feat: default to old PyTorch weight_norm for RVC fork compat + fix FAISS AVX2

BF667 · BF667 · commit de1b2aaee0aa · 2026-05-29T17:06:01.000Z
- Add arvc/engine/models/weight_norm.py wrapper module:
  - Default: torch.nn.utils.weight_norm (old-style, .weight_g/.weight_v)
    compatible with virtually all RVC forks
  - New mode: torch.nn.utils.parametrizations.weight_norm (PyTorch 2.0+)
    Enable by setting new_pytorch_weight_norm: true in config.json

- Update 20+ model files to use weight_norm wrapper instead of
  direct torch.nn.utils.parametrizations imports

- Make checkpoint key conversion conditional:
  - Old mode: no conversion needed (model uses .weight_g/.weight_v directly)
  - New mode: convert between old checkpoint format and new parametrizations
    format on load/save (same behavior as before)

- Fix ModuleNotFoundError for faiss.swigfaiss_avx2:
  Add fallback import in models/utils.py and create_index.py that
  catches the AVX2 module error and falls back to non-AVX2 faiss
diff --git a/arvc/configs/config.json b/arvc/configs/config.json
@@ -614,5 +614,6 @@
     "pretrained_v2_url": "https://huggingface.co/buckets/R-Kentaren/Ultimate-RVC-Models/resolve/pretrained_v2/",
     "pretrained_v1_fallback_url": "https://huggingface.co/buckets/R-Kentaren/Ultimate-RVC-Models/resolve/pretrained_v1/",
     "pretrained_v2_fallback_url": "https://huggingface.co/buckets/R-Kentaren/Ultimate-RVC-Models/resolve/pretrained_v2/",
-    "search_api_url": "https://voice-models.com/fetch_data.php"
+    "search_api_url": "https://voice-models.com/fetch_data.php",
+    "new_pytorch_weight_norm": false
 }
diff --git a/arvc/engine/models/algorithms/discriminators.py b/arvc/engine/models/algorithms/discriminators.py
@@ -2,7 +2,8 @@
 import torch.nn.functional as F
 
 from torch.utils.checkpoint import checkpoint
-from torch.nn.utils.parametrizations import spectral_norm, weight_norm
+from torch.nn.utils.parametrizations import spectral_norm
+from arvc.engine.models.weight_norm import weight_norm
 
 from arvc.engine.models.algorithms.commons import get_padding
 from arvc.engine.models.algorithms.residuals import LRELU_SLOPE
diff --git a/arvc/engine/models/algorithms/generators/pcph_gan.py b/arvc/engine/models/algorithms/generators/pcph_gan.py
@@ -7,8 +7,7 @@
 from torch import Tensor
 from torch.nn import Conv1d, ConvTranspose1d
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.amp import autocast
 from torch.utils.checkpoint import checkpoint
 
diff --git a/arvc/engine/models/algorithms/generators/ringformer.py b/arvc/engine/models/algorithms/generators/ringformer.py
@@ -4,8 +4,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.nn import Conv1d, ConvTranspose1d
 from torch.amp import autocast
 from torch.utils.checkpoint import checkpoint
diff --git a/arvc/engine/models/algorithms/modules.py b/arvc/engine/models/algorithms/modules.py
@@ -2,6 +2,7 @@
 
 import torch.nn.utils.parametrize as parametrize
 
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from .commons import fused_add_tanh_sigmoid_multiply
 
 class WaveNet(torch.nn.Module):
@@ -28,7 +29,7 @@ def __init__(
         self.drop = torch.nn.Dropout(p_dropout)
 
         if gin_channels: 
-            self.cond_layer = torch.nn.utils.parametrizations.weight_norm(
+            self.cond_layer = weight_norm(
                 torch.nn.Conv1d(
                     gin_channels, 
                     2 * hidden_channels * n_layers, 
@@ -42,7 +43,7 @@ def __init__(
 
         for i in range(n_layers):
             self.in_layers.append(
-                torch.nn.utils.parametrizations.weight_norm(
+                weight_norm(
                     torch.nn.Conv1d(
                         hidden_channels, 
                         2 * hidden_channels, 
@@ -57,7 +58,7 @@ def __init__(
             res_skip_channels = (hidden_channels if i == n_layers - 1 else 2 * hidden_channels)
 
