Avoid YuEGP pinned-memory exhaustion on 32GB Windows hosts

scripts/yuegp-worker.py

@@ -61,6 +61,16 @@ def load_codec_on_cpu(ns, torch):
     return codec
 
 
+def memory_profile_options(profile, compile_enabled, platform, total_ram):
+    options = dict(profile_no=profile, quantizeTransformer=profile == 3,
+                   compile=compile_enabled, verboseLevel=1)
+    # Windows CUDA pinned allocations can fail before inference on 32GB hosts.
+    # This affects transfer staging only, not weight precision or VRAM budgets.
+    if platform == 'win32' and total_ram < 48 * 1024 ** 3:
+        options['pinnedMemory'] = False
+    return options
+
+
 def main():
     import psutil
     parent = psutil.Process(os.getppid())
@@ -126,8 +136,16 @@ def main():
         model2.generation_config.cache_implementation = None
     model._validate_model_kwargs = lambda _: None
     model2._validate_model_kwargs = lambda _: None
-    offloader = ns['offload'].profile({'transformer': model, 'stage2': model2}, profile_no=cli.profile,
-                         quantizeTransformer=cli.profile == 3, compile=compile_enabled, verboseLevel=1)
+    profile_options = memory_profile_options(cli.profile, compile_enabled, sys.platform, psutil.virtual_memory().total)
+    emit(stage='loading', message='Preparing YuEGP memory', profileOptions=profile_options)
+    offloader = ns['offload'].profile({'transformer': model, 'stage2': model2}, **profile_options)
+    def memory_checkpoint(stage):
+        torch.cuda.synchronize()
+        free, total = torch.cuda.mem_get_info()
+        emit(stage=stage, message=f'YuEGP {stage}', gpuFreeMiB=round(free / 1024 ** 2),
+             gpuTotalMiB=round(total / 1024 ** 2), gpuAllocatedMiB=round(torch.cuda.memory_allocated() / 1024 ** 2),
+             ramAvailableMiB=round(psutil.virtual_memory().available / 1024 ** 2))
+    memory_checkpoint('memory-ready')
     args = SimpleNamespace(use_audio_prompt=False, use_dual_tracks_prompt=False, rescale=True,
                            output_dir=str(output), cuda_idx=0)
     ns.update(model=model, model_stage2=model2, device=device, codec_model=None, args=args,
@@ -150,6 +168,7 @@ def main():
              progress=round(100 * done / max(1, total), 1))
     emit(stage='stage1', message='Generating song tokens', progress=0)
     stems = ns['stage1_inference'](tags, lyrics, 1, max_tokens, seed, state, callback)
+    memory_checkpoint('stage1-finished')
     emit(stage='stage2', message='Generating audio detail', progress=0)
     results = ns['stage2_inference'](model2, stems, str(output / 'stage2'),
                                     batch_size=20 if cli.profile == 1 else 4, state=state, callback=callback)