Tighten YuE for 16GB VRAM and stop the false Waiting status.

Cap Stage-A segments and use a heavier mmgp profile so long lyrics stop OOMing mid-run, and show that Comfy is actually working.

Co-authored-by: Cursor <cursoragent@cursor.com>

server/assets/workflow_yue.json

@@ -5,9 +5,9 @@
       "xcodec_ckpt": "ckpt_00360000.pth",
       "quantization_model": "fp16",
       "use_mmgp": true,
-      "stage1_cache_size": 16384,
+      "stage1_cache_size": 8192,
       "exllamav2_cache_mode": "FP16",
-      "mmgp_profile": 2
+      "mmgp_profile": 4
     },
     "class_type": "YUE_Stage_A_Loader",
     "_meta": {
@@ -23,7 +23,7 @@
       "repetition_penalty": 1.1,
       "prompt_start_time": 0,
       "prompt_end_time": 30,
-      "max_new_tokens": 3000,
+      "max_new_tokens": 2000,
       "use_dual_tracks_prompt": false,
       "use_audio_prompt": false,
       "offload_model": true,
@@ -38,8 +38,8 @@
   "16": {
     "inputs": {
       "stage_B_repo": "m-a-p/YuE-s2-1B-general",
-      "stage2_cache_size": 8192,
-      "stage2_batch_size": 2,
+      "stage2_cache_size": 4096,
+      "stage2_batch_size": 1,
       "exllamav2_cache_mode": "FP16",
       "use_mmgp": true,
       "info": ["8", 1]