Merge e6ae350999 into b3d7b85376

cmake/FindNCCL.cmake

@@ -0,0 +1,64 @@
+# Find the nccl libraries
+#
+# The following variables are optionally searched for defaults NCCL_ROOT_DIR:
+# Base directory where all NCCL components are found NCCL_INCLUDE_DIR: Directory
+# where NCCL header is found NCCL_LIB_DIR: Directory where NCCL library is found
+#
+# The following are set after configuration is done: NCCL_FOUND
+# NCCL_INCLUDE_DIRS NCCL_LIBRARIES
+#
+# The path hints include CUDA_TOOLKIT_ROOT_DIR seeing as some folks install NCCL
+# in the same location as the CUDA toolkit. See
+# https://github.com/caffe2/caffe2/issues/1601
+
+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()

mlx/backend/cuda/jit_module.cpp

@@ -37,36 +37,46 @@ void check_cu_error(const char* name, CUresult err) {
 }
 
 // Return the location of the CUDA toolkit.
-const char* cuda_home() {
-  const char* home = std::getenv("CUDA_HOME");
-  if (home) {
-    return home;
-  }
-  home = std::getenv("CUDA_PATH");
-  if (home) {
-    return home;
-  }
+const std::string& cuda_home() {
+  static std::string home = []() -> std::string {
+    const char* home = std::getenv("CUDA_HOME");
+    if (home) {
+      return home;
+    }
+    home = std::getenv("CUDA_PATH");
+    if (home) {
+      return home;
+    }
 #if defined(__linux__)
-  home = "/usr/local/cuda";
-  if (std::filesystem::exists(home)) {
-    return home;
-  }
+    home = "/usr/local/cuda";
+    if (std::filesystem::exists(home)) {
+      return home;
+    }
 #endif
-  throw std::runtime_error(
-      "Environment variable CUDA_HOME or CUDA_PATH is not set.");
+    throw std::runtime_error(
+        "Environment variable CUDA_HOME or CUDA_PATH is not set.");
+  }();
+  return home;
 }
 
 // Get the cache directory for storing compiled results.
-bool get_ptx_cache_dir(std::filesystem::path* result) {
-  auto path = std::filesystem::temp_directory_path() / "mlx" / "ptx";
-  if (!std::filesystem::is_directory(path)) {
-    std::error_code error;
-    if (!std::filesystem::create_directories(path, error)) {
-      return false;
+const std::filesystem::path& ptx_cache_dir() {
+  static std::filesystem::path cache = []() -> std::filesystem::path {
+    std::filesystem::path cache;
+    if (auto c = std::getenv("MLX_PTX_CACHE"); c) {
+      cache = c;
+    } else {
+      cache = std::filesystem::temp_directory_path() / "mlx" / "ptx";
     }
-  }
-  *result = path;
-  return true;
+    if (!std::filesystem::exists(cache)) {
+      std::error_code error;
+      if (!std::filesystem::create_directories(cache, error)) {
+        return std::filesystem::path();
+      }
+    }
+    return cache;
+  }();
+  return cache;
 }
 
 // Try to read the cached |ptx| and |ptx_kernels| from |cache_dir|.
@@ -75,6 +85,10 @@ bool read_cached_ptx(
     const std::string& module_name,
     std::vector<char>* ptx,
     std::vector<std::pair<std::string, std::string>>* ptx_kernels) {
+  if (cache_dir.empty()) {
+    return false;
+  }
+
   auto ptx_path = cache_dir / (module_name + ".ptx");
   std::error_code error;
   auto ptx_size = std::filesystem::file_size(ptx_path, error);
@@ -105,6 +119,10 @@ void write_cached_ptx(
     const std::string& module_name,
     const std::vector<char>& ptx,
     const std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
+  if (cache_dir.empty()) {
+    return;
+  }
+
   std::ofstream ptx_file(cache_dir / (module_name + ".ptx"), std::ios::binary);
   if (!ptx.empty()) {
     ptx_file.write(&ptx.front(), ptx.size());
@@ -184,11 +202,9 @@ JitModule::JitModule(
     const std::string& module_name,
     const KernelBuilder& builder) {
   // Check cache.
-  std::filesystem::path cache_dir;
   std::vector<char> ptx;
   std::vector<std::pair<std::string, std::string>> ptx_kernels;
-  if (!get_ptx_cache_dir(&cache_dir) ||
-      !read_cached_ptx(cache_dir, module_name, &ptx, &ptx_kernels)) {
+  if (!read_cached_ptx(ptx_cache_dir(), module_name, &ptx, &ptx_kernels)) {
     // Create program.
     auto [source_code, kernel_names] = builder();
     nvrtcProgram prog;
@@ -246,7 +262,7 @@ JitModule::JitModule(
     } else {
       CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
     }
-    write_cached_ptx(cache_dir, module_name, ptx, ptx_kernels);
+    write_cached_ptx(ptx_cache_dir(), module_name, ptx, ptx_kernels);
   }
 
