Module: Ignis::Collective::ArrayOps

Defined in:
lib/nvruby/collective/array_ops.rb

Overview

NvArray Integration for Collective Operations

Provides tensor-aware wrappers around collective primitives that work directly with NvArray tensors.

Class Method Summary collapse

Class Method Details

.all_gather(tensors, comm:, stream: nil) ⇒ Array<NvArray>

AllGather - each GPU contributes, all receive all

Parameters:

  • tensors (Array<NvArray>) —

    One NvArray per GPU (local chunks)

  • comm (Communicator) —

    Communicator instance

  • stream (CUDA::Stream, nil) (defaults to: nil) —

    Optional CUDA stream

Returns:

  • (Array<NvArray>) —

    NvArrays on all GPUs with gathered data



75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
 
# File 'lib/nvruby/collective/array_ops.rb', line 75

def all_gather(tensors, comm:, stream: nil)
  validate_tensors!(tensors)

  # Calculate total size after gathering
  n_gpus = tensors.size
  chunk_shape = tensors[0].shape.dup
  gathered_shape = chunk_shape.dup
  gathered_shape[0] = chunk_shape[0] * n_gpus  # Concatenate along first axis

  gathered_pointers = comm.all_gather(tensors, stream: stream)

  # Wrap results in NvArrays
  gathered_pointers.map.with_index do |ptr_or_arr, i|
    if ptr_or_arr.is_a?(Array)
      # Already NvArrays
      ptr_or_arr[i]
    else
      create_nvarray_from_ptr(
        ptr: ptr_or_arr,
        shape: gathered_shape,
        dtype: tensors[0].dtype,
        device: comm.gpu_ids[i]
      )
    end
  end
end

.all_reduce(tensors, op: :sum, comm:, stream: nil) ⇒ Array<NvArray>

AllReduce on NvArray tensors

Parameters:

  • tensors (Array<NvArray>) —

    One NvArray per GPU

  • op (Symbol) (defaults to: :sum) —

    Reduction operation (:sum, :prod, :min, :max)

  • comm (Communicator) —

    Communicator instance

  • stream (CUDA::Stream, nil) (defaults to: nil) —

    Optional CUDA stream

Returns:

  • (Array<NvArray>) —

    Modified tensors with reduced values



18
19
20
21
22
23
 
# File 'lib/nvruby/collective/array_ops.rb', line 18

def all_reduce(tensors, op: :sum, comm:, stream: nil)
  validate_tensors!(tensors)
  validate_devices!(tensors, comm)

  comm.all_reduce(tensors, op: op, stream: stream)
end

.broadcast(tensor, root: 0, comm:, stream: nil) ⇒ Array<NvArray>

Broadcast NvArray from root to all GPUs

Parameters:

  • tensor (NvArray) —

    Tensor on root GPU

  • root (Integer) (defaults to: 0) —

    Root rank

  • comm (Communicator) —

    Communicator instance

  • stream (CUDA::Stream, nil) (defaults to: nil) —

    Optional CUDA stream

Returns:

  • (Array<NvArray>) —

    NvArrays on all GPUs with same data



32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
 
# File 'lib/nvruby/collective/array_ops.rb', line 32

def broadcast(tensor, root: 0, comm:, stream: nil)
  validate_single_tensor!(tensor)

  # Comm.broadcast returns FFI::Pointers, wrap in NvArrays
  pointers = comm.broadcast(tensor, root: root, stream: stream)

  pointers.map.with_index do |ptr, i|
    if i == root
      tensor
    else
      create_nvarray_from_ptr(
        ptr: ptr,
        shape: tensor.shape,
        dtype: tensor.dtype,
        device: comm.gpu_ids[i]
      )
    end
  end
end

.reduce(tensors, root: 0, op: :sum, comm:, stream: nil) ⇒ NvArray

Reduce to root GPU

Parameters:

