Replace the hardcoded `_force_fp32_cpu_compute` helper with`_force_supported_compute_dtype`, which delegates to`comfy.model_management.unet_manual_cast(weight_dtype, device)`. The interrogator already encodes per-device dtype support (CPU returns False for fp16/bf16, older GPUs may not support bf16, pre-14 MPS doesn't support bf16, etc.) and returns None when no cast is needed.For SelectModelDevice -> CPU on an fp16/bf16 model, behavior is unchanged: `unet_manual_cast` returns `torch.float32` and `set_model_compute_dtype` casts at use without touching peak memory. As a bonus the same code path now handles other `weight_dtype not supported on device` cases (e.g. bf16 weights on pre-Ampere NVIDIA, bf16 on pre-macOS-14 MPS) without growing the code surface, so the call site no longer needs the `if resolved.type == 'cpu':` gate.
Amp-Thread-ID: https://ampcode.com/threads/T-019e61db-ffb1-73a6-b2a8-3d23d7b05792
Co-authored-by: Amp <amp@ampcode.com>
