Enables dynamic GPU allocation for local workloads

[allenwang28]

Context for this PR - I want to be able to demo a service that spawns N replicas of vLLM. For that to work, we need dynamic GPU allocation. Our current approach hard codes GPU assignment.

So this PR does a few things:

• Introduces GpuResourceManager, a singleton actor (or "controller") responsible for tracking and releasing GPUs. This only works for local proc mesh for now, will expand to multi-host once/if needed. Since it's a singleton, there's no concern about race conditions etc.
• Updates proc and service types to now track num_gpus too, which is the signal for receiving GPU ids
• Adds stubs for the new hostmesh APIs - currently broken for some NYI stuff in Monarch. Probably need to work through a re-design of procmesh management/allocation
• Updates applications (sft_v2, rl, grpo_main, vllm) accordingly

Next up:

• Create a policy specific Replica which will use get_proc_mesh() correctly

[Copilot]

Pull Request Overview

This PR introduces dynamic GPU allocation for local workloads by implementing a GPU resource manager that tracks and distributes GPU resources instead of hard-coding GPU assignments. This enables spawning multiple replicas of GPU-intensive services like vLLM.

• Adds GpuResourceManager singleton actor to track and allocate GPU resources
• Updates process and service types to include num_gpus field for GPU allocation requests
• Integrates GPU allocation into process mesh creation and service configuration

Reviewed Changes

Copilot reviewed 16 out of 19 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/unit_tests/test_service.py	Updates test configurations to specify gpus_per_replica=0 for CPU-only tests
tests/unit_tests/test_gpu_manager.py	Comprehensive test suite for GPU manager allocation and release functionality
src/forge/types.py	Adds num_gpus field to ProcessConfig and gpus_per_replica to ServiceConfig
src/forge/controller/service/spawn.py	Updates import paths after service module restructuring
src/forge/controller/service/service.py	Updates import paths after service module restructuring
src/forge/controller/service/metrics.py	Updates import paths after service module restructuring
src/forge/controller/service/init.py	Creates service module package with proper exports
src/forge/controller/proc_mesh.py	Integrates GPU allocation into process mesh creation and environment setup
src/forge/controller/custom_actors/service_registry.py	Stub implementation for future service tracking functionality
src/forge/controller/custom_actors/gpu_manager.py	Core GPU manager actor implementation with allocation and release logic
src/forge/controller/custom_actors/init.py	Exports GPU manager utility functions
src/forge/controller/init.py	Updates exports after service module restructuring
apps/sft_v2/main.py	Updates comment to reflect correct module path
apps/sft_v2/llama3_8b.yaml	Adds GPU allocation configuration and fixes tokenizer path
apps/rl/llama3_8b.yaml	Adds GPU allocation configuration for trainer and replay buffer
apps/grpo/main.py	Adds GPU allocation to service configuration and removes hardcoded device assignment

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[pbontrager]

Theses are a lot of changes but they seem to be mostly renaming and import changes. I like the addition of the GPU manager.