ignis-collective 0.0.1

data/lib/nvruby/collective.rb

@@ -0,0 +1,156 @@
+# frozen_string_literal: true
+
+require_relative 'collective/p2p_bindings'
+require_relative 'collective/vmm_bindings'
+require_relative 'collective/topology'
+require_relative 'collective/device_manager'
+require_relative 'collective/transport/base'
+require_relative 'collective/transport/p2p_transport'
+require_relative 'collective/transport/ipc_transport'
+require_relative 'collective/transport/vmm_ipc_transport'
+require_relative 'collective/transport/vmm_ipc_structs'
+require_relative 'collective/transport/host_staged_transport'
+require_relative 'collective/transport_selector'
+require_relative 'collective/algorithms/reduction_ops'
+require_relative 'collective/algorithms/ring'
+require_relative 'collective/algorithms/tree'
+require_relative 'collective/communicator'
+require_relative 'collective/array_ops'
+require_relative 'collective/nvarray_adapter'
+require_relative 'collective/health_monitor'
+require_relative 'collective/communicator_healer'
+require_relative 'collective/resilient_transport'
+require_relative 'collective/dynamic_optimizer'
+
+require "ignis"
+
+module Ignis
+  module Collective
+    # NvCCL (Ignis Collective Communications Library) version
+    # Note: NvCCL is NOT NCCL — this is an original design.
+    VERSION = '0.1.0'
+
+    # Custom error classes
+    class Error < StandardError; end
+    class TopologyError < Error; end
+    class TransportError < Error; end
+    class CommunicatorError < Error; end
+
+    @booted = false
+
+    class << self
+      # Boot the NvCCL collective layer.
+      #
+      # Enumerates CUDA devices, detects topology, registers
+      # RecoveryProtocol callbacks, subscribes to EventBus events,
+      # and starts the HealthMonitor.
+      #
+      # @return [void]
+      def boot!
+        return if @booted
+
+        register_recovery_callbacks!
+        subscribe_event_bus!
+        start_health_monitor!
+
+        @booted = true
+        $stderr.puts "[NvCCL] Booted — #{device_count} GPU(s) detected"
+      end
+
+      # @return [Boolean]
+      def booted?
+        @booted
+      end
+
+      # Create a new communicator for the specified GPUs.
+      #
+      # @param gpu_ids [Array<Integer>] GPU device IDs to include
+      # @param rank [Integer] Process rank (for multi-process, default 0)
+      # @param world_size [Integer] Total processes (default 1)
+      # @return [Communicator]
+      def create_communicator(gpu_ids:, rank: 0, world_size: 1)
+        Communicator.new(gpu_ids: gpu_ids, rank: rank, world_size: world_size)
+      end
+
+      # Detect GPU topology for all available devices.
+      # @return [Topology::Detector]
+      def detect_topology
+        Topology::Detector.new
+      end
+
+      # Get list of available GPUs.
+      # @return [Array<CUDA::Device>]
+      def available_devices
+        CUDA::Device.list
+      rescue StandardError
+        []
+      end
+
+      # @return [Integer]
+      def device_count
+        CUDA::Device.count
+      rescue StandardError
+        0
+      end
+
+      # Check if multi-GPU is available.
+      # @return [Boolean]
+      def multi_gpu_available?
+        device_count > 1
+      end
+
+      private
+
+      # Register NvCCL's RecoveryProtocol callbacks.
+      # @return [void]
+      def register_recovery_callbacks!
+        Ignis::Shared::RecoveryProtocol.register(
+          layer: :nvccl,
+          on_degraded: ->(**kwargs) {
+            gpu_id = kwargs[:gpu_id]
+            $stderr.puts "[NvCCL] GPU #{gpu_id} degraded — suspending collective operations"
+          },
+          on_recovering: ->(**kwargs) {
+            $stderr.puts "[NvCCL] Entering recovery — rebuilding communicator"
+          },
+          on_healthy: ->(**kwargs) {
+            recovered = kwargs[:recovered_gpus] || []
+            $stderr.puts "[NvCCL] Recovery complete — active GPUs: #{recovered}"
+            Ignis::Shared::EventBus.publish(:topology_changed, payload: { active_gpus: recovered })
+          },
+          on_failed: ->(**_kwargs) {
+            $stderr.puts "[NvCCL] CRITICAL — recovery failed, NvCCL entering FAILED state"
+          }
+        )
+      end
+
+      # Subscribe to EventBus events that NvCCL handles.
+      # @return [void]
+      def subscribe_event_bus!
+        # When compute is done, NvCCL may need to do an all-reduce
+        Ignis::Shared::EventBus.subscribe(:compute_done, handler_id: :nvccl_compute) do |payload|
+          # NvCCL will check if an all-reduce is needed based on training config
+          # For now, just track the event
+        end
+
+        # WNAIS backpressure signals
+        Ignis::Shared::EventBus.subscribe(:backpressure_on, handler_id: :nvccl_bp_on) do |_payload|
+          $stderr.puts "[NvCCL] Backpressure received — throttling collective operations"
+        end
+
+        Ignis::Shared::EventBus.subscribe(:backpressure_off, handler_id: :nvccl_bp_off) do |_payload|
+          $stderr.puts "[NvCCL] Backpressure released — resuming collective operations"
+        end
+      end
+
+      # Start the HealthMonitor background thread.
+      # @return [void]
+      def start_health_monitor!
+        if defined?(HealthMonitor)
+          @health_monitor = HealthMonitor.new
+          # HealthMonitor will publish :health_alert and :gpu_failed events
+        end
+      end
+    end
+  end
+end

