remove unused unary file (#2672)

* Modified sort behavior when running CPU or Metal to match NumPy/JAX sorting behavior.

* Modified sort behavior when running CPU or Metal to match NumPy/JAX

* nits

---------

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

.circleci/config.yml

@@ -18,13 +18,14 @@ jobs:
         type: boolean
         default: false
     macos:
-      xcode: "16.2.0"
-    resource_class: m2pro.medium
+      xcode: "26.0.0"
+    resource_class: m4pro.medium
     steps:
       - checkout
       - run:
           name: Install
           command: |
+            xcodebuild -downloadComponent MetalToolchain
             brew install python@3.9
             brew install doxygen
             python3.9 -m venv env
@@ -89,7 +90,8 @@ jobs:
           command: |
             uv venv
             uv pip install cmake
-            uv pip install -e ".[dev]" -v
+            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+              uv pip install -e ".[dev]" -v
       - run:
           name: Generate package stubs
           command: |
@@ -118,7 +120,7 @@ jobs:
     parameters:
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "26.0.0"
       macosx_deployment_target:
         type: string
         default: ""
@@ -126,12 +128,13 @@ jobs:
       xcode: << parameters.xcode_version >>
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m2pro.medium
+    resource_class: m4pro.medium
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
+            xcodebuild -downloadComponent MetalToolchain
             HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
               brew install openmpi uv
       - run:
@@ -196,7 +199,7 @@ jobs:
           name: Run Python tests with JIT
           command: |
             CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e .
+              uv pip install -e . -v
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
               uv run --no-project python -m xmlrunner discover \
@@ -222,15 +225,20 @@ jobs:
             sudo apt-get update
             sudo apt-get install libcudnn9-dev-cuda-12
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install libnccl2 libnccl-dev
             curl -sL https://github.com/ccache/ccache/releases/download/v4.11.3/ccache-4.11.3-linux-x86_64.tar.xz | tar xJf -
             sudo mv ccache-4.11.3-linux-x86_64/ccache /usr/bin/ccache
             rm -rf ccache-4.11.3-linux-x86_64
             curl -LsSf https://astral.sh/uv/install.sh | sh
+      - run:
+          name: Set CCache size
+          command: ccache --max-size 1G
       - run:
           name: Install Python package
           command: |
             uv venv
-            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+            uv pip install cmake
+            DEBUG=1 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               uv pip install -e ".[dev]" -v
       - run:
           name: Run Python tests
@@ -238,12 +246,23 @@ jobs:
             source .venv/bin/activate
             LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
             LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
+      - run:
+          name: Build CPP only
+          command: |
+            source .venv/bin/activate
+            cmake . -B build \
+              -DMLX_BUILD_CUDA=ON \
+              -DCMAKE_CUDA_COMPILER=`which nvcc` \
+              -DCMAKE_BUILD_TYPE=DEBUG
+            cmake --build build -j `nproc`
+      - run:
+          name: Run CPP tests
+          command: ./build/tests/tests -sfe="*fft_tests.cpp,*linalg_tests.cpp"
       - run:
           name: CCache report
           command: |
             ccache --show-stats
             ccache --zero-stats
-            ccache --max-size 400MB
             ccache --cleanup
       - save_cache:
           key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
@@ -257,7 +276,7 @@ jobs:
         default: "3.9"
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "26.0.0"
       build_env:
         type: string
         default: ""
@@ -266,7 +285,7 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: m2pro.medium
+    resource_class: m4pro.medium
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
     steps:
@@ -274,11 +293,15 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            brew install python@<< parameters.python_version >>
-            brew install openmpi
-            python<< parameters.python_version >> -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
+            xcodebuild -downloadComponent MetalToolchain
+            mkdir -p ~/miniconda3
+            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
+            bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+            rm ~/miniconda3/miniconda.sh
+            source ~/miniconda3/bin/activate
+            conda init --all
+            conda create -n env python=<< parameters.python_version >> -y
+            conda activate env
             pip install --upgrade cmake
             pip install nanobind==2.4.0
             pip install --upgrade setuptools
@@ -288,19 +311,19 @@ jobs:
       - run:
           name: Install Python package
           command: |
-            source env/bin/activate
+            conda activate env
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
-            source env/bin/activate
+            conda activate env
             pip install typing_extensions
             python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
-            source env/bin/activate
+            conda activate env
             python setup.py clean --all
             << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
       - when:
@@ -310,7 +333,7 @@ jobs:
             - run:
                 name: Build common package
                 command: |
-                  source env/bin/activate
+                  conda activate env
                   python setup.py clean --all
                   << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
@@ -319,7 +342,7 @@ jobs:
             - run:
                 name: Upload package
                 command: |
-                  source env/bin/activate
+                  conda activate env
                   twine upload dist/*
       - store_artifacts:
           path: dist/
@@ -392,7 +415,7 @@ jobs:
         default: ""
     machine:
       image: ubuntu-2204:current
-      resource_class: large
+      resource_class: xlarge
     steps:
       - checkout
       - run:
@@ -439,7 +462,7 @@ workflows:
       - mac_build_and_test:
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              macosx_deployment_target: ["13.5", "15.0"]
       - linux_build_and_test
       - cuda_build_and_test:
           matrix:
@@ -464,68 +487,7 @@ workflows:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
+              xcode_version: ["26.0.0"]
       - build_documentation:
           filters:
             tags:
@@ -567,7 +529,7 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              macosx_deployment_target: ["13.5", "15.0"]
       - linux_build_and_test:
           requires: [ hold ]
       - cuda_build_and_test:
@@ -586,53 +548,7 @@ workflows:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
+              xcode_version: ["26.0.0"]
       - build_linux_release:
           matrix:
             parameters:
@@ -651,68 +567,7 @@ workflows:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
+              xcode_version: ["26.0.0"]
       - build_linux_release:
           matrix:
             parameters:

ACKNOWLEDGMENTS.md

@@ -19,12 +19,17 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
-- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer, and the `ReLU²` activation function.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>
 
+# Organizations
+
+MLX has received contributions from the following companies:
+- NVIDIA Corporation & Affiliates
+
 # Third-Party Software
 
 MLX leverages several third-party software, listed here together with

CMakeLists.txt

@@ -26,6 +26,7 @@ set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_INSTALL_MESSAGE NEVER)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 # ----------------------------- Configuration -----------------------------
 option(MLX_BUILD_TESTS "Build tests for mlx" ON)
@@ -87,22 +88,21 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
-if(MLX_BUILD_METAL)
-  set(METAL_LIB "-framework Metal")
-  set(FOUNDATION_LIB "-framework Foundation")
-  set(QUARTZ_LIB "-framework QuartzCore")
-endif()
-
 if(MLX_BUILD_CUDA)
   enable_language(CUDA)
 endif()
 
-if(MLX_BUILD_METAL AND NOT METAL_LIB)
-  message(STATUS "Metal not found. Unable to build GPU")
-  set(MLX_BUILD_METAL OFF)
-  set(MLX_METAL_DEBUG OFF)
-elseif(MLX_BUILD_METAL)
-  message(STATUS "Building METAL sources")
+if(MLX_BUILD_METAL)
+  find_library(METAL_LIB Metal)
+  find_library(FOUNDATION_LIB Foundation)
+  find_library(QUARTZ_LIB QuartzCore)
+  if(METAL_LIB)
+    message(STATUS "Metal found ${METAL_LIB}")
+  else()
+    message(
+      FATAL_ERROR
+        "Metal not found. Set MLX_BUILD_METAL=OFF to build without GPU")
+  endif()
 
   if(MLX_METAL_DEBUG)
     add_compile_definitions(MLX_METAL_DEBUG)
@@ -111,7 +111,8 @@ elseif(MLX_BUILD_METAL)
   # Throw an error if xcrun not found
   execute_process(
     COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
-    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)
+    OUTPUT_VARIABLE MACOS_SDK_VERSION
+    OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
 
   if(${MACOS_SDK_VERSION} LESS 14.0)
     message(
@@ -140,6 +141,12 @@ elseif(MLX_BUILD_METAL)
   target_link_libraries(mlx PUBLIC ${METAL_LIB} ${FOUNDATION_LIB} ${QUARTZ_LIB})
 endif()
 
+if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+  # With newer clang/gcc versions following libs are implicitly linked, but when
+  # building on old distributions they need to be explicitly listed.
+  target_link_libraries(mlx PRIVATE dl pthread)
+endif()
+
 if(WIN32)
   if(MSVC)
     # GGUF does not build with MSVC.
@@ -167,7 +174,7 @@ if(MLX_BUILD_CPU)
     message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
     set(MLX_BUILD_ACCELERATE ON)
   else()
-    message(STATUS "Accelerate or arm neon not found, using default backend.")
+    message(STATUS "Accelerate not found, using default backend.")
     set(MLX_BUILD_ACCELERATE OFF)
   endif()
 

README.md

@@ -2,7 +2,7 @@
 
 [**Quickstart**](#quickstart) | [**Installation**](#installation) |
 [**Documentation**](https://ml-explore.github.io/mlx/build/html/index.html) |
-[**Examples**](#examples) 
+[**Examples**](#examples)
 
 [![CircleCI](https://circleci.com/gh/ml-explore/mlx.svg?style=svg)](https://circleci.com/gh/ml-explore/mlx)
 
@@ -11,37 +11,37 @@ brought to you by Apple machine learning research.
 
 Some key features of MLX include:
 
- - **Familiar APIs**: MLX has a Python API that closely follows NumPy. MLX
+- **Familiar APIs**: MLX has a Python API that closely follows NumPy. MLX
    also has fully featured C++, [C](https://github.com/ml-explore/mlx-c), and
    [Swift](https://github.com/ml-explore/mlx-swift/) APIs, which closely mirror
    the Python API. MLX has higher-level packages like `mlx.nn` and
    `mlx.optimizers` with APIs that closely follow PyTorch to simplify building
    more complex models.
 
- - **Composable function transformations**: MLX supports composable function
-   transformations for automatic differentiation, automatic vectorization,
-   and computation graph optimization.
+- **Composable function transformations**: MLX supports composable function
+  transformations for automatic differentiation, automatic vectorization,
+  and computation graph optimization.
 
- - **Lazy computation**: Computations in MLX are lazy. Arrays are only
-   materialized when needed.
+- **Lazy computation**: Computations in MLX are lazy. Arrays are only
+  materialized when needed.
 
- - **Dynamic graph construction**: Computation graphs in MLX are constructed
-   dynamically. Changing the shapes of function arguments does not trigger
-   slow compilations, and debugging is simple and intuitive.
+- **Dynamic graph construction**: Computation graphs in MLX are constructed
+  dynamically. Changing the shapes of function arguments does not trigger
+  slow compilations, and debugging is simple and intuitive.
 
- - **Multi-device**: Operations can run on any of the supported devices
-   (currently the CPU and the GPU).
+- **Multi-device**: Operations can run on any of the supported devices
+  (currently the CPU and the GPU).
 
- - **Unified memory**: A notable difference from MLX and other frameworks
-   is the *unified memory model*. Arrays in MLX live in shared memory.
-   Operations on MLX arrays can be performed on any of the supported
-   device types without transferring data.
+- **Unified memory**: A notable difference from MLX and other frameworks
+  is the *unified memory model*. Arrays in MLX live in shared memory.
+  Operations on MLX arrays can be performed on any of the supported
+  device types without transferring data.
 
 MLX is designed by machine learning researchers for machine learning
 researchers. The framework is intended to be user-friendly, but still efficient
 to train and deploy models. The design of the framework itself is also
 conceptually simple. We intend to make it easy for researchers to extend and
-improve MLX with the goal of quickly exploring new ideas. 
+improve MLX with the goal of quickly exploring new ideas.
 
 The design of MLX is inspired by frameworks like
 [NumPy](https://numpy.org/doc/stable/index.html),
@@ -91,7 +91,7 @@ Checkout the
 [documentation](https://ml-explore.github.io/mlx/build/html/install.html#)
 for more information on building the C++ and Python APIs from source.
 
-## Contributing 
+## Contributing
 
 Check out the [contribution guidelines](https://github.com/ml-explore/mlx/tree/main/CONTRIBUTING.md) for more information
 on contributing to MLX. See the
@@ -110,7 +110,7 @@ Hannun, Jagrit Digani, Angelos Katharopoulos, and Ronan Collobert. If you find
 MLX useful in your research and wish to cite it, please use the following
 BibTex entry:
 
-```
+```text
 @software{mlx2023,
   author = {Awni Hannun and Jagrit Digani and Angelos Katharopoulos and Ronan Collobert},
   title = {{MLX}: Efficient and flexible machine learning on Apple silicon},

benchmarks/python/blas/bench_gemm.py

@@ -142,9 +142,7 @@ def bench_shape(B, M, N, K, np_dtype, transpose="nn"):
     t_b = (0, 1, 2) if transpose[1] == "n" else (0, 2, 1)
 
     c_mlx = a_mx.transpose(t_a) @ b_mx.transpose(t_b)
-    c_npy = a_np.transpose(t_a).astype(np.float32) @ b_np.transpose(t_b).astype(
-        np.float32
-    )
+    c_npy = a_np.transpose(t_a).astype(np_dtype) @ b_np.transpose(t_b).astype(np_dtype)
 
     atol = 1e-5 if np_dtype == np.float32 else 1e-4
 
@@ -163,7 +161,7 @@ def get_gflop_count(B, M, N, K):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Run gemm benchmarks")
 
-    dtypes = ("float32", "float16")
+    dtypes = ("float32", "float16", "complex64")
     transposes = ("nn", "nt", "tn")
     shapes = (
         (16, 234, 768, 3072),
@@ -187,7 +185,7 @@ if __name__ == "__main__":
                 diff = gflops_mx / gflops_pt - 1.0
 
                 print(
-                    f"{B:3d}, {M:4d}, {N:4d}, {K:4d}, {dtype}, {transpose}, {gflops_pt:05.3f}, {gflops_mx:05.3f}, {100. * diff:+5.2f}%"
+                    f"{B:3d}, {M:4d}, {N:4d}, {K:4d}, {dtype}, {transpose}, {gflops_pt:05.3f}, {gflops_mx:05.3f}, {100.0 * diff:+5.2f}%"
                 )
                 if gflops_pt >= 2.0 * gflops_mx:
                     print("ATTENTION ^^^^^^^")

benchmarks/python/blas/bench_gemv.py

@@ -196,7 +196,7 @@ def bench_with_out_len(ax, out_vec_len, in_vector_lens, dtype, transpose):
 
 
 for transpose in (False, True):
-    for dtype in ("float32", "float16"):
+    for dtype in ("float32", "float16", "complex64"):
         fig, axs = plt.subplots(
             len(in_vec_sizes), 2, figsize=(8.5, 11), layout="constrained"
         )
@@ -215,7 +215,7 @@ for transpose in (False, True):
         fig.suptitle(f"{device_name}: {dtype} {op_name}")
         fig.savefig(
             os.path.join(
-                results_dir, f'{device_name.replace(" ", "_")}_{dtype}_{op_name}.pdf'
+                results_dir, f"{device_name.replace(' ', '_')}_{dtype}_{op_name}.pdf"
             )
         )
         plt.close(fig)

cmake/FindNCCL.cmake

@@ -0,0 +1,54 @@
+# FindNCCL.cmake This module finds the NVIDIA NCCL library and its include
+# directories.
+
+set(NCCL_ROOT_DIR
+    $ENV{NCCL_ROOT_DIR}
+    CACHE PATH "Folder contains NVIDIA NCCL")
+
+find_path(
+  NCCL_INCLUDE_DIRS
+  NAMES nccl.h
+  HINTS ${NCCL_INCLUDE_DIR} ${NCCL_ROOT_DIR} ${NCCL_ROOT_DIR}/include
+        ${CUDA_TOOLKIT_ROOT_DIR}/include)
+
+if($ENV{USE_STATIC_NCCL})
+  message(
+    STATUS "USE_STATIC_NCCL detected. Linking against static NCCL library")
+  set(NCCL_LIBNAME "libnccl_static.a")
+else()
+  set(NCCL_LIBNAME "nccl")
+endif()
+
+find_library(
+  NCCL_LIBRARIES
+  NAMES ${NCCL_LIBNAME}
+  HINTS ${NCCL_LIB_DIR}
+        ${NCCL_ROOT_DIR}
+        ${NCCL_ROOT_DIR}/lib
+        ${NCCL_ROOT_DIR}/lib/x86_64-linux-gnu
+        ${NCCL_ROOT_DIR}/lib64
+        ${CUDA_TOOLKIT_ROOT_DIR}/lib
+        ${CUDA_TOOLKIT_ROOT_DIR}/lib64)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIRS
+                                  NCCL_LIBRARIES)
+
+if(NCCL_FOUND)
+  set(NCCL_HEADER_FILE "${NCCL_INCLUDE_DIRS}/nccl.h")
+  message(
+    STATUS "Determining NCCL version from the header file: ${NCCL_HEADER_FILE}")
+  file(
+    STRINGS ${NCCL_HEADER_FILE} NCCL_MAJOR_VERSION_DEFINED
+    REGEX "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+[0-9]+.*$"
+    LIMIT_COUNT 1)
+  if(NCCL_MAJOR_VERSION_DEFINED)
+    string(REGEX REPLACE "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+" ""
+                         NCCL_MAJOR_VERSION ${NCCL_MAJOR_VERSION_DEFINED})
+    message(STATUS "NCCL_MAJOR_VERSION: ${NCCL_MAJOR_VERSION}")
+  endif()
+  message(
+    STATUS
+      "Found NCCL (include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES})")
+  mark_as_advanced(NCCL_ROOT_DIR NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
+endif()

docs/src/dev/custom_metal_kernels.rst

@@ -127,7 +127,7 @@ relying on a copy from ``ensure_row_contiguous``:
       name="myexp_strided",
       input_names=["inp"],
       output_names=["out"],
-      source=source
+      source=source,
       ensure_row_contiguous=False,
   )
 

docs/src/install.rst

@@ -271,7 +271,7 @@ and the CUDA toolkit. For example on Ubuntu, run the following:
    dpkg -i cuda-keyring_1.1-1_all.deb
    apt-get update -y
    apt-get -y install cuda-toolkit-12-9
-   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev libcudnn9-dev-cuda-12 -y
 
 
 When building either the Python or C++ APIs make sure to pass the cmake flag

docs/src/python/nn/functions.rst

@@ -27,6 +27,7 @@ simple functions.
    mish
    prelu
    relu
+   relu2
    relu6
    selu
    sigmoid

docs/src/python/nn/layers.rst

@@ -50,6 +50,7 @@ Layers
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU2
    ReLU6
    RNN
    RoPE

docs/src/usage/compile.rst

@@ -130,8 +130,8 @@ Now make an array, and benchmark both functions:
 .. code-block:: python
 
   x = mx.random.uniform(shape=(32, 1000, 4096))
-  timeit(nn.gelu, x)
-  timeit(mx.compile(nn.gelu), x)
+  timeit(gelu, x)
+  timeit(mx.compile(gelu), x)
 
 On an M1 Max the times are 15.5 and 3.1 milliseconds. The compiled ``gelu`` is
 five times faster.

docs/src/usage/distributed.rst

@@ -184,7 +184,7 @@ almost identical to the example above:
 
     def step(model, x, y):
         loss, grads = loss_grad_fn(model, x, y)
-        grads = mlx.nn.average_gradients(grads) # <---- This line was added
+        grads = mx.nn.average_gradients(grads)  # <---- This line was added
         optimizer.update(model, grads)
         return loss
 

docs/src/usage/export.rst

@@ -164,11 +164,11 @@ to export a function which can be used for inputs with variable shapes:
 
 .. code-block:: python
 
-  mx.export_function("fun.mlxfn", mx.abs, mx.array(0.0), shapeless=True)
+  mx.export_function("fun.mlxfn", mx.abs, mx.array([0.0]), shapeless=True)
   imported_abs = mx.import_function("fun.mlxfn")
 
   # Ok
-  out, = imported_abs(mx.array(-1.0))
+  out, = imported_abs(mx.array([-1.0]))
 
   # Also ok
   out, = imported_abs(mx.array([-1.0, -2.0]))

docs/src/usage/indexing.rst

@@ -107,8 +107,20 @@ same array:
   >>> a
   array([1, 2, 0], dtype=int32)
 
+Note that unlike NumPy, slicing an array creates a copy, not a view. So
+mutating it does not mutate the original array:
 
-Note, unlike NumPy, updates to the same location are nondeterministic:
+.. code-block:: shell
+
+  >>> a = mx.array([1, 2, 3])
+  >>> b = a[:]
+  >>> b[2] = 0
+  >>> b
+  array([1, 2, 0], dtype=int32)
+  >>> a
+  array([1, 2, 3], dtype=int32)
+
+Also unlike NumPy, updates to the same location are nondeterministic:
 
 .. code-block:: shell
 

mlx/backend/common/matmul.h

@@ -13,7 +13,7 @@ inline std::tuple<Shape, Strides, Strides> collapse_batches(
     const array& a,
     const array& b) {
   if (a.ndim() == 2) {
-    return {{1}, {0}, {0}};
+    return {Shape{1}, Strides{0}, Strides{0}};
   }
 
   Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
@@ -38,7 +38,7 @@ inline std::tuple<Shape, Strides, Strides> collapse_batches(
 inline std::tuple<Shape, Strides, Strides, Strides>
 collapse_batches(const array& a, const array& b, const array& c) {
   if (a.ndim() == 2) {
-    return {{1}, {0}, {0}, {0}};
+    return {Shape{1}, Strides{0}, Strides{0}, Strides{0}};
   }
 
   Shape A_bshape{a.shape().begin(), a.shape().end() - 2};

mlx/backend/common/ternary.h

@@ -11,6 +11,8 @@ namespace mlx::core {
 enum class TernaryOpType {
   ScalarScalarScalar,
   VectorVectorVector,
+  VectorVectorScalar,
+  VectorScalarVector,
   General,
 };
 
@@ -25,6 +27,14 @@ get_ternary_op_type(const array& a, const array& b, const array& c) {
       (a.flags().col_contiguous && b.flags().col_contiguous &&
        c.flags().col_contiguous)) {
     topt = TernaryOpType::VectorVectorVector;
+  } else if (
+      b.data_size() == 1 && a.flags().row_contiguous &&
+      c.flags().row_contiguous) {
+    topt = TernaryOpType::VectorScalarVector;
+  } else if (
+      c.data_size() == 1 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    topt = TernaryOpType::VectorVectorScalar;
   } else {
     topt = TernaryOpType::General;
   }
@@ -59,6 +69,8 @@ inline void set_ternary_op_output_data(
             b.flags());
       }
       break;
+    case TernaryOpType::VectorVectorScalar:
+    case TernaryOpType::VectorScalarVector:
     case TernaryOpType::General:
       // Try to donate an input which is row_contiguous
       if (!((a.flags().row_contiguous && maybe_donate(a)) ||

mlx/backend/cpu/compiled.cpp

@@ -15,6 +15,7 @@
 #include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
+#include "mlx/version.h"
 
 namespace mlx::core {
 
@@ -94,7 +95,11 @@ void* compile(
     kernel_file_name = kernel_name;
   }
 
-  auto output_dir = std::filesystem::temp_directory_path();
+  auto output_dir =
+      std::filesystem::temp_directory_path() / "mlx" / version() / "cpu";
+  if (!std::filesystem::exists(output_dir)) {
+    std::filesystem::create_directories(output_dir);
+  }
 
   std::string shared_lib_name = "lib" + kernel_file_name + ".so";
   auto shared_lib_path = (output_dir / shared_lib_name).string();

mlx/backend/cpu/gemms/cblas.cpp

@@ -88,4 +88,47 @@ void matmul<double>(
   }
 }
 
+template <>
+void matmul<complex64_t>(
+    const complex64_t* a,
+    const complex64_t* b,
+    complex64_t* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
+  auto calpha = static_cast<complex64_t>(alpha);
+  auto cbeta = static_cast<complex64_t>(beta);
+
+  for (int i = 0; i < batch_size; ++i) {
+    cblas_cgemm(
+        CblasRowMajor,
+        a_transposed ? CblasTrans : CblasNoTrans, // transA
+        b_transposed ? CblasTrans : CblasNoTrans, // transB
+        M,
+        N,
+        K,
+        &calpha,
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
+        lda,
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
+        ldb,
+        &cbeta,
+        out + M * N * i,
+        ldc);
+  }
+}
+
 } // namespace mlx::core

mlx/backend/cpu/matmul.cpp

@@ -108,6 +108,9 @@ void matmul_general(
   } else if (out.dtype() == float64) {
     matmul_dispatch<double>(
         a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
+  } else if (out.dtype() == complex64) {
+    matmul_dispatch<complex64_t>(
+        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
   } else {
     throw std::runtime_error("[Matmul::eval_cpu] Invalid type.");
   }
@@ -128,10 +131,6 @@ void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
 }
 
 void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[AddMM::eval_cpu] Currently only supports float32.");
-  }
   if (out.size() == 0) {
     out.set_data(allocator::malloc(out.nbytes()));
     return;

mlx/backend/cpu/quantized.cpp

@@ -1,7 +1,5 @@
 // Copyright © 2023 Apple Inc.
 
