Update the mlx.launch and mlx.distributed_config docs

Co-authored-by: Awni Hannun <awni@apple.com>

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

@@ -2,9 +2,13 @@ name: 'Build CUDA wheel'
 description: 'Build CUDA wheel'
 
 inputs:
-  toolkit:
-    description: 'The CUDA toolkit'
+  arch:
+    description: 'Platform architecture tag'
     required: true
+    type: choice
+    options:
+      - x86_64
+      - aarch64
 
 runs:
   using: "composite"
@@ -12,9 +16,9 @@ runs:
     - name: Build package
       shell: bash
       env:
-        CMAKE_ARGS: -DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=/usr/local/${{ inputs.toolkit }}/bin/nvcc
+        CMAKE_ARGS: -DMLX_BUILD_CUDA=ON
       run: |
         pip install auditwheel build patchelf setuptools
         python setup.py clean --all
         MLX_BUILD_STAGE=2 python -m build -w
-        bash python/scripts/repair_cuda.sh
+        bash python/scripts/repair_cuda.sh ${{ inputs.arch }}

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

@@ -1,26 +0,0 @@
-name: 'Build and Test with CUDA'
-description: 'Build and test MLX with CUDA'
-
-inputs:
-  toolkit:
-    description: 'The CUDA toolkit'
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Install Python package
-      shell: bash
-      env:
-        DEBUG: 1
-        CMAKE_ARGS: -DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=/usr/local/${{ inputs.toolkit }}/bin/nvcc
-      run: pip install --no-build-isolation -e ".[dev]" -v
-
-    - name: Build CPP only
-      shell: bash
-      run: |
-        cmake . -B build \
-          -DMLX_BUILD_CUDA=ON \
-          -DCMAKE_CUDA_COMPILER=/usr/local/${{ inputs.toolkit }}/bin/nvcc \
-          -DCMAKE_BUILD_TYPE=DEBUG
-        cmake --build build -j $(nproc)

.github/actions/build-docs/action.yml

@@ -1,25 +1,19 @@
 name: 'Build Documentation'
-description: 'Build documentation on a mac'
+description: 'Build documentation'
 
 runs:
   using: "composite"
   steps:
     - name: Setup machine
-      uses: ./.github/actions/setup-macos
-
-    - name: Setup uv
-      uses: astral-sh/setup-uv@v6
-      with:
-          python-version: "3.10"
-          activate-environment: true
+      uses: ./.github/actions/setup-linux
 
     - name: Install dependencies
-      shell: sh
+      shell: bash
       run: |
-        brew install doxygen
-        uv pip install --upgrade pip cmake
-        uv pip install -r docs/requirements.txt
-        uv pip install . -v
+        sudo apt-get install -y doxygen
+        source .venv/bin/activate
+        pip install -r docs/requirements.txt
+        pip install . -v
   
     - name: Build documentation
       shell: bash
@@ -30,8 +24,8 @@ runs:
         make html O=-W
     
     - name: Create artifact tar
-      shell: sh
-      run: tar -cf artifact.tar --cd docs --dereference build/html index.html
+      shell: bash
+      run: tar -cf artifact.tar -C docs --dereference build/html index.html
 
     # Do it manually because upload-pages-artifact requires gtar
     - name: Upload artifact

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

@@ -1,15 +1,32 @@
 name: 'Build and Test on Linux'
-description: 'Build and test MLX on Linux'
+
+inputs:
+  toolkit:
+    description: 'The toolkit to build with'
+    required: false
+    default: 'cpu'
 
 runs:
   using: "composite"
   steps:
     - name: Install Python package
+      id: python_build
       shell: sh
       env:
-        CMAKE_ARGS: "-DCMAKE_COMPILE_WARNING_AS_ERROR=ON"
         DEBUG: 1
-      run: pip install --no-build-isolation -e ".[dev]" -v
+        CMAKE_ARGS: >-
+          -DCMAKE_COMPILE_WARNING_AS_ERROR=ON
+          -DMLX_BUILD_CUDA=${{ startsWith(inputs.toolkit, 'cuda') && 'ON' || 'OFF' }}
+      run: |
+        if ${{ startsWith(inputs.toolkit, 'cuda') && runner.arch == 'arm64' }} ; then
+          # There is no GPU in arm64 runner, use a common arch.
+          CMAKE_ARGS="$CMAKE_ARGS -DMLX_CUDA_ARCHITECTURES=90a"
+          # Can not build tests when the built executables can not run.
+          CMAKE_ARGS="$CMAKE_ARGS -DMLX_BUILD_TESTS=OFF"
+        fi
+        pip install --no-build-isolation -e ".[dev]" -v
+        # Pass the CMAKE_ARGS to following steps.
+        echo CMAKE_ARGS="$CMAKE_ARGS" >> $GITHUB_OUTPUT
 
     - name: Generate package stubs
       shell: sh
@@ -20,6 +37,5 @@ runs:
     - name: Build CPP only
       shell: bash
       run: |
-        mkdir -p build && cd build
-        cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
-        make -j $(nproc)
+        cmake . -B build -DCMAKE_BUILD_TYPE=Debug ${{ steps.python_build.outputs.CMAKE_ARGS }}
+        cmake --build build -j $(nproc)

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

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

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

@@ -15,6 +15,7 @@ runs:
   using: "composite"
   steps:
     - name: Use ccache
+      if: ${{ runner.arch == 'x86_64' }}
       uses: hendrikmuhs/ccache-action@v1.2
       with:
         key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ inputs.toolkit }}-py${{ inputs.python-version }}
@@ -35,7 +36,7 @@ runs:
       run: |
         python -m venv .venv
         source .venv/bin/activate
-        pip install cmake nanobind==2.4.0
+        pip install setuptools cmake nanobind==2.4.0
         echo PATH=$PATH >> $GITHUB_ENV
         # Make cmake search .venv for nanobind
         echo PYTHONPATH=`python -c 'import sys; print(sys.path[-1])'` >> $GITHUB_ENV
@@ -51,23 +52,23 @@ runs:
         # Note: the CI machine does not meet CUDA 13's driver requirement.
         # Compatibility matrix:
         # https://docs.nvidia.com/deeplearning/cudnn/backend/latest/reference/support-matrix.html
-        # The `nvcc` is installed into `/usr/local/cuda-VERSION/bin/nvcc` - but
-        # it's *not* on the default toolkit path.
         PACKAGES: |
           {
             "cuda-12.6": "libcudnn9-dev-cuda-12 cuda-toolkit-12-6",
-            "cuda-12.8": "libcudnn9-dev-cuda-12 cuda-toolkit-12-8",
             "cuda-12.9": "libcudnn9-dev-cuda-12 cuda-toolkit-12-9",
             "cuda-13.0": "libcudnn9-dev-cuda-13 cuda-toolkit-13-0"
           }
       run: |
-        export ARCH=${{ runner.arch == 'arm64' && 'arm64' || 'x86_64' }}
+        # The CUDA binaries are hosted in the "sbsa" repo, the "arm64" repo is
+        # Jetson specific. SBSA means Arm Server Base System Architecture.
+        ARCH=${{ runner.arch == 'arm64' && 'sbsa' || 'x86_64' }}
         wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/$ARCH/cuda-keyring_1.1-1_all.deb
         sudo dpkg -i cuda-keyring_1.1-1_all.deb
         sudo apt-get update
         sudo apt-get install -y \
             libnccl2 libnccl-dev \
             ${{ fromJson(env.PACKAGES)[inputs.toolkit] }}
+        echo "/usr/local/${{ inputs.toolkit }}/bin" >> $GITHUB_PATH
 
     - name: CUDA packages and driver report
       if: ${{ startsWith(inputs.toolkit, 'cuda') }}

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

@@ -1,8 +1,8 @@
 name: 'Run Linux tests'
 
 inputs:
-  cpu-only:
-    description: 'Skip GPU tests'
+  has-gpu:
+    description: 'Run GPU tests'
     required: false
     default: false
 
@@ -17,7 +17,7 @@ runs:
         echo "::endgroup::"
 
     - name: Run distributed tests
-      if: ${{ inputs.cpu-only == 'true' }}
+      if: ${{ inputs.has-gpu == 'false' }}
       shell: bash
       run: |
         echo "::group::Distributed tests"
@@ -30,6 +30,7 @@ runs:
         echo "::endgroup::"
 
     - name: Run Python tests - CPU
+      if: ${{ inputs.has-gpu == 'false' }}
       shell: bash
       env:
         DEVICE: cpu
@@ -39,7 +40,7 @@ runs:
         echo "::endgroup::"
 
     - name: Run Python tests - GPU
-      if: ${{ inputs.cpu-only == 'false' }}
+      if: ${{ inputs.has-gpu == 'true' }}
       shell: bash
       env:
         DEVICE: gpu
@@ -58,7 +59,7 @@ runs:
         echo "::endgroup::"
 
     - name: Run CPP tests - GPU
-      if: ${{ inputs.cpu-only == 'false' }}
+      if: ${{ inputs.has-gpu == 'true' }}
       shell: bash
       env:
         DEVICE: gpu

.github/workflows/build_and_test.yml

@@ -1,39 +1,67 @@
 name: Build and Test
 
-on: [pull_request, push]
+on:
+  pull_request:
+  push:
+    branches:
+      - main
+      # For testing CI without starting a pull request:
+      - test/*
 
 permissions:
   contents: read
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
+  cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
 jobs:
   check_lint:
+    name: Check Lint
     runs-on: ubuntu-22.04
     steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
+      - uses: actions/checkout@v6
       - uses: pre-commit/action@v3.0.1
 
   linux_build_and_test:
+    name: Linux (cpu, ${{ matrix.arch }})
+    needs: check_lint
     strategy:
-      matrix:
-        runner:
-          - ubuntu-22.04
-          - ubuntu-22.04-arm
       fail-fast: false
-    runs-on: ${{ matrix.runner }}
+      matrix:
+        arch: ['x86_64', 'aarch64']
+    runs-on: ${{ matrix.arch == 'x86_64' && 'ubuntu-22.04' || 'ubuntu-22.04-arm' }}
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
       - uses: ./.github/actions/build-linux
       - uses: ./.github/actions/test-linux
+
+  cuda_build_and_test:
+    name: Linux (${{ matrix.toolkit }}, ${{ matrix.arch }})
+    if: github.repository == 'ml-explore/mlx'
+    needs: check_lint
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: ['x86_64', 'aarch64']
+        toolkit: ['cuda-12.6', 'cuda-12.9']
+    runs-on: ${{ matrix.arch == 'x86_64' && 'gpu-t4-4-core' || 'ubuntu-22.04-arm' }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
         with:
-          cpu-only: true
+          toolkit: ${{ matrix.toolkit }}
+      - uses: ./.github/actions/build-linux
+        with:
+          toolkit: ${{ matrix.toolkit }}
+      - uses: ./.github/actions/test-linux
+        if: matrix.arch == 'x86_64'
+        with:
+          has-gpu: true
 
   mac_build_and_test:
+    name: macOS (${{ matrix.macos-target }})
     if: github.repository == 'ml-explore/mlx'
     strategy:
       matrix:
@@ -43,38 +71,22 @@ jobs:
       MACOSX_DEPLOYMENT_TARGET: ${{ matrix.macos-target }}
     needs: check_lint
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-macos
       - uses: ./.github/actions/build-macos
 
-  cuda_build_and_test:
-    if: github.repository == 'ml-explore/mlx'
-    strategy:
-      fail-fast: false
-      matrix:
-        toolkit: ['cuda-12.8', 'cuda-12.9']
-    runs-on: gpu-t4-4-core
-    needs: check_lint
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-        with:
-          toolkit: ${{ matrix.toolkit }}
-      - uses: ./.github/actions/build-cuda
-        with:
-          toolkit: ${{ matrix.toolkit }}
-      - uses: ./.github/actions/test-linux
-
   build_documentation:
+    name: Build Documentation
     if: github.repository == 'ml-explore/mlx'
-    runs-on: [self-hosted, macos]
+    runs-on: ubuntu-22.04
     needs: check_lint
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/build-docs
 
   linux_fedora_build_cpp:
-    name: Linux Fedora CPP Build (${{ matrix.arch }})
+    name: Linux Fedora (${{ matrix.arch }})
+    needs: check_lint
     strategy:
       fail-fast: false
       matrix:
@@ -89,7 +101,7 @@ jobs:
       image: fedora:42
     steps:
       - name: Checkout code
-        uses: actions/checkout@v5
+        uses: actions/checkout@v6
 
       - name: CPP Build Test - No Release
         run: |

.github/workflows/documentation.yml

@@ -8,9 +8,9 @@ permissions:
 
 jobs:
   build:
-    runs-on: [self-hosted, macos]
+    runs-on: ubuntu-22.04
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/build-docs
       
   deploy:

.github/workflows/nightly.yml

@@ -16,7 +16,7 @@ jobs:
         python_version: ["3.10", "3.14"]
     runs-on: ubuntu-22.04
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
       - uses: ./.github/actions/build-linux-release
         with:
@@ -40,20 +40,18 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+        python_version: ["3.11", "3.12", "3.13", "3.14"]
         runner:
           - ubuntu-22.04
           - ubuntu-22.04-arm
     runs-on: ${{ matrix.runner }}
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
         with:
           python-version: ${{ matrix.python_version }}
       - uses: ./.github/actions/build-linux
       - uses: ./.github/actions/test-linux
-        with:
-          cpu-only: true
 
   build_mac_release:
     if: github.repository == 'ml-explore/mlx'
@@ -62,7 +60,7 @@ jobs:
         python-version: ["3.10", "3.13"]
     runs-on: [self-hosted, macos]
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-macos
         with:
           python-version: ${{ matrix.python-version }}
@@ -78,28 +76,11 @@ jobs:
           macos-target: 14.0
           build-backend: ${{ matrix.python-version == '3.10' }}
 
-  build_cuda_with_tests:
-    if: github.repository == 'ml-explore/mlx'
-    strategy:
-      fail-fast: false
-      matrix:
-        toolkit: ['cuda-12.8', 'cuda-12.9']
-    runs-on: gpu-t4-4-core
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-        with:
-          toolkit: ${{ matrix.toolkit }}
-      - uses: ./.github/actions/build-cuda
-        with:
-          toolkit: ${{ matrix.toolkit }}
-      - uses: ./.github/actions/test-linux
-
   build_cuda_release:
     if: github.repository == 'ml-explore/mlx'
     runs-on: ubuntu-22-large
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
         with:
           toolkit: 'cuda-12.9'
@@ -113,25 +94,3 @@ jobs:
           name: mlx-cuda
           path: wheelhouse/mlx_cuda-*.whl
           retention-days: 7
-
-  linux_fedora_build_cpp:
-    name: Linux Fedora CPP Build (${{ matrix.arch }})
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - host: ubuntu-22.04
-            arch: x86_64
-          - host: ubuntu-22.04-arm
-            arch: aarch64
-
-    runs-on: ${{ matrix.host }}
-    container:
-      image: fedora:42
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v5
-
-      - name: CPP Build Test - No Release
-        run: |
-          bash ./.github/scripts/setup+build-cpp-linux-fedora-container.sh

.github/workflows/release.yml

@@ -23,9 +23,9 @@ jobs:
 
   build_documentation:
     if: github.repository == 'ml-explore/mlx'
-    runs-on: [self-hosted, macos]
+    runs-on: ubuntu-22.04
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/build-docs
     
   deploy_documentation:
@@ -53,7 +53,7 @@ jobs:
       PYPI_RELEASE: 1
       DEV_RELEASE: ${{ github.event.inputs.dev_release == 'true' && 1 || 0 }}
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
         with:
           python-version: ${{ matrix.python_version }}
@@ -65,14 +65,14 @@ jobs:
         uses: actions/upload-artifact@v5
         with:
           overwrite: true
-          name: linux-wheels-${{ matrix.python_version }}
+          name: linux-wheels-${{ matrix.python_version }}-${{ matrix.arch }}
           path: wheelhouse/mlx-*.whl
       - name: Upload CPU artifacts
         if: matrix.python_version == '3.10'
         uses: actions/upload-artifact@v5
         with:
           overwrite: true
-          name: mlx-cpu
+          name: mlx-cpu-${{ matrix.arch }}
           path: wheelhouse/mlx_cpu-*.whl
   
   build_mac_release:
@@ -86,7 +86,7 @@ jobs:
       DEV_RELEASE: ${{ github.event.inputs.dev_release == 'true' && 1 || 0 }}
 
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-macos
         with:
           python-version: ${{ matrix.python-version }}
@@ -128,19 +128,23 @@ jobs:
 
   build_cuda_release:
     if: github.repository == 'ml-explore/mlx'
-    runs-on: ubuntu-22-large
+    strategy:
+      matrix:
+        arch: ['x86_64', 'aarch64']
+        toolkit: ['cuda-12.9', 'cuda-13.0']
+    runs-on: ${{ matrix.arch == 'x86_64' && 'ubuntu-22-large' || 'ubuntu-22-large-arm' }}
     env:
       PYPI_RELEASE: 1
       DEV_RELEASE: ${{ github.event.inputs.dev_release == 'true' && 1 || 0 }}
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
         with:
-          toolkit: 'cuda-12.9'
+          toolkit: ${{ matrix.toolkit }}
       - name: Build Python package
         uses: ./.github/actions/build-cuda-release
         with:
-          toolkit: 'cuda-12.9'
+          arch: ${{ matrix.arch }}
       - name: Upload artifacts
         uses: actions/upload-artifact@v5
         with:
@@ -208,7 +212,8 @@ jobs:
     steps:
       - uses: actions/download-artifact@v6
         with:
-          name: mlx-cpu
+          pattern: mlx-cpu-*
+          merge-multiple: true
           path: dist
       - name: Display structure of downloaded files
         run: ls -R dist

benchmarks/python/blas/bench_gemv.py

@@ -1,6 +1,5 @@
 # Copyright © 2023 Apple Inc.
 
-import argparse
 import os
 import subprocess
 import time

benchmarks/python/masked_scatter.py

@@ -0,0 +1,212 @@
+import math
+import os
+import subprocess
+import time
+from copy import copy
+from functools import partial
+
+import matplotlib.pyplot as plt
+import mlx.core as mx
+import numpy as np
+import torch
+from matplotlib.ticker import FuncFormatter
+
+RESULTS_DIR = "./results"
+
+
+if not os.path.isdir(RESULTS_DIR):
+    os.mkdir(RESULTS_DIR)
+
+DEVICE_NAME = subprocess.check_output(["sysctl", "-n", "machdep.cpu.brand_string"])
+DEVICE_NAME = DEVICE_NAME.decode("utf-8").strip("\n")
+
+TORCH_DEVICE = torch.device(
+    "mps"
+    if torch.backends.mps.is_available()
+    else ("cuda" if torch.cuda.is_available() else "cpu")
+)
+
+
+N_WARMUP = 5
+N_ITER_BENCH = 50
+N_ITER_FUNC = 20
+
+VECTOR_LENGTHS = [4096 * (2**i) for i in range(10)]
+MASK_DENSITIES = [0.01, 0.1, 0.25, 0.5]
+D_TYPES = ("float32", "float16")
+
+
+def _power_of_two_formatter(value, _position):
+    if value <= 0:
+        return ""
+    exponent = int(round(math.log2(value)))
+    if abs(value - (1 << exponent)) / value > 1e-6:
+        return f"{value:g}"
+    return f"$2^{{{exponent}}}$"
+
+
+def torch_sync():
+    if TORCH_DEVICE.type == "cuda":
+        torch.cuda.synchronize()
+    elif TORCH_DEVICE.type == "mps":
+        torch.mps.synchronize()
+
+
+def masked_scatter_mlx(self_arr, mask_arr, src_arr):
+    outs = []
+    for _ in range(N_ITER_FUNC):
+        out = copy(self_arr)
+        out[mask_arr] = src_arr
+        outs.append(out)
+    mx.eval(outs)
+    return outs
+
+
+@torch.no_grad()
+def masked_scatter_torch(self_tensor, mask_tensor, src_tensor):
+    outs = []
+    for _ in range(N_ITER_FUNC):
+        out = self_tensor.clone()
+        out.masked_scatter_(mask_tensor, src_tensor)
+        outs.append(out)
+    torch_sync()
+    return outs
+
+
+def measure(fn):
+    for _ in range(N_WARMUP):
+        fn()
+    start = time.perf_counter_ns()
+    for _ in range(N_ITER_BENCH):
+        fn()
+    end = time.perf_counter_ns()
+    return (end - start) * 1e-9
+
+
+def bytes_touched(length, true_count, item_size):
+    mask_bytes = length
+    self_bytes = length * item_size * 2  # read + write
+    src_bytes = true_count * item_size
+    return (mask_bytes + self_bytes + src_bytes) * N_ITER_FUNC * N_ITER_BENCH
+
+
+def build_case(length, density, np_dtype, torch_dtype):
+    true_count = max(1, int(round(length * density)))
+
+    rng = np.random.default_rng()
+    self_np = rng.normal(0.0, 1.0, length).astype(np_dtype)
+    mask_np = np.zeros(length, dtype=bool)
+    mask_np[:true_count] = True
+    rng.shuffle(mask_np)
+    src_np = rng.normal(0.0, 1.0, true_count).astype(np_dtype)
+
+    self_mlx = mx.array(self_np)
+    mask_mlx = mx.array(mask_np)
+    src_mlx = mx.array(src_np)
+
+    self_torch = torch.from_numpy(self_np).to(device=TORCH_DEVICE, dtype=torch_dtype)
+    mask_torch = torch.from_numpy(mask_np).to(device=TORCH_DEVICE)
+    src_torch = torch.from_numpy(src_np).to(device=TORCH_DEVICE, dtype=torch_dtype)
+
+    # Correctness check once per configuration
+    mx_out = mx.array(self_np)
+    mx_out[mask_mlx] = src_mlx
+    mx.eval(mx_out)
+    torch_out = self_torch.clone()
+    torch_out.masked_scatter_(mask_torch, src_torch)
+
+    atol = 5e-3 if np_dtype == np.float16 else 1e-5
+    if not np.allclose(np.array(mx_out), torch_out.cpu().numpy(), atol=atol):
+        raise AssertionError("masked_scatter results diverged between MLX and Torch")
+
+    return (self_mlx, mask_mlx, src_mlx, self_torch, mask_torch, src_torch, true_count)
+
+
+def bench_case(length, density, dtype):
+    np_dtype = getattr(np, dtype)
+    torch_dtype = getattr(torch, dtype)
+    (
+        self_mlx,
+        mask_mlx,
+        src_mlx,
+        self_torch,
+        mask_torch,
+        src_torch,
+        true_count,
+    ) = build_case(length, density, np_dtype, torch_dtype)
+
+    time_mlx = measure(partial(masked_scatter_mlx, self_mlx, mask_mlx, src_mlx))
+    time_torch = measure(
+        partial(masked_scatter_torch, self_torch, mask_torch, src_torch)
+    )
+
+    total_bytes = bytes_touched(length, true_count, np_dtype().itemsize)
+    bytes_per_gb = float(1024**3)
+    mlx_gbps = (total_bytes / bytes_per_gb) / time_mlx
+    torch_gbps = (total_bytes / bytes_per_gb) / time_torch
+
+    return time_mlx, time_torch, mlx_gbps, torch_gbps
+
+
+def plot_density(ax_perf, ax_speedup, density, dtype):
+    mlx_gbps = []
+    torch_gbps = []
+    mlx_times = []
+    torch_times = []
+
+    for length in VECTOR_LENGTHS:
+        t_mlx, t_torch, gbps_mlx, gbps_torch = bench_case(length, density, dtype)
+        mlx_gbps.append(gbps_mlx)
+        torch_gbps.append(gbps_torch)
+        mlx_times.append(t_mlx)
+        torch_times.append(t_torch)
+
+    ax_perf.plot(VECTOR_LENGTHS, mlx_gbps, "tab:blue", label="MLX")
+    ax_perf.plot(VECTOR_LENGTHS, torch_gbps, "tab:red", label="Torch")
+    ax_perf.set_xscale("log", base=2)
+    ax_perf.set_xticks(VECTOR_LENGTHS)
+    formatter = FuncFormatter(_power_of_two_formatter)
+    ax_perf.xaxis.set_major_formatter(formatter)
+    ax_perf.set_title(f"density={density:.2f}")
+    ax_perf.set_ylabel("GB/s")
+    ax_perf.grid(True, which="both", linestyle=":", alpha=0.4)
+    ax_perf.legend()
+
+    speedup = np.array(torch_times) / np.array(mlx_times)
+    ax_speedup.plot(VECTOR_LENGTHS, speedup, "tab:green")
+    ax_speedup.axhline(1.0, color="tab:gray", linestyle="--")
+    ax_speedup.set_xscale("log", base=2)
+    ax_speedup.set_xticks(VECTOR_LENGTHS)
+    ax_speedup.xaxis.set_major_formatter(formatter)
+    ax_speedup.set_ylabel("Speedup (Torch_t / MLX_t)")
+    ax_speedup.grid(True, which="both", linestyle=":", alpha=0.4)
+
+
+def main():
+    for dtype in D_TYPES:
+        fig, axs = plt.subplots(
+            len(MASK_DENSITIES),
+            2,
+            figsize=(10, 12),
+            layout="constrained",
+            sharex=True,
+        )
+
+        for i, density in enumerate(MASK_DENSITIES):
+            plot_density(axs[i][0], axs[i][1], density, dtype)
+            axs[i][0].set_xlabel("vector length")
+            axs[i][1].set_xlabel("vector length")
+
+        fig.suptitle(
+            f"{DEVICE_NAME.replace('Apple ', '')} ({TORCH_DEVICE.type}) | dtype={dtype}"
+        )
+        output_path = os.path.join(
+            RESULTS_DIR,
+            f"{DEVICE_NAME.replace(' ', '_')}_masked_scatter_{dtype}.pdf",
+        )
+        fig.savefig(output_path)
+        plt.close(fig)
+
+
+if __name__ == "__main__":
+    main()

docs/src/install.rst

@@ -29,17 +29,20 @@ MLX has a CUDA backend which you can install with:
 
 .. code-block:: shell
 
-    pip install mlx[cuda]
+    pip install mlx[cuda12]
+
 
 To install the CUDA package from PyPi your system must meet the following
 requirements:
 
-- Nvidia architecture >= SM 7.0 (Volta)
+- Nvidia architecture >= SM 7.5
 - Nvidia driver >= 550.54.14
 - CUDA toolkit >= 12.0
 - Linux distribution with glibc >= 2.35
 - Python >= 3.10
 
+For CUDA 13 use ``pip install mlx[cuda13]``. The CUDA 13 package requires
+an Nvidia driver >= 580 or an appropriate CUDA compatibility package.
 
 CPU-only (Linux)
 ^^^^^^^^^^^^^^^^

docs/src/usage/distributed.rst

@@ -7,22 +7,29 @@ Distributed Communication
 
 MLX supports distributed communication operations that allow the computational cost
 of training or inference to be shared across many physical machines. At the
-moment we support three different communication backends:
+moment we support several different communication backends introduced below.
+
+.. list-table::
+   :widths: 20 80
+   :header-rows: 1
+
+   * - Backend
+     - Description
+   * - :ref:`MPI <mpi_section>`
+     - A full featured and mature distributed communications library.
+   * - :ref:`RING <ring_section>`
+     - Ring all reduce and all gather over TCP sockets. Always available and
+       usually faster than MPI.
+   * - :ref:`JACCL <ring_section>`
+     - Low latency communication with RDMA over thunderbolt. Necessary for
+       things like tensor parallelism.
+   * - :ref:`NCCL <nccl_section>`
+     - The backend of choice for CUDA environments.
 
-* `MPI <https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ a
-  full-featured and mature distributed communications library
-* A **ring** backend of our own that uses native TCP sockets. It should be
-  faster for thunderbolt connections, but it also works over Ethernet.
-* `nccl <https://developer.nvidia.com/nccl>`_, for use in CUDA environments.
 
 The list of all currently supported operations and their documentation can be
 seen in the :ref:`API docs<distributed>`.
 
-.. note::
-   Some operations may not be supported or not as fast as they should be.
-   We are adding more and tuning the ones we have as we are figuring out the
-   best way to do distributed computing on Macs using MLX.
-
 Getting Started
 ---------------
 
@@ -85,7 +92,7 @@ Selecting Backend
 ^^^^^^^^^^^^^^^^^
 
 You can select the backend you want to use when calling :func:`init` by passing
-one of ``{'any', 'ring', 'mpi', 'nccl'}``. When passing ``any``, MLX will try all
+one of ``{'any', 'ring', 'jaccl', 'mpi', 'nccl'}``. When passing ``any``, MLX will try all
 available backends. If they all fail then a singleton group is created.
 
 .. note::
@@ -110,6 +117,8 @@ The following examples aim to clarify the backend initialization logic in MLX:
     world_ring = mx.distributed.init(backend="ring")
     world_any = mx.distributed.init()  # same as MPI because it was initialized first!
 
+.. _training_example:
+
 Training Example
 ----------------
 
@@ -192,16 +201,273 @@ almost identical to the example above:
         loss = step(model, x, y)
         mx.eval(loss, model.parameters())
 
