From dfa0aef6d0171fc7ced791412d544924623e899c Mon Sep 17 00:00:00 2001
From: Alexander-Ger-Reich
 <50119493+Alexander-Ger-Reich@users.noreply.github.com>
Date: Sun, 12 Apr 2026 19:29:14 +0200
Subject: [PATCH 1/3] AMD rocm Fix 1

---
 pyproject.toml | 26 ++++++++++++++++++++++++--
 1 file changed, 24 insertions(+), 2 deletions(-)

diff --git a/pyproject.toml b/pyproject.toml
index c0f0c32e..2d44c807 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -83,10 +83,16 @@ url = "https://download.pytorch.org/whl/cu128"
 # Only use this index when explicitly requested by `tool.uv.sources`.
 explicit = true
 
+[[tool.uv.index]]
+name = "pytorch-rocm"
+# For AMD (e.g., RX 7900 series / GFX 1100 architecture)
+url = "https://rocm.nightlies.amd.com/v2/gfx110X-all/"
+explicit = true
+
 [tool.uv]
 constraint-dependencies = [
-    "torch==2.8.0",
-    "torchaudio==2.8.0",
+    "torch",
+    "torchaudio",
 ]
 
 [tool.hatch.build.targets.sdist]
@@ -94,3 +100,19 @@ include = ["omnivoice"]
 
 [tool.hatch.build.targets.wheel]
 packages = ["omnivoice"]
+
+# =========================================================
+# IMPORTANT: Select the correct graphics card here!
+# Comment out the block for your current card 
+# and uncomment the other one.
+# =========================================================
+
+# FOR NVIDIA:
+[tool.uv.sources]
+torch = { index = "pytorch-cuda" }
+torchaudio = { index = "pytorch-cuda" }
+
+# FOR AMD:
+# [tool.uv.sources]
+# torch = { index = "pytorch-rocm" }
+# torchaudio = { index = "pytorch-rocm" }

From 0261153c80bef9fcd5be3b8d5c556120fcc015e0 Mon Sep 17 00:00:00 2001
From: Alexander-Ger-Reich
 <50119493+Alexander-Ger-Reich@users.noreply.github.com>
Date: Sun, 12 Apr 2026 20:50:28 +0200
Subject: [PATCH 2/3] AMD rocm Fix 2

---
 omnivoice/utils/audio.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/omnivoice/utils/audio.py b/omnivoice/utils/audio.py
index 9fcb2c2c..661e720b 100644
--- a/omnivoice/utils/audio.py
+++ b/omnivoice/utils/audio.py
@@ -25,6 +25,7 @@
 import numpy as np
 import torch
 import torchaudio
+import soundfile as sf
 from pydub import AudioSegment
 from pydub.silence import detect_leading_silence, detect_nonsilent, split_on_silence
 
@@ -42,9 +43,8 @@ def load_audio(audio_path: str, sampling_rate: int):
         PyTorch tensor of shape (1, T)
     """
     try:
-        waveform, prompt_sampling_rate = torchaudio.load(
-            audio_path, backend="soundfile"
-        )
+        data, prompt_sampling_rate = sf.read(audio_path, dtype="float32")
+        waveform = torch.from_numpy(data.T if data.ndim > 1 else data.reshape(1, -1))
     except (RuntimeError, OSError):
         # Fallback via pydub+ffmpeg for formats torchaudio can't handle
         aseg = AudioSegment.from_file(audio_path)

From 338282894724365484ed24cba24197e8e3829888 Mon Sep 17 00:00:00 2001
From: Alexander-Ger-Reich
 <50119493+Alexander-Ger-Reich@users.noreply.github.com>
Date: Sun, 12 Apr 2026 20:54:05 +0200
Subject: [PATCH 3/3] AMD rocm Fix 3

---
 pyproject.toml | 38 ++++++++++++++++----------------------
 1 file changed, 16 insertions(+), 22 deletions(-)

diff --git a/pyproject.toml b/pyproject.toml
index 2d44c807..0c1255fb 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -68,6 +68,14 @@ Repository = "https://github.com/k2-fsa/OmniVoice"
 # Install PyTorch with CUDA support on Linux/Windows (CUDA doesn't exist for Mac).
 # NOTE: We must explicitly request them as `dependencies` above. These improved
 # versions will not be selected if they're only third-party dependencies.
+
+# =========================================================
+# IMPORTANT: Select the correct graphics card here!
+# Comment out the block for your current card 
+# and uncomment the other one.
+# =========================================================
+
+# FOR NVIDIA:
 torch = [
   { index = "pytorch-cuda", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
 ]
@@ -75,6 +83,14 @@ torchaudio = [
   { index = "pytorch-cuda", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
 ]
 
+# FOR AMD:
+#torch = [
+#  { index = "pytorch-rocm", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+#]
+#torchaudio = [
+#  { index = "pytorch-rocm", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
+#]
+
 [[tool.uv.index]]
 name = "pytorch-cuda"
 # Use PyTorch built for NVIDIA Toolkit version 12.8.
@@ -89,30 +105,8 @@ name = "pytorch-rocm"
 url = "https://rocm.nightlies.amd.com/v2/gfx110X-all/"
 explicit = true
 
-[tool.uv]
-constraint-dependencies = [
-    "torch",
-    "torchaudio",
-]
-
 [tool.hatch.build.targets.sdist]
 include = ["omnivoice"]
 
 [tool.hatch.build.targets.wheel]
 packages = ["omnivoice"]
-
-# =========================================================
-# IMPORTANT: Select the correct graphics card here!
-# Comment out the block for your current card 
-# and uncomment the other one.
-# =========================================================
-
-# FOR NVIDIA:
-[tool.uv.sources]
-torch = { index = "pytorch-cuda" }
-torchaudio = { index = "pytorch-cuda" }
-
-# FOR AMD:
-# [tool.uv.sources]
-# torch = { index = "pytorch-rocm" }
-# torchaudio = { index = "pytorch-rocm" }