-#include <cassert>
-
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
@@ -13,6 +11,35 @@ namespace mlx::core {
 
 namespace {
 
+const static float MXFP4_LUT[16] = {
+    +0.0f,
+    +0.5f,
+    +1.0f,
+    +1.5f,
+    +2.0f,
+    +3.0f,
+    +4.0f,
+    +6.0f,
+    -0.0f,
+    -0.5f,
+    -1.0f,
+    -1.5f,
+    -2.0f,
+    -3.0f,
+    -4.0f,
+    -6.0f};
+
+template <typename T>
+static inline T dequantize_scale(uint8_t s) {
+  using FOrI = union {
+    bfloat16_t f;
+    uint16_t i;
+  };
+  FOrI out;
+  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
+  return static_cast<T>(out.f);
+}
+
 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
   return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
 }
@@ -407,6 +434,231 @@ void _qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_qmm(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int bytes_per_pack = get_bytes_per_pack(4);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    std::fill(result, result + N, 0);
+
+    for (int k = 0; k < K; k++) {
+      T* result_local = result;
+      T xi = *x++;
+
+      for (int n = 0; n < N; n += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+        for (int ng = 0; ng < packs_in_group; ng++) {
+          uint8_t wi = *w_local++;
+#pragma clang loop unroll(full)
+          for (int p = 0; p < pack_factor; p++) {
+            (*result_local++) +=
+                xi * scale * static_cast<T>(MXFP4_LUT[wi & 0xf]);
+            wi >>= 4;
+          }
+        }
+      }
+    }
+
+    result += N;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_t(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int bytes_per_pack = get_bytes_per_pack(4);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      const T* x_local = x;
+      T sum = 0;
+      for (int k = 0; k < K; k += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+
+        T gsum = 0;
+        for (int kw = 0; kw < packs_in_group; kw++) {
+          uint8_t wi = *w_local++;
+#pragma clang loop unroll(full)
+          for (int p = 0; p < pack_factor; p++) {
+            gsum += (*x_local++) * static_cast<T>(MXFP4_LUT[wi & 0xf]);
+            wi >>= 4;
+          }
+        }
+        sum += scale * gsum;
+      }
+      *result = sum;
+      result++;
+    }
+
+    x += K;
+  }
+}
+
+template <int S>
+simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
+  if constexpr (S == 8) {
+    constexpr std::array<uint32_t, 8> shifts_ = {{0, 4, 8, 12, 16, 20, 24, 28}};
+    auto shifts(*(simd::Simd<uint32_t, S>*)&shifts_);
+    auto wi = simd::Simd<uint32_t, S>(*w);
+    wi = wi >> shifts;
+    wi = wi & 0xf;
+    simd::Simd<float, S> w_out;
+    for (int i = 0; i < S; ++i) {
+      w_out[i] = MXFP4_LUT[wi[i]];
+    }
+    return w_out;
+  } else {
+    // Appease compiler.. but should never get here
+    throw std::runtime_error("Unsupported combination for simd qmm.");
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_t_simd(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = 32 / 4;
+  constexpr int packs_in_group = group_size / pack_factor;
+  constexpr int S = simd::max_size<T>;
+  static_assert(
+      S % pack_factor == 0, "SIMD size must be divisible by pack factor");
+  constexpr int packs_per_simd = S / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint32_t* w_local = w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      simd::Simd<float, S> acc(0);
+      auto x_local = x;
+      for (int k = 0; k < K; k += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+
+        simd::Simd<float, S> g_acc(0);
+        for (int kw = 0; kw < packs_in_group; kw += packs_per_simd) {
+          // Extract bits
+          auto wf = mxfp4_extract_bits_simd<S>(w_local);
+          w_local += packs_per_simd;
+          simd::Simd<float, S> x_simd = simd::load<T, S>(x_local);
+          g_acc = g_acc + x_simd * wf;
+          x_local += S;
+        }
+        acc = acc + scale * g_acc;
+      }
+
+      *result = T(simd::sum(acc));
+      result++;
+    }
+    x += K;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_transpose(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K,
+    bool transposed_w) {
+  if (transposed_w) {
+    // the simd size must be a multiple of the number of elements per word
+    if constexpr (simd::max_size<T> % 8 == 0) {
+      mxfp4_qmm_t_simd<T>(result, x, w, scales, M, N, K);
+    } else {
+      mxfp4_qmm_t<T>(result, x, w, scales, M, N, K);
+    }
+  } else {
+    mxfp4_qmm<T>(result, x, w, scales, M, N, K);
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.ndim() > 1 ? x.shape(-2) : 1;
+  int N = out.shape(-1);
+  int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
+  int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
+  int batch_size = x.size() / (K * M);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  for (int i = 0; i < batch_size; i++) {
+    mxfp4_qmm_dispatch_transpose<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(i * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(i * w_els, w.shape(), w.strides()),
+        scales_ptr + elem_to_loc(i * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case bfloat16:
+      mxfp4_qmm_dispatch_typed<bfloat16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float16:
+      mxfp4_qmm_dispatch_typed<float16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float32:
+      mxfp4_qmm_dispatch_typed<float>(out, x, w, scales, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 template <typename T>
 void _bs_qmm_dispatch_typed(
     array& out,
@@ -513,115 +765,198 @@ void _bs_qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_bs_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.shape(-2);
+  int N = out.shape(-1);
+
+  int w_els = w.shape(-1) * w.shape(-2);
+  int g_els = scales.shape(-1) * scales.shape(-2);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
+  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
+
+  for (int i = 0; i < lhs_indices.size(); i++) {
+    int x_idx = lhs_indices_ptr[elem_to_loc(
+        i, lhs_indices.shape(), lhs_indices.strides())];
+    int w_idx = rhs_indices_ptr[elem_to_loc(
+        i, rhs_indices.shape(), rhs_indices.strides())];
+    mxfp4_qmm_dispatch_transpose<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
+        scales_ptr +
+            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_bs_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case float32:
+      mxfp4_bs_qmm_dispatch_typed<float>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case float16:
+      mxfp4_bs_qmm_dispatch_typed<float16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case bfloat16:
+      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 } // namespace
 
 void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 4);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
 
-  std::vector<array> temps;
-  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
     if (arr.flags().row_contiguous) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy_cpu(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      encoder.add_temporary(arr_cpy);
+      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous(x_pre);
   auto w = ensure_row_contiguous(w_pre);
   auto scales = ensure_row_contiguous(scales_pre);
-  auto biases = ensure_row_contiguous(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out),
-                    x = array::unsafe_weak_copy(x),
-                    w = array::unsafe_weak_copy(w),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous(inputs[3]);
+    encoder.set_input_array(biases);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      transpose_ = transpose_]() mutable {
+      _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
+    });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_qmm_dispatch(out, x, w, scales, transpose_);
+    });
+  }
 }
 
 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 6);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
-  auto& lhs_indices = inputs[4];
-  auto& rhs_indices = inputs[5];
+  auto& lhs_indices = inputs[inputs.size() - 2];
+  auto& rhs_indices = inputs[inputs.size() - 1];
 
-  std::vector<array> temps;
+  auto& encoder = cpu::get_command_encoder(stream());
   auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &temps](const array& arr) {
+                                          &encoder](const array& arr) {
     auto stride_0 = arr.strides()[arr.ndim() - 2];
     auto stride_1 = arr.strides()[arr.ndim() - 1];
     if (stride_0 == arr.shape(-1) && stride_1 == 1) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy_cpu(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      encoder.add_temporary(arr_cpy);
+      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous_last_dims(x_pre);
   auto w = ensure_row_contiguous_last_dims(w_pre);
   auto scales = ensure_row_contiguous_last_dims(scales_pre);
-  auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_input_array(lhs_indices);
   encoder.set_input_array(rhs_indices);
   encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out),
-                    x = array::unsafe_weak_copy(x),
-                    w = array::unsafe_weak_copy(w),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    lhs_indices = array::unsafe_weak_copy(lhs_indices),
-                    rhs_indices = array::unsafe_weak_copy(rhs_indices),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _bs_qmm_dispatch(
-        out,
-        x,
-        w,
-        scales,
-        biases,
-        lhs_indices,
-        rhs_indices,
-        group_size_,
-        bits_,
-        transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous_last_dims(inputs[3]);
+    encoder.set_input_array(biases);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      transpose_ = transpose_]() mutable {
+      _bs_qmm_dispatch(
+          out,
+          x,
+          w,
+          scales,
+          biases,
+          lhs_indices,
+          rhs_indices,
+          group_size_,
+          bits_,
+          transpose_);
+    });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_bs_qmm_dispatch(
+          out, x, w, scales, lhs_indices, rhs_indices, transpose_);
+    });
+  }
 }
 
 template <typename T, typename U>
@@ -705,7 +1040,7 @@ void dispatch_quantize(
       w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }
 
-void fast::AffineQuantize::eval_cpu(
+void fast::Quantize::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   auto ensure_row_contiguous = [s = stream()](const array& arr) {
@@ -764,7 +1099,7 @@ void fast::AffineQuantize::eval_cpu(
       }
     } else {
       throw std::runtime_error(
-          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
+          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
     }
   });
 }

mlx/backend/cpu/simd/accelerate_simd.h

@@ -9,7 +9,7 @@
 
 #include "mlx/backend/cpu/simd/base_simd.h"
 
-// There seems to be a bug in sims/base.h
+// There seems to be a bug in simd/base_simd.h
 // __XROS_2_0 is not defined, the expression evaluates
 // to true instead of false setting the SIMD library
 // higher than it should be even on macOS < 15
@@ -234,6 +234,7 @@ Simd<T, N> remainder(Simd<T, N> a, Simd<T, N> b) {
 
 template <typename MaskT, typename T1, typename T2, int N>
 Simd<T1, N> select(Simd<MaskT, N> mask, Simd<T1, N> x, Simd<T2, N> y) {
+  static_assert(std::is_same_v<MaskT, bool>);
   if constexpr (sizeof(T1) == 1) {
     return asd::bitselect(y.value, x.value, asd::convert<char>(mask.value));
   } else if constexpr (sizeof(T1) == 2) {
@@ -251,9 +252,13 @@ Simd<T, N> pow(Simd<T, N> base, Simd<T, N> exp) {
     return asd::pow(base.value, exp.value);
   } else {
     Simd<T, N> res = 1;
-    while (any(exp)) {
-      res = select(exp & 1, res * base, res);
-      base = select(exp, base * base, base);
+    // Raising an integer to a negative power is undefined
+    if (any(exp < 0)) {
+      return 0;
+    }
+    while (any(exp > 0)) {
+      res = select((exp & 1) != 0, res * base, res);
+      base = select(exp > 0, base * base, base);
       exp = exp >> 1;
     }
     return res;

mlx/backend/cpu/sort.cpp

@@ -15,6 +15,18 @@ namespace mlx::core {
 
 namespace {
 
+// NaN-aware comparator that places NaNs at the end
+template <typename T>
+bool nan_aware_less(T a, T b) {
+  if constexpr (std::is_floating_point_v<T> || std::is_same_v<T, complex64_t>) {
+    if (std::isnan(a))
+      return false;
+    if (std::isnan(b))
+      return true;
+  }
+  return a < b;
+}
+
 template <typename T>
 struct StridedIterator {
   using iterator_category = std::random_access_iterator_tag;
@@ -130,7 +142,7 @@ void sort(array& out, int axis) {
     StridedIterator st(data_ptr, axis_stride, 0);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
 
-    std::stable_sort(st, ed);
+    std::stable_sort(st, ed, nan_aware_less<T>);
     src_it.step();
   }
 }
@@ -184,6 +196,15 @@ void argsort(const array& in, array& out, int axis) {
     std::stable_sort(st, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
       auto v1 = data_ptr[a * in_stride];
       auto v2 = data_ptr[b * in_stride];
+
+      // Handle NaNs (place them at the end)
+      if (std::is_floating_point<T>::value) {
+        if (std::isnan(v1))
+          return false;
+        if (std::isnan(v2))
+          return true;
+      }
+
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }
@@ -219,7 +240,7 @@ void partition(array& out, int axis, int kth) {
     StridedIterator md(data_ptr, axis_stride, kth);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
 
-    std::nth_element(st, md, ed);
+    std::nth_element(st, md, ed, nan_aware_less<T>);
   }
 }
 
@@ -276,6 +297,15 @@ void argpartition(const array& in, array& out, int axis, int kth) {
     std::nth_element(st, md, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
       auto v1 = data_ptr[a * in_stride];
       auto v2 = data_ptr[b * in_stride];
+
+      // Handle NaNs (place them at the end)
+      if (std::is_floating_point<T>::value) {
+        if (std::isnan(v1))
+          return false;
+        if (std::isnan(v2))
+          return true;
+      }
+
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }

mlx/backend/cpu/unary_ops.h

@@ -77,7 +77,8 @@ struct Real {
 struct Sigmoid {
   template <int N, typename T>
   Simd<T, N> operator()(Simd<T, N> x) {
-    return 1.0f / (1.0f + simd::exp(-x));
+    auto y = 1.0f / (1.0f + simd::exp(simd::abs(x)));
+    return simd::select(x < Simd<T, N>{0}, y, Simd<T, N>{1} - y);
   }
   SINGLE()
 };

mlx/backend/cuda/CMakeLists.txt

@@ -22,6 +22,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/cudnn_utils.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/custom_kernel.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp

mlx/backend/cuda/allocator.cpp

@@ -30,8 +30,20 @@ SmallSizePool::SmallSizePool() {
   next_free_ = buffer_;
 
   CHECK_CUDA_ERROR(cudaMallocManaged(&data_, small_pool_size));
-  CHECK_CUDA_ERROR(
-      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, 0));
+
+  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
+    CHECK_CUDA_ERROR(
+        cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetAccessedBy, loc));
+  }
 
   auto curr = next_free_;
   for (size_t i = 1; i < num_blocks; ++i) {
@@ -79,7 +91,7 @@ CudaAllocator::CudaAllocator()
   // TODO: Set memory limit for multi-device.
   size_t free, total;
   CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
-  memory_limit_ = total * 0.8;
+  memory_limit_ = total * 0.95;
   max_pool_size_ = memory_limit_;
 }
 

mlx/backend/cuda/arange.cu

@@ -6,23 +6,33 @@
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
+#include <cooperative_groups.h>
 #include <nvtx3/nvtx3.hpp>
-#include <thrust/device_ptr.h>
-#include <thrust/transform.h>
 
 namespace mlx::core {
 
 namespace cu {
 
-template <typename T>
-struct Arange {
-  const T start;
-  const T step;
+namespace cg = cooperative_groups;
 
-  __device__ T operator()(uint32_t i) const {
-    return start + i * step;
+template <typename T, typename IdxT, int N_WRITES>
+__global__ void arange(T* out, IdxT size, T start, T step) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_WRITES > size) {
+    for (IdxT i = index * N_WRITES; i < size; ++i) {
+      out[i] = start + i * step;
+    }
+  } else {
+    AlignedVector<T, N_WRITES> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_WRITES; ++i) {
+      out_vec[i] = start + (index * N_WRITES + i) * step;
+    }
+
+    store_vector<N_WRITES>(out, index, out_vec);
   }
-};
+}
 
 } // namespace cu
 
@@ -36,19 +46,23 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& encoder = cu::get_command_encoder(stream());
   encoder.set_output_array(out);
 
-  auto capture = encoder.capture_context();
   dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
     using CTYPE = MLX_GET_TYPE(type_tag);
     using OutType = cuda_type_t<CTYPE>;
-    CTYPE step =
-        static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
-    thrust::transform(
-        cu::thrust_policy(encoder.stream()),
-        thrust::counting_iterator<uint32_t>(0),
-        thrust::counting_iterator<uint32_t>(out.data_size()),
-        thrust::device_pointer_cast(out.data<OutType>()),
-        cu::Arange<OutType>{
-            static_cast<OutType>(start_), static_cast<OutType>(step)});
+    constexpr int N_WRITES = 16 / sizeof(OutType);
+    dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
+      using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+      auto [num_blocks, block_dims] = get_launch_args(out, large(), N_WRITES);
+      encoder.add_kernel_node(
+          cu::arange<OutType, IdxT, N_WRITES>,
+          num_blocks,
+          block_dims,
+          0,
+          out.data<OutType>(),
+          out.data_size(),
+          static_cast<CTYPE>(start_),
+          static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_));
+    });
   });
 }
 

mlx/backend/cuda/compiled.cpp

@@ -332,9 +332,9 @@ void Compiled::eval_gpu(
     encoder.set_output_array(out);
   }
 
-  auto kernel = mod.get_kernel(kernel_name);
+  auto [kernel, max_block_dims] = mod.get_kernel_and_dims(kernel_name);
   auto [num_blocks, block_dims] =
-      get_launch_args(outputs[0], large, work_per_thread);
+      get_launch_args(outputs[0], large, work_per_thread, max_block_dims);
   encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 

mlx/backend/cuda/conv.cpp

@@ -47,7 +47,7 @@ auto& conv_cache() {
       std::pair<
           cudnnBackendDescriptorType_t,
           std::optional<cudnn_frontend::ExecutionPlan>>>
-      cache(/* capacity */ 128);
+      cache("MLX_CUDA_CONV_CACHE_SIZE", /* default_capacity */ 128);
   return cache;
 }
 
@@ -382,20 +382,19 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   }
 
   if (op_graph) {
-    // Setup inputs and outputs.
-    register_args(encoder, backend_type, in, wt, out, out_);
-
     // 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) {
-      throw std::runtime_error("[conv] Unable to find an execution plan.");
-    }
-    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 (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;
+      }
     }
   }
 

mlx/backend/cuda/copy.cu

@@ -15,8 +15,8 @@ void copy_gpu_inplace(
     int64_t offset_out,
     CopyType ctype,
     const Stream& s,
-    const std::optional<array>& dynamic_offset_in,
-    const std::optional<array>& dynamic_offset_out) {
+    std::optional<array> dynamic_offset_in,
+    std::optional<array> dynamic_offset_out) {
   if (out.size() == 0) {
     return;
   }
@@ -44,6 +44,16 @@ void copy_gpu_inplace(
           strides_vec[0]);
     } else {
       if (dynamic_offset_in || dynamic_offset_out) {
+        if (!dynamic_offset_in) {
+          dynamic_offset_in = array(0, int64);
+          encoder.add_temporary(*dynamic_offset_in);
+        }
+        if (!dynamic_offset_out) {
+          dynamic_offset_out = array(0, int64);
+          encoder.add_temporary(*dynamic_offset_out);
+        }
+        encoder.set_input_array(*dynamic_offset_in);
+        encoder.set_input_array(*dynamic_offset_out);
         copy_general_dynamic(
             encoder,
             ctype,
@@ -54,8 +64,8 @@ void copy_gpu_inplace(
             shape_collapsed,
             strides_vec[0],
             strides_vec[1],
-            dynamic_offset_in ? *dynamic_offset_in : array(0, int64),
-            dynamic_offset_out ? *dynamic_offset_out : array(0, int64));
+            *dynamic_offset_in,
+            *dynamic_offset_out);
       } else {
         copy_general(
             encoder,

mlx/backend/cuda/cudnn_utils.cpp

@@ -210,6 +210,9 @@ std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
     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);
 }
 

mlx/backend/cuda/device.cpp

@@ -14,10 +14,6 @@ namespace mlx::core::cu {
 
 namespace {
 
-// Can be tuned with MLX_MAX_OPS_PER_BUFFER
-// This should be less than 255
-constexpr int default_max_nodes_per_graph = 20;
-
 #define CHECK_CUDNN_ERROR(cmd) check_cudnn_error(#cmd, (cmd))
 
 void check_cudnn_error(const char* name, cudnnStatus_t err) {
@@ -27,11 +23,11 @@ void check_cudnn_error(const char* name, cudnnStatus_t err) {
   }
 }
 
-int cuda_graph_cache_size() {
-  static int cache_size = []() {
-    return env::get_var("MLX_CUDA_GRAPH_CACHE_SIZE", 100);
+bool use_cuda_graphs() {
+  static bool use_graphs = []() {
+    return env::get_var("MLX_USE_CUDA_GRAPHS", true);
   }();
-  return cache_size;
+  return use_graphs;
 }
 
 } // namespace
@@ -68,8 +64,8 @@ Device::~Device() {
 
 void Device::make_current() {
   // We need to set/get current CUDA device very frequently, cache it to reduce
-  // actual calls of CUDA APIs. This function assumes single-thread in host.
-  static int current = 0;
+  // actual calls of CUDA APIs.
+  static thread_local int current = 0;
   if (current != device_) {
     CHECK_CUDA_ERROR(cudaSetDevice(device_));
     current = device_;
@@ -86,11 +82,19 @@ CommandEncoder& Device::get_command_encoder(Stream s) {
 
 CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
   enc.device().make_current();
+  if (!use_cuda_graphs()) {
+    return;
+  }
   CHECK_CUDA_ERROR(
       cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
 }
 
 CommandEncoder::CaptureContext::~CaptureContext() {
+  if (!use_cuda_graphs()) {
+    enc.node_count_++;
+    return;
+  }
+
   graph.end_capture(enc.stream());
   if (discard) {
     return;
@@ -105,6 +109,9 @@ CommandEncoder::ConcurrentContext::ConcurrentContext(CommandEncoder& enc)
 
 CommandEncoder::ConcurrentContext::~ConcurrentContext() {
   enc.in_concurrent_ = false;
+  if (!use_cuda_graphs()) {
+    return;
+  }
 
   // Use an empty graph node for synchronization
   CommandEncoder::GraphNode empty{NULL, 'E', std::to_string(enc.node_count_++)};
@@ -186,35 +193,43 @@ CommandEncoder::CommandEncoder(Device& d)
     : device_(d),
       stream_(d),
       graph_(d),
-      graph_cache_(cuda_graph_cache_size()) {}
+      worker_(d),
+      graph_cache_("MLX_CUDA_GRAPH_CACHE_SIZE", /* default_capacity */ 400) {}
 
 void CommandEncoder::add_completed_handler(std::function<void()> task) {
   worker_.add_task(std::move(task));
 }
 
 void CommandEncoder::set_input_array(const array& arr) {
+  if (!use_cuda_graphs()) {
+    return;
+  }
   auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
   active_deps_.push_back(id);
 }
 
 void CommandEncoder::set_output_array(const array& arr) {
+  if (!use_cuda_graphs()) {
+    return;
+  }
+
   auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
   active_deps_.push_back(id);
   active_outputs_.push_back(id);
 }
 
-void CommandEncoder::maybe_commit() {
-  if (node_count_ >= env::max_ops_per_buffer(default_max_nodes_per_graph)) {
-    commit();
-  }
-}
-
 void CommandEncoder::add_kernel_node(
     void* func,
     dim3 grid_dim,
     dim3 block_dim,
     uint32_t smem_bytes,
     void** params) {
+  if (!use_cuda_graphs()) {
+    node_count_++;
+    CHECK_CUDA_ERROR(cudaLaunchKernel(
+        func, grid_dim, block_dim, params, smem_bytes, stream()));
+    return;
+  }
   cudaKernelNodeParams kernel_params = {0};
   kernel_params.func = func;
   kernel_params.gridDim = grid_dim;
@@ -230,6 +245,23 @@ void CommandEncoder::add_kernel_node(
     dim3 block_dim,
     uint32_t smem_bytes,
     void** params) {
+  if (!use_cuda_graphs()) {
+    node_count_++;
+    CHECK_CUDA_ERROR(cuLaunchKernel(
+        func,
+        grid_dim.x,
+        grid_dim.y,
+        grid_dim.z,
+        block_dim.x,
+        block_dim.y,
+        block_dim.z,
+        smem_bytes,
+        stream(),
+        params,
+        nullptr));
+    return;
+  }
+
   CUDA_KERNEL_NODE_PARAMS kernel_params = {0};
   kernel_params.func = func;
   kernel_params.gridDimX = grid_dim.x;
@@ -256,20 +288,38 @@ void CommandEncoder::add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params) {
 }
 
 void CommandEncoder::add_graph_node(cudaGraph_t child) {
+  if (!use_cuda_graphs()) {
+    node_count_++;
+    CudaGraphExec graph_exec;
+    graph_exec.instantiate(child);
+    device_.make_current();
+    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream()));
+    return;
+  }
   cudaGraphNode_t node;
   CHECK_CUDA_ERROR(cudaGraphAddChildGraphNode(&node, graph_, NULL, 0, child));
   insert_graph_dependencies(GraphNode{node, 'G'});
 }
 
+int CommandEncoder::get_num_ops() {
+  return node_count_;
+}
+
 void CommandEncoder::commit() {
   nvtx3::scoped_range r("CommandEncoder::commit");
   if (!temporaries_.empty()) {
     add_completed_handler([temporaries = std::move(temporaries_)]() {});
   }
-  if (node_count_ > 0) {
+  if (use_cuda_graphs() && node_count_ > 0) {
     if (!from_nodes_.empty()) {
       CHECK_CUDA_ERROR(cudaGraphAddDependencies(
-          graph_, from_nodes_.data(), to_nodes_.data(), from_nodes_.size()));
+          graph_,
+          from_nodes_.data(),
+          to_nodes_.data(),
+#if CUDART_VERSION >= 13000
+          nullptr, // edgeData
+#endif // CUDART_VERSION >= 13000
+          from_nodes_.size()));
     }
 
     graph_key_ += ".";
@@ -303,7 +353,6 @@ void CommandEncoder::commit() {
     CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
 
     // Reset state
-    node_count_ = 0;
     graph_node_count_ = 0;
     empty_node_count_ = 0;
     from_nodes_.clear();
@@ -315,6 +364,7 @@ void CommandEncoder::commit() {
 
   // Put completion handlers in a batch.
   worker_.commit(stream_);
+  node_count_ = 0;
 }
 
 void CommandEncoder::synchronize() {

mlx/backend/cuda/device.h

@@ -76,9 +76,6 @@ class CommandEncoder {
       uint32_t smem_bytes,
       void** params);
 
-  // Low-level graph helpers.
-  void add_kernel_node(const cudaKernelNodeParams& params);
-  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
   void add_graph_node(cudaGraph_t child);
 
   void add_temporary(const array& arr) {
@@ -86,7 +83,7 @@ class CommandEncoder {
   }
 
   void add_completed_handler(std::function<void()> task);
-  void maybe_commit();
+  int get_num_ops();
   void commit();
 
   Device& device() {
@@ -101,6 +98,9 @@ class CommandEncoder {
   void synchronize();
 
  private:
+  void add_kernel_node(const cudaKernelNodeParams& params);
+  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
+
   struct GraphNode {
     cudaGraphNode_t node;
     // K = kernel
@@ -140,7 +140,7 @@ class Device {
   Device(const Device&) = delete;
   Device& operator=(const Device&) = delete;
 
-  // Make this device the current cuda device, required by some cuda calls.
+  // Make this device the current cuda device, this method is thread-safe.
   void make_current();
 
   CommandEncoder& get_command_encoder(Stream s);

mlx/backend/cuda/device/binary_ops.cuh

@@ -204,6 +204,12 @@ struct Power {
   __device__ T operator()(T base, T exp) {
     if constexpr (cuda::std::is_integral_v<T>) {
       T res = 1;
+      // Raising an integer to a negative power is undefined
+      if constexpr (cuda::std::is_signed_v<T>) {
+        if (exp < 0) {
+          return 0;
+        }
+      }
       while (exp) {
         if (exp & 1) {
           res *= base;

mlx/backend/cuda/device/cast_op.cuh

@@ -6,7 +6,6 @@
 
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
-#include <thrust/iterator/transform_iterator.h>
 
 namespace mlx::core::cu {
 
@@ -116,15 +115,4 @@ inline __host__ __device__ auto cast_to(SrcT x) {
   return CastOp<SrcT, DstT>{}(x);
 }
 
-// Return an iterator that cast the value to DstT using CastOp.
-template <typename DstT, typename Iterator>
-inline __host__ __device__ auto make_cast_iterator(Iterator it) {
-  using SrcT = typename cuda::std::iterator_traits<Iterator>::value_type;
-  if constexpr (std::is_same_v<SrcT, DstT>) {
-    return it;
-  } else {
-    return thrust::make_transform_iterator(it, CastOp<SrcT, DstT>{});
-  }
-}
-
 } // namespace mlx::core::cu

mlx/backend/cuda/device/unary_ops.cuh

@@ -257,8 +257,8 @@ struct Round {
 struct Sigmoid {
   template <typename T>
   __device__ T operator()(T x) {
-    T y = 1 / (1 + exp(-abs(x)));
-    return (x < 0) ? 1 - y : y;
+    T y = 1 / (1 + exp(abs(x)));
+    return (x < 0) ? y : 1 - y;
   }
 };
 

mlx/backend/cuda/device/utils.cuh

@@ -1,6 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
-// This file must not include any host-only code, utilies that work under both
+// This file must not include any host-only code, utilities that work under both
 // host and device can be put here.
 //
 // See more about the requirements at:
@@ -202,7 +202,7 @@ struct Limits<
   }
 };
 
-// CUDA 11 does not have host side arithmatic operators for half types.
+// CUDA 11 does not have host side arithmetic operators for half types.
 template <typename T>
 struct Limits<
     T,

mlx/backend/cuda/distributed.cu

@@ -0,0 +1,56 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/distributed/primitives.h"
+#include "mlx/primitives.h"
+
+#include <cassert>
+
+namespace mlx::core::distributed {
+void AllReduce::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto set_input_output =
+      [s = stream()](const array& in, array& out) -> std::pair<array, array> {
+    if (!in.flags().row_contiguous) {
+      copy_gpu(in, out, CopyType::General, s);
+      return {out, out};
+    } else if (in.is_donatable()) {
+      out.copy_shared_buffer(in);
+      return {in, out};
+    } else {
+      out.set_data(allocator::malloc(out.nbytes()));
+      return {in, out};
+    }
+  };
+
+  auto [input, output] = set_input_output(inputs[0], outputs[0]);
+
+  auto& encoder = cu::get_command_encoder(stream());
+  encoder.set_input_array(input);
+  encoder.set_output_array(output);
+
+  auto capture = encoder.capture_context();
+  auto& s = stream();
+
+  switch (reduce_type_) {
+    case Sum:
+      distributed::detail::all_sum(group(), input, output, s);
+      break;
+    case Max:
+      distributed::detail::all_max(group(), input, output, s);
+      break;
+    case Min:
+      distributed::detail::all_min(group(), input, output, s);
+      break;
+    default:
+      throw std::runtime_error(
+          "Only all reduce sum, max, and min are supported.");
+  }
+}
+} // namespace mlx::core::distributed

mlx/backend/cuda/eval.cpp

@@ -5,18 +5,24 @@
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/gpu/available.h"
 #include "mlx/primitives.h"
+#include "mlx/scheduler.h"
 
 #include <nvtx3/nvtx3.hpp>
 
 namespace mlx::core::gpu {
 
+// Can be tuned with MLX_MAX_OPS_PER_BUFFER
+constexpr int default_max_nodes_per_graph = 20;
+
 bool is_available() {
   return true;
 }
 
 void new_stream(Stream s) {
-  // Force initalization of cuda, so cuda runtime get destroyed at last.
+  // Force initalization of CUDA, so CUDA runtime get destroyed at last.
   cudaFree(nullptr);
+  // Make sure CUDA event pool get destroyed after device and stream.
+  cu::CudaEvent::init_pool();
   // Ensure the static stream objects get created.
   cu::get_command_encoder(s);
 }
@@ -34,7 +40,8 @@ void eval(array& arr) {
     arr.primitive().eval_gpu(arr.inputs(), outputs);
   }
 
-  auto& encoder = cu::get_command_encoder(arr.primitive().stream());
+  auto& stream = arr.primitive().stream();
+  auto& encoder = cu::get_command_encoder(stream);
   // Keep used buffers alive until kernel finishes running.
   for (auto& in : arr.inputs()) {
     // Except for the donated one.
@@ -45,7 +52,14 @@ void eval(array& arr) {
   for (auto& s : arr.siblings()) {
     encoder.add_temporary(s);
   }
-  encoder.maybe_commit();
+
+  if (encoder.get_num_ops() >=
+      env::max_ops_per_buffer(default_max_nodes_per_graph)) {
+    scheduler::notify_new_task(stream);
+    encoder.add_completed_handler(
+        [stream]() { scheduler::notify_task_completion(stream); });
+    encoder.commit();
+  }
 }
 
 void finalize(Stream s) {

mlx/backend/cuda/event.cu

@@ -3,10 +3,12 @@
 #include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/event.h"
-#include "mlx/backend/cuda/utils.h"
 #include "mlx/event.h"
 #include "mlx/scheduler.h"
 
+#include <map>
+#include <vector>
+
 #include <nvtx3/nvtx3.hpp>
 
 namespace mlx::core {
@@ -17,104 +19,180 @@ namespace cu {
 // CudaEvent implementations
 ///////////////////////////////////////////////////////////////////////////////
 
