytes()));
+    } else {
+      // Copy C into out and set C to out
+      ldc = c.shape(-1);
+      copy_gpu(c, out, CopyType::General, s);
+      c = out;
+    }
+  }
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Check and collapse batch dimensions
+
+  auto [batch_shape, a_batch_strides, b_batch_strides, c_batch_strides] =
+      collapse_batches(a, b, c);
+
+  auto batch_count = out.size() / (M * N);
+
+  // Collapse batches into M if needed
+  if (batch_count > 1 && !a_transposed && batch_shape.size() == 1 &&
+      a.strides()[a.ndim() - 2] == K && a_batch_strides.back() == M * K &&
+      c_batch_strides.back() == M * c.strides()[c.ndim() - 2] &&
+      b_batch_strides.back() == 0) {
+    M *= batch_shape.back();
+    batch_count = 1;
+
+    a_batch_strides = {0};
+    b_batch_strides = {0};
+    c_batch_strides = {0};
+    batch_shape = {1};
+  }
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Invoke cublasLt
+
+  cu::MatMul matmul(
+      cu::device(s.device),
+      a.dtype(),
+      a_transposed,
+      M,
+      K,
+      lda,
+      b_transposed,
+      K,
+      N,
+      ldb,
+      ldc,
+      batch_shape.back(),
+      a_batch_strides.back(),
+      b_batch_strides.back(),
+      c_batch_strides.back());
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  auto nbatch = batch_count / batch_shape.back();
+  if (nbatch == 1) {
+    matmul.run(
+        encoder,
+        out.data<int8_t>(),
+        a.data<int8_t>(),
+        b.data<int8_t>(),
+        c.data<int8_t>(),
+        alpha_,
+        beta_);
+    return;
+  }
+
+  ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
+  ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
+  ContiguousIterator c_it(batch_shape, c_batch_strides, batch_shape.size() - 1);
+  auto concurrent = encoder.concurrent_context();
+  for (size_t i = 0; i < nbatch; ++i) {
+    matmul.run(
+        encoder,
+        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M * N,
+        a.data<int8_t>() + a.itemsize() * a_it.loc,
+        b.data<int8_t>() + b.itemsize() * b_it.loc,
+        c.data<int8_t>() + c.itemsize() * c_it.loc,
+        alpha_,
+        beta_);
+    a_it.step();
+    b_it.step();
+    c_it.step();
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cuda/matmul/mma.cuh

@@ -0,0 +1,108 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/matmul/tiles.cuh"
+
+namespace mlx::core::cu {
+
+template <typename U, typename T>
+__device__ inline void
+mma_t(Tile16x16<U>& C, Tile16x16<T>& A, Tile16x16<T>& B) {}
+
+/**
+ * Multiply the 16x16 bfloat16 tiles and accumulate the result in one 16x16
+ * float tile.
