diff --git a/Dockerfile b/Dockerfile
index 003ff4b1..adf8c7ea 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -84,12 +84,20 @@ RUN mkdir -p /app/data/generations \
 
 WORKDIR /app
 
-# Install only runtime system dependencies (gosu drops root in the entrypoint)
+# Install only runtime system dependencies (gosu drops root in the entrypoint).
+# sox is required by qwen-tts's speech_vq X-vector extractor, which shells
+# out to it via the `sox` Python bindings for reference-audio normalization.
+# gcc/g++ are required by Triton (used by some torch ops, e.g.
+# bmm_outer_product) to JIT-compile its CUDA driver shim on first use --
+# without them, PyTorch's CUDA-capable build crashes with "Failed to find
+# C compiler" even when a GPU is actually present.
 RUN apt-get update && apt-get install -y --no-install-recommends \
     ffmpeg \
+    sox \
+    gcc \
+    g++ \
     curl \
     gosu \
-    sox \
     && rm -rf /var/lib/apt/lists/*
 
 # Copy installed Python packages from builder stage
diff --git a/backend/app.py b/backend/app.py
index c7a4029b..3d892e3c 100644
--- a/backend/app.py
+++ b/backend/app.py
@@ -49,7 +49,7 @@ if not os.environ.get("HSA_OVERRIDE_GFX_VERSION"):
 # Only set HSA_OVERRIDE_GFX_VERSION for older GPUs that need it.
 # RDNA 3+ (gfx1100+) and RDNA 4 (gfx1200+) are natively supported by ROCm
 # and the override can cause suboptimal performance or errors.
-if not os.environ.get("HSA_OVERRIDE_GFX_VERSION"):
+if not os.environ.get("HSA_OVERRIDE_GFX_VERSION") and Path("/dev/kfd").exists():
     try:
         result = subprocess.run(
             ["rocminfo"],
diff --git a/backend/requirements.txt b/backend/requirements.txt
index ddc9bb6c..92621822 100644
--- a/backend/requirements.txt
+++ b/backend/requirements.txt
@@ -58,7 +58,11 @@ librosa>=0.10.0
 soundfile>=0.12.0
 numpy>=1.24.0,<2.0
 numba>=0.60.0,<0.61.0
-pedalboard>=0.9.0
+# pedalboard>=0.9.21 ships a Linux wheel with AVX-512 (zmm) instructions baked
+# into its native extension, which SIGILLs on import on any CPU without
+# AVX-512 (e.g. Zen 2/3, pre-Ice-Lake Intel). Pin below the regression until
+# upstream fixes it: https://github.com/spotify/pedalboard/issues/454
+pedalboard>=0.9.0,<0.9.21
 
 # HTTP client (for CUDA backend download)
 httpx>=0.27.0