             self.res_skip_layers.append(
-                torch.nn.utils.parametrizations.weight_norm(
+                weight_norm(
                     torch.nn.Conv1d(
                         hidden_channels, 
                         res_skip_channels, 
@@ -86,12 +87,12 @@ def forward(self, x, x_mask, g=None):
     def remove_weight_norm(self):
         if self.gin_channels != 0: 
             if hasattr(self.cond_layer, "parametrizations") and "weight" in self.cond_layer.parametrizations: parametrize.remove_parametrizations(self.cond_layer, "weight", leave_parametrized=True)
-            else: torch.nn.utils.remove_weight_norm(self.cond_layer)
+            else: remove_weight_norm(self.cond_layer)
 
         for l in self.in_layers:
             if hasattr(l, "parametrizations") and "weight" in l.parametrizations: parametrize.remove_parametrizations(l, "weight", leave_parametrized=True)
-            else: torch.nn.utils.remove_weight_norm(l)
+            else: remove_weight_norm(l)
 
         for l in self.res_skip_layers:
             if hasattr(l, "parametrizations") and "weight" in l.parametrizations: parametrize.remove_parametrizations(l, "weight", leave_parametrized=True)
-            else: torch.nn.utils.remove_weight_norm(l)
+            else: remove_weight_norm(l)
diff --git a/arvc/engine/models/algorithms/normalization.py b/arvc/engine/models/algorithms/normalization.py
@@ -6,8 +6,7 @@
 import torch.nn.functional as F
 
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 
 
 class LayerNorm(torch.nn.Module):
diff --git a/arvc/engine/models/algorithms/normalizing_flows.py b/arvc/engine/models/algorithms/normalizing_flows.py
@@ -5,8 +5,7 @@
 import torch.nn as nn
 
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 
 from arvc.engine.models.algorithms.modules import WaveNet
 from arvc.engine.models.algorithms.encoders import Encoder
diff --git a/arvc/engine/models/algorithms/residuals.py b/arvc/engine/models/algorithms/residuals.py
@@ -3,8 +3,7 @@
 import torch.nn.utils.parametrize as parametrize
 
 from itertools import chain
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 
 from .modules import WaveNet
 from .commons import get_padding, init_weights
diff --git a/arvc/engine/models/algorithms/wavenet.py b/arvc/engine/models/algorithms/wavenet.py
@@ -2,6 +2,7 @@
 
 import torch.nn.utils.parametrize as parametrize
 
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from .commons import fused_add_tanh_sigmoid_multiply
 
 class WaveNet(torch.nn.Module):
@@ -28,7 +29,7 @@ def __init__(
         self.drop = torch.nn.Dropout(p_dropout)
 
         if gin_channels: 
-            self.cond_layer = torch.nn.utils.parametrizations.weight_norm(
+            self.cond_layer = weight_norm(
                 torch.nn.Conv1d(
                     gin_channels, 
                     2 * hidden_channels * n_layers, 
@@ -42,7 +43,7 @@ def __init__(
 
         for i in range(n_layers):
             self.in_layers.append(
-                torch.nn.utils.parametrizations.weight_norm(
+                weight_norm(
                     torch.nn.Conv1d(
                         hidden_channels, 
                         2 * hidden_channels, 
@@ -57,7 +58,7 @@ def __init__(
             res_skip_channels = (hidden_channels if i == n_layers - 1 else 2 * hidden_channels)
 
             self.res_skip_layers.append(
-                torch.nn.utils.parametrizations.weight_norm(
+                weight_norm(
                     torch.nn.Conv1d(
                         hidden_channels, 
                         res_skip_channels, 
@@ -86,12 +87,12 @@ def forward(self, x, x_mask, g=None):
     def remove_weight_norm(self):
         if self.gin_channels != 0: 
             if hasattr(self.cond_layer, "parametrizations") and "weight" in self.cond_layer.parametrizations: parametrize.remove_parametrizations(self.cond_layer, "weight", leave_parametrized=True)
-            else: torch.nn.utils.remove_weight_norm(self.cond_layer)
+            else: remove_weight_norm(self.cond_layer)
 