-// Cuda event managed with RAII.
-class CudaEventHandle {
+namespace {
+
+// Manage cached cudaEvent_t objects.
+class CudaEventPool {
  public:
-  CudaEventHandle() {
-    CHECK_CUDA_ERROR(cudaEventCreateWithFlags(
-        &event_, cudaEventDisableTiming | cudaEventBlockingSync));
+  CudaEventHandle create(Device& d, int flags) {
+    if (!on_creation_thread()) {
+      return CudaEventHandle(d, flags);
+    }
+    auto& cache = cache_for(d, flags);
+    if (cache.empty()) {
+      return CudaEventHandle(d, flags);
+    } else {
+      CudaEventHandle ret = std::move(cache.back());
+      cache.pop_back();
+      return ret;
+    }
   }
 
-  ~CudaEventHandle() {
-    CHECK_CUDA_ERROR(cudaEventDestroy(event_));
-  }
-
-  CudaEventHandle(const CudaEventHandle&) = delete;
-  CudaEventHandle& operator=(const CudaEventHandle&) = delete;
-
-  operator cudaEvent_t() const {
-    return event_;
+  void release(CudaEventHandle event) {
+    if (!on_creation_thread()) {
+      // Event will be destroyed directly instead of getting moved to cache.
+      return;
+    }
+    cache_for(event.device, event.flags).push_back(std::move(event));
   }
 
  private:
-  cudaEvent_t event_;
+  std::vector<CudaEventHandle>& cache_for(Device& d, int flags) {
+    return cache_[d.cuda_device()][flags];
+  }
+
+  bool on_creation_thread() {
+    return std::this_thread::get_id() == thread_id_;
+  }
+
+  // The CudaEvent may be created and destroyed on different threads (for
+  // example when waiting on GPU work in CPU stream), we don't want to make
+  // the cache thread-safe as it adds overhead, so we just skip cache when
+  // using events in worker threads.
+  std::thread::id thread_id_{std::this_thread::get_id()};
+
+  // {device: {flags: [events]}}
+  std::map<int, std::map<int, std::vector<CudaEventHandle>>> cache_;
 };
 
-CudaEvent::CudaEvent() : event_(std::make_shared<CudaEventHandle>()) {}
+CudaEventPool& cuda_event_pool() {
+  static CudaEventPool pool;
+  return pool;
+}
+
+} // namespace
+
+CudaEventHandle::CudaEventHandle(Device& d, int flags)
+    : device(d), flags(flags) {
+  device.make_current();
+  CHECK_CUDA_ERROR(cudaEventCreateWithFlags(&handle_, flags));
+  assert(handle_ != nullptr);
+}
+
+CudaEvent::CudaEvent(Device& d, int flags)
+    : event_(cuda_event_pool().create(d, flags)) {}
+
+CudaEvent::~CudaEvent() {
+  cuda_event_pool().release(std::move(event_));
+}
 
 void CudaEvent::wait() {
   nvtx3::scoped_range r("cu::CudaEvent::wait");
-  if (!recorded_) {
-    throw std::runtime_error("Should not wait on a CudaEvent before record.");
-  }
-  cudaEventSynchronize(*event_);
+  event_.device.make_current();
+  cudaEventSynchronize(event_);
 }
 
 void CudaEvent::wait(cudaStream_t stream) {
-  if (!recorded_) {
-    throw std::runtime_error("Should not wait on a CudaEvent before record.");
-  }
-  cudaStreamWaitEvent(stream, *event_);
-}
-
-void CudaEvent::wait(Stream s) {
-  if (s.device == mlx::core::Device::cpu) {
-    scheduler::enqueue(s, [*this]() mutable { wait(); });
-  } else {
-    auto& enc = cu::get_command_encoder(s);
-    enc.commit();
-    wait(enc.stream());
-  }
+  event_.device.make_current();
+  cudaStreamWaitEvent(stream, event_);
 }
 
 void CudaEvent::record(cudaStream_t stream) {
-  cudaEventRecord(*event_, stream);
-  recorded_ = true;
-}
-
-void CudaEvent::record(Stream s) {
-  if (s.device == mlx::core::Device::cpu) {
-    throw std::runtime_error("CudaEvent can not wait on cpu stream.");
-  } else {
-    auto& enc = cu::get_command_encoder(s);
-    enc.commit();
-    record(enc.stream());
-  }
+  event_.device.make_current();
+  cudaEventRecord(event_, stream);
 }
 
 bool CudaEvent::completed() const {
-  return cudaEventQuery(*event_) == cudaSuccess;
+  // Note: cudaEventQuery can be safely called from any device.
+  return cudaEventQuery(event_) == cudaSuccess;
 }
 
+// static
+void CudaEvent::init_pool() {
+  cuda_event_pool();
+}
+
+// Wraps CudaEvent with a few features:
+// 1. The class can be copied.
+// 2. Make wait/record work with CPU streams.
+// 3. Add checks for waiting on un-recorded event.
+class CopyableCudaEvent {
+ public:
+  explicit CopyableCudaEvent(Device& d)
+      : event_(std::make_shared<CudaEvent>(
+            d,
+            cudaEventDisableTiming | cudaEventBlockingSync)) {}
+
+  void wait() {
+    event_->wait();
+  }
+
+  void wait(Stream s) {
+    if (s.device == mlx::core::Device::cpu) {
+      scheduler::enqueue(s, [*this]() mutable {
+        check_recorded();
+        event_->wait();
+      });
+    } else {
+      check_recorded();
+      auto& encoder = cu::get_command_encoder(s);
+      encoder.commit();
+      event_->wait(encoder.stream());
+    }
+  }
+
+  void record(Stream s) {
+    if (s.device == mlx::core::Device::cpu) {
+      throw std::runtime_error("CudaEvent can not wait on CPU stream.");
+    } else {
+      auto& encoder = cu::get_command_encoder(s);
+      encoder.commit();
+      event_->record(encoder.stream());
+      recorded_ = true;
+    }
+  }
+
+  bool is_signaled() const {
+    return recorded_ && event_->completed();
+  }
+
+ private:
+  void check_recorded() const {
+    if (!recorded_) {
+      throw std::runtime_error(
+          "Should not wait on a CudaEvent before recording.");
+    }
+  }
+
+  std::shared_ptr<CudaEvent> event_;
+  bool recorded_{false};
+};
+
 ///////////////////////////////////////////////////////////////////////////////
-// SharedEvent implementations
+// AtomicEvent implementations
 ///////////////////////////////////////////////////////////////////////////////
 
-__host__ __device__ void event_wait(SharedEvent::Atomic* ac, uint64_t value) {
+__host__ __device__ void event_wait(AtomicEvent::Atomic* ac, uint64_t value) {
   uint64_t current;
   while ((current = ac->load()) < value) {
     ac->wait(current);
   }
 }
 
-__host__ __device__ void event_signal(SharedEvent::Atomic* ac, uint64_t value) {
+__host__ __device__ void event_signal(AtomicEvent::Atomic* ac, uint64_t value) {
   ac->store(value);
   ac->notify_all();
 }
 
-__global__ void event_wait_kernel(SharedEvent::Atomic* ac, uint64_t value) {
+__global__ void event_wait_kernel(AtomicEvent::Atomic* ac, uint64_t value) {
   event_wait(ac, value);
 }
 
-__global__ void event_signal_kernel(SharedEvent::Atomic* ac, uint64_t value) {
+__global__ void event_signal_kernel(AtomicEvent::Atomic* ac, uint64_t value) {
   event_signal(ac, value);
 }
 
-SharedEvent::Atomic* to_atomic(std::shared_ptr<Buffer> buf) {
-  return static_cast<SharedEvent::Atomic*>(buf->raw_ptr());
-}
-
-SharedEvent::SharedEvent() {
+AtomicEvent::AtomicEvent() {
   buf_ = std::shared_ptr<Buffer>(
       new Buffer{allocator().malloc(sizeof(Atomic))}, [](Buffer* ptr) {
         allocator().free(*ptr);
@@ -123,17 +201,17 @@ SharedEvent::SharedEvent() {
   *static_cast<uint64_t*>(buf_->raw_ptr()) = 0;
 }
 
-void SharedEvent::wait(uint64_t value) {
-  nvtx3::scoped_range r("cu::SharedEvent::wait");
-  event_wait(to_atomic(buf_), value);
+void AtomicEvent::wait(uint64_t value) {
+  nvtx3::scoped_range r("cu::AtomicEvent::wait");
+  event_wait(atomic(), value);
 }
 
-void SharedEvent::wait(cudaStream_t stream, uint64_t value) {
-  event_wait_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
+void AtomicEvent::wait(cudaStream_t stream, uint64_t value) {
+  event_wait_kernel<<<1, 1, 0, stream>>>(atomic(), value);
 }
 
-void SharedEvent::wait(Stream s, uint64_t value) {
-  nvtx3::scoped_range r("cu::SharedEvent::wait(s)");
+void AtomicEvent::wait(Stream s, uint64_t value) {
+  nvtx3::scoped_range r("cu::AtomicEvent::wait(s)");
   if (s.device == mlx::core::Device::cpu) {
     scheduler::enqueue(s, [*this, value]() mutable { wait(value); });
   } else {
@@ -144,17 +222,17 @@ void SharedEvent::wait(Stream s, uint64_t value) {
   }
 }
 
-void SharedEvent::signal(uint64_t value) {
-  nvtx3::scoped_range r("cu::SharedEvent::signal");
-  event_signal(to_atomic(buf_), value);
+void AtomicEvent::signal(uint64_t value) {
+  nvtx3::scoped_range r("cu::AtomicEvent::signal");
+  event_signal(atomic(), value);
 }
 
-void SharedEvent::signal(cudaStream_t stream, uint64_t value) {
-  event_signal_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
+void AtomicEvent::signal(cudaStream_t stream, uint64_t value) {
+  event_signal_kernel<<<1, 1, 0, stream>>>(atomic(), value);
 }
 
-void SharedEvent::signal(Stream s, uint64_t value) {
-  nvtx3::scoped_range r("cu::SharedEvent::signal(s)");
+void AtomicEvent::signal(Stream s, uint64_t value) {
+  nvtx3::scoped_range r("cu::AtomicEvent::signal(s)");
   if (s.device == mlx::core::Device::cpu) {
     // Signal through a GPU stream so the atomic is updated in GPU - updating
     // the atomic in CPU sometimes does not get GPU notified.
@@ -168,14 +246,14 @@ void SharedEvent::signal(Stream s, uint64_t value) {
   }
 }
 
-bool SharedEvent::is_signaled(uint64_t value) const {
-  nvtx3::scoped_range r("cu::SharedEvent::is_signaled");
-  return to_atomic(buf_)->load() >= value;
+bool AtomicEvent::is_signaled(uint64_t value) const {
+  nvtx3::scoped_range r("cu::AtomicEvent::is_signaled");
+  return atomic()->load() >= value;
 }
 
-uint64_t SharedEvent::value() const {
-  nvtx3::scoped_range r("cu::SharedEvent::value");
-  return to_atomic(buf_)->load();
+uint64_t AtomicEvent::value() const {
+  nvtx3::scoped_range r("cu::AtomicEvent::value");
+  return atomic()->load();
 }
 
 } // namespace cu
@@ -188,14 +266,14 @@ namespace {
 
 struct EventImpl {
   // CudaEvent is preferred when possible because it is fast, however we have
-  // to fallback to SharedEvent in following cases:
+  // to fallback to AtomicEvent in following cases:
   // 1. the event is used to wait/signal a cpu stream;
   // 2. signal value other than 1 has been specified.
-  std::unique_ptr<cu::CudaEvent> cuda;
-  std::unique_ptr<cu::SharedEvent> shared;
+  std::unique_ptr<cu::CopyableCudaEvent> cuda;
+  std::unique_ptr<cu::AtomicEvent> atomic;
 
   bool is_created() const {
-    return cuda || shared;
+    return cuda || atomic;
   }
 
   void ensure_created(Stream s, uint64_t signal_value) {
@@ -203,10 +281,10 @@ struct EventImpl {
       return;
     }
     if (s.device == mlx::core::Device::cpu || signal_value > 1) {
-      nvtx3::mark("Using slow SharedEvent");
-      shared = std::make_unique<cu::SharedEvent>();
+      nvtx3::mark("Using slow AtomicEvent");
+      atomic = std::make_unique<cu::AtomicEvent>();
     } else {
-      cuda = std::make_unique<cu::CudaEvent>();
+      cuda = std::make_unique<cu::CopyableCudaEvent>(cu::device(s.device));
     }
   }
 };
@@ -225,7 +303,7 @@ void Event::wait() {
     assert(value() == 1);
     event->cuda->wait();
   } else {
-    event->shared->wait(value());
+    event->atomic->wait(value());
   }
 }
 
@@ -236,7 +314,7 @@ void Event::wait(Stream s) {
     assert(value() == 1);
     event->cuda->wait(s);
   } else {
-    event->shared->wait(s, value());
+    event->atomic->wait(s, value());
   }
 }
 
@@ -247,7 +325,7 @@ void Event::signal(Stream s) {
     assert(value() == 1);
     event->cuda->record(s);
   } else {
-    event->shared->signal(s, value());
+    event->atomic->signal(s, value());
   }
 }
 
@@ -258,9 +336,9 @@ bool Event::is_signaled() const {
   }
   if (event->cuda) {
     assert(value() == 1);
-    return event->cuda->recorded() && event->cuda->completed();
+    return event->cuda->is_signaled();
   } else {
-    return event->shared->is_signaled(value());
+    return event->atomic->is_signaled(value());
   }
 }
 

mlx/backend/cuda/event.h

@@ -3,49 +3,60 @@
 #pragma once
 
 #include "mlx/allocator.h"
+#include "mlx/backend/cuda/utils.h"
 #include "mlx/stream.h"
 
+#include <memory>
+
 #include <cuda_runtime.h>
 #include <cuda/atomic>
 
-#include <memory>
-
 namespace mlx::core::cu {
 
-class CudaEventHandle;
+class Device;
+
+// RAII-managed move-only wrapper of cudaEvent_t.
+struct CudaEventHandle : public CudaHandle<cudaEvent_t, cudaEventDestroy> {
+  CudaEventHandle(Device& d, int flags);
+  Device& device;
+  int flags;
+};
 
 // Wrapper of native cuda event. It can synchronize between GPU streams, or wait
 // on GPU stream in CPU stream, but can not wait on CPU stream.
 class CudaEvent {
  public:
-  CudaEvent();
+  CudaEvent(Device& d, int flags);
+  ~CudaEvent();
+
+  CudaEvent(CudaEvent&&) = default;
+  CudaEvent& operator=(CudaEvent&&) = default;
+
+  CudaEvent(const CudaEvent&) = delete;
+  CudaEvent& operator=(const CudaEvent&) = delete;
 
   void wait();
   void wait(cudaStream_t stream);
-  void wait(Stream s);
   void record(cudaStream_t stream);
-  void record(Stream s);
 
   // Return whether the recorded kernels have completed. Note that this method
   // returns true if record() has not been called.
   bool completed() const;
 
-  bool recorded() const {
-    return recorded_;
-  }
+  // Internal: make sure event pool is initialized.
+  static void init_pool();
 
  private:
-  bool recorded_{false};
-  std::shared_ptr<CudaEventHandle> event_;
+  CudaEventHandle event_;
 };
 
 // Event that can synchronize between CPU and GPU. It is much slower than
 // CudaEvent so the latter should always be preferred when possible.
-class SharedEvent {
+class AtomicEvent {
  public:
   using Atomic = cuda::atomic<uint64_t>;
 
-  SharedEvent();
+  AtomicEvent();
 
   void wait(uint64_t value);
   void wait(cudaStream_t stream, uint64_t value);
@@ -57,7 +68,11 @@ class SharedEvent {
   uint64_t value() const;
 
  private:
-  std::shared_ptr<mlx::core::allocator::Buffer> buf_;
+  Atomic* atomic() const {
+    return static_cast<AtomicEvent::Atomic*>(buf_->raw_ptr());
+  }
+
+  std::shared_ptr<allocator::Buffer> buf_;
 };
 
 } // namespace mlx::core::cu

mlx/backend/cuda/fence.cpp

@@ -7,7 +7,7 @@ namespace mlx::core {
 
 struct FenceImpl {
   uint32_t count;
-  cu::SharedEvent event;
+  cu::AtomicEvent event;
 };
 
 Fence::Fence(Stream s) {

mlx/backend/cuda/gemms/cublas_gemm.cpp

@@ -50,8 +50,10 @@ cublasComputeType_t dtype_to_compute_type(Dtype dtype) {
       return mlx::core::env::enable_tf32() ? CUBLAS_COMPUTE_32F_FAST_TF32
                                            : CUBLAS_COMPUTE_32F;
     case float64:
-    case complex64:
       return CUBLAS_COMPUTE_64F;
+    case complex64:
+      return mlx::core::env::enable_tf32() ? CUBLAS_COMPUTE_32F_FAST_TF32
+                                           : CUBLAS_COMPUTE_32F;
     default:
       throw std::runtime_error(fmt::format(
           "Unsupported dtype in CublasGemm: {}.", dtype_to_string(dtype)));
@@ -85,10 +87,10 @@ cublasLtMatrixLayout_t create_matrix_layout(
     int32_t batch_count,
     int64_t batch_stride) {
   cublasLtMatrixLayout_t desc;
+  if (transposed) {
+    std::swap(rows, cols);
+  }
   CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutCreate(&desc, type, rows, cols, ld));
-  cublasLtOrder_t order = transposed ? CUBLASLT_ORDER_COL : CUBLASLT_ORDER_ROW;
-  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-      desc, CUBLASLT_MATRIX_LAYOUT_ORDER, &order, sizeof(cublasLtOrder_t)));
   if (batch_count > 1) {
     CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
         desc,
@@ -126,37 +128,47 @@ CublasGemm::CublasGemm(
       N_(b_cols) {
   heuristic_.state = CUBLAS_STATUS_NOT_INITIALIZED;
 
-  auto scale_type = dtype_to_cublas_type(dtype);
+  scale_type_ = dtype_to_cublas_type(dtype);
   if (dtype == bfloat16 || dtype == float16) {
-    scale_type = CUDA_R_32F;
+    scale_type_ = CUDA_R_32F;
   }
+
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescCreate(
-      &matmul_desc_, dtype_to_compute_type(dtype), scale_type));
+      &matmul_desc_, dtype_to_compute_type(dtype), scale_type_));
   int32_t pointer_mode = CUBLASLT_POINTER_MODE_HOST;
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
       matmul_desc_,
       CUBLASLT_MATMUL_DESC_POINTER_MODE,
       &pointer_mode,
       sizeof(int32_t)));
-  cublasOperation_t op = CUBLAS_OP_N;
+
+  // In cublasLt matrices use column-major layout, while it is possible to use
+  // the CUBLASLT_ORDER_ROW option to switch to row-major layout, the bias
+  // epilogue does not work with the option. So instead we swap A and B to make
+  // cublasLt return the row-major result, which works because:
+  // - the data of a matrix in row-major layout is identical to its transpose in
+  //   column-major layout
+  // - C^T = (A @ B)^T = B^T @ A^T
+  cublasOperation_t a_op = b_transposed ? CUBLAS_OP_T : CUBLAS_OP_N;
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
       matmul_desc_,
       CUBLASLT_MATMUL_DESC_TRANSA,
-      &op,
+      &a_op,
       sizeof(cublasOperation_t)));
+  cublasOperation_t b_op = a_transposed ? CUBLAS_OP_T : CUBLAS_OP_N;
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
       matmul_desc_,
       CUBLASLT_MATMUL_DESC_TRANSB,
-      &op,
+      &b_op,
       sizeof(cublasOperation_t)));
 
   auto type = dtype_to_cublas_type(dtype);
   a_desc_ = create_matrix_layout(
-      type, a_rows, a_cols, a_transposed, lda, batch_count, a_batch_stride);
+      type, b_cols, b_rows, b_transposed, ldb, batch_count, b_batch_stride);
   b_desc_ = create_matrix_layout(
-      type, b_rows, b_cols, b_transposed, ldb, batch_count, b_batch_stride);
+      type, a_cols, a_rows, a_transposed, lda, batch_count, a_batch_stride);
   out_desc_ = create_matrix_layout(
-      type, a_rows, b_cols, false, b_cols, batch_count, a_rows * b_cols);
+      type, b_cols, a_rows, false, b_cols, batch_count, a_rows * b_cols);
 }
 
 CublasGemm::CublasGemm(
@@ -191,7 +203,7 @@ CublasGemm::CublasGemm(
           b_batch_stride) {
   auto type = dtype_to_cublas_type(dtype);
   c_desc_ = create_matrix_layout(
-      type, a_rows, b_cols, false, ldc, batch_count, c_batch_stride);
+      type, b_cols, a_rows, false, ldc, batch_count, c_batch_stride);
 }
 
 CublasGemm::~CublasGemm() {
@@ -213,14 +225,30 @@ void CublasGemm::set_out(
   CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(out_desc_));
   out_desc_ = create_matrix_layout(
       dtype_to_cublas_type(dtype),
-      rows,
       cols,
+      rows,
       transposed,
       ld,
       batch_count,
       batch_stride);
 }
 
+void CublasGemm::set_bias(cu::CommandEncoder& encoder, const array& bias) {
+  encoder.set_input_array(bias);
+  cublasLtEpilogue_t epilogue = CUBLASLT_EPILOGUE_BIAS;
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_EPILOGUE,
+      &epilogue,
+      sizeof(epilogue)));
+  auto* bias_ptr = bias.data<void>();
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_BIAS_POINTER,
+      &bias_ptr,
+      sizeof(bias_ptr)));
+}
+
 void CublasGemm::run(
     cu::CommandEncoder& encoder,
     array& out,
@@ -228,11 +256,19 @@ void CublasGemm::run(
     const array& b,
     const Shape& batch_shape,
     const Strides& a_batch_strides,
-    const Strides& b_batch_strides) {
+    const Strides& b_batch_strides,
+    float alpha) {
   int batch_count = out.size() / (M_ * N_);
   if (batch_count / batch_shape.back() > 1) {
     run_batched(
-        encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides);
+        encoder,
+        out,
+        a,
+        b,
+        batch_shape,
+        a_batch_strides,
+        b_batch_strides,
+        alpha);
     return;
   }
 
@@ -240,7 +276,13 @@ void CublasGemm::run(
   encoder.set_input_array(b);
   encoder.set_output_array(out);
 
-  execute(encoder, out.data<void>(), a.data<void>(), b.data<void>(), nullptr);
+  execute(
+      encoder,
+      out.data<void>(),
+      a.data<void>(),
+      b.data<void>(),
+      nullptr,
+      alpha);
 }
 
 void CublasGemm::run(
@@ -313,6 +355,16 @@ void CublasGemm::execute(
     }
   }
 
+  const void* alpha_ptr = &alpha;
+  const void* beta_ptr = &beta;
+  complex64_t alpha_c, beta_c;
+  if (scale_type_ == CUDA_C_32F) {
+    alpha_c = complex64_t{alpha, 0.0f};
+    beta_c = complex64_t{beta, 0.0f};
+    alpha_ptr = &alpha_c;
+    beta_ptr = &beta_c;
+  }
+
   void* workspace_ptr = nullptr;
   if (heuristic_.workspaceSize > 0) {
     // Ensure workspace is 256-byte aligned
@@ -329,12 +381,12 @@ void CublasGemm::execute(
   CHECK_CUBLAS_ERROR(cublasLtMatmul(
       handle_,
       matmul_desc_,
-      &alpha,
-      a,
+      alpha_ptr,
+      b, // a and b are swapped
       a_desc_,
-      b,
+      a,
       b_desc_,
-      &beta,
+      beta_ptr,
       c ? c : out,
       c ? c_desc_ : out_desc_,
       out,

mlx/backend/cuda/gemms/cublas_gemm.h

@@ -55,6 +55,8 @@ class CublasGemm {
       int32_t batch_count,
       int64_t batch_stride);
 
+  void set_bias(cu::CommandEncoder& encoder, const array& bias);
+
   void run(
       cu::CommandEncoder& encoder,
       array& out,
@@ -62,7 +64,8 @@ class CublasGemm {
       const array& b,
       const Shape& batch_shape,
       const Strides& a_batch_strides,
-      const Strides& b_batch_strides);
+      const Strides& b_batch_strides,
+      float alpha = 1.0f);
 
   void run(
       cu::CommandEncoder& encoder,
@@ -85,7 +88,8 @@ class CublasGemm {
       const array& b,
       const Shape& batch_shape,
       const Strides& a_batch_strides,
-      const Strides& b_batch_strides);
+      const Strides& b_batch_strides,
+      float alpha);
 
   void run_batched(
       cu::CommandEncoder& encoder,
@@ -111,6 +115,7 @@ class CublasGemm {
 
   uint64_t M_;
   uint64_t N_;
+  cudaDataType_t scale_type_;
   cublasLtMatmulPreference_t pref_{nullptr};
   cublasLtHandle_t handle_{nullptr};
   cublasLtMatmulDesc_t matmul_desc_{nullptr};

mlx/backend/cuda/gemms/cublas_gemm_batched_12_0.cpp

@@ -13,7 +13,8 @@ void CublasGemm::run_batched(
     const array& b,
     const Shape& batch_shape,
     const Strides& a_batch_strides,
-    const Strides& b_batch_strides) {
+    const Strides& b_batch_strides,
+    float alpha) {
   encoder.set_input_array(a);
   encoder.set_input_array(b);
   encoder.set_output_array(out);
@@ -27,7 +28,8 @@ void CublasGemm::run_batched(
         out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M_ * N_,
         a.data<int8_t>() + a.itemsize() * a_it.loc,
         b.data<int8_t>() + b.itemsize() * b_it.loc,
-        nullptr);
+        nullptr,
+        alpha);
     a_it.step();
     b_it.step();
   }

mlx/backend/cuda/gemms/cublas_gemm_batched_12_9.cu

@@ -154,7 +154,8 @@ void CublasGemm::run_batched(
     const array& b,
     const Shape& batch_shape,
     const Strides& a_batch_strides,
-    const Strides& b_batch_strides) {
+    const Strides& b_batch_strides,
+    float alpha) {
   int batch_count = out.size() / (M_ * N_);
   set_pointer_mode(a_desc_, batch_count);
   set_pointer_mode(b_desc_, batch_count);
@@ -226,7 +227,8 @@ void CublasGemm::run_batched(
       reinterpret_cast<void*>(out_pointers),
       reinterpret_cast<void*>(a_pointers),
       reinterpret_cast<void*>(b_pointers),
-      nullptr);
+      nullptr,
+      alpha);
 }
 
 void CublasGemm::run_batched(

mlx/backend/cuda/gemms/gemv.cu

@@ -13,6 +13,37 @@ namespace cg = cooperative_groups;
 
 static constexpr int rows_per_block = 8;
 
+// Accumulator type selection per input element type T.
+template <typename T>
+struct GemvAccType {
+  using type = T;
+};
+
+template <>
+struct GemvAccType<__half> {
+  using type = float;
+};
+
+template <>
+struct GemvAccType<__nv_bfloat16> {
+  using type = float;
+};
+
+template <>
+struct GemvAccType<float> {
+  using type = float;
+};
+
+template <>
+struct GemvAccType<double> {
+  using type = double;
+};
+
+template <>
+struct GemvAccType<cu::complex64_t> {
+  using type = cu::complex64_t;
+};
+
 template <typename T, int rows_per_block, int n_per_thread>
 __device__ void
 gemv_impl(const T* mat, const T* vec, T* out, int rows, int cols) {
@@ -24,7 +55,8 @@ gemv_impl(const T* mat, const T* vec, T* out, int rows, int cols) {
   int row = g_idx.x * rows_per_block + t_idx.y;
 
   if (row < rows) {
-    float sum = 0.0f;
+    using Acc = typename GemvAccType<T>::type;
+    Acc sum = Acc(0);
     for (int col = n_per_thread * warp.thread_rank(); col < cols;
          col += (WARP_SIZE * n_per_thread)) {
       auto local_mat =
@@ -32,12 +64,11 @@ gemv_impl(const T* mat, const T* vec, T* out, int rows, int cols) {
       auto local_vec = unsafe_load_vector<n_per_thread>(vec + col, 0);
 #pragma unroll
       for (int j = 0; j < n_per_thread; ++j) {
-        sum +=
-            static_cast<float>(local_mat[j]) * static_cast<float>(local_vec[j]);
+        sum += static_cast<Acc>(local_mat[j]) * static_cast<Acc>(local_vec[j]);
       }
     }
 
-    sum = cg::reduce(warp, sum, cg::plus<float>{});
+    sum = cg::reduce(warp, sum, cg::plus<Acc>{});
     if (warp.thread_rank() == 0) {
       out[row] = static_cast<T>(sum);
     }
@@ -107,7 +138,7 @@ void gemv(
   encoder.set_input_array(a);
   encoder.set_input_array(b);
   encoder.set_output_array(out);
-  dispatch_float_types(out.dtype(), "gemv", [&](auto type_tag) {
+  dispatch_inexact_types(out.dtype(), "gemv", [&](auto type_tag) {
     using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
     dim3 block_dims{WARP_SIZE, rows_per_block};
     const DataType* mat;

mlx/backend/cuda/indexing.cpp

@@ -110,7 +110,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   args.append<int32_t>(src.ndim());
   args.append_ndim(slice_sizes_);
   args.append(slice_size);
-  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
+  args.append(axes_);
   append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
@@ -211,7 +211,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   args.append_ndim(out.shape());
   args.append_ndim(out.strides());
   args.append<int32_t>(out.ndim());
-  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
+  args.append(axes_);
   append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(

mlx/backend/cuda/jit_module.cpp

@@ -67,9 +67,11 @@ const std::string& cccl_dir() {
       return path.string();
     }
     // Finally check the environment variable.
-    path = std::getenv("MLX_CCCL_DIR");
-    if (!path.empty() && std::filesystem::exists(path)) {
-      return path.string();
+    if (const char* env = std::getenv("MLX_CCCL_DIR"); env) {
+      path = env;
+      if (!path.empty() && std::filesystem::exists(path)) {
+        return path.string();
+      }
     }
     return std::string();
   }();
@@ -97,6 +99,30 @@ const std::filesystem::path& ptx_cache_dir() {
   return cache;
 }
 
