diff --git a/install.bat b/install.bat
index d98c06f..a4c917d 100644
--- a/install.bat
+++ b/install.bat
@@ -1,73 +1,75 @@
 @echo off
-setlocal
-title Installer
+setlocal enabledelayedexpansion
+title Applio Installer
+
+echo Welcome to the RVC CLI Installer!
+echo.
 
 set "principal=%cd%"
-set "URL_EXTRA=https://huggingface.co/IAHispano/applio/resolve/main"
 set "CONDA_ROOT_PREFIX=%UserProfile%\Miniconda3"
 set "INSTALL_ENV_DIR=%principal%\env"
 set "MINICONDA_DOWNLOAD_URL=https://repo.anaconda.com/miniconda/Miniconda3-py39_23.9.0-0-Windows-x86_64.exe"
 set "CONDA_EXECUTABLE=%CONDA_ROOT_PREFIX%\Scripts\conda.exe"
 
-if not exist "%cd%\env.zip" (
-    echo Downloading the fairseq build...
-    curl -s -LJO %URL_EXTRA%/env.zip -o env.zip
-)
-
-if not exist "%cd%\env.zip" (
-    echo Download failed, trying with the powershell method
-    powershell -Command "& {Invoke-WebRequest -Uri '%URL_EXTRA%/env.zip' -OutFile 'env.zip'}"
-)
-
-if not exist "%cd%\env" (
-    echo Extracting the file...
-    powershell -command "& { Add-Type -AssemblyName System.IO.Compression.FileSystem ; [System.IO.Compression.ZipFile]::ExtractToDirectory('%cd%\env.zip', '%cd%') }"
-)
-
-if not exist "%cd%\env" (
-    echo Extracting failed trying with the tar method...
-    tar -xf %cd%\env.zip
-)
-
-if exist "%cd%\env" (
-    del env.zip
-) else (
-    echo Theres a problem extracting the file please download the file and extract it manually.
-    echo https://huggingface.co/IAHispano/applio/resolve/main/env.zip
-    pause
-    exit
+echo Cleaning up unnecessary files...
+for %%F in (Makefile Dockerfile docker-compose.yaml *.sh) do (
+    if exist "%%F" del "%%F"
 )
+echo Cleanup complete.
+echo.
 
 if not exist "%CONDA_EXECUTABLE%" (
-    echo Downloading Miniconda from %MINICONDA_DOWNLOAD_URL%...
-    curl %MINICONDA_DOWNLOAD_URL% -o miniconda.exe
-
-    if not exist "%principal%\miniconda.exe" (
-        echo Download failed trying with the powershell method.
-        powershell -Command "& {Invoke-WebRequest -Uri '%MINICONDA_DOWNLOAD_URL%' -OutFile 'miniconda.exe'}"
+    echo Miniconda not found. Starting download and installation...
+    echo Downloading Miniconda...
+    powershell -Command "& {Invoke-WebRequest -Uri '%MINICONDA_DOWNLOAD_URL%' -OutFile 'miniconda.exe'}"
+    if not exist "miniconda.exe" (
+        echo Download failed. Please check your internet connection and try again.
+        goto :error
     )
 
-    echo Installing Miniconda to %CONDA_ROOT_PREFIX%...
+    echo Installing Miniconda...
     start /wait "" miniconda.exe /InstallationType=JustMe /RegisterPython=0 /S /D=%CONDA_ROOT_PREFIX%
+    if errorlevel 1 (
+        echo Miniconda installation failed.
+        goto :error
+    )
     del miniconda.exe
+    echo Miniconda installation complete.
+) else (
+    echo Miniconda already installed. Skipping installation.
 )
+echo.
 
+echo Creating Conda environment...
 call "%CONDA_ROOT_PREFIX%\_conda.exe" create --no-shortcuts -y -k --prefix "%INSTALL_ENV_DIR%" python=3.9
+if errorlevel 1 goto :error
+echo Conda environment created successfully.
+echo.
 