+ *
+ * We actually perform C += A @ B.T
+ */
+__device__ inline void mma_t(
+    Tile16x16<float>& C,
+    Tile16x16<__nv_bfloat16>& A,
+    Tile16x16<__nv_bfloat16>& B) {
+  asm volatile(
+      "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+      "{%0, %1, %2, %3}, "
+      "{%4, %5, %6, %7}, "
+      "{%8, %9}, "
+      "{%10, %11, %12, %13};"
+
+      // D matrix
+      : "+f"(C.values[0].x),
+        "+f"(C.values[0].y),
+        "+f"(C.values[1].x),
+        "+f"(C.values[1].y)
+
+      // A matrix
+      : "r"(*(uint32_t*)(&A.values[0])),
+        "r"(*(uint32_t*)(&A.values[1])),
+        "r"(*(uint32_t*)(&A.values[2])),
+        "r"(*(uint32_t*)(&A.values[3])),
+
+        // B matrix
+        "r"(*(uint32_t*)(&B.values[0])),
+        "r"(*(uint32_t*)(&B.values[2])),
+
+        // C matrix
+        "f"(C.values[0].x),
+        "f"(C.values[0].y),
+        "f"(C.values[1].x),
+        "f"(C.values[1].y));
+  asm volatile(
+      "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+      "{%0, %1, %2, %3}, "
+      "{%4, %5, %6, %7}, "
+      "{%8, %9}, "
+      "{%10, %11, %12, %13};"
+
+      // D matrix
+      : "+f"(C.values[2].x),
+        "+f"(C.values[2].y),
+        "+f"(C.values[3].x),
+        "+f"(C.values[3].y)
+
+      // A matrix
+      : "r"(*(uint32_t*)(&A.values[0])),
+        "r"(*(uint32_t*)(&A.values[1])),
+        "r"(*(uint32_t*)(&A.values[2])),
+        "r"(*(uint32_t*)(&A.values[3])),
+
+        // B matrix
+        "r"(*(uint32_t*)(&B.values[1])),
+        "r"(*(uint32_t*)(&B.values[3])),
+
+        // C matrix
+        "f"(C.values[2].x),
+        "f"(C.values[2].y),
+        "f"(C.values[3].x),
+        "f"(C.values[3].y));
+}
+
+/**
+ * Multiply larger register tiles by delegating to mma_t.
+ */
+template <typename U, typename T, int M, int N, int K>
+__device__ inline void mma_t(
+    RegisterTile<U, M, N>& C,
+    RegisterTile<T, M, K>& A,
+    RegisterTile<T, N, K>& B) {
+  constexpr int TILES_M = RegisterTile<T, M, K>::TILES_Y;
+  constexpr int TILES_K = RegisterTile<T, M, K>::TILES_X;
+  constexpr int TILES_N = RegisterTile<T, N, K>::TILES_Y;
+
+  MLX_UNROLL
+  for (int k = 0; k < TILES_K; k++) {
+    MLX_UNROLL
+    for (int m = 0; m < TILES_M; m++) {
+      MLX_UNROLL
+      for (int n = 0; n < TILES_N; n++) {
+        mma_t(
+            C.data[m * TILES_N + n],
+            A.data[m * TILES_K + k],
+            B.data[n * TILES_K + k]);
+      }
+    }
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/matmul/tiles.cuh

@@ -0,0 +1,419 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#define MLX_UNROLL _Pragma("unroll")
+
+namespace mlx::core::cu {
+
+// Map types to their vector of 2 type float -> float2, double -> double2 etc
+template <typename T>
+struct Vector2;
+template <>
+struct Vector2<double> {
+  using type = double2;
+};
+template <>
+struct Vector2<float> {
+  using type = float2;
+};
+template <>
+struct Vector2<__half> {
+  using type = __half2;
+};
+template <>
+struct Vector2<__nv_bfloat16> {
+  using type = __nv_bfloat162;
+};
+template <typename T>
+using Vector2_t = typename Vector2<T>::type;
+
+/**
+ * The basic building block for Ampere mmas. A 16x16 tile distributed across
+ * the warp.
+ *
+ * Each thread holds 8 values. They are distributed according to
+ * https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-fragment-mma-16816-float
+ *
+ * For use instructions see the individual methods eg load().
+ */
+template <typename T>
+struct Tile16x16 {
+  using T2 = Vector2_t<T>;
+
+  T2 values[4];
+
+  __device__ inline void fill(T v) {
+    T2 v2 = {v, v};
+    for (int i = 0; i < 4; i++) {
+      values[i] = v2;
+    }
+  }
+
+  /**
+   * Load a 16x16 tile from shared memory.
+   *
+   * The instruction is a bit weird in the sense that the address provided by
+   * each thread and the elements loaded are not the same.
+   *
+   * We load 4 8x8 tiles. The tile rows are stored contiguously in memory. As a
+   * result the warp provides 4*8 = 32 addresses one per row.
