diff --git a/ddistro_install.sh b/ddistro_install.sh
index 90e78b7..efa508f 100755
--- a/ddistro_install.sh
+++ b/ddistro_install.sh
@@ -6,31 +6,109 @@ source /home/dwemer/python-tts/bin/activate
 
 echo "This will take a while so please wait."
 
-# Ask user about GPU
-read -p "Are you using a GT10XX series GPU? (yes/no): " gpu_answer
-if [[ "$gpu_answer" =~ ^[Yy][Ee][Ss]$ || "$gpu_answer" =~ ^[Yy]$ ]]; then
-    cu_tag="cu118"
-    torch_url="https://download.pytorch.org/whl/${cu_tag}"
-    torch_ver="2.2.2"
-    torchaudio_ver="2.2.2"
+# Detect GPU generation and choose appropriate torch/torchaudio + CUDA tag
+# Default to modern RTX (Ada and newer) stable cu124
+cu_tag="cu124"
+torch_ver="2.5.1"
+torchaudio_ver="2.5.1"
+torch_url="https://download.pytorch.org/whl/${cu_tag}"
+use_nightly=0
+
+# Try to detect compute capability via nvidia-smi; fall back to user prompt for 10-series
+compute_cap=""
+if command -v nvidia-smi >/dev/null 2>&1; then
+    compute_cap=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | head -n1 2>/dev/null | tr -d '\r')
+fi
+
+# If user has Pascal (GTX 10xx) fallback to CUDA 11.8 + torch 2.2.2
+if [[ -z "$compute_cap" ]]; then
+    read -p "Are you using a GTX 10xx (Pascal) GPU? (yes/no): " gpu_answer
+    if [[ "$gpu_answer" =~ ^[Yy][Ee][Ss]$ || "$gpu_answer" =~ ^[Yy]$ ]]; then
+        cu_tag="cu118"
+        torch_url="https://download.pytorch.org/whl/${cu_tag}"
+        torch_ver="2.2.2"
+        torchaudio_ver="2.2.2"
+    fi
 else
-    cu_tag="cu124"
-    torch_url="https://download.pytorch.org/whl/${cu_tag}"
-    torch_ver="2.4.0"
-    torchaudio_ver="2.4.0"
+    # Parse major.minor like 8.9 (Ada) or 9.0 (Hopper/Blackwell)
+    cap_major=${compute_cap%%.*}
+    cap_minor=${compute_cap#*.}
+    # For 50xx Blackwell (compute capability >= 9), use nightly cu128 wheels
+    if [[ "$cap_major" -ge 9 ]]; then
+        cu_tag="cu128"
+        torch_url="https://download.pytorch.org/whl/nightly/${cu_tag}"
+        use_nightly=1
+    # For older (<= 7.x Pascal/Turing), force cu118 + torch 2.2.2
+    elif [[ "$cap_major" -le 7 ]]; then
+        cu_tag="cu118"
+        torch_url="https://download.pytorch.org/whl/${cu_tag}"
+        torch_ver="2.2.2"
+        torchaudio_ver="2.2.2"
+    fi
 fi
 
 python3 -m pip install --upgrade pip wheel ninja virtualenv
 pip install setuptools==68.1.2
-# Install app requirements without auto-pulling torch/torchaudio from deps
-pip install --no-deps -r requirements.txt
+# Install app requirements excluding torch/torchaudio/torchvision/triton and nvidia-cu12 split libs
+# so our chosen CUDA wheels stay consistent
+grep -viE '^(torch|torchaudio|torchvision|triton|nvidia-[a-z0-9-]+-cu[0-9]+)\b' requirements.txt > /tmp/requirements.no-torch.txt
+pip install --no-deps -r /tmp/requirements.no-torch.txt
 # Pin to stable, CUDA-tagged PyTorch/Torchaudio that do not require TorchCodec
 pip cache purge || true
-pip uninstall -y torch torchaudio torchcodec torchvision || true
-pip install --no-deps --no-cache-dir --index-url "$torch_url" "torch==${torch_ver}+${cu_tag}" "torchaudio==${torchaudio_ver}+${cu_tag}"
+# Drop local caches that can interfere with extension/wheel resolution
+rm -rf "$HOME/.cache/torch_extensions" "$HOME/.cache/torch" "$HOME/.cache/pip" 2>/dev/null || true
+pip uninstall -y torch torchaudio torchcodec torchvision triton || true
+if [[ "$use_nightly" -eq 1 ]]; then
+    # Blackwell/50xx path: use nightly cu128 wheels, let dependencies resolve
+    pip install --pre -U --no-cache-dir --index-url "$torch_url" torch torchaudio
+else
+    pip install --no-deps --no-cache-dir --index-url "$torch_url" "torch==${torch_ver}+${cu_tag}" "torchaudio==${torchaudio_ver}+${cu_tag}"
+    # Install a Triton version that matches torch (skip for nightly)
+    if [[ "${torch_ver}" == 2.2.* ]]; then
+        pip install --no-cache-dir "triton==2.2.0"
+    elif [[ "${torch_ver}" == 2.4.* ]]; then
+        pip install --no-cache-dir "triton==3.0.0"
+    else
+        # torch 2.5.x
+        pip install --no-cache-dir "triton==3.1.0"
+    fi
+fi
 pip check || true
 # Ensure fallback audio loader is available
 pip install --no-cache-dir soundfile
+# Verify critical versions and backend
+python - <<'PY'
+import sys
+try:
+    import torch, torchaudio
+    print("torch:", torch.__version__)
+    print("torchaudio:", torchaudio.__version__)
+    try:
+        import triton
+        print("triton:", triton.__version__)
+    except Exception as e:
+        print("triton import error:", e)
+    try:
+        import torchcodec  # noqa: F401
+        print("torchcodec: present (should be absent)")
+    except Exception:
+        print("torchcodec: not installed (OK)")
+    try:
+        print("audio_backend:", torchaudio.get_audio_backend())
+    except Exception:
+        print("audio_backend: unknown (OK)")
+    try:
+        if torch.cuda.is_available():
+            d = torch.cuda.get_device_properties(0)
+            print("cuda: available, device=", d.name, "cc=", f"{d.major}.{d.minor}")
+        else:
+            print("cuda: not available")
+    except Exception as e:
+        print("cuda check error:", e)
+except Exception as e:
+    print("xtts-preflight-error:", e)
+    sys.exit(0)
+PY
 #pip install xtts-api-server #Fails
 
