[Copilot]

Setting tensor_parallel_size higher than available XPU devices causes a segfault (null deref in Level Zero driver) because the device count validation in ParallelConfig only checked CUDA platforms. XPU workers spawn on non-existent device IDs with no guard.

• Add device_count() classmethod to base Platform class (returns 0; already overridden on CUDA/ROCm/XPU subclasses)
• Extend the world_size > device_count validation and multi-node nnodes > 1 check to cover XPU
• Replace CUDA-specific cuda_device_count_stateless() with generic current_platform.device_count() in the validation path

With this change, XPU users get a clear ValueError instead of a driver-level segfault:

ValueError: World size (4) is larger than the number of available GPUs (2) in this node.

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• astral.sh
  • Triggering command: /usr/bin/curl curl -LsSf REDACTED s/all_reduce_utils.py s/__�� s/custom_all_reduce.py s/cuda_communicator.py s/quick_all_reduce.py s/symm_mem.py s/cuda_wrapper.p--norc s/base_device_co--noprofile s/all2all.py s/sh�� s/ray_communicator.py s/cpu_communicator.py (dns block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)

---

⚡ Quickly spin up Copilot coding agent tasks from anywhere on your macOS or Windows machine with Raycast.