+   *
+   * Threads 0-7 provide the addresses for the first tile, 8-15 for the second
+   * and so on. For instance to load a non swizzled tile we would do
+   *
+   *    base_addr + (laneid % 16) * BK + (laneid / 2) * 8
+   *
+   * See
+   * https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-instructions-ldmatrix
+   */
+  __device__ inline void load(uint32_t row_address) {
+    if constexpr (
+        std::is_same_v<T2, __nv_bfloat162> || std::is_same_v<T2, __half2>) {
+      asm volatile(
+          "ldmatrix.sync.aligned.m8n8.x4.shared::cta.b16 {%0, %1, %2, %3}, [%4];\n"
+          : "=r"(*(uint32_t*)&(values[0])),
+            "=r"(*(uint32_t*)&(values[1])),
+            "=r"(*(uint32_t*)&(values[2])),
+            "=r"(*(uint32_t*)&(values[3]))
+          : "r"(row_address));
+    }
+  }
+
+  /**
+   * Store the tile to the address pointed to by `x`.
+   *
+   * The provided pointer is a generic pointer but this is meant to be used to
+   * store to global memory. For storing to shared memory we should use
+   * `stmatrix`.
+   *
+   * This also showcases the format of the tile quite nicely. Each register is
+   * holding to adjacent values. The indices are
+   *
+   *    row + 0, col + 0
+   *    row + 8, col + 0
+   *    row + 0, col + 8
+   *    row + 8, col + 8
+   *
+   * Given that we are dealing with Vector2_t<U> the column offsets are 4
+   * instead of 8.
+   */
+  template <typename U>
+  __device__ inline void store_global(U* x, int N) {
+    using U2 = Vector2_t<U>;
+    U2* x2 = reinterpret_cast<U2*>(x);
+    const int laneid = threadIdx.x % 32;
+    const int row = laneid / 4;
+    const int col = laneid % 4;
+    if constexpr (std::is_same_v<U2, T2>) {
+      x2[(row + 0) * (N / 2) + col + 0] = values[0];
+      x2[(row + 0) * (N / 2) + col + 4] = values[2];
+      x2[(row + 8) * (N / 2) + col + 0] = values[1];
+      x2[(row + 8) * (N / 2) + col + 4] = values[3];
+    } else if constexpr (
+        std::is_same_v<T2, float2> && std::is_same_v<U, __nv_bfloat16>) {
+      x2[(row + 0) * (N / 2) + col + 0] =
+          __floats2bfloat162_rn(values[0].x, values[0].y);
+      x2[(row + 0) * (N / 2) + col + 4] =
+          __floats2bfloat162_rn(values[2].x, values[2].y);
+      x2[(row + 8) * (N / 2) + col + 0] =
+          __floats2bfloat162_rn(values[1].x, values[1].y);
+      x2[(row + 8) * (N / 2) + col + 4] =
+          __floats2bfloat162_rn(values[3].x, values[3].y);
+    }
+  }
+
+  template <typename U>
+  __device__ inline void store_global_safe(U* x, int N, int max_rows) {
+    const int laneid = threadIdx.x % 32;
+    const int row = laneid / 4;
+    const int col = laneid % 4;
+    if (row < max_rows) {
+      x[(row + 0) * N + 2 * col + 0] = static_cast<U>(values[0].x);
+      x[(row + 0) * N + 2 * col + 1] = static_cast<U>(values[0].y);
+      x[(row + 0) * N + 2 * col + 8] = static_cast<U>(values[2].x);
+      x[(row + 0) * N + 2 * col + 9] = static_cast<U>(values[2].y);
+    }
+    if (row + 8 < max_rows) {
+      x[(row + 8) * N + 2 * col + 0] = static_cast<U>(values[1].x);
+      x[(row + 8) * N + 2 * col + 1] = static_cast<U>(values[1].y);
+      x[(row + 8) * N + 2 * col + 8] = static_cast<U>(values[3].x);
+      x[(row + 8) * N + 2 * col + 9] = static_cast<U>(values[3].y);
+    }
+  }
+};
+
+/**
+ * A simple container of multiple Tile16x16.
