Merge remote-tracking branch 'origin/main' into fix/dependabot-security

CHANGELOG.md

@@ -44,6 +44,7 @@ The bundled TTS model package (`pyproject.toml`) is versioned independently.
 - Patched dependency releases now cover 35 Python and Rust security advisories without weakening VoiceStudio's GPU or offline-runtime compatibility. (#1456, #1472, #1473, #1474, #1475, #1476, #1477)
 - YouTube imports that require a signed-in session can now use an explicitly selected `cookies.txt` export for one import; VoiceStudio never reads browser cookies silently and makes two best-effort attempts to delete its temporary copy. (#1429, #1432) — thanks @dongqing1968-sudo and @phamvandu9595-tech!
 - First-run source builds no longer stop after uv was successfully downloaded just because its installer failed during a later shell-profile step; app-private uv installs no longer touch shell profiles at all. (#1438) — thanks @AdrianoCahete!
+- ROCm Docker now installs and starts the backend with the same Python whose AMD torch build was validated, instead of launching a second CUDA-only environment and silently running on CPU. (#1274) — thanks @simmessa and @spicchio72!
 - An error whose text merely contained the digits 401 — a file path, a byte count, a job id — no longer tells you to fix your Hugging Face token. (#1427)
 - Custom MLX model IDs and saved voice instructions are now validated in bounded time, so malformed input cannot stall the backend. (#1446)
 - Streaming and provider failures now return stable recovery guidance without exposing exception details. (#1462)

deploy/Dockerfile

@@ -32,7 +32,6 @@ WORKDIR /app
 # Enable unbuffered logs and optimizations
 ENV PYTHONDONTWRITEBYTECODE=1
 ENV PYTHONUNBUFFERED=1
-ENV UV_SYSTEM_PYTHON=1
 ENV HF_HOME=/app/omnivoice_data/huggingface
 # Allow bare imports (from core.config, from services.*, etc.) when
 # uvicorn is started as `backend.main:app` from WORKDIR /app.
@@ -54,9 +53,9 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     && rm -rf /var/lib/apt/lists/*
 
 # PEP 668: the ROCm base (Ubuntu 24.04) marks its system Python
-# EXTERNALLY-MANAGED, which would refuse `pip install` / `uv pip install
-# --system`. Inside a single-purpose container image installing into the
-# base env is exactly what we want. No-ops on the conda-based CUDA image.
+# EXTERNALLY-MANAGED, which would refuse installing into the selected base
+# interpreter. Inside a single-purpose container image that is exactly what
+# we want. No-ops on the conda-based CUDA image.
 ENV PIP_BREAK_SYSTEM_PACKAGES=1
 ENV UV_BREAK_SYSTEM_PACKAGES=1
 
@@ -71,6 +70,11 @@ COPY deploy/torch-constraints.txt ./deploy/torch-constraints.txt
 # Install the project (non-editable — no need for -e in containers).
 # Uses `uv` for exponentially faster resolution than plain pip.
 #
+# Target the exact interpreter selected by the base image. The ROCm image has
+# both /opt/venv/bin/python3 (ROCm torch) and /usr/bin/python (a CUDA-default
+# environment); `--system` used the latter while the build guard used the
+# former, so a green image launched a CPU-only backend on AMD (#1274).
+#
 # NOTE: `uv pip install` (without --upgrade) keeps already-installed packages
 # that satisfy the requirements, so the base image's GPU-built torch/torchaudio
 # (2.8.0, satisfying our `torch>=2.4`) survive this step instead of being
@@ -83,7 +87,8 @@ COPY deploy/torch-constraints.txt ./deploy/torch-constraints.txt
 # stays put — an ABI mismatch at import (#1357). The pins carry no local
 # segment, so they match the base image's +cu128 / +rocm6.4 builds rather than
 # replacing them.
-RUN uv pip install --system --no-cache --constraint deploy/torch-constraints.txt .
+RUN uv pip install --python "$(command -v python3)" --no-cache \
+    --constraint deploy/torch-constraints.txt .
 
 # Guard (fails the build, not the user at runtime): assert the dependency
 # install did NOT replace the base image's GPU torch. A future dep bump that
@@ -121,5 +126,6 @@ HEALTHCHECK --interval=30s --timeout=5s --start-period=120s --retries=5 \
 # Mount points for persistent data (sqlite db, user voices, huggingface cache)
 VOLUME ["/app/omnivoice_data"]
 
-# Bind to 0.0.0.0 for external access
-ENTRYPOINT ["uvicorn", "backend.main:app", "--host", "0.0.0.0", "--port", "3900"]
+# Bind to 0.0.0.0 for external access. `python3 -m` keeps runtime imports on
+# the same interpreter whose torch flavor the build guard validated (#1274).
+ENTRYPOINT ["python3", "-m", "uvicorn", "backend.main:app", "--host", "0.0.0.0", "--port", "3900"]

docs/install/docker.md

@@ -109,16 +109,26 @@ the CUDA tags exactly.
 Verify the container sees the GPU:
 
 ```bash
-docker exec omnivoice python3 -c \
-  "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"
+docker exec <container> python3 -c \
+  "import torch; ok = torch.cuda.is_available(); print(ok, torch.cuda.get_device_name(0) if ok else 'unavailable')"
 ```
 