+.. _ring_section:
+
+Getting Started with Ring
+-------------------------
+
+The ring backend does not depend on any third party library so it is always
+available. It uses TCP sockets so the nodes need to be reachable via a network.
+As the name suggests the nodes are connected in a ring which means that rank 1
+can only communicate with rank 0 and rank 2, rank 2 only with rank 1 and rank 3
+and so on and so forth. As a result :func:`send` and :func:`recv` with
+arbitrary sender and receiver is not supported in the ring backend.
+
+Defining a Ring
+^^^^^^^^^^^^^^^
+
+The easiest way to define and use a ring is via a JSON hostfile and the
+``mlx.launch`` :doc:`helper script <launching_distributed>`. For each node one
+defines a hostname to ssh into to run commands on this node and one or more IPs
+that this node will listen to for connections.
+
+For example the hostfile below defines a 4 node ring. ``hostname1`` will be
+rank 0, ``hostname2`` rank 1 etc.
+
+.. code:: json
+
+    [
+        {"ssh": "hostname1", "ips": ["123.123.123.1"]},
+        {"ssh": "hostname2", "ips": ["123.123.123.2"]},
+        {"ssh": "hostname3", "ips": ["123.123.123.3"]},
+        {"ssh": "hostname4", "ips": ["123.123.123.4"]}
+    ]
+
+Running ``mlx.launch --hostfile ring-4.json my_script.py`` will ssh into each
+node, run the script which will listen for connections in each of the provided
+IPs. Specifically, ``hostname1`` will connect to ``123.123.123.2`` and accept a
+connection from ``123.123.123.4`` and so on and so forth.
+
+Thunderbolt Ring
+^^^^^^^^^^^^^^^^
+
+Although the ring backend can have benefits over MPI even for Ethernet, its
+main purpose is to use Thunderbolt rings for higher bandwidth communication.
+Setting up such thunderbolt rings can be done manually, but is a relatively
+tedious process. To simplify this, we provide the utility ``mlx.distributed_config``.
+
+To use ``mlx.distributed_config`` your computers need to be accessible by ssh via
+Ethernet or Wi-Fi. Subsequently, connect them via thunderbolt cables and then call the
+utility as follows:
+
+.. code:: shell
+
+   mlx.distributed_config --verbose --hosts host1,host2,host3,host4 --backend ring
+
+By default the script will attempt to discover the thunderbolt ring and provide
+you with the commands to configure each node as well as the ``hostfile.json``
+to use with ``mlx.launch``. If password-less ``sudo`` is available on the nodes
+then ``--auto-setup`` can be used to configure them automatically.
+
+If you want to go through the process manually, the steps are as follows:
+
+* Disable the thunderbolt bridge interface
+* For the cable connecting rank ``i`` to rank ``i + 1`` find the interfaces
+  corresponding to that cable in nodes ``i`` and ``i + 1``.
+* Set up a unique subnetwork connecting the two nodes for the corresponding
+  interfaces. For instance if the cable corresponds to ``en2`` on node ``i``
+  and ``en2`` also on node ``i + 1`` then we may assign IPs ``192.168.0.1`` and
+  ``192.168.0.2`` respectively to the two nodes. For more details you can see
+  the commands prepared by the utility script.
+
+.. _jaccl_section:
+
+Getting Started with RDMA over Thunderbolt
+------------------------------------------
+
+Starting from version 26.2 RDMA over thunderbolt is available in MacOS and
+enables low-latency communication between Macs with thunderbolt 5. MLX provides
+the JACCL backend that uses this functionality to achieve communication latency
+an order of magnitude lower than the ring backend.
+
+.. note::
+
+   The name JACCL (pronounced Jackal) stands for *Jack and Angelos' Collective
+   Communication Library* and it is an obvious pun to Nvidia's NCCL but also
+   tribute to *Jack Beasley* who led the development of RDMA over Thunderbolt
+   at Apple.
+
+Enabling RDMA
+^^^^^^^^^^^^^
+
+Until the feature matures, enabling RDMA over thunderbolt is slightly more
+involved and **cannot** be done remotely even with sudo. In fact, it has to be
+done in macOS recovery:
+
+1. `Start your computer in recovery <https://support.apple.com/en-us/102518>`_.
+2. Open the Terminal by going to Utilities -> Terminal.
+3. Run ``rdma_ctl enable``.
+4. Reboot.
+
+To verify that you have successfully enabled Thunderbolt RDMA you can run
+``ibv_devices`` which should produce something like the following for an M3 Ultra.
+
+.. code-block:: bash
+
+    ~ % ibv_devices
+    device          	   node GUID
+    ------          	----------------
+    rdma_en2        	8096a9d9edbaac05
+    rdma_en3        	8196a9d9edbaac05
+    rdma_en5        	8396a9d9edbaac05
+    rdma_en4        	8296a9d9edbaac05
+    rdma_en6        	8496a9d9edbaac05
+    rdma_en7        	8596a9d9edbaac05
+
+Defining a Mesh
+^^^^^^^^^^^^^^^
+
+The JACCL backend supports only fully connected topologies. Namely, there needs
+to be a thunderbolt cable connecting all pairs of Macs directly. For example, in
+the following topology visualizations, the left one is valid because there is a
+connection from any node to any other node, while for the one on the right M3
+Ultra 1 is not connected to M3 Ultra 2.
+
+.. raw:: html
+
+   <div style="display: flex; text-align: center; align-items: end; font-size: 80%;">
+     <div>
+       <img src="/_static/distributed/m3-ultra-mesh.png" alt="M3 Ultra thunderbolt mesh" style="width: 55%">
+       <p>Fully connected mesh of four M3 Ultra.</p>
+     </div>
+     <div>
+       <img src="/_static/distributed/m3-ultra-mesh-broken.png" alt="M3 Ultra broken thunderbolt mesh" style="width: 55%">
+       <p>Not a valid mesh (M3 Ultra 1 is not connected to M3 Ultra 2).</p>
+     </div>
+   </div>
+
+Similar to the ring backend, the easiest way to use JACCL with MLX is to write
+a JSON hostfile that will be used by ``mlx.launch``. The hostfile needs to contain
+
+- Hostnames to use for launching scripts via ssh
+- An IP for rank 0 that is reachable by all nodes
+- A list of rdma devices that connect each node to each other node
+
+The following JSON defines the valid 4-node mesh from the image above.
+
+.. code-block:: json
+
+    [
+        {
+            "ssh": "m3-ultra-1",
+            "ips": ["123.123.123.1"],
+            "rdma": [null, "rdma_en5", "rdma_en4", "rdma_en3"]
+        },
+        {
+            "ssh": "m3-ultra-2",
+            "ips": [],
+            "rdma": ["rdma_en5", null, "rdma_en3", "rdma_en4"]
+        },
+        {
+            "ssh": "m3-ultra-3",
+            "ips": [],
+            "rdma": ["rdma_en4", "rdma_en3", null, "rdma_en5"]
+        },
+        {
+            "ssh": "m3-ultra-4",
+            "ips": [],
+            "rdma": ["rdma_en3", "rdma_en4", "rdma_en5", null]
+        }
+    ]
+
+Even though TCP/IP is not used when communicating with Thunderbolt RDMA,
+disabling the thunderbolt bridge is still required as well as setting up
+isolated local networks for each thunderbolt connection.
+
+All of the above can be done instead via ``mlx.distributed_config``. This helper
+script will
+
+- ssh into each node
+- extract the thunderbolt connectivity
+- check for a valid mesh
+- provide the commands to configure each node (or run them if sudo is available)
+- generate the hostfile to be used with ``mlx.launch``
+
+Putting it All Together
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+For example launching a distributed MLX script that uses JACCL is fairly simple
+if the nodes are reachable via ssh and have password-less sudo.
+
+First, connect all the thunderbolt cables. Then we can verify the connections
+by using the ``mlx.distributed_config`` script to visualize them.
+
+.. code-block::
+
+   mlx.distributed_config --verbose \
+        --hosts m3-ultra-1,m3-ultra-2,m3-ultra-3,m3-ultra-4 \
+        --over thunderbolt --dot | dot -Tpng | open -f -a Preview
+
+After making sure that everything looks right we can auto-configure the nodes
+and save the hostfile to ``m3-ultra-jaccl.json`` by running:
+
+.. code-block::
+
+   mlx.distributed_config --verbose \
+        --hosts m3-ultra-1,m3-ultra-2,m3-ultra-3,m3-ultra-4 \
+        --over thunderbolt --backend jaccl \
+        --auto-setup --output m3-ultra-jaccl.json
+
+And now we are ready to run a distributed MLX script such as distributed inference
+of a gigantic model using MLX-LM.
+
+.. code-block::
+
+   mlx.launch --verbose --backend jaccl --hostfile m3-ultra-jaccl.json \
+        --env MLX_METAL_FAST_SYNCH=1 -- \  # <--- important
+        /path/to/remote/python -m mlx_lm chat --model mlx-community/DeepSeek-V3.2-8bit --shard
+
+.. note::
+
+   Defining the environment variable ``MLX_METAL_FAST_SYNCH=1`` enables a
+   different, faster way of synchronizing between the GPU and the CPU. It is
+   not specific to the JACCL backend and can be used in all cases where the CPU
+   and GPU need to collaborate for some computation and is pretty critical for
+   low-latency communication since the communication is done by the CPU.
+
+.. _nccl_section:
+
+Getting Started with NCCL
+-------------------------
+
+MLX on CUDA environments ships with the ability to talk to `NCCL
+<https://developer.nvidia.com/nccl>`_ which is a high-performance collective
+communication library that supports both multi-gpu and multi-node setups.
+
+For CUDA environments, NCCL is the default backend for ``mlx.launch`` and all
+it takes to run a distributed job is
+
+.. code-block::
+
+   mlx.launch -n 8 test.py
+
+   # perfect for interactive scripts
+   mlx.launch -n 8 python -m mlx_lm chat --model my-model --shard
+
+You can also use ``mlx.launch`` to ssh to a remote node and launch a script
+with the same ease
+
+.. code-block::
+
+   mlx.launch --hosts my-cuda-node -n 8 test.py
+
+In many cases you may not want to use ``mlx.launch`` with the NCCL backend
+because the cluster scheduler will be the one launching the processes. You can
+:ref:`see which environment variables need to be defined <no_mlx_launch>` in
+order for the MLX NCCL backend to be initialized correctly.
+
+.. _mpi_section:
 
 Getting Started with MPI
 ------------------------
 
-MLX already comes with the ability to "talk" to MPI if it is installed on the
-machine. Launching distributed MLX programs that use MPI can be done with
-``mpirun`` as expected. However, in the following examples we will be using
-``mlx.launch --backend mpi`` which takes care of some nuisances such as setting
-absolute paths for the ``mpirun`` executable and the ``libmpi.dyld`` shared
-library.
+MLX already comes with the ability to "talk" to `MPI
+<https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ if it is installed
+on the machine. Launching distributed MLX programs that use MPI can be done
+with ``mpirun`` as expected. However, in the following examples we will be
+using ``mlx.launch --backend mpi`` which takes care of some nuisances such as
+setting absolute paths for the ``mpirun`` executable and the ``libmpi.dyld``
+shared library.
 
 The simplest possible usage is the following which, assuming the minimal
 example in the beginning of this page, should result in:
@@ -269,78 +535,116 @@ Force MPI to use the most performant network interface by setting ``--mca
 btl_tcp_if_include <iface>`` where ``<iface>`` should be the interface you want
 to use.
 
-Getting Started with Ring
--------------------------
+.. _no_mlx_launch:
 
-The ring backend does not depend on any third party library so it is always
-available. It uses TCP sockets so the nodes need to be reachable via a network.
-As the name suggests the nodes are connected in a ring which means that rank 1
-can only communicate with rank 0 and rank 2, rank 2 only with rank 1 and rank 3
-and so on and so forth. As a result :func:`send` and :func:`recv` with
-arbitrary sender and receiver is not supported in the ring backend.
+Distributed Without ``mlx.launch``
+----------------------------------
 
-Defining a Ring
-^^^^^^^^^^^^^^^
+None of the implementations of the distributed backends require launching with
+``mlx.launch``. The script simply connects to each host. Starts a process per
+rank and sets up the necessary environment variables before delegating to your
+MLX script. See the :doc:`dedicated documentation page <launching_distributed>`
+for more details.
 
-The easiest way to define and use a ring is via a JSON hostfile and the
-``mlx.launch`` :doc:`helper script <launching_distributed>`. For each node one
-defines a hostname to ssh into to run commands on this node and one or more IPs
-that this node will listen to for connections.
+For many use-cases this will be the easiest way to perform distributed
+computations in MLX. However, there may be reasons that you cannot or should
+not use ``mlx.launch``. A common such case is the use of a scheduler that
+starts all the processes for you on machines undetermined at the time of
+scheduling the job.
 
-For example the hostfile below defines a 4 node ring. ``hostname1`` will be
-rank 0, ``hostname2`` rank 1 etc.
+Below we list the environment variables required to use each backend.
 
-.. code:: json
+Ring
+^^^^^^
 
-    [
-        {"ssh": "hostname1", "ips": ["123.123.123.1"]},
-        {"ssh": "hostname2", "ips": ["123.123.123.2"]},
-        {"ssh": "hostname3", "ips": ["123.123.123.3"]},
-        {"ssh": "hostname4", "ips": ["123.123.123.4"]}
-    ]
+**MLX_RANK** should contain a single 0-based integer that defines the rank of
+the process.
 
-Running ``mlx.launch --hostfile ring-4.json my_script.py`` will ssh into each
-node, run the script which will listen for connections in each of the provided
-IPs. Specifically, ``hostname1`` will connect to ``123.123.123.2`` and accept a
-connection from ``123.123.123.4`` and so on and so forth.
+**MLX_HOSTFILE** should contain the path to a json file that contains IPs and
+ports for each rank to listen to, something like the following:
 
-Thunderbolt Ring
-^^^^^^^^^^^^^^^^
+.. code-block:: json
 
-Although the ring backend can have benefits over MPI even for Ethernet, its
-main purpose is to use Thunderbolt rings for higher bandwidth communication.
-Setting up such thunderbolt rings can be done manually, but is a relatively
-tedious process. To simplify this, we provide the utility ``mlx.distributed_config``.
+   [
+     ["123.123.1.1:5000", "123.123.1.2:5000"],
+     ["123.123.2.1:5000", "123.123.2.2:5000"],
+     ["123.123.3.1:5000", "123.123.3.2:5000"],
+     ["123.123.4.1:5000", "123.123.4.2:5000"]
+   ]
 
-To use ``mlx.distributed_config`` your computers need to be accessible by ssh via
-Ethernet or Wi-Fi. Subsequently, connect them via thunderbolt cables and then call the
-utility as follows:
+**MLX_RING_VERBOSE** is optional and if set to 1 it enables some more logging
+from the distributed backend.
 
-.. code:: shell
+JACCL
+^^^^^
 
-   mlx.distributed_config --verbose --hosts host1,host2,host3,host4
+**MLX_RANK** should contain a single 0-based integer that defines the rank of
+the process.
 
-By default the script will attempt to discover the thunderbolt ring and provide
-you with the commands to configure each node as well as the ``hostfile.json``
-to use with ``mlx.launch``. If password-less ``sudo`` is available on the nodes
-then ``--auto-setup`` can be used to configure them automatically.
+**MLX_JACCL_COORDINATOR** should contain the IP and port that rank 0 can listen
+to all the other ranks connect to in order to establish the RDMA connections.
 
-To validate your connection without configuring anything
-``mlx.distributed_config`` can also plot the ring using DOT format.
+**MLX_IBV_DEVICES** should contain the path to a json file that contains the
+ibverbs device names that connect each node to each other node, something like
+the following:
 
-.. code:: shell
+.. code-block:: json
 
-   mlx.distributed_config --verbose --hosts host1,host2,host3,host4 --dot >ring.dot
-   dot -Tpng ring.dot >ring.png
-   open ring.png
+   [
+      [null, "rdma_en5", "rdma_en4", "rdma_en3"],
+      ["rdma_en5", null, "rdma_en3", "rdma_en4"],
+      ["rdma_en4", "rdma_en3", null, "rdma_en5"],
+      ["rdma_en3", "rdma_en4", "rdma_en5", null]
+   ]
 
-If you want to go through the process manually, the steps are as follows:
 
-* Disable the thunderbolt bridge interface
-* For the cable connecting rank ``i`` to rank ``i + 1`` find the interfaces
-  corresponding to that cable in nodes ``i`` and ``i + 1``.
-* Set up a unique subnetwork connecting the two nodes for the corresponding
-  interfaces. For instance if the cable corresponds to ``en2`` on node ``i``
-  and ``en2`` also on node ``i + 1`` then we may assign IPs ``192.168.0.1`` and
-  ``192.168.0.2`` respectively to the two nodes. For more details you can see
-  the commands prepared by the utility script.
+NCCL
+^^^^^
+
+**MLX_RANK** should contain a single 0-based integer that defines the rank of
+the process.
+
+**MLX_WORLD_SIZE** should contain the total number of processes that will be
+launched.
+
+**NCCL_HOST_IP** and **NCCL_PORT** should contain the IP and port that all
+hosts can connect to to establish the NCCL communication.
+
+**CUDA_VISIBLE_DEVICES** should contain the local index of the gpu that
+corresponds to this process.
+
+Of course any `other environment variable
+<https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html>`_ that is
+used by NCCL can be set.
+
+.. _tips_and_tricks:
+
+Tips and Tricks
+----------------
+
+This is a small collection of tips to help you utilize better the distributed
+communication capabilities of MLX.
+
+- *Test locally first.*
+
+  You can use the pattern ``mlx.launch -n2 -- my_script.py`` to run a small
+  scale test on a single node first.
+
+- *Batch your communication.*
+
+  As described in the :ref:`training example <training_example>`, performing a
+  lot of small communication can hurt performance. Copy the approach of
+  :func:`mlx.nn.average_gradients` to gather many small communications in a
+  single large one.
+
+- *Visualize the connectivity.*
+
+  Use ``mlx.distributed_config --hosts h1,h2,h3 --over thunderbolt --dot`` to
+  visualize the connnections and make sure that the cables are connected
+  correctly. See the :ref:`JACCL section <jaccl_section>` for examples.
+
+- *Use the debugger.*
+
+  ``mlx.launch`` is meant for interactive use. It broadcasts stdin to all
+  processes and gathers stdout from all processes. This makes using ``pdb`` a
+  breeze.

docs/src/usage/indexing.rst

@@ -70,7 +70,8 @@ Differences from NumPy
 
   * Indexing does not perform bounds checking. Indexing out of bounds is
     undefined behavior.
-  * Boolean mask based indexing is not yet supported.
+  * Boolean mask based indexing is supported for assignment only (see
+    :ref:`boolean-mask-assignment`).
 
 The reason for the lack of bounds checking is that exceptions cannot propagate
 from the GPU. Performing bounds checking for array indices before launching the
@@ -143,3 +144,51 @@ expected. For example:
 
 In the above ``dfdx`` will have the correct gradient, namely zeros at ``idx``
 and ones elsewhere.
+
+.. _boolean-mask-assignment:
+
+Boolean Mask Assignment
+-----------------------
+
+MLX supports boolean indices using NumPy syntax. A mask must already be
+a :class:`bool_` MLX :class:`array` or a NumPy ``ndarray`` with ``dtype=bool``.
+Other index types are routed through the standard scatter code.
+
+.. code-block:: shell
+
+   >>> a = mx.array([1.0, 2.0, 3.0])
+   >>> mask = mx.array([True, False, True])
+   >>> updates = mx.array([5.0, 6.0])
+   >>> a[mask] = updates
+   >>> a
+   array([5.0, 2.0, 6.0], dtype=float32)
+
+Scalar assignments broadcast to every ``True`` entry in ``mask``. For non-scalar
+assignments, ``updates`` must provide at least as many elements as there are
+``True`` entries in ``mask``.
+
+.. code-block:: shell
+
+   >>> a = mx.zeros((2, 3))
+   >>> mask = mx.array([[True, False, True],
+                        [False, False, True]])
+   >>> a[mask] = 1.0
+   >>> a
+   array([[1.0, 0.0, 1.0],
+          [0.0, 0.0, 1.0]], dtype=float32)
+
+Boolean masks follow NumPy semantics:
+
+- The mask shape must match the shape of the axes it indexes exactly. The only
+  exception is a scalar boolean mask, which broadcasts to the full array.
+- Any axes not covered by the mask are taken in full.
+
+.. code-block:: shell
+
+   >>> a = mx.arange(1000).reshape(10, 10, 10)
+   >>> a[mx.random.randn(10, 10) > 0.0] = 0  # valid: mask covers axes 0 and 1
+
+The mask of shape ``(10, 10)`` applies to the first two axes, so ``a[mask]``
+selects the 1-D slices ``a[i, j, :]`` where ``mask[i, j]`` is ``True``.
+Shapes such as ``(1, 10, 10)`` or ``(10, 10, 1)`` do not match the indexed
+axes and therefore raise errors.

docs/src/usage/launching_distributed.rst

@@ -7,13 +7,106 @@ Launching Distributed Programs
 
 .. currentmodule:: mlx.core.distributed
 
-Installing the MLX python package provides a helper script ``mlx.launch`` that
-can be used to run python scripts distributed on several nodes. It allows
-launching using either the MPI backend or the ring backend. See the
-:doc:`distributed docs <distributed>` for the different backends.
+Installing the MLX python package provides two utilities to help you configure
+your Macs for distributed computation and also launch distributed programs on
+multiple nodes or with many processes in a single node. These utilities are aptly named
 
-Usage
------
+- ``mlx.launch``
+- ``mlx.distributed_config``
+
+See the :doc:`distributed docs <distributed>` for an introduction and
+getting-started guides to the various backends.
+
+``mlx.distributed_config`` 
+---------------------------
+
+Unless you are launching distributed jobs locally for development or multi-gpu
+CUDA environments, then you have several Macs that you need to configure for
+distributed communication with MLX.
+
+``mlx.distributed_config`` aims to automate the process of configuring the
+network interfaces (especially for communication over thunderbolt) and also
+creating the hostfile to be used with ``mlx.launch``.
+
+We will analyse 3 cases of using ``mlx.distributed_config``
+
+1. RDMA over thunderbolt using JACCL
+2. TCP/IP over thunderbolt using the ring backend
+3. TCP/IP over ethernet using the ring backend
+
+JACCL
+^^^^^^^
+
+After following :ref:`the steps to enable RDMA <jaccl_section>` you can run the
+following command to configure the nodes and create the hostfile.
+
+.. code-block::
+
+   mlx.distributed_config --verbose --backend jaccl \
+        --hosts m3-ultra-1,m3-ultra-2,m3-ultra-3,m3-ultra-4 --over thunderbolt \
+        --auto-setup --output m3-ultra-jaccl.json
+
+Let's walk through the steps that the script takes to configure the nodes.
+
+1. Ssh to all nodes to verify that they are reachable
+2. Extract the thunderbolt connectivity. Namely run commands on each node to
+   calculate which node is connected to which other node.
+3. Verify that we have a valid fully connected mesh
+4. Check that RDMA is enabled
+5. Extract the ethernet IP from interface en0
+6. Disable the thunderbolt bridge and set up peer to peer networks for each
+   thunderbolt cable
+7. Write the hostfile
+
+Knowing the above steps allows you to manually configure the nodes but also
+debug any configuration issue. For instance changing the Ethernet IP to a
+different interface directly in the config is possible (as long as it is
+reachable from all nodes).
+
+The ``--auto-setup`` argument requires password-less sudo on each node. If it
+isn't available then the configuration script will print commands to be run on
+each node.
+
+Ring over thunderbolt
+^^^^^^^^^^^^^^^^^^^^^
+
+Setting up a ring backend over thunderbolt only requires changing the
+``--backend`` from ``jaccl`` to ``ring``.
+
+The steps are very similar with the main difference being that instead of
+verifying that the nodes are fully connected, the script attempts to identify a
+ring topology (or multiple rings).
+
+Ring over Ethernet
+^^^^^^^^^^^^^^^^^^
+
+Configuring the ring backend over ethernet doesn't require setting up network
+interface and as such it simply extracts the ``en0`` IP from each node and
+writes the hostfile.
+
+Debugging cable connections
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+``mlx.distributed_config`` can help you debug the connectivity of your nodes
+over thunderbolt by exporting a graph of the connections.
+
+Running
+
+.. code-block::
+
+   mlx.distributed_config --verbose \
+        --hosts host1,host2,host3,host4 \
+        --over thunderbolt --dot
+
+will export a `GraphViz <https://graphviz.org>`_ representation of the
+connections between the nodes which makes it very easy to figure out which
+cable is not connected correctly.
+
+See :ref:`the JACCL section <jaccl_section>` for an example.
+
+
+``mlx.launch``
+--------------
 
 The minimal usage example of ``mlx.launch`` is simply
 
@@ -33,6 +126,10 @@ the rest if one of them fails unexpectedly or if ``mlx.launch`` is terminated.
 It also takes care of forwarding the output of each remote process to stdout
 and stderr respectively.
 
+Importantly, it also broadcasts stdin to each process which enables interactive
+programs to work in distributed mode as well as debugging using the interactive
+debugger.
+
 Providing Hosts
 ^^^^^^^^^^^^^^^^
 
@@ -63,10 +160,62 @@ host and on the same path. A good checklist to debug errors is the following:
   ``mlx.launch --print-python`` to see what that path is.
 * the script you want to run is available on all hosts at the same path
 
+If you are launching from a node with a completely different setup than the
+nodes that the program will run on, you can specify ``--no-verify-script`` so
+that ``mlx.launch`` does not attempt to verify that the executable and script
+exist locally before launching the distributed job.
+
+.. _ring_specifics:
+
+Ring Specifics
+^^^^^^^^^^^^^^
+
+The :ref:`ring <ring_section>` backend, which is also the default
+backend, can be explicitly selected with the argument ``--backend ring``. The
+ring backend has some specific requirements and arguments that are different to
+other backends:
+
+* The argument ``--hosts`` only accepts IPs and not hostnames. If we need to
+  ssh to a hostname that does not correspond to the IP we want to bind to we
+  have to provide a hostfile.
+* ``--starting-port`` defines the port to bind to on the remote hosts.
+  Specifically rank 0 for the first IP will use this port and each subsequent
+  IP or rank will add 1 to this port.
+* ``--connections-per-ip`` allows us to increase the number of connections
+  between neighboring nodes. This corresponds to ``--mca btl_tcp_links 2`` for
+  ``mpirun``.
+
+.. _jaccl_specifics:
+
+JACCL Specifics
+^^^^^^^^^^^^^^^^
+
+The :ref:`JACCL <jaccl_section>` backend can be selected with the argument
+``--backend jaccl``. A hostfile is necessary to launch with this backend
+because it needs to contain the RDMA devices connecting each node to each other
+node.
+
+NCCL Specifics
+^^^^^^^^^^^^^^
+
+The :ref:`NCCL <nccl_section>` backend is the default backend for CUDA
+environments. When launching from a Mac to a Linux machine with CUDA then the
+backend should be selected using ``--backend nccl``.
+
+The ``--repeat-hosts, -n`` argument should be used to launch multi-node and
+multi-gpu jobs. For instance
+
+.. code-block::
+
+   mlx.launch --backend nccl --hosts linux-1,linux-2 -n 8 --no-verify-script -- ./my-job.sh
+
+will attempt to launch 16 processes, 8 on each node that will all run
+``my-job.sh``.
+
 .. _mpi_specifics:
 
 MPI Specifics
--------------
+^^^^^^^^^^^^^
 
 One can use MPI by passing ``--backend mpi`` to ``mlx.launch``. In that case,
 ``mlx.launch`` is a thin wrapper over ``mpirun``. Moreover,
@@ -83,23 +232,3 @@ to choose a specific interface for the byte-transfer-layer of MPI we can call
 .. code:: shell
 
     mlx.launch --backend mpi --mpi-arg '--mca btl_tcp_if_include en0' --hostfile hosts.json my_script.py
-
-
-.. _ring_specifics:
-
-Ring Specifics
---------------
-
-The ring backend, which is also the default backend, can be explicitly selected
-with the argument ``--backend ring``. The ring backend has some specific
-requirements and arguments that are different to MPI:
-
-* The argument ``--hosts`` only accepts IPs and not hostnames. If we need to
-  ssh to a hostname that does not correspond to the IP we want to bind to we
-  have to provide a hostfile.
-* ``--starting-port`` defines the port to bind to on the remote hosts.
-  Specifically rank 0 for the first IP will use this port and each subsequent
-  IP or rank will add 1 to this port.
-* ``--connections-per-ip`` allows us to increase the number of connections
-  between neighboring nodes. This corresponds to ``--mca btl_tcp_links 2`` for
-  ``mpirun``.

mlx/CMakeLists.txt

@@ -1,7 +1,6 @@
 target_sources(
   mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/array.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/array.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/compile.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/dtype.cpp

mlx/allocator.cpp

@@ -1,24 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cstdlib>
-#include <sstream>
-
-#include "mlx/allocator.h"
-
-namespace mlx::core::allocator {
-
-Buffer malloc(size_t size) {
-  auto buffer = allocator().malloc(size);
-  if (size && !buffer.ptr()) {
-    std::ostringstream msg;
-    msg << "[malloc] Unable to allocate " << size << " bytes.";
-    throw std::runtime_error(msg.str());
-  }
-  return buffer;
-}
-
-void free(Buffer buffer) {
-  allocator().free(buffer);
-}
-
-} // namespace mlx::core::allocator

mlx/allocator.h

@@ -28,16 +28,16 @@ class Buffer {
   };
 };
 
-Buffer malloc(size_t size);
-
-void free(Buffer buffer);
-
 class Allocator {
   /** Abstract base class for a memory allocator. */
  public:
   virtual Buffer malloc(size_t size) = 0;
   virtual void free(Buffer buffer) = 0;
   virtual size_t size(Buffer buffer) const = 0;
+  virtual Buffer make_buffer(void* ptr, size_t size) {
+    return Buffer{nullptr};
+  };
+  virtual void release(Buffer buffer) {}
 
   Allocator() = default;
   Allocator(const Allocator& other) = delete;
@@ -49,4 +49,25 @@ class Allocator {
 
 Allocator& allocator();
 
+inline Buffer malloc(size_t size) {
+  return allocator().malloc(size);
+}
+
+inline void free(Buffer buffer) {
+  allocator().free(buffer);
+}
+
+// Make a Buffer from a raw pointer of the given size without a copy.  If a
+// no-copy conversion is not possible then the returned buffer.ptr() will be
+// nullptr. Any buffer created with this function must be released with
+// release(buffer)
+inline Buffer make_buffer(void* ptr, size_t size) {
+  return allocator().make_buffer(ptr, size);
+};
+
+// Release a buffer from the allocator made with make_buffer
+inline void release(Buffer buffer) {
+  allocator().release(buffer);
+}
+
 } // namespace mlx::core::allocator

mlx/array.cpp

@@ -82,6 +82,28 @@ array::array(std::initializer_list<int> data, Dtype dtype)
   init(data.begin());
 }
 
+array::array(
+    void* data,
+    Shape shape,
+    Dtype dtype,
+    const std::function<void(void*)>& deleter)
+    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
+  auto buffer = allocator::make_buffer(data, nbytes());
+  if (buffer.ptr() == nullptr) {
+    set_data(allocator::malloc(nbytes()));
+    auto ptr = static_cast<char*>(data);
+    std::copy(ptr, ptr + nbytes(), this->data<char>());
+    deleter(data);
+  } else {
+    auto wrapped_deleter = [deleter](allocator::Buffer buffer) {
+      auto ptr = buffer.ptr();
+      allocator::release(buffer);
+      return deleter(ptr);
+    };
+    set_data(buffer, std::move(wrapped_deleter));
+  }
+}
+
 /* Build an array from a shared buffer */
 array::array(allocator::Buffer data, Shape shape, Dtype dtype, Deleter deleter)
     : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {

mlx/array.h

@@ -57,6 +57,16 @@ class array {
       Shape shape,
       Dtype dtype = TypeToDtype<T>());
 
+  /* Build an array from a raw pointer. The constructor will attempt to use the
+   * input data without a copy. The deleter will be called when the array no
+   * longer needs the underlying memory - after the array is destroyed in the
+   * no-copy case and after the copy otherwise. */
+  explicit array(
+      void* data,
+      Shape shape,
+      Dtype dtype,
+      const std::function<void(void*)>& deleter);
+
   /* Build an array from a buffer */
   explicit array(
       allocator::Buffer data,

mlx/backend/cpu/eig.cpp

@@ -12,6 +12,167 @@ namespace mlx::core {
 
 namespace {
 
+template <typename T>
+complex64_t to_complex(T r, T i) {
+  return {static_cast<float>(r), static_cast<float>(i)};
+}
+
+template <typename T, class Enable = void>
+struct EigWork {};
+
+template <typename T>
+struct EigWork<
+    T,
+    typename std::enable_if<std::is_floating_point<T>::value>::type> {
+  using O = complex64_t;
+
+  char jobl;
+  char jobr;
+  int N;
+  int lwork;
+  int info;
+  std::vector<array::Data> buffers;
+
+  EigWork(char jobl_, char jobr_, int N_, bool compute_eigenvectors)
+      : jobl(jobl_), jobr(jobr_), N(N_), lwork(-1) {
+    T work;
+    int n_vecs_l = compute_eigenvectors ? N_ : 1;
+    int n_vecs_r = 1;
+    geev<T>(
+        &jobl,
+        &jobr,
+        &N,
+        nullptr,
+        &N,
+        nullptr,
+        nullptr,
+        nullptr,
+        &n_vecs_l,
+        nullptr,
+        &n_vecs_r,
+        &work,
+        &lwork,
+        &info);
+    lwork = static_cast<int>(work);
+
+    buffers.emplace_back(allocator::malloc(sizeof(T) * N * 2));
+    if (compute_eigenvectors) {
+      buffers.emplace_back(allocator::malloc(sizeof(T) * N * N * 2));
+    }
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+  }
+
+  void run(T* a, O* values, O* vectors) {
+    auto eig_tmp = static_cast<T*>(buffers[0].buffer.raw_ptr());
+    T* vec_tmp = nullptr;
+    if (vectors) {
+      vec_tmp = static_cast<T*>(buffers[1].buffer.raw_ptr());
+    }
+    auto work = static_cast<T*>(buffers.back().buffer.raw_ptr());
+
+    int n_vecs_l = vectors ? N : 1;
+    int n_vecs_r = 1;
+    geev<T>(
+        &jobl,
+        &jobr,
+        &N,
+        a,
+        &N,
+        eig_tmp,
+        eig_tmp + N,
+        vectors ? vec_tmp : nullptr,
+        &n_vecs_l,
+        nullptr,
+        &n_vecs_r,
+        work,
+        &lwork,
+        &info);
+
+    for (int i = 0; i < N; ++i) {
+      values[i] = to_complex(eig_tmp[i], eig_tmp[N + i]);
+    }
+
+    if (vectors) {
+      for (int i = 0; i < N; ++i) {
+        if (values[i].imag() != 0) {
+          for (int j = 0; j < N; ++j) {
+            vectors[i * N + j] =
+                to_complex(vec_tmp[i * N + j], -vec_tmp[(i + 1) * N + j]);
+            vectors[(i + 1) * N + j] =
+                to_complex(vec_tmp[i * N + j], vec_tmp[(i + 1) * N + j]);
+          }
+          i += 1;
+        } else {
+          for (int j = 0; j < N; ++j) {
+            vectors[i * N + j] = to_complex(vec_tmp[i * N + j], T(0.0));
+          }
+        }
+      }
+    }
+  }
+};
+
+template <>
+struct EigWork<std::complex<float>> {
+  using T = std::complex<float>;
+  using R = float;
+  using O = T;
+
+  char jobl;
+  char jobr;
+  int N;
+  int lwork;
+  int lrwork;
+  int info;
+  std::vector<array::Data> buffers;
+
+  EigWork(char jobl_, char jobr_, int N_, bool compute_eigenvectors)
+      : jobl(jobl_), jobr(jobr_), N(N_), lwork(-1), lrwork(2 * N_) {
+    T work;
+    R rwork;
+    int n_vecs_l = compute_eigenvectors ? N_ : 1;
+    int n_vecs_r = 1;
+    geev<T>(
+        &jobl,
+        &jobr,
+        &N,
+        nullptr,
+        &N,
+        nullptr,
+        nullptr,
+        &n_vecs_l,
+        nullptr,
+        &n_vecs_r,
+        &work,
+        &lwork,
+        &rwork,
+        &info);
+    lwork = static_cast<int>(work.real());
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+    buffers.emplace_back(allocator::malloc(sizeof(R) * lrwork));
+  }
+
+  void run(T* a, T* values, T* vectors) {
+    int n_vecs_l = vectors ? N : 1;
+    int n_vecs_r = 1;
+    geev<T>(
+        &jobl,
+        &jobr,
+        &N,
+        a,
+        &N,
+        values,
+        vectors,
+        &n_vecs_l,
+        nullptr,
+        &n_vecs_r,
+        static_cast<T*>(buffers[0].buffer.raw_ptr()),
+        &lwork,
+        static_cast<R*>(buffers[1].buffer.raw_ptr()),
+        &info);
+  }
+};
+
 template <typename T>
 void eig_impl(
     array& a,
@@ -19,101 +180,39 @@ void eig_impl(
     array& values,
     bool compute_eigenvectors,
     Stream stream) {
-  using OT = std::complex<T>;
   auto a_ptr = a.data<T>();
-  auto eig_ptr = values.data<OT>();
+  auto val_ptr = values.data<complex64_t>();
 
   auto& encoder = cpu::get_command_encoder(stream);
   encoder.set_input_array(a);
   encoder.set_output_array(values);
-  OT* vec_ptr = nullptr;
+  complex64_t* vec_ptr = nullptr;
   if (compute_eigenvectors) {
     encoder.set_output_array(vectors);
-    vec_ptr = vectors.data<OT>();
+    vec_ptr = vectors.data<complex64_t>();
   }
   encoder.dispatch([a_ptr,
+                    val_ptr,
                     vec_ptr,
-                    eig_ptr,
                     compute_eigenvectors,
                     N = vectors.shape(-1),
                     size = vectors.size()]() mutable {
-    // Work query
     char jobr = 'N';
     char jobl = compute_eigenvectors ? 'V' : 'N';
-    int n_vecs_r = 1;
-    int n_vecs_l = compute_eigenvectors ? N : 1;
-    int lwork = -1;
-    int info;
-    {
-      T work;
-      geev<T>(
-          &jobl,
-          &jobr,
-          &N,
-          nullptr,
-          &N,
-          nullptr,
-          nullptr,
-          nullptr,
-          &n_vecs_l,
-          nullptr,
-          &n_vecs_r,
-          &work,
-          &lwork,
-          &info);
-      lwork = static_cast<int>(work);
-    }
 