+ *
+ * Provides utility functions for loading and manipulating collections of basic
+ * tiles.
+ */
+template <typename T, int ROWS_, int COLS_>
+struct RegisterTile {
+  static constexpr int ROWS = ROWS_;
+  static constexpr int COLS = COLS_;
+  static constexpr int TILES_X = COLS / 16;
+  static constexpr int TILES_Y = ROWS / 16;
+
+  Tile16x16<T> data[TILES_X * TILES_Y];
+
+  __device__ inline void fill(T v) {
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        data[i * TILES_X + j].fill(v);
+      }
+    }
+  }
+
+  template <typename Tile>
+  __device__ inline void
+  load(Tile& tile, uint32_t base_address, int row, int col) {
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        data[i * TILES_X + j].load(
+            tile.loc(base_address, row + i * 16, col + j * 16));
+      }
+    }
+  }
+
+  template <typename U>
+  __device__ inline void store_global(U* x, int N, int row, int col) {
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        data[i * TILES_X + j].store_global(
+            x + (row + i * 16) * N + col + j * 16, N);
+      }
+    }
+  }
+
+  template <typename U>
+  __device__ inline void
+  store_global_safe(U* x, int N, int row, int col, int max_rows) {
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        data[i * TILES_X + j].store_global_safe(
+            x + (row + i * 16) * N + col + j * 16, N, max_rows - row - i * 16);
+      }
+    }
+  }
+};
+
+template <typename T, int ROWS_, int COLS_>
+struct SharedTile {
+  static constexpr int ROWS = ROWS_;
+  static constexpr int COLS = COLS_;
+  static constexpr int TILES_X = COLS / 16;
+  static constexpr int TILES_Y = ROWS / 16;
+  static constexpr int NUMEL = ROWS * COLS;
+
+  // Swizzle taken from ThunderKittens.
+  //
+  // See inludes/types/shared/st.cuh
+  //
+  // I do feel that it is too math heavy and can be improved. Also the math is
+  // done every time although the addresses don't change from load to load. I
+  // guess we are expecting the compiler to figure that out.
+  static constexpr int swizzle_bytes =
+      (sizeof(T) == 2 ? (TILES_X % 4 == 0 ? 128 : (TILES_X % 2 == 0 ? 64 : 32))
+                      : (sizeof(T) == 4 ? (TILES_X % 2 == 0 ? 128 : 64) : 0));
+
+  T data[ROWS * COLS];
+
+  // Return a pointer to the element at (row, col) using the swizzle.
+  __device__ static inline T* ptr(T* ptr, int row, int col) {
+    if constexpr (swizzle_bytes > 0) {
+      static constexpr int swizzle_repeat = swizzle_bytes * 8;
+      static constexpr int subtile_cols = swizzle_bytes / sizeof(T);
+      const int outer_idx = col / subtile_cols;
+      const uint64_t addr =
+          (uint64_t)(&ptr
+                         [outer_idx * ROWS * subtile_cols + row * subtile_cols +
+                          col % subtile_cols]);
+      const int swizzle = ((addr % swizzle_repeat) >> 7) << 4;
+      return (T*)(addr ^ swizzle);
+    } else {
+      return ptr + row * COLS + col;
+    }
+  }
+
+  // Return the location of the element at (row, col) using the swizzle.
+  __device__ static inline uint32_t loc(uint32_t ptr, int row, int col) {
+    if constexpr (swizzle_bytes > 0) {
+      static constexpr int swizzle_repeat = swizzle_bytes * 8;
+      static constexpr int subtile_cols = swizzle_bytes / sizeof(T);
+      const int outer_idx = col / subtile_cols;
+      const uint32_t addr = ptr +
+          sizeof(T) *
+              (outer_idx * ROWS * subtile_cols + row * subtile_cols +
+               col % subtile_cols);
+      const int swizzle = ((addr % swizzle_repeat) >> 7) << 4;
+      return (addr ^ swizzle);
+    } else {
+      return ptr + sizeof(T) * (row * COLS + col);
+    }
+  }
+
+  // Convenience functions to edit elements going through the swizzle.
