torch/csrc/distributed/c10d/Functional.cpp

#include <ATen/ATen.h>
#include <ATen/core/op_registration/op_registration.h>
#include <c10/core/DispatchKey.h>
#include <torch/csrc/autograd/custom_function.h>
#include <torch/csrc/autograd/function.h>
#include <torch/csrc/distributed/c10d/GroupRegistry.hpp>
#include <torch/csrc/distributed/c10d/ProcessGroup.hpp>
#include <torch/csrc/distributed/c10d/RankLocal.hpp>
#include <utility>

namespace {

class WorkRegistry {
 public:
  void register_work(
      const at::Tensor& tensor,
      const c10::intrusive_ptr<c10d::Work>& work) {
    auto storage = tensor.storage().getWeakStorageImpl();
    std::unique_lock lock(lock_);
    auto [it, inserted] = registry_.try_emplace(std::move(storage), work);
    TORCH_CHECK(
        inserted || it->second != work,
        "The tensor storage is already associated with another work.");
  }

  c10::intrusive_ptr<c10d::Work> pop_work(const at::Tensor& tensor) {
    const auto storage = tensor.storage().getWeakStorageImpl();
    std::unique_lock lock(lock_);
    auto it = registry_.find(storage);
    if (it == registry_.end()) {
      return nullptr;
    }
    auto work = it->second;
    registry_.erase(it);
    return work;
  }

  ~WorkRegistry() {
    // If there are still unwaited work objects, their corresponding process
    // groups should have already been destroyed at this stage. Any attempts to
    // wait for these work objects or to destroy them will only result in
    // confusing errors. Therefore, we simply issue a warning and intentionally
    // allow the unwaited work objects to leak.
    if (!registry_.empty()) {
      TORCH_WARN(
          "At the time of process termination, there are still ",
          registry_.size(),
          " unwaited c10d_functional collective calls. "
          "Please review your program to ensure c10d_functional.wait_tensor() "
          "is invoked on all tensors returned from c10d_functional collective "
          "ops before they are used.");
    }
    for (auto& it : registry_) {
      it.second.release();
    }
  }

 private:
  std::unordered_map<
      c10::weak_intrusive_ptr<c10::StorageImpl>,
      c10::intrusive_ptr<c10d::Work>>
      registry_;
  std::mutex lock_;
};

static WorkRegistry process_registry;

void register_work(
    const at::Tensor& tensor,
    const c10::intrusive_ptr<c10d::Work>& work) {
  if (c10d::get_thread_isolation_mode()) {
    c10d::RankLocal<WorkRegistry>::get().register_work(tensor, work);
  } else {
    process_registry.register_work(tensor, work);
  }
}

c10::intrusive_ptr<c10d::Work> pop_work(const at::Tensor& tensor) {
  if (c10d::get_thread_isolation_mode()) {
    return c10d::RankLocal<WorkRegistry>::get().pop_work(tensor);
  } else {
    return process_registry.pop_work(tensor);
  }
}

const std::unordered_map<std::string, c10d::ReduceOp> str_to_reduce_op = {
    {"sum", c10d::ReduceOp(c10d::ReduceOp::RedOpType::SUM)},
    {"avg", c10d::ReduceOp(c10d::ReduceOp::RedOpType::AVG)},
    {"product", c10d::ReduceOp(c10d::ReduceOp::RedOpType::PRODUCT)},
    {"min", c10d::ReduceOp(c10d::ReduceOp::RedOpType::MIN)},
    {"max", c10d::ReduceOp(c10d::ReduceOp::RedOpType::MAX)},
    {"band", c10d::ReduceOp(c10d::ReduceOp::RedOpType::BAND)},
    {"bor", c10d::ReduceOp(c10d::ReduceOp::RedOpType::BOR)},
    {"bxor", c10d::ReduceOp(c10d::ReduceOp::RedOpType::BXOR)},
    // TODO: support premul_sum
    // {"premul_sum", c10d::ReduceOp(c10d::ReduceOp::RedOpType::PREMUL_SUM)},
    {"unused", c10d::ReduceOp(c10d::ReduceOp::RedOpType::UNUSED)}};

c10d::ReduceOp to_reduce_op(const std::string& reduce_op) {
  auto it = str_to_reduce_op.find(reduce_op);
  TORCH_CHECK(
      it != str_to_reduce_op.end(), "Unrecognized reduce_op: ", reduce_op);
  return it->second;
}

at::Tensor& all_reduce_(
    at::Tensor& input,
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::string reduce_op,
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::string group_name) {
  c10d::AllreduceOptions opts;
  opts.reduceOp = to_reduce_op(reduce_op);