-if exist "%cd%\env\python.exe" (
-    echo Installing pip version less than 24.1...
-    "%cd%\env\python.exe" -m pip install "pip<24.1"
+if exist "%INSTALL_ENV_DIR%\python.exe" (
+    echo Installing specific pip version...
+    "%INSTALL_ENV_DIR%\python.exe" -m pip install "pip<24.1"
+    if errorlevel 1 goto :error
+    echo Pip installation complete.
+    echo.
 )
 
-echo Installing the dependencies...
-call "%CONDA_ROOT_PREFIX%\condabin\conda.bat" activate "%INSTALL_ENV_DIR%"
-pip install --upgrade setuptools
-pip install -r "%principal%\requirements.txt"
+echo Installing dependencies...
+call "%CONDA_ROOT_PREFIX%\condabin\conda.bat" activate "%INSTALL_ENV_DIR%" || goto :error
+pip install --upgrade setuptools || goto :error
+pip install -r "%principal%\requirements.txt" || goto :error
 pip uninstall torch torchvision torchaudio -y
-pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu121
+pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu121 || goto :error
 call "%CONDA_ROOT_PREFIX%\condabin\conda.bat" deactivate
-echo.
+echo Dependencies installation complete.
+echo
 
 echo RVC CLI has been installed successfully!
+echo.
+pause
+exit /b 0
+
+:error
+echo An error occurred during installation. Please check the output above for details.
 pause
-cls
\ No newline at end of file
+exit /b 1
\ No newline at end of file
diff --git a/requirements.txt b/requirements.txt
index 2be70ba..7232ce3 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -21,8 +21,6 @@ pedalboard
 
 # Machine learning
 omegaconf==2.0.5; sys_platform == 'darwin'
-git+https://github.com/IAHispano/fairseq; sys_platform == 'linux'
-fairseq==0.12.2; sys_platform == 'darwin' or sys_platform == 'win32'
 numba; sys_platform == 'linux'
 numba==0.57.0; sys_platform == 'darwin' or sys_platform == 'win32'
 torchaudio==2.1.1
@@ -32,6 +30,7 @@ torchvision==0.16.1
 einops
 libf0
 torchfcpe
+transformers==4.44.2
 
 # Visualization
 matplotlib==3.7.2
diff --git a/rvc/configs/config.py b/rvc/configs/config.py
index 8b5ec68..51eb0e2 100644
--- a/rvc/configs/config.py
+++ b/rvc/configs/config.py
@@ -131,6 +131,13 @@ def device_config(self) -> tuple:
     def set_cuda_config(self):
         i_device = int(self.device.split(":")[-1])
         self.gpu_name = torch.cuda.get_device_name(i_device)
+        # Zluda
+        if self.gpu_name.endswith("[ZLUDA]"):
+            print("Zluda compatibility enabled, experimental feature.")
+            torch.backends.cudnn.enabled = False
+            torch.backends.cuda.enable_flash_sdp(False)
+            torch.backends.cuda.enable_math_sdp(True)
+            torch.backends.cuda.enable_mem_efficient_sdp(False)
         low_end_gpus = ["16", "P40", "P10", "1060", "1070", "1080"]
         if (
             any(gpu in self.gpu_name for gpu in low_end_gpus)
diff --git a/rvc/infer/infer.py b/rvc/infer/infer.py
index 352ab78..56cfd6a 100644
--- a/rvc/infer/infer.py
+++ b/rvc/infer/infer.py
@@ -53,6 +53,7 @@ def __init__(self):
         self.hubert_model = (
             None  # Initialize the Hubert model (for embedding extraction)
         )
+        self.last_embedder_model = None  # Last used embedder model
         self.tgt_sr = None  # Target sampling rate for the output audio
         self.net_g = None  # Generator network for voice conversion
         self.vc = None  # Voice conversion pipeline instance
@@ -69,8 +70,8 @@ def load_hubert(self, embedder_model: str, embedder_model_custom: str = None):
             embedder_model (str): Path to the pre-trained HuBERT model.
             embedder_model_custom (str): Path to the custom HuBERT model.
         """
-        models, _, _ = load_embedding(embedder_model, embedder_model_custom)
-        self.hubert_model = models[0].to(self.config.device)
+        self.hubert_model = load_embedding(embedder_model, embedder_model_custom)
+        self.hubert_model.to(self.config.device)
         self.hubert_model = (
             self.hubert_model.half()
             if self.config.is_half
@@ -333,8 +334,9 @@ def convert_audio(
             if audio_max > 1:
                 audio /= audio_max
 