 sed -i 's/checkpoint = load_fsspec(model_path, map_location=torch.device("cpu"))\["model"\]/checkpoint = load_fsspec(model_path, map_location=torch.device("cpu"), weights_only=False)["model"]/' /home/dwemer/python-tts/lib/python3.11/site-packages/TTS/tts/models/xtts.py
diff --git a/start-deepspeed.sh b/start-deepspeed.sh
index 41a3fa3..d0c5d91 100755
--- a/start-deepspeed.sh
+++ b/start-deepspeed.sh
@@ -1,7 +1,11 @@
 #!/bin/bash
-# Add CUDA to PATH if the directory exists
-if [ -d "/usr/local/cuda-12.8/bin" ]; then
-  export PATH="/usr/local/cuda-12.8/bin:$PATH"
+# Add CUDA to PATH (detect highest installed cuda-*/bin)
+cuda_bin=""
+if [ -d "/usr/local" ]; then
+  cuda_bin=$(ls -d /usr/local/cuda-*/bin 2>/dev/null | sort -V | tail -n1)
+fi
+if [ -n "$cuda_bin" ] && [ -d "$cuda_bin" ]; then
+  export PATH="$cuda_bin:$PATH"
 fi
 
 cd /home/dwemer/xtts-api-server/
diff --git a/start-lowvram.sh b/start-lowvram.sh
index 8308c2f..e9c34f0 100755
--- a/start-lowvram.sh
+++ b/start-lowvram.sh
@@ -1,8 +1,12 @@
 #!/bin/bash
 
-# Add CUDA to PATH if the directory exists (matches deepspeed launcher)
-if [ -d "/usr/local/cuda-12.8/bin" ]; then
-  export PATH="/usr/local/cuda-12.8/bin:$PATH"
+# Add CUDA to PATH (detect highest installed cuda-*/bin)
+cuda_bin=""
+if [ -d "/usr/local" ]; then
+  cuda_bin=$(ls -d /usr/local/cuda-*/bin 2>/dev/null | sort -V | tail -n1)
+fi
+if [ -n "$cuda_bin" ] && [ -d "$cuda_bin" ]; then
+  export PATH="$cuda_bin:$PATH"
 fi
 
 cd /home/dwemer/xtts-api-server/
diff --git a/start-regular.sh b/start-regular.sh
index a661b95..857dc6b 100755
--- a/start-regular.sh
+++ b/start-regular.sh
@@ -1,8 +1,12 @@
 #!/bin/bash
 
-# Add CUDA to PATH if the directory exists (matches deepspeed launcher)
-if [ -d "/usr/local/cuda-12.8/bin" ]; then
-  export PATH="/usr/local/cuda-12.8/bin:$PATH"
+# Add CUDA to PATH (detect highest installed cuda-*/bin)
+cuda_bin=""
+if [ -d "/usr/local" ]; then
+  cuda_bin=$(ls -d /usr/local/cuda-*/bin 2>/dev/null | sort -V | tail -n1)
+fi
+if [ -n "$cuda_bin" ] && [ -d "$cuda_bin" ]; then
+  export PATH="$cuda_bin:$PATH"
 fi
 
 cd /home/dwemer/xtts-api-server/