+std::filesystem::path get_ptx_path(
+    const std::filesystem::path& cache_dir,
+    const std::string& module_name) {
+#ifdef _WIN32
+  constexpr int max_file_name_length = 140;
+#else
+  constexpr int max_file_name_length = 245;
+#endif
+
+  if (module_name.size() <= max_file_name_length) {
+    return cache_dir / (module_name + ".ptx");
+  }
+
+  auto ptx_path = cache_dir;
+  int offset = 0;
+  while (module_name.size() - offset > max_file_name_length) {
+    ptx_path /= module_name.substr(offset, max_file_name_length);
+    offset += max_file_name_length;
+  }
+  ptx_path /= module_name.substr(offset) + ".ptx";
+
+  return ptx_path;
+}
+
 // Try to read the cached |ptx| and |ptx_kernels| from |cache_dir|.
 bool read_cached_ptx(
     const std::filesystem::path& cache_dir,
@@ -107,7 +133,7 @@ bool read_cached_ptx(
     return false;
   }
 
-  auto ptx_path = cache_dir / (module_name + ".ptx");
+  auto ptx_path = get_ptx_path(cache_dir, module_name);
   std::error_code error;
   auto ptx_size = std::filesystem::file_size(ptx_path, error);
   if (error) {
@@ -120,7 +146,7 @@ bool read_cached_ptx(
   ptx.resize(ptx_size);
   ptx_file.read(ptx.data(), ptx_size);
 
-  std::ifstream txt_file(cache_dir / (module_name + ".txt"), std::ios::binary);
+  std::ifstream txt_file(ptx_path.replace_extension(".txt"), std::ios::binary);
   std::string line;
   while (std::getline(txt_file, line)) {
     auto tab = line.find('\t');
@@ -142,16 +168,26 @@ void write_cached_ptx(
     return;
   }
 
-  std::ofstream ptx_file(cache_dir / (module_name + ".ptx"), std::ios::binary);
+  auto ptx_path = get_ptx_path(cache_dir, module_name);
+
+  // Ensure that the directory exists
+  auto parent = ptx_path.parent_path();
+  if (parent != cache_dir) {
+    std::filesystem::create_directories(parent);
+  }
+
+  // Write the compiled code and mangled names
+  std::ofstream ptx_file(ptx_path, std::ios::binary);
   if (!ptx.empty()) {
     ptx_file.write(&ptx.front(), ptx.size());
   }
-  std::ofstream txt_file(cache_dir / (module_name + ".txt"), std::ios::binary);
+  std::ofstream txt_file(ptx_path.replace_extension(".txt"), std::ios::binary);
   for (const auto& [name, mangled] : ptx_kernels) {
     txt_file << name << "\t" << mangled << std::endl;
   }
 
-  std::ofstream source_file(cache_dir / (module_name + ".cu"));
+  // Write the generated code
+  std::ofstream source_file(ptx_path.replace_extension(".cu"));
   source_file << source_code;
 }
 
@@ -295,7 +331,8 @@ void load_module(
     const std::string& ptx,
     const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
     CUmodule& module_,
-    std::unordered_map<std::string, std::pair<CUfunction, bool>>& kernels) {
+    std::unordered_map<std::string, std::tuple<CUfunction, bool, uint>>&
+        kernels) {
   // Load module.
   char jit_log[4089] = {};
   CUjit_option options[] = {
@@ -312,7 +349,7 @@ void load_module(
   for (const auto& [name, mangled] : ptx_kernels) {
     CUfunction kernel;
     CHECK_CUDA_ERROR(cuModuleGetFunction(&kernel, module_, mangled.c_str()));
-    kernels[name] = std::make_pair(kernel, false);
+    kernels[name] = std::make_tuple(kernel, false, 0);
   }
 }
 
@@ -356,7 +393,7 @@ JitModule::~JitModule() {
   CHECK_CUDA_ERROR(cuModuleUnload(module_));
 }
 
-CUfunction JitModule::get_kernel(
+std::pair<CUfunction, uint> JitModule::get_kernel_and_dims(
     const std::string& kernel_name,
     std::function<void(CUfunction)> configure_kernel) {
   auto it = kernels_.find(kernel_name);
@@ -367,14 +404,22 @@ CUfunction JitModule::get_kernel(
 
   // If it is the first time we run this kernel then configure it. Do it only
   // once!
-  if (!it->second.second) {
+  auto kernel = std::get<0>(it->second);
+  if (!std::get<1>(it->second)) {
     if (configure_kernel) {
-      configure_kernel(it->second.first);
+      configure_kernel(kernel);
     }
-    it->second.second = true;
+    std::get<1>(it->second) = true;
+    std::get<2>(it->second) = max_occupancy_block_dim(kernel);
   }
 
-  return it->second.first;
+  return {kernel, std::get<2>(it->second)};
+}
+
+CUfunction JitModule::get_kernel(
+    const std::string& kernel_name,
+    std::function<void(CUfunction)> configure_kernel) {
+  return get_kernel_and_dims(kernel_name, std::move(configure_kernel)).first;
 }
 
 std::unordered_map<std::string, JitModule>& get_jit_module_cache() {

mlx/backend/cuda/jit_module.h

@@ -46,6 +46,11 @@ struct KernelArgs {
     append_ptr(std::get<SmallVector<T>>(storage_.back()).data());
   }
 
+  template <typename T>
+  void append(const std::vector<T>& vec) {
+    append(SmallVector<T>(vec.begin(), vec.end()));
+  }
+
   // Make sure the arg is copied to an array with size of NDIM.
   template <size_t NDIM = MAX_NDIM, typename T>
   void append_ndim(SmallVector<T> vec) {
@@ -94,10 +99,13 @@ class JitModule {
   CUfunction get_kernel(
       const std::string& kernel_name,
       std::function<void(CUfunction)> configure_kernel = nullptr);
+  std::pair<CUfunction, uint> get_kernel_and_dims(
+      const std::string& kernel_name,
+      std::function<void(CUfunction)> configure_kernel = nullptr);
 
  private:
   CUmodule module_{nullptr};
-  std::unordered_map<std::string, std::pair<CUfunction, bool>> kernels_;
+  std::unordered_map<std::string, std::tuple<CUfunction, bool, uint>> kernels_;
 };
 
 std::unordered_map<std::string, JitModule>& get_jit_module_cache();

mlx/backend/cuda/kernel_utils.cu

@@ -35,12 +35,10 @@ std::tuple<dim3, uint> get_launch_args(
     const Shape& shape,
     const Strides& strides,
     bool large,
-    int work_per_thread) {
+    int work_per_thread /* = 1 */,
+    uint max_block_dim /* = 1024 */) {
   size_t nthreads = cuda::ceil_div(size, work_per_thread);
-  uint block_dim = 1024;
-  if (block_dim > nthreads) {
-    block_dim = nthreads;
-  }
+  uint block_dim = max_block_dim < nthreads ? max_block_dim : nthreads;
   dim3 num_blocks;
   if (large) {
     num_blocks = get_2d_grid_dims(shape, strides, work_per_thread);

mlx/backend/cuda/kernel_utils.cuh

@@ -1,8 +1,8 @@
 // Copyright © 2025 Apple Inc.
 
-// This file includes host-only utilies for writing CUDA kernels, the difference
-// from backend/cuda/device/utils.cuh is that the latter file only include
-// device-only code.
+// This file includes host-only utilities for writing CUDA kernels, the
+// difference from backend/cuda/device/utils.cuh is that the latter file only
+// include device-only code.
 
 #pragma once
 
@@ -120,19 +120,28 @@ dim3 get_2d_grid_dims(
     size_t divisor);
 std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2);
 
-// Get the num_blocks and block_dims that maximize occupancy for |kernel|,
-// assuming each thread handles |work_per_thread| elements of |arr|.
+// Get the num_blocks and block_dims assuming each thread handles
+// |work_per_thread| elements of |arr|.
 std::tuple<dim3, uint> get_launch_args(
     size_t size,
     const Shape& shape,
     const Strides& strides,
     bool large,
-    int work_per_thread = 1);
+    int work_per_thread = 1,
+    uint max_block_dim = 1024);
 
-inline std::tuple<dim3, uint>
-get_launch_args(const array& arr, bool large, int work_per_thread = 1) {
+inline std::tuple<dim3, uint> get_launch_args(
+    const array& arr,
+    bool large,
+    int work_per_thread = 1,
+    uint max_block_dim = 1024) {
   return get_launch_args(
-      arr.size(), arr.shape(), arr.strides(), large, work_per_thread);
+      arr.size(),
+      arr.shape(),
+      arr.strides(),
+      large,
+      work_per_thread,
+      max_block_dim);
 }
 
 } // namespace mlx::core

mlx/backend/cuda/lru_cache.h

@@ -2,11 +2,15 @@
 
 #pragma once
 
+#include "mlx/utils.h"
+
 #include <cstring>
 #include <list>
 #include <unordered_map>
 #include <utility>
 
+#include <fmt/format.h>
+
 namespace mlx::core {
 
 template <
@@ -27,6 +31,14 @@ class LRUCache {
     }
   }
 
+  // Initialize with capacity read from |env_name|.
+  LRUCache(const char* env_name, int default_capacity)
+      : LRUCache(env::get_var(env_name, default_capacity)) {
+    if (env::get_var("MLX_ENABLE_CACHE_THRASHING_CHECK", 1)) {
+      env_name_ = env_name;
+    }
+  }
+
   size_t size() const {
     return map_.size();
   }
@@ -76,6 +88,14 @@ class LRUCache {
       return {it->second, false};
     }
 
+    if (env_name_ && ++cache_misses_ > 2 * capacity_) {
+      throw std::runtime_error(fmt::format(
+          "Cache thrashing is happening, please set the environment variable "
+          "{} to a larger value than {} to fix degraded performance.",
+          env_name_,
+          capacity_));
+    }
+
     vlist_.emplace_front(key, std::forward<U>(value));
     map_[key] = vlist_.begin();
 
@@ -106,6 +126,9 @@ class LRUCache {
     }
   }
 
+  const char* env_name_{nullptr};
+  size_t cache_misses_{0};
+
   list_type vlist_;
   map_type map_;
   size_t capacity_;

mlx/backend/cuda/matmul.cpp

@@ -11,6 +11,7 @@
 #include <numeric>
 
 namespace mlx::core {
+
 namespace {
 
 std::tuple<bool, int64_t, array>
@@ -28,6 +29,80 @@ check_transpose(cu::CommandEncoder& enc, const Stream& s, const array& arr) {
   }
 }
 
+void gemm_and_bias(
+    cu::CommandEncoder& encoder,
+    int M,
+    int N,
+    int K,
+    bool a_transposed,
+    int64_t lda,
+    bool b_transposed,
+    int64_t ldb,
+    array& out,
+    const array& a,
+    const array& b,
+    const std::optional<array>& bias = std::nullopt,
+    float alpha = 1.0f) {
+  // Check and collapse batch dimensions
+  auto [batch_shape, a_batch_strides, b_batch_strides] = collapse_batches(a, b);
+
+  auto batch_count = out.size() / (M * N);
+
+  // Collapse batches into M if needed
+  if (batch_count > 1 && !a_transposed && batch_shape.size() == 1 &&
+      a.strides()[a.ndim() - 2] == K && a_batch_strides.back() == M * K &&
+      b_batch_strides.back() == 0) {
+    M *= batch_shape.back();
+    batch_count = 1;
+
+    a_batch_strides = {0};
+    b_batch_strides = {0};
+    batch_shape = {1};
+  }
+
+  // Use gemmv when possible
+  if (!bias && cu::can_use_gemv(M, N, K, a_transposed, b_transposed)) {
+    cu::gemv(
+        a,
+        b,
+        out,
+        M,
+        N,
+        K,
+        batch_count,
+        batch_shape,
+        a_batch_strides,
+        b_batch_strides,
+        encoder);
+    return;
+  }
+
+  // Invoke cublasLt
+  CublasGemm gemm(
+      encoder.device(),
+      a.dtype(),
+      a_transposed,
+      M,
+      K,
+      lda,
+      b_transposed,
+      K,
+      N,
+      ldb,
+      batch_shape.back(),
+      a_batch_strides.back(),
+      b_batch_strides.back());
+  if (bias) {
+    if (a.dtype() == complex64) {
+      throw std::runtime_error(
+          "[gemm_and_bias] complex64 bias epilogue isn’t supported in cublasLtMatmul.");
+    }
+    gemm.set_bias(encoder, *bias);
+  }
+  gemm.run(
+      encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides, alpha);
+}
+
 } // namespace
 
 void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -48,9 +123,6 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  /////////////////////////////////////////////////////////////////////////////
-  // Init checks and prep
-
   int M = a_pre.shape(-2);
   int N = b_pre.shape(-1);
   int K = a_pre.shape(-1);
@@ -60,58 +132,8 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
   auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
 
-  /////////////////////////////////////////////////////////////////////////////
-  // Check and collapse batch dimensions
-
-  auto [batch_shape, a_batch_strides, b_batch_strides] = collapse_batches(a, b);
-
-  auto batch_count = out.size() / (M * N);
-
-  // Collapse batches into M if needed
-  if (batch_count > 1 && !a_transposed && batch_shape.size() == 1 &&
-      a.strides()[a.ndim() - 2] == K && a_batch_strides.back() == M * K &&
-      b_batch_strides.back() == 0) {
-    M *= batch_shape.back();
-    batch_count = 1;
-
-    a_batch_strides = {0};
-    b_batch_strides = {0};
-    batch_shape = {1};
-  }
-
-  if (cu::can_use_gemv(M, N, K, a_transposed, b_transposed)) {
-    cu::gemv(
-        a,
-        b,
-        out,
-        M,
-        N,
-        K,
-        batch_count,
-        batch_shape,
-        a_batch_strides,
-        b_batch_strides,
-        encoder);
-    return;
-  }
-
-  /////////////////////////////////////////////////////////////////////////////
-  // Invoke cublasLt
-  CublasGemm gemm(
-      cu::device(s.device),
-      a.dtype(),
-      a_transposed,
-      M,
-      K,
-      lda,
-      b_transposed,
-      K,
-      N,
-      ldb,
-      batch_shape.back(),
-      a_batch_strides.back(),
-      b_batch_strides.back());
-  gemm.run(encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides);
+  gemm_and_bias(
+      encoder, M, N, K, a_transposed, lda, b_transposed, ldb, out, a, b);
 }
 
 void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -136,6 +158,29 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
   auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
 
+  /////////////////////////////////////////////////////////////////////////////
+  // Dispatch to GEMM with epilogue or AddMM
+
+  if (beta_ == 1 && a.dtype() != complex64 && c.strides(-1) == 1 &&
+      c.data_size() == out.shape(-1)) {
+    out.set_data(allocator::malloc(out.nbytes()));
+    gemm_and_bias(
+        encoder,
+        M,
+        N,
+        K,
+        a_transposed,
+        lda,
+        b_transposed,
+        ldb,
+        out,
+        a,
+        b,
+        c,
+        alpha_);
+    return;
+  }
+
   int64_t ldc;
   {
     auto stx = c.strides()[c.ndim() - 2];
@@ -177,7 +222,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
 
   /////////////////////////////////////////////////////////////////////////////
-  // Invoke cublasLt
+  // Invoke cublasLt with AddMM settings
 
   CublasGemm gemm(
       cu::device(s.device),

mlx/backend/cuda/primitives.cpp

@@ -24,8 +24,6 @@ namespace mlx::core {
   }
 
 NO_GPU(BlockMaskedMM)
-NO_GPU(DynamicSlice)
-NO_GPU(DynamicSliceUpdate)
 NO_GPU(FFT)
 NO_GPU(GatherMM)
 NO_GPU(GatherQMM)
@@ -42,7 +40,6 @@ NO_GPU_MULTI(Eig)
 NO_GPU_MULTI(Eigh)
 
 namespace distributed {
-NO_GPU_MULTI(AllReduce)
 NO_GPU_MULTI(AllGather)
 NO_GPU_MULTI(Send)
 NO_GPU_MULTI(Recv)

mlx/backend/cuda/quantized/quantized.cpp

@@ -46,10 +46,10 @@ inline array ensure_row_contiguous_matrix(
 
 } // namespace
 
-void fast::AffineQuantize::eval_gpu(
+void fast::Quantize::eval_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
-  nvtx3::scoped_range r("AffineQuantize::eval_gpu");
+  nvtx3::scoped_range r("Quantize::eval_gpu");
   auto& s = stream();
   auto& d = cu::device(s.device);
   auto& enc = d.get_command_encoder(s);

mlx/backend/cuda/reduce/col_reduce.cu

@@ -181,6 +181,47 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
   }
 }
 
+template <typename T, typename U, typename Op, int N_READS = 4>
+__global__ void col_reduce_small(
+    const T* in,
+    U* out,
+    const __grid_constant__ ColReduceArgs args,
+    size_t total) {
+  Op op;
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+
+  const auto idx = grid.thread_rank() * N_READS;
+  const auto before_axis = idx / args.reduction_stride;
+  const auto after_axis = idx % args.reduction_stride;
+  const auto offset =
+      before_axis * args.reduction_stride * args.reduction_size + after_axis;
+
+  if (idx >= total) {
+    return;
+  }
+
+  in += offset;
+  out += idx;
+
+  AlignedVector<U, N_READS> accumulator;
+  for (int i = 0; i < N_READS; i++) {
+    accumulator[i] = ReduceInit<Op, T>::value();
+  }
+
+  for (int i = 0; i < args.reduction_size; i++) {
+    auto values = load_vector<N_READS>(in, 0);
+
+    for (int j = 0; j < N_READS; j++) {
+      accumulator[j] = op(accumulator[j], cast_to<U>(values[j]));
+    }
+
+    in += args.reduction_stride;
+  }
+
+  store_vector(out, 0, accumulator);
+}
+
 } // namespace cu
 
 inline auto output_grid_for_col_reduce(
@@ -206,7 +247,7 @@ void col_reduce_looped(
     Reduce::ReduceType reduce_type,
     const std::vector<int>& axes,
     const ReductionPlan& plan,
-    cu::ColReduceArgs args) {
+    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);
@@ -230,12 +271,55 @@ void col_reduce_looped(
         auto kernel =
             cu::col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS>;
         encoder.add_kernel_node(
-            kernel, grid, blocks, 0, indata, out.data<U>(), args);
+            kernel,
+            grid,
+            blocks,
+            0,
+            indata,
+            out.data<U>(),
+            static_cast<cu::ColReduceArgs>(args));
       });
     });
   });
 }
 
+void col_reduce_small(
+    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.set_input_array(in);
+  encoder.set_output_array(out);
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+      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 = 16 / sizeof(T);
+      auto tmp_grid = get_2d_grid_dims(out.shape(), out.strides());
+      auto [grid, block] = get_grid_and_block(tmp_grid.x, tmp_grid.y, 1);
+      auto kernel = cu::col_reduce_small<T, U, OP, N_READS>;
+      encoder.add_kernel_node(
+          kernel,
+          grid,
+          block,
+          0,
+          in.data<T>(),
+          out.data<U>(),
+          static_cast<cu::ColReduceArgs>(args),
+          out.size());
+    });
+  });
+}
+
 void col_reduce(
     cu::CommandEncoder& encoder,
     const array& in,
@@ -258,6 +342,13 @@ void col_reduce(
   // Make the args struct to help route to the best kernel
   cu::ColReduceArgs args(in, plan, axes);
 
+  // Small col reduce with a single or contiguous reduction axis
+  if (args.non_col_reductions == 1 && args.reduction_size <= 32 &&
+      args.reduction_stride % (16 / in.itemsize()) == 0) {
+    col_reduce_small(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/row_reduce.cu

@@ -7,8 +7,6 @@
 
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
-#include <cub/block/block_load.cuh>
-#include <cub/block/block_reduce.cuh>
 
 namespace mlx::core {
 
@@ -83,7 +81,8 @@ struct RowReduceArgs {
 };
 
 template <typename T, typename U, typename ReduceOp, int N = 4, int M = 1>
-__global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
+__global__ void
+row_reduce_simple(const T* in, U* out, size_t n_rows, int size) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
   auto warp = cg::tiled_partition<WARP_SIZE>(block);
@@ -91,8 +90,8 @@ __global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
   const U init = cu::ReduceInit<ReduceOp, T>::value();
   ReduceOp op;
 
-  T vals[M][N];
-  U accs[M];
+  AlignedVector<T, N> vals[M];
+  AlignedVector<U, M> accs;
   for (int i = 0; i < M; i++) {
     accs[i] = init;
   }
@@ -101,43 +100,31 @@ __global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
       min(n_rows - M, static_cast<size_t>(grid.block_rank() * M));
   const size_t full_blocks = size / (block.size() * N);
   const size_t final_offset = full_blocks * (block.size() * N);
-  in += start_row * size;
+  in += start_row * size + block.thread_rank() * N;
   out += start_row;
 
-  if (size % N == 0) {
-    for (size_t r = 0; r < full_blocks; r++) {
-      for (int k = 0; k < M; k++) {
-        cub::LoadDirectBlockedVectorized<T, N>(
-            block.thread_rank(),
-            in + k * size + r * (block.size() * N),
-            vals[k]);
-        for (int j = 0; j < N; j++) {
-          accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
-        }
-      }
-    }
-  } else {
-    for (size_t r = 0; r < full_blocks; r++) {
-      for (int k = 0; k < M; k++) {
-        cub::LoadDirectBlocked(
-            block.thread_rank(),
-            in + k * size + r * (block.size() * N),
-            vals[k]);
-        for (int j = 0; j < N; j++) {
-          accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
-        }
+  for (size_t r = 0; r < full_blocks; r++) {
+    for (int k = 0; k < M; k++) {
+      vals[k] = load_vector<N>(in + k * size, 0);
+    }
+    for (int k = 0; k < M; k++) {
+      for (int j = 0; j < N; j++) {
+        accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
       }
     }
+
+    in += block.size() * N;
   }
 
   if (final_offset < size) {
     for (int k = 0; k < M; k++) {
-      cub::LoadDirectBlocked(
-          block.thread_rank(),
-          in + k * size + final_offset,
-          vals[k],
-          size,
-          cast_to<T>(init));
+      for (int i = 0; i < N; i++) {
+        vals[k][i] = ((final_offset + block.thread_rank() * N + i) < size)
+            ? in[k * size + i]
+            : cast_to<T>(init);
+      }
+    }
+    for (int k = 0; k < M; k++) {
       for (int j = 0; j < N; j++) {
         accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
       }
@@ -145,13 +132,11 @@ __global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
   }
 
   __shared__ U shared_accumulators[32 * M];
-  block_reduce(block, warp, accs, shared_accumulators, op, init);
+  block_reduce(block, warp, accs.val, shared_accumulators, op, init);
 
   if (block.thread_rank() == 0) {
     if (grid.block_rank() * M + M <= n_rows) {
-      for (int i = 0; i < M; i++) {
-        out[i] = accs[i];
-      }
+      store_vector(out, 0, accs);
     } else {
       short offset = grid.block_rank() * M + M - n_rows;
       for (int i = offset; i < M; i++) {
@@ -161,17 +146,10 @@ __global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
   }
 }
 
-template <
-    typename T,
-    typename U,
-    typename Op,
-    int NDIM,
-    int BLOCK_DIM,
-    int N_READS = 4>
+template <typename T, typename U, typename Op, int NDIM, int N_READS = 4>
 __global__ void row_reduce_looped(
-    T* in,
+    const T* in,
     U* out,
-    size_t out_size,
     const __grid_constant__ RowReduceArgs args) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
@@ -185,36 +163,60 @@ __global__ void row_reduce_looped(
   U init = ReduceInit<Op, T>::value();
   total[0] = init;
   LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-  size_t full_blocks = args.row_size / (BLOCK_DIM * N_READS);
-  size_t final_offset = full_blocks * BLOCK_DIM * N_READS;
+  const size_t full_blocks = args.row_size / (block.size() * N_READS);
+  const size_t final_offset = full_blocks * (block.size() * N_READS);
 
   in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
+  in += block.thread_rank() * N_READS;
 
-  for (size_t n = 0; n < args.non_row_reductions; n++) {
-    for (size_t r = 0; r < full_blocks; r++) {
-      T vals[N_READS];
-      cub::LoadDirectBlockedVectorized<T, N_READS>(
-          block.thread_rank(),
-          in + loop.location() + r * BLOCK_DIM * N_READS,
-          vals);
-      for (int i = 0; i < N_READS; i++) {
-        total[0] = op(total[0], cast_to<U>(vals[i]));
-      }
+  // Unaligned reduce
+  if (final_offset < args.row_size) {
+    bool mask[N_READS];
+    for (int i = 0; i < N_READS; i++) {
+      mask[i] =
+          (final_offset + block.thread_rank() * N_READS + i) < args.row_size;
     }
-    if (final_offset < args.row_size) {
-      T vals[N_READS];
-      cub::LoadDirectBlocked(
-          block.thread_rank(),
-          in + loop.location() + final_offset,
-          vals,
-          args.row_size - final_offset,
-          cast_to<T>(init));
-      for (int i = 0; i < N_READS; i++) {
-        total[0] = op(total[0], cast_to<U>(vals[i]));
+
+    for (size_t n = 0; n < args.non_row_reductions; n++) {
+      const T* inlocal = in + loop.location();
+
+      for (size_t r = 0; r < full_blocks; r++) {
+        auto vals = load_vector<N_READS>(inlocal, 0);
+        for (int i = 0; i < N_READS; i++) {
+          total[0] = op(total[0], cast_to<U>(vals[i]));
+        }
+        inlocal += block.size() * N_READS;
       }
+
+      {
+        T vals[N_READS];
+        for (int i = 0; i < N_READS; i++) {
+          vals[i] = mask[i] ? inlocal[i] : cast_to<T>(init);
+        }
+        for (int i = 0; i < N_READS; i++) {
+          total[0] = op(total[0], cast_to<U>(vals[i]));
+        }
+      }
+
+      loop.next(args.reduce_shape.data(), args.reduce_strides.data());
+    }
+  }
+
+  // Aligned case
+  else {
+    for (size_t n = 0; n < args.non_row_reductions; n++) {
+      const T* inlocal = in + loop.location();
+
+      for (size_t r = 0; r < full_blocks; r++) {
+        auto vals = load_vector<N_READS>(inlocal, 0);
+        for (int i = 0; i < N_READS; i++) {
+          total[0] = op(total[0], cast_to<U>(vals[i]));
+        }
+        inlocal += block.size() * N_READS;
+      }
+
+      loop.next(args.reduce_shape.data(), args.reduce_strides.data());
     }
-    // TODO: Maybe block.sync() here?
-    loop.next(args.reduce_shape.data(), args.reduce_strides.data());
   }
 
   __shared__ U shared_accumulators[32];
@@ -234,8 +236,6 @@ void row_reduce_simple(
     Reduce::ReduceType reduce_type,
     const std::vector<int>& axes,
     const ReductionPlan& plan) {
-  constexpr int N_READS = 8;
-
   // Allocate data for the output using in's layout to avoid elem_to_loc in the
   // kernel.
   allocate_same_layout(out, in, axes);
@@ -250,14 +250,15 @@ void row_reduce_simple(
       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*>(in.data<T>());
+      constexpr int N_READS = 16 / sizeof(T);
 
       // Calculate the grid and block dims
       size_t reductions = (plan.shape.back() + N_READS - 1) / N_READS;
       dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
-      int threads = std::min(1024UL, reductions);
-      threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
+      int warps = (reductions + WARP_SIZE - 1) / WARP_SIZE;
+      warps /= 4;
+      warps = std::max(std::min(warps, 32), 1);
+      int threads = warps * WARP_SIZE;
       dim3 block(threads, 1, 1);
 
       // Pick the kernel
@@ -267,6 +268,7 @@ void row_reduce_simple(
         kernel = cu::row_reduce_simple<T, U, OP, N_READS, 2>;
       }
 
+      T* indata = const_cast<T*>(in.data<T>());
       int size = plan.shape.back();
       encoder.add_kernel_node(
           kernel, grid, block, 0, indata, out.data<U>(), out.size(), size);
@@ -282,8 +284,6 @@ void row_reduce_looped(
     const std::vector<int>& axes,
     const ReductionPlan& plan,
     cu::RowReduceArgs args) {
-  constexpr int N_READS = 8;
-
   // Allocate data for the output using in's layout to access them as
   // contiguously as possible.
   allocate_same_layout(out, in, axes);
@@ -295,34 +295,27 @@ void row_reduce_looped(
       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*>(in.data<T>());
+
+      constexpr int N_READS = 16 / sizeof(T);
 