-    auto eig_tmp_data = array::Data{allocator::malloc(sizeof(T) * N * 2)};
-    auto vec_tmp_data =
-        array::Data{allocator::malloc(vec_ptr ? sizeof(T) * N * N * 2 : 0)};
-    auto eig_tmp = static_cast<T*>(eig_tmp_data.buffer.raw_ptr());
-    auto vec_tmp = static_cast<T*>(vec_tmp_data.buffer.raw_ptr());
-    auto work_buf = array::Data{allocator::malloc(sizeof(T) * lwork)};
+    EigWork<T> work(jobl, jobr, N, compute_eigenvectors);
+
     for (size_t i = 0; i < size / (N * N); ++i) {
-      geev<T>(
-          &jobl,
-          &jobr,
-          &N,
-          a_ptr,
-          &N,
-          eig_tmp,
-          eig_tmp + N,
-          vec_tmp,
-          &n_vecs_l,
-          nullptr,
-          &n_vecs_r,
-          static_cast<T*>(work_buf.buffer.raw_ptr()),
-          &lwork,
-          &info);
-      for (int i = 0; i < N; ++i) {
-        eig_ptr[i] = {eig_tmp[i], eig_tmp[N + i]};
-      }
+      work.run(a_ptr, val_ptr, vec_ptr);
+      a_ptr += N * N;
+      val_ptr += N;
       if (vec_ptr) {
-        for (int i = 0; i < N; ++i) {
-          if (eig_ptr[i].imag() != 0) {
-            // This vector and the next are a pair
-            for (int j = 0; j < N; ++j) {
-              vec_ptr[i * N + j] = {
-                  vec_tmp[i * N + j], -vec_tmp[(i + 1) * N + j]};
-              vec_ptr[(i + 1) * N + j] = {
-                  vec_tmp[i * N + j], vec_tmp[(i + 1) * N + j]};
-            }
-            i += 1;
-          } else {
-            for (int j = 0; j < N; ++j) {
-              vec_ptr[i * N + j] = {vec_tmp[i * N + j], 0};
-            }
-          }
-        }
         vec_ptr += N * N;
       }
-      a_ptr += N * N;
-      eig_ptr += N;
-      if (info != 0) {
+      if (work.info != 0) {
         std::stringstream msg;
         msg << "[Eig::eval_cpu] Eigenvalue decomposition failed with error code "
-            << info;
+            << work.info;
         throw std::runtime_error(msg.str());
       }
     }
@@ -165,8 +264,17 @@ void Eig::eval_cpu(
     case float32:
       eig_impl<float>(a_copy, vectors, values, compute_eigenvectors_, stream());
       break;
+    case float64:
+      eig_impl<double>(
+          a_copy, vectors, values, compute_eigenvectors_, stream());
+      break;
+    case complex64:
+      eig_impl<std::complex<float>>(
+          a_copy, vectors, values, compute_eigenvectors_, stream());
+      break;
     default:
-      throw std::runtime_error("[Eig::eval_cpu] only supports float32.");
+      throw std::runtime_error(
+          "[Eig::eval_cpu] only supports float32, float64, or complex64.");
   }
 }
 

mlx/backend/cpu/indexing.cpp

@@ -747,4 +747,108 @@ void ScatterAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
   });
 }
 
+template <typename T>
+void masked_scatter_impl(const array& mask, const array& src, array& out) {
+  ContiguousIterator mask_it(mask);
+  ContiguousIterator src_it(src);
+  ContiguousIterator out_it(out);
+
+  const bool* mask_ptr = mask.data<bool>();
+  const T* src_ptr = src.data<T>();
+  T* dst_ptr = out.data<T>();
+
+  const size_t batch_count = mask.shape(0);
+  const size_t mask_batch_size = mask.size() / batch_count;
+  const size_t src_batch_size = src.size() / batch_count;
+
+  for (uint b = 0; b < batch_count; ++b) {
+    size_t src_consumed = 0;
+    src_it.seek(b * src_batch_size);
+
+    for (size_t i = 0; i < mask_batch_size; ++i) {
+      if (mask_ptr[mask_it.loc]) {
+        if (src_consumed >= src_batch_size) {
+          throw std::runtime_error(
+              "[MaskedScatter::eval_cpu] Source does not have enough elements for mask.");
+        }
+        dst_ptr[out_it.loc] = src_ptr[src_it.loc];
+        src_it.step();
+        ++src_consumed;
+      }
+      mask_it.step();
+      out_it.step();
+    }
+  }
+}
+
+void MaskedScatter::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 3);
+
+  auto& dst = inputs[0];
+  auto& mask = inputs[1];
+  auto& src = inputs[2];
+
+  // Copy src into out (copy allocates memory for out)
+  auto ctype =
+      dst.flags().row_contiguous ? CopyType::Vector : CopyType::General;
+  copy_cpu(dst, out, ctype, stream());
+
+  if (mask.size() == 0) {
+    return;
+  }
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(mask);
+  encoder.set_input_array(src);
+  encoder.set_output_array(out);
+  encoder.dispatch([mask = array::unsafe_weak_copy(mask),
+                    src = array::unsafe_weak_copy(src),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        masked_scatter_impl<bool>(mask, src, out);
+        break;
+      case uint8:
+        masked_scatter_impl<uint8_t>(mask, src, out);
+        break;
+      case uint16:
+        masked_scatter_impl<uint16_t>(mask, src, out);
+        break;
+      case uint32:
+        masked_scatter_impl<uint32_t>(mask, src, out);
+        break;
+      case uint64:
+        masked_scatter_impl<uint64_t>(mask, src, out);
+        break;
+      case int8:
+        masked_scatter_impl<int8_t>(mask, src, out);
+        break;
+      case int16:
+        masked_scatter_impl<int16_t>(mask, src, out);
+        break;
+      case int32:
+        masked_scatter_impl<int32_t>(mask, src, out);
+        break;
+      case int64:
+        masked_scatter_impl<int64_t>(mask, src, out);
+        break;
+      case float16:
+        masked_scatter_impl<float16_t>(mask, src, out);
+        break;
+      case float32:
+        masked_scatter_impl<float>(mask, src, out);
+        break;
+      case float64:
+        masked_scatter_impl<double>(mask, src, out);
+        break;
+      case bfloat16:
+        masked_scatter_impl<bfloat16_t>(mask, src, out);
+        break;
+      case complex64:
+        masked_scatter_impl<complex64_t>(mask, src, out);
+        break;
+    }
+  });
+}
+
 } // namespace mlx::core

mlx/backend/cpu/lapack.h

@@ -45,9 +45,7 @@
 INSTANTIATE_LAPACK_REAL(geqrf)
 INSTANTIATE_LAPACK_REAL(orgqr)
 INSTANTIATE_LAPACK_REAL(syevd)
-INSTANTIATE_LAPACK_REAL(geev)
 INSTANTIATE_LAPACK_REAL(potrf)
-INSTANTIATE_LAPACK_REAL(gesdd)
 INSTANTIATE_LAPACK_REAL(getrf)
 INSTANTIATE_LAPACK_REAL(getri)
 INSTANTIATE_LAPACK_REAL(trtri)
@@ -63,3 +61,20 @@ INSTANTIATE_LAPACK_REAL(trtri)
   }
 
 INSTANTIATE_LAPACK_COMPLEX(heevd)
+
+#define INSTANTIATE_LAPACK_ALL(FUNC)                                \
+  template <typename T, typename... Args>                           \
+  void FUNC(Args... args) {                                         \
+    if constexpr (std::is_same_v<T, float>) {                       \
+      MLX_LAPACK_FUNC(s##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, double>) {               \
+      MLX_LAPACK_FUNC(d##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, std::complex<float>>) {  \
+      MLX_LAPACK_FUNC(c##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, std::complex<double>>) { \
+      MLX_LAPACK_FUNC(z##FUNC)(std::forward<Args>(args)...);        \
+    }                                                               \
+  }
+
+INSTANTIATE_LAPACK_ALL(geev)
+INSTANTIATE_LAPACK_ALL(gesdd)

mlx/backend/cpu/svd.cpp

@@ -8,6 +8,183 @@
 
 namespace mlx::core {
 
+template <typename T, class Enable = void>
+struct SVDWork {};
+
+template <typename T>
+struct SVDWork<
+    T,
+    typename std::enable_if<std::is_floating_point<T>::value>::type> {
+  using R = T;
+
+  int N;
+  int M;
+  int K;
+  int lda;
+  int ldu;
+  int ldvt;
+  char jobz;
+  std::vector<array::Data> buffers;
+  int lwork;
+
+  SVDWork(int N, int M, int K, char jobz)
+      : N(N), M(M), K(K), lda(N), ldu(N), ldvt(M), jobz(jobz) {
+    T workspace_dimension = 0;
+
+    // Will contain the indices of eigenvectors that failed to converge (not
+    // used here but required by lapack).
+    buffers.emplace_back(allocator::malloc(sizeof(int) * 8 * K));
+
+    int lwork_query = -1;
+    int info;
+
+    // Compute workspace size.
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ nullptr,
+        /* lda = */ &lda,
+        /* s = */ nullptr,
+        /* u = */ nullptr,
+        /* ldu = */ &ldu,
+        /* vt = */ nullptr,
+        /* ldvt = */ &ldvt,
+        /* work = */ &workspace_dimension,
+        /* lwork = */ &lwork_query,
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    lwork = workspace_dimension;
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+  }
+
+  void run(T* a, R* s, T* u, T* vt) {
+    int info;
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ a,
+        /* lda = */ &lda,
+        /* s = */ s,
+        // According to the identity above, lapack will write Vᵀᵀ as U.
+        /* u = */ u,
+        /* ldu = */ &ldu,
+        // According to the identity above, lapack will write Uᵀ as Vᵀ.
+        /* vt = */ vt,
+        /* ldvt = */ &ldvt,
+        /* work = */ static_cast<T*>(buffers[1].buffer.raw_ptr()),
+        /* lwork = */ &lwork,
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "svd_impl: sgesvdx_ failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+  }
+};
+
+template <>
+struct SVDWork<std::complex<float>> {
+  using T = std::complex<float>;
+  using R = float;
+
+  int N;
+  int M;
+  int K;
+  int lda;
+  int ldu;
+  int ldvt;
+  char jobz;
+  std::vector<array::Data> buffers;
+  int lwork;
+
+  SVDWork(int N, int M, int K, char jobz)
+      : N(N), M(M), K(K), lda(N), ldu(N), ldvt(M), jobz(jobz) {
+    T workspace_dimension = 0;
+
+    // Will contain the indices of eigenvectors that failed to converge (not
+    // used here but required by lapack).
+    buffers.emplace_back(allocator::malloc(sizeof(int) * 8 * K));
+
+    const int lrwork =
+        jobz == 'A' ? std::max(1, 5 * K * K + 5 * K) : std::max(1, 7 * K);
+    buffers.emplace_back(allocator::malloc(sizeof(float) * lrwork));
+
+    int lwork_query = -1;
+    int work_query = -1;
+    int info;
+
+    // Compute workspace size.
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ nullptr,
+        /* lda = */ &lda,
+        /* s = */ nullptr,
+        /* u = */ nullptr,
+        /* ldu = */ &ldu,
+        /* vt = */ nullptr,
+        /* ldvt = */ &ldvt,
+        /* work = */ &workspace_dimension,
+        /* lwork = */ &lwork_query,
+        /* rwork = */ static_cast<float*>(buffers[1].buffer.raw_ptr()),
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    lwork = workspace_dimension.real();
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+  }
+
+  void run(T* a, R* s, T* u, T* vt) {
+    int info;
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ a,
+        /* lda = */ &lda,
+        /* s = */ s,
+        // According to the identity above, lapack will write Vᵀᵀ as U.
+        /* u = */ u,
+        /* ldu = */ &ldu,
+        // According to the identity above, lapack will write Uᵀ as Vᵀ.
+        /* vt = */ vt,
+        /* ldvt = */ &ldvt,
+        /* work = */ static_cast<T*>(buffers[2].buffer.raw_ptr()),
+        /* lwork = */ &lwork,
+        /* rwork = */ static_cast<float*>(buffers[1].buffer.raw_ptr()),
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "svd_impl: sgesvdx_ failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+  }
+};
+
 template <typename T>
 void svd_impl(
     const array& a,
@@ -27,6 +204,8 @@ void svd_impl(
   const int N = a.shape(-1);
   const int K = std::min(M, N);
 
+  using R = typename SVDWork<T>::R;
+
   size_t num_matrices = a.size() / (M * N);
 
   // lapack clobbers the input, so we have to make a copy.
@@ -42,7 +221,7 @@ void svd_impl(
   encoder.set_input_array(a);
   auto in_ptr = in.data<T>();
   T* u_ptr;
-  T* s_ptr;
+  R* s_ptr;
   T* vt_ptr;
 
   if (compute_uv) {
@@ -58,7 +237,7 @@ void svd_impl(
     encoder.set_output_array(s);
     encoder.set_output_array(vt);
 
-    s_ptr = s.data<T>();
+    s_ptr = s.data<R>();
     u_ptr = u.data<T>();
     vt_ptr = vt.data<T>();
   } else {
@@ -68,96 +247,26 @@ void svd_impl(
 
     encoder.set_output_array(s);
 
-    s_ptr = s.data<T>();
+    s_ptr = s.data<R>();
     u_ptr = nullptr;
     vt_ptr = nullptr;
   }
 
   encoder.dispatch([in_ptr, u_ptr, s_ptr, vt_ptr, M, N, K, num_matrices]() {
-    // A of shape M x N. The leading dimension is N since lapack receives Aᵀ.
-    const int lda = N;
-    // U of shape M x M. (N x N in lapack).
-    const int ldu = N;
-    // Vᵀ of shape N x N. (M x M in lapack).
-    const int ldvt = M;
-
-    auto jobz = (u_ptr) ? "A" : "N";
-
-    T workspace_dimension = 0;
-
-    // Will contain the indices of eigenvectors that failed to converge (not
-    // used here but required by lapack).
-    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
-
-    static const int lwork_query = -1;
-
-    int info;
-
-    // Compute workspace size.
-    gesdd<T>(
-        /* jobz = */ jobz,
-        // M and N are swapped since lapack expects column-major.
-        /* m = */ &N,
-        /* n = */ &M,
-        /* a = */ nullptr,
-        /* lda = */ &lda,
-        /* s = */ nullptr,
-        /* u = */ nullptr,
-        /* ldu = */ &ldu,
-        /* vt = */ nullptr,
-        /* ldvt = */ &ldvt,
-        /* work = */ &workspace_dimension,
-        /* lwork = */ &lwork_query,
-        /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
-      throw std::runtime_error(ss.str());
-    }
-
-    const int lwork = workspace_dimension;
-    auto scratch = array::Data{allocator::malloc(sizeof(T) * lwork)};
-
+    auto jobz = (u_ptr) ? 'A' : 'N';
+    SVDWork<T> svd_work(N, M, K, jobz);
     // Loop over matrices.
     for (int i = 0; i < num_matrices; i++) {
-      gesdd<T>(
-          /* jobz = */ jobz,
-          // M and N are swapped since lapack expects column-major.
-          /* m = */ &N,
-          /* n = */ &M,
-          /* a = */ in_ptr + M * N * i,
-          /* lda = */ &lda,
-          /* s = */ s_ptr + K * i,
-          // According to the identity above, lapack will write Vᵀᵀ as U.
-          /* u = */ vt_ptr ? vt_ptr + N * N * i : nullptr,
-          /* ldu = */ &ldu,
-          // According to the identity above, lapack will write Uᵀ as Vᵀ.
-          /* vt = */ u_ptr ? u_ptr + M * M * i : nullptr,
-          /* ldvt = */ &ldvt,
-          /* work = */ static_cast<T*>(scratch.buffer.raw_ptr()),
-          /* lwork = */ &lwork,
-          /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-          /* info = */ &info);
-
-      if (info != 0) {
-        std::stringstream ss;
-        ss << "svd_impl: sgesvdx_ failed with code " << info;
-        throw std::runtime_error(ss.str());
-      }
+      svd_work.run(
+          in_ptr + M * N * i,
+          s_ptr + K * i,
+          vt_ptr ? vt_ptr + N * N * i : nullptr,
+          u_ptr ? u_ptr + M * M * i : nullptr);
     }
   });
   encoder.add_temporary(in);
 }
 
-template <typename T>
-void compute_svd(
-    const array& a,
-    bool compute_uv,
-    std::vector<array>& outputs,
-    Stream stream) {}
-
 void SVD::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
@@ -168,9 +277,12 @@ void SVD::eval_cpu(
     case float64:
       svd_impl<double>(inputs[0], outputs, compute_uv_, stream());
       break;
+    case complex64:
+      svd_impl<std::complex<float>>(inputs[0], outputs, compute_uv_, stream());
+      break;
     default:
       throw std::runtime_error(
-          "[SVD::eval_cpu] only supports float32 or float64.");
+          "[SVD::eval_cpu] only supports float32, float64, or complex64.");
   }
 }
 

mlx/backend/cuda/CMakeLists.txt

@@ -123,14 +123,21 @@ if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
     mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--compress-mode=size>")
 endif()
 
-# Compute capability >= 7.0 is required for synchronization between CPU/GPU with
-# managed memory.
+# Use native CUDA arch by default.
 if(NOT DEFINED MLX_CUDA_ARCHITECTURES)
   execute_process(
-    COMMAND bash detect_cuda_arch.sh
-    WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+    COMMAND __nvcc_device_query
     OUTPUT_VARIABLE MLX_CUDA_ARCHITECTURES
     OUTPUT_STRIP_TRAILING_WHITESPACE)
+  set(UPGRADABLE_ARCHITECTURES "90;100;121")
+  if(MLX_CUDA_ARCHITECTURES STREQUAL "")
+    message(
+      FATAL_ERROR
+        "Can not get native CUDA arch, must set MLX_CUDA_ARCHITECTURES")
+  elseif(MLX_CUDA_ARCHITECTURES IN_LIST UPGRADABLE_ARCHITECTURES)
+    # Use arch-specific compute capability whenever possible.
+    set(MLX_CUDA_ARCHITECTURES "${MLX_CUDA_ARCHITECTURES}a")
+  endif()
 endif()
 message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
@@ -142,6 +149,7 @@ FetchContent_Declare(
   URL "https://github.com/NVIDIA/cccl/releases/download/v2.8.1/cccl-v2.8.1.zip")
 FetchContent_MakeAvailable(cccl)
 target_include_directories(mlx BEFORE PRIVATE "${cccl_SOURCE_DIR}/include")
+set_target_properties(mlx PROPERTIES CCCL_DIR "${cccl_SOURCE_DIR}/include")
 
 # Use fixed version of NVTX.
 FetchContent_Declare(
@@ -167,7 +175,7 @@ target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 FetchContent_Declare(
   cudnn
   GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
-  GIT_TAG v1.14.0
+  GIT_TAG v1.16.0
   GIT_SHALLOW TRUE
   EXCLUDE_FROM_ALL)
 set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)

mlx/backend/cuda/allocator.cpp

@@ -20,6 +20,19 @@ constexpr int page_size = 16384;
 // Any allocations smaller than this will try to use the small pool
 constexpr int small_block_size = 8;
 
+#if CUDART_VERSION >= 13000
+inline cudaMemLocation cuda_mem_loc(int i) {
+  cudaMemLocation loc;
+  loc.type = cudaMemLocationTypeDevice;
+  loc.id = i;
+  return loc;
+}
+#else
+inline int cuda_mem_loc(int i) {
+  return i;
+}
+#endif // CUDART_VERSION >= 13000
+
 // 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;
@@ -35,13 +48,7 @@ SmallSizePool::SmallSizePool() {
   int device_count = 0;
   CHECK_CUDA_ERROR(cudaGetDeviceCount(&device_count));
   for (int i = 0; i < device_count; ++i) {
-#if CUDART_VERSION >= 13000
-    cudaMemLocation loc;
-    loc.type = cudaMemLocationTypeDevice;
-    loc.id = i;
-#else
-    int loc = i;
-#endif // CUDART_VERSION >= 13000
+    auto loc = cuda_mem_loc(i);
     CHECK_CUDA_ERROR(
         cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetAccessedBy, loc));
   }
@@ -90,9 +97,10 @@ CudaAllocator::CudaAllocator()
           page_size,
           [](CudaBuffer* buf) { return buf->size; },
           [this](CudaBuffer* buf) { cuda_free(buf); }) {
-  size_t free, total;
-  CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
-  memory_limit_ = total * 0.95;
+  size_t free;
+  CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total_memory_));
+  memory_limit_ = total_memory_ * 0.95;
+  free_limit_ = total_memory_ - memory_limit_;
   max_pool_size_ = memory_limit_;
 
   int device_count = 0;
@@ -104,6 +112,10 @@ CudaAllocator::CudaAllocator()
     cudaStream_t s;
     CHECK_CUDA_ERROR(cudaStreamCreateWithFlags(&s, cudaStreamNonBlocking));
     free_streams_.push_back(s);
+
+    cudaMemPool_t mem_pool;
+    CHECK_CUDA_ERROR(cudaDeviceGetDefaultMemPool(&mem_pool, i));
+    mem_pools_.push_back(mem_pool);
   }
   CHECK_CUDA_ERROR(cudaSetDevice(curr));
 }
@@ -119,7 +131,8 @@ void copy_to_managed(CudaBuffer& buf) {
   buf.data = new_data;
 }
 
-Buffer CudaAllocator::malloc_impl(size_t size, cudaStream_t stream) {
+Buffer
+CudaAllocator::malloc_async(size_t size, int device, cudaStream_t stream) {
   if (size == 0) {
     return Buffer{new CudaBuffer{nullptr, 0, -1}};
   }
@@ -134,9 +147,8 @@ Buffer CudaAllocator::malloc_impl(size_t size, cudaStream_t stream) {
     size = page_size * ((size + page_size - 1) / page_size);
   }
 
-  int device = -1;
-  if (size > small_block_size && stream != nullptr) {
-    CHECK_CUDA_ERROR(cudaStreamGetDevice(stream, &device));
+  if (size <= small_block_size || stream == nullptr) {
+    device = -1;
   }
 
   CudaBuffer* buf = buffer_cache_.reuse_from_cache(size);
@@ -154,19 +166,35 @@ Buffer CudaAllocator::malloc_impl(size_t size, cudaStream_t stream) {
     }
     lock.unlock();
     if (!buf) {
-      buf = new CudaBuffer{nullptr, size, device};
-      cudaError_t err;
+      void* data = nullptr;
       if (device == -1) {
-        err = cudaMallocManaged(&buf->data, size);
+        CHECK_CUDA_ERROR(cudaMallocManaged(&data, size));
       } else {
-        err = cudaMallocAsync(&buf->data, size, stream);
+        CHECK_CUDA_ERROR(cudaMallocAsync(&data, size, stream));
       }
-      if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
-        throw std::runtime_error(fmt::format(
-            "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+      if (!data) {
+        std::ostringstream msg;
+        msg << "[malloc] Unable to allocate " << size << " bytes.";
+        throw std::runtime_error(msg.str());
       }
+      buf = new CudaBuffer{data, size, device};
     }
     lock.lock();
+
+    // If any cuda memory pool has too much reserved memory, clear some
+    // memory from the cache. This prevents graph / kernel execution failing
+    // from OOM
+    if (get_cache_memory() > 0) {
+      for (auto p : mem_pools_) {
+        size_t used = 0;
+        CHECK_CUDA_ERROR(cudaMemPoolGetAttribute(
+            p, cudaMemPoolAttrReservedMemCurrent, &used));
+        if (used > (total_memory_ - free_limit_)) {
+          buffer_cache_.release_cached_buffers(free_limit_);
+          break;
+        }
+      }
+    }
   }
   active_memory_ += buf->size;
   peak_memory_ = std::max(active_memory_, peak_memory_);
@@ -176,18 +204,14 @@ Buffer CudaAllocator::malloc_impl(size_t size, cudaStream_t stream) {
     buffer_cache_.release_cached_buffers(get_cache_memory() - max_pool_size_);
   }
   // Copy to managed here if the buffer is not on the right device
-  if (buf->device != device) {
+  if (buf->device >= 0 && buf->device != device) {
     copy_to_managed(*buf);
   }
   return Buffer{buf};
 }
 
-Buffer CudaAllocator::malloc_async(size_t size, cudaStream_t stream) {
-  return malloc_impl(size, stream);
-}
-
 Buffer CudaAllocator::malloc(size_t size) {
-  return malloc_impl(size, nullptr);
+  return malloc_async(size, -1, nullptr);
 }
 
 void CudaAllocator::free(Buffer buffer) {
@@ -223,9 +247,9 @@ void CudaAllocator::cuda_free(CudaBuffer* buf) {
     scalar_pool_.free(buf);
   } else {
     if (buf->device >= 0) {
-      cudaFreeAsync(buf->data, free_streams_[buf->device]);
+      CHECK_CUDA_ERROR(cudaFreeAsync(buf->data, free_streams_[buf->device]));
     } else {
-      cudaFree(buf->data);
+      CHECK_CUDA_ERROR(cudaFree(buf->data));
     }
     delete buf;
   }
@@ -277,8 +301,9 @@ CudaAllocator& allocator() {
   return *allocator_;
 }
 
-Buffer malloc_async(size_t size, cudaStream_t stream) {
-  auto buffer = allocator().malloc_async(size, stream);
+Buffer malloc_async(size_t size, CommandEncoder& encoder) {
+  auto buffer = allocator().malloc_async(
+      size, encoder.device().cuda_device(), encoder.stream());
   if (size && !buffer.ptr()) {
     std::ostringstream msg;
     msg << "[malloc_async] Unable to allocate " << size << " bytes.";

mlx/backend/cuda/allocator.h

@@ -13,6 +13,8 @@
 
 namespace mlx::core::cu {
 
+class CommandEncoder;
+
 using allocator::Buffer;
 
 // Stores cuda-managed unified memory.
@@ -48,7 +50,7 @@ class SmallSizePool {
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
-  Buffer malloc_async(size_t size, cudaStream_t stream);
+  Buffer malloc_async(size_t size, int device, cudaStream_t stream);
   void free(Buffer buffer) override;
   size_t size(Buffer buffer) const override;
 
@@ -62,7 +64,6 @@ class CudaAllocator : public allocator::Allocator {
   void clear_cache();
 
  private:
-  Buffer malloc_impl(size_t size, cudaStream_t stream);
   void cuda_free(CudaBuffer* buf);
 
   CudaAllocator();
@@ -70,16 +71,19 @@ class CudaAllocator : public allocator::Allocator {
 
   std::mutex mutex_;
   size_t memory_limit_;
+  size_t free_limit_;
+  size_t total_memory_;
   size_t max_pool_size_;
   BufferCache<CudaBuffer> buffer_cache_;
   size_t active_memory_{0};
   size_t peak_memory_{0};
   std::vector<cudaStream_t> free_streams_;
+  std::vector<cudaMemPool_t> mem_pools_;
   SmallSizePool scalar_pool_;
 };
 
 CudaAllocator& allocator();
 
-Buffer malloc_async(size_t size, cudaStream_t stream);
+Buffer malloc_async(size_t size, CommandEncoder& encoder);
 
 } // namespace mlx::core::cu

mlx/backend/cuda/arange.cu

@@ -42,7 +42,7 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
     return;
   }
   auto& encoder = cu::get_command_encoder(stream());
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   encoder.set_output_array(out);
 
   dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {

mlx/backend/cuda/arg_reduce.cu

@@ -143,7 +143,7 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   // Prepare the shapes, strides and axis arguments.
   Shape shape = remove_index(in.shape(), axis_);

mlx/backend/cuda/binary/binary.cuh

@@ -367,9 +367,8 @@ void binary_op_gpu(
   auto bopt = get_binary_op_type(a, b);
   auto& encoder = cu::get_command_encoder(s);
 
-  set_binary_op_output_data(a, b, out, bopt, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  set_binary_op_output_data(
+      a, b, out, bopt, [&](auto n) { return cu::malloc_async(n, encoder); });
   binary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 

mlx/backend/cuda/binary_two.cu

@@ -246,12 +246,10 @@ void binary_two_op_gpu_inplace(
   auto& out_b = outputs[1];
   auto bopt = get_binary_op_type(a, b);
   auto& encoder = cu::get_command_encoder(s);
-  set_binary_op_output_data(a, b, out_a, bopt, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
-  set_binary_op_output_data(a, b, out_b, bopt, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  set_binary_op_output_data(
+      a, b, out_a, bopt, [&](auto n) { return cu::malloc_async(n, encoder); });
+  set_binary_op_output_data(
+      a, b, out_b, bopt, [&](auto n) { return cu::malloc_async(n, encoder); });
 
   if (out_a.size() == 0) {
     return;

mlx/backend/cuda/compiled.cpp

@@ -298,7 +298,7 @@ void Compiled::eval_gpu(
   // Put outputs.
   compiled_allocate_outputs(
       inputs, outputs, is_constant_, contiguous, [&](auto n) {
-        return cu::malloc_async(n, encoder.stream());
+        return cu::malloc_async(n, encoder);
       });
   for (auto& x : outputs) {
     args.append(x);

mlx/backend/cuda/conv.cpp

@@ -15,19 +15,16 @@ namespace mlx::core {
 
 namespace {
 
-// Alias for better readability.
-#define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
-#define CONV_BACKWARD_INPUT \
-  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_DATA_DESCRIPTOR
-#define CONV_BACKWARD_WEIGHT \
-  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
-
-// Custom placeholder representing fallback kernel.
-#define CONV_FALLBACK static_cast<cudnnBackendDescriptorType_t>(-1)
+enum ConvBackendType {
+  CONV_FALLBACK,
+  CONV_FORWARD,
+  CONV_BACKWARD_INPUT,
+  CONV_BACKWARD_WEIGHT,
+};
 
 struct ConvCacheKey {
   int device_id;
-  cudnnDataType_t cudnn_dtype;
+  fe::DataType_t cudnn_dtype;
   std::array<int, MAX_NDIM> input_shape;
   std::array<int, MAX_NDIM> weight_shape;
   std::array<int, MAX_NDIM> stride;
@@ -44,15 +41,13 @@ struct ConvCacheKey {
 auto& conv_cache() {
   static LRUBytesKeyCache<
       ConvCacheKey,
-      std::pair<
-          cudnnBackendDescriptorType_t,
-          std::optional<cudnn_frontend::ExecutionPlan>>>
+      std::pair<ConvBackendType, std::optional<DnnGraph>>>
       cache("MLX_CUDA_CONV_CACHE_SIZE", /* default_capacity */ 128);
   return cache;
 }
 
-auto get_conv_op_settings(
-    cudnnBackendDescriptorType_t backend_type,
+auto get_conv_settings(
+    ConvBackendType backend_type,
     array& x,
     array& w,
     array& y,
@@ -68,8 +63,8 @@ auto get_conv_op_settings(
     for (int i = 0; i < padding_lo.size(); ++i) {
       int wt_size = 1 + kernel_dilation[i] * (w.shape(1 + i) - 1);
       padding_lo[i] = wt_size - padding_lo[i] - 1;
-      int in_size = 1 + kernel_strides[i] * (x.shape(1 + i) - 1);
-      int out_size = 1 + input_dilation[i] * (y.shape(1 + i) - 1);
+      int in_size = 1 + kernel_strides[i] * (y.shape(1 + i) - 1);
+      int out_size = 1 + input_dilation[i] * (x.shape(1 + i) - 1);
       padding_hi[i] = out_size - in_size + padding_hi[i];
     }
     return std::make_tuple(
@@ -95,49 +90,57 @@ auto get_conv_op_settings(
   }
 }
 
-std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
+std::optional<DnnGraph> build_conv_graph(
     cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
+    ConvBackendType backend_type,
     Dtype dtype,
     array& x,
     array& w,
     array& y,
-    const SmallVector<int64_t>& stride,
-    const SmallVector<int64_t>& padding_lo,
-    const SmallVector<int64_t>& padding_hi,
-    const SmallVector<int64_t>& dilation) {
-  try {
-    auto compute_dtype = (dtype == float16 || dtype == bfloat16)
-        ? CUDNN_DATA_FLOAT
-        : dtype_to_cudnn_type(dtype);
-    auto conv_desc = cudnn_frontend::ConvDescBuilder()
-                         .setDataType(compute_dtype)
-                         .setMathMode(CUDNN_CROSS_CORRELATION)
-                         .setNDims(stride.size())
-                         .setStrides(stride.size(), stride.data())
-                         .setPrePadding(padding_lo.size(), padding_lo.data())
-                         .setPostPadding(padding_hi.size(), padding_hi.data())
-                         .setDilation(dilation.size(), dilation.data())
-                         .build();
+    const std::vector<int64_t>& stride,
+    const std::vector<int64_t>& padding_lo,
+    const std::vector<int64_t>& padding_hi,
+    const std::vector<int64_t>& dilation) {
+  auto compute_dtype =
+      (dtype == float16 || dtype == bfloat16) ? float32 : dtype;
+  DnnGraph graph(encoder.device().cudnn_handle(), dtype, compute_dtype);
+  auto x_ = graph.tensor_nchw("X", 'x', x);
+  auto w_ = graph.tensor_nchw("W", 'w', w);
 
-    auto op = cudnn_frontend::OperationBuilder(backend_type)
-                  .setxDesc(build_cudnn_tensor_nchw('x', x))
-                  .setwDesc(build_cudnn_tensor_nchw('w', w))
-                  .setyDesc(build_cudnn_tensor_nchw('y', y))
-                  .setcDesc(conv_desc)
-                  .build();
+  auto set_options = [&](auto& options) {
+    options.set_compute_data_type(dtype_to_cudnn_type(compute_dtype))
+        .set_convolution_mode(fe::ConvolutionMode_t::CROSS_CORRELATION)
+        .set_stride(stride)
+        .set_pre_padding(padding_lo)
+        .set_post_padding(padding_hi)
+        .set_dilation(dilation);
+  };
 