   // Load module.

mlx/backend/cuda/primitives.cu

@@ -54,6 +54,28 @@ bool fast::ScaledDotProductAttention::use_fallback(
   return true;
 }
 
+namespace distributed {
+void AllReduce::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  // Here I assume for now that in is donatable and contiguous.
+  // TODO
+
+  auto& input = inputs[0];
+  auto& output = outputs[0];
+
+  output.copy_shared_buffer(input);
+  auto& s = stream();
+  switch (reduce_type_) {
+    case Sum:
+      distributed::detail::all_sum(group(), input, output, s);
+      break;
+    default:
+      throw std::runtime_error("Only all reduce sum is supported for now");
+  }
+}
+} // namespace distributed
+
 #define NO_GPU_MULTI(func)                                             \
   void func::eval_gpu(                                                 \
       const std::vector<array>& inputs, std::vector<array>& outputs) { \
@@ -97,7 +119,6 @@ NO_GPU_MULTI(CustomKernel)
 } // namespace fast
 
 namespace distributed {
-NO_GPU_MULTI(AllReduce)
 NO_GPU_MULTI(AllGather)
 NO_GPU_MULTI(Send)
 NO_GPU_MULTI(Recv)

mlx/distributed/CMakeLists.txt

@@ -6,3 +6,4 @@ target_sources(
 
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/mpi)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ring)
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/nccl)

mlx/distributed/distributed.cpp

@@ -5,6 +5,7 @@
 #include "mlx/distributed/distributed.h"
 #include "mlx/distributed/distributed_impl.h"
 #include "mlx/distributed/mpi/mpi.h"
+#include "mlx/distributed/nccl/nccl.h"
 #include "mlx/distributed/ring/ring.h"
 
 namespace mlx::core::distributed {
@@ -111,6 +112,8 @@ Group init(bool strict /* = false */, const std::string& bk /* = "any" */) {
     group = mpi::init(strict);
   } else if (bk == "ring") {
     group = ring::init(strict);
+  } else if (bk == "nccl") {
+    group = nccl::init(strict);
   } else if (bk == "any") {
     group = ring::init(false);
     bk_ = "ring";

mlx/distributed/nccl/CMakeLists.txt

@@ -0,0 +1,8 @@
+if(MLX_BUILD_CUDA)
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/nccl.cpp)
+  find_package(NCCL REQUIRED)
+  target_link_libraries(mlx PRIVATE ${NCCL_LIBRARIES})
+  target_include_directories(mlx PRIVATE ${NCCL_INCLUDE_DIRS})
+else()
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/no_nccl.cpp)
+endif()