       // Calculate the grid and block dims
       args.sort_access_pattern(in, axes);
       dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
       size_t reductions = (args.row_size + N_READS - 1) / N_READS;
-      int threads = std::min(1024UL, reductions);
-      threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
+      int warps = (reductions + WARP_SIZE - 1) / WARP_SIZE;
+      warps /= 4;
+      warps = std::max(std::min(warps, 32), 1);
+      int threads = warps * WARP_SIZE;
       dim3 block(threads, 1, 1);
 
       // Pick the kernel
-      auto kernel = cu::row_reduce_looped<T, U, OP, 1, 32, N_READS>;
+      auto kernel = cu::row_reduce_looped<T, U, OP, 1, N_READS>;
       dispatch_reduce_ndim(args.reduce_ndim, [&](auto reduce_ndim) {
-        dispatch_block_dim(threads, [&](auto threads_constant) {
-          kernel = cu::row_reduce_looped<
-              T,
-              U,
-              OP,
-              reduce_ndim.value,
-              threads_constant.value,
-              N_READS>;
-          block.x = threads_constant.value;
-        });
+        kernel = cu::row_reduce_looped<T, U, OP, reduce_ndim.value, N_READS>;
       });
 
       encoder.add_kernel_node(
-          kernel, grid, block, 0, indata, out.data<U>(), out.size(), args);
+          kernel, grid, block, 0, in.data<T>(), out.data<U>(), args);
     });
   });
 }

mlx/backend/cuda/rope.cu

@@ -103,15 +103,21 @@ template <typename T, bool traditional, bool forward, int N = 4>
 __device__ void rope_impl(
     const T* in,
     T* out,
-    int offset,
+    const int* offset,
     float inv_freq,
     float scale,
     const cuda::std::array<int64_t, 3> strides,
     const cuda::std::array<int64_t, 3> out_strides,
-    int64_t n_batch,
+    int64_t offset_stride,
+    int n_head,
     uint3 pos,
     uint3 dims) {
-  float L = scale * static_cast<float>(pos.y + offset);
+  auto n_head_up = N * ((n_head + N - 1) / N);
+  auto head_idx = static_cast<int>((pos.z * N) % n_head_up);
+  auto batch_idx = (pos.z * N) / n_head_up;
+  auto batch_offset = offset[batch_idx * offset_stride];
+  float L = scale * static_cast<float>(pos.y + batch_offset);
+  auto mat_idx = batch_idx * n_head + head_idx;
 
   // Compute costheta, sintheta
   float theta = L * inv_freq;
@@ -123,20 +129,19 @@ __device__ void rope_impl(
   size_t out_index_1, out_index_2;
   if (traditional) {
     out_index_1 = 2 * pos.x * out_strides[2] + pos.y * out_strides[1] +
-        N * pos.z * out_strides[0];
+        mat_idx * out_strides[0];
     out_index_2 = out_index_1 + 1;
     in_index_1 =
-        2 * pos.x * strides[2] + pos.y * strides[1] + N * pos.z * strides[0];
+        2 * pos.x * strides[2] + pos.y * strides[1] + mat_idx * strides[0];
     in_index_2 = in_index_1 + strides[2];
   } else {
     out_index_1 = pos.x * out_strides[2] + pos.y * out_strides[1] +
-        N * pos.z * out_strides[0];
+        mat_idx * out_strides[0];
     out_index_2 = out_index_1 + dims.x * out_strides[2];
-    in_index_1 =
-        pos.x * strides[2] + pos.y * strides[1] + N * pos.z * strides[0];
+    in_index_1 = pos.x * strides[2] + pos.y * strides[1] + mat_idx * strides[0];
     in_index_2 = in_index_1 + dims.x * strides[2];
   }
-  for (int i = 0; i < N && pos.z * N + i < n_batch; ++i) {
+  for (int i = 0; i < N && head_idx + i < n_head; ++i) {
     // Read and write the output
     float x1 = static_cast<float>(in[in_index_1]);
     float x2 = static_cast<float>(in[in_index_2]);
@@ -167,7 +172,8 @@ __global__ void rope(
     float base,
     const __grid_constant__ cuda::std::array<int64_t, 3> strides,
     const __grid_constant__ cuda::std::array<int64_t, 3> out_strides,
-    int64_t n_batch,
+    int64_t offset_stride,
+    int n_head,
     uint3 dims) {
   uint3 pos = make_uint3(
       blockIdx.x * blockDim.x + threadIdx.x,
@@ -182,12 +188,13 @@ __global__ void rope(
   rope_impl<T, traditional, forward>(
       in,
       out,
-      *offset,
+      offset,
       inv_freq,
       scale,
       strides,
       out_strides,
-      n_batch,
+      offset_stride,
+      n_head,
       pos,
       dims);
 }
@@ -202,7 +209,8 @@ __global__ void rope_freqs(
     float base,
     const __grid_constant__ cuda::std::array<int64_t, 3> strides,
     const __grid_constant__ cuda::std::array<int64_t, 3> out_strides,
-    int64_t n_batch,
+    int64_t offset_stride,
+    int n_head,
     uint3 dims,
     int64_t freq_stride) {
   uint3 pos = make_uint3(
@@ -217,12 +225,13 @@ __global__ void rope_freqs(
   rope_impl<T, traditional, forward>(
       in,
       out,
-      *offset,
+      offset,
       inv_freq,
       scale,
       strides,
       out_strides,
-      n_batch,
+      offset_stride,
+      n_head,
       pos,
       dims);
 }
@@ -245,23 +254,28 @@ void RoPE::eval_gpu(
   auto& offset = inputs[1];
   auto& out = outputs[0];
 
-  if (in.ndim() < 3) {
-    throw std::runtime_error("[RoPE] Input must have at least 3 dimensions");
-  }
-
   cuda::std::array<int64_t, 3> strides;
   cuda::std::array<int64_t, 3> out_strides;
   bool donated = false;
   int ndim = in.ndim();
-  int dispatch_ndim = in.ndim();
+
+  int B = in.shape(0);
+  int T = in.shape(-2);
+  int D = in.shape(-1);
+  size_t mat_size = T * D;
+  int dispatch_ndim = ndim;
   while (in.shape(-dispatch_ndim) == 1 && dispatch_ndim > 3) {
     dispatch_ndim--;
   }
-  size_t mat_size = in.shape(-2) * in.shape(-1);
+
+  int N = 1;
+  for (int i = 1; i < (ndim - 2); ++i) {
+    N *= in.shape(i);
+  }
 
   // We apply rope to less that the whole vector so copy to output and then
   // apply in-place.
-  if (dims_ < in.shape(-1)) {
+  if (dims_ < D) {
     donated = true;
     auto ctype =
         (in.flags().row_contiguous) ? CopyType::Vector : CopyType::General;
@@ -302,7 +316,7 @@ void RoPE::eval_gpu(
   out_strides[2] = out.strides()[ndim - 1];
 
   // Some flags to help us dispatch below
-  bool single = in.flags().row_contiguous && (mat_size == in.shape(-1));
+  bool single = in.flags().row_contiguous && B == 1 && T == 1;
   bool with_freqs = inputs.size() == 3;
 
   auto& encoder = cu::get_command_encoder(s);
@@ -319,7 +333,7 @@ void RoPE::eval_gpu(
         if (single && !with_freqs) {
           auto kernel =
               cu::rope_single<DataType, traditional.value, forward.value>;
-          uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
+          uint2 dims = make_uint2(dims_ / 2, N);
           auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
           encoder.add_kernel_node(
               kernel,
@@ -336,7 +350,7 @@ void RoPE::eval_gpu(
         } else if (single) {
           auto kernel =
               cu::rope_single_freqs<DataType, traditional.value, forward.value>;
-          uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
+          uint2 dims = make_uint2(dims_ / 2, N);
           auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
           encoder.add_kernel_node(
               kernel,
@@ -354,10 +368,14 @@ void RoPE::eval_gpu(
         } else if (with_freqs) {
           auto kernel =
               cu::rope_freqs<DataType, traditional.value, forward.value>;
-          uint3 dims =
-              make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
-          dims.z = (dims.z + 3) / 4;
+          int n_per_thread = 4;
+          uint32_t dimz = B * ((N + n_per_thread - 1) / n_per_thread);
+          uint3 dims = make_uint3(dims_ / 2, T, dimz);
           auto [grid, block] = get_grid_and_block(dims.x, dims.y, dims.z);
+          int64_t offset_stride = 0;
+          if (inputs[1].ndim() > 0) {
+            offset_stride = inputs[1].strides()[0];
+          }
           encoder.add_kernel_node(
               kernel,
               grid,
@@ -371,15 +389,20 @@ void RoPE::eval_gpu(
               std::log2(base_),
               strides,
               out_strides,
-              in.size() / mat_size,
+              offset_stride,
+              N,
               dims,
               inputs[2].strides(0));
         } else {
           auto kernel = cu::rope<DataType, traditional.value, forward.value>;
-          uint3 dims =
-              make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
-          dims.z = (dims.z + 3) / 4;
+          int n_per_thread = 4;
+          uint32_t dimz = B * ((N + n_per_thread - 1) / n_per_thread);
+          uint3 dims = make_uint3(dims_ / 2, T, dimz);
           auto [grid, block] = get_grid_and_block(dims.x, dims.y, dims.z);
+          int64_t offset_stride = 0;
+          if (inputs[1].ndim() > 0) {
+            offset_stride = inputs[1].strides()[0];
+          }
           encoder.add_kernel_node(
               kernel,
               grid,
@@ -392,7 +415,8 @@ void RoPE::eval_gpu(
               std::log2(base_),
               strides,
               out_strides,
-              in.size() / mat_size,
+              offset_stride,
+              N,
               dims);
         }
       });

mlx/backend/cuda/scaled_dot_product_attention.cu

@@ -4,7 +4,6 @@
 #include "mlx/backend/cuda/device/config.h"
 #include "mlx/backend/cuda/device/utils.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/dtype_utils.h"
 #include "mlx/fast_primitives.h"
@@ -46,6 +45,7 @@ __global__ void kernel_sdpav_1pass(
     const T* K,
     const T* V,
     T* O,
+    const T* sinks,
     __grid_constant__ const AttnParams params) {
   constexpr int BN = 32;
   constexpr int BD = 32;
@@ -65,7 +65,7 @@ __global__ void kernel_sdpav_1pass(
   __shared__ U max_scores[BN];
   __shared__ U sum_exp_scores[BN];
 
-  const U scale_log2 = params.scale * 1.44269504089f;
+  const U scale_log2 = params.scale * M_LOG2E;
 
   auto block = cg::this_thread_block();
   auto warp = cg::tiled_partition<32>(block);
@@ -108,8 +108,12 @@ __global__ void kernel_sdpav_1pass(
     o[i] = 0.f;
   }
 
-  U max_score = -INFINITY;
+  U max_score = Limits<U>::finite_min();
   U sum_exp_score = 0.f;
+  if (sinks && warp_idx == 0) {
+    max_score = M_LOG2E * static_cast<U>(sinks[head_idx]);
+    sum_exp_score = 1.f;
+  }
 
   // For each key
   for (int i = kv_seq_idx; i < params.kL; i += BN) {
@@ -167,7 +171,7 @@ __global__ void kernel_sdpav_1pass(
   U factor = exp2f(max_score - new_max);
   sum_exp_score =
       cg::reduce(warp, sum_exp_scores[lane_idx] * factor, cg::plus<U>());
-  sum_exp_score = __frcp_rn(sum_exp_score);
+  sum_exp_score = sum_exp_score == 0 ? 0 : __frcp_rn(sum_exp_score);
 
   // Now we need to aggregate all the outputs
   PRAGMA_LOOP_UNROLL
@@ -193,6 +197,7 @@ __global__ void kernel_sdpav_2pass_1(
     const T* Q,
     const T* K,
     const T* V,
+    const T* sinks,
     float* partials,
     float* sums,
     float* maxs,
@@ -268,8 +273,12 @@ __global__ void kernel_sdpav_2pass_1(
     o[i] = 0.f;
   }
 
-  U max_score = -1e9;
+  U max_score = Limits<U>::finite_min();
   U sum_exp_score = 0.f;
+  if (sinks && warp_idx == 0 && block_idx == 0) {
+    max_score = M_LOG2E * static_cast<U>(sinks[head_idx]);
+    sum_exp_score = 1.f;
+  }
 
   // For each key
   for (int i = kv_seq_idx; i < params.kL; i += blocks * BN) {
@@ -410,7 +419,7 @@ __global__ void kernel_sdpav_2pass_2(
   U new_max = cg::reduce(warp, max_score, cg::greater<U>());
   U factor = exp2f(max_score - new_max);
   U sum_exp_score = cg::reduce(warp, sums[lane_idx] * factor, cg::plus<U>());
-  sum_exp_score = __frcp_rn(sum_exp_score);
+  sum_exp_score = sum_exp_score == 0 ? 0 : __frcp_rn(sum_exp_score);
 
   PRAGMA_LOOP_UNROLL
   for (int i = 0; i < v_per_thread; i++) {
@@ -463,10 +472,14 @@ void sdpa_vector_1pass_fallback(
     const array& v,
     const float scale,
     array& o,
-    bool do_causal_ = false) {
+    bool do_causal,
+    const std::optional<array>& sinks) {
   encoder.set_input_array(q);
   encoder.set_input_array(k);
   encoder.set_input_array(v);
+  if (sinks) {
+    encoder.set_input_array(*sinks);
+  }
   encoder.set_output_array(o);
 
   cu::AttnParams params{
@@ -489,7 +502,7 @@ void sdpa_vector_1pass_fallback(
   dim3 block_dim(1024, 1, 1);
 
   dispatch_float_types(o.dtype(), "kernel_sdpav_1pass", [&](auto type_tag) {
-    dispatch_bool(do_causal_, [&](auto do_causal) {
+    dispatch_bool(do_causal, [&](auto do_causal) {
       dispatch_headdim(params.D, [&](auto headdim) {
         using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
 
@@ -504,6 +517,7 @@ void sdpa_vector_1pass_fallback(
             k.data<DataType>(),
             v.data<DataType>(),
             o.data<DataType>(),
+            sinks ? (*sinks).data<DataType>() : nullptr,
             params);
       });
     });
@@ -518,7 +532,8 @@ void sdpa_vector_2pass_fallback(
     const array& v,
     const float scale,
     array& o,
-    bool do_causal_ = false) {
+    bool do_causal,
+    const std::optional<array>& sinks) {
   cu::AttnParams params{
       /* int B = */ q.shape(0),
       /* int H = */ q.shape(1),
@@ -559,7 +574,7 @@ void sdpa_vector_2pass_fallback(
   encoder.add_temporary(maxs);
 
   dispatch_float_types(o.dtype(), "kernel_sdpav_2pass", [&](auto type_tag) {
-    dispatch_bool(do_causal_, [&](auto do_causal) {
+    dispatch_bool(do_causal, [&](auto do_causal) {
       dispatch_headdim(params.D, [&](auto headdim) {
         using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
 
@@ -570,6 +585,10 @@ void sdpa_vector_2pass_fallback(
           encoder.set_input_array(q);
           encoder.set_input_array(k);
           encoder.set_input_array(v);
+          if (sinks) {
+            encoder.set_input_array(*sinks);
+          }
+
           encoder.set_output_array(intermediate);
           encoder.set_output_array(sums);
           encoder.set_output_array(maxs);
@@ -585,6 +604,7 @@ void sdpa_vector_2pass_fallback(
               q.data<DataType>(),
               k.data<DataType>(),
               v.data<DataType>(),
+              sinks ? (*sinks).data<DataType>() : nullptr,
               intermediate.data<float>(),
               sums.data<float>(),
               maxs.data<float>(),
@@ -627,15 +647,16 @@ void sdpa_vector_fallback(
     const array& v,
     const float scale,
     array& o,
-    bool do_causal_ = false) {
+    bool do_causal,
+    const std::optional<array>& sinks) {
   int kL = k.shape(2);
 
   if (kL > 1024) {
     return sdpa_vector_2pass_fallback(
-        s, encoder, q, k, v, scale, o, do_causal_);
+        s, encoder, q, k, v, scale, o, do_causal, sinks);
   } else {
     return sdpa_vector_1pass_fallback(
-        s, encoder, q, k, v, scale, o, do_causal_);
+        s, encoder, q, k, v, scale, o, do_causal, sinks);
   }
 }
 
@@ -691,7 +712,7 @@ void ScaledDotProductAttention::eval_gpu(
 
   // Define some copy functions to ensure the layout of the inputs is as
   // expected.
-  copies.reserve(3);
+  copies.reserve(inputs.size());
   auto copy_unless = [&copies, &s](
                          auto predicate, const array& arr) -> const array& {
     if (!predicate(arr)) {
@@ -703,6 +724,16 @@ void ScaledDotProductAttention::eval_gpu(
     }
   };
 
+  // Checks that the headdim dimension has stride 1.
+  auto is_matrix_contiguous = [](const array& arr) {
+    return arr.strides(-1) == 1;
+  };
+
+  std::optional<array> sinks = std::nullopt;
+  if (has_sinks_) {
+    sinks = copy_unless(is_matrix_contiguous, inputs.back());
+  }
+
   // We are in vector mode ie single query
   if (q_pre.shape(2) < 4) {
     auto q_copy_unless = [](const array& arr) {
@@ -740,10 +771,6 @@ void ScaledDotProductAttention::eval_gpu(
     const auto& k = copy_unless(kv_copy_unless, k_pre);
     const auto& v = copy_unless(kv_copy_unless, v_pre);
 
-    for (const auto& cp : copies) {
-      encoder.add_temporary(cp);
-    }
-
     // Donate the query if possible
     if (q.is_donatable() && q.flags().row_contiguous && q.size() == o.size()) {
       o.copy_shared_buffer(q);
@@ -752,22 +779,26 @@ void ScaledDotProductAttention::eval_gpu(
       int64_t str_oH = o.shape(3);
       int64_t str_oL = o.shape(1) * str_oH;
       int64_t str_oB = o.shape(2) * str_oL;
-      size_t data_size = o.shape(0) * str_oB;
 
       array::Flags flags{
           /* bool contiguous = */ 1,
           /* bool row_contiguous = */ o.shape(2) == 1,
-          /* bool col_contiguous = */ 0,
+          /* bool col_contiguous = */ o.size() == o.shape(3),
       };
 
       o.set_data(
           allocator::malloc(o.nbytes()),
-          data_size,
+          o.size(),
           {str_oB, str_oH, str_oL, str_oD},
           flags);
     }
 
-    return sdpa_vector_fallback(s, encoder, q, k, v, scale_, o, do_causal_);
+    for (const auto& cp : copies) {
+      encoder.add_temporary(cp);
+    }
+
+    return sdpa_vector_fallback(
+        s, encoder, q, k, v, scale_, o, do_causal_, sinks);
   }
 
   // Full attention mode should never reach here

mlx/backend/cuda/slicing.cpp

@@ -1,8 +1,11 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/common/slicing.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/jit_module.h"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/gpu/slicing.h"
+#include "mlx/dtype_utils.h"
 
 #include <numeric>
 
@@ -27,8 +30,7 @@ void concatenate_gpu(
   flags.row_contiguous = false;
   flags.col_contiguous = false;
   flags.contiguous = false;
-  // TODO: Handle concurrent outputs:
-  // https://github.com/ml-explore/mlx/pull/2145#discussion_r2070753816
+  auto concurrent = cu::get_command_encoder(s).concurrent_context();
   for (int i = 0; i < inputs.size(); i++) {
     array out_slice(inputs[i].shape(), out.dtype(), nullptr, {});
     size_t data_offset = strides[axis] * sizes[i];
@@ -38,4 +40,71 @@ void concatenate_gpu(
   }
 }
 
+array compute_dynamic_offset(
+    const array& indices,
+    const Strides& strides,
+    const std::vector<int>& axes,
+    const Stream& s) {
+  Dtype dtype = indices.dtype();
+  int nidx = axes.size();
+
+  std::string module_name =
+      fmt::format("compute_dynamic_offset_{}_{}", dtype_to_string(dtype), nidx);
+  std::string kernel_name = fmt::format(
+      "mlx::core::cu::compute_dynamic_offset<{}, {}>",
+      dtype_to_cuda_type(dtype),
+      nidx);
+
+  cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
+    std::string source = R"(
+        #include "mlx/backend/cuda/device/utils.cuh"
+
+        namespace mlx::core::cu {
+
+        template <typename T, int NIDX>
+        __global__ void compute_dynamic_offset(
+            const T* indices,
+            int64_t* offset,
+            const __grid_constant__ Strides strides,
+            const __grid_constant__ cuda::std::array<int, NIDX> axes) {
+          int64_t acc = 0;
+          #pragma unroll
+          for (int i = 0; i < NIDX; ++i) {
+            acc += indices[i] * strides[axes[i]];
+          }
+          *offset = acc;
+        }
+
+        } // namespace mlx::core::cu
+    )";
+    return std::make_tuple(false, std::move(source), std::vector{kernel_name});
+  });
+
+  // Prepare output.
+  array offset({1}, int64, nullptr, {});
+  bool donate = indices.is_donatable() &&
+      (indices.data_size() * indices.itemsize()) >= offset.itemsize();
+  if (donate) {
+    offset.copy_shared_buffer(indices);
+  } else {
+    offset.set_data(allocator::malloc(offset.itemsize()));
+  }
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.add_temporary(offset);
+  encoder.set_input_array(indices);
+  encoder.set_output_array(offset);
+
+  cu::KernelArgs args;
+  args.append(indices);
+  args.append(offset);
+  args.append_ndim(strides);
+  args.append(axes);
+
+  auto kernel = mod.get_kernel(kernel_name);
+  encoder.add_kernel_node(kernel, 1, 1, 0, args.args());
+
+  return offset;
+}
+
 } // namespace mlx::core

mlx/backend/cuda/sort.cu

@@ -9,7 +9,7 @@
 #include <nvtx3/nvtx3.hpp>
 #include <thrust/device_ptr.h>
 #include <thrust/transform.h>
-#include <cub/device/device_segmented_sort.cuh>
+#include <cub/device/device_segmented_radix_sort.cuh>
 
 #include <cassert>
 
@@ -79,7 +79,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
         encoder.add_temporary(discard);
 
         size_t size;
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortPairs(
             nullptr,
             size,
             in.data<Type>(),
@@ -90,6 +90,8 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
+            0,
+            sizeof(Type) * 8,
             stream));
 
         array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
@@ -104,7 +106,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             thrust::device_pointer_cast(indices.data<uint32_t>()),
             ModOp<uint32_t>{static_cast<uint32_t>(nsort)});
 
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortPairs(
             temp.data<void>(),
             size,
             in.data<Type>(),
@@ -115,10 +117,12 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
+            0,
+            sizeof(Type) * 8,
             stream));
       } else {
         size_t size;
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortKeys(
             nullptr,
             size,
             in.data<Type>(),
@@ -127,6 +131,8 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
+            0,
+            sizeof(Type) * 8,
             stream));
 
         array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
@@ -134,7 +140,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
 
         // Start capturing after allocations
         auto capture = encoder.capture_context();
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortKeys(
             temp.data<void>(),
             size,
             in.data<Type>(),
@@ -143,6 +149,8 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
+            0,
+            sizeof(Type) * 8,
             stream));
       }
     } else {

mlx/backend/cuda/ternary.cu

@@ -156,7 +156,25 @@ void ternary_op_gpu_inplace(
     using DType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
 
     auto topt = get_ternary_op_type(a, b, c);
-    if (topt == TernaryOpType::General) {
+    if (topt == TernaryOpType::VectorVectorVector ||
+        topt == TernaryOpType::ScalarScalarScalar) {
+      dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
+        using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+        constexpr int N_READS = 16 / sizeof(DType);
+        auto [num_blocks, block_dims] = get_launch_args(
+            out.data_size(), out.shape(), out.strides(), large(), N_READS);
+        encoder.add_kernel_node(
+            cu::ternary_v<Op, DType, IdxT, N_READS>,
+            num_blocks,
+            block_dims,
+            0,
+            a.data<bool>(),
+            b.data<DType>(),
+            c.data<DType>(),
+            out.data<DType>(),
+            out.data_size());
+      });
+    } else {
       dispatch_bool(
           a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
               c.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
@@ -225,23 +243,6 @@ void ternary_op_gpu_inplace(
                   ndim);
             }
           });
-    } else {
-      dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
-        using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-        constexpr int N_READS = 16 / sizeof(DType);
-        auto [num_blocks, block_dims] = get_launch_args(
-            out.data_size(), out.shape(), out.strides(), large(), N_READS);
-        encoder.add_kernel_node(
-            cu::ternary_v<Op, DType, IdxT, N_READS>,
-            num_blocks,
-            block_dims,
-            0,
-            a.data<bool>(),
-            b.data<DType>(),
-            c.data<DType>(),
-            out.data<DType>(),
-            out.data_size());
-      });
     }
   });
 }

mlx/backend/cuda/unary.cu

@@ -1,284 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/common/unary.h"
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/unary_ops.cuh"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/dtype_utils.h"
-#include "mlx/primitives.h"
-
-#include <cooperative_groups.h>
-#include <nvtx3/nvtx3.hpp>
-
-namespace mlx::core {
-
-namespace cu {
-
-namespace cg = cooperative_groups;
-
-template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
-__global__ void unary_v(const In* in, Out* out, IdxT size) {
-  IdxT index = cg::this_grid().thread_rank();
-
-  if ((index + 1) * N_READS > size) {
-    for (IdxT i = index * N_READS; i < size; ++i) {
-      out[i] = Op{}(in[i]);
-    }
-  } else {
-    auto in_vec = load_vector<N_READS>(in, index);
-
-    AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-    for (int i = 0; i < N_READS; ++i) {
-      out_vec[i] = Op{}(in_vec[i]);
-    }
-
-    store_vector<N_READS>(out, index, out_vec);
-  }
-}
-
-template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
-__global__ void unary_g(
-    const In* in,
-    Out* out,
-    IdxT size_rest,
-    const __grid_constant__ Shape shape,
-    const __grid_constant__ Strides strides,
-    int ndim) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[ndim - 1];
-  auto stride_x = strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
-  auto idx =
-      elem_to_loc(index_rest * shape_x, shape.data(), strides.data(), ndim);
-  auto in_vec =
-      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = Op{}(in_vec[i]);
-  }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
-}
-
-template <typename Op, typename In, typename Out>
-constexpr bool supports_unary_op() {
-  if (std::is_same_v<Op, Abs> || std::is_same_v<Op, Negative> ||
-      std::is_same_v<Op, Sign> || std::is_same_v<Op, Square>) {
-    return std::is_same_v<In, Out>;
-  }
-  if (std::is_same_v<Op, ArcCosh> || std::is_same_v<Op, ArcSinh> ||
-      std::is_same_v<Op, ArcTanh> || std::is_same_v<Op, Erf> ||
-      std::is_same_v<Op, ErfInv> || std::is_same_v<Op, Expm1> ||
-      std::is_same_v<Op, Sigmoid>) {
-    return std::is_same_v<In, Out> && is_floating_v<In>;
-  }
-  if (std::is_same_v<Op, BitwiseInvert>) {
-    return std::is_same_v<In, Out> && std::is_integral_v<In> &&
-        !std::is_same_v<In, bool>;
-  }
-  if (std::is_same_v<Op, Ceil> || std::is_same_v<Op, Floor>) {
-    return std::is_same_v<In, Out> && !mlx::core::is_complex_v<In>;
-  }
-  if (std::is_same_v<Op, Conjugate>) {
-    return std::is_same_v<In, Out> && mlx::core::is_complex_v<In>;
-  }
-  if (std::is_same_v<Op, ArcCos> || std::is_same_v<Op, ArcSin> ||
-      std::is_same_v<Op, ArcTan> || std::is_same_v<Op, Cos> ||
-      std::is_same_v<Op, Cosh> || std::is_same_v<Op, Exp> ||
-      std::is_same_v<Op, Log> || std::is_same_v<Op, Log2> ||
-      std::is_same_v<Op, Log10> || std::is_same_v<Op, Log1p> ||
-      std::is_same_v<Op, Round> || std::is_same_v<Op, Rsqrt> ||
-      std::is_same_v<Op, Sqrt> || std::is_same_v<Op, Sin> ||
-      std::is_same_v<Op, Sinh> || std::is_same_v<Op, Tan> ||
-      std::is_same_v<Op, Tanh>) {
-    return std::is_same_v<In, Out> && is_inexact_v<In>;
-  }
-  if (std::is_same_v<Op, Imag> || std::is_same_v<Op, Real>) {
-    return mlx::core::is_complex_v<In> && std::is_same_v<Out, float>;
-  }
-  if (std::is_same_v<Op, LogicalNot>) {
-    return std::is_same_v<In, Out> && std::is_same_v<In, bool>;
-  }
-  return false;
-}
-
-} // namespace cu
-
-template <typename Op>
-void unary_op_gpu_inplace(
-    const std::vector<array>& inputs,
-    array& out,
-    const char* op,
-    const Stream& s) {
-  auto& in = inputs[0];
-  if (in.size() == 0) {
-    return;
-  }
-  bool contig = in.flags().contiguous;
-  bool large;
-  if (!contig) {
-    large = in.data_size() > INT32_MAX || out.size() > INT32_MAX;
-  } else {
-    large = in.data_size() > UINT32_MAX;
-  }
-
-  auto& encoder = cu::get_command_encoder(s);
-  encoder.set_input_array(in);
-  encoder.set_output_array(out);
-  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
-    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
-      using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
-      using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
-      if constexpr (cu::supports_unary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
-        dispatch_bool(large, [&](auto large) {
-          using InType = cuda_type_t<CTYPE_IN>;
-          using OutType = cuda_type_t<CTYPE_OUT>;
-          if (contig) {
-            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            constexpr int N_READS = 16 / sizeof(OutType);
-            auto [num_blocks, block_dims] = get_launch_args(
-                out.data_size(), out.shape(), out.strides(), large, N_READS);
-            encoder.add_kernel_node(
-                cu::unary_v<Op, InType, OutType, IdxT, N_READS>,
-                num_blocks,
-                block_dims,
-                0,
-                in.data<InType>(),
-                out.data<OutType>(),
-                out.data_size());
-          } else {
-            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-            auto [shape, strides] = collapse_contiguous_dims(in);
-            auto ndim = shape.size();
-            int work_per_thread = 1;
-            auto kernel = cu::unary_g<Op, InType, OutType, IdxT, 1>;
-            auto dim0 = ndim > 0 ? shape.back() : 1;
-            auto rest = out.size() / dim0;
-            if (dim0 >= 4) {
-              kernel = cu::unary_g<Op, InType, OutType, IdxT, 4>;
-              work_per_thread = 4;
-            }
-            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-            auto block_dims = get_block_dims(dim0, rest, 1);
-            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-            encoder.add_kernel_node(
-                kernel,
-                {num_blocks_x, num_blocks_y},
-                block_dims,
-                0,
-                in.data<InType>(),
-                out.data<OutType>(),
-                rest,
-                const_param(shape),
-                const_param(strides),
-                ndim);
-          }
-        });
-      } else {
-        throw std::runtime_error(fmt::format(
-            "Can not do unary op {} on input of {} with output of {}.",
-            op,
-            dtype_to_string(in.dtype()),
-            dtype_to_string(out.dtype())));
-      }
-    });
-  });
-}
-
-template <typename Op>
-void unary_op_gpu(
-    const std::vector<array>& inputs,
-    array& out,
-    const char* op,
-    const Stream& s) {
-  set_unary_output_data(inputs[0], out);
-  unary_op_gpu_inplace<Op>(inputs, out, op, s);
-}
-
-#define UNARY_GPU(func)                                               \
-  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
-    nvtx3::scoped_range r(#func "::eval_gpu");                        \
-    auto& s = out.primitive().stream();                               \
-    unary_op_gpu<cu::func>(inputs, out, name(), s);                   \
-  }
-
-UNARY_GPU(Abs)
-UNARY_GPU(ArcCos)
-UNARY_GPU(ArcCosh)
-UNARY_GPU(ArcSin)
-UNARY_GPU(ArcSinh)
-UNARY_GPU(ArcTan)
-UNARY_GPU(ArcTanh)
-UNARY_GPU(BitwiseInvert)
-UNARY_GPU(Ceil)
-UNARY_GPU(Conjugate)
-UNARY_GPU(Cos)
-UNARY_GPU(Cosh)
-UNARY_GPU(Erf)
-UNARY_GPU(ErfInv)
-UNARY_GPU(Exp)
-UNARY_GPU(Expm1)
-UNARY_GPU(Floor)
-UNARY_GPU(Imag)
-UNARY_GPU(Log1p)
-UNARY_GPU(LogicalNot)
-UNARY_GPU(Negative)
-UNARY_GPU(Real)
-UNARY_GPU(Sigmoid)
-UNARY_GPU(Sign)
-UNARY_GPU(Sin)
-UNARY_GPU(Sinh)
-UNARY_GPU(Square)
-UNARY_GPU(Tan)
-UNARY_GPU(Tanh)
-
-void Log::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Log::eval_gpu");
-  auto& s = out.primitive().stream();
-  switch (base_) {
-    case Base::e:
-      unary_op_gpu<cu::Log>(inputs, out, name(), s);
-      break;
-    case Base::two:
-      unary_op_gpu<cu::Log2>(inputs, out, name(), s);
-      break;
-    case Base::ten:
-      unary_op_gpu<cu::Log10>(inputs, out, name(), s);
-      break;
-  }
-}
-
-void Round::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Round::eval_gpu");
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  auto& s = out.primitive().stream();
-  if (issubdtype(in.dtype(), inexact)) {
-    unary_op_gpu<cu::Round>(inputs, out, name(), s);
-  } else {
-    // No-op integer types
-    out.copy_shared_buffer(in);
-  }
-}
-
-void Sqrt::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Sort::eval_gpu");
-  auto& s = out.primitive().stream();
-  if (recip_) {
-    unary_op_gpu<cu::Rsqrt>(inputs, out, "Rsqrt", s);
-  } else {
-    unary_op_gpu<cu::Sqrt>(inputs, out, "Sqrt", s);
-  }
-}
-
-} // namespace mlx::core

mlx/backend/cuda/utils.h

@@ -1,6 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
-// This file include utilies that are used by C++ code (i.e. .cpp files).
+// This file include utilities that are used by C++ code (i.e. .cpp files).
 
 #pragma once
 
@@ -12,6 +12,7 @@ namespace mlx::core {
 
 namespace cu {
 class Device;
+
 }
 
 struct Dtype;
@@ -86,4 +87,17 @@ class CudaStream : public CudaHandle<cudaStream_t, cudaStreamDestroy> {
   explicit CudaStream(cu::Device& device);
 };
 
+template <typename T>
+inline uint max_occupancy_block_dim(T kernel) {
+  int _, block_dim;
+  if constexpr (std::is_same_v<T, CUfunction>) {
+    CHECK_CUDA_ERROR(
+        cuOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel, 0, 0, 0));
+  } else {
+    CHECK_CUDA_ERROR(
+        cudaOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel));
+  }
+  return block_dim;
+}
+
 } // namespace mlx::core

mlx/backend/cuda/worker.cpp

@@ -5,8 +5,9 @@
 
 namespace mlx::core::cu {
 
-Worker::Worker()
-    : signal_stream_(device(mlx::core::Device::gpu)),
+Worker::Worker(Device& d)
+    : signal_stream_(d),
+      signal_event_(d, cudaEventDisableTiming | cudaEventBlockingSync),
       worker_(&Worker::thread_fn, this) {}
 
 Worker::~Worker() {

mlx/backend/cuda/worker.h

@@ -3,7 +3,6 @@
 #pragma once
 
 #include "mlx/backend/cuda/event.h"
-#include "mlx/backend/cuda/utils.h"
 
 #include <condition_variable>
 #include <functional>
@@ -16,7 +15,7 @@ namespace mlx::core::cu {
 // Run tasks in worker thread, synchronized with cuda stream.
 class Worker {
  public:
-  Worker();
+  explicit Worker(Device& d);
   ~Worker();
 
   Worker(const Worker&) = delete;

mlx/backend/gpu/copy.h

@@ -20,8 +20,8 @@ void copy_gpu_inplace(
     int64_t o_offset,
     CopyType ctype,
     const Stream& s,
-    const std::optional<array>& dynamic_i_offset = std::nullopt,
-    const std::optional<array>& dynamic_o_offset = std::nullopt);
+    std::optional<array> dynamic_i_offset = std::nullopt,
+    std::optional<array> dynamic_o_offset = std::nullopt);
 
 void copy_gpu(const array& src, array& out, CopyType ctype, const Stream& s);
 void copy_gpu(const array& src, array& out, CopyType ctype);

mlx/backend/gpu/primitives.cpp

@@ -80,6 +80,74 @@ void Depends::eval_gpu(
   eval(inputs, outputs);
 }
 
+void DynamicSlice::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("DynamicSlice::eval_gpu");
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  auto& in = inputs[0];
+  auto& start = inputs[1];
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto s = stream();
+  auto in_offset = compute_dynamic_offset(start, in.strides(), axes_, s);
+  copy_gpu_inplace(
+      /* const array& src = */ in,
+      /* array& dst = */ out,
+      /* const Shape& data_shape = */ out.shape(),
+      /* const Strides& i_strides = */ in.strides(),
+      /* const Strides& o_strides = */ out.strides(),
+      /* int64_t i_offset = */ 0,
+      /* int64_t o_offset = */ 0,
+      /* CopyType ctype = */ CopyType::GeneralGeneral,
+      /* const Stream& s = */ s,
+      /* std::optional<array> dynamic_i_offset = */ std::move(in_offset),
+      /* std::optional<array> dynamic_o_offset = */ std::nullopt);
+}
+
+void DynamicSliceUpdate::eval_gpu(
+    const std::vector<array>& inputs,
+    array& out) {
+  MLX_PROFILER_RANGE("DynamicSliceUpdate::eval_gpu");
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  auto& in = inputs[0];
+  auto& upd = inputs[1];
+  auto& start_indices = inputs[2];
+
+  if (upd.size() == 0) {
+    out.copy_shared_buffer(in);
+    return;
+  }
+
+  // Copy or donate input to output
+  auto s = stream();
+  auto ctype = in.flags().contiguous && in.size() == in.data_size()
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_gpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, s);
+
+  auto out_offset =
+      compute_dynamic_offset(start_indices, out.strides(), axes_, s);
+  copy_gpu_inplace(
+      /* const array& src = */ upd,
+      /* array& dst = */ out,
+      /* const Shape& data_shape = */ upd.shape(),
+      /* const Strides& i_strides = */ upd.strides(),
+      /* const Strides& o_strides = */ out.strides(),
+      /* int64_t i_offset = */ 0,
+      /* int64_t o_offset = */ 0,
+      /* CopyType ctype = */ CopyType::GeneralGeneral,
+      /* const Stream& s = */ s,
+      /* std::optional<array> dynamic_i_offset = */ std::nullopt,
+      /* std::optional<array> dynamic_o_offset = */ std::move(out_offset));
+}
+
 void ExpandDims::eval_gpu(const std::vector<array>& inputs, array& out) {
   MLX_PROFILER_RANGE("ExpandDims::eval_gpu");
   eval(inputs, out);

mlx/backend/gpu/slicing.h

@@ -27,4 +27,10 @@ void pad_gpu(
     const Shape& low_pad_size,
     const Stream& s);
 
+array compute_dynamic_offset(
+    const array& indices,
+    const Strides& strides,
+    const std::vector<int>& axes,
+    const Stream& s);
+
 } // namespace mlx::core

mlx/backend/metal/CMakeLists.txt

@@ -33,10 +33,11 @@ make_jit_source(unary_ops kernels/erf.h kernels/expm1f.h)
 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(scatter kernels/indexing.h)
-make_jit_source(gather kernels/indexing.h)
-make_jit_source(gather_axis)
-make_jit_source(scatter_axis)
+make_jit_source(indexing/scatter kernels/indexing/indexing.h)
+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)
+make_jit_source(indexing/scatter_axis)
 make_jit_source(hadamard)
 
 if(MLX_METAL_JIT)
@@ -77,7 +78,10 @@ if(MLX_METAL_JIT)
   make_jit_source(steel/conv/kernels/steel_conv)
   make_jit_source(steel/conv/kernels/steel_conv_general kernels/steel/defines.h
                   kernels/steel/conv/loaders/loader_general.h)
-  make_jit_source(quantized)
+
+  make_jit_source(quantized_utils)
+  make_jit_source(quantized kernels/quantized_utils.h)
+  make_jit_source(fp4_quantized kernels/quantized_utils.h)
   make_jit_source(gemv_masked)
 else()
   target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/nojit_kernels.cpp)

mlx/backend/metal/conv.cpp

@@ -2,7 +2,6 @@
 #include <algorithm>
 #include <cassert>
 #include <numeric>
-#include <sstream>
 
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
@@ -39,10 +38,11 @@ void explicit_gemm_conv_ND_gpu(
   in_unfolded.set_data(allocator::malloc(in_unfolded.nbytes()));
 
   // Prepare unfolding kernel
-  std::ostringstream kname;
-  kname << "naive_unfold_nd_" << type_to_name(in_unfolded) << "_" << N;
+  std::string kname;
+  kname.reserve(32);
+  concatenate(kname, "naive_unfold_nd_", type_to_name(in_unfolded), "_", N);
   auto& compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  auto kernel = d.get_kernel(kname);
   compute_encoder.set_compute_pipeline_state(kernel);
 
   compute_encoder.set_input_array(in, 0);
@@ -117,11 +117,12 @@ void explicit_gemm_conv_group_ND_gpu(
   in_unfolded.set_data(allocator::malloc(in_unfolded.nbytes()));
 
   // Prepare unfolding kernel
-  std::ostringstream kname;
-  kname << "naive_unfold_transpose_nd_" << type_to_name(in_unfolded) << "_"
-        << N;
+  std::string kname;
+  kname.reserve(32);
+  concatenate(
+      kname, "naive_unfold_transpose_nd_", type_to_name(in_unfolded), "_", N);
   auto& compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  auto kernel = d.get_kernel(kname);
   compute_encoder.set_compute_pipeline_state(kernel);
 
   compute_encoder.set_input_array(in, 0);
@@ -252,18 +253,32 @@ void implicit_gemm_conv_2D_gpu(
       /* const int swizzle_log = */ swizzle_log};
 
   // Determine kernel
-  std::ostringstream kname;
-  kname << "implicit_gemm_conv_2d_" << type_to_name(out) << "_bm" << bm << "_bn"
-        << bn << "_bk" << bk << "_wm" << wm << "_wn" << wn << "_channel_"
-        << (n_channel_specialization ? std::to_string(n_channel_specialization)
-                                     : "l")
-        << "_filter_" << (small_filter ? 's' : 'l');
+  std::string kname;
+  kname.reserve(64);
+  concatenate(
+      kname,
+      "implicit_gemm_conv_2d_",
+      type_to_name(out),
+      "_bm",
+      bm,
+      "_bn",
+      bn,
+      "_bk",
+      bk,
+      "_wm",
+      wm,
+      "_wn",
+      wn,
+      "_channel_",
+      n_channel_specialization ? std::to_string(n_channel_specialization) : "l",
+      "_filter_",
+      small_filter ? 's' : 'l');
 
   // Encode and dispatch kernel
   auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = get_steel_conv_kernel(
       d,
-      kname.str(),
+      kname,
       out,
       bm,
       bn,
@@ -559,11 +574,16 @@ void winograd_conv_2D_gpu(
   {
     int bc = 32;
     int bo = 4;
-    std::ostringstream kname;
-    kname << "winograd_conv_2d_weight_transform_" << type_to_name(out) << "_bc"
-          << bc;
+    std::string kname;
+    kname.reserve(32);
+    concatenate(
+        kname,
+        "winograd_conv_2d_weight_transform_",
+        type_to_name(out),
+        "_bc",
+        bc);
     auto& compute_encoder = d.get_command_encoder(s.index);
-    auto kernel = d.get_kernel(kname.str());
+    auto kernel = d.get_kernel(kname);
     compute_encoder.set_compute_pipeline_state(kernel);
 
     compute_encoder.set_input_array(wt, 0);
@@ -587,11 +607,16 @@ void winograd_conv_2D_gpu(
     int bc = 32;
     int wm = 2;
     int wn = 2;
-    std::ostringstream kname;
-    kname << "winograd_conv_2d_input_transform_" << type_to_name(out) << "_bc"
-          << bc;
+    std::string kname;
+    kname.reserve(32);
+    concatenate(
+        kname,
+        "winograd_conv_2d_input_transform_",
+        type_to_name(out),
+        "_bc",
+        bc);
     auto& compute_encoder = d.get_command_encoder(s.index);
-    auto kernel = d.get_kernel(kname.str());
+    auto kernel = d.get_kernel(kname);
     compute_encoder.set_compute_pipeline_state(kernel);
 
     compute_encoder.set_input_array(in_padded, 0);
@@ -634,11 +659,16 @@ void winograd_conv_2D_gpu(
     int bc = 32;
     int wm = 2;
     int wn = 2;
-    std::ostringstream kname;
-    kname << "winograd_conv_2d_output_transform_" << type_to_name(out) << "_bo"
-          << bc;
+    std::string kname;
+    kname.reserve(32);
+    concatenate(
+        kname,
+        "winograd_conv_2d_output_transform_",
+        type_to_name(out),
+        "_bo",
+        bc);
     auto& compute_encoder = d.get_command_encoder(s.index);
-    auto kernel = d.get_kernel(kname.str());
+    auto kernel = d.get_kernel(kname);
     compute_encoder.set_compute_pipeline_state(kernel);
 
     compute_encoder.set_input_array(out_wg, 0);
@@ -660,9 +690,9 @@ void depthwise_conv_2D_gpu(
     const array& wt,
     array out,
     const MLXConvParams<2>& conv_params) {
-  std::ostringstream kname;
-  kname << "depthwise_conv_2d_" << type_to_name(out);
-  std::string base_name = kname.str();
+  std::string base_name;
+  base_name.reserve(32);
+  concatenate(base_name, "depthwise_conv_2d_", type_to_name(out));
 
   const int N = conv_params.N;
   const int ker_h = conv_params.wS[0];
@@ -685,15 +715,18 @@ void depthwise_conv_2D_gpu(
   };
 
   // clang-format off
-  kname << "_ker_h_" << ker_h
-        << "_ker_w_" << ker_w
-        << "_str_h_" << str_h
-        << "_str_w_" << str_w
-        << "_tgp_h_" << th
-        << "_tgp_w_" << tw
-        << "_do_flip_" << (do_flip ? 't' : 'n'); // clang-format on
-
-  std::string hash_name = kname.str();
+  std::string hash_name;
+  hash_name.reserve(64);
+  concatenate(
+      hash_name,
+      base_name,
+  "_ker_h_", ker_h,
+  "_ker_w_", ker_w,
+  "_str_h_", str_h,
+  "_str_w_", str_w,
+  "_tgp_h_", th,
+  "_tgp_w_", tw,
+  "_do_flip_", do_flip ? 't' : 'n'); // clang-format on
 
   auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(base_name, hash_name, func_consts);
@@ -774,6 +807,56 @@ void dispatch_conv_2D_gpu(
   }
 }
 
+void depthwise_conv_1D_gpu(
+    const Stream& s,
+    metal::Device& d,
+    const array& in,
+    array wt,
+    array out) {
+  bool large = in.size() > INT32_MAX || in.data_size() > INT32_MAX;
+  std::string base_name;
+  base_name.reserve(32);
+  concatenate(
+      base_name,
+      "depthwise_conv_1d_",
+      large ? "_large" : "",
+      type_to_name(out));
+
+  if (!wt.flags().row_contiguous) {
+    wt = contiguous_copy_gpu(wt, s);
+    d.add_temporary(wt, s.index);
+  }
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel = d.get_kernel(base_name);
+  compute_encoder.set_compute_pipeline_state(kernel);
+
+  auto B = in.shape(0);
+  auto Tout = out.shape(1);
+  auto D = in.shape(2);
+  auto K = wt.shape(1);
+
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_input_array(wt, 1);
+  compute_encoder.set_output_array(out, 2);
+  if (large) {
+    int64_t strides[3] = {in.strides(0), in.strides(1), in.strides(2)};
+    compute_encoder.set_bytes(strides, 3, 3);
+
+  } else {
+    int strides[3] = {
+        static_cast<int>(in.strides(0)),
+        static_cast<int>(in.strides(1)),
+        static_cast<int>(in.strides(2))};
+    compute_encoder.set_bytes(strides, 3, 3);
+  }
+
+  compute_encoder.set_bytes(K, 4);
+  auto group_dims = get_block_dims(D, Tout, B);
+  MTL::Size grid_dims = MTL::Size(D, Tout, B);
+
+  compute_encoder.dispatch_threads(grid_dims, group_dims);
+}
+
 void conv_1D_gpu(
     const Stream& s,
     metal::Device& d,
@@ -790,8 +873,15 @@ void conv_1D_gpu(
   bool is_idil_one = in_dilation[0] == 1;
   int C = in.shape(2);
   int O = wt.shape(0);
-  const int C_per_group = in.shape(2) / groups;
-  const int O_per_group = wt.shape(0) / groups;
+  // Fast path for fully separable 1D convolution
+  if (is_idil_one && (groups == C) && groups == O && wt_strides[0] == 1 &&
+      wt_dilation[0] == 1 && padding[0] == 0 && !flip) {
+    depthwise_conv_1D_gpu(s, d, in, wt, out);
+    return;
+  }
+
+  const int C_per_group = C / groups;
+  const int O_per_group = O / groups;
 
   // Direct to implicit gemm conv
   if (is_idil_one && (C_per_group <= 4 || C_per_group % 16 == 0) &&

mlx/backend/metal/copy.cpp

@@ -20,8 +20,8 @@ void copy_gpu_inplace(
     int64_t out_offset,
     CopyType ctype,
     const Stream& s,
-    const std::optional<array>& dynamic_i_offset /* = std::nullopt */,
-    const std::optional<array>& dynamic_o_offset /* = std::nullopt */) {
+    std::optional<array> dynamic_i_offset /* = std::nullopt */,
+    std::optional<array> dynamic_o_offset /* = std::nullopt */) {
   if (out.size() == 0) {
     return;
   }

mlx/backend/metal/device.cpp

@@ -108,7 +108,7 @@ std::pair<MTL::Library*, NS::Error*> load_swiftpm_library(
 }
 
 MTL::Library* load_default_library(MTL::Device* device) {
-  NS::Error* error[4];
+  NS::Error* error[5];
   MTL::Library* lib;
   // First try the colocated mlx.metallib
   std::tie(lib, error[0]) = load_colocated_library(device, "mlx");
@@ -127,12 +127,19 @@ MTL::Library* load_default_library(MTL::Device* device) {
     return lib;
   }
 
+  // Try lo load resources from Framework resources if SwiftPM wrapped as a
+  // dynamic framework.
+  std::tie(lib, error[3]) = load_colocated_library(device, "Resources/default");
+  if (lib) {
+    return lib;
+  }
+
   // Finally try default_mtllib_path
-  std::tie(lib, error[3]) = load_library_from_path(device, default_mtllib_path);
+  std::tie(lib, error[4]) = load_library_from_path(device, default_mtllib_path);
   if (!lib) {
     std::ostringstream msg;
     msg << "Failed to load the default metallib. ";
-    for (int i = 0; i < 4; i++) {
+    for (int i = 0; i < 5; i++) {
       if (error[i] != nullptr) {
         msg << error[i]->localizedDescription()->utf8String() << " ";
       }
@@ -464,6 +471,10 @@ void Device::end_encoding(int index) {
 CommandEncoder& Device::get_command_encoder(int index) {
   auto& stream = get_stream_(index);
   if (stream.encoder == nullptr) {
+    // Ensure there is an active command buffer
+    if (stream.buffer == nullptr) {
+      get_command_buffer(index);
+    }
     stream.encoder = std::make_unique<CommandEncoder>(stream);
     stream.fence = std::make_shared<Fence>(device_->newFence());
   }

mlx/backend/metal/indexing.cpp

@@ -52,8 +52,10 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& s = stream();
   auto& d = metal::device(s.device);
 
-  int idx_ndim = nidx ? inputs[1].ndim() : 0;
-  size_t ndim = src.ndim();
+  size_t slice_size = 1;
+  for (auto s : slice_sizes_) {
+    slice_size *= s;
+  }
 
   bool large_index = nidx && inputs[1].size() > INT32_MAX;
   bool large_src = src.size() > INT32_MAX;
@@ -61,6 +63,55 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   bool large = large_index || large_src || large_out;
 
   std::string idx_type_name = nidx ? type_to_name(inputs[1]) : "";
+
+  if (src.flags().row_contiguous && nidx == 1 && axes_[0] == 0 &&
+      inputs[1].flags().row_contiguous && slice_size == src.strides()[0]) {
+    int work_per_thread = (slice_size > 8 && src.dtype().size() < 4) ? 2 : 1;
+    auto& indices = inputs[1];
+    std::string kernel_name = fmt::format(
+        "gather_front{0}_{1}_{2}_{3}",
+        type_to_name(out),
+        idx_type_name,
+        large ? "int64_t" : "int",
+        work_per_thread);
+    std::string lib_name = kernel_name;
+
+    auto lib = d.get_library(lib_name, [&]() {
+      std::string kernel_source = metal::utils();
+      kernel_source += metal::gather_front();
+      kernel_source += get_template_definition(
+          kernel_name,
+          "gather_front",
+          get_type_string(out.dtype()),
+          get_type_string(indices.dtype()),
+          large ? "int64_t" : "int",
+          work_per_thread);
+
+      return kernel_source;
+    });
+
+    auto& compute_encoder = d.get_command_encoder(s.index);
+    auto kernel = d.get_kernel(kernel_name, lib);
+    compute_encoder.set_compute_pipeline_state(kernel);
+
+    size_t dim_x = (slice_size + work_per_thread - 1) / work_per_thread;
+    size_t dim_y = indices.size();
+    auto group_dims = get_block_dims(dim_x, dim_y, 1);
+    MTL::Size grid_dims = MTL::Size(dim_x, dim_y, 1);
+
+    compute_encoder.set_input_array(src, 0);
+    compute_encoder.set_input_array(indices, 1);
+    compute_encoder.set_output_array(out, 2);
+    compute_encoder.set_bytes(slice_size, 3);
+    compute_encoder.set_bytes(src.shape(0), 4);
+    compute_encoder.dispatch_threads(grid_dims, group_dims);
+
+    return;
+  }
+
+  int idx_ndim = nidx ? inputs[1].ndim() : 0;
+  size_t ndim = src.ndim();
+
   std::string kernel_name = fmt::format(
       "gather{0}{1}_{2}_{3}_{4}",
       type_to_name(out),
@@ -96,11 +147,6 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto kernel = d.get_kernel(kernel_name, lib);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  size_t slice_size = 1;
-  for (auto s : slice_sizes_) {
-    slice_size *= s;
-  }
-
   // Launch 3D grid of threads
   // First two dimensions for the indices, the last one for the slice
   size_t dim0 = 1;
@@ -332,7 +378,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
 
   if (upd_ndim == 0) {
-    // Need placeholders so Metal doesn't compalain
+    // Need placeholders so Metal doesn't complain
     int shape_ = 0;
     int64_t stride_ = 0;
     compute_encoder.set_bytes(shape_, 3);
@@ -347,7 +393,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   // Set output info
   size_t out_ndim = out.ndim();
   if (out_ndim == 0) {
-    // Need placeholders so Metal doesn't compalain
+    // Need placeholders so Metal doesn't complain
     int shape_ = 0;
     int64_t stride_ = 0;
     compute_encoder.set_bytes(shape_, 7);

mlx/backend/metal/jit/includes.h

@@ -19,9 +19,12 @@ const char* binary_two();
 const char* copy();
 const char* fft();
 const char* gather_axis();
+const char* gather_front();
 const char* hadamard();
 const char* logsumexp();
+const char* quantized_utils();
 const char* quantized();
+const char* fp4_quantized();
 const char* ternary();
 const char* scan();
 const char* scatter_axis();

mlx/backend/metal/jit_kernels.cpp

@@ -144,8 +144,7 @@ MTL::ComputePipelineState* get_ternary_kernel(
     auto t_str = get_type_string(type);
     std::string kernel_source = metal::utils();
     concatenate(kernel_source, metal::ternary_ops(), metal::ternary());
-    const std::array<std::pair<std::string, std::string>, 4> kernel_types = {{
-        {"v2", "ternary_v2"},
+    const std::array<std::pair<std::string, std::string>, 3> kernel_types = {{
         {"g1large", "ternary_g_nd1"},
         {"g2large", "ternary_g_nd2"},
         {"g3large", "ternary_g_nd3"},
@@ -154,13 +153,29 @@ MTL::ComputePipelineState* get_ternary_kernel(
       kernel_source +=
           get_template_definition(name + "_" + lib_name, func, t_str, op);
     }
+
+    kernel_source += get_template_definition(
+        "v2_" + lib_name, "ternary_v2", t_str, op, false, false);
+    kernel_source += get_template_definition(
+        "sv2_" + lib_name, "ternary_v2", t_str, op, true, false);
+    kernel_source += get_template_definition(
+        "vs2_" + lib_name, "ternary_v2", t_str, op, false, true);
+
     if (get_work_per_thread(type) > 1) {
-      kernel_source +=
-          get_template_definition("vn_" + lib_name, "ternary_v", t_str, op);
+      kernel_source += get_template_definition(
+          "vn_" + lib_name, "ternary_v", t_str, op, false, false);
+      kernel_source += get_template_definition(
+          "svn_" + lib_name, "ternary_v", t_str, op, true, false);
+      kernel_source += get_template_definition(
+          "vsn_" + lib_name, "ternary_v", t_str, op, false, true);
     }
 
