[RFC] Fine-tuning on Intel hardware in the PyTorch ecosystem — where to invest?

[DamianSzwichtenberg]

Motivation

We (Intel) want to enable fine-tuning on Intel GPU hardware within the PyTorch ecosystem. We started with torchforge and have been contributing device-agnostic improvements. Now we need to understand where to focus next.

Work completed

What	Status	Description
#749 — Make SFT hardware-agnostic	Merged	Replaced torch.cuda.* with torch.accelerator.*, introduced DeviceProxy for device counting and env var mapping across backends. Updated tests accordingly.
#760 — XPU: add install script and docs	Open (awaiting review)	Adds scripts/install_xpu.sh following the install_rocm.sh pattern.
#759 — SFT checkpoint bug	Open	Bug found during testing, reported upstream.

Next goal: GRPO — but where?

Our next target is enabling the GRPO workflow on Intel hardware. GRPO has a deeper stack than SFT — it relies on Monarch actors, TorchStore (RDMA-based weight sync), and vLLM. Some of these have CUDA-specific paths.

The question is where this work should land. We've noticed that torchforge activity has slowed down, while torchtitan has added its own RL support via experiments/rl — an alternative GRPO implementation using the same core dependencies (Monarch, TorchStore, vLLM) but directly within torchtitan.

We'd like to understand: is torchforge still the right place to invest, or should we shift our fine-tuning enablement efforts toward torchtitan?

Any guidance from maintainers would be greatly appreciated.

/cc @felipemello1