  std::vector<at::Tensor> inputs{input};
  auto group = c10d::resolve_process_group(group_name);
  auto work = group->allreduce(inputs, opts);
  c10d::RankLocal<WorkRegistry>::get().register_work(input, work);
  return input;
}

at::Tensor all_reduce(
    const at::Tensor& input,
    std::string reduce_op,
    std::string group_name) {
  auto output = input.clone(at::MemoryFormat::Contiguous);
  return all_reduce_(output, std::move(reduce_op), std::move(group_name));
}

std::vector<at::Tensor> all_reduce_coalesced_(
    std::vector<at::Tensor> inputs,
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::string reduce_op,
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::string group_name) {
  c10d::AllreduceCoalescedOptions opts;
  opts.reduceOp = to_reduce_op(reduce_op);

  auto group = c10d::resolve_process_group(group_name);
  auto work = group->allreduce_coalesced(inputs, opts);
  for (const auto& tensor : inputs) {
    c10d::RankLocal<WorkRegistry>::get().register_work(tensor, work);
  }
  return inputs;
}

std::vector<at::Tensor> all_reduce_coalesced(
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::vector<at::Tensor> inputs,
    std::string reduce_op,
    std::string group_name) {
  std::vector<at::Tensor> outputs;
  outputs.reserve(inputs.size());
  for (const auto& tensor : inputs) {
    outputs.push_back(tensor.clone(at::MemoryFormat::Contiguous));
  }
  return all_reduce_coalesced_(
      outputs, std::move(reduce_op), std::move(group_name));
}

at::Tensor allocate_all_gather_output(
    const at::Tensor& input,
    int64_t group_size) {
  auto output_size = input.sizes().vec();
  output_size[0] *= group_size;
  return at::empty(
      output_size,
      at::TensorOptions().dtype(input.dtype()).device(input.device()));
}

std::vector<at::Tensor> all_gather_into_tensor_coalesced(
    std::vector<at::Tensor> inputs,
    int64_t group_size,
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::string group_name) {
  std::vector<at::Tensor> outputs;
  outputs.reserve(inputs.size());
  for (const auto& tensor : inputs) {
    outputs.push_back(allocate_all_gather_output(tensor, group_size));
  }

  auto group = c10d::resolve_process_group(group_name);
  auto work = group->allgather_into_tensor_coalesced(outputs, inputs);
  for (const auto& tensor : outputs) {
    c10d::RankLocal<WorkRegistry>::get().register_work(tensor, work);
  }
  return outputs;
}

at::Tensor all_gather_into_tensor(
    const at::Tensor& input,
    int64_t group_size,
    std::string group_name) {
  std::vector<at::Tensor> inputs{input};
  return all_gather_into_tensor_coalesced(
      inputs, group_size, std::move(group_name))[0];
}

at::Tensor& all_gather_into_tensor_out(
    at::Tensor& input,
    int64_t group_size,
    std::string group_name,
    at::Tensor& output) {
  c10d::AllgatherOptions opts;

  auto group = c10d::resolve_process_group(group_name);
  auto work = group->_allgather_base(output, input, opts);
  c10d::RankLocal<WorkRegistry>::get().register_work(output, work);
  return output;
}

at::Tensor allocate_reduce_scatter_output(
    const at::Tensor& input,
    const int64_t group_size) {
  auto output_size = input.sizes().vec();
  if (output_size[0] % group_size != 0) {
    LOG(WARNING) << "The first dimension of the reduce_scatter input ("
                 << output_size[0] << ") is not divisible by the group size ("
                 << group_size << ").";
  }
  output_size[0] /= group_size;
  return at::empty(
      output_size,
      at::TensorOptions().dtype(input.dtype()).device(input.device()));
}

std::vector<at::Tensor> reduce_scatter_tensor_coalesced(
    std::vector<at::Tensor> inputs,
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::string reduce_op,
    int64_t group_size,
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::string group_name) {
  c10d::ReduceScatterOptions opts;
  opts.reduceOp = to_reduce_op(reduce_op);
  std::vector<at::Tensor> outputs;
  outputs.reserve(inputs.size());
  for (const auto& tensor : inputs) {
    outputs.push_back(allocate_reduce_scatter_output(tensor, group_size));
  }

  auto group = c10d::resolve_process_group(group_name);
  auto work = group->reduce_scatter_tensor_coalesced(outputs, inputs, opts);
  for (const auto& tensor : outputs) {
    c10d::RankLocal<WorkRegistry>::get().register_work(tensor, work);
  }
  return outputs;
}

