fix: harden RDNA device property patch against OOM and attr loss

[on22s]

Summary

Three bugs in _patch_rdna_device_properties and the batch warmup paths:

• Device key TypeError — int(device) raises TypeError when get_device_properties is called with a torch.device object (e.g. torch.device('cuda:0')). Fixed by using dev.index, with fallbacks for unindexed devices and errors.

• SimpleNamespace silently drops C-extension attrs — dir(props) does not enumerate all C-extension attributes on the real props object, so any property not surfaced by dir() is missing from the patched result. Replaced with _RDNADeviceProps, a thin proxy that delegates all attribute lookups to the real props object and only overrides multi_processor_count and warp_size. Future PyTorch device property fields are forwarded automatically with no maintenance required.

• Warmup loads second model into VRAM (OOM) — _local_batch_clone and _local_batch_lora both called self._init_local_custom() as the warmup model target while the clone/LoRA model was already loaded. Two full models resident simultaneously causes OOM on 12–16 GB cards. Fixed by using the already-loaded model for warmup. For the LoRA path, the warmup block is moved inside the adapter loop so the LoRA model is resident before warmup runs.

Test plan

• Verify get_device_properties called with torch.device('cuda:0') no longer raises TypeError on RDNA hardware
• Verify patched props object forwards arbitrary attributes (e.g. props.total_memory, props.name) correctly to the underlying device
• Verify batch clone generation completes on a 12–16 GB RDNA card without OOM (previously required a restart after warmup)
• Verify batch LoRA generation completes with warmup running against the LoRA model, not a separately loaded CustomVoice model

🤖 Generated with Claude Code