-            if not self.hubert_model:
+            if not self.hubert_model or embedder_model != self.last_embedder_model:
                 self.load_hubert(embedder_model, embedder_model_custom)
+                self.last_embedder_model = embedder_model
 
             file_index = (
                 index_path.strip()
@@ -637,8 +639,9 @@ def convert_audio_batch(
         with open(pid_file_path, "w") as pid_file:
             pid_file.write(str(pid))
         try:
-            if not self.hubert_model:
+            if not self.hubert_model or embedder_model != self.last_embedder_model:
                 self.load_hubert(embedder_model, embedder_model_custom)
+                self.last_embedder_model = embedder_model
             self.get_vc(model_path, sid)
             file_index = (
                 index_path.strip()
diff --git a/rvc/infer/pipeline.py b/rvc/infer/pipeline.py
index cb4f6bd..e3eae01 100644
--- a/rvc/infer/pipeline.py
+++ b/rvc/infer/pipeline.py
@@ -425,14 +425,11 @@ def voice_conversion(
         feats = feats.view(1, -1)
         padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
 
-        inputs = {
-            "source": feats.to(self.device),
-            "padding_mask": padding_mask,
-            "output_layer": 9 if version == "v1" else 12,
-        }
         with torch.no_grad():
-            logits = model.extract_features(**inputs)
-            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+            feats = model(feats.to(self.device))["last_hidden_state"]
+            feats = (
+                model.final_proj(feats[0]).unsqueeze(0) if version == "v1" else feats
+            )
         if protect < 0.5 and pitch != None and pitchf != None:
             feats0 = feats.clone()
         if (
diff --git a/rvc/lib/algorithm/commons.py b/rvc/lib/algorithm/commons.py
index 2524abc..c76328c 100644
--- a/rvc/lib/algorithm/commons.py
+++ b/rvc/lib/algorithm/commons.py
@@ -157,6 +157,24 @@ def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
     return acts
 
 
+# Zluda, same as previous, but without jit.script
+def fused_add_tanh_sigmoid_multiply_no_jit(input_a, input_b, n_channels):
+    """
+    Fused add tanh sigmoid multiply operation.
+
+    Args:
+        input_a: The first input tensor.
+        input_b: The second input tensor.
+        n_channels: The number of channels.
+    """
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
 def convert_pad_shape(pad_shape: List[List[int]]) -> List[int]:
     """
     Convert the pad shape to a list of integers.
diff --git a/rvc/lib/algorithm/modules.py b/rvc/lib/algorithm/modules.py
index 1038356..7a9bc54 100644
--- a/rvc/lib/algorithm/modules.py
+++ b/rvc/lib/algorithm/modules.py
@@ -1,5 +1,8 @@
 import torch
-from rvc.lib.algorithm.commons import fused_add_tanh_sigmoid_multiply
+from rvc.lib.algorithm.commons import (
+    fused_add_tanh_sigmoid_multiply_no_jit,
+    fused_add_tanh_sigmoid_multiply,
+)
 
 
 class WaveNet(torch.nn.Module):
@@ -85,6 +88,11 @@ def forward(self, x, x_mask, g=None, **kwargs):
         if g is not None:
             g = self.cond_layer(g)
 
+        # Zluda
+        is_zluda = x.device.type == "cuda" and torch.cuda.get_device_name().endswith(
+            "[ZLUDA]"
+        )
+
         for i in range(self.n_layers):
             x_in = self.in_layers[i](x)
             if g is not None:
@@ -93,7 +101,14 @@ def forward(self, x, x_mask, g=None, **kwargs):
             else:
                 g_l = torch.zeros_like(x_in)
 
-            acts = fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
+            # Preventing HIP crash by not using jit-decorated function
+            if is_zluda:
+                acts = fused_add_tanh_sigmoid_multiply_no_jit(
+                    x_in, g_l, n_channels_tensor
+                )
+            else:
+                acts = fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
+
             acts = self.drop(acts)
 