at::Tensor reduce_scatter_tensor(
    const at::Tensor& input,
    std::string reduce_op,
    int64_t group_size,
    std::string group_name) {
  std::vector<at::Tensor> inputs{input};
  return reduce_scatter_tensor_coalesced(
      inputs, std::move(reduce_op), group_size, std::move(group_name))[0];
}

at::Tensor all_to_all_single(
    const at::Tensor& input,
    std::vector<int64_t> output_split_sizes,
    std::vector<int64_t> input_split_sizes,
    // NOLINTNEXTLINE(performance-unnecessary-value-param)
    std::string group_name) {
  std::vector<int64_t> output_sizes = input.sizes().vec();
  output_sizes[0] = std::accumulate(
      output_split_sizes.begin(), output_split_sizes.end(), int64_t(0));
  auto output = input.new_empty(output_sizes);

  auto group = c10d::resolve_process_group(group_name);
  auto work = group->alltoall_base(
      output,
      // NOLINTNEXTLINE(cppcoreguidelines-pro-type-const-cast)
      const_cast<at::Tensor&>(input),
      output_split_sizes,
      input_split_sizes);
  c10d::RankLocal<WorkRegistry>::get().register_work(output, work);
  return output;
}

// NOLINTNEXTLINE(performance-unnecessary-value-param)
at::Tensor& broadcast_(at::Tensor& input, int64_t src, std::string group_name) {
  c10d::BroadcastOptions opts;
  opts.rootRank = src;
  std::vector<at::Tensor> inputs{input};

  auto group = c10d::resolve_process_group(group_name);
  auto work = group->broadcast(inputs, opts);
  c10d::RankLocal<WorkRegistry>::get().register_work(input, work);
  return input;
}

at::Tensor broadcast(
    const at::Tensor& input,
    int64_t src,
    std::string group_name) {
  auto output = input.clone(at::MemoryFormat::Contiguous);
  return broadcast_(output, src, std::move(group_name));
}

at::Tensor wait_tensor(const at::Tensor& tensor) {
  auto work = c10d::RankLocal<WorkRegistry>::get().pop_work(tensor);
  if (work != nullptr) {
    work->wait();
  }
  return tensor;
}

} // namespace

TORCH_LIBRARY(_c10d_functional, m) {
  m.def(
      "all_reduce(Tensor input, str reduce_op, str group_name) -> Tensor",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::all_reduce),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "all_reduce_(Tensor(a!) input, str reduce_op, str group_name) -> Tensor(a!)",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::all_reduce_),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "all_reduce_coalesced(Tensor[] inputs, str reduce_op, str group_name) -> Tensor[]",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::all_reduce_coalesced),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "all_reduce_coalesced_(Tensor[](a!) inputs, str reduce_op, str group_name) -> Tensor[](a!)",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::all_reduce_coalesced_),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "all_gather_into_tensor_out(Tensor input, int group_size, str group_name, *, Tensor(a!) out) -> Tensor(a!)",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd,
          ::all_gather_into_tensor_out),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "all_gather_into_tensor(Tensor input, int group_size, str group_name) -> Tensor",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd,
          ::all_gather_into_tensor),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "all_gather_into_tensor_coalesced(Tensor[] inputs, int group_size, str group_name) -> Tensor[]",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd,
          ::all_gather_into_tensor_coalesced),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "reduce_scatter_tensor(Tensor input, str reduce_op, int group_size, str group_name) -> Tensor",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::reduce_scatter_tensor),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "reduce_scatter_tensor_coalesced(Tensor[] inputs, str reduce_op, int group_size, str group_name) -> Tensor[]",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd,
          ::reduce_scatter_tensor_coalesced),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "all_to_all_single("
      "Tensor input, "
      "SymInt[] output_split_sizes, "
      "SymInt[] input_split_sizes, "
      "str group_name) -> Tensor",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::all_to_all_single),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "broadcast(Tensor input, int src, str group_name) -> Tensor",
      torch::dispatch(c10::DispatchKey::CompositeExplicitAutograd, ::broadcast),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "broadcast_(Tensor(a!) input, int src, str group_name) -> Tensor(a!)",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::broadcast_),
      {at::Tag::pt2_compliant_tag});

  m.def(
      "wait_tensor(Tensor tensor) -> Tensor",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::wait_tensor),
      {at::Tag::pt2_compliant_tag});
}

