Persistent thread pool for multi-GPU CFG splitting

[Kosinkadink]

Summary

Replaces the per-step thread create/destroy pattern in _calc_cond_batch_multigpu with a persistent MultiGPUThreadPool. Each worker thread calls torch.cuda.set_device() once at startup, preserving compiled kernel caches (inductor/triton) across diffusion steps.

Motivation

For models like SD3.5 large (fp8), thread creation/destruction every diffusion step causes cold compilation caches, leading to significant GPU idle time. Persistent threads keep the CUDA context and compiled kernels warm.

Changes

• comfy/multigpu.py: Added MultiGPUThreadPool class with one persistent worker per extra GPU, using queue.Queue for dispatch and result collection
• comfy/samplers.py:
  • Pool created in CFGGuider.outer_sample() and stored in model_options["multigpu_thread_pool"]
  • Pool shut down in the finally block alongside cleanup
  • _calc_cond_batch_multigpu submits extra GPU work to pool, main thread handles its own device directly (parallel execution)
  • Falls back gracefully if no pool is available

Testing

Benchmarked with NetaYumev35 (Lumina2) on 2x RTX 4090, 1024x1024, 30 steps:

Metric	Before (thread-per-step)	After (persistent pool)	Δ
it/s	5.55	5.73	+3.2%
Time	5.83s	5.65s	-3.1%

5 consecutive runs with zero errors.

API Node PR Checklist

Scope

• Is API Node Change

Pricing & Billing

• Need pricing update
• No pricing update

If Need pricing update:

• Metronome rate cards updated
• Auto‑billing tests updated and passing

QA

• QA done
• QA not required

Comms

• Informed Kosinkadink