+Use `omnivoice` for the `docker run` examples above. Docker Compose names the
+ROCm container `omnivoice-studio-rocm` (CPU: `omnivoice-studio`, NVIDIA:
+`omnivoice-studio-gpu`); `docker compose ps` shows the exact active name.
+
 (ROCm-built PyTorch reports through `torch.cuda.*` — `True` plus your card's
 name means torch can see the GPU.) That check alone isn't proof the app is
 using it: **Settings → System** shows the device VoiceStudio actually resolved.
 If it reads `cpu` while the command above prints `True`, the backend log line
 starting `Falling back to CPU:` names the architecture mismatch it hit.
 
+The image installs and launches VoiceStudio through that same `python3`
+interpreter. To verify this invariant on an older or custom image, compare
+`docker exec <container> python3 -c "import sys, torch; print(sys.executable,
+torch.version.hip)"` with `docker exec <container> sh -c 'tr "\\0" " "
+</proc/1/cmdline'`; PID 1 must begin with `python3 -m uvicorn`.
+
 If the command prints `False`, **Settings → System** now says why, and the
 three answers need different fixes:
 
@@ -203,7 +213,7 @@ Two paths are worth persisting across container restarts:
   pushes. Pull the image again after the fix is merged: `docker pull ghcr.io/debpalash/omnivoice-studio:latest`.
   The running version is now shown in **Settings → About → Version** (read live
   from the backend), so the web UI no longer displays a dash in Docker.
-- **Checking which version is running:** `docker exec omnivoice python -c "import importlib.metadata; print(importlib.metadata.version('omnivoice'))"`, or hit the `/health` endpoint — it returns `{"status": "ok", "device": ..., "version": "0.3.x"}`.
+- **Checking which version is running:** `docker exec <container> python3 -c "import importlib.metadata; print(importlib.metadata.version('omnivoice'))"`, or hit the `/health` endpoint — it returns `{"status": "ok", "device": ..., "version": "0.3.x"}`. Use the container name listed by `docker compose ps` (or `omnivoice` for the `docker run` examples).
 - **"Loopback origin required" errors (and a blank version):** the desktop
   build restricts the `/system/*` and `/api/settings/*` routes to a loopback
   origin, but Docker's NAT makes every request look non-loopback, so the gate

tests/test_torch_constraints_are_applied.py

@@ -6,7 +6,7 @@ The coupling lives in `[tool.uv] constraint-dependencies`.
 
 That setting is part of the **project** API — `uv sync`, `uv lock`, `uv run`.
 `uv pip install` is the pip-compatible interface and ignores it. Both install
-paths that use `uv pip install --system` (the Colab notebook and the Docker
+paths that use `uv pip install` (the Colab notebook and the Docker
 image) therefore resolved the trio on its bare lower bounds, free to upgrade
 torch while leaving a torchvision built against an older ABI in place:
 
@@ -99,13 +99,33 @@ def test_the_pins_carry_no_local_version(file_constraints):
 
 def test_the_dockerfile_passes_the_constraint():
     text = _DOCKERFILE.read_text(encoding="utf-8")
-    install = [ln for ln in text.splitlines() if "uv pip install" in ln and "--system" in ln]
-    assert install, "no `uv pip install --system` line found in the Dockerfile"
-    for line in install:
-        assert "--constraint" in line and "torch-constraints.txt" in line, (
-            f"Docker installs without the torch constraint, so the trio can "
-            f"drift again:\n  {line.strip()}"
-        )
+    install_start = text.index("RUN uv pip install")
+    install_end = text.index("\n\n", install_start)
+    install = text[install_start:install_end]
+    assert "--constraint" in install and "torch-constraints.txt" in install, (
+        "Docker installs without the torch constraint, so the trio can "
+        f"drift again:\n{install}"
+    )
+
+
+def test_docker_install_and_runtime_use_the_guarded_python():
+    """#1274: ROCm's `python3` had HIP torch, while `--system` installed and
+    bare `uvicorn` launched through `/usr/bin/python` with CUDA torch."""
+    text = _DOCKERFILE.read_text(encoding="utf-8")
+    install_start = text.index("RUN uv pip install")
+    install_end = text.index("\n\n", install_start)
+    install = text[install_start:install_end]
+    assert '--python "$(command -v python3)"' in install
+    assert "--system" not in install
+    assert 'ENTRYPOINT ["python3", "-m", "uvicorn"' in text
+
+
+def test_docker_docs_do_not_assume_the_run_name_for_compose():
+    docs = (_ROOT / "docs" / "install" / "docker.md").read_text(encoding="utf-8")
+    assert "docker exec <container> python3" in docs
+    assert "torch.cuda.get_device_name(0) if ok else 'unavailable'" in docs
+    for compose_name in ("omnivoice-studio", "omnivoice-studio-gpu", "omnivoice-studio-rocm"):
+        assert compose_name in docs
 
 
 def test_the_dockerfile_copies_the_constraints_file():