             res_skip_acts = self.res_skip_layers[i](acts)
diff --git a/rvc/lib/predictors/RMVPE.py b/rvc/lib/predictors/RMVPE.py
index f4d9175..7e9f6dd 100644
--- a/rvc/lib/predictors/RMVPE.py
+++ b/rvc/lib/predictors/RMVPE.py
@@ -408,6 +408,15 @@ def forward(self, audio, keyshift=0, speed=1, center=True):
             self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(
                 audio.device
             )
+
+        # Zluda, fall-back to CPU for FFTs since HIP SDK has no cuFFT alternative
+        source_device = audio.device
+        if audio.device.type == "cuda" and torch.cuda.get_device_name().endswith(
+            "[ZLUDA]"
+        ):
+            audio = audio.to("cpu")
+            self.hann_window[keyshift_key] = self.hann_window[keyshift_key].to("cpu")
+
         fft = torch.stft(
             audio,
             n_fft=n_fft_new,
@@ -416,7 +425,8 @@ def forward(self, audio, keyshift=0, speed=1, center=True):
             window=self.hann_window[keyshift_key],
             center=center,
             return_complex=True,
-        )
+        ).to(source_device)
+
         magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
         if keyshift != 0:
             size = self.n_fft // 2 + 1
diff --git a/rvc/lib/tools/prerequisites_download.py b/rvc/lib/tools/prerequisites_download.py
index dbc9f06..e383562 100644
--- a/rvc/lib/tools/prerequisites_download.py
+++ b/rvc/lib/tools/prerequisites_download.py
@@ -44,7 +44,7 @@
     )
 ]
 models_list = [("predictors/", ["rmvpe.pt", "fcpe.pt"])]
-embedders_list = [("embedders/", ["contentvec_base.pt"])]
+embedders_list = [("embedders/contentvec/", ["pytorch_model.bin", "config.json"])]
 linux_executables_list = [("formant/", ["stftpitchshift"])]
 executables_list = [
     ("", ["ffmpeg.exe", "ffprobe.exe"]),
@@ -54,7 +54,7 @@
 folder_mapping_list = {
     "pretrained_v1/": "rvc/models/pretraineds/pretrained_v1/",
     "pretrained_v2/": "rvc/models/pretraineds/pretrained_v2/",
-    "embedders/": "rvc/models/embedders/",
+    "embedders/contentvec/": "rvc/models/embedders/contentvec/",
     "predictors/": "rvc/models/predictors/",
     "formant/": "rvc/models/formant/",
 }
diff --git a/rvc/lib/utils.py b/rvc/lib/utils.py
index a53810d..429a172 100644
--- a/rvc/lib/utils.py
+++ b/rvc/lib/utils.py
@@ -4,16 +4,23 @@
 import numpy as np
 import re
 import unicodedata
-from fairseq import checkpoint_utils
 import wget
 import subprocess
 from pydub import AudioSegment
 import tempfile
+from torch import nn
 
 import logging
+from transformers import HubertModel
+import warnings
 
-logging.getLogger("fairseq").setLevel(logging.WARNING)
-logging.getLogger("faiss.loader").setLevel(logging.WARNING)
+# Remove this to see warnings about transformers models
+warnings.filterwarnings("ignore")
+
+logging.getLogger("fairseq").setLevel(logging.ERROR)
+logging.getLogger("faiss.loader").setLevel(logging.ERROR)
+logging.getLogger("transformers").setLevel(logging.ERROR)
+logging.getLogger("torch").setLevel(logging.ERROR)
 
 now_dir = os.getcwd()
 sys.path.append(now_dir)
@@ -22,6 +29,12 @@
 stft = base_path + ".exe" if sys.platform == "win32" else base_path
 
 
+class HubertModelWithFinalProj(HubertModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.final_proj = nn.Linear(config.hidden_size, config.classifier_proj_size)
+
+
 def load_audio(file, sample_rate):
     try:
         file = file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
@@ -96,36 +109,45 @@ def format_title(title):
 def load_embedding(embedder_model, custom_embedder=None):
     embedder_root = os.path.join(now_dir, "rvc", "models", "embedders")
     embedding_list = {
-        "contentvec": os.path.join(embedder_root, "contentvec_base.pt"),
-        "chinese-hubert-base": os.path.join(embedder_root, "chinese_hubert_base.pt"),
-        "japanese-hubert-base": os.path.join(embedder_root, "japanese_hubert_base.pt"),
-        "korean-hubert-base": os.path.join(embedder_root, "korean_hubert_base.pt"),
+        "contentvec": os.path.join(embedder_root, "contentvec"),
+        "chinese-hubert-base": os.path.join(embedder_root, "chinese_hubert_base"),
+        "japanese-hubert-base": os.path.join(embedder_root, "japanese_hubert_base"),
+        "korean-hubert-base": os.path.join(embedder_root, "korean_hubert_base"),
     }
 