  • tensors (Array<NvArray>) —

    One NvArray per GPU

  • root (Integer) (defaults to: 0) —

    Root rank

  • op (Symbol) (defaults to: :sum) —

    Reduction operation

  • comm (Communicator) —

    Communicator instance

  • stream (CUDA::Stream, nil) (defaults to: nil) —

    Optional CUDA stream

Returns:

  • (NvArray) —

    Reduced tensor on root GPU



60
61
62
63
64
65
66
67
 
# File 'lib/nvruby/collective/array_ops.rb', line 60

def reduce(tensors, root: 0, op: :sum, comm:, stream: nil)
  validate_tensors!(tensors)

  result = comm.reduce(tensors, root: root, op: op, stream: stream)

  # result is already the root tensor or pointer
  result.is_a?(NvArray) ? result : tensors[root]
end

.reduce_scatter(tensors, op: :sum, comm:, stream: nil) ⇒ Array<NvArray>

ReduceScatter - reduce then scatter

Parameters:

  • tensors (Array<NvArray>) —

    One NvArray per GPU (full size)

  • op (Symbol) (defaults to: :sum) —

    Reduction operation

  • comm (Communicator) —

    Communicator instance

  • stream (CUDA::Stream, nil) (defaults to: nil) —

    Optional CUDA stream

Returns:

  • (Array<NvArray>) —

    Chunk NvArrays (1/N size each)



109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
 
# File 'lib/nvruby/collective/array_ops.rb', line 109

def reduce_scatter(tensors, op: :sum, comm:, stream: nil)
  validate_tensors!(tensors)

  n_gpus = tensors.size
  full_shape = tensors[0].shape.dup
  chunk_shape = full_shape.dup
  chunk_shape[0] = full_shape[0] / n_gpus

  chunk_pointers = comm.reduce_scatter(tensors, op: op, stream: stream)

  chunk_pointers.map.with_index do |ptr, i|
    create_nvarray_from_ptr(
      ptr: ptr,
      shape: chunk_shape,
      dtype: tensors[0].dtype,
      device: comm.gpu_ids[i]
    )
  end
end

.scatter(tensor, root: 0, comm:, stream: nil) ⇒ Array<NvArray>

Distribute tensor across GPUs (scatter)

Parameters:

  • tensor (NvArray) —

    Source tensor on root GPU

  • root (Integer) (defaults to: 0) —

    Root rank

  • comm (Communicator) —

    Communicator instance

  • stream (CUDA::Stream, nil) (defaults to: nil) —

    Optional CUDA stream

Returns:

  • (Array<NvArray>) —

    Chunks on each GPU



136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
 
# File 'lib/nvruby/collective/array_ops.rb', line 136

def scatter(tensor, root: 0, comm:, stream: nil)
  validate_single_tensor!(tensor)

  n_gpus = comm.gpu_ids.size
  full_shape = tensor.shape.dup
  chunk_shape = full_shape.dup
  chunk_shape[0] = full_shape[0] / n_gpus
  chunk_size = tensor.nbytes / n_gpus

  chunks = []

  n_gpus.times do |i|
    if i == root
      # Root keeps first chunk
      chunk_ptr = tensor.device_ptr
      chunks << create_nvarray_from_ptr(
        ptr: chunk_ptr,
        shape: chunk_shape,
        dtype: tensor.dtype,
        device: comm.gpu_ids[i]
      )
    else
      # Allocate and send chunk
      chunk_ptr = allocate_on_device(comm.gpu_ids[i], chunk_size)
      offset = i * chunk_size
      src_ptr = FFI::Pointer.new(:uint8, tensor.device_ptr.address + offset)

      comm.send_recv(
        src_ptr,
        src_rank: root,
        dst_buffer: chunk_ptr,
        dst_rank: i,
        size: chunk_size,
        stream: stream
      )

      chunks << create_nvarray_from_ptr(
        ptr: chunk_ptr,
        shape: chunk_shape,
        dtype: tensor.dtype,
        device: comm.gpu_ids[i]
      )
    end
  end

  chunks
end