-    std::array<cudnn_frontend::Operation const*, 1> ops = {&op};
-    return cudnn_frontend::OperationGraphBuilder()
-        .setHandle(encoder.device().cudnn_handle())
-        .setOperationGraph(ops.size(), ops.data())
-        .build();
-  } catch (cudnn_frontend::cudnnException& error) {
-    if (error.getCudnnStatus() != CUDNN_STATUS_BAD_PARAM) {
-      throw;
-    }
+  std::shared_ptr<fe::graph::Tensor_attributes> y_;
+  if (backend_type == CONV_FORWARD) {
+    auto options = fe::graph::Conv_fprop_attributes();
+    set_options(options);
+    y_ = graph.conv_fprop(x_, w_, options);
+  } else if (backend_type == CONV_BACKWARD_INPUT) {
+    auto options = fe::graph::Conv_dgrad_attributes();
+    set_options(options);
+    y_ = graph.conv_dgrad(x_, w_, options);
+  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
+    auto options = fe::graph::Conv_wgrad_attributes();
+    set_options(options);
+    y_ = graph.conv_wgrad(w_, x_, options);
+  }
+  graph.tensor_nchw(y_, 'y', y)->set_output(true);
+
+  if (graph.prepare().is_bad()) {
     return std::nullopt;
   }
+  graph.deselect_numeric_notes({fe::NumericalNote_t::DOWN_CONVERT_INPUTS});
+  if (dtype == float32 && !env::enable_tf32()) {
+    graph.deselect_numeric_notes({fe::NumericalNote_t::TENSOR_CORE});
+  }
+  CHECK_CUDNN_FE_ERROR(graph.build());
+  return graph;
 }
 
 // Transpose from (C_out, H, W, C_in / groups) to (C_in, H, W, C_out / groups).
@@ -181,7 +184,7 @@ array group_transpose(
 // eval_gpu, with cost of possible redundant copies.
 std::tuple<array, array, array> prepare_args(
     cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
+    ConvBackendType backend_type,
     array in,
     array wt,
     array out,
@@ -221,27 +224,11 @@ std::tuple<array, array, array> prepare_args(
   return {std::move(in), std::move(wt), std::move(out)};
 }
 
-// Get the x/w/y args from the in/wt/out args depending on backend type.
-inline std::tuple<array&, array&, array&> dispatch_args(
-    cudnnBackendDescriptorType_t backend_type,
-    array& in,
-    array& wt,
-    array& out) {
-  switch (backend_type) {
-    case CONV_BACKWARD_INPUT:
-      return {out, wt, in};
-    case CONV_BACKWARD_WEIGHT:
-      return {in, out, wt};
-    default:
-      return {in, wt, out};
-  }
-}
-
 // Register inputs and outputs before actually running conv op. Can only be
 // called once per eval_gpu.
 void register_args(
     cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
+    ConvBackendType backend_type,
     array& in,
     array& wt,
     array& intermediate_out,
@@ -277,7 +264,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   array in = inputs[0];
   array wt = inputs[1];
   array out = out_;
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   Dtype dtype = out.dtype();
 
   // Search cache.
@@ -297,16 +284,19 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       get_alignment(wt),
       get_alignment(out)};
   if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
-    auto& [backend_type, plan] = it->second;
-    if (plan) {
-      // Run cached plan.
+    auto& [backend_type, graph] = it->second;
+    if (graph) {
+      // Run cached graph.
       std::tie(in, wt, out) =
           prepare_args(encoder, backend_type, in, wt, out, groups_, s);
       register_args(encoder, backend_type, in, wt, out, out_);
-      auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-      if (!encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
-        throw std::runtime_error("[conv] Cached plan failed to execute.");
-      }
+      CHECK_CUDNN_FE_ERROR(graph->encode_capturing(
+          encoder,
+          {
+              {'x', gpu_ptr<void>(in)},
+              {'w', gpu_ptr<void>(wt)},
+              {'y', gpu_ptr<void>(out)},
+          }));
     } else {
       // Run fallback kernel.
       gemm_conv(
@@ -327,7 +317,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
 
   // There is no reliable way to deduce the proper cuDNN backend for the
   // convolution, so we make a best guess and then try.
-  SmallVector<cudnnBackendDescriptorType_t, 2> try_backends;
+  SmallVector<ConvBackendType, 2> try_backends;
   if (flip_) {
     // When weight is flipped, we assume it is backward input convolution.
     try_backends.push_back(CONV_BACKWARD_INPUT);
@@ -345,13 +335,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   }
 
   // Try to build op graph.
-  cudnnBackendDescriptorType_t backend_type;
-  std::optional<cudnn_frontend::OperationGraph> op_graph;
+  ConvBackendType backend_type;
+  std::optional<DnnGraph> graph;
   for (auto try_backend : try_backends) {
-    auto [in_copy, wt_copy, out_copy] =
+    auto [x, w, y] =
         prepare_args(encoder, try_backend, in, wt, out, groups_, s);
-    auto [x, w, y] = dispatch_args(try_backend, in_copy, wt_copy, out_copy);
-    auto [stride, padding_lo, padding_hi, dilation] = get_conv_op_settings(
+    auto [stride, padding_lo, padding_hi, dilation] = get_conv_settings(
         try_backend,
         x,
         w,
@@ -361,7 +350,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_hi_,
         kernel_dilation_,
         input_dilation_);
-    op_graph = build_conv_op_graph(
+    graph = build_conv_graph(
         encoder,
         try_backend,
         dtype,
@@ -372,30 +361,27 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_lo,
         padding_hi,
         dilation);
-    if (op_graph) {
+    if (graph) {
       backend_type = try_backend;
-      in = std::move(in_copy);
-      wt = std::move(wt_copy);
-      out = std::move(out_copy);
+      in = std::move(x);
+      wt = std::move(w);
+      out = std::move(y);
       break;
     }
   }
 
-  if (op_graph) {
-    // Find a plan for the graph and execute it.
-    auto plan = find_cudnn_plan_from_op_graph(
-        encoder.device().cudnn_handle(), backend_type, dtype, *op_graph);
-    if (plan) {
-      // Setup inputs and outputs.
-      register_args(encoder, backend_type, in, wt, out, out_);
-
-      auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-      if (encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
-        conv_cache().emplace(
-            cache_key, std::make_pair(backend_type, std::move(*plan)));
-        return;
-      }
-    }
+  if (graph) {
+    register_args(encoder, backend_type, in, wt, out, out_);
+    CHECK_CUDNN_FE_ERROR(graph->encode_capturing(
+        encoder,
+        {
+            {'x', gpu_ptr<void>(in)},
+            {'w', gpu_ptr<void>(wt)},
+            {'y', gpu_ptr<void>(out)},
+        }));
+    conv_cache().emplace(
+        cache_key, std::make_pair(backend_type, std::move(*graph)));
+    return;
   }
 
   // Use fallback kernel for settings not supported by cuDNN.

mlx/backend/cuda/conv/gemm_conv.cu

@@ -86,7 +86,7 @@ array unfold_inputs_nd(
     int mat_N,
     ConvParams<NDIM>& params) {
   array unfolded({mat_M, mat_K}, in.dtype(), nullptr, {});
-  unfolded.set_data(cu::malloc_async(unfolded.nbytes(), encoder.stream()));
+  unfolded.set_data(cu::malloc_async(unfolded.nbytes(), encoder));
   encoder.add_temporary(unfolded);
 
   int filter_size = params.C;

mlx/backend/cuda/conv/gemm_grouped_conv.cu

@@ -89,7 +89,7 @@ array grouped_unfold_transpose_inputs_nd(
     int mat_N,
     ConvParams<NDIM>& params) {
   array unfolded({mat_M, mat_K * params.groups}, in.dtype(), nullptr, {});
-  unfolded.set_data(cu::malloc_async(unfolded.nbytes(), encoder.stream()));
+  unfolded.set_data(cu::malloc_async(unfolded.nbytes(), encoder));
   encoder.add_temporary(unfolded);
 
   int filter_size = params.C;

mlx/backend/cuda/copy.cu

@@ -7,9 +7,8 @@ namespace mlx::core {
 
 void copy_gpu(const array& in, array& out, CopyType ctype, const Stream& s) {
   auto& encoder = cu::get_command_encoder(s);
-  bool donated = set_copy_output_data(in, out, ctype, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  bool donated = set_copy_output_data(
+      in, out, ctype, [&](auto n) { return cu::malloc_async(n, encoder); });
   if (donated && in.dtype() == out.dtype()) {
     // If the output has the same type as the input then there is nothing to
     // copy, just use the buffer.
@@ -104,7 +103,7 @@ void fill_gpu(const array& in, array& out, const Stream& s) {
     return;
   }
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   encoder.set_input_array(in);
   encoder.set_output_array(out);
   copy_contiguous(encoder, CopyType::Scalar, in, out, 0, 0);
@@ -114,7 +113,7 @@ void reshape_gpu(const array& in, array& out, Stream s) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
   if (copy_necessary) {
     auto& encoder = cu::get_command_encoder(s);
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
     copy_gpu_inplace(
         in,
         out,

mlx/backend/cuda/copy/copy_general_input.cu

@@ -95,11 +95,14 @@ void copy_general_input(
             const InType* in_ptr = gpu_ptr<InType>(in) + offset_in;
             OutType* out_ptr = gpu_ptr<OutType>(out) + offset_out;
             int ndim = shape.size();
-            int work_per_thread = 1;
+
+            int work_per_thread = 8;
             auto dim0 = ndim > 0 ? shape.back() : 1;
             auto rest = out.size() / dim0;
-            if (dim0 >= 4) {
+            if (dim0 >= 4 && dim0 < 8) {
               work_per_thread = 4;
+            } else if (dim0 < 4) {
+              work_per_thread = 1;
             }
             dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
             auto block_dims = get_block_dims(dim0, rest, 1);
@@ -110,7 +113,10 @@ void copy_general_input(
               dispatch_1_2_3(ndim, [&](auto dims_constant) {
                 auto kernel =
                     cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 1>;
-                if (work_per_thread == 4) {
+                if (work_per_thread == 8) {
+                  kernel =
+                      cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 8>;
+                } else if (work_per_thread == 4) {
                   kernel =
                       cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 4>;
                 }
@@ -127,7 +133,9 @@ void copy_general_input(
               });
             } else { // ndim >= 4
               auto kernel = cu::copy_g<InType, OutType, IdxT, 1>;
-              if (work_per_thread == 4) {
+              if (work_per_thread == 8) {
+                kernel = cu::copy_g<InType, OutType, IdxT, 8>;
+              } else if (work_per_thread == 4) {
                 kernel = cu::copy_g<InType, OutType, IdxT, 4>;
               }
               encoder.add_kernel_node(

mlx/backend/cuda/cuda_utils.h

@@ -5,6 +5,7 @@
 #include <cublasLt.h>
 #include <cuda.h>
 #include <cuda_runtime.h>
+#include <cudnn.h>
 
 namespace mlx::core {
 
@@ -12,10 +13,12 @@ namespace mlx::core {
 void check_cublas_error(const char* name, cublasStatus_t err);
 void check_cuda_error(const char* name, cudaError_t err);
 void check_cuda_error(const char* name, CUresult err);
+void check_cudnn_error(const char* name, cudnnStatus_t err);
 
 // The macro version that prints the command that failed.
 #define CHECK_CUBLAS_ERROR(cmd) check_cublas_error(#cmd, (cmd))
 #define CHECK_CUDA_ERROR(cmd) check_cuda_error(#cmd, (cmd))
+#define CHECK_CUDNN_ERROR(cmd) check_cudnn_error(#cmd, (cmd))
 
 // Base class for RAII managed CUDA resources.
 template <typename Handle, cudaError_t (*Destroy)(Handle)>
@@ -29,6 +32,10 @@ class CudaHandle {
   }
 
   ~CudaHandle() {
+    // Skip if there was an error to avoid throwing in the destructors
+    if (cudaPeekAtLastError() != cudaSuccess) {
+      return;
+    }
     reset();
   }
 

mlx/backend/cuda/cudnn_utils.cpp

@@ -7,32 +7,26 @@ namespace mlx::core {
 
 namespace {
 
-// Create a cudnn tensor descriptor.
-template <typename Vec>
-inline cudnn_frontend::Tensor build_cudnn_tensor(
-    int64_t id,
-    const array& x,
-    const Vec& shape,
-    const Vec& strides) {
-  return cudnn_frontend::TensorBuilder()
-      .setDim(shape.size(), shape.data())
-      .setStrides(strides.size(), strides.data())
-      .setId(id)
-      .setAlignment(get_alignment(x))
-      .setDataType(dtype_to_cudnn_type(x.dtype()))
-      .build();
-}
+#define RETURN_IF_ERROR(cmd)          \
+  if (auto ret = cmd; ret.is_bad()) { \
+    return ret;                       \
+  }
 
 // In MLX a singleton dim (shape[dim] == 1) can have any stride, but in cuDNN
 // whether a tensor is contiguous is determined with:
 // shape[dim] == shape[dim + 1] * strides[dim + 1]
 // So a contiguous array with singleton dims in MLX may be mistakenly treated
 // as strided in cuDNN, and we work around it by normalizing the strides.
-Strides normalized_strides(const array& x) {
-  if (!x.flags().row_contiguous || x.ndim() < 2) {
-    return x.strides();
+std::vector<int64_t> normalized_strides(const array& x) {
+  std::vector<int64_t> strides(x.strides().begin(), x.strides().end());
+  if (std::all_of(
+          strides.begin(), strides.end(), [](int64_t s) { return s == 0; })) {
+    strides.back() = 1;
+    return strides;
+  }
+  if (!x.flags().row_contiguous || x.ndim() < 2) {
+    return strides;
   }
-  Strides strides = x.strides();
   for (int i = x.ndim() - 2; i >= 0; --i) {
     if (x.shape(i) == 1) {
       strides[i] = x.shape(i + 1) * strides[i + 1];
@@ -42,7 +36,9 @@ Strides normalized_strides(const array& x) {
 }
 
 // Return the shape and strides after transposing from NHWC to NCHW.
-auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
+inline auto nhwc_to_nchw(const array& x) {
+  auto shape = convert_vector<int64_t>(x.shape());
+  auto strides = normalized_strides(x);
   assert(shape.size() >= 3);
   shape.insert(shape.begin() + 1, shape.back());
   shape.erase(shape.end() - 1);
@@ -51,228 +47,95 @@ auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
   return std::make_tuple(std::move(shape), std::move(strides));
 }
 
-inline auto nhwc_to_nchw(const array& x) {
-  return nhwc_to_nchw(
-      convert_vector<int64_t>(x.shape()), normalized_strides(x));
-}
-
-// Return available engines for a |op_graph|.
-cudnn_frontend::EngineConfigList get_cudnn_engine_configs(
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph,
-    bool use_fallback = true) {
-  SmallVector<cudnn_frontend::GeneratorSource, 2> sources;
-  sources.push_back([](auto& op_graph) {
-    auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
-                          .setOperationGraph(op_graph)
-                          .setHeurMode(CUDNN_HEUR_MODE_A)
-                          .build();
-    return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
-  });
-  if (use_fallback) {
-    sources.push_back([&backend_type](auto& op_graph) {
-      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
-                          .setOperationGraph(op_graph)
-                          .setOperation(backend_type)
-                          .build();
-      return fallback.getFallbackList();
-    });
-  }
-
-  auto configs =
-      cudnn_frontend::EngineConfigGenerator(sources.size(), sources.data())
-          .generate_engine_config(op_graph);
-
-  cudnn_frontend::EngineConfigList filtered_configs;
-  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
-    if (cudnn_frontend::hasNumericalNote<
-            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
-      return true;
-    }
-    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
-        dtype == float32 && !env::enable_tf32()) {
-      return true;
-    }
-    return false;
-  });
-  return filtered_configs;
-}
-
-// Take |engine_configs| and |op_graph| and find a working execution plans
-// from them.
-std::optional<cudnn_frontend::ExecutionPlan>
-find_cudnn_plan_from_engine_configs(
-    cudnnHandle_t handle,
-    const cudnn_frontend::EngineConfigList& engine_configs,
-    const cudnn_frontend::OperationGraph& op_graph) {
-  auto op_graph_tag = op_graph.getTag();
-  for (const auto& config : engine_configs) {
-    try {
-      return cudnn_frontend::ExecutionPlanBuilder()
-          .setHandle(handle)
-          .setEngineConfig(config, op_graph_tag)
-          .build();
-    } catch (cudnn_frontend::cudnnException& error) {
-      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
-        throw;
-      }
-    }
-  }
-  return std::nullopt;
-}
-
-// Prepare workspace and args to execute plan.
-template <typename F>
-bool prepare_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs,
-    F&& execute) {
-  int workspace_size = plan.getWorkspaceSize();
-  void* workspace_ptr = nullptr;
-  if (workspace_size > 0) {
-    array workspace(
-        cu::malloc_async(workspace_size, encoder.stream()),
-        {workspace_size},
-        uint8);
-    encoder.add_temporary(workspace);
-    workspace_ptr = gpu_ptr<void>(workspace);
-  }
-
-  auto args = cudnn_frontend::VariantPackBuilder()
-                  .setWorkspacePointer(workspace_ptr)
-                  .setDataPointers(num_args, data_ptrs)
-                  .setUids(num_args, uids)
-                  .build();
-
-  auto handle = encoder.device().cudnn_handle();
-  cudnnSetStream(handle, encoder.stream());
-
-  if (!execute(handle, plan.get_raw_desc(), args.get_raw_desc())) {
-    return false;
-  }
-
-  return true;
-}
-
 } // namespace
 
-cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x) {
-  auto shape = convert_vector<int64_t>(x.shape());
-  return build_cudnn_tensor(id, x, shape, normalized_strides(x));
+fe::error_t DnnGraph::prepare() {
+  RETURN_IF_ERROR(validate());
+  try {
+    RETURN_IF_ERROR(build_operation_graph(handle_));
+  } catch (cudnn_frontend::cudnnException& error) {
+    // cuDNN bug: they did not catch all exceptions in the API.
+    return {fe::error_code_t::CUDNN_BACKEND_API_FAILED, error.what()};
+  }
+  RETURN_IF_ERROR(create_execution_plans({fe::HeurMode_t::A}));
+  return {};
 }
 
-cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x) {
+fe::error_t DnnGraph::build() {
+  RETURN_IF_ERROR(check_support(handle_));
+  RETURN_IF_ERROR(build_plans(handle_));
+  return {};
+}
+
+fe::error_t DnnGraph::encode_graph(
+    cu::CommandEncoder& encoder,
+    std::unordered_map<int64_t, void*> variant_pack) {
+  cudnnSetStream(handle_, encoder.stream());
+  CudaGraph cuda_graph(encoder.device());
+  RETURN_IF_ERROR(populate_cuda_graph(
+      handle_, variant_pack, prepare_workspace(encoder), cuda_graph));
+  encoder.add_graph_node(cuda_graph);
+  return {};
+}
+
+fe::error_t DnnGraph::encode_capturing(
+    cu::CommandEncoder& encoder,
+    std::unordered_map<int64_t, void*> variant_pack) {
+  auto* workspace_ptr = prepare_workspace(encoder);
+  auto capture = encoder.capture_context();
+  cudnnSetStream(handle_, encoder.stream());
+  auto ret = execute(handle_, variant_pack, workspace_ptr);
+  if (ret.is_bad()) {
+    capture.discard = true;
+  }
+  return ret;
+}
+
+void* DnnGraph::prepare_workspace(cu::CommandEncoder& encoder) {
+  int64_t workspace_size = 0;
+  CHECK_CUDNN_FE_ERROR(get_workspace_size(workspace_size));
+  if (workspace_size > 0) {
+    array workspace(
+        cu::malloc_async(workspace_size, encoder),
+        {static_cast<int>(workspace_size)},
+        uint8);
+    encoder.add_temporary(workspace);
+    return gpu_ptr<void>(workspace);
+  }
+  return nullptr;
+}
+
+void DnnGraph::set_tensor_attrs(
+    std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+    int64_t uid,
+    const array& x,
+    const std::vector<int64_t>& shape,
+    const std::vector<int64_t>& strides) {
+  tensor->set_uid(uid)
+      .set_alignment(get_alignment(x))
+      .set_data_type(dtype_to_cudnn_type(x.dtype()))
+      .set_dim(shape)
+      .set_stride(strides);
+}
+
+void DnnGraph::set_tensor_attrs(
+    std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+    int64_t uid,
+    const array& x) {
+  set_tensor_attrs(
+      tensor,
+      uid,
+      x,
+      convert_vector<int64_t>(x.shape()),
+      normalized_strides(x));
+}
+
+void DnnGraph::set_tensor_attrs_nchw(
+    std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+    int64_t uid,
+    const array& x) {
   auto [shape, strides] = nhwc_to_nchw(x);
-  return build_cudnn_tensor(id, x, shape, strides);
+  set_tensor_attrs(tensor, uid, x, shape, strides);
 }
 
-cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x) {
-  if (x.ndim() == 0) {
-    SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
-    return build_cudnn_tensor(id, x, scalar_dims, scalar_dims);
-  }
-  if (x.ndim() == 1) {
-    int64_t s = x.shape(0);
-    SmallVector<int64_t, 4> shape = {1, x.shape(0), 1, 1};
-    SmallVector<int64_t, 4> strides = {s, 1, s, s};
-    return build_cudnn_tensor(id, x, shape, strides);
-  }
-  if (x.ndim() == 2) {
-    int64_t s =
-        x.flags().row_contiguous ? x.shape(1) * x.strides(1) : x.strides(0);
-    SmallVector<int64_t, 4> shape = {x.shape(0), x.shape(1), 1, 1};
-    SmallVector<int64_t, 4> strides = {s, x.strides(1), s, s};
-    return build_cudnn_tensor(id, x, shape, strides);
-  }
-  if (x.ndim() == 3 || x.ndim() == 4) {
-    return build_cudnn_tensor_nchw(id, x);
-  }
-  throw std::runtime_error(
-      fmt::format("Unsupported array with {} dims.", x.ndim()));
-}
-
-cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype) {
-  SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
-  return cudnn_frontend::TensorBuilder()
-      .setDim(scalar_dims.size(), scalar_dims.data())
-      .setStrides(scalar_dims.size(), scalar_dims.data())
-      .setId(id)
-      .setAlignment(16)
-      .setDataType(dtype_to_cudnn_type(dtype))
-      .setByValue(true)
-      .build();
-}
-
-std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
-    cudnnHandle_t handle,
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph) {
-  auto engine_configs = get_cudnn_engine_configs(backend_type, dtype, op_graph);
-  if (engine_configs.empty()) {
-    return std::nullopt;
-  }
-  return find_cudnn_plan_from_engine_configs(handle, engine_configs, op_graph);
-}
-
-bool encode_cudnn_plan_with_capturing(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs) {
-  return prepare_cudnn_plan(
-      encoder,
-      plan,
-      num_args,
-      uids,
-      data_ptrs,
-      [&](auto handle, auto plan, auto args) {
-        auto capture = encoder.capture_context();
-        if (cudnnBackendExecute(handle, plan, args) != CUDNN_STATUS_SUCCESS) {
-          // Discard the captured graph when failed.
-          capture.discard = true;
-          return false;
-        }
-        return true;
-      });
-}
-
-#if CUDNN_VERSION >= 90500
-bool encode_cudnn_plan_with_graph_api(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs) {
-  return prepare_cudnn_plan(
-      encoder,
-      plan,
-      num_args,
-      uids,
-      data_ptrs,
-      [&](auto handle, auto plan, auto args) {
-        if (!graph) {
-          graph = CudaGraph(encoder.device());
-          if (cudnnBackendPopulateCudaGraph(handle, plan, args, graph) !=
-              CUDNN_STATUS_SUCCESS) {
-            return false;
-          }
-        } else {
-          if (cudnnBackendUpdateCudaGraph(handle, plan, args, graph) !=
-              CUDNN_STATUS_SUCCESS) {
-            return false;
-          }
-        }
-        encoder.add_graph_node(graph);
-        return true;
-      });
-}
-#endif
-
 } // namespace mlx::core

mlx/backend/cuda/cudnn_utils.h

@@ -2,25 +2,30 @@
 
 #pragma once
 
-#include "mlx/array.h"
-#include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/device/config.h"
 #include "mlx/backend/cuda/utils.h"
 #include "mlx/dtype_utils.h"
 
 #include <cudnn_frontend.h>
-#include <cudnn_frontend_find_plan.h>
 #include <fmt/format.h>
 
-#include <algorithm>
-#include <array>
-
 namespace mlx::core {
 
 namespace cu {
 class CommandEncoder;
 }
 
+namespace fe = cudnn_frontend;
+
+#define CHECK_CUDNN_FE_ERROR(cmd)                                    \
+  do {                                                               \
+    auto error = cmd;                                                \
+    if (!error.is_good()) {                                          \
+      throw std::runtime_error(                                      \
+          fmt::format("{} failed: {}.", #cmd, error.get_message())); \
+    }                                                                \
+  } while (0)
+
 // Return pointer alignment of |x|'s data.
 inline uint8_t get_alignment(const array& x) {
   uint8_t alignment = 1;
@@ -35,8 +40,31 @@ inline uint8_t get_alignment(const array& x) {
 
 // Convert the type of elements in |vec| to |T|.
 template <typename T, typename Vec>
-inline SmallVector<T> convert_vector(const Vec& vec) {
-  return SmallVector<T>(vec.begin(), vec.end());
+inline std::vector<T> convert_vector(const Vec& vec) {
+  return std::vector<T>(vec.begin(), vec.end());
+}
+
+// Map dtype to cudnn data type.
+inline fe::DataType_t dtype_to_cudnn_type(Dtype dtype) {
+  switch (dtype) {
+    case int8:
+      return fe::DataType_t::INT8;
+    case int32:
+      return fe::DataType_t::INT32;
+    case uint8:
+      return fe::DataType_t::UINT8;
+    case float16:
+      return fe::DataType_t::HALF;
+    case bfloat16:
+      return fe::DataType_t::BFLOAT16;
+    case float32:
+      return fe::DataType_t::FLOAT;
+    case float64:
+      return fe::DataType_t::DOUBLE;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in cuDNN: {}.", dtype_to_string(dtype)));
+  }
 }
 
 // Return an array that can be used as map key for |vec| with size <= MAX_NDIM.
@@ -55,111 +83,89 @@ inline std::array<T, NDIM> vector_key(const Vec<T>& vec) {
   return result;
 }
 
-// Helpers used by get_data_ptrs to get pointers.
-inline void* get_data_ptr(const array& arr) {
-  return const_cast<void*>(gpu_ptr<void>(arr));
-}
-
-template <typename T, typename = std::enable_if_t<std::is_scalar_v<T>>>
-inline void* get_data_ptr(T& scalar) {
-  return &scalar;
-}
-
-// Return an array filled with data pointers of args.
-template <typename... Args>
-inline std::array<void*, sizeof...(Args)> get_data_ptrs(Args&... args) {
-  return {get_data_ptr(args)...};
-}
-
-// Map dtype to cudnn data type.
-inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
-  switch (dtype) {
-    case int8:
-      return CUDNN_DATA_INT8;
-    case int32:
-      return CUDNN_DATA_INT32;
-    case uint8:
-      return CUDNN_DATA_UINT8;
-    case float16:
-      return CUDNN_DATA_HALF;
-    case bfloat16:
-      return CUDNN_DATA_BFLOAT16;
-    case float32:
-      return CUDNN_DATA_FLOAT;
-    case float64:
-      return CUDNN_DATA_DOUBLE;
-    default:
-      throw std::runtime_error(fmt::format(
-          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
+// Extends cuDNN graph with helpers.
+class DnnGraph : public fe::graph::Graph {
+ public:
+  DnnGraph(cudnnHandle_t handle, Dtype io_dtype, Dtype compute_dtype = float32)
+      : handle_(handle) {
+    set_io_data_type(dtype_to_cudnn_type(io_dtype));
+    set_intermediate_data_type(dtype_to_cudnn_type(compute_dtype));
+    set_compute_data_type(dtype_to_cudnn_type(compute_dtype));
   }
-}
 
-// Create a tensor descriptor from |x|.
-cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x);
+  // Create a cuDNN tensor description from MLX array |x|.
+  auto& tensor(
+      std::shared_ptr<fe::graph::Tensor_attributes>& attrs,
+      int64_t uid,
+      const array& x) {
+    set_tensor_attrs(attrs, uid, x);
+    return attrs;
+  }
+  auto tensor(const char* name, int64_t uid, const array& x) {
+    auto attrs = Graph::tensor(fe::graph::Tensor_attributes().set_name(name));
+    tensor(attrs, uid, x);
+    return attrs;
+  }
 
-// Create a tensor descriptor from |x|, and transpose from NHWC to NCHW.
-cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x);
+  // Create a cuDNN tensor description from MLX array |x|, and transpose it from
+  // NHWC layout to NCHW.
+  auto& tensor_nchw(
+      std::shared_ptr<fe::graph::Tensor_attributes>& attrs,
+      int64_t uid,
+      const array& x) {
+    set_tensor_attrs_nchw(attrs, uid, x);
+    return attrs;
+  }
+  auto tensor_nchw(const char* name, int64_t uid, const array& x) {
+    auto attrs = Graph::tensor(fe::graph::Tensor_attributes().set_name(name));
+    tensor_nchw(attrs, uid, x);
+    return attrs;
+  }
 
-// Create a tensor descriptor from |x|, make sure it is 4D, and transpose it
-// from NHWC to NCHW.
-cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x);
+  // Create a cuDNN tensor for scalar.
+  auto scalar(const char* name, int64_t uid, Dtype dtype) {
+    return Graph::tensor(fe::graph::Tensor_attributes()
+                             .set_name(name)
+                             .set_uid(uid)
+                             .set_dim({1, 1, 1, 1})
+                             .set_stride({1, 1, 1, 1})
+                             .set_is_pass_by_value(true)
+                             .set_data_type(dtype_to_cudnn_type(dtype)));
+  }
 
-// Create a 4D scalar tensor descriptor, which is passed by value.
-cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype);
+  // Call this before setting notes.
+  fe::error_t prepare();
+  // Call this after setting notes.
+  fe::error_t build();
 
-// Find a working plan for |op_graph|.
-std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
-    cudnnHandle_t handle,
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph);
+  // Add cuDNN graph to CUDA graph, using native CUDA graph API.
+  fe::error_t encode_graph(
+      cu::CommandEncoder& encoder,
+      std::unordered_map<int64_t, void*> variant_pack);
+  // Add cuDNN graph to CUDA graph, using stream capture.
+  fe::error_t encode_capturing(
+      cu::CommandEncoder& encoder,
+      std::unordered_map<int64_t, void*> variant_pack);
 
-// Encode the plan to command buffer by capturing.
-bool encode_cudnn_plan_with_capturing(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs);
+ private:
+  void* prepare_workspace(cu::CommandEncoder& encoder);
 
-#if CUDNN_VERSION >= 90500
-// Encode the plan to command buffer by using native graph api of cudnn. If the
-// |graph| is empty it will be populated, otherwise it will be updated.
-bool encode_cudnn_plan_with_graph_api(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs);
-#endif
+  void set_tensor_attrs(
+      std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+      int64_t uid,
+      const array& x,
+      const std::vector<int64_t>& shape,
+      const std::vector<int64_t>& strides);
+  void set_tensor_attrs(
+      std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+      int64_t uid,
+      const array& x);
+  void set_tensor_attrs_nchw(
+      std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+      int64_t uid,
+      const array& x);
 
-// Helpers to make calls like encode_cudnn_plan(..., {'x', 'y', 'z'}, x, y, z).
-template <typename... Args>
-bool encode_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    std::initializer_list<int64_t> uids,
-    Args&... args) {
-  assert(uids.size() == sizeof...(args));
-  auto data_ptrs = get_data_ptrs(args...);
-  return encode_cudnn_plan_with_capturing(
-      encoder, plan, uids.size(), uids.begin(), data_ptrs.data());
-}
-
-#if CUDNN_VERSION >= 90500
-template <typename... Args>
-bool encode_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    std::initializer_list<int64_t> uids,
-    Args&... args) {
-  assert(uids.size() == sizeof...(args));
-  auto data_ptrs = get_data_ptrs(args...);
-  return encode_cudnn_plan_with_graph_api(
-      encoder, plan, graph, uids.size(), uids.begin(), data_ptrs.data());
-}
-#endif
+  cudnnHandle_t handle_;
+};
 
 } // namespace mlx::core

mlx/backend/cuda/custom_kernel.cpp

@@ -289,7 +289,7 @@ void CustomKernel::eval_gpu(
       copies.emplace_back(init_value_.value(), out.dtype());
       fill_gpu(copies.back(), out, s);
     } else {
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     }
   }
 

mlx/backend/cuda/detect_cuda_arch.sh

@@ -1,13 +0,0 @@
-#!/bin/bash
-
-arch=`__nvcc_device_query`
-case "$arch" in
-    "90")
-        echo "90a" ;;
-    "100")
-        echo "100a" ;;
-    "121")
-        echo "121a" ;;
-    *)
-        echo "native" ;;
-esac

mlx/backend/cuda/device.cpp

@@ -14,20 +14,20 @@ namespace mlx::core::cu {
 
 namespace {
 
-#define CHECK_CUDNN_ERROR(cmd) check_cudnn_error(#cmd, (cmd))
-
-void check_cudnn_error(const char* name, cudnnStatus_t err) {
-  if (err != CUDNN_STATUS_SUCCESS) {
-    throw std::runtime_error(
-        fmt::format("{} failed: {}.", name, cudnnGetErrorString(err)));
-  }
+bool use_cuda_graphs() {
+  static bool use_graphs = env::get_var("MLX_USE_CUDA_GRAPHS", true);
+  return use_graphs;
 }
 
-bool use_cuda_graphs() {
-  static bool use_graphs = []() {
-    return env::get_var("MLX_USE_CUDA_GRAPHS", true);
+const char* save_cuda_graphs_dot_file() {
+  static const char* filename = []() -> const char* {
+    const char* env = std::getenv("MLX_SAVE_CUDA_GRAPHS_DOT_FILE");
+    if (env && std::strlen(env) == 0) {
+      return nullptr;
+    }
+    return env;
   }();
-  return use_graphs;
+  return filename;
 }
 