     online_embedders = {
-        "chinese-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/chinese_hubert_base.pt",
-        "japanese-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/japanese_hubert_base.pt",
-        "korean-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/korean_hubert_base.pt",
+        "contentvec": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/contentvec/pytorch_model.bin",
+        "chinese-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/chinese_hubert_base/pytorch_model.bin",
+        "japanese-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/japanese_hubert_base/pytorch_model.bin",
+        "korean-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/korean_hubert_base/pytorch_model.bin",
+    }
+
+    config_files = {
+        "contentvec": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/contentvec/config.json",
+        "chinese-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/chinese_hubert_base/config.json",
+        "japanese-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/japanese_hubert_base/config.json",
+        "korean-hubert-base": "https://huggingface.co/IAHispano/Applio/resolve/main/Resources/embedders/korean_hubert_base/config.json",
     }
 
     if embedder_model == "custom":
-        model_path = custom_embedder
-        if not custom_embedder and os.path.exists(custom_embedder):
+        if os.path.exists(custom_embedder):
+            model_path = custom_embedder
+        else:
+            print(f"Custom embedder not found: {custom_embedder}, using contentvec")
             model_path = embedding_list["contentvec"]
     else:
         model_path = embedding_list[embedder_model]
-        if embedder_model in online_embedders:
-            if not os.path.exists(model_path):
-                url = online_embedders[embedder_model]
-                print(f"Downloading {url} to {model_path}...")
-                wget.download(url, out=model_path)
-        else:
-            model_path = embedding_list["contentvec"]
-
-    models = checkpoint_utils.load_model_ensemble_and_task(
-        [model_path],
-        suffix="",
-    )
-
-    # print(f"Embedding model {embedder_model} loaded successfully.")
+        bin_file = os.path.join(model_path, "pytorch_model.bin")
+        json_file = os.path.join(model_path, "config.json")
+        os.makedirs(model_path, exist_ok=True)
+        if not os.path.exists(bin_file):
+            url = online_embedders[embedder_model]
+            print(f"Downloading {url} to {model_path}...")
+            wget.download(url, out=bin_file)
+        if not os.path.exists(json_file):
+            url = config_files[embedder_model]
+            print(f"Downloading {url} to {model_path}...")
+            wget.download(url, out=json_file)
+
+    models = HubertModelWithFinalProj.from_pretrained(model_path)
     return models
diff --git a/rvc/train/extract/extract.py b/rvc/train/extract/extract.py
index 8f2e5ff..b9fe041 100644
--- a/rvc/train/extract/extract.py
+++ b/rvc/train/extract/extract.py
@@ -229,7 +229,7 @@ def run_pitch_extraction(exp_dir, f0_method, hop_length, num_processes, gpus):
     print(f"Pitch extraction completed in {elapsed_time:.2f} seconds.")
 
 
-def process_file_embedding(file, wav_path, out_path, model, device, version, saved_cfg):
+def process_file_embedding(file, wav_path, out_path, model, device, version):
     """Process a single audio file for embedding extraction."""
     wav_file_path = os.path.join(wav_path, file)
     out_file_path = os.path.join(out_path, file.replace("wav", "npy"))
@@ -237,21 +237,14 @@ def process_file_embedding(file, wav_path, out_path, model, device, version, sav
     if os.path.exists(out_file_path):
         return
 
-    feats = read_wave(wav_file_path, normalize=saved_cfg.task.normalize)
-    dtype = torch.float16 if device.startswith("cuda") else torch.float32
+    feats = read_wave(wav_file_path)
+    dtype = torch.float16 if config.is_half else torch.float32
     feats = feats.to(dtype).to(device)
-
-    padding_mask = torch.BoolTensor(feats.shape).fill_(False).to(dtype).to(device)
-    inputs = {
-        "source": feats,
-        "padding_mask": padding_mask,
-        "output_layer": 9 if version == "v1" else 12,
-    }
+    model = model.to(dtype).to(device)
 
     with torch.no_grad():
-        model = model.to(device).to(dtype)
-        logits = model.extract_features(**inputs)
-        feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+        feats = model(feats)["last_hidden_state"]
+        feats = model.final_proj(feats[0]).unsqueeze(0) if version == "v1" else feats
 
     feats = feats.squeeze(0).float().cpu().numpy()
     if not np.isnan(feats).any():
@@ -269,8 +262,16 @@ def run_embedding_extraction(
     print("Starting embedding extraction...")
     start_time = time.time()
 