metadata

@@ -0,0 +1,92 @@
+--- !ruby/object:Gem::Specification
+name: ignis-collective
+version: !ruby/object:Gem::Version
+  version: 0.0.1
+platform: ruby
+authors:
+- NNW / Ignis contributors
+bindir: bin
+cert_chain: []
+date: 1980-01-02 00:00:00.000000000 Z
+dependencies:
+- !ruby/object:Gem::Dependency
+  name: ignis
+  requirement: !ruby/object:Gem::Requirement
+    requirements:
+    - - '='
+      - !ruby/object:Gem::Version
+        version: 0.0.1
+  type: :runtime
+  prerelease: false
+  version_requirements: !ruby/object:Gem::Requirement
+    requirements:
+    - - '='
+      - !ruby/object:Gem::Version
+        version: 0.0.1
+description: |
+  ignis-collective brings NCCL-style collective communication to Ruby: ring/tree
+  all-reduce, and P2P / IPC (VMM) / host-staged / TCP transports, on the Ignis
+  foundation. EXPERIMENTAL — the transports require multiple GPUs / nodes and
+  cannot be exercised on a single GPU, so this is shipped separately from the
+  verified single-GPU stack.
+executables: []
+extensions: []
+extra_rdoc_files: []
+files:
+- README.md
+- lib/ignis-collective.rb
+- lib/nvruby/collective.rb
+- lib/nvruby/collective/algorithms/double_binary_tree.rb
+- lib/nvruby/collective/algorithms/pipeliner.rb
+- lib/nvruby/collective/algorithms/reduction_ops.rb
+- lib/nvruby/collective/algorithms/ring.rb
+- lib/nvruby/collective/algorithms/topology_router.rb
+- lib/nvruby/collective/algorithms/tree.rb
+- lib/nvruby/collective/array_ops.rb
+- lib/nvruby/collective/communicator.rb
+- lib/nvruby/collective/communicator_healer.rb
+- lib/nvruby/collective/device_manager.rb
+- lib/nvruby/collective/dynamic_optimizer.rb
+- lib/nvruby/collective/health_monitor.rb
+- lib/nvruby/collective/net/nd_adapter.rb
+- lib/nvruby/collective/net/nd_bindings.rb
+- lib/nvruby/collective/net/rdma_transport.rb
+- lib/nvruby/collective/nvarray_adapter.rb
+- lib/nvruby/collective/p2p_bindings.rb
+- lib/nvruby/collective/resilient_transport.rb
+- lib/nvruby/collective/topology.rb
+- lib/nvruby/collective/transport/base.rb
+- lib/nvruby/collective/transport/host_staged_transport.rb
+- lib/nvruby/collective/transport/ipc_transport.rb
+- lib/nvruby/collective/transport/p2p_transport.rb
+- lib/nvruby/collective/transport/rdma_transports.rb
+- lib/nvruby/collective/transport/rio_transport.rb
+- lib/nvruby/collective/transport/tcp_transport.rb
+- lib/nvruby/collective/transport/vmm_ipc_structs.rb
+- lib/nvruby/collective/transport/vmm_ipc_transport.rb
+- lib/nvruby/collective/transport_selector.rb
+- lib/nvruby/collective/vmm_bindings.rb
+homepage: https://github.com/tigel-agm/Ignis
+licenses:
+- MIT
+metadata:
+  source_code_uri: https://github.com/tigel-agm/Ignis
+  rubygems_mfa_required: 'true'
+rdoc_options: []
+require_paths:
+- lib
+required_ruby_version: !ruby/object:Gem::Requirement
+  requirements:
+  - - ">="
+    - !ruby/object:Gem::Version
+      version: '3.1'
+required_rubygems_version: !ruby/object:Gem::Requirement
+  requirements:
+  - - ">="
+    - !ruby/object:Gem::Version
+      version: '0'
+requirements: []
+rubygems_version: 3.6.9
+specification_version: 4
+summary: Multi-GPU collective communication (NvCCL) for Ruby — EXPERIMENTAL
+test_files: []