namespace {
class AllToAllSingle : public torch::autograd::Function<AllToAllSingle> {
 public:
  static torch::autograd::Variable forward(
      torch::autograd::AutogradContext* ctx,
      const at::Tensor& input,
      // NOLINTNEXTLINE(performance-unnecessary-value-param)
      std::vector<int64_t> output_split_sizes,
      // NOLINTNEXTLINE(performance-unnecessary-value-param)
      std::vector<int64_t> input_split_sizes,
      // NOLINTNEXTLINE(performance-unnecessary-value-param)
      std::string group_name) {
    // swap sizes for backwards pass
    ctx->saved_data["output_split_sizes"] = input_split_sizes;
    ctx->saved_data["input_split_sizes"] = output_split_sizes;
    ctx->saved_data["group_name"] = group_name;

    return c10::Dispatcher::singleton()
        .findSchemaOrThrow("_c10d_functional::all_to_all_single", "")
        .typed<decltype(all_to_all_single)>()
        .call(input, output_split_sizes, input_split_sizes, group_name);
  }

  static torch::autograd::variable_list backward(
      torch::autograd::AutogradContext* ctx,
      torch::autograd::variable_list grad_out_list) {
    const std::vector<int64_t>& output_split_sizes =
        ctx->saved_data["output_split_sizes"].toIntVector();
    const std::vector<int64_t>& input_split_sizes =
        ctx->saved_data["input_split_sizes"].toIntVector();
    const std::string& group_name = ctx->saved_data["group_name"].toStringRef();

    DCHECK(grad_out_list.size() == 1);
    auto grad_out = grad_out_list[0].contiguous();

    auto out =
        c10::Dispatcher::singleton()
            .findSchemaOrThrow("_c10d_functional::all_to_all_single", "")
            .typed<decltype(all_to_all_single)>()
            .call(grad_out, output_split_sizes, input_split_sizes, group_name);

    // do an explicit wait to avoid cuda stream issues
    // TODO: track active cuda stream in wait
    out = c10::Dispatcher::singleton()
              .findSchemaOrThrow("_c10d_functional::wait_tensor", "")
              .typed<decltype(wait_tensor)>()
              .call(out);

    return {out, at::Tensor(), at::Tensor(), at::Tensor()};
  }
};

at::Tensor all_to_all_single_autograd(
    const at::Tensor& input,
    const std::vector<int64_t>& output_split_sizes,
    const std::vector<int64_t>& input_split_sizes,
    const std::string& group_name) {
  return AllToAllSingle::apply(
      input, output_split_sizes, input_split_sizes, group_name);
}

class ReduceScatterTensor
    : public torch::autograd::Function<ReduceScatterTensor> {
 public:
  static torch::autograd::Variable forward(
      torch::autograd::AutogradContext* ctx,
      const at::Tensor& input,
      std::string reduce_op,
      int64_t group_size,
      std::string group_name) {
    TORCH_CHECK(reduce_op == "sum", "Only sum reduce op is supported");

    ctx->saved_data["group_size"] = group_size;
    ctx->saved_data["group_name"] = group_name;

    return c10::Dispatcher::singleton()
        .findSchemaOrThrow("_c10d_functional::reduce_scatter_tensor", "")
        .typed<decltype(reduce_scatter_tensor)>()
        .call(input, reduce_op, group_size, group_name);
  }

  static torch::autograd::variable_list backward(
      torch::autograd::AutogradContext* ctx,
      torch::autograd::variable_list grad_out_list) {
    const int64_t group_size = ctx->saved_data["group_size"].toInt();
    const std::string& group_name = ctx->saved_data["group_name"].toStringRef();

    DCHECK(grad_out_list.size() == 1);
    auto grad_out = grad_out_list[0];

    auto out =
        c10::Dispatcher::singleton()
            .findSchemaOrThrow("_c10d_functional::all_gather_into_tensor", "")
            .typed<decltype(all_gather_into_tensor)>()
            .call(grad_out, group_size, group_name);

    // do an explicit wait to avoid cuda stream issues
    // TODO: track active cuda stream in wait
    out = c10::Dispatcher::singleton()
              .findSchemaOrThrow("_c10d_functional::wait_tensor", "")
              .typed<decltype(wait_tensor)>()
              .call(out);

    return {
        out,
        at::Tensor(),
        at::Tensor(),
        at::Tensor(),
    };
  }
};

at::Tensor reduce_scatter_tensor_autograd(
    const at::Tensor& input,
    std::string reduce_op,
    int64_t group_size,
    std::string group_name) {
  return ReduceScatterTensor::apply(input, reduce_op, group_size, group_name);
}

class AllGatherIntoTensor
    : public torch::autograd::Function<AllGatherIntoTensor> {
 public:
  static torch::autograd::Variable forward(
      torch::autograd::AutogradContext* ctx,
      const at::Tensor& input,
      int64_t group_size,
      std::string group_name) {
    ctx->saved_data["group_size"] = group_size;
    ctx->saved_data["group_name"] = group_name;