-    kernel_source +=
-        get_template_definition("v_" + lib_name, "ternary_v", t_str, op, 1);
+    kernel_source += get_template_definition(
+        "v_" + lib_name, "ternary_v", t_str, op, false, false, 1);
+    kernel_source += get_template_definition(
+        "sv_" + lib_name, "ternary_v", t_str, op, true, false, 1);
+    kernel_source += get_template_definition(
+        "vs_" + lib_name, "ternary_v", t_str, op, false, true, 1);
     kernel_source += get_template_definition(
         "g1_" + lib_name, "ternary_g_nd1", t_str, op, "int");
     kernel_source += get_template_definition(
@@ -804,13 +819,19 @@ MTL::ComputePipelineState* get_fft_kernel(
 MTL::ComputePipelineState* get_quantized_kernel(
     metal::Device& d,
     const std::string& kernel_name,
-    const std::string& template_def) {
+    const std::string& template_def,
+    const std::string& mode) {
   const auto& lib_name = kernel_name;
   auto lib = d.get_library(lib_name, [&]() {
-    std::ostringstream kernel_source;
-    kernel_source << metal::utils() << metal::gemm() << metal::quantized()
-                  << template_def;
-    return kernel_source.str();
+    std::string kernel_source;
+    concatenate(
+        kernel_source,
+        metal::utils(),
+        metal::gemm(),
+        metal::quantized_utils(),
+        (mode == "affine") ? metal::quantized() : metal::fp4_quantized(),
+        template_def);
+    return kernel_source;
   });
   return d.get_kernel(kernel_name, lib);
 }
@@ -823,6 +844,7 @@ MTL::ComputePipelineState* get_gather_qmm_kernel(
     const array& x,
     int group_size,
     int bits,
+    const std::string& mode,
     int bm,
     int bn,
     int bk,
@@ -833,22 +855,40 @@ MTL::ComputePipelineState* get_gather_qmm_kernel(
   auto lib = d.get_library(lib_name, [&]() {
     std::string kernel_source;
     concatenate(
-        kernel_source,
-        metal::utils(),
-        metal::gemm(),
-        metal::quantized(),
-        get_template_definition(
-            lib_name,
-            "gather_qmm_rhs",
-            get_type_string(x.dtype()),
-            group_size,
-            bits,
-            bm,
-            bn,
-            bk,
-            wm,
-            wn,
-            transpose));
+        kernel_source, metal::utils(), metal::quantized_utils(), metal::gemm());
+    if (mode == "affine") {
+      concatenate(
+          kernel_source,
+          metal::quantized(),
+          get_template_definition(
+              lib_name,
+              mode + "_gather_qmm_rhs",
+              get_type_string(x.dtype()),
+              group_size,
+              bits,
+              bm,
+              bn,
+              bk,
+              wm,
+              wn,
+              transpose));
+    } else {
+      concatenate(
+          kernel_source,
+          metal::fp4_quantized(),
+          get_template_definition(
+              lib_name,
+              mode + "_gather_qmm_rhs",
+              get_type_string(x.dtype()),
+              group_size,
+              "uint8_t",
+              bm,
+              bn,
+              bk,
+              wm,
+              wn,
+              transpose));
+    }
     return kernel_source;
   });
   return d.get_kernel(kernel_name, lib, hash_name, func_consts);

mlx/backend/metal/kernels.h

@@ -238,7 +238,8 @@ MTL::ComputePipelineState* get_fft_kernel(
 MTL::ComputePipelineState* get_quantized_kernel(
     metal::Device& d,
     const std::string& kernel_name,
-    const std::string& template_def);
+    const std::string& template_def,
+    const std::string& mode);
 
 MTL::ComputePipelineState* get_gather_qmm_kernel(
     metal::Device& d,
@@ -248,6 +249,7 @@ MTL::ComputePipelineState* get_gather_qmm_kernel(
     const array& x,
     int group_size,
     int bits,
+    const std::string& mode,
     int bm,
     int bn,
     int bk,

mlx/backend/metal/kernels/CMakeLists.txt

@@ -108,7 +108,8 @@ if(NOT MLX_METAL_JIT)
     reduction/reduce_all.h
     reduction/reduce_col.h
     reduction/reduce_row.h)
-  build_kernel(quantized quantized.h ${STEEL_HEADERS})
+  build_kernel(quantized quantized.h quantized_utils.h ${STEEL_HEADERS})
+  build_kernel(fp4_quantized fp4_quantized.h quantized_utils.h ${STEEL_HEADERS})
   build_kernel(scan scan.h)
   build_kernel(softmax softmax.h)
   build_kernel(logsumexp logsumexp.h)

mlx/backend/metal/kernels/binary_ops.h

@@ -223,6 +223,11 @@ struct Power {
   template <typename T>
   metal::enable_if_t<metal::is_integral_v<T>, T> operator()(T base, T exp) {
     T res = 1;
+    // Undefined to raise integer to negative power
+    if (exp < 0) {
+      return 0;
+    }
+
     while (exp) {
       if (exp & 1) {
         res *= base;

mlx/backend/metal/kernels/complex.h

@@ -104,6 +104,27 @@ constexpr bool operator==(complex64_t a, complex64_t b) {
 constexpr complex64_t operator+(complex64_t a, complex64_t b) {
   return {a.real + b.real, a.imag + b.imag};
 }
+
+constexpr thread complex64_t& operator+=(thread complex64_t& a, complex64_t b) {
+  a.real += b.real;
+  a.imag += b.imag;
+  return a;
+}
+
+constexpr threadgroup complex64_t& operator+=(
+    threadgroup complex64_t& a,
+    complex64_t b) {
+  a.real += b.real;
+  a.imag += b.imag;
+  return a;
+}
+
+constexpr device complex64_t& operator+=(device complex64_t& a, complex64_t b) {
+  a.real += b.real;
+  a.imag += b.imag;
+  return a;
+}
+
 constexpr complex64_t operator+(float a, complex64_t b) {
   return {a + b.real, b.imag};
 }

mlx/backend/metal/kernels/conv.metal

@@ -288,6 +288,40 @@ instantiate_depthconv2d(float32, float);
 instantiate_depthconv2d(float16, half);
 instantiate_depthconv2d(bfloat16, bfloat16_t);
 
+template <typename T, typename IdxT>
+[[kernel]] void depthwise_conv_1d(
+    const device T* in [[buffer(0)]],
+    const device T* w [[buffer(1)]],
+    device T* out [[buffer(2)]],
+    constant const IdxT strides[3],
+    constant const int& kernel_size,
+    uint3 tid [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]) {
+  out += (tid.z * static_cast<IdxT>(grid_dim.y) + tid.y) * grid_dim.x + tid.x;
+  in += tid.z * strides[0] + tid.y * strides[1] + tid.x * strides[2];
+  w += tid.x * kernel_size;
+
+  float acc = 0.0;
+  for (int i = 0; i < kernel_size; ++i) {
+    acc += static_cast<float>(in[0]) * w[i];
+    in += strides[1];
+  }
+  *out = static_cast<T>(acc);
+}
+
+#define instantiate_depthconv1d(iname, itype)                         \
+  instantiate_kernel(                                                 \
+      "depthwise_conv_1d_" #iname, depthwise_conv_1d, itype, int32_t) \
+      instantiate_kernel(                                             \
+          "depthwise_conv_1d_" #iname "_large",                       \
+          depthwise_conv_1d,                                          \
+          itype,                                                      \
+          int64_t)
+
+instantiate_depthconv1d(float32, float);
+instantiate_depthconv1d(float16, half);
+instantiate_depthconv1d(bfloat16, bfloat16_t);
+
 ///////////////////////////////////////////////////////////////////////////////
 /// Winograd kernels
 ///////////////////////////////////////////////////////////////////////////////

mlx/backend/metal/kernels/fp4_quantized.metal

@@ -0,0 +1,127 @@
+// 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/fp4_quantized.h"
+
+#define instantiate_quantized(name, type) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4", \
+      name, \
+      type, \
+      32,   \
+      uint8_t)
+
+#define instantiate_quantized_batched(name, type, batched) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_batch_" #batched, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      batched)
+
+#define instantiate_quantized_aligned(name, type, aligned) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_alN_" #aligned, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      aligned)
+
+#define instantiate_quantized_aligned_batched(name, type, aligned, batched) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_alN_" #aligned "_batch_" #batched, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      aligned, \
+      batched)
+
+#define instantiate_quantized_quad(name, type, D, batched) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_d_" #D "_batch_" #batched, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      D,       \
+      batched)
+
+#define instantiate_quantized_split_k(name, type, split_k) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_spk_" #split_k, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      split_k)
+
+#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,      \
+      uint8_t, \
+      bm,      \
+      bn,      \
+      bk,      \
+      wm,      \
+      wn,      \
+      transpose)
+
+#define instantiate_quantized_batched_wrap(name, type) \
+  instantiate_quantized_batched(name, type, 1)         \
+  instantiate_quantized_batched(name, type, 0)
+
+#define instantiate_quantized_all_batched(type) \
+  instantiate_quantized_batched_wrap(mxfp4_qmv_fast, type) \
+  instantiate_quantized_batched_wrap(mxfp4_qmv, type)      \
+  instantiate_quantized_batched_wrap(mxfp4_qvm, type)      \
+  instantiate_quantized_batched_wrap(mxfp4_qmm_n, type)
+
+#define instantiate_quantized_all_single(type) \
+  instantiate_quantized(mxfp4_gather_qmv_fast, type) \
+  instantiate_quantized(mxfp4_gather_qmv, type)      \
+  instantiate_quantized(mxfp4_gather_qvm, type)      \
+  instantiate_quantized(mxfp4_gather_qmm_n, type)
+
+#define instantiate_quantized_all_aligned(type) \
+  instantiate_quantized_aligned(mxfp4_gather_qmm_t, type, true)      \
+  instantiate_quantized_aligned(mxfp4_gather_qmm_t, type, false)     \
+  instantiate_quantized_aligned_batched(mxfp4_qmm_t, type, true, 1)  \
+  instantiate_quantized_aligned_batched(mxfp4_qmm_t, type, true, 0)  \
+  instantiate_quantized_aligned_batched(mxfp4_qmm_t, type, false, 1) \
+  instantiate_quantized_aligned_batched(mxfp4_qmm_t, type, false, 0)
+
+#define instantiate_quantized_all_quad(type) \
+  instantiate_quantized_quad(mxfp4_qmv_quad, type, 64, 1)  \
+  instantiate_quantized_quad(mxfp4_qmv_quad, type, 64, 0)  \
+  instantiate_quantized_quad(mxfp4_qmv_quad, type, 128, 1) \
+  instantiate_quantized_quad(mxfp4_qmv_quad, type, 128, 0)
+
+#define instantiate_quantized_all_splitk(type) \
+  instantiate_quantized_split_k(mxfp4_qvm_split_k, type, 8) \
+  instantiate_quantized_split_k(mxfp4_qvm_split_k, type, 32)
+
+#define instantiate_quantized_all_rhs(type) \
+  instantiate_gather_qmm_rhs(mxfp4_gather_qmm_rhs, mxfp4_gather_qmm_rhs_nt, type, 16, 32, 32, 1, 2, true) \
+  instantiate_gather_qmm_rhs(mxfp4_gather_qmm_rhs, mxfp4_gather_qmm_rhs_nn, type, 16, 32, 32, 1, 2, false)
+
+#define instantiate_quantized_types(type) \
+  instantiate_quantized_all_batched(type) \
+  instantiate_quantized_all_quad(type)    \
+  instantiate_quantized_all_splitk(type)  \
+  instantiate_quantized_all_single(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/gemv.metal

@@ -15,6 +15,15 @@ using namespace metal;
 
 #define MLX_MTL_CONST static constant constexpr const
 
+template <typename U>
+struct DefaultAccT {
+  using type = float;
+};
+template <>
+struct DefaultAccT<complex64_t> {
+  using type = complex64_t;
+};
+
 template <
     typename T,
     const int BM, /* Threadgroup rows (in simdgroups) */
@@ -24,8 +33,10 @@ template <
     const int TM, /* Thread rows (in elements) */
     const int TN, /* Thread cols (in elements) */
     const bool kDoAxpby, /* Do out = alpha * out + beta * bias */
-    typename AccT = float>
+    typename AccT = typename DefaultAccT<T>::type>
 struct GEMVKernel {
+  using acc_type = AccT;
+
   MLX_MTL_CONST int threadsM = BM * SM;
   MLX_MTL_CONST int threadsN = BN * SN;
 
@@ -35,8 +46,8 @@ struct GEMVKernel {
   static_assert(SM * SN == 32, "simdgroup can only have 32 threads");
 
   static_assert(
-      SN == 8 || SN == 16 || SN == 32,
-      "gemv block must have a width of 8, 16, or 32");
+      SN == 4 || SN == 8 || SN == 16 || SN == 32,
+      "gemv block must have a width of 4, 8, 16, or 32");
 
   // - The matrix of size (M = out_vec_size, K = in_vec_size) is divided up
   //   into blocks of (blockM, blockN) divided among threadgroups
@@ -246,8 +257,10 @@ template <
     const int TM, /* Thread rows (in elements) */
     const int TN, /* Thread cols (in elements) */
     const bool kDoAxpby, /* Do out = alpha * out + beta * bias */
-    typename AccT = float>
+    typename AccT = typename DefaultAccT<T>::type>
 struct GEMVTKernel {
+  using acc_type = AccT;
+
   MLX_MTL_CONST int threadsM = BM * SM;
   MLX_MTL_CONST int threadsN = BN * SN;
 
@@ -453,7 +466,7 @@ template <
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
   using gemv_kernel = GEMVKernel<T, BM, BN, SM, SN, TM, TN, kDoAxpby>;
-  threadgroup float tgp_memory
+  threadgroup typename gemv_kernel::acc_type tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   // Update batch offsets
@@ -511,21 +524,26 @@ template <
       axpby)
 
 // clang-format off
-#define instantiate_gemv(name, itype, bm, bn, tm, tn)              \
-  instantiate_gemv_helper(name, itype, bm, 1, 1, bn, tm, tn, 0, 0) \
-  instantiate_gemv_helper(name, itype, bm, 1, 1, bn, tm, tn, 0, 1) \
-  instantiate_gemv_helper(name, itype, bm, 1, 1, bn, tm, tn, 1, 0) \
-  instantiate_gemv_helper(name, itype, bm, 1, 1, bn, tm, tn, 1, 1) // clang-format on
+#define instantiate_gemv(name, itype, bm, bn, sm, sn, tm, tn)        \
+  instantiate_gemv_helper(name, itype, bm, bn, sm, sn, tm, tn, 0, 0) \
+  instantiate_gemv_helper(name, itype, bm, bn, sm, sn, tm, tn, 0, 1) \
+  instantiate_gemv_helper(name, itype, bm, bn, sm, sn, tm, tn, 1, 0) \
+  instantiate_gemv_helper(name, itype, bm, bn, sm, sn, tm, tn, 1, 1) // clang-format on
 
 // clang-format off
 #define instantiate_gemv_blocks(name, itype) \
-  instantiate_gemv(name, itype, 4, 32, 1, 4) \
-  instantiate_gemv(name, itype, 4, 32, 4, 4) \
-  instantiate_gemv(name, itype, 8, 32, 4, 4) // clang-format on
+  instantiate_gemv(name, itype, 1,  8, 1, 32, 4, 4) \
+  instantiate_gemv(name, itype, 1,  8, 1, 32, 1, 4) \
+  instantiate_gemv(name, itype, 1,  1, 8,  4, 4, 4) \
+  instantiate_gemv(name, itype, 1,  1, 8,  4, 1, 4) \
+  instantiate_gemv(name, itype, 4,  1, 1, 32, 1, 4) \
+  instantiate_gemv(name, itype, 4,  1, 1, 32, 4, 4) \
+  instantiate_gemv(name, itype, 8,  1, 1, 32, 4, 4) // clang-format on
 
 instantiate_gemv_blocks(float32, float);
 instantiate_gemv_blocks(float16, half);
 instantiate_gemv_blocks(bfloat16, bfloat16_t);
+instantiate_gemv_blocks(complex64, complex64_t);
 
 template <
     typename T,
@@ -561,7 +579,7 @@ template <
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
   using gemv_kernel = GEMVKernel<T, BM, BN, SM, SN, TM, TN, false>;
-  threadgroup float tgp_memory
+  threadgroup typename gemv_kernel::acc_type tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   uint32_t indx_vec;
@@ -632,6 +650,7 @@ template <
 instantiate_gemv_bs_blocks(float32, float);
 instantiate_gemv_bs_blocks(float16, half);
 instantiate_gemv_bs_blocks(bfloat16, bfloat16_t);
+instantiate_gemv_bs_blocks(complex64, complex64_t);
 
 ///////////////////////////////////////////////////////////////////////////////
 /// Vector matrix multiplication
@@ -668,7 +687,7 @@ template <
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
   using gemv_kernel = GEMVTKernel<T, BM, BN, SM, SN, TM, TN, kDoAxpby>;
-  threadgroup float tgp_memory
+  threadgroup typename gemv_kernel::acc_type tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   // Update batch offsets
@@ -734,7 +753,8 @@ template <
 // clang-format off
 instantiate_gemv_t_blocks(float32, float);
 instantiate_gemv_t_blocks(float16, half);
-instantiate_gemv_t_blocks(bfloat16, bfloat16_t); // clang-format on
+instantiate_gemv_t_blocks(bfloat16, bfloat16_t);
+instantiate_gemv_t_blocks(complex64, complex64_t); // clang-format on
 
 template <
     typename T,
@@ -769,8 +789,8 @@ template <
     uint3 lid [[thread_position_in_threadgroup]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
-  using gemv_kernel = GEMVTKernel<T, BM, BN, SM, SN, TM, TN, false, float>;
-  threadgroup float tgp_memory
+  using gemv_kernel = GEMVTKernel<T, BM, BN, SM, SN, TM, TN, false>;
+  threadgroup typename gemv_kernel::acc_type tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   uint32_t indx_vec;
@@ -844,4 +864,5 @@ template <
 // clang-format off
 instantiate_gemv_t_bs_blocks(float32, float);
 instantiate_gemv_t_bs_blocks(float16, half);
-instantiate_gemv_t_bs_blocks(bfloat16, bfloat16_t); // clang-format on
+instantiate_gemv_t_bs_blocks(bfloat16, bfloat16_t);
+instantiate_gemv_t_bs_blocks(complex64, complex64_t); // clang-format on

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

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "mlx/backend/metal/kernels/indexing.h"
+#include "mlx/backend/metal/kernels/indexing/indexing.h"
 
 template <typename T, typename IdxT, int NIDX, int IDX_NDIM, typename LocT>
 METAL_FUNC void gather_impl(

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

@@ -0,0 +1,24 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/metal/kernels/indexing/indexing.h"
+
+template <typename T, typename IdxT, typename LocT, int N>
+[[kernel]] void gather_front(
+    const device T* src,
+    const device IdxT* indices,
+    device T* out,
+    const constant int64_t& stride,
+    const constant int& size,
+    uint2 index [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]) {
+  auto idx = offset_neg_idx(indices[index.y], size);
+  LocT src_idx = static_cast<LocT>(stride) * idx;
+  LocT out_idx = static_cast<LocT>(stride) * index.y;
+
+  int s_idx = N * index.x;
+  for (int i = 0; i < N && s_idx < stride; ++i, ++s_idx) {
+    out[out_idx + s_idx] = src[src_idx + s_idx];
+  }
+}

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

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "mlx/backend/metal/kernels/indexing.h"
+#include "mlx/backend/metal/kernels/indexing/indexing.h"
 
 template <
     typename T,

mlx/backend/metal/kernels/quantized.h

@@ -1434,7 +1434,7 @@ METAL_FUNC void adjust_matrix_offsets(
 }
 
 template <typename T, int group_size, int bits, int D, bool batched>
-[[kernel]] void qmv_quad(
+[[kernel]] void affine_qmv_quad(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1486,7 +1486,7 @@ template <typename T, int group_size, int bits, int D, bool batched>
 }
 
 template <typename T, int group_size, int bits, bool batched>
-[[kernel]] void qmv_fast(
+[[kernel]] void affine_qmv_fast(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1538,7 +1538,7 @@ template <typename T, int group_size, int bits, bool batched>
 }
 
 template <typename T, const int group_size, const int bits, bool batched>
-[[kernel]] void qmv(
+[[kernel]] void affine_qmv(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1590,7 +1590,7 @@ template <typename T, const int group_size, const int bits, bool batched>
 }
 
 template <typename T, const int group_size, const int bits, bool batched>
-[[kernel]] void qvm(
+[[kernel]] void affine_qvm(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1642,7 +1642,7 @@ template <typename T, const int group_size, const int bits, bool batched>
 }
 
 template <typename T, const int group_size, const int bits, int split_k = 32>
-[[kernel]] void qvm_split_k(
+[[kernel]] void affine_qvm_split_k(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1706,7 +1706,7 @@ template <
     const int BM = 32,
     const int BK = 32,
     const int BN = 32>
-[[kernel]] void qmm_t(
+[[kernel]] void affine_qmm_t(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1764,7 +1764,7 @@ template <
     const int BM = 32,
     const int BK = 32,
     const int BN = 32>
-[[kernel]] void qmm_n(
+[[kernel]] void affine_qmm_n(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1817,7 +1817,7 @@ template <
 }
 
 template <typename T, int group_size, int bits>
-[[kernel]] void gather_qmv_fast(
+[[kernel]] void affine_gather_qmv_fast(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1879,7 +1879,7 @@ template <typename T, int group_size, int bits>
 }
 
 template <typename T, int group_size, int bits>
-[[kernel]] void gather_qmv(
+[[kernel]] void affine_gather_qmv(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1941,7 +1941,7 @@ template <typename T, int group_size, int bits>
 }
 
 template <typename T, int group_size, int bits>
-[[kernel]] void gather_qvm(
+[[kernel]] void affine_gather_qvm(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -2010,7 +2010,7 @@ template <
     const int BM = 32,
     const int BK = 32,
     const int BN = 32>
-[[kernel]] void gather_qmm_t(
+[[kernel]] void affine_gather_qmm_t(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -2077,7 +2077,7 @@ template <
     const int BM = 32,
     const int BK = 32,
     const int BN = 32>
-[[kernel]] void gather_qmm_n(
+[[kernel]] void affine_gather_qmm_n(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -2138,92 +2138,6 @@ template <
       w, scales, biases, x, y, Xs, Ws, K, N, M, tid, lid, simd_gid, simd_lid);
 }
 
-template <typename T, typename mma_t, typename loader_a_t, typename loader_b_t>
-METAL_FUNC void gemm_loop_aligned(
-    threadgroup T* As,
-    threadgroup T* Bs,
-    thread mma_t& mma_op,
-    thread loader_a_t& loader_a,
-    thread loader_b_t& loader_b,
-    const int k_iterations) {
-  for (int k = 0; k < k_iterations; k++) {
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Load elements into threadgroup memory
-    loader_a.load_unsafe();
-    loader_b.load_unsafe();
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Multiply and accumulate threadgroup elements
-    mma_op.mma(As, Bs);
-
-    // Prepare for next iteration
-    loader_a.next();
-    loader_b.next();
-  }
-}
-
-template <
-    bool rows_aligned,
-    bool cols_aligned,
-    bool transpose,
-    typename T,
-    typename mma_t,
-    typename loader_a_t,
-    typename loader_b_t>
-METAL_FUNC void gemm_loop_unaligned(
-    threadgroup T* As,
-    threadgroup T* Bs,
-    thread mma_t& mma_op,
-    thread loader_a_t& loader_a,
-    thread loader_b_t& loader_b,
-    const int k_iterations,
-    const short tgp_bm,
-    const short tgp_bn,
-    const short tgp_bk) {
-  for (int k = 0; k < k_iterations; k++) {
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Load elements into threadgroup memory
-    if (rows_aligned) {
-      loader_a.load_unsafe();
-    } else {
-      loader_a.load_safe(short2(tgp_bk, tgp_bm));
-    }
-    if (cols_aligned) {
-      loader_b.load_unsafe();
-    } else {
-      loader_b.load_safe(
-          transpose ? short2(tgp_bk, tgp_bn) : short2(tgp_bn, tgp_bk));
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Multiply and accumulate threadgroup elements
-    mma_op.mma(As, Bs);
-
-    // Prepare for next iteration
-    loader_a.next();
-    loader_b.next();
-  }
-}
-
-template <typename T, typename mma_t, typename loader_a_t, typename loader_b_t>
-METAL_FUNC void gemm_loop_finalize(
-    threadgroup T* As,
-    threadgroup T* Bs,
-    thread mma_t& mma_op,
-    thread loader_a_t& loader_a,
-    thread loader_b_t& loader_b,
-    const short2 tile_a,
-    const short2 tile_b) {
-  loader_a.load_safe(tile_a);
-  loader_b.load_safe(tile_b);
-  threadgroup_barrier(mem_flags::mem_threadgroup);
-  mma_op.mma(As, Bs);
-}
-
 template <
     typename T,
     int group_size,
@@ -2234,7 +2148,7 @@ template <
     int WM,
     int WN,
     bool transpose>
-[[kernel]] void gather_qmm_rhs(
+[[kernel]] void affine_gather_qmm_rhs(
     const device T* x [[buffer(0)]],
     const device uint32_t* w [[buffer(1)]],
     const device T* scales [[buffer(2)]],

mlx/backend/metal/kernels/quantized.metal

@@ -3,6 +3,7 @@
 // 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/quantized.h"
 
 #define instantiate_quantized(name, type, group_size, bits)     \
@@ -79,40 +80,40 @@
   instantiate_quantized_batched(name, type, group_size, bits, 0)
 
 #define instantiate_quantized_all_batched(type, group_size, bits) \
-  instantiate_quantized_batched_wrap(qmv_fast, type, group_size, bits)     \
-  instantiate_quantized_batched_wrap(qmv, type, group_size, bits)     \
-  instantiate_quantized_batched_wrap(qvm, type, group_size, bits)     \
-  instantiate_quantized_batched_wrap(qmm_n, type, group_size, bits)
+  instantiate_quantized_batched_wrap(affine_qmv_fast, type, group_size, bits)     \
+  instantiate_quantized_batched_wrap(affine_qmv, type, group_size, bits)     \
+  instantiate_quantized_batched_wrap(affine_qvm, type, group_size, bits)     \
+  instantiate_quantized_batched_wrap(affine_qmm_n, type, group_size, bits)
 
 #define instantiate_quantized_all_single(type, group_size, bits) \
   instantiate_quantized(affine_quantize, type, group_size, bits) \
   instantiate_quantized(affine_dequantize, type, group_size, bits)     \
-  instantiate_quantized(gather_qmv_fast, type, group_size, bits)     \
-  instantiate_quantized(gather_qmv, type, group_size, bits)     \
-  instantiate_quantized(gather_qvm, type, group_size, bits)     \
-  instantiate_quantized(gather_qmm_n, type, group_size, bits)
+  instantiate_quantized(affine_gather_qmv_fast, type, group_size, bits)     \
+  instantiate_quantized(affine_gather_qmv, type, group_size, bits)     \
+  instantiate_quantized(affine_gather_qvm, type, group_size, bits)     \
+  instantiate_quantized(affine_gather_qmm_n, type, group_size, bits)
 
 #define instantiate_quantized_all_aligned(type, group_size, bits)   \
-  instantiate_quantized_aligned(gather_qmm_t, type, group_size, bits, true) \
-  instantiate_quantized_aligned(gather_qmm_t, type, group_size, bits, false) \
-  instantiate_quantized_aligned_batched(qmm_t, type, group_size, bits, true, 1) \
-  instantiate_quantized_aligned_batched(qmm_t, type, group_size, bits, true, 0) \
-  instantiate_quantized_aligned_batched(qmm_t, type, group_size, bits, false, 1) \
-  instantiate_quantized_aligned_batched(qmm_t, type, group_size, bits, false, 0)
+  instantiate_quantized_aligned(affine_gather_qmm_t, type, group_size, bits, true) \
+  instantiate_quantized_aligned(affine_gather_qmm_t, type, group_size, bits, false) \
+  instantiate_quantized_aligned_batched(affine_qmm_t, type, group_size, bits, true, 1) \
+  instantiate_quantized_aligned_batched(affine_qmm_t, type, group_size, bits, true, 0) \
+  instantiate_quantized_aligned_batched(affine_qmm_t, type, group_size, bits, false, 1) \
+  instantiate_quantized_aligned_batched(affine_qmm_t, type, group_size, bits, false, 0)
 
 #define instantiate_quantized_all_quad(type, group_size, bits)   \
-  instantiate_quantized_quad(qmv_quad, type, group_size, bits, 64, 1)   \
-  instantiate_quantized_quad(qmv_quad, type, group_size, bits, 64, 0)  \
-  instantiate_quantized_quad(qmv_quad, type, group_size, bits, 128, 1)  \
-  instantiate_quantized_quad(qmv_quad, type, group_size, bits, 128, 0)
+  instantiate_quantized_quad(affine_qmv_quad, type, group_size, bits, 64, 1)   \
+  instantiate_quantized_quad(affine_qmv_quad, type, group_size, bits, 64, 0)   \
+  instantiate_quantized_quad(affine_qmv_quad, type, group_size, bits, 128, 1)  \
+  instantiate_quantized_quad(affine_qmv_quad, type, group_size, bits, 128, 0)
 