         for l in self.in_layers:
             if hasattr(l, "parametrizations") and "weight" in l.parametrizations: parametrize.remove_parametrizations(l, "weight", leave_parametrized=True)
-            else: torch.nn.utils.remove_weight_norm(l)
+            else: remove_weight_norm(l)
 
         for l in self.res_skip_layers:
             if hasattr(l, "parametrizations") and "weight" in l.parametrizations: parametrize.remove_parametrizations(l, "weight", leave_parametrized=True)
-            else: torch.nn.utils.remove_weight_norm(l)
+            else: remove_weight_norm(l)
diff --git a/arvc/engine/models/embedders/fairseq.py b/arvc/engine/models/embedders/fairseq.py
@@ -4,6 +4,8 @@
 import uuid
 import torch
 import types
+
+from arvc.engine.models.weight_norm import weight_norm as _weight_norm, remove_weight_norm as _remove_weight_norm
 import contextlib
 
 import numpy as np
@@ -558,7 +560,7 @@ def make_conv_pos(e, k, g):
     dropout = 0
     nn.init.normal_(pos_conv.weight, mean=0, std=math.sqrt((4 * (1.0 - dropout)) / (k * e)))
     nn.init.constant_(pos_conv.bias, 0)
-    return nn.Sequential(nn.utils.parametrizations.weight_norm(pos_conv, name="weight", dim=2), SamePad(k), nn.GELU())
+    return nn.Sequential(_weight_norm(pos_conv, name="weight", dim=2), SamePad(k), nn.GELU())
 
 def is_xla_tensor(tensor):
     return torch.is_tensor(tensor) and tensor.device.type == "xla"
@@ -1183,7 +1185,7 @@ def make_generation_fast_(self, **kwargs):
 
         def apply_remove_weight_norm(module):
             try:
-                nn.utils.remove_weight_norm(module)
+                _remove_weight_norm(module)
             except (AttributeError, ValueError):
                 return
 
diff --git a/arvc/engine/models/generators/bigvgan.py b/arvc/engine/models/generators/bigvgan.py
@@ -5,8 +5,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 
 class AntiAliasActivation(nn.Module):
     def __init__(
diff --git a/arvc/engine/models/generators/fullband_mrf.py b/arvc/engine/models/generators/fullband_mrf.py
@@ -30,8 +30,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.utils.checkpoint import checkpoint
 
 LRELU_SLOPE = 0.1
diff --git a/arvc/engine/models/generators/hifigan.py b/arvc/engine/models/generators/hifigan.py
@@ -3,8 +3,7 @@
 import torch
 
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 
 
 from arvc.engine.models.algorithms.commons import init_weights
diff --git a/arvc/engine/models/generators/hifigan_v3.py b/arvc/engine/models/generators/hifigan_v3.py
@@ -25,8 +25,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.utils.checkpoint import checkpoint
 
 LRELU_SLOPE = 0.1
diff --git a/arvc/engine/models/generators/jvsf_hifigan.py b/arvc/engine/models/generators/jvsf_hifigan.py
@@ -28,8 +28,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.utils.checkpoint import checkpoint
 
 LRELU_SLOPE = 0.1
diff --git a/arvc/engine/models/generators/mrf_hifigan.py b/arvc/engine/models/generators/mrf_hifigan.py
@@ -5,8 +5,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.utils.checkpoint import checkpoint
 
 LRELU_SLOPE = 0.1
diff --git a/arvc/engine/models/generators/nsf_apnet.py b/arvc/engine/models/generators/nsf_apnet.py
@@ -26,8 +26,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.utils.checkpoint import checkpoint
 
 LRELU_SLOPE = 0.1
diff --git a/arvc/engine/models/generators/nsf_hifigan.py b/arvc/engine/models/generators/nsf_hifigan.py
@@ -7,8 +7,7 @@
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
 
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.utils.checkpoint import checkpoint
 
 
diff --git a/arvc/engine/models/generators/refinegan.py b/arvc/engine/models/generators/refinegan.py
@@ -8,8 +8,7 @@
 import torch.nn.functional as F
 from torch.utils.checkpoint import checkpoint
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 
 
 from arvc.engine.models.algorithms.commons import init_weights, get_padding
diff --git a/arvc/engine/models/generators/vocos.py b/arvc/engine/models/generators/vocos.py
@@ -15,8 +15,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 from torch.utils.checkpoint import checkpoint
 
 
diff --git a/arvc/engine/models/generators/waveglow.py b/arvc/engine/models/generators/waveglow.py
@@ -30,8 +30,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm, weight_norm_v
 from torch.utils.checkpoint import checkpoint
 