mlx/distributed/nccl/nccl.cpp

@@ -0,0 +1,382 @@
+#include <arpa/inet.h>
+#include <cuda_runtime.h>
+#include <nccl.h>
+#include <netdb.h>
+#include <sys/socket.h>
+#include <unistd.h>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <iostream>
+#include <mutex>
+#include <stdexcept>
+#include <string>
+#include <type_traits>
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/distributed/distributed.h"
+#include "mlx/distributed/distributed_impl.h"
+
+namespace mlx::core::distributed::nccl {
+
+#define CHECK_CUDA(cmd)              \
+  do {                               \
+    cudaError_t e = cmd;             \
+    if (e != cudaSuccess) {          \
+      fprintf(                       \
+          stderr,                    \
+          "CUDA error %s:%d '%s'\n", \
+          __FILE__,                  \
+          __LINE__,                  \
+          cudaGetErrorString(e));    \
+      exit(1);                       \
+    }                                \
+  } while (0)
+
+#define CHECK_NCCL(cmd)              \
+  do {                               \
+    ncclResult_t r = cmd;            \
+    if (r != ncclSuccess) {          \
+      fprintf(                       \
+          stderr,                    \
+          "NCCL error %s:%d '%s'\n", \
+          __FILE__,                  \
+          __LINE__,                  \
+          ncclGetErrorString(r));    \
+      exit(1);                       \
+    }                                \
+  } while (0)
+
+namespace detail {
+
+inline void sendAll(int sock, const void* buf, size_t len) {
+  const char* ptr = reinterpret_cast<const char*>(buf);
+  while (len > 0) {
+    ssize_t sent = send(sock, ptr, len, 0);
+    if (sent <= 0) {
+      perror("send");
+      exit(1);
+    }
+    ptr += sent;
+    len -= sent;
+  }
+}
+
+inline void recvAll(int sock, void* buf, size_t len) {
+  char* ptr = reinterpret_cast<char*>(buf);
+  while (len > 0) {
+    ssize_t rec = recv(sock, ptr, len, 0);
+    if (rec <= 0) {
+      perror("recv");
+      exit(1);
+    }
+    ptr += rec;
+    len -= rec;
+  }
+}
+
+inline void bootstrap_unique_id(
+    ncclUniqueId& id,
+    int rank,
+    int size,
+    const std::string& initMethod) {
+
+  if (initMethod.rfind("tcp://", 0) != 0)
+    throw;
+  auto hostport = initMethod.substr(6);
+  auto colon = hostport.find(':');
+  std::string host = hostport.substr(0, colon);
+  int port = std::stoi(hostport.substr(colon + 1));
+
+  if (rank == 0) {
+    CHECK_NCCL(ncclGetUniqueId(&id));
+
+    int sock = socket(AF_INET, SOCK_STREAM, 0);
+
+    if (sock < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] Couldn't create socket (error: " << errno << ")";
+      throw std::runtime_error(msg.str());
+    }
+
+    sockaddr_in serv = {};
+    serv.sin_family = AF_INET;
+    serv.sin_addr.s_addr = htonl(INADDR_ANY);
+    serv.sin_port = htons(port);
+
+    int reuse = 1;
+    if (setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, &reuse, sizeof(reuse)) < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] setsockopt() failed: " << strerror(errno);
+      throw std::runtime_error(msg.str());
+    }
+
+    if (bind(sock, reinterpret_cast<sockaddr*>(&serv), sizeof(serv)) < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] bind() failed: " << strerror(errno);
+      throw std::runtime_error(msg.str());
+    }
+    if (listen(sock, size - 1) < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] listen() failed: " << strerror(errno);
+      throw std::runtime_error(msg.str());
+    }
+
+    for (int peer = 1; peer < size; ++peer) {
+      int conn = accept(sock, nullptr, nullptr);
+      if (conn < 0) {
+        std::ostringstream msg;
+        msg << "[nccl] accept() failed: " << strerror(errno);
+        throw std::runtime_error(msg.str());
+      }
+      sendAll(conn, &id, sizeof(id));
+      close(conn);
+    }
+    close(sock);
+
+  } else {
+    // Here just wanted to make show that rank 0 has enough time to bind
+    // so we will retry to connect until max attempts
+
+    int sock = socket(AF_INET, SOCK_STREAM, 0);