 } // namespace
@@ -87,7 +87,7 @@ CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
     return;
   }
   CHECK_CUDA_ERROR(
-      cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
+      cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeThreadLocal));
 }
 
 CommandEncoder::CaptureContext::~CaptureContext() {
@@ -115,18 +115,17 @@ CommandEncoder::ConcurrentContext::~ConcurrentContext() {
   }
 
   // Use an empty graph node for synchronization
-  CommandEncoder::GraphNode empty{NULL, 'E', std::to_string(enc.node_count_++)};
-  enc.empty_node_count_++;
+  CommandEncoder::GraphNode empty{NULL, "E", std::to_string(enc.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;
+    enc.graph_deps_key_ += from.id;
+    enc.graph_deps_key_ += "-";
+    enc.graph_deps_key_ += empty.id;
+    enc.graph_deps_key_ += "-";
   }
 
   // Insert the input -> concurrent node dependencies without updating output
@@ -141,9 +140,6 @@ CommandEncoder::ConcurrentContext::~ConcurrentContext() {
 }
 
 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));
@@ -155,6 +151,10 @@ void CommandEncoder::insert_graph_dependencies(GraphNode node) {
 }
 
 void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
+  for (auto& node : nodes) {
+    graph_nodes_key_ += node.node_type;
+    graph_nodes_key_ += "-";
+  }
   std::vector<GraphNode> deps;
   {
     // Dependencies must be added in the same order to produce a consistent
@@ -182,10 +182,10 @@ void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
     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;
+      graph_deps_key_ += from.id;
+      graph_deps_key_ += "-";
+      graph_deps_key_ += to.id;
+      graph_deps_key_ += "-";
     }
   }
 }
@@ -309,13 +309,61 @@ void CommandEncoder::add_kernel_node(
 void CommandEncoder::add_kernel_node(const cudaKernelNodeParams& params) {
   cudaGraphNode_t node;
   CHECK_CUDA_ERROR(cudaGraphAddKernelNode(&node, graph_, NULL, 0, &params));
-  insert_graph_dependencies(GraphNode{node, 'K'});
+  insert_graph_dependencies(GraphNode{node, "K"});
 }
 
 void CommandEncoder::add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params) {
   CUgraphNode node;
   CHECK_CUDA_ERROR(cuGraphAddKernelNode(&node, graph_, NULL, 0, &params));
-  insert_graph_dependencies(GraphNode{node, 'K'});
+  insert_graph_dependencies(GraphNode{node, "K"});
+}
+
+std::pair<std::string, bool> subgraph_to_key(cudaGraph_t graph) {
+  // Constructs a key representing the nodes of a sub-graph.
+  // Also checks if the sub-graph is updatable as CUDA graphs do not get
+  // updated correctly if a kernel node getting updated has a different cluster
+  // shape than the node it's being updated with.
+  std::string key = "(";
+  size_t num_nodes = 0;
+  CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, nullptr, &num_nodes));
+  if (num_nodes == 0) {
+    return {key + ")", true};
+  }
+  bool is_updatable = true;
+  std::vector<cudaGraphNode_t> nodes(num_nodes);
+  CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, nodes.data(), &num_nodes));
+  for (const auto& node : nodes) {
+    if (!is_updatable) {
+      break;
+    }
+    cudaGraphNodeType type;
+    CHECK_CUDA_ERROR(cudaGraphNodeGetType(node, &type));
+    if (type == cudaGraphNodeTypeGraph) {
+      // Try to be updatable for a structure like graph -> graph -> kernel
+      cudaGraph_t child;
+      CHECK_CUDA_ERROR(cudaGraphChildGraphNodeGetGraph(node, &child));
+      auto [subkey, sub_is_updatable] = subgraph_to_key(child);
+      is_updatable &= sub_is_updatable;
+      key += subkey;
+    } else if (type == cudaGraphNodeTypeMemset) {
+      key += "M";
+    } else if (type != cudaGraphNodeTypeKernel) {
+      is_updatable = false;
+    } else {
+      cudaLaunchAttributeValue cluster_dim;
+      CHECK_CUDA_ERROR(cudaGraphKernelNodeGetAttribute(
+          node, cudaLaunchAttributeClusterDimension, &cluster_dim));
+      // Only allow dim.x to be greater than 1
+      if (cluster_dim.clusterDim.y > 1 || cluster_dim.clusterDim.z > 1) {
+        is_updatable = false;
+      } else {
+        key += "K";
+        key += std::to_string(cluster_dim.clusterDim.x);
+      }
+    }
+  }
+  key += ")";
+  return {key, is_updatable};
 }
 
 void CommandEncoder::add_graph_node(cudaGraph_t child) {
@@ -328,8 +376,10 @@ void CommandEncoder::add_graph_node(cudaGraph_t child) {
     return;
   }
   cudaGraphNode_t node;
+  auto [sub_graph_key, is_updatable] = subgraph_to_key(child);
+  is_graph_updatable_ &= is_updatable;
   CHECK_CUDA_ERROR(cudaGraphAddChildGraphNode(&node, graph_, NULL, 0, child));
-  insert_graph_dependencies(GraphNode{node, 'G'});
+  insert_graph_dependencies(GraphNode{node, sub_graph_key});
 }
 
 bool CommandEncoder::needs_commit() {
@@ -354,44 +404,53 @@ void CommandEncoder::commit() {
           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& 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
-        graph_exec.reset();
-      }
-    }
-    if (graph_exec == nullptr) {
-      graph_exec.instantiate(graph_);
-    }
     device_.make_current();
-    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
+
+    if (!is_graph_updatable_) {
+      CudaGraphExec graph_exec;
+      graph_exec.instantiate(graph_);
+      CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
+    } else {
+      auto graph_key = graph_nodes_key_ + ":" + graph_deps_key_;
+      auto& 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
+          graph_exec.reset();
+        }
+      }
+      if (graph_exec == nullptr) {
+        graph_exec.instantiate(graph_);
+      }
+
+      CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
+    }
+
+    // Save cuda graph to dot file
+    if (const char* filename = save_cuda_graphs_dot_file(); filename) {
+      static int count = 0;
+      auto path = fmt::format("{}_{}.dot", filename, ++count);
+      CHECK_CUDA_ERROR(cudaGraphDebugDotPrint(graph_, path.c_str(), 0));
+    }
 
     // Reset state
-    graph_node_count_ = 0;
-    empty_node_count_ = 0;
     from_nodes_.clear();
     to_nodes_.clear();
-    graph_key_.clear();
+    graph_deps_key_.clear();
+    graph_nodes_key_.clear();
     node_map_.clear();
     graph_ = CudaGraph(device_);
+    is_graph_updatable_ = true;
   }
 
   // Put completion handlers in a batch.

mlx/backend/cuda/device.h

@@ -106,8 +106,9 @@ class CommandEncoder {
     cudaGraphNode_t node;
     // K = kernel
     // E = empty
-    // G = subgraph
-    char node_type;
+    // () = subgraph (with metadata)
+    // Symbols ':', '-' are reserved as separators
+    std::string node_type;
     std::string id;
   };
 
@@ -119,12 +120,11 @@ class CommandEncoder {
   CudaGraph 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::string graph_nodes_key_;
+  std::string graph_deps_key_;
   std::vector<GraphNode> concurrent_nodes_;
   std::vector<std::shared_ptr<array::Data>> temporaries_;
   LRUCache<std::string, CudaGraphExec> graph_cache_;
@@ -132,6 +132,7 @@ class CommandEncoder {
   std::vector<std::uintptr_t> active_outputs_;
   std::unordered_map<std::uintptr_t, GraphNode> node_map_;
   size_t bytes_in_graph_{0};
+  bool is_graph_updatable_{true};
   int max_ops_per_graph_;
   int max_mb_per_graph_;
 };

mlx/backend/cuda/distributed.cu

@@ -26,7 +26,7 @@ void AllReduce::eval_gpu(
       out.copy_shared_buffer(in);
       return {in, out};
     } else {
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
       return {in, out};
     }
   };
@@ -74,7 +74,7 @@ void AllGather::eval_gpu(
   };
 
   auto input = ensure_contiguous(inputs[0]);
-  outputs[0].set_data(cu::malloc_async(outputs[0].nbytes(), encoder.stream()));
+  outputs[0].set_data(cu::malloc_async(outputs[0].nbytes(), encoder));
 
   encoder.set_input_array(input);
   encoder.set_output_array(outputs[0]);
@@ -103,7 +103,7 @@ void ReduceScatter::eval_gpu(
   };
 
   auto input = ensure_contiguous(inputs[0]);
-  outputs[0].set_data(cu::malloc_async(outputs[0].nbytes(), encoder.stream()));
+  outputs[0].set_data(cu::malloc_async(outputs[0].nbytes(), encoder));
 
   encoder.set_input_array(input);
   encoder.set_output_array(outputs[0]);

mlx/backend/cuda/event.cu

@@ -305,6 +305,7 @@ void Event::wait() {
   } else {
     event->atomic->wait(value());
   }
+  CHECK_CUDA_ERROR(cudaPeekAtLastError());
 }
 
 void Event::wait(Stream s) {

mlx/backend/cuda/gemms/cublas_gemm.cpp

@@ -370,7 +370,7 @@ void CublasGemm::execute(
     // Ensure workspace is 256-byte aligned
     int nbytes = cuda::ceil_div(heuristic_.workspaceSize, 256) * 256;
     array workspace(
-        cu::malloc_async(nbytes, encoder.stream()),
+        cu::malloc_async(nbytes, encoder),
         {static_cast<int>(heuristic_.workspaceSize)},
         int8);
     encoder.add_temporary(workspace);

mlx/backend/cuda/gemms/cublas_gemm_batched_12_9.cu

@@ -163,7 +163,7 @@ void CublasGemm::run_batched(
 
   // Launch kernel to set device offsets
   auto pointers = array(
-      cu::malloc_async(batch_count * sizeof(void*) * 3, encoder.stream()),
+      cu::malloc_async(batch_count * sizeof(void*) * 3, encoder),
       {batch_count * 3},
       uint64);
 
@@ -251,7 +251,7 @@ void CublasGemm::run_batched(
 
   // Launch kernel to set device offsets
   auto pointers = array(
-      cu::malloc_async(batch_count * sizeof(uint64_t) * 4, encoder.stream()),
+      cu::malloc_async(batch_count * sizeof(uint64_t) * 4, encoder),
       {batch_count * 4},
       uint64);
 

mlx/backend/cuda/indexing.cpp

@@ -61,7 +61,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   if (out.size() == 0) {
     return;
   }
@@ -241,7 +241,7 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   if (out.size() == 0) {
     return;
   }

mlx/backend/cuda/layer_norm.cu

@@ -244,7 +244,7 @@ void LayerNorm::eval_gpu(
         out.copy_shared_buffer(x);
       } else {
         out.set_data(
-            cu::malloc_async(x.data_size() * x.itemsize(), encoder.stream()),
+            cu::malloc_async(x.data_size() * x.itemsize(), encoder),
             x.data_size(),
             x.strides(),
             x.flags());
@@ -335,7 +335,7 @@ void LayerNormVJP::eval_gpu(
     gx.copy_shared_buffer(g);
     g_in_gx = true;
   } else {
-    gx.set_data(cu::malloc_async(gx.nbytes(), encoder.stream()));
+    gx.set_data(cu::malloc_async(gx.nbytes(), encoder));
   }
   if (g_copied && !g_in_gx) {
     encoder.add_temporary(g);
@@ -355,7 +355,7 @@ void LayerNormVJP::eval_gpu(
       g_in_gw = true;
       gw_temp.copy_shared_buffer(g);
     } else {
-      gw_temp.set_data(cu::malloc_async(gw_temp.nbytes(), encoder.stream()));
+      gw_temp.set_data(cu::malloc_async(gw_temp.nbytes(), encoder));
       encoder.add_temporary(gw_temp);
     }
   }

mlx/backend/cuda/load.cpp

@@ -32,7 +32,7 @@ void Load::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& encoder = cu::get_command_encoder(stream());
   auto size = out.size();
   auto nbytes = size * out.itemsize();
-  out.set_data(cu::malloc_async(nbytes, encoder.stream()));
+  out.set_data(cu::malloc_async(nbytes, encoder));
   auto out_ptr = malloc(nbytes);
   reader_->read(static_cast<char*>(out_ptr), nbytes, offset_);
   if (swap_endianness_) {

mlx/backend/cuda/logsumexp.cu

@@ -115,7 +115,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto in = ensure_contiguous(inputs[0]);
   if (in.flags().row_contiguous) {
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
   } else {
     auto n = in.shape(-1);
     auto flags = in.flags();
@@ -130,7 +130,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
     }
     flags.col_contiguous = col_contig;
     out.set_data(
-        cu::malloc_async(in.nbytes() / n, encoder.stream()),
+        cu::malloc_async(in.nbytes() / n, encoder),
         in.data_size() / n,
         std::move(strides),
         flags);

mlx/backend/cuda/matmul.cpp

@@ -121,7 +121,7 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
     return;
   }
 
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   int M = a_pre.shape(-2);
   int N = b_pre.shape(-1);
@@ -163,7 +163,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   if (beta_ == 1 && a.dtype() != complex64 && c.strides(-1) == 1 &&
       c.data_size() == out.shape(-1)) {
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
     gemm_and_bias(
         encoder,
         M,
@@ -187,10 +187,10 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
     auto sty = c.strides()[c.ndim() - 1];
     if (sty == 1 && stx == c.shape(-1)) {
       ldc = stx;
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     } else if (sty == 1 && stx == 0) {
       ldc = 0;
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     } else {
       // Copy C into out and set C to out
       ldc = c.shape(-1);

mlx/backend/cuda/primitives.cpp

@@ -37,6 +37,7 @@ NO_GPU(Inverse)
 NO_GPU(Cholesky)
 NO_GPU_MULTI(Eig)
 NO_GPU_MULTI(Eigh)
+NO_GPU(MaskedScatter)
 
 namespace distributed {
 NO_GPU_MULTI(Send)

mlx/backend/cuda/quantized/quantized.cpp

@@ -59,7 +59,7 @@ void fast::Quantize::eval_gpu(
     auto scales = ensure_row_contiguous(inputs[1], enc, s);
     auto& w = outputs[0];
 
-    w.set_data(cu::malloc_async(w.nbytes(), enc.stream()));
+    w.set_data(cu::malloc_async(w.nbytes(), enc));
 
     if (mode_ == QuantizationMode::Affine) {
       auto biases = ensure_row_contiguous(inputs[2], enc, s);
@@ -72,11 +72,11 @@ void fast::Quantize::eval_gpu(
     auto& wq = outputs[0];
     auto& scales = outputs[1];
 
-    wq.set_data(cu::malloc_async(wq.nbytes(), enc.stream()));
-    scales.set_data(cu::malloc_async(scales.nbytes(), enc.stream()));
+    wq.set_data(cu::malloc_async(wq.nbytes(), enc));
+    scales.set_data(cu::malloc_async(scales.nbytes(), enc));
     if (mode_ == QuantizationMode::Affine) {
       auto& biases = outputs[2];
-      biases.set_data(cu::malloc_async(biases.nbytes(), enc.stream()));
+      biases.set_data(cu::malloc_async(biases.nbytes(), enc));
       affine_quantize(w, wq, scales, biases, group_size_, bits_, enc, s);
     } else {
       fp_quantize(w, wq, scales, group_size_, bits_, enc, s);

mlx/backend/cuda/random.cu

@@ -139,30 +139,36 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   // keys has shape (N1, ..., NK, 2)
   // out has shape (N1, ..., NK, M1, M2, ...)
   auto& keys = inputs[0];
-  uint32_t num_keys = keys.size() / 2;
+  size_t num_keys = keys.size() / 2;
 
-  uint32_t elems_per_key = out.size() / num_keys;
-  uint32_t bytes_per_key = out.itemsize() * elems_per_key;
+  size_t elems_per_key = out.size() / num_keys;
+  size_t bytes_per_key = out.itemsize() * elems_per_key;
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   if (out.size() == 0) {
     return;
   }
 
-  uint32_t out_per_key = (bytes_per_key + 4 - 1) / 4;
-  uint32_t half_size = out_per_key / 2;
+  size_t out_per_key = (bytes_per_key + 4 - 1) / 4;
+  size_t half_size = out_per_key / 2;
+
   bool odd = out_per_key % 2;
+  if ((half_size + odd) >= UINT32_MAX || num_keys >= UINT32_MAX) {
+    throw std::runtime_error("[RandomBits::eval_gpu] Large size unsupported");
+  }
 
   encoder.set_input_array(keys);
   encoder.set_output_array(out);
-  dim3 grid_dims{num_keys, half_size + odd};
-  int64_t total = grid_dims.x * grid_dims.y;
-  int32_t threads_y = 1;
-  while ((total / threads_y) >= (1U << 31)) {
+  int64_t total = num_keys * (half_size + odd);
+  uint32_t threads_y = 1;
+  while ((total / threads_y) >= UINT_MAX) {
     threads_y *= 2;
   }
-  int32_t threads_x = cuda::ceil_div(total, threads_y);
+  uint32_t threads_x = cuda::ceil_div(total, threads_y);
+
+  dim3 grid_dims{
+      static_cast<uint32_t>(num_keys), static_cast<uint32_t>(half_size + odd)};
   auto [grid, block] = get_grid_and_block(threads_x, threads_y, 1);
   auto& stream = encoder.stream();
   if (keys.flags().row_contiguous) {

mlx/backend/cuda/reduce/all_reduce.cu

@@ -66,7 +66,7 @@ void all_reduce(
     Reduce::ReduceType reduce_type) {
   constexpr int N_READS = 8;
 
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   auto get_args = [](size_t size, int N) {
     int threads = std::min(512UL, (size + N - 1) / N);
@@ -107,8 +107,7 @@ void all_reduce(
   encoder.set_input_array(in);
   if (blocks > 1) {
     array intermediate({blocks}, out.dtype(), nullptr, {});
-    intermediate.set_data(
-        cu::malloc_async(intermediate.nbytes(), encoder.stream()));
+    intermediate.set_data(cu::malloc_async(intermediate.nbytes(), encoder));
     encoder.add_temporary(intermediate);
     encoder.set_output_array(intermediate);
     dispatch_all_types(dt, [&](auto type_tag) {

mlx/backend/cuda/reduce/col_reduce.cu

@@ -89,9 +89,13 @@ template <
     int NDIM,
     int BM,
     int BN,
-    int N_READS = 4>
-__global__ void
-col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
+    int N_READS = 4,
+    int BLOCKS = 1>
+__global__ void col_reduce_looped(
+    T* in,
+    U* out,
+    const __grid_constant__ ColReduceArgs args,
+    int64_t out_size) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
   auto warp = cg::tiled_partition<WARP_SIZE>(block);
@@ -102,6 +106,8 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
   size_t tile_idx = grid.block_rank();
   size_t tile_x = tile_idx % ((args.reduction_stride + BN - 1) / BN);
   size_t tile_y = tile_idx / ((args.reduction_stride + BN - 1) / BN);
+  size_t tile_out = tile_y / out_size;
+  tile_y = tile_y % out_size;
 
   // Compute the indices for the thread within the tile
   short thread_x = block.thread_rank() % threads_per_row;
@@ -118,12 +124,23 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
     totals[i] = ReduceInit<Op, T>::value();
   }
 
-  LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-  loop.next(thread_y, args.reduce_shape.data(), args.reduce_strides.data());
   size_t total = args.non_col_reductions * args.reduction_size;
+  size_t per_block, start, end;
+  if constexpr (BLOCKS > 1) {
+    per_block = (total + BLOCKS - 1) / BLOCKS;
+    start = tile_out * per_block + thread_y;
+    end = min((tile_out + 1) * per_block, total);
+  } else {
+    per_block = total;
+    start = thread_y;
+    end = total;
+  }
+
+  LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
+  loop.next(start, args.reduce_shape.data(), args.reduce_strides.data());
   if (tile_x * BN + BN <= args.reduction_stride) {
     if (args.reduction_stride % N_READS == 0) {
-      for (size_t r = thread_y; r < total; r += BM) {
+      for (size_t r = start; r < end; r += BM) {
         T vals[N_READS];
         cub::LoadDirectBlockedVectorized(thread_x, in + loop.location(), vals);
         for (int i = 0; i < N_READS; i++) {
@@ -132,7 +149,7 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
         loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
       }
     } else {
-      for (size_t r = thread_y; r < total; r += BM) {
+      for (size_t r = start; r < end; r += BM) {
         T vals[N_READS];
         cub::LoadDirectBlocked(thread_x, in + loop.location(), vals);
         for (int i = 0; i < N_READS; i++) {
@@ -142,7 +159,7 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
       }
     }
   } else {
-    for (size_t r = thread_y; r < total; r += BM) {
+    for (size_t r = start; r < end; r += BM) {
       T vals[N_READS];
       cub::LoadDirectBlocked(
           thread_x,
@@ -173,6 +190,9 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
 
   // Write result.
   if (warp.thread_rank() == 0) {
+    if (BLOCKS > 1) {
+      out += tile_out * out_size * args.reduction_stride;
+    }
     cub::StoreDirectBlocked(
         warp.meta_group_rank(),
         out + tile_y * args.reduction_stride + tile_x * BN,
@@ -227,11 +247,12 @@ __global__ void col_reduce_small(
 inline auto output_grid_for_col_reduce(
     const array& out,
     const cu::ColReduceArgs& args,
-    int bn) {
+    int bn,
+    int outer = 1) {
   int gx, gy = 1;
   size_t n_inner_blocks = cuda::ceil_div(args.reduction_stride, bn);
   size_t n_outer_blocks = out.size() / args.reduction_stride;
-  size_t n_blocks = n_outer_blocks * n_inner_blocks;
+  size_t n_blocks = n_outer_blocks * n_inner_blocks * outer;
   while (n_blocks / gy > INT32_MAX) {
     gy *= 2;
   }
@@ -277,7 +298,8 @@ void col_reduce_looped(
             0,
             indata,
             gpu_ptr<U>(out),
-            static_cast<cu::ColReduceArgs>(args));
+            static_cast<cu::ColReduceArgs>(args),
+            out.size() / args.reduction_stride);
       });
     });
   });
@@ -320,6 +342,117 @@ void col_reduce_small(
   });
 }
 
+void col_reduce_two_pass(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type,
+    const std::vector<int>& axes,
+    const ReductionPlan& plan,
+    const cu::ColReduceArgs& args) {
+  // Allocate data for the output using in's layout to access them as
+  // contiguously as possible.
+  allocate_same_layout(out, in, axes, encoder);
+
+  // Allocate an intermediate array to hold the 1st pass result
+  constexpr int outer = 32;
+
+  Shape intermediate_shape;
+  intermediate_shape.push_back(outer);
+  intermediate_shape.insert(
+      intermediate_shape.end(), out.shape().begin(), out.shape().end());
+
+  Strides intermediate_strides;
+  intermediate_strides.push_back(out.size());
+  intermediate_strides.insert(
+      intermediate_strides.end(), out.strides().begin(), out.strides().end());
+
+  array intermediate(intermediate_shape, out.dtype(), nullptr, {});
+  auto [data_size, rc, cc] =
+      check_contiguity(intermediate_shape, intermediate_strides);
+  auto fl = out.flags();
+  fl.row_contiguous = rc;
+  fl.col_contiguous = cc;
+  fl.contiguous = true;
+  intermediate.set_data(
+      cu::malloc_async(intermediate.nbytes(), encoder),
+      data_size,
+      intermediate_strides,
+      fl,
+      allocator::free);
+
+  encoder.add_temporary(intermediate);
+  encoder.set_input_array(in);
+  encoder.set_output_array(intermediate);
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+      dispatch_reduce_ndim(args.reduce_ndim, [&](auto reduce_ndim) {
+        using OP = MLX_GET_TYPE(reduce_type_tag);
+        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        using U = typename cu::ReduceResult<OP, T>::type;
+        // Cub doesn't like const pointers for vectorized loads. (sigh)
+        T* indata = const_cast<T*>(gpu_ptr<T>(in));
+
+        constexpr int N_READS = 4;
+        constexpr int BM = 32;
+        constexpr int BN = 32;
+        dim3 grid = output_grid_for_col_reduce(out, args, BN, outer);
+        int blocks = BM * BN / N_READS;
+        auto kernel = cu::
+            col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS, outer>;
+        encoder.add_kernel_node(
+            kernel,
+            grid,
+            blocks,
+            0,
+            indata,
+            gpu_ptr<U>(intermediate),
+            static_cast<cu::ColReduceArgs>(args),
+            out.size() / args.reduction_stride);
+      });
+    });
+  });
+
+  // Prepare the reduction arguments for the 2nd pass
+  cu::ColReduceArgs second_args = args;
+  second_args.reduction_size = outer;
+  second_args.reduction_stride = out.size();
+  second_args.ndim = 0;
+  second_args.reduce_shape[0] = outer;
+  second_args.reduce_strides[0] = out.size();
+  second_args.reduce_ndim = 1;
+  second_args.non_col_reductions = 1;
+
+  encoder.set_input_array(intermediate);
+  encoder.set_output_array(out);
+  dispatch_all_types(intermediate.dtype(), [&](auto type_tag) {
+    dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+      dispatch_reduce_ndim(second_args.reduce_ndim, [&](auto reduce_ndim) {
+        using OP = MLX_GET_TYPE(reduce_type_tag);
+        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        using U = typename cu::ReduceResult<OP, T>::type;
+
+        constexpr int N_READS = 4;
+        constexpr int BM = 32;
+        constexpr int BN = 32;
+        dim3 grid = output_grid_for_col_reduce(out, second_args, BN);
+        int blocks = BM * BN / N_READS;
+        auto kernel =
+            cu::col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS>;
+        encoder.add_kernel_node(
+            kernel,
+            grid,
+            blocks,
+            0,
+            gpu_ptr<T>(intermediate),
+            gpu_ptr<U>(out),
+            second_args,
+            second_args.reduction_stride);
+      });
+    });
+  });
+}
+
 void col_reduce(
     cu::CommandEncoder& encoder,
     const array& in,
@@ -334,6 +467,18 @@ void col_reduce(
   //   It is a general strided reduce. Each threadblock computes the output for
   //   a subrow of the fast moving axis. For instance 32 elements.
   //
+  // - col_reduce_small
+  //
+  //  It is a column reduce for small columns. Each thread loops over the whole
+  //  column without communicating with any other thread.
+  //
+  // - col_reduce_two_pass
+  //
+  //  It is a reduce for long columns. To increase parallelism, we split the
+  //  reduction in two passes. First we do a column reduce where many
+  //  threadblocks operate on different parts of the reduced axis. Then we
+  //  perform a final column reduce.
+  //
   // Notes: As in row reduce we opt to read as much in order as possible and
   //        leave transpositions as they are (contrary to our Metal backend).
   //
@@ -349,6 +494,14 @@ void col_reduce(
     return;
   }
 
+  // Long column with smallish row
+  size_t total_sums = args.non_col_reductions * args.reduction_size;
+  size_t approx_threads = out.size();
+  if (total_sums / approx_threads > 32) {
+    col_reduce_two_pass(encoder, in, out, reduce_type, axes, plan, args);
+    return;
+  }
+
   // Fallback col reduce
   col_reduce_looped(encoder, in, out, reduce_type, axes, plan, args);
 }

mlx/backend/cuda/reduce/init_reduce.cu

@@ -28,7 +28,7 @@ void init_reduce(
     Reduce::ReduceType reduce_type) {
   // Allocate if needed
   if (out.data_shared_ptr() == nullptr) {
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
   }
 
   encoder.set_output_array(out);

mlx/backend/cuda/reduce/reduce_utils.cuh

@@ -96,7 +96,7 @@ inline void allocate_same_layout(
     const std::vector<int>& axes,
     cu::CommandEncoder& encoder) {
   if (in.flags().row_contiguous) {
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
     return;
   }
 
@@ -135,7 +135,7 @@ inline void allocate_same_layout(
   fl.col_contiguous = cc;
   fl.contiguous = true;
   out.set_data(
-      cu::malloc_async(out.nbytes(), encoder.stream()),
+      cu::malloc_async(out.nbytes(), encoder),
       data_size,
       final_strides,
       fl,

mlx/backend/cuda/rms_norm.cu

@@ -22,26 +22,28 @@ inline __device__ float2 plus_f2(const float2& a, const float2& b) {
 }
 
 // Similar to cub::BlockReduce, but result is broadcasted to every thread.
-template <typename T, int BLOCK_DIM>
+template <typename T, int BLOCK_DIM, int GROUP_DIM = WARP_SIZE>
 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];
+  using TempStorage = T[std::max(BLOCK_DIM / WARP_SIZE, 1)];
 
   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);
+    auto warp = cg::tiled_partition<GROUP_DIM>(block);
     T x = cg::reduce(warp, input, op);
-    if (warp.thread_rank() == 0) {
-      temp[warp.meta_group_rank()] = x;
+    if constexpr (BLOCK_DIM > GROUP_DIM) {
+      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);
+    } else {
+      return 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) {
@@ -49,6 +51,52 @@ struct BlockBroadcastReduce {
   }
 };
 
+template <typename T, int BLOCK_DIM, int REDUCE_DIM, int N_READS = 4>
+__global__ void rms_norm_small(
+    const T* x,
+    const T* w,
+    T* out,
+    float eps,
+    uint32_t axis_size,
+    uint32_t n_rows,
+    int64_t w_stride) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+
+  using BlockReduceT = BlockBroadcastReduce<float, BLOCK_DIM, REDUCE_DIM>;
+  __shared__ typename BlockReduceT::TempStorage temp;
+
+  auto row =
+      (grid.block_rank() * block.dim_threads().y) + block.thread_index().y;
+  if (row >= n_rows) {
+    return;
+  }
+  x += row * axis_size;
+  out += row * axis_size;
+
+  // Normalizer.
+  float normalizer = 0;
+  auto index = block.thread_index().x;
+  auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    float t = static_cast<float>(xn[i]);
+    normalizer += t * t;
+  }
+
+  normalizer = BlockReduceT{block, temp}.Sum(normalizer);
+  normalizer = rsqrt(normalizer / axis_size + eps);
+
+  // Outputs.
+  auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    float y = static_cast<float>(xn[i]) * normalizer;
+    xn[i] = wn[i] * static_cast<T>(y);
+  }
+  store_vector<N_READS>(out, index, xn, axis_size);
+}
+
 template <typename T, int BLOCK_DIM, int N_READS = 4>
 __global__ void rms_norm(
     const T* x,
@@ -94,6 +142,74 @@ __global__ void rms_norm(
   }
 }
 
+template <
+    typename T,
+    bool HAS_W,
+    int BLOCK_DIM,
+    int REDUCE_DIM,
+    int N_READS = 4>
+__global__ void rms_norm_vjp_small(
+    const T* x,
+    const T* w,
+    const T* g,
+    T* gx,
+    T* gw,
+    float eps,
+    int32_t axis_size,
+    int32_t n_rows,
+    int64_t w_stride) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+
+  using BlockReduceF2 = BlockBroadcastReduce<float2, BLOCK_DIM, REDUCE_DIM>;
+  __shared__ typename BlockReduceF2::TempStorage temp;
+
+  auto row =
+      (grid.block_rank() * block.dim_threads().y) + block.thread_index().y;
+  if (row >= n_rows) {
+    return;
+  }
+
+  x += row * axis_size;
+  g += row * axis_size;
+  gx += row * axis_size;
+  gw += row * axis_size;
+
+  // Normalizer.
+  float2 factors = {};
+  auto index = block.thread_index().x;
+  auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+  auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
+  auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+  for (int i = 0; i < N_READS; i++) {
+    float t = static_cast<float>(xn[i]);
+    float wi = wn[i];
+    float gi = gn[i];
+    float wg = wi * gi;
+    factors = plus_f2(factors, {wg * t, t * t});
+  }
+
+  factors = BlockReduceF2{block, temp}.Reduce(factors, plus_f2, {});
+  float meangwx = factors.x / axis_size;
+  float normalizer = rsqrt(factors.y / axis_size + eps);
+  float normalizer3 = normalizer * normalizer * normalizer;
+
+  // Outputs.
+  for (int i = 0; i < N_READS; i++) {
+    float xi = xn[i];
+    float wi = wn[i];
+    float gi = gn[i];
+    xn[i] = static_cast<T>(normalizer * wi * gi - xi * meangwx * normalizer3);
+    if constexpr (HAS_W) {
+      wn[i] = static_cast<T>(gi * xi * normalizer);
+    }
+  }
+  store_vector<N_READS>(gx, index, xn, axis_size);
+  if constexpr (HAS_W) {
+    store_vector<N_READS>(gw, index, wn, axis_size);
+  }
+}
+
 template <typename T, bool HAS_W, int BLOCK_DIM, int N_READS = 4>
 __global__ void rms_norm_vjp(
     const T* x,
@@ -107,12 +223,8 @@ __global__ void rms_norm_vjp(
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
 
-  using BlockReduceF = BlockBroadcastReduce<float, BLOCK_DIM>;
   using BlockReduceF2 = BlockBroadcastReduce<float2, BLOCK_DIM>;
-  __shared__ union {
-    typename BlockReduceF::TempStorage f;
-    typename BlockReduceF2::TempStorage f2;
-  } temp;
+  __shared__ typename BlockReduceF2::TempStorage temp;
 
   x += grid.block_rank() * axis_size;
   g += grid.block_rank() * axis_size;
@@ -134,7 +246,7 @@ __global__ void rms_norm_vjp(
       factors = plus_f2(factors, {wg * t, t * t});
     }
   }
-  factors = BlockReduceF2{block, temp.f2}.Reduce(factors, plus_f2, {});
+  factors = BlockReduceF2{block, temp}.Reduce(factors, plus_f2, {});
   float meangwx = factors.x / axis_size;
   float normalizer = rsqrt(factors.y / axis_size + eps);
   float normalizer3 = normalizer * normalizer * normalizer;
@@ -169,6 +281,43 @@ bool RMSNorm::use_fallback(Stream s) {
   return s.device == Device::cpu;
 }
 