    return c10::Dispatcher::singleton()
        .findSchemaOrThrow("_c10d_functional::all_gather_into_tensor", "")
        .typed<decltype(all_gather_into_tensor)>()
        .call(input, group_size, group_name);
  }

  static torch::autograd::variable_list backward(
      torch::autograd::AutogradContext* ctx,
      torch::autograd::variable_list grad_out_list) {
    const int64_t group_size = ctx->saved_data["group_size"].toInt();
    const std::string& group_name = ctx->saved_data["group_name"].toStringRef();

    DCHECK(grad_out_list.size() == 1);
    auto grad_out = grad_out_list[0];

    auto out =
        c10::Dispatcher::singleton()
            .findSchemaOrThrow("_c10d_functional::reduce_scatter_tensor", "")
            .typed<decltype(reduce_scatter_tensor)>()
            .call(grad_out, "sum", group_size, group_name);

    // do an explicit wait to avoid cuda stream issues
    // TODO: track active cuda stream in wait
    out = c10::Dispatcher::singleton()
              .findSchemaOrThrow("_c10d_functional::wait_tensor", "")
              .typed<decltype(wait_tensor)>()
              .call(out);

    return {
        out,
        at::Tensor(),
        at::Tensor(),
    };
  }
};

at::Tensor all_gather_into_tensor_autograd(
    const at::Tensor& input,
    int64_t group_size,
    std::string group_name) {
  return AllGatherIntoTensor::apply(input, group_size, group_name);
}

} // namespace

TORCH_LIBRARY(_c10d_functional_autograd, m) {
  m.def(
      "all_to_all_single("
      "Tensor input, "
      "SymInt[] output_split_sizes, "
      "SymInt[] input_split_sizes, "
      "str group_name) -> Tensor",
      torch::dispatch(c10::DispatchKey::Autograd, ::all_to_all_single_autograd),
      {at::Tag::pt2_compliant_tag});
  m.def(
      "reduce_scatter_tensor("
      "Tensor input, "
      "str reduce_op, "
      "int group_size, "
      "str group_name) -> Tensor",
      torch::dispatch(
          c10::DispatchKey::Autograd, ::reduce_scatter_tensor_autograd),
      {at::Tag::pt2_compliant_tag});
  m.def(
      "all_gather_into_tensor("
      "Tensor input, "
      "int group_size, "
      "str group_name) -> Tensor",
      torch::dispatch(
          c10::DispatchKey::Autograd, ::all_gather_into_tensor_autograd),
      {at::Tag::pt2_compliant_tag});
}

namespace {
// DTensor related comm operations, sharing code with functional collective for
// now
at::Tensor shard_dim_alltoall(
    const at::Tensor& input,
    int64_t gather_dim,
    int64_t shard_dim,
    std::string group_name) {
  auto group = c10d::resolve_process_group(group_name);
  auto group_size = group->getSize();
  std::vector<int64_t> output_sizes = input.sizes().vec();
  if (output_sizes[shard_dim] % group_size != 0) {
    LOG(WARNING) << "The first dimension of the shard_dim_alltoall input ("
                 << output_sizes[shard_dim]
                 << ") is not divisible by the group size (" << group_size
                 << ").";
  }
  output_sizes[shard_dim] = output_sizes[shard_dim] / group_size;
  std::vector<at::Tensor> inputs;
  auto length = output_sizes[shard_dim];
  for (int i = 0; i < group_size; i++) {
    inputs.push_back(input.narrow(shard_dim, i * length, length).contiguous());
  }
  // allocate outputs
  std::vector<at::Tensor> outputs;
  for (int i = 0; i < group_size; i++) {
    outputs.push_back(input.new_empty(output_sizes).contiguous());
  }
  auto work = group->alltoall(outputs, inputs);

  work->wait();
  // TODO: it's very tricky to get the current async behavior work for shard dim
  // alltoall so for now we just keep this comm op to be synchronous. We can
  // revisit later how to support the async case with the Work registry.
  return at::cat(outputs, gather_dim);
}
} // namespace

// DTensor comm op registry
TORCH_LIBRARY(_dtensor, m) {
  m.def(
      "shard_dim_alltoall(Tensor input, int gather_dim, int shard_dim, str group_name) -> Tensor",
      torch::dispatch(
          c10::DispatchKey::CompositeExplicitAutograd, ::shard_dim_alltoall),
      {at::Tag::pt2_compliant_tag});
}