+  __device__ inline T& operator()(int row, int col) {
+    return *ptr(data, row, col);
+  }
+  __device__ inline void store(float4& v, int row, int col) {
+    *(reinterpret_cast<float4*>(ptr(data, row, col))) = v;
+  }
+  __device__ inline void store(float2& v, int row, int col) {
+    *(reinterpret_cast<float2*>(ptr(data, row, col))) = v;
+  }
+  __device__ inline void store(float& v, int row, int col) {
+    *(reinterpret_cast<float*>(ptr(data, row, col))) = v;
+  }
+  template <int N>
+  __device__ inline void store(T (&v)[N], int row, int col) {
+    if constexpr (sizeof(T) * N == 4) {
+      store(*(reinterpret_cast<float*>(&v[0])), row, col);
+    } else if constexpr (sizeof(T) * N == 8) {
+      store(*(reinterpret_cast<float2*>(&v[0])), row, col);
+    } else if constexpr (sizeof(T) * N == 16) {
+      store(*(reinterpret_cast<float4*>(&v[0])), row, col);
+    } else {
+      MLX_UNROLL
+      for (int i = 0; i < N; i++) {
+        *ptr(data, row, col + i) = v[i];
+      }
+    }
+  }
+};
+
+/**
+ * Load the tile from global memory by loading 16 bytes at a time and storing
+ * them immediately.
+ */
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void load(Tile& tile, const T* x, int N) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+  const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+  x += row * N + col * ELEMENTS_PER_LOAD;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    float4 tmp;
+    tmp = *(reinterpret_cast<const float4*>(&x[i * STEP_ROWS * N]));
+    tile.store(tmp, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD);
+  }
+}
+
+/**
+ * Copy 16 bytes from the globale memory address pointed to by x to the smem
+ * address pointed to by row_address.
+ *
+ * A simple wrapper over the PTX.
+ */
+template <typename T>
+__device__ inline void cp_async_16(uint32_t row_address, const T* x) {
+  asm volatile(
+      "cp.async.ca.shared::cta.global [%0], [%1], 16;\n" ::"r"(row_address),
+      "l"(reinterpret_cast<const int4*>(x)));
+}
+
+/**
+ * Submit all the previous async copies to be executed.
+ */
+__device__ inline void cp_async_commit() {
+  asm volatile("cp.async.commit_group;\n" ::);
+}
+
+/**
+ * Wait for all the async copies to finish.
+ */
+__device__ inline void cp_async_wait_all() {
+  asm volatile("cp.async.wait_all;\n" ::);
+}
+
+/**
+ * The asynchronous equivalent of load.
+ *
+ * Loads the tile from global memory by submitting a bunch of async copy
+ * instructions. The copy won't start until commit is called and we don't have
+ * a guarantee it will finish until wait is called.
+ *
+ * It should be used as follows
+ *
+ *    load(...)
+ *    load(...)
+ *    cp_async_commit()
+ *    do_other_stuff()
+ *    cp_async_wait_all()
+ *    do_stuff_with_shmem()
+ */
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void
+load_async(Tile& tile, uint32_t base_address, const T* x, int N) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+  const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+  x += row * N + col * ELEMENTS_PER_LOAD;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    cp_async_16(
+        tile.loc(base_address, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD),
+        x + i * STEP_ROWS * N);
+  }
+}
+
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void load_async_safe(
+    Tile& tile,
+    uint32_t base_address,
+    const T* x,
+    int N,
+    int max_rows) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+  const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+  x += row * N + col * ELEMENTS_PER_LOAD;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    if (row + i * STEP_ROWS < max_rows) {
+      cp_async_16(
+          tile.loc(base_address, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD),
+          x + i * STEP_ROWS * N);
+    } else {
+      float4 tmp = {0, 0, 0, 0};
+      tile.store(tmp, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD);
+    }
+  }
+}
+
+} // namespace mlx::core::cu