+template <int n_per_thread, typename F>
+void dispatch_group_dim(int axis_size, F&& f) {
+  if (axis_size <= n_per_thread * 8) {
+    f(std::integral_constant<int, 8>{},
+      std::integral_constant<int, 1>(),
+      std::integral_constant<int, 16>());
+  } else if (axis_size <= n_per_thread * 16) {
+    f(std::integral_constant<int, 16>{},
+      std::integral_constant<int, 1>(),
+      std::integral_constant<int, 8>());
+  } else if (axis_size <= n_per_thread * 32) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 1>(),
+      std::integral_constant<int, 4>());
+  } else if (axis_size <= n_per_thread * 32 * 2) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 2>(),
+      std::integral_constant<int, 2>());
+  } else if (axis_size <= n_per_thread * 32 * 4) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 4>(),
+      std::integral_constant<int, 1>());
+  } else if (axis_size <= n_per_thread * 32 * 8) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 8>(),
+      std::integral_constant<int, 1>());
+  } else if (axis_size <= n_per_thread * 32 * 16) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 16>(),
+      std::integral_constant<int, 1>());
+  } else {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 32>(),
+      std::integral_constant<int, 1>());
+  }
+}
+
 // TODO: There are duplicate code with backend/metal/normalization.cpp
 void RMSNorm::eval_gpu(
     const std::vector<array>& inputs,
@@ -190,7 +339,7 @@ void RMSNorm::eval_gpu(
         out.copy_shared_buffer(x);
       } else {
         out.set_data(
-            cu::malloc_async(x.data_size() * x.itemsize(), encoder.stream()),
+            cu::malloc_async(x.data_size() * x.itemsize(), encoder),
             x.data_size(),
             x.strides(),
             x.flags());
@@ -216,12 +365,33 @@ void RMSNorm::eval_gpu(
   dispatch_float_types(out.dtype(), "rms_norm", [&](auto type_tag) {
     using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
     constexpr int N_READS = 16 / sizeof(DataType);
-    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
-      auto kernel = cu::rms_norm<DataType, block_dim(), N_READS>;
+    if (axis_size <= N_READS * 1024) {
+      dispatch_group_dim<N_READS>(
+          axis_size, [&](auto group_dim, auto n_groups, auto groups_per_block) {
+            constexpr int block_dim = n_groups() * group_dim();
+            auto kernel =
+                cu::rms_norm_small<DataType, block_dim, group_dim(), N_READS>;
+            auto n_blocks =
+                (n_rows + groups_per_block() - 1) / groups_per_block();
+            encoder.add_kernel_node(
+                kernel,
+                n_blocks,
+                {block_dim, groups_per_block()},
+                0,
+                gpu_ptr<DataType>(x),
+                gpu_ptr<DataType>(w),
+                gpu_ptr<DataType>(out),
+                eps_,
+                axis_size,
+                n_rows,
+                w_stride);
+          });
+    } else {
+      auto kernel = cu::rms_norm<DataType, 1024, N_READS>;
       encoder.add_kernel_node(
           kernel,
           n_rows,
-          block_dim(),
+          1024,
           0,
           gpu_ptr<DataType>(x),
           gpu_ptr<DataType>(w),
@@ -229,7 +399,7 @@ void RMSNorm::eval_gpu(
           eps_,
           axis_size,
           w_stride);
-    });
+    }
   });
 }
 
@@ -274,7 +444,7 @@ void RMSNormVJP::eval_gpu(
     gx.copy_shared_buffer(g);
     g_in_gx = true;
   } else {
-    gx.set_data(cu::malloc_async(gx.nbytes(), encoder.stream()));
+    gx.set_data(cu::malloc_async(gx.nbytes(), encoder));
   }
   if (g_copied && !g_in_gx) {
     encoder.add_temporary(g);
@@ -292,7 +462,7 @@ void RMSNormVJP::eval_gpu(
     if (!g_in_gx && donate_g) {
       gw_temp.copy_shared_buffer(g);
     } else {
-      gw_temp.set_data(cu::malloc_async(gw_temp.nbytes(), encoder.stream()));
+      gw_temp.set_data(cu::malloc_async(gw_temp.nbytes(), encoder));
       encoder.add_temporary(gw_temp);
     }
   }
@@ -306,27 +476,51 @@ void RMSNormVJP::eval_gpu(
     dispatch_bool(has_w, [&](auto has_w_constant) {
       using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
       constexpr int N_READS = 16 / sizeof(DataType);
-      dispatch_block_dim(
-          cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
-            auto kernel = cu::rms_norm_vjp<
-                DataType,
-                has_w_constant.value,
-                block_dim(),
-                N_READS>;
-            encoder.add_kernel_node(
-                kernel,
-                n_rows,
-                block_dim(),
-                0,
-                gpu_ptr<DataType>(x),
-                gpu_ptr<DataType>(w),
-                gpu_ptr<DataType>(g),
-                gpu_ptr<DataType>(gx),
-                gpu_ptr<DataType>(gw_temp),
-                eps_,
-                axis_size,
-                w_stride);
-          });
+      if (axis_size <= N_READS * 1024) {
+        dispatch_group_dim<N_READS>(
+            axis_size,
+            [&](auto group_dim, auto n_groups, auto groups_per_block) {
+              constexpr int block_dim = group_dim() * n_groups();
+              auto kernel = cu::rms_norm_vjp_small<
+                  DataType,
+                  has_w_constant.value,
+                  block_dim,
+                  group_dim(),
+                  N_READS>;
+              auto n_blocks =
+                  (n_rows + groups_per_block() - 1) / groups_per_block();
+              encoder.add_kernel_node(
+                  kernel,
+                  n_blocks,
+                  {block_dim, groups_per_block()},
+                  0,
+                  gpu_ptr<DataType>(x),
+                  gpu_ptr<DataType>(w),
+                  gpu_ptr<DataType>(g),
+                  gpu_ptr<DataType>(gx),
+                  gpu_ptr<DataType>(gw_temp),
+                  eps_,
+                  axis_size,
+                  n_rows,
+                  w_stride);
+            });
+      } else {
+        auto kernel =
+            cu::rms_norm_vjp<DataType, has_w_constant.value, 1024, N_READS>;
+        encoder.add_kernel_node(
+            kernel,
+            n_rows,
+            1024,
+            0,
+            gpu_ptr<DataType>(x),
+            gpu_ptr<DataType>(w),
+            gpu_ptr<DataType>(g),
+            gpu_ptr<DataType>(gx),
+            gpu_ptr<DataType>(gw_temp),
+            eps_,
+            axis_size,
+            w_stride);
+      }
     });
   });
 

mlx/backend/cuda/rope.cu

@@ -292,14 +292,14 @@ void RoPE::eval_gpu(
       donated = true;
       out.copy_shared_buffer(in);
     } else {
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     }
     strides[0] = mat_size;
     strides[1] = in.strides()[ndim - 2];
     strides[2] = in.strides()[ndim - 1];
   } else if (dispatch_ndim == 3) {
     // Handle non-contiguous 3D inputs
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
     strides[0] = in.strides()[ndim - 3];
     strides[1] = in.strides()[ndim - 2];
     strides[2] = in.strides()[ndim - 1];

mlx/backend/cuda/scaled_dot_product_attention.cpp

@@ -5,47 +5,13 @@
 #include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/fast_primitives.h"
-#include "mlx/transforms_impl.h"
 
 #include <nvtx3/nvtx3.hpp>
 
 namespace mlx::core {
 
-namespace fe = cudnn_frontend;
-
 namespace {
 
-#define CHECK_CUDNN_FE_ERROR(cmd)                                    \
-  do {                                                               \
-    auto error = cmd;                                                \
-    if (!error.is_good()) {                                          \
-      throw std::runtime_error(                                      \
-          fmt::format("{} failed: {}.", #cmd, error.get_message())); \
-    }                                                                \
-  } while (0)
-
-std::vector<int64_t> normalized_strides(const array& x) {
-  std::vector<int64_t> strides(x.strides().begin(), x.strides().end());
-  if (!x.flags().row_contiguous || x.ndim() < 2) {
-    return strides;
-  }
-  for (int i = x.ndim() - 2; i >= 0; --i) {
-    if (x.shape(i) == 1) {
-      strides[i] = x.shape(i + 1) * strides[i + 1];
-    }
-  }
-  return strides;
-}
-
-void set_tensor_attrs(
-    std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
-    int64_t uid,
-    const array& x) {
-  tensor->set_uid(uid)
-      .set_dim({x.shape().begin(), x.shape().end()})
-      .set_stride(normalized_strides(x));
-}
-
 array prepare_sdpa_input(const array& x, Stream s) {
   // SDPA kernel's requirements on inputs:
   // 1. last dim's stride be 1;
@@ -59,11 +25,43 @@ array prepare_sdpa_input(const array& x, Stream s) {
   return x;
 }
 
+void malloc_with_same_layout(
+    cu::CommandEncoder& encoder,
+    array& o,
+    const array& q) {
+  if (q.flags().row_contiguous) {
+    o.set_data(cu::malloc_async(o.nbytes(), encoder));
+    return;
+  }
+  // fill_order = argsort(q.strides())
+  Shape fill_order(q.ndim());
+  std::iota(fill_order.begin(), fill_order.end(), 0);
+  std::stable_sort(
+      fill_order.begin(), fill_order.end(), [&q](int idx1, int idx2) {
+        auto s1 = q.strides(idx1) > 0 ? q.strides(idx1) : 1;
+        auto s2 = q.strides(idx2) > 0 ? q.strides(idx2) : 1;
+        return s1 < s2;
+      });
+  // Generate o_strides with fill_order
+  Strides o_strides(q.ndim());
+  int64_t stride = 1;
+  for (int i : fill_order) {
+    o_strides[i] = stride;
+    stride *= o.shape(i);
+  }
+  // o is a transposed contiguous array
+  o.set_data(
+      cu::malloc_async(o.nbytes(), encoder),
+      o.size(),
+      o_strides,
+      {true, false, false});
+}
+
 constexpr int QKV_NDIM = 4;
 
 struct SDPACacheKey {
   int device_id;
-  cudnnDataType_t cudnn_dtype;
+  fe::DataType_t cudnn_dtype;
   std::array<int, QKV_NDIM> q_shape;
   std::array<int, QKV_NDIM> k_shape;
   std::array<int, QKV_NDIM> v_shape;
@@ -71,11 +69,50 @@ struct SDPACacheKey {
   std::array<int64_t, QKV_NDIM> k_strides;
   std::array<int64_t, QKV_NDIM> v_strides;
   bool do_causal;
+  std::array<int, QKV_NDIM> mask_shape;
+  std::array<int64_t, QKV_NDIM> mask_strides;
+  bool output_logsumexp;
 };
 
+inline BytesKey<SDPACacheKey> build_sdpa_cache_key(
+    cu::CommandEncoder& encoder,
+    const array& q,
+    const array& k,
+    const array& v,
+    bool do_causal,
+    const std::optional<array>& mask_arr,
+    bool output_logsumexp = true) {
+  BytesKey<SDPACacheKey> cache_key;
+  cache_key.pod = {
+      encoder.device().cuda_device(),
+      dtype_to_cudnn_type(q.dtype()),
+      vector_key<QKV_NDIM>(q.shape()),
+      vector_key<QKV_NDIM>(k.shape()),
+      vector_key<QKV_NDIM>(v.shape()),
+      vector_key<QKV_NDIM>(q.strides()),
+      vector_key<QKV_NDIM>(k.strides()),
+      vector_key<QKV_NDIM>(v.strides()),
+      do_causal,
+      {},
+      {},
+      output_logsumexp,
+  };
+  if (mask_arr) {
+    cache_key.pod.mask_shape = vector_key<QKV_NDIM>(mask_arr->shape());
+    cache_key.pod.mask_strides = vector_key<QKV_NDIM>(mask_arr->strides());
+  }
+  return cache_key;
+}
+
 auto& sdpa_cache() {
-  static LRUBytesKeyCache<SDPACacheKey, fe::graph::Graph> cache(
-      "MLX_CUDA_SDPA_CACHE_SIZE", /* default_capacity */ 128);
+  static LRUBytesKeyCache<SDPACacheKey, DnnGraph> cache(
+      "MLX_CUDA_SDPA_CACHE_SIZE", /* default_capacity */ 64);
+  return cache;
+}
+
+auto& sdpa_backward_cache() {
+  static LRUBytesKeyCache<SDPACacheKey, DnnGraph> cache(
+      "MLX_CUDA_SDPA_BACKWARD_CACHE_SIZE", /* default_capacity */ 64);
   return cache;
 }
 
@@ -84,59 +121,106 @@ enum UIDS {
   K,
   V,
   SCALE,
+  BIAS,
   O,
+  STATS,
+  // Backward graph:
+  D_Q,
+  D_K,
+  D_V,
+  D_O,
 };
 
-fe::graph::Graph build_sdpa_graph(
+DnnGraph build_sdpa_graph(
     cudnnHandle_t handle,
     const array& q,
     const array& k,
     const array& v,
     bool do_causal,
-    const array& o) {
-  auto dtype = fe::DataType_t::HALF;
-  if (q.dtype() == bfloat16) {
-    dtype = fe::DataType_t::BFLOAT16;
+    const std::optional<array>& mask_arr,
+    bool output_logsumexp,
+    const array& o,
+    const array& stats) {
+  DnnGraph graph(handle, q.dtype());
+
+  auto q_ = graph.tensor("Q", Q, q);
+  auto k_ = graph.tensor("K", K, k);
+  auto v_ = graph.tensor("V", V, v);
+
+  auto options = fe::graph::SDPA_attributes()
+                     .set_name("sdpa_cudnn")
+                     .set_attn_scale(graph.scalar("Scale", SCALE, float32))
+                     .set_generate_stats(output_logsumexp);
+  if (do_causal) {
+    if (q.shape(2) > k.shape(2)) {
+      options.set_causal_mask(do_causal);
+    } else {
+      options.set_causal_mask_bottom_right(do_causal);
+    }
+  }
+  if (mask_arr) {
+    options.set_bias(graph.tensor("BIAS", BIAS, *mask_arr));
   }
 
-  fe::graph::Graph graph;
-  graph.set_io_data_type(dtype)
-      .set_intermediate_data_type(fe::DataType_t::FLOAT)
-      .set_compute_data_type(fe::DataType_t::FLOAT);
+  auto [o_, stats_] = graph.sdpa(q_, k_, v_, options);
+  graph.tensor(o_, O, o)->set_output(true);
+  if (output_logsumexp) {
+    graph.tensor(stats_, STATS, stats)->set_output(true);
+  }
 
-  auto q_ = graph.tensor(fe::graph::Tensor_attributes().set_name("Q"));
-  auto k_ = graph.tensor(fe::graph::Tensor_attributes().set_name("K"));
-  auto v_ = graph.tensor(fe::graph::Tensor_attributes().set_name("V"));
-  set_tensor_attrs(q_, Q, q);
-  set_tensor_attrs(k_, K, k);
-  set_tensor_attrs(v_, V, v);
-
-  auto scale = graph.tensor(fe::graph::Tensor_attributes()
-                                .set_name("Scale")
-                                .set_uid(SCALE)
-                                .set_dim({1, 1, 1, 1})
-                                .set_stride({1, 1, 1, 1})
-                                .set_is_pass_by_value(true)
-                                .set_data_type(fe::DataType_t::FLOAT));
-
-  auto sdpa_options = fe::graph::SDPA_attributes()
-                          .set_name("sdpa_cudnn")
-                          .set_attn_scale(scale)
-                          .set_causal_mask(do_causal)
-                          .set_generate_stats(false);
-
-  auto [o_, _] = graph.sdpa(q_, k_, v_, sdpa_options);
-  o_->set_output(true);
-  set_tensor_attrs(o_, O, o);
-
-  CHECK_CUDNN_FE_ERROR(graph.validate());
-  CHECK_CUDNN_FE_ERROR(graph.build_operation_graph(handle));
-  CHECK_CUDNN_FE_ERROR(graph.create_execution_plans({fe::HeurMode_t::A}));
+  CHECK_CUDNN_FE_ERROR(graph.prepare());
   graph.select_behavior_notes(
       {fe::BehaviorNote_t::SUPPORTS_CUDA_GRAPH_NATIVE_API});
-  CHECK_CUDNN_FE_ERROR(graph.check_support(handle));
-  CHECK_CUDNN_FE_ERROR(graph.build_plans(handle));
+  CHECK_CUDNN_FE_ERROR(graph.build());
+  return graph;
+}
 
+DnnGraph build_sdpa_backward_graph(
+    cudnnHandle_t handle,
+    const array& q,
+    const array& k,
+    const array& v,
+    bool do_causal,
+    const std::optional<array>& mask_arr,
+    const array& o,
+    const array& d_o,
+    const array& stats,
+    array& d_q,
+    array& d_k,
+    array& d_v) {
+  DnnGraph graph(handle, q.dtype());
+
+  auto q_ = graph.tensor("Q", Q, q);
+  auto k_ = graph.tensor("K", K, k);
+  auto v_ = graph.tensor("V", V, v);
+  auto o_ = graph.tensor("O", O, o);
+  auto d_o_ = graph.tensor("D_O", D_O, d_o);
+  auto stats_ = graph.tensor("STATS", STATS, stats);
+
+  auto options = fe::graph::SDPA_backward_attributes()
+                     .set_name("sdpa_backward_cudnn")
+                     .set_attn_scale(graph.scalar("Scale", SCALE, float32));
+  if (do_causal) {
+    if (q.shape(2) > k.shape(2)) {
+      options.set_causal_mask(do_causal);
+    } else {
+      options.set_causal_mask_bottom_right(do_causal);
+    }
+  }
+  if (mask_arr) {
+    options.set_bias(graph.tensor("BIAS", BIAS, *mask_arr));
+  }
+
+  auto [d_q_, d_k_, d_v_] =
+      graph.sdpa_backward(q_, k_, v_, o_, d_o_, stats_, options);
+  graph.tensor(d_q_, D_Q, d_q)->set_output(true);
+  graph.tensor(d_k_, D_K, d_k)->set_output(true);
+  graph.tensor(d_v_, D_V, d_v)->set_output(true);
+
+  CHECK_CUDNN_FE_ERROR(graph.prepare());
+  graph.select_behavior_notes(
+      {fe::BehaviorNote_t::SUPPORTS_CUDA_GRAPH_NATIVE_API});
+  CHECK_CUDNN_FE_ERROR(graph.build());
   return graph;
 }
 
@@ -146,7 +230,6 @@ bool supports_sdpa_cudnn(
     const array& q,
     const array& k,
     const array& v,
-    bool has_mask,
     bool do_causal,
     Stream s) {
   static bool enabled = env::get_var("MLX_CUDA_USE_CUDNN_SPDA", 1);
@@ -159,19 +242,8 @@ bool supports_sdpa_cudnn(
     return false;
   }
 
-  if (has_mask) {
-    // TODO: Support array masks.
-    if (!do_causal) {
-      return false;
-    }
-    // FIXME: Causal mask generates wrong results when L_Q != L_K.
-    if (q.shape(2) != k.shape(2)) {
-      return false;
-    }
-  }
-
-  // Only use cuDNN for prefilling.
-  if (q.shape(2) != k.shape(2)) {
+  // Only use cuDNN for prefilling (T_q > 1) and training (T_q == T_kv).
+  if ((q.shape(2) == 1) && (q.shape(2) != k.shape(2))) {
     return false;
   }
 
@@ -191,66 +263,115 @@ void sdpa_cudnn(
     const array& v,
     float scale,
     array& o,
+    array& stats,
     bool do_causal,
+    const std::optional<array>& mask_arr,
+    bool output_logsumexp,
     Stream s) {
   auto& encoder = cu::get_command_encoder(s);
-  // TODO: Handle donation.
-  // TODO: Make O use same memory layout with Q.
-  o.set_data(cu::malloc_async(o.nbytes(), encoder.stream()));
+  auto handle = encoder.device().cudnn_handle();
+
+  malloc_with_same_layout(encoder, o, q);
 
   encoder.set_input_array(q);
   encoder.set_input_array(k);
   encoder.set_input_array(v);
   encoder.set_output_array(o);
-
-  auto handle = encoder.device().cudnn_handle();
-  cudnnSetStream(handle, encoder.stream());
+  if (mask_arr) {
+    encoder.set_input_array(*mask_arr);
+  }
+  if (output_logsumexp) {
+    stats.set_data(cu::malloc_async(stats.nbytes(), encoder));
+    encoder.set_output_array(stats);
+  }
 
   // Search cache.
-  BytesKey<SDPACacheKey> cache_key;
-  cache_key.pod = {
-      encoder.device().cuda_device(),
-      dtype_to_cudnn_type(q.dtype()),
-      vector_key<QKV_NDIM>(q.shape()),
-      vector_key<QKV_NDIM>(k.shape()),
-      vector_key<QKV_NDIM>(v.shape()),
-      vector_key<QKV_NDIM>(q.strides()),
-      vector_key<QKV_NDIM>(k.strides()),
-      vector_key<QKV_NDIM>(v.strides()),
-      do_causal,
-  };
+  auto cache_key = build_sdpa_cache_key(
+      encoder, q, k, v, do_causal, mask_arr, output_logsumexp);
   auto it = sdpa_cache().find(cache_key);
   if (it == sdpa_cache().end()) {
-    it =
-        sdpa_cache()
-            .emplace(cache_key, build_sdpa_graph(handle, q, k, v, do_causal, o))
-            .first;
+    auto graph = build_sdpa_graph(
+        handle, q, k, v, do_causal, mask_arr, output_logsumexp, o, stats);
+    it = sdpa_cache().emplace(cache_key, std::move(graph)).first;
   }
   auto& graph = it->second;
 
   std::unordered_map<int64_t, void*> variant_pack{
-      {Q, const_cast<void*>(gpu_ptr<void>(q))},
-      {K, const_cast<void*>(gpu_ptr<void>(k))},
-      {V, const_cast<void*>(gpu_ptr<void>(v))},
+      {Q, gpu_ptr<void>(q)},
+      {K, gpu_ptr<void>(k)},
+      {V, gpu_ptr<void>(v)},
       {SCALE, &scale},
       {O, gpu_ptr<void>(o)}};
-
-  int64_t workspace_size = 0;
-  CHECK_CUDNN_FE_ERROR(graph.get_workspace_size(workspace_size));
-  void* workspace_ptr = nullptr;
-  if (workspace_size > 0) {
-    array workspace(
-        cu::malloc_async(workspace_size, encoder.stream()),
-        {static_cast<int>(workspace_size)},
-        uint8);
-    encoder.add_temporary(workspace);
-    workspace_ptr = gpu_ptr<void>(workspace);
+  if (mask_arr) {
+    variant_pack[BIAS] = gpu_ptr<void>(*mask_arr);
+  }
+  if (output_logsumexp) {
+    variant_pack[STATS] = gpu_ptr<void>(stats);
   }
 
-  CudaGraph cuda_graph(encoder.device());
-  CHECK_CUDNN_FE_ERROR(graph.populate_cuda_graph(
-      handle, variant_pack, workspace_ptr, cuda_graph));
-  encoder.add_graph_node(cuda_graph);
+  CHECK_CUDNN_FE_ERROR(graph.encode_graph(encoder, std::move(variant_pack)));
+}
+
+void sdpa_backward_cudnn(
+    const array& q,
+    const array& k,
+    const array& v,
+    float scale,
+    const array& o,
+    const array& stats,
+    bool do_causal,
+    const std::optional<array>& mask_arr,
+    const array& d_o,
+    array& d_q,
+    array& d_k,
+    array& d_v,
+    Stream s) {
+  auto& encoder = cu::get_command_encoder(s);
+  auto handle = encoder.device().cudnn_handle();
+
+  malloc_with_same_layout(encoder, d_q, q);
+  malloc_with_same_layout(encoder, d_k, k);
+  malloc_with_same_layout(encoder, d_v, v);
+
+  encoder.set_input_array(q);
+  encoder.set_input_array(k);
+  encoder.set_input_array(v);
+  encoder.set_input_array(o);
+  encoder.set_input_array(stats);
+  encoder.set_input_array(d_o);
+  encoder.set_output_array(d_q);
+  encoder.set_output_array(d_k);
+  encoder.set_output_array(d_v);
+  if (mask_arr) {
+    encoder.set_input_array(*mask_arr);
+  }
+
+  // Search cache.
+  auto cache_key = build_sdpa_cache_key(encoder, q, k, v, do_causal, mask_arr);
+  auto it = sdpa_backward_cache().find(cache_key);
+  if (it == sdpa_backward_cache().end()) {
+    auto graph = build_sdpa_backward_graph(
+        handle, q, k, v, do_causal, mask_arr, o, d_o, stats, d_q, d_k, d_v);
+    it = sdpa_backward_cache().emplace(cache_key, std::move(graph)).first;
+  }
+  auto& graph = it->second;
+
+  std::unordered_map<int64_t, void*> variant_pack{
+      {Q, gpu_ptr<void>(q)},
+      {K, gpu_ptr<void>(k)},
+      {V, gpu_ptr<void>(v)},
+      {SCALE, &scale},
+      {O, gpu_ptr<void>(o)},
+      {STATS, gpu_ptr<void>(stats)},
+      {D_O, gpu_ptr<void>(d_o)},
+      {D_Q, gpu_ptr<void>(d_q)},
+      {D_K, gpu_ptr<void>(d_k)},
+      {D_V, gpu_ptr<void>(d_v)}};
+  if (mask_arr) {
+    variant_pack[BIAS] = gpu_ptr<void>(*mask_arr);
+  }
+
+  CHECK_CUDNN_FE_ERROR(graph.encode_graph(encoder, std::move(variant_pack)));
 }
 
 // Defined in scaled_dot_product_attention.cu file.
@@ -260,7 +381,8 @@ bool supports_sdpa_vector(
     const array& v,
     bool has_mask,
     bool has_arr_mask,
-    bool do_causal);
+    bool do_causal,
+    bool output_logsumexp);
 void sdpa_vector(
     const array& q,
     const array& k,
@@ -280,21 +402,25 @@ bool ScaledDotProductAttention::use_fallback(
     bool has_mask,
     bool has_arr_mask,
     bool do_causal,
+    bool is_training,
+    bool output_logsumexp,
     Stream s) {
-  if (detail::in_grad_tracing()) {
-    return true;
-  }
   if (s.device == Device::cpu) {
     return true;
   }
 
-  return !supports_sdpa_vector(q, k, v, has_mask, has_arr_mask, do_causal) &&
-      !supports_sdpa_cudnn(q, k, v, has_mask, do_causal, s);
+  return !supports_sdpa_vector(
+             q, k, v, has_mask, has_arr_mask, do_causal, output_logsumexp) &&
+      !supports_sdpa_cudnn(q, k, v, do_causal, s);
+}
+
+bool ScaledDotProductAttention::supports_bool_mask() {
+  return false;
 }
 
 void ScaledDotProductAttention::eval_gpu(
     const std::vector<array>& inputs,
-    array& out) {
+    std::vector<array>& outputs) {
   nvtx3::scoped_range r("ScaledDotProductAttention::eval_gpu");
 
   auto& s = stream();
@@ -302,20 +428,79 @@ void ScaledDotProductAttention::eval_gpu(
   array q = prepare_sdpa_input(inputs[0], s);
   array k = prepare_sdpa_input(inputs[1], s);
   array v = prepare_sdpa_input(inputs[2], s);
+  auto& out = outputs[0];
+  auto& stats = outputs[1];
   bool has_mask = inputs.size() - has_sinks_ > 3;
   bool has_arr_mask = has_mask && !do_causal_;
 
-  if (supports_sdpa_vector(q, k, v, has_mask, has_arr_mask, do_causal_)) {
+  std::optional<array> mask_arr;
+  if (has_arr_mask) {
+    mask_arr = prepare_sdpa_input(inputs[3], s);
+  }
+
+  if (supports_sdpa_vector(
+          q, k, v, has_mask, has_arr_mask, do_causal_, output_logsumexp_)) {
     if (has_sinks_) {
       sdpa_vector(q, k, v, scale_, out, do_causal_, inputs.back(), s);
     } else {
       sdpa_vector(q, k, v, scale_, out, do_causal_, std::nullopt, s);
     }
   } else {
-    sdpa_cudnn(q, k, v, scale_, out, do_causal_, s);
+    sdpa_cudnn(
+        q,
+        k,
+        v,
+        scale_,
+        out,
+        stats,
+        do_causal_,
+        mask_arr,
+        output_logsumexp_,
+        s);
   }
 }
 
+bool ScaledDotProductAttentionVJP::use_fallback(const array& q, Stream s) {
+  // The frontend adds a padding mask when sequence length is not a multiple of
+  // tile size.
+  if (q.shape(2) % 128 != 0) {
+    return true;
+  }
+  return s.device == Device::cpu;
+}
+
+void ScaledDotProductAttentionVJP::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("ScaledDotProductAttentionVJP::eval_gpu");
+
+  auto& s = stream();
+
+  assert(inputs.size() >= 6);
+  int primals_size = inputs.size() - 3;
+  bool has_arr_mask = primals_size > 3 + has_sinks_;
+
+  array q = prepare_sdpa_input(inputs[0], s);
+  array k = prepare_sdpa_input(inputs[1], s);
+  array v = prepare_sdpa_input(inputs[2], s);
+  array o = prepare_sdpa_input(inputs[primals_size], s);
+  array stats = prepare_sdpa_input(inputs[primals_size + 1], s);
+  array d_o = prepare_sdpa_input(inputs[primals_size + 2], s);
+
+  std::optional<array> mask_arr;
+  if (has_arr_mask) {
+    mask_arr = prepare_sdpa_input(inputs[3], s);
+  }
+
+  assert(outputs.size() == 3);
+  auto& d_q = outputs[0];
+  auto& d_k = outputs[1];
+  auto& d_v = outputs[2];
+
+  sdpa_backward_cudnn(
+      q, k, v, scale_, o, stats, do_causal_, mask_arr, d_o, d_q, d_k, d_v, s);
+}
+
 } // namespace fast
 
 } // namespace mlx::core

mlx/backend/cuda/scaled_dot_product_attention.cu

@@ -561,10 +561,9 @@ void sdpa_vector_2pass_fallback(
   array sums(intermediate_shape, float32, nullptr, {});
   array maxs(std::move(intermediate_shape), float32, nullptr, {});
 
-  intermediate.set_data(
-      cu::malloc_async(intermediate.nbytes(), encoder.stream()));
-  sums.set_data(cu::malloc_async(sums.nbytes(), encoder.stream()));
-  maxs.set_data(cu::malloc_async(maxs.nbytes(), encoder.stream()));
+  intermediate.set_data(cu::malloc_async(intermediate.nbytes(), encoder));
+  sums.set_data(cu::malloc_async(sums.nbytes(), encoder));
+  maxs.set_data(cu::malloc_async(maxs.nbytes(), encoder));
 
   encoder.add_temporary(intermediate);
   encoder.add_temporary(sums);
@@ -665,7 +664,12 @@ bool supports_sdpa_vector(
     const array& v,
     bool has_mask,
     bool has_arr_mask,
-    bool do_causal) {
+    bool do_causal,
+    bool output_logsumexp) {
+  if (output_logsumexp) {
+    return false;
+  }
+
   const int value_head_dim = v.shape(-1);
   const int query_head_dim = q.shape(-1);
   const int query_sequence_length = q.shape(2);
@@ -769,7 +773,7 @@ void sdpa_vector(
       };
 
       o.set_data(
-          cu::malloc_async(o.nbytes(), encoder.stream()),
+          cu::malloc_async(o.nbytes(), encoder),
           o.size(),
           {str_oB, str_oH, str_oL, str_oD},
           flags);

mlx/backend/cuda/scan.cu

@@ -374,7 +374,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
       out.copy_shared_buffer(in);
     } else {
       out.set_data(
-          cu::malloc_async(in.data_size() * out.itemsize(), encoder.stream()),
+          cu::malloc_async(in.data_size() * out.itemsize(), encoder),
           in.data_size(),
           in.strides(),
           in.flags());

mlx/backend/cuda/slicing.cpp

@@ -24,7 +24,7 @@ void concatenate_gpu(
   std::partial_sum(sizes.cbegin(), sizes.cend(), sizes.begin());
 
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   auto strides = out.strides();
   auto flags = out.flags();
@@ -89,7 +89,7 @@ array compute_dynamic_offset(
   if (donate) {
     offset.copy_shared_buffer(indices);
   } else {
-    offset.set_data(cu::malloc_async(offset.itemsize(), encoder.stream()));
+    offset.set_data(cu::malloc_async(offset.itemsize(), encoder));
   }
 
   encoder.add_temporary(offset);

mlx/backend/cuda/softmax.cu

@@ -118,7 +118,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
         out.copy_shared_buffer(x);
       } else {
         out.set_data(
-            cu::malloc_async(x.data_size() * x.itemsize(), encoder.stream()),
+            cu::malloc_async(x.data_size() * x.itemsize(), encoder),
             x.data_size(),
             x.strides(),
             x.flags());

mlx/backend/cuda/sort.cu

@@ -49,14 +49,12 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
     array trans = swapaxes_in_eval(in, axis, last_dim);
     in = contiguous_copy_gpu(trans, s);
     encoder.add_temporary(in);
-    out = array(
-        cu::malloc_async(out.nbytes(), encoder.stream()),
-        in.shape(),
-        out.dtype());
+    out =
+        array(cu::malloc_async(out.nbytes(), encoder), in.shape(), out.dtype());
     encoder.add_temporary(out);
   } else {
     out.set_data(
-        cu::malloc_async(in.data_size() * out.itemsize(), encoder.stream()),
+        cu::malloc_async(in.data_size() * out.itemsize(), encoder),
         in.data_size(),
         in.strides(),
         in.flags());
@@ -74,17 +72,13 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
       if (argsort) {
         // Indices in the sorted dimension.
         array indices(
-            cu::malloc_async(out.nbytes(), encoder.stream()),
-            in.shape(),
-            out.dtype());
+            cu::malloc_async(out.nbytes(), encoder), in.shape(), out.dtype());
         encoder.add_temporary(indices);
 