 #define instantiate_quantized_all_splitk(type, group_size, bits)   \
-  instantiate_quantized_split_k(qvm_split_k, type, group_size, bits, 8)   \
-  instantiate_quantized_split_k(qvm_split_k, type, group_size, bits, 32)
+  instantiate_quantized_split_k(affine_qvm_split_k, type, group_size, bits, 8)   \
+  instantiate_quantized_split_k(affine_qvm_split_k, type, group_size, bits, 32)
 
 #define instantiate_quantized_all_rhs(type, group_size, bits) \
-  instantiate_gather_qmm_rhs(gather_qmm_rhs, gather_qmm_rhs_nt, type, group_size, bits, 16, 32, 32, 1, 2, true) \
-  instantiate_gather_qmm_rhs(gather_qmm_rhs, gather_qmm_rhs_nn, type, group_size, bits, 16, 32, 32, 1, 2, false)
+  instantiate_gather_qmm_rhs(affine_gather_qmm_rhs, affine_gather_qmm_rhs_nt, type, group_size, bits, 16, 32, 32, 1, 2, true) \
+  instantiate_gather_qmm_rhs(affine_gather_qmm_rhs, affine_gather_qmm_rhs_nn, type, group_size, bits, 16, 32, 32, 1, 2, false)
 
 #define instantiate_quantized_funcs(type, group_size, bits) \
   instantiate_quantized_all_single(type, group_size, bits)  \

mlx/backend/metal/kernels/quantized_utils.h

@@ -0,0 +1,90 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <metal_simdgroup>
+#include <metal_stdlib>
+
+template <typename T, typename mma_t, typename loader_a_t, typename loader_b_t>
+METAL_FUNC void gemm_loop_aligned(
+    threadgroup T* As,
+    threadgroup T* Bs,
+    thread mma_t& mma_op,
+    thread loader_a_t& loader_a,
+    thread loader_b_t& loader_b,
+    const int k_iterations) {
+  for (int k = 0; k < k_iterations; k++) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Load elements into threadgroup memory
+    loader_a.load_unsafe();
+    loader_b.load_unsafe();
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Multiply and accumulate threadgroup elements
+    mma_op.mma(As, Bs);
+
+    // Prepare for next iteration
+    loader_a.next();
+    loader_b.next();
+  }
+}
+
+template <
+    bool rows_aligned,
+    bool cols_aligned,
+    bool transpose,
+    typename T,
+    typename mma_t,
+    typename loader_a_t,
+    typename loader_b_t>
+METAL_FUNC void gemm_loop_unaligned(
+    threadgroup T* As,
+    threadgroup T* Bs,
+    thread mma_t& mma_op,
+    thread loader_a_t& loader_a,
+    thread loader_b_t& loader_b,
+    const int k_iterations,
+    const short tgp_bm,
+    const short tgp_bn,
+    const short tgp_bk) {
+  for (int k = 0; k < k_iterations; k++) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Load elements into threadgroup memory
+    if (rows_aligned) {
+      loader_a.load_unsafe();
+    } else {
+      loader_a.load_safe(short2(tgp_bk, tgp_bm));
+    }
+    if (cols_aligned) {
+      loader_b.load_unsafe();
+    } else {
+      loader_b.load_safe(
+          transpose ? short2(tgp_bk, tgp_bn) : short2(tgp_bn, tgp_bk));
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Multiply and accumulate threadgroup elements
+    mma_op.mma(As, Bs);
+
+    // Prepare for next iteration
+    loader_a.next();
+    loader_b.next();
+  }
+}
+
+template <typename T, typename mma_t, typename loader_a_t, typename loader_b_t>
+METAL_FUNC void gemm_loop_finalize(
+    threadgroup T* As,
+    threadgroup T* Bs,
+    thread mma_t& mma_op,
+    thread loader_a_t& loader_a,
+    thread loader_b_t& loader_b,
+    const short2 tile_a,
+    const short2 tile_b) {
+  loader_a.load_safe(tile_a);
+  loader_b.load_safe(tile_b);
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  mma_op.mma(As, Bs);
+}

mlx/backend/metal/kernels/rope.metal

@@ -3,14 +3,19 @@
 #include <metal_math>
 
 #include "mlx/backend/metal/kernels/utils.h"
-template <typename T, bool traditional, bool forward>
+
+constant bool forward [[function_constant(1)]];
+constant bool traditional [[function_constant(2)]];
+constant bool hs_transpose [[function_constant(3)]];
+
+template <typename T>
 void rope_single_impl(
     const device T* in,
     device T* out,
     constant const int& offset,
     const float inv_freq,
     constant const float& scale,
-    constant const size_t& stride,
+    constant const int64_t& stride,
     uint2 pos,
     uint2 grid) {
   float L = scale * static_cast<float>(offset);
@@ -46,76 +51,85 @@ void rope_single_impl(
   out[index_2] = static_cast<T>(rx2);
 }
 
-template <typename T, bool traditional, bool forward>
+template <typename T>
 [[kernel]] void rope_single(
     const device T* in [[buffer(0)]],
     device T* out [[buffer(1)]],
     constant const int& offset,
     constant const float& scale,
-    constant const size_t& stride,
+    constant const int64_t& stride,
     constant const float& base [[buffer(10)]],
     uint2 pos [[thread_position_in_grid]],
     uint2 grid [[threads_per_grid]]) {
   float d = static_cast<float>(pos.x) / static_cast<float>(grid.x);
   float inv_freq = metal::exp2(-d * base);
-  rope_single_impl<T, traditional, forward>(
-      in, out, offset, inv_freq, scale, stride, pos, grid);
+  rope_single_impl<T>(in, out, offset, inv_freq, scale, stride, pos, grid);
 }
 
-template <typename T, bool traditional, bool forward>
+template <typename T>
 [[kernel]] void rope_single_freqs(
     const device T* in [[buffer(0)]],
     device T* out [[buffer(1)]],
     constant const int& offset,
     constant const float& scale,
-    constant const size_t& stride,
+    constant const int64_t& stride,
     const device float* freqs [[buffer(10)]],
-    constant const size_t& freq_stride [[buffer(11)]],
+    constant const int64_t& freq_stride [[buffer(11)]],
     uint2 pos [[thread_position_in_grid]],
     uint2 grid [[threads_per_grid]]) {
   float inv_freq = 1.0 / (freqs[freq_stride * pos.x]);
-  rope_single_impl<T, traditional, forward>(
-      in, out, offset, inv_freq, scale, stride, pos, grid);
+  rope_single_impl<T>(in, out, offset, inv_freq, scale, stride, pos, grid);
 }
 
-template <typename T, bool traditional, bool forward, int N = 4>
+template <typename T, typename IdxT, int N = 4>
 void rope_impl(
     const device T* in,
     device T* out,
-    constant const int& offset,
+    const device int* offset,
     const float inv_freq,
     constant const float& scale,
-    constant const size_t strides[3],
-    constant const size_t out_strides[3],
-    constant const size_t& n_batch,
+    constant const int64_t strides[3],
+    constant const int64_t out_strides[3],
+    constant const int64_t& offset_stride,
+    constant const int& n_head,
     uint3 pos,
     uint3 grid) {
-  float L = scale * static_cast<float>(pos.y + offset);
+  auto n_head_up = N * ((n_head + N - 1) / N);
+  auto head_idx = static_cast<int>((pos.z * N) % n_head_up);
+  auto batch_idx = (pos.z * N) / n_head_up;
+  auto batch_offset = offset[batch_idx * offset_stride];
+  float L = scale * static_cast<float>(pos.y + batch_offset);
+  auto mat_idx = batch_idx * n_head + head_idx;
 
   // Compute costheta, sintheta
   float theta = L * inv_freq;
   float costheta = metal::fast::cos(theta);
   float sintheta = metal::fast::sin(theta);
-
   // Compute the input and output indices
-  size_t in_index_1, in_index_2;
-  size_t out_index_1, out_index_2;
-  if (traditional) {
-    out_index_1 = 2 * pos.x * out_strides[2] + pos.y * out_strides[1] +
-        N * pos.z * out_strides[0];
-    out_index_2 = out_index_1 + 1;
+  IdxT in_index_1;
+  if (hs_transpose) {
+    IdxT batch_stride = grid.y * IdxT(strides[1]);
     in_index_1 =
-        2 * pos.x * strides[2] + pos.y * strides[1] + N * pos.z * strides[0];
-    in_index_2 = in_index_1 + strides[2];
+        batch_idx * batch_stride + pos.y * strides[1] + head_idx * strides[0];
   } else {
-    out_index_1 = pos.x * out_strides[2] + pos.y * out_strides[1] +
-        N * pos.z * out_strides[0];
-    out_index_2 = out_index_1 + grid.x * out_strides[2];
-    in_index_1 =
-        pos.x * strides[2] + pos.y * strides[1] + N * pos.z * strides[0];
-    in_index_2 = in_index_1 + grid.x * strides[2];
+    in_index_1 = pos.y * IdxT(strides[1]) + mat_idx * IdxT(strides[0]);
   }
-  for (int i = 0; i < N && pos.z * N + i < n_batch; ++i) {
+  IdxT in_index_2;
+  IdxT out_index_1 =
+      pos.y * IdxT(out_strides[1]) + mat_idx * IdxT(out_strides[0]);
+  IdxT out_index_2;
+  if (traditional) {
+    out_index_1 += 2 * pos.x * IdxT(out_strides[2]);
+    out_index_2 = out_index_1 + 1;
+    in_index_1 += 2 * pos.x * IdxT(strides[2]);
+    in_index_2 = in_index_1 + IdxT(strides[2]);
+  } else {
+    out_index_1 += pos.x * IdxT(out_strides[2]);
+    out_index_2 = out_index_1 + grid.x * IdxT(out_strides[2]);
+    in_index_1 += pos.x * IdxT(strides[2]);
+    in_index_2 = in_index_1 + grid.x * IdxT(strides[2]);
+  }
+  for (int i = 0; i < N && head_idx + i < n_head; ++i) {
     // Read and write the output
     float x1 = static_cast<float>(in[in_index_1]);
     float x2 = static_cast<float>(in[in_index_2]);
@@ -130,28 +144,29 @@ void rope_impl(
     }
     out[out_index_1] = static_cast<T>(rx1);
     out[out_index_2] = static_cast<T>(rx2);
-    in_index_1 += strides[0];
-    in_index_2 += strides[0];
-    out_index_1 += out_strides[0];
-    out_index_2 += out_strides[0];
+    in_index_1 += IdxT(strides[0]);
+    in_index_2 += IdxT(strides[0]);
+    out_index_1 += IdxT(out_strides[0]);
+    out_index_2 += IdxT(out_strides[0]);
   }
 }
 
-template <typename T, bool traditional, bool forward, int N = 4>
+template <typename T, typename IdxT, int N = 4>
 [[kernel]] void rope(
     const device T* in [[buffer(0)]],
     device T* out [[buffer(1)]],
-    constant const int& offset,
+    const device int* offset,
     constant const float& scale,
-    constant const size_t strides[3],
-    constant const size_t out_strides[3],
-    constant const size_t& n_batch,
+    constant const int64_t strides[3],
+    constant const int64_t out_strides[3],
+    constant const int64_t& offset_stride,
+    constant const int& n_head,
     constant const float& base [[buffer(10)]],
     uint3 pos [[thread_position_in_grid]],
     uint3 grid [[threads_per_grid]]) {
   float d = static_cast<float>(pos.x) / static_cast<float>(grid.x);
   float inv_freq = metal::exp2(-d * base);
-  rope_impl<T, traditional, forward, N>(
+  rope_impl<T, IdxT, N>(
       in,
       out,
       offset,
@@ -159,26 +174,28 @@ template <typename T, bool traditional, bool forward, int N = 4>
       scale,
       strides,
       out_strides,
-      n_batch,
+      offset_stride,
+      n_head,
       pos,
       grid);
 }
 
-template <typename T, bool traditional, bool forward, int N = 4>
+template <typename T, typename IdxT, int N = 4>
 [[kernel]] void rope_freqs(
     const device T* in [[buffer(0)]],
     device T* out [[buffer(1)]],
-    constant const int& offset,
+    const device int* offset,
     constant const float& scale,
-    constant const size_t strides[3],
-    constant const size_t out_strides[3],
-    constant const size_t& n_batch,
+    constant const int64_t strides[3],
+    constant const int64_t out_strides[3],
+    constant const int64_t& offset_stride,
+    constant const int& n_head,
     const device float* freqs [[buffer(10)]],
-    constant const size_t& freq_stride [[buffer(11)]],
+    constant const int64_t& freq_stride [[buffer(11)]],
     uint3 pos [[thread_position_in_grid]],
     uint3 grid [[threads_per_grid]]) {
   float inv_freq = 1.0 / (freqs[freq_stride * pos.x]);
-  rope_impl<T, traditional, forward, N>(
+  rope_impl<T, IdxT, N>(
       in,
       out,
       offset,
@@ -186,75 +203,27 @@ template <typename T, bool traditional, bool forward, int N = 4>
       scale,
       strides,
       out_strides,
-      n_batch,
+      offset_stride,
+      n_head,
       pos,
       grid);
 }
 
 // clang-format off
-#define instantiate_rope_g(name, type, traditional, forward) \
-  template [[host_name("rope_" #name)]] [[kernel]] void      \
-  rope<type, traditional, forward>(                          \
-      const device type* in [[buffer(0)]],                   \
-      device type* out [[buffer(1)]],                        \
-      constant const int& offset,                            \
-      constant const float& scale,                           \
-      constant const size_t strides[3],                      \
-      constant const size_t out_strides[3],                  \
-      constant const size_t& n_batch,                        \
-      constant const float& base [[buffer(10)]],             \
-      uint3 pos [[thread_position_in_grid]],                 \
-      uint3 grid [[threads_per_grid]]);                      \
-  template [[host_name("rope_freqs_" #name)]]                \
-  [[kernel]] void rope_freqs<type, traditional, forward>(    \
-      const device type* in [[buffer(0)]],                   \
-      device type* out [[buffer(1)]],                        \
-      constant const int& offset,                            \
-      constant const float& scale,                           \
-      constant const size_t strides[3],                      \
-      constant const size_t out_strides[3],                  \
-      constant const size_t& n_batch,                        \
-      const device float* freqs [[buffer(10)]],              \
-      constant const size_t& freq_stride [[buffer(11)]],     \
-      uint3 pos [[thread_position_in_grid]],                 \
-      uint3 grid [[threads_per_grid]]);
+#define instantiate_rope_g(name, type) \
+  instantiate_kernel("rope_" #name, rope, type, int32_t) \
+  instantiate_kernel("rope_freqs_" #name, rope_freqs, type, int32_t) \
+  instantiate_kernel("rope_large_" #name, rope, type, int64_t) \
+  instantiate_kernel("rope_freqs_large_" #name, rope_freqs, type, int64_t)
 
-#define instantiate_rope_s(name, type, traditional, forward)     \
-  template [[host_name("rope_single_" #name)]] [[kernel]] void   \
-  rope_single<type, traditional, forward>(                       \
-      const device type* in [[buffer(0)]],                       \
-      device type* out [[buffer(1)]],                            \
-      constant const int& offset,                                \
-      constant const float& scale,                               \
-      constant const size_t& stride,                             \
-      constant const float& base [[buffer(10)]],                 \
-      uint2 pos [[thread_position_in_grid]],                     \
-      uint2 grid [[threads_per_grid]]);                          \
-  template [[host_name("rope_single_freqs_" #name)]]             \
-  [[kernel]] void rope_single_freqs<type, traditional, forward>( \
-      const device type* in [[buffer(0)]],                       \
-      device type* out [[buffer(1)]],                            \
-      constant const int& offset,                                \
-      constant const float& scale,                               \
-      constant const size_t& stride,                             \
-      const device float* freqs [[buffer(10)]],                  \
-      constant const size_t& freq_stride [[buffer(11)]],         \
-      uint2 pos [[thread_position_in_grid]],                     \
-      uint2 grid [[threads_per_grid]]);
+#define instantiate_rope_s(name, type) \
+  instantiate_kernel("rope_single_" #name, rope_single, type) \
+  instantiate_kernel("rope_single_freqs_" #name, rope_single_freqs, type)
 
-#define instantiate_rope(name, type, traditional, forward) \
-  instantiate_rope_s(name, type, traditional, forward)     \
-  instantiate_rope_g(name, type, traditional, forward)
+#define instantiate_rope(name, type) \
+  instantiate_rope_s(name, type)     \
+  instantiate_rope_g(name, type)
 
-instantiate_rope(traditional_float16, half, true, true)
-instantiate_rope(traditional_bfloat16, bfloat16_t, true, true)
-instantiate_rope(traditional_float32, float, true, true)
-instantiate_rope(float16, half, false, true)
-instantiate_rope(bfloat16, bfloat16_t, false, true)
-instantiate_rope(float32, float, false, true)
-instantiate_rope(vjp_traditional_float16, half, true, false)
-instantiate_rope(vjp_traditional_bfloat16, bfloat16_t, true, false)
-instantiate_rope(vjp_traditional_float32, float, true, false)
-instantiate_rope(vjp_float16, half, false, false)
-instantiate_rope(vjp_bfloat16, bfloat16_t, false, false)
-instantiate_rope(vjp_float32, float, false, false) // clang-format on
+instantiate_rope(float16, half)
+instantiate_rope(bfloat16, bfloat16_t)
+instantiate_rope(float32, float) // clang-format on

mlx/backend/metal/kernels/scan.h

@@ -306,6 +306,7 @@ template <
     U prev_thread = op.simd_exclusive_scan(values[N_READS - 1]);
 
     // Write simdgroup_sums to SM
+    threadgroup_barrier(mem_flags::mem_threadgroup);
     if (simd_lane_id == simd_size - 1) {
       simdgroup_sums[simd_group_id] = op(prev_thread, values[N_READS - 1]);
     }
@@ -440,6 +441,7 @@ template <
     }
 
     // Read in SM
+    threadgroup_barrier(mem_flags::mem_threadgroup);
     if (check_index_y < axis_size && (read_offset_x + N_READS) < stride_limit) {
       for (int i = 0; i < N_READS; i++) {
         read_into[i] = in[index_y * stride + i];

mlx/backend/metal/kernels/sdpa_vector.h

@@ -9,6 +9,7 @@ constant bool query_transposed [[function_constant(21)]];
 constant bool do_causal [[function_constant(22)]];
 constant bool bool_mask [[function_constant(23)]];
 constant bool float_mask [[function_constant(24)]];
+constant bool has_sinks [[function_constant(25)]];
 
 template <typename T, int D, int V = D>
 [[kernel]] void sdpa_vector(
@@ -31,6 +32,9 @@ template <typename T, int D, int V = D>
     [[buffer(14), function_constant(has_mask)]],
     const constant int& mask_head_stride
     [[buffer(15), function_constant(has_mask)]],
+    const device T* sinks [[buffer(16), function_constant(has_sinks)]],
+    const constant int& num_q_heads
+    [[buffer(17), function_constant(has_sinks)]],
     uint3 tid [[threadgroup_position_in_grid]],
     uint3 tpg [[threadgroups_per_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
@@ -53,24 +57,24 @@ template <typename T, int D, int V = D>
   threadgroup U sum_exp_scores[BN];
 
   // Adjust positions
-  const int head_idx = tid.x;
+  const int q_batch_head_idx = tid.x;
   const int q_seq_idx = tid.y;
-  const int kv_head_idx = head_idx / gqa_factor;
-  const int o_offset = head_idx * tpg.y + q_seq_idx;
+  const int kv_head_idx = q_batch_head_idx / gqa_factor;
+  const int o_offset = q_batch_head_idx * tpg.y + q_seq_idx;
   const int q_offset =
-      query_transposed ? tpg.x * q_seq_idx + head_idx : o_offset;
+      query_transposed ? tpg.x * q_seq_idx + q_batch_head_idx : o_offset;
   queries += q_offset * D + simd_lid * qk_per_thread;
   keys += kv_head_idx * k_head_stride + simd_gid * k_seq_stride +
       simd_lid * qk_per_thread;
   values += kv_head_idx * v_head_stride + simd_gid * v_seq_stride +
       simd_lid * v_per_thread;
   if (bool_mask) {
-    bmask += head_idx * mask_head_stride + simd_gid * mask_kv_seq_stride +
-        q_seq_idx * mask_q_seq_stride;
+    bmask += q_batch_head_idx * mask_head_stride +
+        simd_gid * mask_kv_seq_stride + q_seq_idx * mask_q_seq_stride;
   }
   if (float_mask) {
-    fmask += head_idx * mask_head_stride + simd_gid * mask_kv_seq_stride +
-        q_seq_idx * mask_q_seq_stride;
+    fmask += q_batch_head_idx * mask_head_stride +
+        simd_gid * mask_kv_seq_stride + q_seq_idx * mask_q_seq_stride;
   }
 
   out += o_offset * V + simd_gid * v_per_thread;
@@ -83,8 +87,12 @@ template <typename T, int D, int V = D>
     o[i] = 0;
   }
 
-  U max_score = -INFINITY;
+  U max_score = Limits<U>::finite_min;
   U sum_exp_score = 0;
+  if (has_sinks && simd_gid == 0) {
+    max_score = static_cast<U>(sinks[q_batch_head_idx % num_q_heads]);
+    sum_exp_score = 1;
+  }
 
   // For each key
   for (int i = simd_gid; i < N; i += BN) {
@@ -93,6 +101,8 @@ template <typename T, int D, int V = D>
       use_key = i <= (N - int(tpg.y) + int(q_seq_idx));
     } else if (bool_mask) {
       use_key = bmask[0];
+    } else if (float_mask) {
+      use_key = (fmask[0] >= Limits<T>::finite_min);
     }
     if (use_key) {
       // Read the key
@@ -107,7 +117,7 @@ template <typename T, int D, int V = D>
       }
       score = simd_sum(score);
       if (float_mask) {
-        score += max(Limits<U>::finite_min, static_cast<U>(fmask[0]));
+        score += static_cast<U>(fmask[0]);
       }
 
       // Update the accumulators
@@ -152,7 +162,8 @@ template <typename T, int D, int V = D>
   for (int i = 0; i < v_per_thread; i++) {
     outputs[simd_lid * BD + simd_gid] = o[i];
     threadgroup_barrier(mem_flags::mem_threadgroup);
-    o[i] = simd_sum(outputs[simd_gid * BD + simd_lid] * factor) / sum_exp_score;
+    o[i] = simd_sum(outputs[simd_gid * BD + simd_lid] * factor);
+    o[i] = sum_exp_score == 0 ? o[i] : (o[i] / sum_exp_score);
     threadgroup_barrier(mem_flags::mem_threadgroup);
   }
 
@@ -187,6 +198,9 @@ template <typename T, int D, int V = D>
     [[buffer(16), function_constant(has_mask)]],
     const constant int& mask_head_stride
     [[buffer(17), function_constant(has_mask)]],
+    const device T* sinks [[buffer(18), function_constant(has_sinks)]],
+    const constant int& num_q_heads
+    [[buffer(19), function_constant(has_sinks)]],
     uint3 tid [[threadgroup_position_in_grid]],
     uint3 tpg [[threadgroups_per_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
@@ -211,12 +225,12 @@ template <typename T, int D, int V = D>
 
   // Adjust positions
   const int block_idx = tid.z;
-  const int head_idx = tid.x;
+  const int q_batch_head_idx = tid.x;
   const int q_seq_idx = tid.y;
-  const int o_offset = head_idx * tpg.y + q_seq_idx;
+  const int o_offset = q_batch_head_idx * tpg.y + q_seq_idx;
   const int q_offset =
-      query_transposed ? tpg.x * q_seq_idx + head_idx : o_offset;
-  const int kv_head_idx = head_idx / gqa_factor;
+      query_transposed ? tpg.x * q_seq_idx + q_batch_head_idx : o_offset;
+  const int kv_head_idx = q_batch_head_idx / gqa_factor;
 
   queries += q_offset * D + simd_lid * qk_per_thread;
   keys += kv_head_idx * k_head_stride +
@@ -225,12 +239,12 @@ template <typename T, int D, int V = D>
       (block_idx * BN + simd_gid) * v_seq_stride + simd_lid * v_per_thread;
   out += o_offset * blocks * V + block_idx * V + simd_lid * v_per_thread;
   if (bool_mask) {
-    bmask += head_idx * mask_head_stride +
+    bmask += q_batch_head_idx * mask_head_stride +
         (block_idx * BN + simd_gid) * mask_kv_seq_stride +
         q_seq_idx * mask_q_seq_stride;
   }
   if (float_mask) {
-    fmask += head_idx * mask_head_stride +
+    fmask += q_batch_head_idx * mask_head_stride +
         (block_idx * BN + simd_gid) * mask_kv_seq_stride +
         q_seq_idx * mask_q_seq_stride;
   }
@@ -245,8 +259,13 @@ template <typename T, int D, int V = D>
     o[i] = 0;
   }
 
-  U max_score = -1e9;
+  U max_score = Limits<U>::finite_min;
   U sum_exp_score = 0;
+  if (has_sinks && block_idx == 0 && simd_gid == 0) {
+    int q_head_idx = q_batch_head_idx % num_q_heads;
+    max_score = static_cast<U>(sinks[q_head_idx]);
+    sum_exp_score = 1;
+  }
 
   // For each key
   for (int i = block_idx * BN + simd_gid; i < N; i += blocks * BN) {
@@ -255,6 +274,8 @@ template <typename T, int D, int V = D>
       use_key = i <= (N - int(tpg.y) + int(q_seq_idx));
     } else if (bool_mask) {
       use_key = bmask[0];
+    } else if (float_mask) {
+      use_key = (fmask[0] >= Limits<T>::finite_min);
     }
     if (use_key) {
       // Read the key
@@ -268,6 +289,7 @@ template <typename T, int D, int V = D>
         score += q[i] * k[i];
       }
       score = simd_sum(score);
+
       if (float_mask) {
         score += fmask[0];
       }
@@ -379,7 +401,8 @@ template <typename T, int D>
   for (int i = 0; i < elem_per_thread; i++) {
     outputs[simd_lid * BD + simd_gid] = o[i];
     threadgroup_barrier(mem_flags::mem_threadgroup);
-    o[i] = simd_sum(outputs[simd_gid * BD + simd_lid] * factor) / sum_exp_score;
+    o[i] = simd_sum(outputs[simd_gid * BD + simd_lid] * factor);
+    o[i] = sum_exp_score == 0 ? o[i] : (o[i] / sum_exp_score);
     threadgroup_barrier(mem_flags::mem_threadgroup);
   }
 

mlx/backend/metal/kernels/sort.h

@@ -19,11 +19,28 @@ METAL_FUNC void thread_swap(thread T& a, thread T& b) {
   b = w;
 }
 
+template <typename T, typename = void>
+struct Init {
+  static constexpr constant T v = Limits<T>::max;
+};
+
+template <typename T>
+struct Init<T, metal::enable_if_t<metal::is_floating_point_v<T>>> {
+  static constexpr constant T v = metal::numeric_limits<T>::quiet_NaN();
+};
+
 template <typename T>
 struct LessThan {
-  static constexpr constant T init = Limits<T>::max;
-
-  METAL_FUNC bool operator()(T a, T b) {
+  static constexpr constant T init = Init<T>::v;
+  METAL_FUNC bool operator()(T a, T b) const {
+    if constexpr (
+        metal::is_floating_point_v<T> || metal::is_same_v<T, complex64_t>) {
+      bool an = isnan(a);
+      bool bn = isnan(b);
+      if (an | bn) {
+        return (!an) & bn;
+      }
+    }
     return a < b;
   }
 };

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

@@ -11,6 +11,7 @@ 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 {
@@ -82,6 +83,7 @@ template <
     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]],
@@ -169,7 +171,7 @@ template <
   VBlockLoader loader_v(
       V, params->V_strides[2], Vs, simd_group_id, simd_lane_id);
 
-  TransformScale<T> ts(static_cast<T>(params->scale * 1.44269504089));
+  TransformScale<T> ts(static_cast<T>(params->scale * M_LOG2E_F));
 
   // Prepare MMA tiles
   constexpr short kFragSize = 8; // MMAFrag size
@@ -232,6 +234,14 @@ template <
     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) {
@@ -350,7 +360,7 @@ template <
               Stile.frag_at(i, j)[jj] =
                   mfrag[jj] ? Stile.frag_at(i, j)[jj] : neg_inf;
             } else {
-              Stile.frag_at(i, j)[jj] += 1.44269504089 * selem_t(mfrag[jj]);
+              Stile.frag_at(i, j)[jj] += M_LOG2E_F * selem_t(mfrag[jj]);
             }
           }
         }

mlx/backend/metal/kernels/steel/conv/kernels/steel_conv.metal

@@ -3,9 +3,8 @@
 #include <metal_stdlib>
 
 // clang-format off
-#include "mlx/backend/metal/kernels/steel/gemm/mma.h"
-
 #include "mlx/backend/metal/kernels/utils.h"
+#include "mlx/backend/metal/kernels/steel/gemm/mma.h"
 #include "mlx/backend/metal/kernels/steel/conv/conv.h"
 #include "mlx/backend/metal/kernels/steel/conv/params.h"
 #include "mlx/backend/metal/kernels/steel/conv/kernels/steel_conv.h"