 LRELU_SLOPE = 0.1
@@ -101,7 +100,7 @@ def __init__(self, channels):
         self.conv = weight_norm(nn.Conv1d(channels, channels, kernel_size=1, bias=False))
 
         # Initialize as a near-identity rotation
-        nn.init.eye_(self.conv.weight_v.data[:channels, :channels, 0])
+        nn.init.eye_(weight_norm_v(self.conv).data[:channels, :channels, 0])
 
     def forward(self, x):
         return self.conv(x)
diff --git a/arvc/engine/models/predictors/FCPE/FCPE.py b/arvc/engine/models/predictors/FCPE/FCPE.py
@@ -9,8 +9,7 @@
 
 from einops import rearrange
 import torch.nn.utils.parametrize as parametrize
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
+from arvc.engine.models.weight_norm import weight_norm, remove_weight_norm
 
 os.environ["LRU_CACHE_CAPACITY"] = "3"
 
diff --git a/arvc/engine/models/utils.py b/arvc/engine/models/utils.py
@@ -3,15 +3,35 @@
 import gc
 import sys
 import torch
-import faiss
-import codecs
 import logging
 
 import numpy as np
 
 from pydub import AudioSegment
 
 
+# ── Safe FAISS import ────────────────────────────────────────────────────
+# faiss-cpu may fail with ModuleNotFoundError("No module named 'faiss.swigfaiss_avx2'")
+# when the AVX2-optimized native module isn't available (e.g. older CPUs, some
+# cloud VMs).  Fall back to the non-AVX2 version gracefully.
+try:
+    import faiss
+except ModuleNotFoundError as _faiss_err:
+    if "swigfaiss_avx2" in str(_faiss_err):
+        import warnings
+        warnings.warn(
+            "faiss.swigfaiss_avx2 not available — falling back to non-AVX2 faiss. "
+            "If you need AVX2 support, install a compatible faiss-cpu wheel.",
+            stacklevel=2,
+        )
+        # Force faiss to use the fallback (non-AVX2) implementation
+        import faiss.loader as _faiss_loader
+        _faiss_loader.toggle_swigfaiss_avx2 = False
+        import faiss
+    else:
+        raise
+
+
 from arvc.utils import huggingface
 from arvc.utils.variables import translations, configs, config, logger, embedders_model, spin_model, whisper_model
 
diff --git a/arvc/engine/models/weight_norm.py b/arvc/engine/models/weight_norm.py
diff --git a/arvc/engine/speaker/ECAPA_TDNN.py b/arvc/engine/speaker/ECAPA_TDNN.py
diff --git a/arvc/engine/training/create_index.py b/arvc/engine/training/create_index.py
diff --git a/arvc/engine/training/runner/extract_model.py b/arvc/engine/training/runner/extract_model.py
diff --git a/arvc/engine/training/runner/train.py b/arvc/engine/training/runner/train.py
diff --git a/arvc/engine/training/runner/utils.py b/arvc/engine/training/runner/utils.py
diff --git a/arvc/utils/variables.py b/arvc/utils/variables.py

Original file line number	Diff line number	Diff line change
`@@ -614,5 +614,6 @@`
`614`	`614`	`"pretrained_v2_url": "https://huggingface.co/buckets/R-Kentaren/Ultimate-RVC-Models/resolve/pretrained_v2/",`
`615`	`615`	`"pretrained_v1_fallback_url": "https://huggingface.co/buckets/R-Kentaren/Ultimate-RVC-Models/resolve/pretrained_v1/",`
`616`	`616`	`"pretrained_v2_fallback_url": "https://huggingface.co/buckets/R-Kentaren/Ultimate-RVC-Models/resolve/pretrained_v2/",`
`617`		`- "search_api_url": "https://voice-models.com/fetch_data.php"`
	`617`	`+ "search_api_url": "https://voice-models.com/fetch_data.php",`
	`618`	`+ "new_pytorch_weight_norm": false`
`618`	`619`	`}`