+    if (sock < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] socket() failed: " << strerror(errno);
+      throw std::runtime_error(msg.str());
+    }
+
+    hostent* he = gethostbyname(host.c_str());
+    if (!he) {
+      throw std::runtime_error("[nccl] lookup failed for host: " + host);
+    }
+    sockaddr_in serv = {};
+    serv.sin_family = AF_INET;
+    memcpy(&serv.sin_addr, he->h_addr_list[0], he->h_length);
+    serv.sin_port = htons(port);
+
+    const int max_retries = 30;
+    int attempt = 0;
+    bool connected = false;
+
+    for (attempt = 0; attempt < max_retries; ++attempt) {
+      if (connect(sock, reinterpret_cast<sockaddr*>(&serv), sizeof(serv)) ==
+          0) {
+        connected = true;
+        std::cout << "[Rank " << rank << "] Connected successfully on attempt "
+                  << attempt + 1 << std::endl;
+        break;
+      }
+      if (errno != ECONNREFUSED) {
+        break;
+      }
+      std::this_thread::sleep_for(std::chrono::milliseconds(500));
+    }
+
+    if (!connected) {
+      std::ostringstream msg;
+      msg << "[Rank " << rank << "] connect() failed after " << attempt
+          << " retries: " << strerror(errno);
+      close(sock);
+      throw std::runtime_error(msg.str());
+    }
+    recvAll(sock, &id, sizeof(id));
+    close(sock);
+  }
+}
+
+template <typename T>
+struct type_identity {
+  using type = T;
+};
+
+template <typename F>
+void dispatch_dtype(const array& arr, F&& f) {
+  switch (arr.dtype()) {
+    case bool_:
+      throw std::invalid_argument("[nccl] Boolean arrays not supported");
+    case int8:
+      f(type_identity<int8_t>{}, ncclChar);
+      break;
+    case uint8:
+      f(type_identity<uint8_t>{}, ncclUint8);
+      break;
+    case int32:
+      f(type_identity<int32_t>{}, ncclInt);
+      break;
+    case uint32:
+      f(type_identity<uint32_t>{}, ncclUint32);
+      break;
+    case int64:
+      f(type_identity<int64_t>{}, ncclInt64);
+      break;
+    case uint64:
+      f(type_identity<uint64_t>{}, ncclUint64);
+      break;
+    case float16:
+      f(type_identity<float16_t>{}, ncclHalf);
+      break;
+    case bfloat16:
+      f(type_identity<bfloat16_t>{}, ncclBfloat16);
+      break;
+    case float32:
+      f(type_identity<float>{}, ncclFloat);
+      break;
+    case float64:
+      f(type_identity<double>{}, ncclDouble);
+      break;
+    default:
+      throw std::invalid_argument("[nccl] Unknown or unsupported dtype");
+  }
+}
+
+} // namespace detail
+
+using GroupImpl = mlx::core::distributed::detail::GroupImpl;
+class NCCLGroup : public GroupImpl {
+ public:
+  NCCLGroup(int worldRank, int worldSize, const std::string initMethod)
+      : rank_(worldRank),
+        size_(worldSize),
+        comm_(nullptr),
+        initMethod_(initMethod) {
+    if (initialized_)
+      return;
+    int ndev;
+    CHECK_CUDA(cudaGetDeviceCount(&ndev));
+    CHECK_CUDA(cudaSetDevice(rank_ % ndev));
+    CHECK_CUDA(cudaStreamCreate(&stream_));
+
+    detail::bootstrapUniqueId(uniqueId_, rank_, size_, initMethod_);
+    CHECK_NCCL(ncclCommInitRank(&comm_, size_, uniqueId_, rank_));
+    initialized_ = true;
+  }
+
+  ~NCCLGroup() {
+    ncclCommDestroy(comm_);
+    ncclGroupEnd();
+    cudaStreamDestroy(stream_);
+    initialized_ = false;
+  }
+
+  int rank() override {
+    return rank_;
+  }
+
+  int size() override {
+    return size_;
+  }
+
+  void all_sum(const array& input, array& output, Stream stream) override {
+    if (input.size() != output.size()) {
+      throw std::runtime_error(
+          "[nccl] Input and output arrays must have the same size.");
+    }
+    detail::dispatch_dtype(input, [&](auto type_tag, ncclDataType_t dt) {
+      using T = typename decltype(type_tag)::type;
+      all_reduce_impl<T>(input, output, stream, dt, ncclSum);
+    });
+  }
+
+  virtual std::shared_ptr<GroupImpl> split(int color, int key = -1) override {
+    throw std::runtime_error("[nccl] Group split not supported.");
+  }
+
+  void all_gather(const array& input, array& output, Stream stream) override {
+    if (input.size() != output.size() / size_) {
+      throw std::runtime_error(
+          "[nccl] Input size must match output size divided by group size.");
+    }
+  }
+
+  void send(const array& input, int dst, Stream stream) override {
+    if (input.size() == 0) {
+      return; // Nothing to send
+    }
+  }
+
+  void recv(array& output, int src, Stream stream) override {
+    if (output.size() == 0) {
+      return; // Nothing to receive
+    }
+  }
+
+  void all_max(const array& input, array& output, Stream stream) override {
+    if (input.size() != output.size()) {
+      throw std::runtime_error(
+          "[nccl] Input and output arrays must have the same size.");
+    }
+    detail::dispatch_dtype(input, [&](auto type_tag, ncclDataType_t dt) {
+      using T = typename decltype(type_tag)::type;
+      all_reduce_impl<T>(input, output, stream, dt, ncclMax);
+    });
+  }
+
+  void all_min(const array& input, array& output, Stream stream) override {
+    if (input.size() != output.size()) {
+      throw std::runtime_error(
+          "[nccl] Input and output arrays must have the same size.");
+    }
+    detail::dispatch_dtype(input, [&](auto type_tag, ncclDataType_t dt) {
+      using T = typename decltype(type_tag)::type;
+      all_reduce_impl<T>(input, output, stream, dt, ncclMin);
+    });
+  }
+
+  template <typename T>
+  void all_reduce_impl(
+      const array& input,
+      array& output,
+      Stream stream,
+      ncclDataType_t dt,
+      ncclRedOp_t op) {
+        
+    CHECK_NCCL(ncclAllReduce(
+        input.data<T>(),
+        output.data<T>(),
+        input.size(),
+        dt,
+        op,
+        comm_,
+        stream_));
+    cudaStreamSynchronize(stream_);
+  }
+
+  int rank_, size_;
+  std::string initMethod_;
+  ncclUniqueId uniqueId_;
+  ncclComm_t comm_;
+  cudaStream_t stream_;
+  bool initialized_ = false;
+};
+
+bool is_available() {
+  return true;
+}
+
+namespace detail {
+static std::string get_env_var_or_throw(const char* env_var_name) {
+  const char* value = std::getenv(env_var_name);
+  if (value == nullptr) {
+    std::ostringstream msg;
+    msg << "[nccl] Required environment variable '" << env_var_name
+        << "' is not set. "
+        << "Please set it before initializing the distributed backend.";
+    throw std::runtime_error(msg.str());
+  }
+  return std::string(value);
+}
+} // namespace detail
+
+std::shared_ptr<GroupImpl> init(bool strict /* = false */) {
+  std::string host = detail::get_env_var_or_throw("NCCL_HOST_IP");
+  std::string port = detail::get_env_var_or_throw("NCCL_PORT");
+  std::string rank_str = detail::get_env_var_or_throw("MLX_RANK");
+  std::string n_nodes_str = detail::get_env_var_or_throw("MLX_WORLD_SIZE");
+
+  int rank = std::stoi(rank_str);
+  int n_nodes = std::stoi(n_nodes_str);
+  std::string init_method = "tcp://" + host + ":" + port;
+
+  return std::make_shared<NCCLGroup>(rank, n_nodes, init_method);
+}
+} // namespace mlx::core::distributed::nccl

mlx/distributed/nccl/nccl.h

@@ -0,0 +1,12 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/distributed/distributed.h"
+
+namespace mlx::core::distributed::nccl {
+
+using GroupImpl = mlx::core::distributed::detail::GroupImpl;
+
+bool is_available();
+std::shared_ptr<GroupImpl> init(bool strict = false);
+
+} // namespace mlx::core::distributed::nccl

mlx/distributed/nccl/no_nccl.cpp

@@ -0,0 +1,20 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/distributed/nccl/nccl.h"
+
+namespace mlx::core::distributed::nccl {
+
+using GroupImpl = mlx::core::distributed::detail::GroupImpl;
+
+bool is_available() {
+  return false;
+}
+
+std::shared_ptr<GroupImpl> init(bool strict /* = false */) {
+  if (strict) {
+    throw std::runtime_error("Cannot initialize nccl distributed backend.");
+  }
+  return nullptr;
+}
+
+} // namespace mlx::core::distributed::nccl

mlx/distributed/ops.cpp

@@ -31,8 +31,7 @@ array all_sum(
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<AllReduce>(
-          to_stream(s, Device::cpu), group, AllReduce::Sum),
+      std::make_shared<AllReduce>(to_stream(s), group, AllReduce::Sum),
       {x});
 }