fix(fairseq): hubert load model error

infer/lib/train/utils.py

@@ -29,6 +29,24 @@ def load_checkpoint(checkpoint_path, model, optimizer=None, load_opt=1):
     checkpoint_dict = torch.load(checkpoint_path, map_location="cpu", weights_only=True)
 
     saved_state_dict = checkpoint_dict["model"]
+    # Convert old-style weight_norm keys (weight_g/weight_v) to new
+    # parametrizations format (parametrizations.weight.original0/original1)
+    # so that checkpoints saved with the deprecated API can still be loaded.
+    _converted = {}
+    for k, v in list(saved_state_dict.items()):
+        if k.endswith(".weight_g"):
+            new_key = k[: -len(".weight_g")] + ".parametrizations.weight.original0"
+            _converted[new_key] = v
+        elif k.endswith(".weight_v"):
+            new_key = k[: -len(".weight_v")] + ".parametrizations.weight.original1"
+            _converted[new_key] = v
+    if _converted:
+        logger.info(
+            "Converting %d old-style weight_norm keys from checkpoint to new parametrizations format",
+            len(_converted),
+        )
+        saved_state_dict.update(_converted)
+
     if hasattr(model, "module"):
         state_dict = model.module.state_dict()
     else:

infer/modules/train/train.py

@@ -29,10 +29,12 @@ try:
 
         GradScaler = gradscaler_init()
         ipex_init()
-    else:
-        from torch.cuda.amp import GradScaler, autocast
 except Exception:
-    from torch.cuda.amp import GradScaler, autocast
+    pass
+finally:
+    if not ('GradScaler' in globals() and 'autocast' in globals()):
+        from torch.amp.grad_scaler import GradScaler
+        from torch.amp.autocast_mode import autocast
 
 torch.backends.cudnn.deterministic = False
 torch.backends.cudnn.benchmark = False
@@ -535,7 +537,7 @@ def train_and_evaluate(
             # wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
 
         # Calculate
-        with autocast(enabled=hps.train.fp16_run):
+        with autocast(device_type="cuda", enabled=hps.train.fp16_run):
             (
                 y_hat,
                 ids_slice,
@@ -554,7 +556,7 @@ def train_and_evaluate(
             y_mel = slice_on_last_dim(
                 mel, ids_slice, hps.train.segment_size // hps.data.hop_length
             )
-            with autocast(enabled=False):
+            with autocast(device_type="cuda", enabled=False):
                 y_hat_mel = mel_spectrogram_torch(
                     y_hat.float().squeeze(1),
                     hps.data.filter_length,
@@ -573,7 +575,7 @@ def train_and_evaluate(
 
             # Discriminator
             y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach())
-            with autocast(enabled=False):
+            with autocast(device_type="cuda", enabled=False):
                 loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(
                     y_d_hat_r, y_d_hat_g
                 )
@@ -583,10 +585,10 @@ def train_and_evaluate(
         grad_norm_d = total_grad_norm(net_d.parameters())
         scaler.step(optim_d)
 
-        with autocast(enabled=hps.train.fp16_run):
+        with autocast(device_type="cuda", enabled=hps.train.fp16_run):
             # Generator
             y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
-            with autocast(enabled=False):
+            with autocast(device_type="cuda", enabled=False):
                 loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
                 loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
                 loss_fm = feature_loss(fmap_r, fmap_g)

infer/modules/vc/hash.py

@@ -10,9 +10,6 @@ from pybase16384 import encode_to_string, decode_from_string
 from configs import CPUConfig
 from rvc.synthesizer import get_synthesizer
 
-from .pipeline import Pipeline
-from .utils import load_hubert
-
 
 class TorchSeedContext:
     def __init__(self, seed):
@@ -95,6 +92,9 @@ def wave_hash(time_field):
 
 
 def model_hash(config, tgt_sr, net_g, if_f0, version):
+    from .pipeline import Pipeline
+    from .utils import load_hubert
+
     pipeline = Pipeline(tgt_sr, config)
     audio = original_audio()
     hbt = load_hubert(config.device, config.is_half)

infer/modules/vc/utils.py

@@ -1,6 +1,7 @@
 import os, pathlib
 
-from fairseq import checkpoint_utils
+import torch
+from fairseq import checkpoint_utils, data
 
 
 def get_index_path_from_model(sid):
@@ -21,10 +22,11 @@ def get_index_path_from_model(sid):
 
 
 def load_hubert(device, is_half):
-    models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
-        ["assets/hubert/hubert_base.pt"],
-        suffix="",
-    )
+    with torch.serialization.safe_globals([data.dictionary.Dictionary]):
+        models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
+            ["assets/hubert/hubert_base.pt"],
+            suffix="",
+        )
     hubert_model = models[0]
     hubert_model = hubert_model.to(device)
     if is_half: