fix: remove useless warmup from __init__ (model on CPU, no compilation)

Warmup at init time runs before Lightning moves the model to GPU,
so Triton cannot compile kernels there. Remove the calls entirely;
compilation happens naturally on the first real training step.

Also fix default glu_type to softsign_glu in both tasks.

Author: KakaruHayate

training/acoustic_task.py

@@ -101,21 +101,13 @@ def __init__(self):
 
         # ── Fuse LYNXNet2 backbone kernels (in-place) ──
         if hparams.get('use_fused_kernels', False):
-            from modules.kernels.integration import patch_diffusion_module, warmup_fused_backbone
+            from modules.kernels.integration import patch_diffusion_module
             from lightning.pytorch.utilities.rank_zero import rank_zero_info
             n = patch_diffusion_module(
                 self.model.diffusion,
                 glu_type=hparams['backbone_args'].get('glu_type', 'softsign_glu'),
             )
             rank_zero_info('Fused kernels: patched %d LYNXNet2 blocks', n)
-            if n > 0:
-                backbone = getattr(self.model.diffusion, 'denoise_fn',
-                                   getattr(self.model.diffusion, 'velocity_fn', None))
-                if backbone is not None:
-                    # Warmup with realistic M matching max_batch_frames
-                    warmup_M = hparams.get('max_batch_frames', 50000)
-                    warmup_fused_backbone(backbone, M=warmup_M)
-                rank_zero_info('Fused kernels: autotune complete')
 
     def _build_model(self):
         return DiffSingerAcoustic(

training/variance_task.py

@@ -125,9 +125,14 @@ def __init__(self):
         # small for Triton fusion to benefit. The compilation overhead outweighs
         # the HBM savings. Fused kernels for acoustic model only (K=1024).
         if hparams.get('use_fused_kernels', False):
-            from modules.kernels.integration import warmup_fused_backbone
+            from modules.kernels.integration import patch_variance_model
             from lightning.pytorch.utilities.rank_zero import rank_zero_info
-            rank_zero_info('Fused kernels: skipping variance model (small backbones, no benefit)')
+            n = patch_variance_model(
+                self.model,
+                glu_type=hparams.get('backbone_args', {}).get('glu_type', 'softsign_glu'),
+            )
+            rank_zero_info('Fused kernels: patched %d LYNXNet2 blocks in variance model', n)
+
 
     def _build_model(self):
         return DiffSingerVariance(