-    models, saved_cfg, _ = load_embedding(embedder_model, embedder_model_custom)
-    model = models[0]
+    models = load_embedding(embedder_model, embedder_model_custom)
+
+    # Zluda
+    if torch.cuda.is_available() and torch.cuda.get_device_name().endswith("[ZLUDA]"):
+        print("Disabling CUDNN for Zluda")
+        torch.backends.cudnn.enabled = False
+        torch.backends.cuda.enable_flash_sdp(False)
+        torch.backends.cuda.enable_math_sdp(True)
+        torch.backends.cuda.enable_mem_efficient_sdp(False)
+
     devices = [get_device(gpu) for gpu in (gpus.split("-") if gpus != "-" else ["cpu"])]
 
     paths = sorted([file for file in os.listdir(wav_path) if file.endswith(".wav")])
@@ -281,7 +282,7 @@ def run_embedding_extraction(
     pbar = tqdm.tqdm(total=len(paths) * len(devices), desc="Embedding Extraction")
 
     tasks = [
-        (file, wav_path, out_path, model, device, version, saved_cfg)
+        (file, wav_path, out_path, models, device, version)
         for file in paths
         for device in devices
     ]
diff --git a/rvc/train/mel_processing.py b/rvc/train/mel_processing.py
index ecad648..1f5e8f3 100644
--- a/rvc/train/mel_processing.py
+++ b/rvc/train/mel_processing.py
@@ -76,6 +76,12 @@ def spectrogram_torch(y, n_fft, hop_size, win_size, center=False):
     )
     y = y.squeeze(1)
 
+    # Zluda, fall-back to CPU for FFTs since HIP SDK has no cuFFT alternative
+    source_device = y.device
+    if y.device.type == "cuda" and torch.cuda.get_device_name().endswith("[ZLUDA]"):
+        y = y.to("cpu")
+        hann_window[wnsize_dtype_device] = hann_window[wnsize_dtype_device].to("cpu")
+
     spec = torch.stft(
         y,
         n_fft,
@@ -87,7 +93,7 @@ def spectrogram_torch(y, n_fft, hop_size, win_size, center=False):
         normalized=False,
         onesided=True,
         return_complex=True,
-    )
+    ).to(source_device)
 
     spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + 1e-6)
 
diff --git a/rvc/train/train.py b/rvc/train/train.py
index a4e5ce9..b4c1120 100644
--- a/rvc/train/train.py
+++ b/rvc/train/train.py
@@ -319,6 +319,14 @@ def run(
     if torch.cuda.is_available():
         torch.cuda.set_device(rank)
 
+    # Zluda
+    if torch.cuda.is_available() and torch.cuda.get_device_name().endswith("[ZLUDA]"):
+        print("Disabling CUDNN for traning with Zluda")
+        torch.backends.cudnn.enabled = False
+        torch.backends.cuda.enable_flash_sdp(False)
+        torch.backends.cuda.enable_math_sdp(True)
+        torch.backends.cuda.enable_mem_efficient_sdp(False)
+
     # Create datasets and dataloaders
     if pitch_guidance == True:
         train_dataset = TextAudioLoaderMultiNSFsid(config.data)
diff --git a/rvc_cli.py b/rvc_cli.py
index 90479f4..bca001c 100644
--- a/rvc_cli.py
+++ b/rvc_cli.py
@@ -86,6 +86,8 @@ def run_infer_script(
     delay: bool = False,
     *sliders: list,
 ):
+    if not sliders:
+        sliders = [0] * 25
     infer_pipeline = import_voice_converter()
     additional_params = {
         "reverb_room_size": sliders[0],
@@ -198,6 +200,8 @@ def run_batch_infer_script(
         f for f in os.listdir(input_folder) if f.endswith((".mp3", ".wav", ".flac"))
     ]
     print(f"Detected {len(audio_files)} audio files for inference.")
+    if not sliders:
+        sliders = [0] * 25
     infer_pipeline = import_voice_converter()
     additional_params = {
         "reverb_room_size": sliders[0],