         // In argsort though we don't need the result of sorted values, the
         // API requires us to provide an array to store it.
         array discard(
-            cu::malloc_async(in.nbytes(), encoder.stream()),
-            in.shape(),
-            in.dtype());
+            cu::malloc_async(in.nbytes(), encoder), in.shape(), in.dtype());
         encoder.add_temporary(discard);
 
         size_t size;
@@ -104,9 +98,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             stream));
 
         array temp(
-            cu::malloc_async(size, encoder.stream()),
-            {static_cast<int>(size)},
-            uint8);
+            cu::malloc_async(size, encoder), {static_cast<int>(size)}, uint8);
         encoder.add_temporary(temp);
 
         // Start capturing after allocations
@@ -148,9 +140,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             stream));
 
         array temp(
-            cu::malloc_async(size, encoder.stream()),
-            {static_cast<int>(size)},
-            uint8);
+            cu::malloc_async(size, encoder), {static_cast<int>(size)}, uint8);
         encoder.add_temporary(temp);
 
         // Start capturing after allocations

mlx/backend/cuda/ternary.cu

@@ -257,9 +257,8 @@ void ternary_op_gpu(
   auto& c = inputs[2];
   auto topt = get_ternary_op_type(a, b, c);
   auto& encoder = cu::get_command_encoder(s);
-  set_ternary_op_output_data(a, b, c, out, topt, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  set_ternary_op_output_data(
+      a, b, c, out, topt, [&](auto n) { return cu::malloc_async(n, encoder); });
   ternary_op_gpu_inplace<Op>(inputs, out, s);
 }
 

mlx/backend/cuda/unary/unary.cuh

@@ -208,9 +208,8 @@ void unary_op_gpu(
     const char* op,
     const Stream& s) {
   auto& encoder = cu::get_command_encoder(s);
-  set_unary_output_data(inputs[0], out, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  set_unary_output_data(
+      inputs[0], out, [&](auto n) { return cu::malloc_async(n, encoder); });
   unary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 

mlx/backend/cuda/utils.cpp

@@ -5,6 +5,7 @@
 #include "mlx/dtype_utils.h"
 
 #include <fmt/format.h>
+#include <vector>
 
 namespace mlx::core {
 
@@ -31,6 +32,13 @@ void check_cuda_error(const char* name, CUresult err) {
   }
 }
 
+void check_cudnn_error(const char* name, cudnnStatus_t err) {
+  if (err != CUDNN_STATUS_SUCCESS) {
+    throw std::runtime_error(
+        fmt::format("{} failed: {}.", name, cudnnGetErrorString(err)));
+  }
+}
+
 const char* dtype_to_cuda_type(const Dtype& dtype) {
   switch (dtype) {
     case bool_:
@@ -72,7 +80,6 @@ CudaGraph::CudaGraph(cu::Device& device) {
 }
 
 void CudaGraph::end_capture(cudaStream_t stream) {
-  assert(handle_ == nullptr);
   CHECK_CUDA_ERROR(cudaStreamEndCapture(stream, &handle_));
 }
 

mlx/backend/cuda/utils.h

@@ -31,8 +31,10 @@ inline T* gpu_ptr(array& arr) {
       arr.offset());
 }
 
+// For const array, keep constness in pointer unless it is untyped.
 template <typename T>
-inline const T* gpu_ptr(const array& arr) {
+inline std::conditional_t<std::is_same_v<T, void>, void*, const T*> gpu_ptr(
+    const array& arr) {
   return gpu_ptr<T>(const_cast<array&>(arr));
 }
 

mlx/backend/gpu/copy.cpp

@@ -7,8 +7,6 @@
 
 namespace mlx::core {
 
-void copy_gpu(const array& in, array& out, CopyType ctype, const Stream& s);
-
 void copy_gpu(const array& in, array& out, CopyType ctype) {
   copy_gpu(in, out, ctype, out.primitive().stream());
 }

mlx/backend/metal/CMakeLists.txt

@@ -28,6 +28,7 @@ make_jit_source(binary_ops)
 make_jit_source(ternary_ops)
 make_jit_source(reduce_utils kernels/atomic.h kernels/reduction/ops.h)
 make_jit_source(indexing/scatter kernels/indexing/indexing.h)
+make_jit_source(indexing/masked_scatter)
 make_jit_source(indexing/gather kernels/indexing/indexing.h)
 make_jit_source(indexing/gather_front kernels/indexing/indexing.h)
 make_jit_source(indexing/gather_axis)

mlx/backend/metal/allocator.cpp

@@ -149,7 +149,9 @@ Buffer MetalAllocator::malloc(size_t size) {
       buf = device_->newBuffer(size, resource_options);
     }
     if (!buf) {
-      return Buffer{nullptr};
+      std::ostringstream msg;
+      msg << "[malloc] Unable to allocate " << size << " bytes.";
+      throw std::runtime_error(msg.str());
     }
     lk.lock();
     num_resources_++;
@@ -201,6 +203,32 @@ size_t MetalAllocator::size(Buffer buffer) const {
   return static_cast<MTL::Buffer*>(buffer.ptr())->length();
 }
 
+Buffer MetalAllocator::make_buffer(void* ptr, size_t size) {
+  auto buf = device_->newBuffer(ptr, size, resource_options, nullptr);
+  if (!buf) {
+    return Buffer{nullptr};
+  }
+  std::unique_lock lk(mutex_);
+  residency_set_.insert(buf);
+  active_memory_ += buf->length();
+  peak_memory_ = std::max(peak_memory_, active_memory_);
+  num_resources_++;
+  return Buffer{static_cast<void*>(buf)};
+}
+
+void MetalAllocator::release(Buffer buffer) {
+  auto buf = static_cast<MTL::Buffer*>(buffer.ptr());
+  if (buf == nullptr) {
+    return;
+  }
+  std::unique_lock lk(mutex_);
+  active_memory_ -= buf->length();
+  num_resources_--;
+  lk.unlock();
+  auto pool = metal::new_scoped_memory_pool();
+  buf->release();
+}
+
 MetalAllocator& allocator() {
   // By creating the |allocator_| on heap, the destructor of MetalAllocator
   // will not be called on exit and buffers in the cache will be leaked. This

mlx/backend/metal/allocator.h

@@ -21,6 +21,9 @@ class MetalAllocator : public allocator::Allocator {
   virtual Buffer malloc(size_t size) override;
   virtual void free(Buffer buffer) override;
   virtual size_t size(Buffer buffer) const override;
+  virtual Buffer make_buffer(void* ptr, size_t size) override;
+  virtual void release(Buffer buffer) override;
+
   size_t get_active_memory() {
     return active_memory_;
   };

mlx/backend/metal/device.h

@@ -265,4 +265,19 @@ Device& device(mlx::core::Device);
 
 std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool();
 
+inline bool is_nax_available() {
+  auto _check_nax = []() {
+    bool can_use_nax = false;
+    if (__builtin_available(
+            macOS 26.2, iOS 26.2, tvOS 26.2, visionOS 26.2, *)) {
+      can_use_nax = true;
+    }
+    can_use_nax &=
+        metal::device(mlx::core::Device::gpu).get_architecture_gen() >= 17;
+    return can_use_nax;
+  };
+  static bool is_nax_available_ = _check_nax();
+  return is_nax_available_;
+}
+
 } // namespace mlx::core::metal

mlx/backend/metal/indexing.cpp

@@ -1,4 +1,5 @@
 // Copyright © 2023-2024 Apple Inc.
+
 #include <fmt/format.h>
 
 #include "mlx/backend/common/compiled.h"
@@ -8,7 +9,9 @@
 #include "mlx/backend/metal/jit/includes.h"
 #include "mlx/backend/metal/jit/indexing.h"
 #include "mlx/backend/metal/kernels.h"
+#include "mlx/backend/metal/scan.h"
 #include "mlx/backend/metal/utils.h"
+#include "mlx/dtype.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
 
@@ -641,4 +644,84 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   compute_encoder.dispatch_threads(grid_dims, group_dims);
 }
 
+void MaskedScatter::eval_gpu(const std::vector<array>& inputs, array& out) {
+  const array& dst = inputs[0];
+  const array& mask = inputs[1];
+  const array& src = inputs[2];
+
+  auto& s = stream();
+  auto& d = metal::device(s.device);
+
+  const size_t total = mask.size();
+  const CopyType ct = (total == 1)
+      ? CopyType::Scalar
+      : (dst.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+  copy_gpu(dst, out, ct, s);
+  if (total == 0) {
+    return;
+  }
+
+  array mask_flat = flatten_in_eval(mask, 1, -1, s);
+  if (mask_flat.data<void>() != mask.data<void>()) {
+    d.add_temporary(mask_flat, s.index);
+  }
+
+  if (!mask_flat.flags().row_contiguous) {
+    mask_flat = contiguous_copy_gpu(mask_flat, s);
+    d.add_temporary(mask_flat, s.index);
+  }
+
+  // Prefix (exclusive) of mask → scatter_offsets
+  array scatter_offsets(mask_flat.shape(), uint32, nullptr, {});
+  scatter_offsets.set_data(allocator::malloc(scatter_offsets.nbytes()));
+  d.add_temporary(scatter_offsets, s.index);
+
+  scan_gpu_inplace(
+      mask_flat,
+      scatter_offsets,
+      Scan::Sum,
+      /*axis=*/1,
+      /*reverse=*/false,
+      /*inclusive=*/false,
+      s);
+
+  // Kernel selection/build
+  static constexpr std::string_view kBaseName = "masked_assign";
+  const std::string dtype_tag = type_to_name(out.dtype());
+  const std::string value_type = get_type_string(out.dtype());
+  const std::string contiguous =
+      (src.flags().row_contiguous) ? "true" : "false";
+  const std::string kernel_name =
+      fmt::format("{}_{}_{}", kBaseName, dtype_tag, contiguous);
+
+  auto lib = d.get_library(kernel_name, [&]() {
+    std::string source = metal::utils();
+    source += metal::masked_scatter();
+    source += fmt::format(
+        std::string(masked_assign_kernel), kernel_name, value_type, contiguous);
+    return source;
+  });
+  auto kernel = d.get_kernel(kernel_name, lib);
+
+  // Binding
+  int bind_idx = 0;
+  const int ndim = static_cast<int>(src.ndim());
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder.set_compute_pipeline_state(kernel);
+  compute_encoder.set_input_array(mask_flat, bind_idx++);
+  compute_encoder.set_input_array(scatter_offsets, bind_idx++);
+  compute_encoder.set_input_array(src, bind_idx++);
+  compute_encoder.set_output_array(out, bind_idx++);
+  compute_encoder.set_vector_bytes(src.shape(), bind_idx++);
+  compute_encoder.set_vector_bytes(src.strides(), bind_idx++);
+  compute_encoder.set_bytes(ndim, bind_idx++);
+  compute_encoder.set_bytes(src.size() / src.shape(0), bind_idx++);
+  compute_encoder.set_bytes(mask_flat.size() / mask.shape(0), bind_idx++);
+
+  // Dispatch
+  auto group_dims = get_block_dims(total, 1, 1);
+  MTL::Size grid_dims(total, 1, 1);
+  compute_encoder.dispatch_threads(grid_dims, group_dims);
+}
+
 } // namespace mlx::core

mlx/backend/metal/jit/includes.h

@@ -11,6 +11,7 @@ const char* ternary_ops();
 const char* reduce_utils();
 const char* gather();
 const char* scatter();
+const char* masked_scatter();
 
 const char* arange();
 const char* unary();

mlx/backend/metal/jit/indexing.h

@@ -70,3 +70,7 @@ constexpr std::string_view scatter_kernels = R"(
       gid);
 }}
 )";
+
+constexpr std::string_view masked_assign_kernel = R"(
+template [[host_name("{0}")]] [[kernel]] decltype(masked_assign_impl<{1}, {2}>) masked_assign_impl<{1}, {2}>;
+)";

mlx/backend/metal/kernels/CMakeLists.txt

@@ -9,7 +9,14 @@ set(BASE_HEADERS
     utils.h)
 
 function(build_kernel_base TARGET SRCFILE DEPS)
-  set(METAL_FLAGS -Wall -Wextra -fno-fast-math -Wno-c++17-extensions)
+  set(METAL_FLAGS
+      -x
+      metal
+      -Wall
+      -Wextra
+      -fno-fast-math
+      -Wno-c++17-extensions
+      -Wno-c++20-extensions)
   if(MLX_METAL_DEBUG)
     set(METAL_FLAGS ${METAL_FLAGS} -gline-tables-only -frecord-sources)
   endif()
@@ -120,6 +127,30 @@ if(NOT MLX_METAL_JIT)
   build_kernel(gemv_masked steel/utils.h)
 endif()
 
+if((MLX_METAL_VERSION GREATER_EQUAL 400) AND (MACOS_SDK_VERSION GREATER_EQUAL
+                                              26.2))
+  set(STEEL_NAX_HEADERS
+      steel/defines.h
+      steel/utils.h
+      steel/gemm/transforms.h
+      steel/gemm/nax.h
+      steel/gemm/gemm_nax.h
+      steel/utils/type_traits.h
+      steel/utils/integral_constant.h)
+
+  build_kernel(steel/gemm/kernels/steel_gemm_fused_nax ${STEEL_NAX_HEADERS})
+  build_kernel(steel/gemm/kernels/steel_gemm_gather_nax ${STEEL_NAX_HEADERS})
+
+  build_kernel(quantized_nax quantized_nax.h ${STEEL_NAX_HEADERS})
+  build_kernel(fp_quantized_nax fp_quantized_nax.h ${STEEL_NAX_HEADERS})
+
+  set(STEEL_NAX_ATTN_HEADERS
+      steel/defines.h steel/utils.h steel/attn/nax.h steel/utils/type_traits.h
+      steel/utils/integral_constant.h)
+
+  build_kernel(steel/attn/kernels/steel_attention_nax ${STEEL_NAX_ATTN_HEADERS})
+endif()
+
 add_custom_command(
   OUTPUT ${MLX_METAL_PATH}/mlx.metallib
   COMMAND xcrun -sdk macosx metallib ${KERNEL_AIR} -o

mlx/backend/metal/kernels/fp_quantized_nax.metal

@@ -0,0 +1,74 @@
+// Copyright © 2025 Apple Inc.
+
+// clang-format off
+#include "mlx/backend/metal/kernels/utils.h"
+#include "mlx/backend/metal/kernels/steel/gemm/gemm.h"
+#include "mlx/backend/metal/kernels/quantized_utils.h"
+#include "mlx/backend/metal/kernels/steel/gemm/nax.h"
+#include "mlx/backend/metal/kernels/fp_quantized_nax.h"
+
+
+#define instantiate_quantized_batched(mode, name, type, bm, bn, bk, wm, wn, batched) \
+  instantiate_kernel( \
+      #mode "_" #name "_" #type "_gs_32_b_4_bm" #bm "_bn" #bn "_bk" #bk "_wm" #wm "_wn" #wn "_batch_" #batched, \
+      fp_ ## name,  \
+      type,         \
+      32,           \
+      4,            \
+      batched)
+
+#define instantiate_quantized_aligned(mode, name, type, bm, bn, bk, wm, wn, aligned) \
+  instantiate_kernel( \
+      #mode "_" #name "_" #type "_gs_32_b_4_bm" #bm "_bn" #bn "_bk" #bk "_wm" #wm "_wn" #wn "_alN_" #aligned, \
+      fp_ ## name, \
+      type,        \
+      32,          \
+      4,           \
+      aligned)
+
+#define instantiate_quantized_aligned_batched(mode, name, type, bm, bn, bk, wm, wn, aligned, batched) \
+  instantiate_kernel( \
+      #mode "_" #name "_" #type "_gs_32_b_4_bm" #bm "_bn" #bn "_bk" #bk "_wm" #wm "_wn" #wn "_alN_" #aligned "_batch_" #batched, \
+      fp_ ## name,    \
+      type,    \
+      32,      \
+      4,       \
+      aligned, \
+      batched)
+
+#define instantiate_gather_qmm_rhs(func, name, type, bm, bn, bk, wm, wn, transpose) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_bm_" #bm "_bn_" #bn "_bk_" #bk "_wm_" #wm "_wn_" #wn, \
+      func,    \
+      type,    \
+      32,      \
+      4,       \
+      bm,      \
+      bn,      \
+      bk,      \
+      wm,      \
+      wn,      \
+      transpose)
+
+
+#define instantiate_quantized_all_aligned(type) \
+  instantiate_quantized_aligned(mxfp4, gather_qmm_t_nax, type, 64, 64, 64, 2, 2, true)      \
+  instantiate_quantized_aligned(mxfp4, gather_qmm_t_nax, type, 64, 64, 64, 2, 2, false)     \
+  instantiate_quantized_aligned_batched(mxfp4, qmm_t_nax, type, 64, 64, 64, 2, 2, true, 1)  \
+  instantiate_quantized_aligned_batched(mxfp4, qmm_t_nax, type, 64, 64, 64, 2, 2, true, 0)  \
+  instantiate_quantized_aligned_batched(mxfp4, qmm_t_nax, type, 64, 64, 64, 2, 2, false, 1) \
+  instantiate_quantized_aligned_batched(mxfp4, qmm_t_nax, type, 64, 64, 64, 2, 2, false, 0)
+
+
+#define instantiate_quantized_all_rhs(type) \
+  instantiate_gather_qmm_rhs(fp_gather_qmm_rhs_nax, mxfp4_gather_qmm_rhs_nax_nt, type, 64, 64, 64, 2, 2, true) \
+  instantiate_gather_qmm_rhs(fp_gather_qmm_rhs_nax, mxfp4_gather_qmm_rhs_nax_nn, type, 64, 64, 64, 2, 2, false) 
+
+#define instantiate_quantized_types(type) \
+  instantiate_quantized_all_aligned(type) \
+  instantiate_quantized_all_rhs(type)
+
+instantiate_quantized_types(float)
+instantiate_quantized_types(bfloat16_t)
+instantiate_quantized_types(float16_t)
+    // clang-format on

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

@@ -0,0 +1,38 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+template <typename T, bool src_contiguous>
+[[kernel]] void masked_assign_impl(
+    const device bool* mask [[buffer(0)]],
+    const device uint* scatter_offsets [[buffer(1)]],
+    const device T* src [[buffer(2)]],
+    device T* out [[buffer(3)]],
+    const constant int* src_shapes [[buffer(4)]],
+    const constant int64_t* src_strides [[buffer(5)]],
+    const constant int& src_ndim [[buffer(6)]],
+    const constant int64_t& src_batch_size [[buffer(7)]],
+    const constant int64_t& mask_batch_size [[buffer(8)]],
+    uint idx [[thread_position_in_grid]]) {
+  const bool mask_value = mask[idx];
+  if (!mask_value) {
+    return;
+  }
+
+  const uint src_index = scatter_offsets[idx];
+  if (src_index >= src_batch_size) {
+    return;
+  }
+
+  const uint batch_idx = idx / mask_batch_size;
+
+  if (src_contiguous) {
+    out[idx] = src[batch_idx * src_batch_size + src_index];
+  } else {
+    out[idx] = src[elem_to_loc<uint>(
+        batch_idx * src_batch_size + src_index,
+        src_shapes,
+        src_strides,
+        src_ndim)];
+  }
+}

mlx/backend/metal/kernels/quantized_nax.metal

@@ -0,0 +1,106 @@
+// Copyright © 2023-2024 Apple Inc.
+
+// clang-format off
+#include "mlx/backend/metal/kernels/utils.h"
+#include "mlx/backend/metal/kernels/steel/gemm/gemm.h"
+#include "mlx/backend/metal/kernels/steel/gemm/nax.h"
+#include "mlx/backend/metal/kernels/steel/gemm/loader.h"
+#include "mlx/backend/metal/kernels/quantized_nax.h"
+
+#define instantiate_quantized(name, type, group_size, bits, bm, bn, bk, wm, wn)  \
+  instantiate_kernel(                                                    \
+      #name "_" #type "_gs_" #group_size "_b_" #bits,                    \
+      name,                                                              \
+      type,                                                              \
+      group_size,                                                        \
+      bits, bm, bk, bn, wm, wn)
+
+#define instantiate_quantized_batched(name, type, group_size, bits, bm, bn, bk, wm, wn, batched)     \
+  instantiate_kernel(                                                    \
+      #name "_" #type "_gs_" #group_size "_b_" #bits "_bm" #bm "_bn" #bn "_bk" #bk "_wm" #wm "_wn" #wn "_batch_" #batched, \
+      name,                                                              \
+      type,                                                              \
+      group_size,                                                        \
+      bits,                                                              \
+      batched, bm, bk, bn, wm, wn)
+
+#define instantiate_quantized_aligned(name, type, group_size, bits, bm, bn, bk, wm, wn, aligned)     \
+  instantiate_kernel(                                                                     \
+      #name "_" #type "_gs_" #group_size "_b_" #bits "_bm" #bm "_bn" #bn "_bk" #bk "_wm" #wm "_wn" #wn "_alN_" #aligned, \
+      name,                                                                  \
+      type,                                                                  \
+      group_size,                                                            \
+      bits,                                                                  \
+      aligned, bm, bk, bn, wm, wn)
+
+#define instantiate_quantized_aligned_batched(name, type, group_size, bits, bm, bn, bk, wm, wn, aligned, batched)     \
+  instantiate_kernel(                                                                     \
+      #name "_" #type "_gs_" #group_size "_b_" #bits "_bm" #bm "_bn" #bn "_bk" #bk "_wm" #wm "_wn" #wn "_alN_" #aligned "_batch_" #batched, \
+      name,                                                                  \
+      type,                                                                  \
+      group_size,                                                            \
+      bits,                                                                  \
+      aligned,                                                               \
+      batched, bm, bk, bn, wm, wn)
+
+#define instantiate_gather_qmm_rhs(func, name, type, group_size, bits, bm, bn, bk, wm, wn, transpose)        \
+  instantiate_kernel(                                                                                        \
+      #name "_" #type "_gs_" #group_size "_b_" #bits "_bm_" #bm "_bn_" #bn "_bk_" #bk "_wm_" #wm "_wn_" #wn, \
+      func,                                                         \
+      type,                                                         \
+      group_size,                                                   \
+      bits,                                                         \
+      bm,                                                           \
+      bn,                                                           \
+      bk,                                                           \
+      wm,                                                           \
+      wn,                                                           \
+      transpose)
+
+#define instantiate_quantized_batched_wrap(name, type, group_size, bits) \
+  instantiate_quantized_batched(name, type, group_size, bits, 64, 64, 64, 2, 2, 1)      \
+  instantiate_quantized_batched(name, type, group_size, bits, 64, 64, 64, 2, 2, 0)
+
+#define instantiate_quantized_all_batched(type, group_size, bits) \
+  instantiate_quantized_batched_wrap(affine_qmm_n_nax, type, group_size, bits)
+
+
+#define instantiate_quantized_all_single(type, group_size, bits) \
+  instantiate_quantized(affine_gather_qmm_n_nax, type, group_size, bits, 64, 64, 64, 2, 2)
+
+#define instantiate_quantized_all_aligned(type, group_size, bits)   \
+  instantiate_quantized_aligned(affine_gather_qmm_t_nax, type, group_size, bits, 64, 64, 64, 2, 2, true) \
+  instantiate_quantized_aligned(affine_gather_qmm_t_nax, type, group_size, bits, 64, 64, 64, 2, 2, false) \
+  instantiate_quantized_aligned_batched(affine_qmm_t_nax, type, group_size, bits, 64, 64, 64, 2, 2, true, 1) \
+  instantiate_quantized_aligned_batched(affine_qmm_t_nax, type, group_size, bits, 64, 64, 64, 2, 2, true, 0) \
+  instantiate_quantized_aligned_batched(affine_qmm_t_nax, type, group_size, bits, 64, 64, 64, 2, 2, false, 1) \
+  instantiate_quantized_aligned_batched(affine_qmm_t_nax, type, group_size, bits, 64, 64, 64, 2, 2, false, 0)
+
+#define instantiate_quantized_all_rhs(type, group_size, bits) \
+  instantiate_gather_qmm_rhs(affine_gather_qmm_rhs_nax, affine_gather_qmm_rhs_nax_nt, type, group_size, bits, 64, 64, 64, 2, 2, true) \
+  instantiate_gather_qmm_rhs(affine_gather_qmm_rhs_nax, affine_gather_qmm_rhs_nax_nn, type, group_size, bits, 64, 64, 64, 2, 2, false)
+
+#define instantiate_quantized_funcs(type, group_size, bits) \
+  instantiate_quantized_all_batched(type, group_size, bits) \
+  instantiate_quantized_all_aligned(type, group_size, bits) \
+  instantiate_quantized_all_rhs(type, group_size, bits)
+
+#define instantiate_quantized_types(group_size, bits)       \
+  instantiate_quantized_funcs(float, group_size, bits)      \
+  instantiate_quantized_funcs(float16_t, group_size, bits)  \
+  instantiate_quantized_funcs(bfloat16_t, group_size, bits)  
+
+#define instantiate_quantized_groups(bits) \
+  instantiate_quantized_types(128, bits)   \
+  instantiate_quantized_types(64, bits)    \
+  instantiate_quantized_types(32, bits)
+
+#define instantiate_quantized_all() \
+  instantiate_quantized_groups(2) \
+  instantiate_quantized_groups(3) \
+  instantiate_quantized_groups(4) \
+  instantiate_quantized_groups(5) \
+  instantiate_quantized_groups(6) \
+  instantiate_quantized_groups(8)
+
+instantiate_quantized_all() // clang-format on

mlx/backend/metal/kernels/scan.metal

@@ -51,6 +51,7 @@ using namespace metal;
   instantiate_strided_scan(reverse_exclusive_##name, itype, otype, op, false, true, nreads)
 
 instantiate_scan_helper(sum_bool__int32,         bool,        int32_t,     CumSum, 4)
+instantiate_scan_helper(sum_bool__uint32,        bool,        uint32_t,    CumSum, 4)
 instantiate_scan_helper(sum_uint8_uint8,         uint8_t,     uint8_t,     CumSum, 4)
 instantiate_scan_helper(sum_uint16_uint16,       uint16_t,    uint16_t,    CumSum, 4)
 instantiate_scan_helper(sum_uint32_uint32,       uint32_t,    uint32_t,    CumSum, 4)

mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_nax.h

@@ -0,0 +1,476 @@
+// Copyright © 2024-25 Apple Inc.
+
+using namespace mlx::steel;
+
+///////////////////////////////////////////////////////////////////////////////
+// GEMM kernels
+///////////////////////////////////////////////////////////////////////////////
+
+constant bool align_Q [[function_constant(200)]];
+constant bool align_K [[function_constant(201)]];
+
+constant bool has_mask [[function_constant(300)]];
+constant bool do_causal [[function_constant(301)]];
+constant bool has_sinks [[function_constant(302)]];
+
+template <typename T>
+struct TransformScale {
+  T scale;
+  METAL_FUNC TransformScale(T scale_) : scale(scale_) {}
+
+  METAL_FUNC T apply(T x) const {
+    return scale * x;
+  }
+};
+
+struct MaxOp {
+  template <typename T>
+  METAL_FUNC static constexpr T apply(T x, T y) {
+    return metal::max(x, y);
+  }
+};
+
+struct SumOp {
+  template <typename T>
+  METAL_FUNC static constexpr T apply(T x, T y) {
+    return x + y;
+  }
+};
+
+struct MulOp {
+  template <typename T>
+  METAL_FUNC static constexpr T apply(T x, T y) {
+    return x * y;
+  }
+};
+
+struct SubOp {
+  template <typename T>
+  METAL_FUNC static constexpr T apply(T x, T y) {
+    return x - y;
+  }
+};
+
+struct ExpSubOp {
+  template <typename T>
+  METAL_FUNC static constexpr T apply(T x, T y) {
+    return fast::exp2(x - y);
+  }
+};
+
+struct DivOp {
+  template <typename T>
+  METAL_FUNC static constexpr T apply(T x, T y) {
+    return x / y;
+  }
+};
+
+// clang-format off
+template <
+    typename T,
+    int BQ,
+    int BK,
+    int BD,
+    int WM,
+    int WN,
+    typename MaskType = float,
+    typename AccumType = float>
+[[kernel, max_total_threads_per_threadgroup(WM * WN * 32)]] void attention_nax(
+    const device T* Q [[buffer(0)]],
+    const device T* K [[buffer(1)]],
+    const device T* V [[buffer(2)]],
+    device T* O [[buffer(3)]],
+    const constant AttnParams* params [[buffer(4)]],
+    const constant AttnMaskParams* mask_params [[buffer(5), function_constant(has_mask)]],
+    const device MaskType* mask [[buffer(6), function_constant(has_mask)]],
+    const device T* sinks [[buffer(7), function_constant(has_sinks)]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]) { // clang-format on
+
+  // Pacifying compiler
+  (void)lid;
+  (void)simd_lane_id;
+
+  // Move to correct block
+  ulong3 tidl{tid.x, tid.y, tid.z};
+
+  Q += tidl.z * params->Q_strides[0] + // Batch
+      tidl.y * params->Q_strides[1] + // Head
+      tidl.x * BQ * params->Q_strides[2]; // Sequence
+
+  ulong kv_head_idx = int(tid.y) / params->gqa_factor;
+  K += tidl.z * params->K_strides[0] + // Batch
+      kv_head_idx * params->K_strides[1]; // Head
+
+  V += tidl.z * params->V_strides[0] + // Batch
+      kv_head_idx * params->V_strides[1]; // Head
+
+  O += tidl.z * params->O_strides[0] + // Batch
+      tidl.y * params->O_strides[1] + // Head
+      tidl.x * BQ * params->O_strides[2]; // Sequence
+
+  if (has_mask) {
+    mask += tidl.z * mask_params->M_strides[0] + // Batch
+        tidl.y * mask_params->M_strides[1]; // Head
+  }
+
+  const metal::uniform<float> scale2 =
+      make_uniform(params->scale) * make_uniform(1.44269504089f);
+
+  // Prepare MMA tiles
+  constexpr short UQ = 16;
+  constexpr short UD = 32;
+
+  constexpr int kNWarps = WM * WN;
+  static_assert(
+      BQ >= (kNWarps * UQ) && BQ % (kNWarps * UQ) == 0,
+      "Each simdgroup must host atleast 1 simdgroup matrix along Q sequence.");
+
+  // Q seq frags per warp
+  constexpr int TQ = BQ / (kNWarps * UQ);
+  // HeadDim frags (all warps load the same frags)
+  constexpr int TD = BD / UD;
+
+  static_assert(TQ == 1, "Check TQ");
+
+  using OSubTile = NAXSubTile<AccumType, UQ, UD>;
+  NAXTile<AccumType, TQ, TD, OSubTile> Otile;
+
+  Otile.clear();
+
+  // Prepare mma tile offsets
+  const short2 simd_coord = OSubTile::NAXFrag_t::get_coord();
+  const short sm = simd_coord.y;
+  const short sn = simd_coord.x;
+  const short tm = UQ * TQ * simd_group_id;
+
+  Q += (tm + sm) * int(params->Q_strides[2]) + sn;
+  K += sm * int(params->K_strides[2]) + sn;
+  V += sm * int(params->V_strides[2]) + sn;
+
+  // Init row reduction variables
+  constexpr short kRowsPT = decltype(Otile)::kRowsPerThread;
+
+  metal::vec<AccumType, kRowsPT> max_score;
+  metal::vec<AccumType, kRowsPT> sum_score{0};
+
+  // Init to -Inf
+  STEEL_PRAGMA_UNROLL
+  for (short i = 0; i < kRowsPT; ++i) {
+    max_score[i] = Limits<AccumType>::finite_min;
+  }
+
+  if (has_sinks) {
+    STEEL_PRAGMA_UNROLL
+    for (short i = 0; i < kRowsPT; ++i) {
+      max_score[i] = M_LOG2E_F * static_cast<AccumType>(sinks[tidl.y]);
+      sum_score[i] = 1;
+    }
+  }
+
+  int kb_lim = params->NK;
+
+  if (do_causal) {
+    int q_max = (tid.x + 1) * BQ + params->qL_off;
+    kb_lim = (q_max + BK - 1) / BK;
+    kb_lim = min(params->NK, kb_lim);
+  }
+
+  const bool is_last_bq = int(tid.x) == (params->NQ_aligned);
+  // const bool is_last_tq = int(simd_group_id) >= (params->qL_rem / UQ);
+  const bool is_last_q = is_last_bq;
+
+  const short lim_rows_q = params->qL_rem - (tm + sm);
+  const short lim_rows_k = params->kL_rem - sm;
+
+  // Loop over KV seq length
+  for (int kb = 0; kb < kb_lim; kb++) {
+    const int is_last_k = (kb == (params->NK_aligned));
+
+    // Do S = Q @ K.T
+    constexpr short UDs = 16;
+    constexpr short UKs = 32;
+
+    constexpr short TDs = BD / UDs;
+    constexpr short TKs = BK / UKs;
+
+    using SSubTile = NAXSubTile<AccumType, UQ, UKs>;
+    using QSubTile = NAXSubTile<T, UQ, UDs>;
+    using KSubTile = NAXSubTile<T, UKs, UDs>;
+
+    NAXTile<AccumType, TQ, TKs, SSubTile> Stile;
+
+    Stile.clear();
+
+    STEEL_PRAGMA_UNROLL
+    for (short iq = 0; iq < TQ; iq++) {
+      STEEL_PRAGMA_UNROLL
+      for (short ik = 0; ik < TKs; ik++) {
+        STEEL_PRAGMA_UNROLL
+        for (short id = 0; id < TDs; id++) {
+          NAXTile<T, 1, 1, QSubTile> Qtile;
+          NAXTile<T, 1, 1, KSubTile> Ktile;
+
+          const int Q_load_off = iq * UQ * int(params->Q_strides[2]) + id * UDs;
+          const int K_load_off =
+              ik * UKs * int(params->K_strides[2]) + id * UDs;
+
+          if (!align_Q && is_last_q) {
+            // Qtile.load_rows(
+            //     Q + Q_load_off,
+            //     int(params->Q_strides[2]),
+            //     lim_rows_q - iq * UQ);
+            Qtile.load_safe(
+                Q + Q_load_off,
+                int(params->Q_strides[2]),
+                short2(BD, lim_rows_q - iq * UQ));
+          } else {
+            Qtile.load(Q + Q_load_off, int(params->Q_strides[2]));
+          }
+
+          if (!align_K && is_last_k) {
+            // Ktile.load_rows(
+            //     K + K_load_off,
+            //     int(params->K_strides[2]),
+            //     lim_rows_k - ik * UKs);
+            Ktile.load_safe(
+                K + K_load_off,
+                int(params->K_strides[2]),
+                short2(BD, lim_rows_k - ik * UKs));
+          } else {
+            Ktile.load(K + K_load_off, int(params->K_strides[2]));
+          }
+
+          subtile_matmad_nax(
+              Stile.subtile_at(iq, ik),
+              Qtile.subtile_at(0, 0),
+              metal::false_type{},
+              Ktile.subtile_at(0, 0),
+              metal::true_type{});
+        }
+      }
+    }
+
+    // Scale S
+    STEEL_PRAGMA_UNROLL
+    for (short ii = 0; ii < decltype(Stile)::kElemsPerTile; ii++) {
+      Stile.elems()[ii] *= float(scale2);
+    }
+
+    // Scale and Retile S
+    constexpr short UK = 16;
+    constexpr short TK = BK / UK;
+    using PSubTile = NAXSubTile<AccumType, UQ, UK>;
+
+    NAXTile<AccumType, TQ, TK, PSubTile> Ptile;
+
+    STEEL_PRAGMA_UNROLL
+    for (short ii = 0; ii < decltype(Stile)::kElemsPerTile; ii++) {
+      Ptile.elems()[ii] = Stile.elems()[ii];
+    }
+
+    // Mask out length sequence
+    if (!align_K && is_last_k) {
+      constexpr auto neg_inf = Limits<AccumType>::finite_min;
+
+      STEEL_PRAGMA_UNROLL
+      for (short iq = 0; iq < TQ; iq++) {
+        STEEL_PRAGMA_UNROLL
+        for (short ik = 0; ik < TK; ik++) {
+          const short col_pos = sn + ik * UK;
+
+          thread auto& fg = Ptile.subtile_at(iq, ik).frag_at(0, 0);
+
+          STEEL_PRAGMA_UNROLL
+          for (short ii = 0; ii < PSubTile::kFragThrRows; ii++) {
+            STEEL_PRAGMA_UNROLL
+            for (short jj = 0; jj < PSubTile::kFragThrCols; jj++) {
+              const auto loc = ii * PSubTile::kFragThrCols + jj;
+              fg[loc] = ((col_pos + jj) >= params->kL_rem) ? neg_inf : fg[loc];
+            }
+          }
+        }
+      }
+    }
+
+    // Mask out if causal
+    if (do_causal && kb >= (kb_lim - ((BQ + BK - 1) / BK) - int(!align_K))) {
+      constexpr auto neg_inf = Limits<AccumType>::finite_min;
+
+      const int base_row = tid.x * BQ + params->qL_off + tm;
+      const int base_col = kb * BK;
+
+      STEEL_PRAGMA_UNROLL
+      for (short iq = 0; iq < TQ; iq++) {
+        STEEL_PRAGMA_UNROLL
+        for (short ik = 0; ik < TK; ik++) {
+          const short row_pos = base_row + iq * UQ;
+          const short col_pos = base_col + ik * UK;
+
+          thread auto& fg = Ptile.subtile_at(iq, ik).frag_at(0, 0);
+
+          STEEL_PRAGMA_UNROLL
+          for (short ii = 0; ii < PSubTile::kFragThrRows; ii++) {
+            STEEL_PRAGMA_UNROLL
+            for (short jj = 0; jj < PSubTile::kFragThrCols; jj++) {
+              const auto r = row_pos + ii * PSubTile::kFragRowsJump + sm;
+              const auto c = col_pos + jj + sn;
+              const auto loc = ii * PSubTile::kFragThrCols + jj;
+              fg[loc] = (r < c) ? neg_inf : fg[loc];
+            }
+          }
+        }
+      }
+    }
+
+    // Other masking as needed
+    if (has_mask) {
+      constexpr auto neg_inf = Limits<AccumType>::finite_min;
+
+      const int base_row = tid.x * BQ + tm;
+      const int base_col = kb * BK;
+
+      constexpr bool is_bool = is_same_v<MaskType, bool>;
+      using melem_t = typename metal::conditional_t<is_bool, bool, AccumType>;
+      using MSubTile = NAXSubTile<melem_t, UQ, UK>;
+
+      STEEL_PRAGMA_UNROLL
+      for (short iq = 0; iq < TQ; iq++) {
+        STEEL_PRAGMA_UNROLL
+        for (short ik = 0; ik < TK; ik++) {
+          const short row_pos = base_row + iq * UQ + sm;
+          const short col_pos = base_col + ik * UK + sn;
+
+          MSubTile mfrag;
+          mfrag.load_safe(
+              mask,
+              int(mask_params->M_strides[2]),
+              Int<1>{},
+              params->qL,
+              params->kL,
+              row_pos,
+              col_pos);
+
+          thread auto& fg = Ptile.subtile_at(iq, ik).frag_at(0, 0);
+
+          STEEL_PRAGMA_UNROLL
+          for (short jj = 0; jj < MSubTile::kElemsPerFrag; jj++) {
+            if constexpr (is_bool) {
+              fg[jj] = mfrag.elems()[jj] ? fg[jj] : neg_inf;
+            } else {
+              fg[jj] += M_LOG2E_F * AccumType(mfrag.elems()[jj]);
+            }
+          }
+        }
+      }
+    }
+
+    // Do softmax
+
+    // Temp variables
+    metal::vec<AccumType, kRowsPT> new_max;
+    metal::vec<AccumType, kRowsPT> factor;
+    STEEL_PRAGMA_UNROLL
+    for (short i = 0; i < kRowsPT; ++i) {
+      new_max[i] = max_score[i];
+    }
+
+    // Row max
+    Ptile.template row_reduce<MaxOp>(new_max);
+
+    // exp(Si - rowmax(Si))
+    Ptile.template row_bin_op<ExpSubOp>(new_max);
+
+    // Factor exp(rowmax(Si) - rowmax(Si-1))
+    STEEL_PRAGMA_UNROLL
+    for (short i = 0; i < kRowsPT; ++i) {
+      factor[i] = fast::exp2(max_score[i] - new_max[i]);
+      max_score[i] = new_max[i];
+    }
+
+    // Row Sum
+    STEEL_PRAGMA_UNROLL
+    for (short i = 0; i < kRowsPT; ++i) {
+      sum_score[i] = sum_score[i] * factor[i];
+    }
+
+    Ptile.template row_reduce<SumOp>(sum_score);
+
+    // Update O
+    Otile.template row_bin_op<MulOp>(factor);
+
+    simdgroup_barrier(mem_flags::mem_none);
+
+    // Do O = P @ V
+    STEEL_PRAGMA_UNROLL
+    for (short iq = 0; iq < TQ; iq++) {
+      STEEL_PRAGMA_UNROLL
+      for (short id = 0; id < TD; id++) {
+        if constexpr (BD == 128) {
+          if (id == 2) {
+            threadgroup_barrier(mem_flags::mem_none);
+          }
+        }
+
+        STEEL_PRAGMA_UNROLL
+        for (short ik = 0; ik < TK; ik++) {
+          using VSubTile = NAXSubTile<T, UK, UD>;
+          NAXTile<T, 1, 1, VSubTile> Vtile;
+
+          const int V_load_off = ik * UK * int(params->V_strides[2]) + id * UD;
+
+          if (!align_K && is_last_k) {
+            // Vtile.load_rows(
+            //     V + V_load_off,
+            //     int(params->V_strides[2]),
+            //     lim_rows_k - ik * UK);
+            Vtile.load_safe(
+                V + V_load_off,
+                int(params->V_strides[2]),
+                short2(BD, lim_rows_k - ik * UK));
+          } else {
+            Vtile.load(V + V_load_off, int(params->V_strides[2]));
+          }
+
+          subtile_matmad_nax(
+              Otile.subtile_at(iq, id),
+              Ptile.subtile_at(iq, ik),
+              metal::bool_constant<false>{},
+              Vtile.subtile_at(0, 0),
+              metal::bool_constant<false>{});
+        }
+      }
+    }
+
+    // Prepare for next iteration
+    K += BK * int(params->K_strides[2]);
+    V += BK * int(params->V_strides[2]);
+  }
+
+  // Normalize output
+
+  threadgroup_barrier(mem_flags::mem_none);
+
+  metal::vec<AccumType, kRowsPT> rcp;
+  STEEL_PRAGMA_UNROLL
+  for (short i = 0; i < kRowsPT; ++i) {
+    rcp[i] = (1.f / sum_score[i]);
+  }
+
+  Otile.template row_bin_op<MulOp>(rcp);
+
+  // Store results
+  O += (tm + sm) * int(params->O_strides[2]) + sn;
+
+  if (!align_Q && is_last_q) {
+    if (lim_rows_q <= 0)
+      return;
+
+    // Otile.store_rows(O, params->O_strides[2], lim_rows_q);
+    Otile.store_safe(O, params->O_strides[2], short2(BD, lim_rows_q));
+  } else {
+    Otile.store(O, int(params->O_strides[2]));
+  }
+}

mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_nax.metal

@@ -0,0 +1,33 @@
+// Copyright © 2024-25 Apple Inc.
+
+// clang-format off
+#include "mlx/backend/metal/kernels/utils.h"
+
+#include "mlx/backend/metal/kernels/steel/attn/nax.h"
+#include "mlx/backend/metal/kernels/steel/attn/params.h"
+#include "mlx/backend/metal/kernels/steel/attn/transforms.h"
+#include "mlx/backend/metal/kernels/steel/utils.h"
+
+#include "mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_nax.h"
+
+#define instantiate_attn(tname, dtype, bq, bk, bd, wm, wn, mname, mtype) \
+  instantiate_kernel(                                                    \
+      "steel_attention_" #tname "_bq" #bq "_bk" #bk "_bd" #bd            \
+      "_wm" #wm "_wn" #wn "_mask" #mname,                                \
+  attention_nax, dtype, bq, bk, bd, wm, wn, mtype, float)
+
+#define instantiate_attn_shapes_helper(iname, itype, mname, mtype)  \
+    instantiate_attn(iname, itype, 64, 32, 128, 4, 1, mname, mtype) \
+    instantiate_attn(iname, itype, 64, 32,  64, 4, 1, mname, mtype) \
+    instantiate_attn(iname, itype, 64, 64, 128, 4, 1, mname, mtype) \
+    instantiate_attn(iname, itype, 64, 64,  64, 4, 1, mname, mtype)
+
+#define instantiate_attn_mask_helper(iname, itype) \
+    instantiate_attn_shapes_helper(iname, itype, iname, itype) \
+    instantiate_attn_shapes_helper(iname, itype, bool_, bool)
+
+instantiate_attn_mask_helper(float16, half);
+instantiate_attn_mask_helper(bfloat16, bfloat);
+
+instantiate_attn_mask_helper(float32, float);
+// clang-format on

mlx/backend/metal/kernels/steel/defines.h

@@ -1,4 +1,7 @@
 // Copyright © 2024 Apple Inc.
 
+#pragma once
+
 #define STEEL_CONST static constant constexpr const
 #define STEEL_PRAGMA_UNROLL _Pragma("clang loop unroll(full)")
+#define STEEL_PRAGMA_NO_UNROLL _Pragma("clang loop unroll(disable)")

mlx/backend/metal/kernels/steel/gemm/gemm_nax.h

@@ -0,0 +1,154 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/metal/kernels/steel/gemm/nax.h"
+#include "mlx/backend/metal/kernels/steel/gemm/params.h"
+
+using namespace metal;
+
+namespace mlx::steel {
+
+template <
+    typename T,
+    short SM,
+    short SN,
+    short SK,
+    short BK,
+    bool transpose_a,
+    bool transpose_b,
+    bool kAlignedM,
+    bool kAlignedN,
+    bool kAlignedK,
+    short UM,
+    short UN,
+    short UK,
+    typename AccumType = float>
+auto gemm_loop(
+    const device T* A,
+    const device T* B,
+    const constant GEMMParams* params [[buffer(4)]],
+    const short sgp_sm,
+    const short sgp_sn) {
+  constexpr short TM = SM / UM;
+  constexpr short TN = SN / UN;
+  constexpr short TK = SK / UK;
+
+  constexpr int RA = transpose_a ? TK : TM;
+  constexpr int CA = transpose_a ? TM : TK;
+
+  constexpr int RB = transpose_b ? TN : TK;
+  constexpr int CB = transpose_b ? TK : TN;
+
+  using DSubTile = NAXSubTile<AccumType, UM, UN>;
+  using ASubTile =
+      NAXSubTile<T, (transpose_a ? UK : UM), (transpose_a ? UM : UK)>;
+  using BSubTile =
+      NAXSubTile<T, (transpose_b ? UN : UK), (transpose_b ? UK : UN)>;
+
+  NAXTile<AccumType, TM, TN, DSubTile> Dtile;
+  Dtile.clear();
+
+  int gemm_k_iterations_ = params->gemm_k_iterations_aligned;
+
+  STEEL_PRAGMA_NO_UNROLL
+  for (int kk0 = 0; kk0 < gemm_k_iterations_; kk0++) {
+    threadgroup_barrier(mem_flags::mem_none);
+
+    STEEL_PRAGMA_NO_UNROLL
+    for (int kk1 = 0; kk1 < BK; kk1 += SK) {
+      NAXTile<T, RA, CA, ASubTile> Atile;
+      NAXTile<T, RB, CB, BSubTile> Btile;
+      const int k = kk1;
+
+      volatile int compiler_barrier;
+
+      const int A_offset = transpose_a ? k * params->lda : k;
+      const int B_offset = transpose_b ? k : k * params->ldb;
+
+      if constexpr (kAlignedM) {
+        Atile.load(A + A_offset, params->lda);
+      } else {
+        const short rmax = transpose_a ? SK : sgp_sm;
+        const short cmax = transpose_a ? sgp_sm : SK;
+        Atile.load_safe(A + A_offset, params->lda, short2(cmax, rmax));
+      }
+
+      if constexpr (kAlignedN) {
+        Btile.load(B + B_offset, params->ldb);
+      } else {
+        const short rmax = transpose_b ? sgp_sn : SK;
+        const short cmax = transpose_b ? SK : sgp_sn;
+        Btile.load_safe(B + B_offset, params->ldb, short2(cmax, rmax));
+      }
+
+      tile_matmad_nax(
+          Dtile,
+          Atile,
+          metal::bool_constant<transpose_a>{},
+          Btile,
+          metal::bool_constant<transpose_b>{});
+
+      (void)compiler_barrier;
+    }
+
+    A += transpose_a ? (BK * params->lda) : BK;
+    B += transpose_b ? BK : (BK * params->ldb);
+  }
+
+  if constexpr (!kAlignedK) {
+    simdgroup_barrier(mem_flags::mem_none);
+
+    const short rem_bk = params->K - gemm_k_iterations_ * BK;
+
+    STEEL_PRAGMA_NO_UNROLL
+    for (int kk1 = 0; kk1 < rem_bk; kk1 += SK) {
+      NAXTile<T, 1, 1, ASubTile> Atile;
+      NAXTile<T, 1, 1, BSubTile> Btile;
+
+      STEEL_PRAGMA_UNROLL
+      for (int mm = 0; mm < TM; mm++) {
+        STEEL_PRAGMA_UNROLL
+        for (int nn = 0; nn < TN; nn++) {
+          STEEL_PRAGMA_UNROLL
+          for (int kk = 0; kk < TK; kk++) {
+            const int m = mm * UM;
+            const int n = nn * UN;
+            const int k = kk1 + kk * UK;
+            const short psk = max(0, rem_bk - k);
+
+            const int A_offset =
+                transpose_a ? (m + k * params->lda) : (m * params->lda + k);
+            const int B_offset =
+                transpose_b ? (k + n * params->ldb) : (k * params->ldb + n);
+
+            {
+              const short psm = kAlignedM ? SM : max(0, sgp_sm - m);
+              const short rmax = transpose_a ? psk : psm;
+              const short cmax = transpose_a ? psm : psk;
+              Atile.load_safe(A + A_offset, params->lda, short2(cmax, rmax));
+            }
+
+            {
+              const short psn = kAlignedN ? SN : max(0, sgp_sn - n);
+              const short rmax = transpose_b ? psn : psk;
+              const short cmax = transpose_b ? psk : psn;
+              Btile.load_safe(B + B_offset, params->ldb, short2(cmax, rmax));
+            }
+
+            subtile_matmad_nax(
+                Dtile.subtile_at(mm, nn),
+                Atile.subtile_at(0, 0),
+                metal::bool_constant<transpose_a>{},
+                Btile.subtile_at(0, 0),
+                metal::bool_constant<transpose_b>{});
+          }
+        }
+      }
+    }
+  }
+
+  return Dtile;
+}
+
+} // namespace mlx::steel

mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused_nax.h

@@ -0,0 +1,207 @@
+// Copyright © 2025 Apple Inc.
+
+using namespace mlx::steel;
+
+constant bool has_batch [[function_constant(10)]];
+
+constant bool use_out_source [[function_constant(100)]];
+constant bool do_axpby [[function_constant(110)]];
+
+constant bool align_M [[function_constant(200)]];
+constant bool align_N [[function_constant(201)]];
+constant bool align_K [[function_constant(202)]];
+
+// clang-format off
+template <
+    bool kAlignedM,
+    bool kAlignedN,
+    typename NAXTile_t,
+    typename T>
+void gemm_epilogue(
+    thread NAXTile_t& Dtile,
+    const device T* C,
+    const constant GEMMParams* params,
+    const constant GEMMAddMMParams* addmm_params,
+    const short sgp_sm, 
+    const short sgp_sn) { // clang-format on
+
+  (void)params;
+
+  constexpr short UM = NAXTile_t::kSubTileRows;
+  constexpr short UN = NAXTile_t::kSubTileCols;
+  using CSubTile = NAXSubTile<T, UM, UN>;
+
+  using V = typename NAXTile_t::elem_type;
+
+  constexpr short TM = NAXTile_t::kTileRows;
+  constexpr short TN = NAXTile_t::kTileCols;
+  constexpr short kElemsPerSubTile = NAXTile_t::kElemsPerSubTile;
+
+  STEEL_PRAGMA_UNROLL
+  for (short mm = 0; mm < TM; mm++) {
+    STEEL_PRAGMA_UNROLL
+    for (short nn = 0; nn < TN; nn++) {
+      const short m = mm * UM;
+      const short n = nn * UN;
+
+      CSubTile CTile;
+
+      if constexpr (kAlignedM && kAlignedN) {
+        CTile.load(C, addmm_params->ldc, addmm_params->fdc, m, n);
+      } else {
+        CTile.load_safe(
+            C, addmm_params->ldc, addmm_params->fdc, sgp_sm, sgp_sn, m, n);
+      }
+
+      auto delems = Dtile.subtile_at(mm, nn).elems();
+      auto celems = CTile.elems();
+
+      STEEL_PRAGMA_UNROLL
+      for (short i = 0; i < kElemsPerSubTile; i++) {
+        if (do_axpby) {
+          delems[i] = addmm_params->alpha * delems[i] +
+              addmm_params->beta * static_cast<V>(celems[i]);
+        } else {
+          delems[i] += static_cast<V>(celems[i]);
+        }
+      }
+    }
+  }
+}
+
+// clang-format off
+template <
+    typename T,
+    int BM,
+    int BN,
+    int BK,
+    int WM,
+    int WN,
+    bool transpose_a,
+    bool transpose_b,
+    typename AccumType = float>
+[[kernel, max_total_threads_per_threadgroup(WM* WN * 32)]] void gemm(
+    const device T* A [[buffer(0)]],
+    const device T* B [[buffer(1)]],
+    const device T* C [[buffer(2), function_constant(use_out_source)]],
+    device T* D [[buffer(3)]],
+    const constant GEMMParams* params [[buffer(4)]],
+    const constant GEMMAddMMParams* addmm_params [[buffer(5), function_constant(use_out_source)]],
+    const constant int* batch_shape [[buffer(6), function_constant(has_batch)]],
+    const constant int64_t* batch_strides [[buffer(7), function_constant(has_batch)]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]]) { // clang-format on
+  // Find block
+  const int tid_y = ((tid.y) << params->swizzle_log) +
+      ((tid.x) & ((1 << params->swizzle_log) - 1));
+  const int tid_x = (tid.x) >> params->swizzle_log;
+
+  // Exit early if out of bounds
+  if (params->tiles_n <= tid_x || params->tiles_m <= tid_y) {
+    return;
+  }
+
+  // Adjust for batch
+  if (has_batch) {
+    const constant auto* A_bstrides = batch_strides;
+    const constant auto* B_bstrides = batch_strides + params->batch_ndim;
+
+    ulong2 batch_offsets = elem_to_loc_broadcast(
+        tid.z, batch_shape, A_bstrides, B_bstrides, params->batch_ndim);
+
+    A += batch_offsets.x;
+    B += batch_offsets.y;
+
+    if (use_out_source) {
+      const constant auto* C_bstrides = B_bstrides + params->batch_ndim;
+      C += elem_to_loc(tid.z, batch_shape, C_bstrides, params->batch_ndim);
+    }
+  } else {
+    A += params->batch_stride_a * tid.z;
+    B += params->batch_stride_b * tid.z;
+
+    if (use_out_source) {
+      C += addmm_params->batch_stride_c * tid.z;
+    }
+  }
+
+  D += params->batch_stride_d * tid.z;
+
+  // Prepare threadgroup memory
+  threadgroup_barrier(mem_flags::mem_none);
+
+  // Find block in A, B, C
+  const int c_row = tid_y * BM;
+  const int c_col = tid_x * BN;
+  const size_t c_row_long = size_t(c_row);
+  const size_t c_col_long = size_t(c_col);
+
+  A += transpose_a ? c_row_long : c_row_long * params->lda;
+  B += transpose_b ? c_col_long * params->ldb : c_col_long;
+  D += c_row_long * params->ldd + c_col_long;
+
+  if (use_out_source) {
+    C += c_row_long * addmm_params->ldc + c_col_long * addmm_params->fdc;
+  }
+
+  constexpr short UM = 16;
+  constexpr short UN = 32;
+  constexpr short UK = 16;
+  constexpr short SM = BM / WM;
+  constexpr short SN = BN / WN;
+  constexpr short SK = 32;
+
+  constexpr short TM = SM / UM;
+  constexpr short TN = SN / UN;
+
+  const short tm = SM * (simd_group_id / WN);
+  const short tn = SN * (simd_group_id % WN);
+
+  const short sgp_sm = align_M ? SM : min(SM, short(params->M - (c_row + tm)));
+  const bool is_unaligned_sm = align_M ? false : (sgp_sm != SM);
+
+  const short sgp_sn = align_N ? SN : min(SN, short(params->N - (c_col + tn)));
+  const bool is_unaligned_sn = align_N ? false : (sgp_sn != SN);
+
+  A += transpose_a ? tm : (tm * params->lda);
+  B += transpose_b ? (tn * params->ldb) : tn;
+  D += tm * params->ldd + tn;
+
+  if (use_out_source) {
+    C += tm * addmm_params->ldc + tn * addmm_params->fdc;
+  }
+
+  using DSubTile = NAXSubTile<AccumType, UM, UN>;
+  NAXTile<AccumType, TM, TN, DSubTile> Dtile;
+
+  dispatch_bool(align_K, [&](auto kAlignedK) {
+    dispatch_bool(align_M || !is_unaligned_sm, [&](auto kAlignedM) {
+      dispatch_bool(align_N || !is_unaligned_sn, [&](auto kAlignedN) {
+        Dtile = gemm_loop<
+            T,
+            SM,
+            SN,
+            SK,
+            BK,
+            transpose_a,
+            transpose_b,
+            kAlignedM.value,
+            kAlignedN.value,
+            kAlignedK.value,
+            UM,
+            UN,
+            UK,
+            AccumType>(A, B, params, sgp_sm, sgp_sn);
+        if (use_out_source) {
+          gemm_epilogue<kAlignedM.value, kAlignedN.value>(
+              Dtile, C, params, addmm_params, sgp_sm, sgp_sn);
+        }
+        if constexpr (kAlignedM && kAlignedN) {
+          Dtile.store(D, int(params->ldd));
+        } else {
+          Dtile.store_safe(D, int(params->ldd), short2(sgp_sn, sgp_sm));
+        }
+      });
+    });
+  });
+}

mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused_nax.metal

@@ -0,0 +1,35 @@
+// Copyright © 2025 Apple Inc.
+
+#include <metal_stdlib>
+
+#include "mlx/backend/metal/kernels/utils.h"
+
+#include "mlx/backend/metal/kernels/steel/gemm/gemm_nax.h"
+#include "mlx/backend/metal/kernels/steel/gemm/nax.h"
+#include "mlx/backend/metal/kernels/steel/gemm/params.h"
+#include "mlx/backend/metal/kernels/steel/gemm/transforms.h"
+#include "mlx/backend/metal/kernels/steel/utils.h"
+
+#include "mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused_nax.h"
+
+// clang-format off
+#define instantiate_gemm(tname, trans_a, trans_b, iname, itype, oname, otype, bm, bn, bk, wm, wn) \
+  instantiate_kernel(                                                                             \
+      "steel_gemm_fused_nax_" #tname "_"  #iname "_" #oname                                       \
+      "_bm" #bm "_bn" #bn "_bk" #bk "_wm" #wm "_wn" #wn,                                          \
+  gemm, itype, bm, bn, bk, wm, wn, trans_a, trans_b, float)
+
+#define instantiate_gemm_transpose_helper(iname, itype, oname, otype, bm, bn, bk, wm, wn) \
+    instantiate_gemm(nn, false, false, iname, itype, oname, otype, bm, bn, bk, wm, wn) \
+    instantiate_gemm(nt, false, true , iname, itype, oname, otype, bm, bn, bk, wm, wn) \
+    instantiate_gemm(tn, true , false, iname, itype, oname, otype, bm, bn, bk, wm, wn) \
+    instantiate_gemm(tt, true , true , iname, itype, oname, otype, bm, bn, bk, wm, wn)
+
+#define instantiate_gemm_shapes_helper(iname, itype, oname, otype) \
+    instantiate_gemm_transpose_helper(iname, itype, oname, otype,  64,  64, 256, 2, 2) \
+    instantiate_gemm_transpose_helper(iname, itype, oname, otype, 128, 128, 512, 4, 4)
+
+instantiate_gemm_shapes_helper(float16, half, float16, half);
+instantiate_gemm_shapes_helper(bfloat16, bfloat, bfloat16, bfloat);
+instantiate_gemm_shapes_helper(float32, float, float32, float);
+// clang-format on

mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather_nax.h

@@ -0,0 +1,132 @@
+// Copyright © 2024 Apple Inc.
+
+using namespace mlx::steel;
+
+constant bool align_M [[function_constant(200)]];
+constant bool align_N [[function_constant(201)]];
+constant bool align_K [[function_constant(202)]];
+
+template <
+    typename T,
+    int BM,
+    int BN,
+    int BK,
+    int WM,
+    int WN,
+    bool transpose_a,
+    bool transpose_b,
+    typename AccumType = float>
+[[kernel, max_total_threads_per_threadgroup(WM* WN * 32)]] void
+gather_mm_rhs_nax(
+    const device T* A [[buffer(0)]],
+    const device T* B [[buffer(1)]],
+    const device uint32_t* rhs_indices [[buffer(2)]],
+    device T* C [[buffer(3)]],
+    const constant GEMMParams* params [[buffer(4)]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]]) {
+  constexpr short UM = 16;
+  constexpr short UN = 32;
+  constexpr short UK = 16;
+  constexpr short SM = BM / WM;
+  constexpr short SN = BN / WN;
+  constexpr short SK = 32;
+  constexpr short TM = SM / UM;
+  constexpr short TN = SN / UN;
+
+  if (params->tiles_n <= static_cast<int>(tid.x) ||
+      params->tiles_m <= static_cast<int>(tid.y)) {
+    return;
+  }
+
+  // Find the block in A, B, C
+  const int c_row = tid.y * BM;
+  const int c_col = tid.x * BN;
+  const size_t c_row_long = size_t(c_row);
+  const size_t c_col_long = size_t(c_col);
+
+  A += transpose_a ? c_row_long : c_row_long * params->lda;
+  B += transpose_b ? c_col_long * params->ldb : c_col_long;
+  C += c_row_long * params->ldd + c_col_long;
+  rhs_indices += c_row;
+
+  const short tm = SM * (simd_group_id / WN);
+  const short tn = SN * (simd_group_id % WN);
+
+  const short sgp_sm = align_M ? SM : min(SM, short(params->M - (c_row + tm)));
+  const bool is_unaligned_sm = align_M ? false : (sgp_sm != SM);
+
+  const short sgp_sn = align_N ? SN : min(SN, short(params->N - (c_col + tn)));
+  const bool is_unaligned_sn = align_N ? false : (sgp_sn != SN);
+
+  A += transpose_a ? tm : (tm * params->lda);
+  B += transpose_b ? (tn * params->ldb) : tn;
+  C += tm * params->ldd + tn;
+  rhs_indices += tm;
+
+  // Do as many matmuls as necessary
+  uint32_t index;
+  short offset;
+  uint32_t index_next = rhs_indices[0];
+  short offset_next = 0;
+  int n = 0;
+  while (n < sgp_sm) {
+    n++;
+    offset = offset_next;
+    index = index_next;
+    offset_next = sgp_sm;
+    for (; n < sgp_sm; n++) {
+      if (rhs_indices[n] != index) {
+        offset_next = n;
+        index_next = rhs_indices[n];
+        break;
+      }
+    }
+    threadgroup_barrier(mem_flags::mem_none);
+
+    using DSubTile = NAXSubTile<AccumType, UM, UN>;
+    NAXTile<AccumType, TM, TN, DSubTile> Ctile;
+
+    dispatch_bool(align_K, [&](auto kAlignedK) {
+      dispatch_bool(align_M || !is_unaligned_sm, [&](auto kAlignedM) {
+        dispatch_bool(align_N || !is_unaligned_sn, [&](auto kAlignedN) {
+          auto do_gemm = gemm_loop<
+              T,
+              SM,
+              SN,
+              SK,
+              BK,
+              transpose_a,
+              transpose_b,
+              kAlignedM.value,
+              kAlignedN.value,
+              kAlignedK.value,
+              UM,
+              UN,
+              UK,
+              AccumType>;
+          Ctile = do_gemm(
+              A, B + index * params->batch_stride_b, params, sgp_sm, sgp_sn);
+
+          if constexpr (kAlignedN.value) {
+            if (offset_next - offset == SM) {
+              Ctile.store(C, int(params->ldd));
+            } else {
+              Ctile.store_slice(
+                  C,
+                  int(params->ldd),
+                  short2(0, offset),
+                  short2(SN, offset_next));
+            }
+          } else {
+            Ctile.store_slice(
+                C,
+                int(params->ldd),
+                short2(0, offset),
+                short2(sgp_sn, offset_next));
+          }
+        });
+      });
+    });
+  }
+}

mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather_nax.metal

@@ -0,0 +1,39 @@
+// Copyright © 2024 Apple Inc.
+
+#include <metal_stdlib>
+
+#include "mlx/backend/metal/kernels/steel/gemm/gemm_nax.h"
+#include "mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather_nax.h"
+#include "mlx/backend/metal/kernels/steel/gemm/nax.h"
+#include "mlx/backend/metal/kernels/steel/gemm/params.h"
+#include "mlx/backend/metal/kernels/steel/utils.h"
+#include "mlx/backend/metal/kernels/utils.h"
+
+// clang-format off
+#define instantiate_gather_mm_rhs(tname, trans_a, trans_b, iname, itype, oname, otype, bm, bn, bk, wm, wn) \
+  instantiate_kernel(                                                             \
+      "steel_gather_mm_rhs_nax_" #tname "_" #iname "_" #oname "_bm" #bm "_bn" #bn \
+      "_bk" #bk "_wm" #wm "_wn" #wn,                                              \
+      gather_mm_rhs_nax,                                                          \
+      itype,                                                                      \
+      bm,                                                                         \
+      bn,                                                                         \
+      bk,                                                                         \
+      wm,                                                                         \
+      wn,                                                                         \
+      trans_a,                                                                    \
+      trans_b,                                                                    \
+      float)
+
+#define instantiate_gather_mm_rhs_transpose_helper(iname, itype, oname, otype, bm, bn, bk, wm, wn) \
+  instantiate_gather_mm_rhs(nn, false, false, iname, itype, oname, otype, bm, bn, bk, wm, wn)  \
+  instantiate_gather_mm_rhs(nt, false,  true, iname, itype, oname, otype, bm, bn, bk, wm, wn)
+
+#define instantiate_gather_mm_shapes_helper(iname, itype, oname, otype)                      \
+  instantiate_gather_mm_rhs_transpose_helper(iname, itype, oname, otype, 16, 128, 128, 1, 4) \
+  instantiate_gather_mm_rhs_transpose_helper(iname, itype, oname, otype, 32, 128, 128, 1, 4) \
+  instantiate_gather_mm_rhs_transpose_helper(iname, itype, oname, otype, 64, 128, 128, 2, 4)
+// clang-format on
+
+instantiate_gather_mm_shapes_helper(float16, half, float16, half);
+instantiate_gather_mm_shapes_helper(bfloat16, bfloat, bfloat16, bfloat);